WO2019016996A1 - Anomaly detection device, anomaly detection method, and computer program - Google Patents
Anomaly detection device, anomaly detection method, and computer program Download PDFInfo
- Publication number
- WO2019016996A1 WO2019016996A1 PCT/JP2018/007061 JP2018007061W WO2019016996A1 WO 2019016996 A1 WO2019016996 A1 WO 2019016996A1 JP 2018007061 W JP2018007061 W JP 2018007061W WO 2019016996 A1 WO2019016996 A1 WO 2019016996A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- vehicle
- anomaly detection
- anomaly
- conditions
- threshold
- Prior art date
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 196
- 238000004590 computer program Methods 0.000 title claims description 7
- 238000009826 distribution Methods 0.000 claims description 21
- 238000005259 measurement Methods 0.000 claims description 20
- 238000000034 method Methods 0.000 claims description 9
- 238000010586 diagram Methods 0.000 description 20
- 238000003860 storage Methods 0.000 description 20
- 238000012423 maintenance Methods 0.000 description 18
- 238000003066 decision tree Methods 0.000 description 17
- 238000001556 precipitation Methods 0.000 description 14
- 238000012545 processing Methods 0.000 description 8
- 230000008859 change Effects 0.000 description 7
- 238000006243 chemical reaction Methods 0.000 description 5
- 238000004891 communication Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000010248 power generation Methods 0.000 description 4
- 230000001172 regenerating effect Effects 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 230000007704 transition Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000000611 regression analysis Methods 0.000 description 3
- 238000010187 selection method Methods 0.000 description 3
- 238000005299 abrasion Methods 0.000 description 2
- 230000001174 ascending effect Effects 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000004378 air conditioning Methods 0.000 description 1
- 230000002547 anomalous effect Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 238000005401 electroluminescence Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000000704 physical effect Effects 0.000 description 1
- 238000003825 pressing Methods 0.000 description 1
- 238000012628 principal component regression Methods 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B61—RAILWAYS
- B61L—GUIDING RAILWAY TRAFFIC; ENSURING THE SAFETY OF RAILWAY TRAFFIC
- B61L15/00—Indicators provided on the vehicle or train for signalling purposes
- B61L15/0081—On-board diagnosis or maintenance
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B61—RAILWAYS
- B61L—GUIDING RAILWAY TRAFFIC; ENSURING THE SAFETY OF RAILWAY TRAFFIC
- B61L27/00—Central railway traffic control systems; Trackside control; Communication systems specially adapted therefor
- B61L27/50—Trackside diagnosis or maintenance, e.g. software upgrades
- B61L27/57—Trackside diagnosis or maintenance, e.g. software upgrades for vehicles or trains, e.g. trackside supervision of train conditions
Definitions
- Embodiments of the present invention relate to an anomaly detection device, an anomaly detection method, and a computer program.
- anomaly detection is performed through comparison of a value measured by a sensor on a railway vehicle with a threshold.
- a model that recreates a normal operation of a railway vehicle is prepared and used to detect an anomaly or an anomaly sign.
- a travel condition dynamically changes in a temporally sequential manner due to, for example, a route slope, a weather change, passenger loading and unloading, and an operation by a driver.
- Embodiments of the present invention provide an anomaly detection device, an anomaly detection method, and a computer program that achieve highly accurate anomaly detection.
- an anomaly detection device includes: a condition generator, a threshold setter and an anomaly detector.
- the generator generates a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model.
- the threshold setter sets a plurality of thresholds for the conditions.
- the anomaly detector performs anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions.
- FIG. 1 is a block diagram of an anomaly detection system according to an embodiment of the present invention.
- FIG. 2 is a diagram illustrating an exemplary configuration of a brake notch, a brake, and an air spring of a railway vehicle.
- FIG. 3 is a diagram illustrating an exemplary configuration of a power generation brake and a regenerative brake of the railway vehicle.
- FIG. 4 is a diagram illustrating an exemplary table related to measurement information and environment information.
- FIG. 5 is a diagram illustrating an exemplary table related to brake information.
- FIG. 6 is a diagram illustrating an exemplary conversion table.
- FIG. 7 is a diagram illustrating an exemplary model database.
- FIG. 8 is a diagram illustrating an exemplary method of determining a threshold by using a normal distribution.
- FIG. 8 is a diagram illustrating an exemplary method of determining a threshold by using a normal distribution.
- FIG. 9 is a diagram illustrating an exemplary detection result database.
- FIG. 10 is a diagram illustrating an exemplary data set used to generate a travel condition.
- FIG. 11 is a diagram illustrating an exemplary decision tree.
- FIG. 12 is a diagram illustrating another exemplary model database.
- FIG. 13 is a diagram illustrating an exemplary operation of an anomaly detection model.
- FIG. 14 is a diagram illustrating an exemplary main screen output from an anomaly detection device.
- FIG. 15 is a diagram illustrating an exemplary anomaly detailed screen output from the anomaly detection device.
- FIG. 16 is a diagram illustrating an exemplary check screen for a driver.
- FIG. 17 is a diagram illustrating a hardware configuration of the anomaly detection device according to the present embodiment of the present invention.
- FIG. 18 is a flowchart of anomaly detection processing according to an embodiment of the present invention.
- FIG. 19 is a flowchart of processing at the anomaly detection device in a learning mode.
- FIG. 20 is
- FIG. 1 is a block diagram illustrating an exemplary anomaly detection system according to a first embodiment of the present invention.
- the anomaly detection system illustrated in FIG. 1 includes an anomaly detection device 100, a vehicle system 500, an environment information system 600, a terminal 700, an input device 800, and a screen display device 900.
- the outline of the anomaly detection system illustrated in FIG. 1 will be described next.
- the anomaly detection device 100 operates in a learning mode and an operational mode.
- the anomaly detection device 100 includes a function (anomaly detection model generator 200) that creates, in the learning mode, an anomaly detection model related to a brake system of a railway vehicle based on at least one of measurement information of a railway vehicle (hereinafter, a vehicle) and environment information of the vehicle.
- the measurement information is acquired from the vehicle system 500.
- the environment information is acquired from the environment information system 600.
- Travel information of the vehicle according to the present embodiment includes at least one of the measurement information of the vehicle, which is acquired from the vehicle system 500, and the environment information of the vehicle, which is acquired from the environment information system 600.
- the anomaly detection model includes a prediction model for predicting the state of the vehicle, and a threshold related to a residual (difference) from a predicted value of the prediction model.
- the state of the vehicle is, as an example, the deceleration of the vehicle.
- the anomaly detection device 100 includes a function (anomaly detector 110) that performs anomaly detection of the vehicle by using the prediction model and the threshold in the operational mode.
- the anomaly detection is determination of whether there is an anomaly.
- the anomaly detection is also called anomaly determination.
- the anomaly detection is performed by comparing, with the threshold, a residual as the difference between the predicted value of the prediction model and the actual value of the state of the vehicle acquired from the vehicle.
- the anomaly detection device 100 includes a function (threshold setter 220) that sets a threshold for each of a plurality of conditions (hereinafter, travel conditions) based on the travel information of the vehicle.
- the anomaly detection device 100 generates one prediction model, and an anomaly detection model including the thresholds corresponding to the respective travel conditions.
- a threshold corresponding to a travel condition under which the vehicle exists in other words, a travel condition that satisfies travel information (current travel information) as a target of the anomaly detection is selected from among the thresholds included in the anomaly detection model, and used together with the prediction model.
- the creation of a plurality of travel conditions and the setting of a plurality of thresholds are performed in the learning mode.
- the travel information of the vehicle and a result for example, the residual between the predicted value of the prediction model and the actual value of the state value
- the learning mode and the operational mode may be switched automatically or by an instruction from, for example, a maintenance, or may be simultaneously executed.
- the anomaly detection device 100 When having detected an anomaly, the anomaly detection device 100 displays, on the screen display device 900, for example, a place where the anomaly is detected, an anomaly detection model used for the anomaly detection, travel information used for the anomaly detection, and a predicted value of a prediction model. This supports monitoring by a railway operator.
- FIG. 2 illustrates an exemplary configuration of a brake notch, and a brake and an air spring for a particular wheel of the vehicle.
- the brake notch is located in a driver room of a train in reality.
- the following describes the brake system of a vehicle, as a target of anomaly detection by the anomaly detection device 100, with reference to FIG. 2.
- a brake lever 10 as an exemplary controller provides a driver with a means for a brake operation.
- the driver brakes the vehicle by moving the brake lever upward.
- the numbers of one to eight that are indicated on the brake lever 10 represent brake notches (brake steps), and a larger number means a stronger braking force applied to the vehicle. This number of notches is exemplary and does not prevent the vehicle from using a larger or smaller number of notches.
- Each brake notch is an exemplary control command value to the vehicle or the brake.
- the brake operation to the vehicle is not limited to an operation performed by a driver.
- the brake operation is performed by the device in place of a driver in some cases.
- a braking command output from the device corresponds to the control command value.
- FIG. 2 illustrates a wheel 30 of a vehicle traveling on a rail 20.
- One means for braking the vehicle with a brake is a tread brake 42.
- a tread brake 42 is provided in this example, only one wheel is illustrated for simplicity of description, but in reality, a plurality of pairs of right and left wheels are provided.
- the tread brake 42 uses an air cylinder as a power source.
- brake cylinder pressure that is the pressure inside an air cylinder 43
- a brake block 41 is pressed against a tread as a surface of the wheel 30, which contacts with the rail. Frictional force between the wheel 30 and the brake block 41 functions as braking force of the tread brake 42.
- the tread brake uses the frictional force of the brake block in this manner, the brake block is abraded by continuous use, which potentially decreases the braking force.
- the tread brake is an exemplary mechanical brake used for a vehicle, and another scheme uses a disk brake that obtains braking force by pressing, with a pad or the like, a disk fixed to a wheel shaft against a wheel.
- the braking force of a brake varies with an abrasion state of, for example, the brake block or the pad.
- a worker or the like may check, for example, the brake block or the disk of the brake system and check actual existence of the anomaly.
- the braking force of a brake also varies with a load on a vehicle in addition to the abrasion of a component thereof.
- a load responsive device 50 is mounted on the vehicle illustrated in FIG. 2.
- the load responsive device 50 includes an air spring 51 and measures a load on the vehicle by sensing air spring pressure of the air spring 51.
- the braking force of the brake may be adjusted in response to the air spring pressure detected by the load responsive device 50. Accordingly, desired deceleration can be achieved irrespective of change of a load on the vehicle.
- FIG. 3 illustrates an exemplary configuration of a power generation brake and a regenerative brake of a vehicle.
- Main electric motors 60a and 60b are mounted on the vehicle illustrated in FIG. 3. When the power generation brake is used, the main electric motors 60a and 60b form a closed circuit with a resistor 70 to convert electrical power of the main electric motors into thermal energy.
- the regenerative brake When the regenerative brake is used, electrical power generated by the main electric motors 60a and 60b is transmitted to a line 90 through a pantograph 80. When a secondary battery is mounted on the vehicle, the generated electrical power may be used to charge the secondary battery. In this manner, the regenerative brake obtains braking force through conversion of kinetic energy into electrical power by using the main electric motors 60a and 60b as electric generators.
- the mechanical brake and the power generation brake are exemplary, and the anomaly detection device 100 can perform anomaly detection on a brake of any other scheme used in the brake system.
- brakes of a plurality of schemes having different characteristics are used in the brake system of a vehicle.
- the braking force of the brake system of a vehicle varies with a load as described above.
- the number of passengers largely varies with a time slot and an operation interval, and accordingly, the braking force of the brake system largely varies in a short duration.
- a load largely varies with the amount of cargo.
- deceleration when the brake operation is performed potentially varies between travel routes and intervals of a vehicle, which have different slope and cant tendencies.
- any difference in weather conditions of precipitation, atmospheric temperature, and the like potentially changes the physical property of a component of the brake system, affecting the characteristic of the brake system.
- drivers perform the brake operation in different manners, and vehicles are manufactured with different brake characteristics.
- highly accurate anomaly detection is easily performed by using a threshold appropriate for situations inside and outside of the vehicle by switching thresholds depending on a travel condition under which a vehicle exists. In this manner, the risk of false anomaly detection is reduced to achieve early anomaly detection and safe and reliable railway operation.
- a brake system as a target of the anomaly detection by the anomaly detection device 100 may be a brake device for a particular wheel of an optional railway vehicle, all of a plurality of brake devices in the entire vehicle, or the entire group of brake devices in a plurality of vehicles or cars in a train.
- the anomaly detection target is not limited to a brake system, but may be a power system, an air-conditioning system, or a power generating system.
- the anomaly detection target is not limited a railway vehicle but may be an optional vehicle including a wheel, such as an automobile, a construction machine, and an aircraft.
- the anomaly detection target also includes any device or system other than a vehicle.
- the anomaly detection device 100 includes a vehicle information collector 101, an environment information collector 102, a data processor 103, the anomaly detection model generator 200, a condition generator 230, the anomaly detector 110, an alarm 120, and a screen generator 130.
- the anomaly detection model generator 200 includes a model generator 210 and the threshold setter 220.
- the vehicle information collector 101 acquires measurement information (also referred to as measure data) related to a vehicle from various sensors of the vehicle system 500 inside of the vehicle.
- the sensors include a sensor configured to detect the brake operation of the vehicle or the like as a control command value, a sensor configured to detect deceleration of the vehicle, a sensor configured to detect a driving speed, and a sensor configured to measure a load applied on the vehicle. Other various sensors may be included.
- the measurement information includes a detected value (control command value, for example) of a sensor, and a measured value (driving speed, load, or deceleration, for example) of a sensor. When the vehicle system 500 calculates deceleration from the value of a speed sensor, the calculated deceleration may be acquired as part of the measurement information.
- the kind of the measurement information to be acquired may be optionally set.
- the measurement information may be acquired in an optionally set period.
- the measurement information related to the driving speed of the vehicle is acquired in a short sampling period in milliseconds.
- the value of the sensor configured to measure a load applied to the vehicle is acquired in a sampling period in minutes.
- the environment information collector 102 acquires environment information (also referred to as travel environment data) of the vehicle from the environment information system 600.
- the environment information include information related to an operation route and information related to weather.
- Examples of the information related to an operation route include a slope and a cant (height difference between inner and outer rails of railway) for each interval.
- Examples of the information related to weather include weather, the atmospheric temperature, the amount of precipitation, the speed of wind, and the atmospheric pressure.
- the acquisition of the environment information may be acquisition of information accumulated in a database in a ground system, or acquisition of information distributed from an external server.
- the kind of the environment information to be acquired and the frequency of the acquisition may be optionally set.
- the anomaly detection device 100 may be installed as a ground device outside of the vehicle, for example, in a facility or an operation command center of a railway operation management company, or may be installed as an on-board device on the vehicle.
- the anomaly detection device 100 is not limited to a particular installation manner.
- the anomaly detection device 100 When the anomaly detection device 100 is installed as a ground device outside of the vehicle, the measurement information of the vehicle system 500 inside of the vehicle and the like are received through, for example, an on-board element, a transponder ground element, and a ground information network. Specifically, the vehicle system 500 transmits data to the ground information network through the ground element and the like, and the anomaly detection system receives the data through the ground information network.
- the ground information network may use, for example, a metallic cable, a coaxial cable, an optical cable, a phone line, a wireless device, or Ethernet (registered trademark), but is not limited to a particular scheme.
- the anomaly detection device 100 receives data from the environment information system 600 through the ground information network.
- the anomaly detection device 100 acquires data from the vehicle system 500 through an information network inside of the vehicle.
- the information network inside of the vehicle is, for example, Ethernet or a wireless local area network (LAN), but may be achieved in any other scheme.
- the anomaly detection device 100 may use the on-board element and the transponder ground element to acquire data from the environment information system 600 connected with the ground information network.
- the input device 800 provides an interface for an operation by a maintenance person.
- the input device 800 includes a mouse, a keyboard, a voice recognition system, an image recognition system, a touch panel, or a combination thereof.
- the maintenance person can perform an operation by inputting various commands or data to the anomaly detection device 100 through the input device 800.
- the screen display device 900 displays, as a still image or a moving image, data or information output from the anomaly detection device 100.
- the screen display device 900 is, for example, a liquid crystal display (LCD), an organic electroluminescence display, or a vacuum fluorescence display (VFD), but may be a display device in any other scheme.
- Each of the input device 800 and the screen display device 900 may be one of a plurality of installed devices.
- the input device 800 and the screen display device 900 may be installed at each of the operation command center and an operation table of the vehicle.
- the input device 800 and the screen display device 900 may be one integrated device.
- a single device may serve as the input device 800 and the screen display device 900.
- the anomaly detection device 100 includes an information database 310, a model database 320, and a detection result database 330.
- the databases 310, 320, and 330 are all arranged inside of the anomaly detection device 100 illustrated in FIG. 1.
- the arrangement of the databases is not limited to a particular method.
- part of the databases may be arranged in an external server or storage device.
- Each database can be implemented by a relational database management system and various NoSQL systems, but may be implemented in any other scheme.
- Each database may employ a storage format of XML, JSON, or CSV, or any other format such as a binary format.
- Not all databases inside of the anomaly detection device 100 need to be achieved by an identical database system and an identical storage format, but the databases may be achieved in a mixture of a plurality of schemes.
- the information database 310 stores the measurement information acquired by the vehicle information collector 101 and the environment information acquired by the environment information collector 102.
- a storage medium such as a memory device storing the measurement information and the environment information may be inserted into the anomaly detection device 100 and used as the information database 310.
- FIGS. 4 and 5 illustrate exemplary information databases 310.
- the travel information (the measurement information and the environment information) is stored in forms of a table 310a illustrated in FIG. 4 and a table 310b illustrated in FIG. 5.
- the "data ID” column of the table 310a illustrated in FIG. 4 stores an identification number of an entry stored in the table 310a.
- a data ID serves as a primary key.
- Each data ID is associated with the table 310b as illustrated in FIG. 5.
- the table 310b is stored in the information database 310.
- the "time" column stores the generation time of an entry. In this example, an entry is generated in each constant sampling time. However, an entry may be generated at an interval set to a rail track in advance or in any other scale.
- the "driver" column of the table 310a stores the name of a driver who performed a brake operation.
- a device such as the ATS, the ATC, or the ATO instead of a driver
- the name of a device that performed the operation or an identifier indicating the device may be stored instead.
- the "weather" column of the table 310a stores information related to weather acquired from the environment information system 600.
- the "atmospheric temperature” column of the table 310a stores information related to the atmospheric temperature acquired from the environment information system 600.
- the information related to the atmospheric temperature may be an actual value or a label classifying the actual value.
- the "atmospheric temperature” column stores the label of any one of classes T1, T2, T3, T4, T5, T6, and T7 into which the atmospheric temperature as a real number is converted by using a conversion table 310c illustrated in FIG. 6.
- the atmospheric temperature at -11°Cis converted into class T1 the atmospheric temperature at 15°C is converted into class T4, and the atmospheric temperature at 33°C is converted into class T6.
- data obtained by converting a class name into an optional integer for example, the class T1 into 1, the class T2 into 2, and the class T3 into 3, may be used as explanatory variables.
- the information database 310 may store processed information obtained by performing calculation or conversion on the measurement information or the environment information.
- the "boarding rate" column of the table 310a stores a boarding rate in percentage as an index for a load applied to the vehicle. Another index may be used to indicate a load.
- the boarding rate is defined with, for example, a ratio of the capacity of a passenger vehicle and the number of passengers in the passenger vehicle.
- the boarding rate is often estimated based on the air spring pressure of the load responsive device. In such a case, the air spring pressure may be directly used as the index.
- the air spring pressure is the actual value of a sensor, which is not a value estimated indirectly by using, for example, a conversion table or a formula unlike the boarding rate, and thus can be used to reduce residual in model generation.
- the air spring pressure takes a value depending on the manufacturer and the model number of the load responsive device mounted on the vehicle, and thus lacks in versatility.
- the difference between vehicles due to different load responsive devices thereof can be absorbed in some cases when a typically used index such as the boarding rate is used.
- the "slope" column of the table 310a stores the slope of a route in a value expressed in the unit of permil.
- the permil is a value obtained by expressing a height difference for the horizontal distance of 1000 m in meters.
- the permil is exemplary, and the "slope" column may store a value in another unit.
- the "cant" column of the table 310a stores a cant in millimeters, but may store a value in another unit.
- the table 310a further includes the "wind speed” column and the “atmospheric pressure” column.
- the table 310a may include columns storing other information such as the current position and the current interval on the rail track.
- the table 310b illustrated in FIG. 5 stores information such as time information, a brake notch, and the actual value of deceleration for a corresponding entry in the table 310a illustrated in FIG. 4.
- the table 310b corresponds to the data ID of 2560 in FIG. 4.
- Entries in the table 310b are generated in a time interval shorter than that for the table 310a.
- the generation interval (sampling interval) of entries in the table 310b may be same as that for the table 310a.
- the table 310b and the table 310a are separately provided in this example but may be integrated with each other.
- Data stored in the information database 310 may be processed.
- the data processor 103 causes the screen display device 900 to display the content of each table stored in the information database 310.
- a maintenance person or a driver performs a fabrication operation on data by using the input device 800.
- the data processor 103 performs data fabrication in accordance with the fabrication operation.
- the interval of acquisition of information or data by the vehicle information collector 101 or the environment information collector 102 may be adjusted.
- the data processor 103 receives an operation to specify the acquisition interval from a maintenance person or a driver through the input device 800, and adjusts the acquisition interval in accordance with the content of the operation.
- the anomaly detection model generator 200 creates an anomaly detection model of the brake system of the vehicle by using data stored in the information database 310.
- the anomaly detection model includes a prediction model and one or a plurality of thresholds.
- the prediction model is generated by the model generator 210, and each threshold is generated by the threshold setter 220.
- the generated anomaly detection model is stored in the model database 320.
- FIG. 7 illustrates an exemplary model database 320.
- the model database 320 can store one or a plurality of anomaly detection models. Each anomaly detection model is identified by a model ID.
- the column of prediction model stores data indicating a prediction model or a memory address (pointer) at which the prediction model is stored.
- the data indicating a prediction model includes, for example, a deceleration model.
- Each cell in the column of threshold stores one or a plurality of thresholds. When a plurality of thresholds are stored, a travel condition corresponding to each threshold is stored together (to be described later in detail).
- the generation of an anomaly detection model is performed in the learning mode, for example, when the anomaly detection device 100 is started or when a system is newly added as an anomaly detection target.
- an anomaly detection model is generated for each system.
- the anomaly detection model is created by using a data sample (feature vector) extracted from the information database 310.
- the data sample (feature vector) includes one or more explanatory variables.
- the explanatory variables include the value (control command value) of a brake notch in the table 310b. Additionally, a value (such as speed) of any other kind in the travel information and specifications of the vehicle (such as the sizes and weight of the vehicle) may be used as explanatory variables.
- Explanatory variables may be generated through calculation with a plurality of items included in the travel information. In this example, the objective variable of a prediction model is deceleration.
- the data sample may be generated for each entry in the brake information table 310b, or a plurality of continuous entries with lower temporal granularity may be collected to generate one data sample.
- the model generator 210 performs multiple regression analysis to obtain Formula (1) for predicting deceleration as the objective variable.
- y represents the objective variable
- x n represents an explanatory variable
- b n represents a partial regression coefficient.
- a standard partial regression coefficient may be used as the partial regression coefficient b n by normalizing the objective variable and all explanatory variables to the average value of 0 and the dispersion of 1.
- the number of explanatory variables may be one or more.
- model generation by the multiple regression analysis is exemplary, and a prediction model for the objective variable may be created by any other method such as support vector regression or autoregression.
- cross validation may be used.
- a data sample can be divided into a plurality of sets so that at least one of the sets is used as test data for validation and the other sets are used for model creation. This allows the performance of a generated model to be checked.
- an information database stores information acquired when a brake system on which anomaly detection is performed is in the normal state.
- a generated prediction model is a model obtained by modeling the behavior of the brake system of the vehicle in the normal state.
- the information database may store information when a particular brake is in failure.
- the number of explanatory variables used to create a prediction model may be reduced by adjusting the number of explanatory variables by, for example, a variable selection method or principal components analysis. This adjustment is effective when there is a correlation between different explanatory variables or when reduction of a calculation time and a processing load is required.
- a model is generated by selecting a subset of explanatory variables effective for prediction from among a set of all explanatory variables.
- Useful explanatory variables may be selected by first generating a model by using one or a small number of explanatory variables and then generating a model to which explanatory variables are added one by one.
- useful explanatory variables may be specified by first generating a model with a large number of explanatory variables and then generating a model obtained by removing explanatory variables one by one.
- the selection of explanatory variables may be performed by using a genetic algorithm.
- variable selection method When the brake system of a vehicle is an anomaly detection target and the deceleration of a brake is the objective variable of a prediction model, for example, a prediction model using only a brake notch as an explanatory variable is created first. The brake notch is presumed to have a highest correlation with the deceleration. Thereafter, any other explanatory variable such as the travel speed of the vehicle, which is thought to have a correlation with the deceleration, is sequentially added to the prediction model, and the accuracy of prediction is checked. An explanatory variable when a desired prediction accuracy is obtained is employed.
- the threshold is used to determine that there is an anomaly when the difference between a predicted value of the objective variable (in this example, a predicted value of the deceleration), which is calculated based on a prediction model, and a measured value (actual value) of the deceleration exceeds the threshold.
- the determination that there is an anomaly is also referred to as detection of an anomaly.
- the difference between the actual value and the predicted value is also referred to as residual.
- the actual value can be larger or smaller than the predicted value, and thus the value of the residual can be positive or negative.
- the residual may be defined to be the absolute value of the difference, which is the absolute value of the distance from the predicted value.
- FIG. 8 is used to describe an exemplary threshold determination method using a normal distribution.
- FIG. 8 illustrates the graph of a normal distribution 400 of residual.
- the horizontal axis represents the residual, and the vertical axis represents a probability density.
- a plurality of residuals between the predicted value of a prediction model and the actual value are acquired to create the normal distribution 400, assuming that these residuals obey a normal distribution.
- Data used to acquire the residuals may be a data sample used to generate the prediction model, test data, travel information unrelated to generation of the prediction model, or an optional combination thereof.
- the distribution has a more spread skirt as illustrated with normal distributions 401 and 402 illustrated with a dashed line.
- the normal distribution 400 is used to set a threshold for the prediction model. For example, when the standard deviation is represented by ⁇ , a multiple of the standard deviation, such as 2 ⁇ or 3 ⁇ , is set as the threshold. When the residual exceeds the threshold that is set to be 2 ⁇ , an anomaly is detected in anomaly detection. When such a threshold is set, about 95% of the actual values are determined to have no anomaly (be normal). As another exemplary threshold setting, the value of the residual corresponding to a predetermined probability (for example, a highest X percent point or a lowest X percent point) or the absolute value thereof may be set as the threshold.
- a predetermined probability for example, a highest X percent point or a lowest X percent point
- the absolute value thereof may be set as the threshold.
- the above-described threshold determination method is exemplary, and any other method may be used. For example, the threshold may be determined by assuming a distribution other than a normal distribution or by a person such as a maintenance person or a driver based on experience thereof
- the anomaly detector 110 performs anomaly detection on a target system by using an anomaly detection model (a prediction model and a threshold) stored in the model database 320.
- a feature vector is generated from travel information used for the anomaly detection, and a deceleration is predicted by using the generated feature vector and the prediction model.
- the residual between the predicted deceleration and the actual deceleration is compared with the threshold. When the residual is equal to or smaller than the threshold, it is determined that the deceleration is normal. When the residual is larger than the threshold, it is determined that the deceleration is anomalous.
- the anomaly detector 110 stores information in the detection result database 330 based on a result of the anomaly detection.
- the anomaly detector 110 notifies information related to the result of the anomaly detection to the screen generator 130 and the alarm 120.
- FIG. 9 illustrates an exemplary detection result database.
- the database stores the brake notch, the actual value of the deceleration, the predicted value of the deceleration based on a prediction model (ID0001), and anomaly detection result (existence or non-existence on anomaly) in a temporally sequential manner. In the example illustrated in FIG. 9, no anomaly is detected at any timing. Travel information (for example, the items of driver, weather, atmospheric temperature, boarding rate, slope, cant, wind speed, and atmospheric pressure) at a corresponding time may be added to the detection result database.
- Travel information for example, the items of driver, weather, atmospheric temperature, boarding rate, slope, cant, wind speed, and atmospheric pressure
- the threshold setter 220 can set, to a prediction model, a plurality of thresholds in accordance with a travel condition.
- the anomaly detector 110 specifies a travel condition satisfied by the current travel information. Then, anomaly detection is performed by using a threshold corresponding to the specified travel condition.
- the travel condition needed by the threshold setter 220 to set a plurality of thresholds is generated by the condition generator 230.
- the condition generator 230 generates, by using the detection result database 330 and the travel information used for anomaly detection, a plurality of travel conditions (a plurality of conditions) for classifying the difference between the predicted value of a prediction model and the actual value in accordance with the value of the difference.
- a plurality of travel conditions a plurality of conditions for classifying the difference between the predicted value of a prediction model and the actual value in accordance with the value of the difference. The following describes operation of the condition generator 230 in detail.
- the condition generator 230 creates, by using the detection result database 330 and the travel information, a classifier (for example, a decision tree) for predicting a class in accordance with the residual.
- a classifier for example, a decision tree
- Each residual between the predicted value and the actual value in the detection result database is classified into a plurality of classes (residual classes). For example, the residual that is equal to or smaller than a threshold A is classified into a residual class A for a small residual, the residual that is larger than the threshold A and smaller than a threshold B is classified into a residual class B for an intermediate residual, or the residual that is equal to or larger than the threshold B is classified into a residual class C for a large residual.
- the threshold A may be a threshold set first to the prediction model but is not limited thereto. Another exemplary classification will be described next. Residuals are arranged in ascending or descending order.
- a data set of smallest 20% residuals is classified into the residual class A
- a data set of next smallest 60% residuals is classified into the residual class B
- a data set of the remaining 20% residuals is classified into the residual class C.
- the ratio of the classes may be optionally determined and is not limited to the above-described values.
- the condition generator 230 selects, for each entry in the detection result database 330 as illustrated in FIG. 9, any one of the residual classes A to C in accordance with the residual of the entry, and allocates the selected residual class to the entry.
- the condition generator 230 specifies travel information corresponding to each entry from the table 310a illustrated in FIG. 4, and associates the travel information with the entry. However, this operation is unnecessary when the travel information is already included in the detection result database 330.
- a data set is generated in which each entry in the detection result database is associated with the residual class and the travel information.
- the data set may be stored in an internal buffer of the condition generator 230, or may be stored in a storage device or database not illustrated. In the data set illustrated in FIG.
- a decision tree is generated by using the data set as learning data, specifically, by setting the residual class as the objective variable and the other items as explanatory variables. Any item unnecessary for producing a data set, in databases (those illustrated in FIGS. 9 and 4, for example) from which the data set is created, does not need to be used. For example, when unnecessary, the item of anomaly detection result does not need to be included in the data set illustrated in FIG. 10.
- the data set includes no data in which an anomaly is detected. However, when there is data in which an anomaly is detected, this data may be excluded from the data set.
- the anomaly detector 110 detects an anomaly, whether this detection result is correct may be checked by a maintenance person. For example, when an anomaly is detected, the screen display device 900 displays a check screen (refer to FIG. 16 to be described later) for checking whether this detection result is correct.
- the maintenance person inputs an instruction for the false detection.
- the condition generator 230 or another processor corrects the detection result in the detection result database 330 based on this instruction. Data in which an anomaly is detected may be excluded from the data set at creation of a decision tree.
- FIG. 11 illustrates an exemplary generated decision tree. This decision tree predicts a residual class corresponding to the objective variable based on two explanatory variables related to the amount of precipitation and the boarding rate.
- Nodes 1001a, 1001b, and 1001c are end nodes corresponding to the objective variable.
- the top node is called a root node. Any node other than the end nodes and the root node is called an intermediate node.
- the root node and the intermediate node are explanatory variable nodes.
- the end nodes are residual class nodes (objective variable nodes).
- the nodes 1001a, 1001b, and 1001c correspond to the residual class A, the residual class B, and the residual class C, respectively.
- This decision tree classifies the residual into the residual class A for "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%", classifies the residual into the residual class B for “sunny weather (no precipitation) and the boarding rate higher than 90%”, or classifies the residual into the residual class C for "rainy weather (no precipitation)”.
- the condition generator 230 acquires, as a travel condition corresponding to each residual class, a condition included in a path from the residual class node (end node) to the root node. Specifically, the condition generator 230 acquires the condition "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%" corresponding to a path from the residual class A to the root node, the condition "sunny weather (no precipitation) and the boarding rate higher than 90%” corresponding to a path from the residual class B to the root node, and the condition "rainy weather (no precipitation)" corresponding to a path from the residual class C to the root node, as travel conditions A, B, and C, respectively.
- the threshold setter 220 sets a threshold to each of the travel conditions A to C. Specifically, the threshold setter 220 classifies travel information used to generate a prediction model (or travel information not used to generate the prediction model) into groups A to C that satisfy the travel conditions A to C. Anomaly detection is performed on the travel information classified into the group A, and the residual is calculated based on a result of the detection. The distribution (for example, normal distribution) of the residual is calculated and used to set the threshold A (refer to the description with reference to FIG. 8). In this manner, the threshold A is set for the travel condition A. Similarly, the threshold B corresponding to the travel condition B is set to the group B, and the threshold C corresponding to the travel condition C is set to the group C.
- the distribution of the residual corresponding to the group A has a small standard deviation, and thus the threshold A is small.
- the distribution of the residual corresponding to the group C has a large standard deviation, and thus the threshold C is large.
- the threshold B corresponding to the group B is between the threshold A and the threshold C.
- the threshold A is used when the weather is sunny and the boarding rate is equal to or lower than 90%.
- the threshold B is used when the weather is sunny and the boarding rate is higher than 90%.
- the threshold C is used when the amount of precipitation is larger than zero.
- the thresholds are switched in accordance with such a travel condition at anomaly detection, thereby achieving the anomaly detection accurately based on a brake characteristic.
- the amount of precipitation and the boarding rate are used as travel conditions above, for example, other items of atmospheric temperature and humidity may be used as travel conditions.
- Examples of algorithms used for learning of a decision tree include ID3 and C4.5, but any algorithm may be used. Pruning of a decision tree may be performed to prevent noise and overlearning.
- the learning of a decision tree is exemplary, and any other classifier may be used. Which item of explanatory variable is employed in a decision tree among a plurality of items of travel information depends on an algorithm and learning data to be used.
- a maintenance person may employ professional knowledge to set a threshold for each travel condition.
- Each pair of a travel condition generated by the condition generator 230 and a threshold set by the threshold setter 220 are stored in a corresponding cell in the column of threshold in the model database 320.
- FIG. 12 illustrates an example in which a plurality of thresholds and corresponding travel conditions are stored for the model 0001.
- the anomaly detector 110 specifies a travel condition that satisfies current travel information among the travel conditions. Then, the anomaly detector 110 performs anomaly detection by using thresholds and a prediction model corresponding to the specified travel condition.
- the following describes a specific example of operation of the anomaly detector 110 in this case.
- FIG. 13 is a diagram for description of exemplary operation of the anomaly detector 110.
- An upper part of FIG. 13 illustrates the brake notch.
- a middle part thereof illustrates the brake deceleration.
- a lower part thereof illustrates the residual between the actual value and the predicted value of the prediction model.
- the brake notch corresponds to an explanatory variable
- the deceleration corresponds to the objective variable.
- an operation to set the brake notch to notch "4" is performed. Having received this operation, the brake system applies braking force to the vehicle. Accordingly, the deceleration of the vehicle increases and thereafter stays near a constant value.
- a predicted value of the deceleration and a measured value (actual value) thereof have a slight difference therebetween but transition substantially at the same value, and the residual between the predicted value and the actual value is smaller than a threshold ⁇ .
- the threshold ⁇ corresponds to the travel condition A.
- the travel condition satisfied by the travel environment of the vehicle changes from A to B.
- the anomaly detector 110 detects the change of travel condition and changes a used threshold to ⁇ .
- Travel environment and change thereof can be detected from a measured value and a control command value included in measurement information, and route data and weather data included in environment information.
- travel environment and change thereof may be detected based on, for example, an explicit command from a driver or a command center, or a radio signal received from a ground element.
- the predicted value of the deceleration is constant, but the actual value varies largely as compared to that between times t1 and t2. Accordingly, the residual exceeds the threshold ⁇ at three timings, and the anomaly detector 110 detects an anomaly at each timing.
- the anomaly detector 110 changes the used threshold from ⁇ to ⁇ at time t4. In the interval between times t4 and t5, the residual is within the range of the threshold ⁇ , and thus no anomaly is detected.
- an operation to cancel the brake is performed. Having received this operation, the brake system further reduces the braking force applied to the vehicle, and accordingly the deceleration of the vehicle further decreases. After time t5, the residual is within the range of the threshold ⁇ , and thus no anomaly is detected.
- the alarm 120 notifies the detection of an anomaly by the anomaly detector 110 to the terminal 700 used by a railway operator, driver, or maintenance person.
- This notification may be performed by transmission of an electronic mail, display of a pop-up message on an operation screen of the terminal 700, or notification by predetermined instrument management protocol, or may be performed by any other means.
- the notification may include detailed information (for example, a place (current value) on a map where the anomaly has occurred, or an identifier of the vehicle to which the anomaly has occurred) of the anomaly.
- An operator or a maintenance person can know the detection of the anomaly and details thereof by receiving the notification.
- the screen generator 130 displays, on the screen display device 900, for example, an anomaly detection result, a current position of the vehicle if an anomaly is detected, an anomaly detection model and a threshold used for the anomaly detection, sensor data, and a predicted value based on a prediction model.
- the screen generator 130 may be included in the anomaly detection device 100 or may be included in a vehicle information system connected with the anomaly detection device 100, or a terminal or a management server on an information network of a ground system.
- FIG. 14 illustrates an exemplary main screen 901 displayed by the screen generator 130.
- the main screen 901 displays information on a plurality of trains.
- the screen display device 900 is installed in a command room for managing and monitoring the trains.
- Information related to vehicle is displayed in a table format at an upper part of the main screen 901.
- Examples of display items include the items of train, anomaly detection result, line number (train identifier), boarding rate, and current position, but any other information may be displayed.
- the "anomaly" column in the table displays an anomaly detection result.
- An exclamation mark "! indicates that an anomaly is detected. Thus, it is indicated that an anomaly is detected at train B.
- This display of an anomaly detection result is exemplary and may be achieved in any other manner.
- a map is displayed at a lower part of the main screen 901, indicating the current position of each train.
- the detection of the anomaly and the name of a used model are displayed in a word balloon for train B at which the anomaly is detected.
- Click on a train of interest on the main screen 901 illustrated in FIG. 14 performs transition to an anomaly detailed screen.
- the scheme of screen transition is not limited thereto but may employ any other scheme such as a predetermined keyboard operation.
- FIG. 15 illustrates an exemplary anomaly detailed screen 902 to which transition is made by clicking train B.
- graphs same as those illustrated FIG. 13 are displayed at a right part of the screen. Specifically, a graph of the brake notch, a graph of the predicted value of the deceleration, a graph of the actual value of the deceleration, and a graph of the residual are displayed. Each graph is displayed for a particular duration that includes a time when an anomaly is detected. A long bar indicating that a threshold is exceeded is displayed to allow visual check of a time when an anomaly is detected.
- a check box is provided at a left part of the screen illustrated in FIG. 15 to allow selection of an item, the graph of which is to be displayed. Another separate means may be provided to allow specification of the time range of the graph display. When such an interface as described above is provided, a railway operator or the like can understand details of the anomaly and take fast measures.
- a check screen may be presented to a maintenance person to allow the maintenance person to check whether a result of the detection is correct.
- FIG. 16 illustrates an exemplary check screen 903.
- the maintenance person inputs an instruction to correct the result.
- the condition generator 230 corrects a detection result in the detection result database based on this instruction.
- FIG. 17 illustrates a hardware configuration of the anomaly detection device according to the present embodiment.
- the anomaly detection device according to the present embodiment is achieved by a computer device 100.
- the computer device 100 includes a CPU 151, an input interface 152, a display device 153, a communication device 154, a main storage device 155, and an external storage device 156. These components are connected with each other through a bus 157.
- the central processing unit (CPU) 151 executes an anomaly detection program as a computer program on the main storage device 155.
- the anomaly detection program is a computer program that achieves each above-described functional component of the anomaly detection device. Each functional component is achieved by the CPU 151 executing the anomaly detection program.
- the input interface 152 is a circuit for inputting, to the anomaly detection device, an operation signal from an input device such as a keyboard, a mouse, or a touch panel.
- the display device 153 displays data or information output from the anomaly detection device.
- the display device 153 is, for example, a liquid crystal display (LCD), a cathode-ray tube (CRT), or a plasma display (PDP), but is not limited thereto.
- the data or information output from the computer device 100 can be displayed by the display device 153.
- the communication device 154 is a circuit that allows the anomaly detection device to communicate with an external device in a wireless or wired manner. Measurement information can be input from the external device through the communication device 154. The measurement information input from the external device can be stored in the information database 310.
- the main storage device 155 stores, for example, the anomaly detection program, data necessary for executing the anomaly detection program, and data generated through execution of the anomaly detection program.
- the anomaly detection program is loaded onto the main storage device 155 and executed.
- the main storage device 155 is, for example, a RAM, a DRAM, or an SRAM, but is not limited thereto.
- the information database 310, the model database 320, and the detection result database 330 may be constructed on the main storage device 155.
- the external storage device 156 stores, for example, the anomaly detection program, data necessary for executing the anomaly detection program, and data generated through execution of the anomaly detection program. These program and data are read onto the main storage device 155 at execution of the anomaly detection program.
- the external storage device 156 is, for example, a hard disk, an optical disk, a flash memory, or a magnetic tape, but is not limited thereto.
- the information database 310, the model database 320, and the detection result database 330 may be constructed on the external storage device 156.
- the anomaly detection program may be installed on the computer device 100 in advance or may be stored in a storage medium such as a CD-ROM. Alternatively, the anomaly detection program may be uploaded on the Internet.
- the computer device 100 may include one or a plurality of CPUs 151, one or a plurality of input interfaces 152, one or a plurality of display devices 153, one or a plurality of communication devices 154, and one or a plurality of main storage devices 155, and may be connected with a peripheral instrument such as a printer or a scanner.
- a peripheral instrument such as a printer or a scanner.
- the anomaly detection device may be achieved by the single computer device 100 or may be configured as a system including a plurality of computer devices 100 connected with each other.
- FIG. 18 is a flowchart of anomaly detection processing performed in the operational mode according to the embodiment of the present invention.
- the processing of the flowchart illustrated in FIG. 18 may be executed upon a certain operation of an anomaly detection target system, may be executed in a constant period, may be executed upon reception of an instruction from a user such as a maintenance person, or may be executed at any other timing.
- the anomaly detector 110 acquires travel information as an anomaly detection target from the information database 310.
- the anomaly detector 110 selects a prediction model corresponding to an anomaly detection target system (in this example, the brake system of a vehicle) from the model database 320.
- the anomaly detector 110 also selects a threshold corresponding to a travel condition satisfied by the acquired travel information among a plurality of travel conditions.
- the prediction model is a model for predicting an objective variable representing the state (for example, the deceleration) of the vehicle from an explanatory variable representing a control command value (for example, the brake notch) to the vehicle.
- the prediction model is a model that associates an explanatory variable representing the control command value to the vehicle with an objective variable representing the state of the vehicle.
- the anomaly detector 110 generates a feature vector from the acquired travel information. For example, the anomaly detector 110 generates a feature vector including the control command value. The number of elements of the feature vector may be one or more.
- the anomaly detector 110 predicts the objective variable (in this example, the deceleration) based on the feature vector and the prediction model. In other words, the anomaly detector 110 calculates a predicted value of the state of the vehicle based on the control command value and the prediction model.
- the anomaly detector 110 calculates a residual as the difference between the predicted deceleration and a deceleration included in the travel information, and compares the calculated residual with the threshold.
- the anomaly detector 110 detects an anomaly and outputs information notifying the detection of the anomaly to the screen display device 900 or the like (S105).
- the anomaly detector 110 detects no anomaly (S106). In other words, the anomaly detector 110 determines that the brake system of the vehicle is normal. When no anomaly is detected, information notifying that the brake system of the vehicle is normal may be output to the screen display device 900 or the like.
- FIG. 19 is a flowchart of threshold setting processing at the anomaly detection device in the learning mode. This processing may be executed in a constant period, may be executed at a timing instructed by a maintenance person, or may be executed at any other timing.
- the following describes an exemplary operation when a plurality of thresholds in accordance with a travel condition is set to a prediction model. Assume that the anomaly detector 110 performs anomaly detection based on a prediction model generated in advance and one threshold and the detection result database 330 stores data related to anomaly detection.
- the condition generator 230 allocates a residual class to the residual between the predicted value and the actual value in accordance with the value of the residual based on the detection result database 330.
- the condition generator 230 generates a data set in which the residual class is associated with travel information (refer to FIG. 10).
- the condition generator 230 sets each item of the data set to be an explanatory variable and the residual class to be an objective variable, and performs, for example, machine learning to generate a classifier that predicts the objective variable from at least one of a plurality of explanatory variables. Specifically, the condition generator 230 generates a classifier that associates a plurality of conditions related to at least one explanatory variable with a plurality of residual classes. In this example, a decision tree (refer to FIG. 11) is generated as the classifier.
- the condition generator 230 acquires the conditions included in the classifier as a plurality of travel conditions.
- a condition included in a path from each residual class node (end node) to the root node is acquired as the travel condition corresponding to a residual class.
- the threshold setter 220 sets a plurality of thresholds to the travel conditions. For example, the threshold setter determines a threshold based on the distribution of the residual classified into a residual class that satisfies each travel condition of the travel information. For example, travel information used to generate the decision tree (or travel information not used to generate the decision tree) is classified into groups that satisfy the travel conditions. Anomaly detection is performed for each group, and the residual is calculated based on a result of the detection. Then, a probability distribution (refer to FIG. 8) of the residual is generated. A value of the residual corresponding to a predetermined probability (such as a higher-level X percent point) in the probability distribution, or a value based on a value two or three times as large as the standard deviation ⁇ is determined as a threshold.
- a predetermined probability such as a higher-level X percent point
- the threshold setter 220 stores a plurality of pairs of the thresholds and the travel conditions in the model database 320 in association with the corresponding prediction model.
- the objective variable of a prediction model is the deceleration of a brake
- a prediction model for predicting another state of the vehicle for example, a braking distance of the brake may be used instead.
- the braking distance may be measured by calculating, for example, a distance since braking starts until the vehicle stops or until a desired deceleration or speed is reached.
- a prediction model for predicting both of the deceleration and the braking distance of the brake may be used.
- Formula (1) is prepared for each of the deceleration and the braking distance.
- the number of objective variables of the prediction model is two.
- a prediction model may have a plurality of objective variables instead of a single objective variable. In such a case, an anomaly may be detected when the residual of each objective variable or any one of the objective variables exceeds a threshold.
- a large number of anomaly detection models applicable to various conditions can be generated by setting a threshold in accordance with a travel condition. It is possible to set a threshold appropriate for a detailed condition of, for example, a plurality of time slots such as morning, afternoon, and night, routes in a plurality of areas such as an urban area, a suburban area, and a mountainous area, all seasons of spring, summer, fall, and winter, and a plurality of weathers such as rainy, snowy, and sunny weathers.
- a threshold appropriate for a detailed condition of, for example, a plurality of time slots such as morning, afternoon, and night, routes in a plurality of areas such as an urban area, a suburban area, and a mountainous area, all seasons of spring, summer, fall, and winter, and a plurality of weathers such as rainy, snowy, and sunny weathers.
- a plurality of thresholds are set to an identical prediction model in accordance with travel conditions.
- a plurality of anomaly detection models (a plurality of pairs of a prediction model and a threshold) may be generated in accordance with travel conditions.
- a travel condition that satisfies current travel information is specified, and an anomaly detection model (a prediction model and a threshold) corresponding to the specified travel condition is used.
- the model generator 210 generates a prediction model for each of a plurality of travel conditions.
- the threshold setter 220 sets a threshold corresponding to each prediction model (in other words, a threshold corresponding to each travel condition).
- the model generator 210 generates a plurality of travel conditions.
- the model generator 210 extracts, from travel information, data that satisfies each travel condition, and generates a prediction model by using the extracted data.
- the prediction model is generated by a method same as that in the above-described embodiment.
- the threshold setter 220 sets a threshold corresponding to each prediction model (in other words, a threshold corresponding to each travel condition) in a manner same as that in the above-described embodiment.
- the generated prediction model, the set threshold, and the corresponding travel condition are stored in the model database 320.
- FIG. 20 illustrates an exemplary model database 320 according to the second embodiment.
- Models 0001_A, 0001_B, and 0001_C are generated in place of the model 0001 illustrated in FIG. 7. In other words, three anomaly detection models are newly generated in place of one anomaly detection model. This generation of a plurality of models in place of one model is referred to as model division.
- the column of travel condition is additionally provided to store the travel condition corresponding to each model.
- the model 0001_A is used when the travel condition "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%" is satisfied.
- the model 0001_B is used when the travel condition “sunny weather (no precipitation) and the boarding rate higher than 90%” is satisfied.
- the model 0001_C is used when the travel condition "rainy weather (no precipitation)" is satisfied.
- a large number of anomaly detection models applicable to various conditions can be generated by recursively repeating the model division on the anomaly detection models generated in the present embodiment. It is possible to generate an anomaly detection model appropriate for a detailed condition of, for example, a plurality of time slots such as morning, afternoon, and night, routes in a plurality of areas such as an urban area, a suburban area, and a mountainous area, all seasons of spring, summer, fall, and winter, and a plurality of weathers such as rainy, snowy, and sunny weathers.
- the present embodiment may be combined with the first embodiment. Specifically, a plurality of thresholds in accordance with travel conditions can be set to each of a plurality of anomaly detection models generated through the model division. This allows generation of an anomaly detection model corresponding to a further detailed condition.
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Mechanical Engineering (AREA)
- Electric Propulsion And Braking For Vehicles (AREA)
- Valves And Accessory Devices For Braking Systems (AREA)
Abstract
Embodiments of the present invention achieve highly accurate anomaly detection. According to one embodiment, an anomaly detection device includes: a condition generator, a threshold setter and an anomaly detector. The generator generates a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model. The threshold setter sets a plurality of thresholds for the conditions. The anomaly detector performs anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions.
Description
Embodiments of the present invention relate to an anomaly detection device, an anomaly detection method, and a computer program.
Maintenance management and inspection of railway vehicles need to be performed daily to maintain safety and reliability of railway operation. Conventional maintenance management has been focused on periodic examination of railway vehicles. To achieve higher railway safety, however, recent development has been made on technologies for achieving early anomaly detection by performing diagnosis and state monitoring by utilizing vehicle information including sensor values and control values acquired from railway vehicles.
In an existing technology, anomaly detection is performed through comparison of a value measured by a sensor on a railway vehicle with a threshold. In another existing technology, a model that recreates a normal operation of a railway vehicle is prepared and used to detect an anomaly or an anomaly sign. However, for a railway vehicle, a travel condition dynamically changes in a temporally sequential manner due to, for example, a route slope, a weather change, passenger loading and unloading, and an operation by a driver. Thus, it is difficult to accurately perform anomaly detection on such a vehicle based on a single threshold.
Embodiments of the present invention provide an anomaly detection device, an anomaly detection method, and a computer program that achieve highly accurate anomaly detection.
According to one embodiment, an anomaly detection device includes: a condition generator, a threshold setter and an anomaly detector. The generator generates a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model. The threshold setter sets a plurality of thresholds for the conditions. The anomaly detector performs anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions.
Embodiments of the present invention will be described below with reference to the accompanying drawings. Any identical component between the drawings is denoted by an identical reference numeral, and description thereof will be omitted as appropriate.
FIG. 1 is a block diagram illustrating an exemplary anomaly detection system according to a first embodiment of the present invention.
The anomaly detection system illustrated in FIG. 1 includes an anomaly detection device 100, a vehicle system 500, an environment information system 600, a terminal 700, an input device 800, and a screen display device 900. The outline of the anomaly detection system illustrated in FIG. 1 will be described next.
The anomaly detection device 100 operates in a learning mode and an operational mode. The anomaly detection device 100 includes a function (anomaly detection model generator 200) that creates, in the learning mode, an anomaly detection model related to a brake system of a railway vehicle based on at least one of measurement information of a railway vehicle (hereinafter, a vehicle) and environment information of the vehicle. The measurement information is acquired from the vehicle system 500. The environment information is acquired from the environment information system 600. Travel information of the vehicle according to the present embodiment includes at least one of the measurement information of the vehicle, which is acquired from the vehicle system 500, and the environment information of the vehicle, which is acquired from the environment information system 600. The anomaly detection model includes a prediction model for predicting the state of the vehicle, and a threshold related to a residual (difference) from a predicted value of the prediction model. The state of the vehicle is, as an example, the deceleration of the vehicle.
The anomaly detection device 100 includes a function (anomaly detector 110) that performs anomaly detection of the vehicle by using the prediction model and the threshold in the operational mode. The anomaly detection is determination of whether there is an anomaly. The anomaly detection is also called anomaly determination. The anomaly detection is performed by comparing, with the threshold, a residual as the difference between the predicted value of the prediction model and the actual value of the state of the vehicle acquired from the vehicle.
The anomaly detection device 100 includes a function (threshold setter 220) that sets a threshold for each of a plurality of conditions (hereinafter, travel conditions) based on the travel information of the vehicle. The anomaly detection device 100 generates one prediction model, and an anomaly detection model including the thresholds corresponding to the respective travel conditions. In the anomaly detection using the anomaly detection model, a threshold corresponding to a travel condition under which the vehicle exists, in other words, a travel condition that satisfies travel information (current travel information) as a target of the anomaly detection is selected from among the thresholds included in the anomaly detection model, and used together with the prediction model.
The creation of a plurality of travel conditions and the setting of a plurality of thresholds are performed in the learning mode. In the creation of travel conditions, the travel information of the vehicle and a result (for example, the residual between the predicted value of the prediction model and the actual value of the state value) of the anomaly detection are used. The learning mode and the operational mode may be switched automatically or by an instruction from, for example, a maintenance, or may be simultaneously executed.
When having detected an anomaly, the anomaly detection device 100 displays, on the screen display device 900, for example, a place where the anomaly is detected, an anomaly detection model used for the anomaly detection, travel information used for the anomaly detection, and a predicted value of a prediction model. This supports monitoring by a railway operator.
The following describes briefly a brake system of a vehicle according to the present embodiment. FIG. 2 illustrates an exemplary configuration of a brake notch, and a brake and an air spring for a particular wheel of the vehicle. The brake notch is located in a driver room of a train in reality. The following describes the brake system of a vehicle, as a target of anomaly detection by the anomaly detection device 100, with reference to FIG. 2.
A brake lever 10 as an exemplary controller provides a driver with a means for a brake operation. The driver brakes the vehicle by moving the brake lever upward. The numbers of one to eight that are indicated on the brake lever 10 represent brake notches (brake steps), and a larger number means a stronger braking force applied to the vehicle. This number of notches is exemplary and does not prevent the vehicle from using a larger or smaller number of notches. Each brake notch is an exemplary control command value to the vehicle or the brake.
The brake operation to the vehicle is not limited to an operation performed by a driver. For example, in a vehicle on which an automatic train stop (ATS), an automatic train control (ATC), or an automatic train operation (ATO) is mounted, the brake operation is performed by the device in place of a driver in some cases. In such a case, for example, a braking command output from the device corresponds to the control command value.
FIG. 2 illustrates a wheel 30 of a vehicle traveling on a rail 20. One means for braking the vehicle with a brake is a tread brake 42. In this example, only one wheel is illustrated for simplicity of description, but in reality, a plurality of pairs of right and left wheels are provided.
The tread brake 42 uses an air cylinder as a power source. When brake cylinder pressure that is the pressure inside an air cylinder 43 is increased, a brake block 41 is pressed against a tread as a surface of the wheel 30, which contacts with the rail. Frictional force between the wheel 30 and the brake block 41 functions as braking force of the tread brake 42.
Since the tread brake uses the frictional force of the brake block in this manner, the brake block is abraded by continuous use, which potentially decreases the braking force. The tread brake is an exemplary mechanical brake used for a vehicle, and another scheme uses a disk brake that obtains braking force by pressing, with a pad or the like, a disk fixed to a wheel shaft against a wheel. The braking force of a brake varies with an abrasion state of, for example, the brake block or the pad. When an anomaly is detected in the brake system by the present anomaly detection device, a worker or the like may check, for example, the brake block or the disk of the brake system and check actual existence of the anomaly.
The braking force of a brake also varies with a load on a vehicle in addition to the abrasion of a component thereof. A load responsive device 50 is mounted on the vehicle illustrated in FIG. 2. The load responsive device 50 includes an air spring 51 and measures a load on the vehicle by sensing air spring pressure of the air spring 51. To control the brake of the vehicle, in addition to the operation of the brake lever 10, the braking force of the brake may be adjusted in response to the air spring pressure detected by the load responsive device 50. Accordingly, desired deceleration can be achieved irrespective of change of a load on the vehicle.
To supplement the braking force of the mechanical brake, an electric brake may be additionally used in the brake system of a vehicle. The electric brake will be described with reference to FIG. 3. FIG. 3 illustrates an exemplary configuration of a power generation brake and a regenerative brake of a vehicle.
Main electric motors 60a and 60b are mounted on the vehicle illustrated in FIG. 3. When the power generation brake is used, the main electric motors 60a and 60b form a closed circuit with a resistor 70 to convert electrical power of the main electric motors into thermal energy.
When the regenerative brake is used, electrical power generated by the main electric motors 60a and 60b is transmitted to a line 90 through a pantograph 80. When a secondary battery is mounted on the vehicle, the generated electrical power may be used to charge the secondary battery. In this manner, the regenerative brake obtains braking force through conversion of kinetic energy into electrical power by using the main electric motors 60a and 60b as electric generators.
The mechanical brake and the power generation brake are exemplary, and the anomaly detection device 100 can perform anomaly detection on a brake of any other scheme used in the brake system.
It is difficult to perform accurate anomaly detection on the brake system of a vehicle because the brake system has a relatively complicated configuration and characteristics of the brake and the brake system vary with a plurality of factors and conditions.
For example, brakes of a plurality of schemes having different characteristics are used in the brake system of a vehicle. In addition, the braking force of the brake system of a vehicle varies with a load as described above. For example, for a passenger vehicle, the number of passengers largely varies with a time slot and an operation interval, and accordingly, the braking force of the brake system largely varies in a short duration. For a cargo vehicle, a load largely varies with the amount of cargo. In addition, deceleration when the brake operation is performed potentially varies between travel routes and intervals of a vehicle, which have different slope and cant tendencies. Moreover, any difference in weather conditions of precipitation, atmospheric temperature, and the like potentially changes the physical property of a component of the brake system, affecting the characteristic of the brake system. As other factors, drivers perform the brake operation in different manners, and vehicles are manufactured with different brake characteristics.
In the present embodiment, highly accurate anomaly detection is easily performed by using a threshold appropriate for situations inside and outside of the vehicle by switching thresholds depending on a travel condition under which a vehicle exists. In this manner, the risk of false anomaly detection is reduced to achieve early anomaly detection and safe and reliable railway operation.
The anomaly detection device 100 illustrated FIG. 1 will be described further in detail below. In the following description, a brake system as a target of the anomaly detection by the anomaly detection device 100 may be a brake device for a particular wheel of an optional railway vehicle, all of a plurality of brake devices in the entire vehicle, or the entire group of brake devices in a plurality of vehicles or cars in a train. The anomaly detection target is not limited to a brake system, but may be a power system, an air-conditioning system, or a power generating system. The anomaly detection target is not limited a railway vehicle but may be an optional vehicle including a wheel, such as an automobile, a construction machine, and an aircraft. The anomaly detection target also includes any device or system other than a vehicle.
The anomaly detection device 100 includes a vehicle information collector 101, an environment information collector 102, a data processor 103, the anomaly detection model generator 200, a condition generator 230, the anomaly detector 110, an alarm 120, and a screen generator 130.
The anomaly detection model generator 200 includes a model generator 210 and the threshold setter 220.
The vehicle information collector 101 acquires measurement information (also referred to as measure data) related to a vehicle from various sensors of the vehicle system 500 inside of the vehicle. Examples of the sensors include a sensor configured to detect the brake operation of the vehicle or the like as a control command value, a sensor configured to detect deceleration of the vehicle, a sensor configured to detect a driving speed, and a sensor configured to measure a load applied on the vehicle. Other various sensors may be included. The measurement information includes a detected value (control command value, for example) of a sensor, and a measured value (driving speed, load, or deceleration, for example) of a sensor. When the vehicle system 500 calculates deceleration from the value of a speed sensor, the calculated deceleration may be acquired as part of the measurement information.
The kind of the measurement information to be acquired (the kind of a sensor or the kind of a control command value) may be optionally set. The measurement information may be acquired in an optionally set period. For example, the measurement information related to the driving speed of the vehicle is acquired in a short sampling period in milliseconds. The value of the sensor configured to measure a load applied to the vehicle is acquired in a sampling period in minutes.
The environment information collector 102 acquires environment information (also referred to as travel environment data) of the vehicle from the environment information system 600. Examples of the environment information include information related to an operation route and information related to weather. Examples of the information related to an operation route include a slope and a cant (height difference between inner and outer rails of railway) for each interval. Examples of the information related to weather include weather, the atmospheric temperature, the amount of precipitation, the speed of wind, and the atmospheric pressure. The acquisition of the environment information may be acquisition of information accumulated in a database in a ground system, or acquisition of information distributed from an external server. The kind of the environment information to be acquired and the frequency of the acquisition may be optionally set.
The anomaly detection device 100 may be installed as a ground device outside of the vehicle, for example, in a facility or an operation command center of a railway operation management company, or may be installed as an on-board device on the vehicle. The anomaly detection device 100 is not limited to a particular installation manner.
When the anomaly detection device 100 is installed as a ground device outside of the vehicle, the measurement information of the vehicle system 500 inside of the vehicle and the like are received through, for example, an on-board element, a transponder ground element, and a ground information network. Specifically, the vehicle system 500 transmits data to the ground information network through the ground element and the like, and the anomaly detection system receives the data through the ground information network. The ground information network may use, for example, a metallic cable, a coaxial cable, an optical cable, a phone line, a wireless device, or Ethernet (registered trademark), but is not limited to a particular scheme. The anomaly detection device 100 receives data from the environment information system 600 through the ground information network.
When the anomaly detection device 100 is an on-board device, the anomaly detection device 100 acquires data from the vehicle system 500 through an information network inside of the vehicle. The information network inside of the vehicle is, for example, Ethernet or a wireless local area network (LAN), but may be achieved in any other scheme. The anomaly detection device 100 may use the on-board element and the transponder ground element to acquire data from the environment information system 600 connected with the ground information network.
The input device 800 provides an interface for an operation by a maintenance person. The input device 800 includes a mouse, a keyboard, a voice recognition system, an image recognition system, a touch panel, or a combination thereof. The maintenance person can perform an operation by inputting various commands or data to the anomaly detection device 100 through the input device 800.
The screen display device 900 displays, as a still image or a moving image, data or information output from the anomaly detection device 100. The screen display device 900 is, for example, a liquid crystal display (LCD), an organic electroluminescence display, or a vacuum fluorescence display (VFD), but may be a display device in any other scheme.
Each of the input device 800 and the screen display device 900 may be one of a plurality of installed devices. For example, the input device 800 and the screen display device 900 may be installed at each of the operation command center and an operation table of the vehicle.
The input device 800 and the screen display device 900 may be one integrated device. For example, when a display with a touch panel function is used, a single device may serve as the input device 800 and the screen display device 900.
The anomaly detection device 100 includes an information database 310, a model database 320, and a detection result database 330.
The databases 310, 320, and 330 are all arranged inside of the anomaly detection device 100 illustrated in FIG. 1. However, the arrangement of the databases is not limited to a particular method. For example, part of the databases may be arranged in an external server or storage device. Each database can be implemented by a relational database management system and various NoSQL systems, but may be implemented in any other scheme. Each database may employ a storage format of XML, JSON, or CSV, or any other format such as a binary format. Not all databases inside of the anomaly detection device 100 need to be achieved by an identical database system and an identical storage format, but the databases may be achieved in a mixture of a plurality of schemes.
The information database 310 stores the measurement information acquired by the vehicle information collector 101 and the environment information acquired by the environment information collector 102. A storage medium such as a memory device storing the measurement information and the environment information may be inserted into the anomaly detection device 100 and used as the information database 310.
FIGS. 4 and 5 illustrate exemplary information databases 310. The travel information (the measurement information and the environment information) is stored in forms of a table 310a illustrated in FIG. 4 and a table 310b illustrated in FIG. 5.
The "data ID" column of the table 310a illustrated in FIG. 4 stores an identification number of an entry stored in the table 310a. A data ID serves as a primary key. Each data ID is associated with the table 310b as illustrated in FIG. 5. The table 310b is stored in the information database 310. The "time" column stores the generation time of an entry. In this example, an entry is generated in each constant sampling time. However, an entry may be generated at an interval set to a rail track in advance or in any other scale.
The "driver" column of the table 310a stores the name of a driver who performed a brake operation. When the brake operation is performed by a device such as the ATS, the ATC, or the ATO instead of a driver, for example, the name of a device that performed the operation or an identifier indicating the device may be stored instead.
The "weather" column of the table 310a stores information related to weather acquired from the environment information system 600.
The "atmospheric temperature" column of the table 310a stores information related to the atmospheric temperature acquired from the environment information system 600. The information related to the atmospheric temperature may be an actual value or a label classifying the actual value. In the example illustrated in FIG. 4, the "atmospheric temperature" column stores the label of any one of classes T1, T2, T3, T4, T5, T6, and T7 into which the atmospheric temperature as a real number is converted by using a conversion table 310c illustrated in FIG. 6. For example, the atmospheric temperature at -11℃is converted into class T1, the atmospheric temperature at 15℃ is converted into class T4, and the atmospheric temperature at 33℃ is converted into class T6. When a prediction model to be described later is created, data obtained by converting a class name into an optional integer, for example, the class T1 into 1, the class T2 into 2, and the class T3 into 3, may be used as explanatory variables.
As exemplarily described with the "atmospheric temperature" column, the information database 310 may store processed information obtained by performing calculation or conversion on the measurement information or the environment information.
The "boarding rate" column of the table 310a stores a boarding rate in percentage as an index for a load applied to the vehicle. Another index may be used to indicate a load. The boarding rate is defined with, for example, a ratio of the capacity of a passenger vehicle and the number of passengers in the passenger vehicle. The boarding rate is often estimated based on the air spring pressure of the load responsive device. In such a case, the air spring pressure may be directly used as the index.
The air spring pressure is the actual value of a sensor, which is not a value estimated indirectly by using, for example, a conversion table or a formula unlike the boarding rate, and thus can be used to reduce residual in model generation. However, the air spring pressure takes a value depending on the manufacturer and the model number of the load responsive device mounted on the vehicle, and thus lacks in versatility. Thus, the difference between vehicles due to different load responsive devices thereof can be absorbed in some cases when a typically used index such as the boarding rate is used.
The "slope" column of the table 310a stores the slope of a route in a value expressed in the unit of permil. The permil is a value obtained by expressing a height difference for the horizontal distance of 1000 m in meters. The permil is exemplary, and the "slope" column may store a value in another unit.
The "cant" column of the table 310a stores a cant in millimeters, but may store a value in another unit.
The table 310a further includes the "wind speed" column and the "atmospheric pressure" column. The table 310a may include columns storing other information such as the current position and the current interval on the rail track.
The table 310b illustrated in FIG. 5 stores information such as time information, a brake notch, and the actual value of deceleration for a corresponding entry in the table 310a illustrated in FIG. 4. In the example illustrated in FIG. 5, the table 310b corresponds to the data ID of 2560 in FIG. 4. Entries in the table 310b are generated in a time interval shorter than that for the table 310a. The generation interval (sampling interval) of entries in the table 310b may be same as that for the table 310a. The table 310b and the table 310a are separately provided in this example but may be integrated with each other.
Data stored in the information database 310 may be processed. For example, the data processor 103 causes the screen display device 900 to display the content of each table stored in the information database 310. A maintenance person or a driver performs a fabrication operation on data by using the input device 800. The data processor 103 performs data fabrication in accordance with the fabrication operation.
The interval of acquisition of information or data by the vehicle information collector 101 or the environment information collector 102 may be adjusted. For example the data processor 103 receives an operation to specify the acquisition interval from a maintenance person or a driver through the input device 800, and adjusts the acquisition interval in accordance with the content of the operation.
The anomaly detection model generator 200 creates an anomaly detection model of the brake system of the vehicle by using data stored in the information database 310. The anomaly detection model includes a prediction model and one or a plurality of thresholds. The prediction model is generated by the model generator 210, and each threshold is generated by the threshold setter 220. The generated anomaly detection model is stored in the model database 320.
FIG. 7 illustrates an exemplary model database 320. The model database 320 can store one or a plurality of anomaly detection models. Each anomaly detection model is identified by a model ID. The column of prediction model stores data indicating a prediction model or a memory address (pointer) at which the prediction model is stored. The data indicating a prediction model includes, for example, a deceleration model. Each cell in the column of threshold stores one or a plurality of thresholds. When a plurality of thresholds are stored, a travel condition corresponding to each threshold is stored together (to be described later in detail).
The generation of an anomaly detection model is performed in the learning mode, for example, when the anomaly detection device 100 is started or when a system is newly added as an anomaly detection target. When there are a plurality of systems as anomaly detection targets, an anomaly detection model is generated for each system.
The anomaly detection model is created by using a data sample (feature vector) extracted from the information database 310.
The data sample (feature vector) includes one or more explanatory variables. Examples of the explanatory variables include the value (control command value) of a brake notch in the table 310b. Additionally, a value (such as speed) of any other kind in the travel information and specifications of the vehicle (such as the sizes and weight of the vehicle) may be used as explanatory variables. Explanatory variables may be generated through calculation with a plurality of items included in the travel information. In this example, the objective variable of a prediction model is deceleration. The data sample may be generated for each entry in the brake information table 310b, or a plurality of continuous entries with lower temporal granularity may be collected to generate one data sample.
The following describes a method of generating a prediction model. It is assumed that a regression model is used as a prediction model. The model generator 210 obtains a feature vector X = (x1, x2, x3, ..., xn) having explanatory variables as elements by using the information database 310.
Subsequently, the model generator 210 performs multiple regression analysis to obtain Formula (1) for predicting deceleration as the objective variable.
In the above formula, y represents the objective variable, xn represents an explanatory variable, and bn represents a partial regression coefficient. To absorb a difference in measurement units between the explanatory variables, a standard partial regression coefficient may be used as the partial regression coefficient bn by normalizing the objective variable and all explanatory variables to the average value of 0 and the dispersion of 1. The number of explanatory variables may be one or more.
The model generation by the multiple regression analysis is exemplary, and a prediction model for the objective variable may be created by any other method such as support vector regression or autoregression.
When a prediction model is created, cross validation may be used. For example, a data sample can be divided into a plurality of sets so that at least one of the sets is used as test data for validation and the other sets are used for model creation. This allows the performance of a generated model to be checked.
In each embodiment of the present invention, an information database stores information acquired when a brake system on which anomaly detection is performed is in the normal state. Thus, a generated prediction model is a model obtained by modeling the behavior of the brake system of the vehicle in the normal state. However, the information database may store information when a particular brake is in failure.
The number of explanatory variables used to create a prediction model may be reduced by adjusting the number of explanatory variables by, for example, a variable selection method or principal components analysis. This adjustment is effective when there is a correlation between different explanatory variables or when reduction of a calculation time and a processing load is required.
In the variable selection method, a model is generated by selecting a subset of explanatory variables effective for prediction from among a set of all explanatory variables. Useful explanatory variables may be selected by first generating a model by using one or a small number of explanatory variables and then generating a model to which explanatory variables are added one by one. Alternatively, useful explanatory variables may be specified by first generating a model with a large number of explanatory variables and then generating a model obtained by removing explanatory variables one by one. The selection of explanatory variables may be performed by using a genetic algorithm.
In the principal components analysis, an eigenvalue problem of a correlation matrix or a variance-covariance matrix of data used for model generation is solved to generate a new explanatory variable, thereby reducing the number of dimensions. Regression analysis using, as the variable xn of Formula (1), the new explanatory variable obtained by the principal components analysis is called principal component regression.
The following describes an example using the variable selection method. When the brake system of a vehicle is an anomaly detection target and the deceleration of a brake is the objective variable of a prediction model, for example, a prediction model using only a brake notch as an explanatory variable is created first. The brake notch is presumed to have a highest correlation with the deceleration. Thereafter, any other explanatory variable such as the travel speed of the vehicle, which is thought to have a correlation with the deceleration, is sequentially added to the prediction model, and the accuracy of prediction is checked. An explanatory variable when a desired prediction accuracy is obtained is employed.
The following describes a threshold set to a prediction model by the threshold setter 220. The threshold is used to determine that there is an anomaly when the difference between a predicted value of the objective variable (in this example, a predicted value of the deceleration), which is calculated based on a prediction model, and a measured value (actual value) of the deceleration exceeds the threshold. The determination that there is an anomaly is also referred to as detection of an anomaly. The difference between the actual value and the predicted value is also referred to as residual. The actual value can be larger or smaller than the predicted value, and thus the value of the residual can be positive or negative. When the sign is not important, the residual may be defined to be the absolute value of the difference, which is the absolute value of the distance from the predicted value.
FIG. 8 is used to describe an exemplary threshold determination method using a normal distribution. FIG. 8 illustrates the graph of a normal distribution 400 of residual. The horizontal axis represents the residual, and the vertical axis represents a probability density. A plurality of residuals between the predicted value of a prediction model and the actual value are acquired to create the normal distribution 400, assuming that these residuals obey a normal distribution. Data used to acquire the residuals may be a data sample used to generate the prediction model, test data, travel information unrelated to generation of the prediction model, or an optional combination thereof. When the residuals have larger variance, the distribution has a more spread skirt as illustrated with normal distributions 401 and 402 illustrated with a dashed line.
The normal distribution 400 is used to set a threshold for the prediction model. For example, when the standard deviation is represented by σ, a multiple of the standard deviation, such as 2σ or 3σ, is set as the threshold. When the residual exceeds the threshold that is set to be 2σ, an anomaly is detected in anomaly detection. When such a threshold is set, about 95% of the actual values are determined to have no anomaly (be normal). As another exemplary threshold setting, the value of the residual corresponding to a predetermined probability (for example, a highest X percent point or a lowest X percent point) or the absolute value thereof may be set as the threshold. The above-described threshold determination method is exemplary, and any other method may be used. For example, the threshold may be determined by assuming a distribution other than a normal distribution or by a person such as a maintenance person or a driver based on experience thereof.
The anomaly detector 110 performs anomaly detection on a target system by using an anomaly detection model (a prediction model and a threshold) stored in the model database 320. A feature vector is generated from travel information used for the anomaly detection, and a deceleration is predicted by using the generated feature vector and the prediction model. The residual between the predicted deceleration and the actual deceleration is compared with the threshold. When the residual is equal to or smaller than the threshold, it is determined that the deceleration is normal. When the residual is larger than the threshold, it is determined that the deceleration is anomalous. The anomaly detector 110 stores information in the detection result database 330 based on a result of the anomaly detection. The anomaly detector 110 notifies information related to the result of the anomaly detection to the screen generator 130 and the alarm 120.
FIG. 9 illustrates an exemplary detection result database. The database stores the brake notch, the actual value of the deceleration, the predicted value of the deceleration based on a prediction model (ID0001), and anomaly detection result (existence or non-existence on anomaly) in a temporally sequential manner. In the example illustrated in FIG. 9, no anomaly is detected at any timing. Travel information (for example, the items of driver, weather, atmospheric temperature, boarding rate, slope, cant, wind speed, and atmospheric pressure) at a corresponding time may be added to the detection result database.
The above description is made on an example in which one threshold is set to a prediction model. In the present embodiment, however, the threshold setter 220 can set, to a prediction model, a plurality of thresholds in accordance with a travel condition. In this case, at anomaly detection, the anomaly detector 110 specifies a travel condition satisfied by the current travel information. Then, anomaly detection is performed by using a threshold corresponding to the specified travel condition. The travel condition needed by the threshold setter 220 to set a plurality of thresholds is generated by the condition generator 230.
The condition generator 230 generates, by using the detection result database 330 and the travel information used for anomaly detection, a plurality of travel conditions (a plurality of conditions) for classifying the difference between the predicted value of a prediction model and the actual value in accordance with the value of the difference. The following describes operation of the condition generator 230 in detail.
The condition generator 230 creates, by using the detection result database 330 and the travel information, a classifier (for example, a decision tree) for predicting a class in accordance with the residual.
Each residual between the predicted value and the actual value in the detection result database is classified into a plurality of classes (residual classes). For example, the residual that is equal to or smaller than a threshold A is classified into a residual class A for a small residual, the residual that is larger than the threshold A and smaller than a threshold B is classified into a residual class B for an intermediate residual, or the residual that is equal to or larger than the threshold B is classified into a residual class C for a large residual. The threshold A may be a threshold set first to the prediction model but is not limited thereto. Another exemplary classification will be described next.
Residuals are arranged in ascending or descending order. When the residuals are arranged in ascending order of size, a data set of smallest 20% residuals is classified into the residual class A, a data set of next smallest 60% residuals is classified into the residual class B, and a data set of the remaining 20% residuals is classified into the residual class C. The ratio of the classes may be optionally determined and is not limited to the above-described values.
Residuals are arranged in ascending or descending order. When the residuals are arranged in ascending order of size, a data set of smallest 20% residuals is classified into the residual class A, a data set of next smallest 60% residuals is classified into the residual class B, and a data set of the remaining 20% residuals is classified into the residual class C. The ratio of the classes may be optionally determined and is not limited to the above-described values.
The condition generator 230 selects, for each entry in the detection result database 330 as illustrated in FIG. 9, any one of the residual classes A to C in accordance with the residual of the entry, and allocates the selected residual class to the entry. In addition, the condition generator 230 specifies travel information corresponding to each entry from the table 310a illustrated in FIG. 4, and associates the travel information with the entry. However, this operation is unnecessary when the travel information is already included in the detection result database 330. In this manner, as illustrated in FIG. 10, a data set is generated in which each entry in the detection result database is associated with the residual class and the travel information. The data set may be stored in an internal buffer of the condition generator 230, or may be stored in a storage device or database not illustrated. In the data set illustrated in FIG. 10, all entries are set to the residual class A. A decision tree is generated by using the data set as learning data, specifically, by setting the residual class as the objective variable and the other items as explanatory variables. Any item unnecessary for producing a data set, in databases (those illustrated in FIGS. 9 and 4, for example) from which the data set is created, does not need to be used. For example, when unnecessary, the item of anomaly detection result does not need to be included in the data set illustrated in FIG. 10.
This example assumes that the data set includes no data in which an anomaly is detected. However, when there is data in which an anomaly is detected, this data may be excluded from the data set. When the anomaly detector 110 detects an anomaly, whether this detection result is correct may be checked by a maintenance person. For example, when an anomaly is detected, the screen display device 900 displays a check screen (refer to FIG. 16 to be described later) for checking whether this detection result is correct. When having determined that the detection result is false, the maintenance person inputs an instruction for the false detection. The condition generator 230 or another processor corrects the detection result in the detection result database 330 based on this instruction. Data in which an anomaly is detected may be excluded from the data set at creation of a decision tree.
FIG. 11 illustrates an exemplary generated decision tree. This decision tree predicts a residual class corresponding to the objective variable based on two explanatory variables related to the amount of precipitation and the boarding rate. Nodes 1001a, 1001b, and 1001c are end nodes corresponding to the objective variable. The top node is called a root node. Any node other than the end nodes and the root node is called an intermediate node. The root node and the intermediate node are explanatory variable nodes. The end nodes are residual class nodes (objective variable nodes). The nodes 1001a, 1001b, and 1001c correspond to the residual class A, the residual class B, and the residual class C, respectively. This decision tree classifies the residual into the residual class A for "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%", classifies the residual into the residual class B for "sunny weather (no precipitation) and the boarding rate higher than 90%", or classifies the residual into the residual class C for "rainy weather (no precipitation)".
The condition generator 230 acquires, as a travel condition corresponding to each residual class, a condition included in a path from the residual class node (end node) to the root node. Specifically, the condition generator 230 acquires the condition "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%" corresponding to a path from the residual class A to the root node, the condition "sunny weather (no precipitation) and the boarding rate higher than 90%" corresponding to a path from the residual class B to the root node, and the condition "rainy weather (no precipitation)" corresponding to a path from the residual class C to the root node, as travel conditions A, B, and C, respectively.
The threshold setter 220 sets a threshold to each of the travel conditions A to C. Specifically, the threshold setter 220 classifies travel information used to generate a prediction model (or travel information not used to generate the prediction model) into groups A to C that satisfy the travel conditions A to C. Anomaly detection is performed on the travel information classified into the group A, and the residual is calculated based on a result of the detection. The distribution (for example, normal distribution) of the residual is calculated and used to set the threshold A (refer to the description with reference to FIG. 8). In this manner, the threshold A is set for the travel condition A. Similarly, the threshold B corresponding to the travel condition B is set to the group B, and the threshold C corresponding to the travel condition C is set to the group C.
The distribution of the residual corresponding to the group A has a small standard deviation, and thus the threshold A is small. The distribution of the residual corresponding to the group C has a large standard deviation, and thus the threshold C is large. The threshold B corresponding to the group B is between the threshold A and the threshold C.
The threshold A is used when the weather is sunny and the boarding rate is equal to or lower than 90%. The threshold B is used when the weather is sunny and the boarding rate is higher than 90%. The threshold C is used when the amount of precipitation is larger than zero. The thresholds are switched in accordance with such a travel condition at anomaly detection, thereby achieving the anomaly detection accurately based on a brake characteristic. Although only the amount of precipitation and the boarding rate are used as travel conditions above, for example, other items of atmospheric temperature and humidity may be used as travel conditions.
Examples of algorithms used for learning of a decision tree include ID3 and C4.5, but any algorithm may be used. Pruning of a decision tree may be performed to prevent noise and overlearning. The learning of a decision tree is exemplary, and any other classifier may be used. Which item of explanatory variable is employed in a decision tree among a plurality of items of travel information depends on an algorithm and learning data to be used.
Although a decision tree is used to set a threshold for each travel condition above, a maintenance person may employ professional knowledge to set a threshold for each travel condition.
Each pair of a travel condition generated by the condition generator 230 and a threshold set by the threshold setter 220 are stored in a corresponding cell in the column of threshold in the model database 320. FIG. 12 illustrates an example in which a plurality of thresholds and corresponding travel conditions are stored for the model 0001.
When an anomaly detection model in which a plurality of thresholds are set for a prediction model in this manner is used at anomaly detection, the anomaly detector 110 specifies a travel condition that satisfies current travel information among the travel conditions. Then, the anomaly detector 110 performs anomaly detection by using thresholds and a prediction model corresponding to the specified travel condition. The following describes a specific example of operation of the anomaly detector 110 in this case.
FIG. 13 is a diagram for description of exemplary operation of the anomaly detector 110. An upper part of FIG. 13 illustrates the brake notch. A middle part thereof illustrates the brake deceleration. A lower part thereof illustrates the residual between the actual value and the predicted value of the prediction model. In the prediction model, the brake notch corresponds to an explanatory variable, and the deceleration corresponds to the objective variable.
At time t1, an operation to set the brake notch to notch "4" is performed. Having received this operation, the brake system applies braking force to the vehicle. Accordingly, the deceleration of the vehicle increases and thereafter stays near a constant value. In the interval between times t1 and t2, a predicted value of the deceleration and a measured value (actual value) thereof have a slight difference therebetween but transition substantially at the same value, and the residual between the predicted value and the actual value is smaller than a threshold α. In this interval, the travel environment of the vehicle satisfies the travel condition A. The threshold α corresponds to the travel condition A.
At time t2, the travel condition satisfied by the travel environment of the vehicle changes from A to B. The anomaly detector 110 detects the change of travel condition and changes a used threshold to β.
Travel environment and change thereof can be detected from a measured value and a control command value included in measurement information, and route data and weather data included in environment information. Alternatively, travel environment and change thereof may be detected based on, for example, an explicit command from a driver or a command center, or a radio signal received from a ground element.
In the interval between times t2 and t3, the predicted value of the deceleration is constant, but the actual value varies largely as compared to that between times t1 and t2. Accordingly, the residual exceeds the threshold βat three timings, and the anomaly detector 110 detects an anomaly at each timing.
At time t3, an operation to change the brake notch from notch "4" to notch "2" is performed. Having received this operation, the brake system reduces the braking force applied to the vehicle, and accordingly, the deceleration of the vehicle decreases. In the interval between times t3 and t4, the residual is within the range of the threshold β, and thus no anomaly is detected.
At time t4, the travel condition satisfied by the travel environment of the vehicle returns from B to A. Having detected the change of travel condition, the anomaly detector 110 changes the used threshold from βto α at time t4. In the interval between times t4 and t5, the residual is within the range of the threshold α, and thus no anomaly is detected.
At time t5, an operation to cancel the brake is performed. Having received this operation, the brake system further reduces the braking force applied to the vehicle, and accordingly the deceleration of the vehicle further decreases.
After time t5, the residual is within the range of the threshold αα, and thus no anomaly is detected.
After time t5, the residual is within the range of the threshold αα, and thus no anomaly is detected.
The alarm 120 notifies the detection of an anomaly by the anomaly detector 110 to the terminal 700 used by a railway operator, driver, or maintenance person. This notification may be performed by transmission of an electronic mail, display of a pop-up message on an operation screen of the terminal 700, or notification by predetermined instrument management protocol, or may be performed by any other means. The notification may include detailed information (for example, a place (current value) on a map where the anomaly has occurred, or an identifier of the vehicle to which the anomaly has occurred) of the anomaly. An operator or a maintenance person can know the detection of the anomaly and details thereof by receiving the notification.
The screen generator 130 displays, on the screen display device 900, for example, an anomaly detection result, a current position of the vehicle if an anomaly is detected, an anomaly detection model and a threshold used for the anomaly detection, sensor data, and a predicted value based on a prediction model. The screen generator 130 may be included in the anomaly detection device 100 or may be included in a vehicle information system connected with the anomaly detection device 100, or a terminal or a management server on an information network of a ground system.
FIG. 14 illustrates an exemplary main screen 901 displayed by the screen generator 130. The main screen 901 displays information on a plurality of trains. In this example, the screen display device 900 is installed in a command room for managing and monitoring the trains.
Information related to vehicle is displayed in a table format at an upper part of the main screen 901. Examples of display items include the items of train, anomaly detection result, line number (train identifier), boarding rate, and current position, but any other information may be displayed. The "anomaly" column in the table displays an anomaly detection result. An exclamation mark "!" indicates that an anomaly is detected. Thus, it is indicated that an anomaly is detected at train B. This display of an anomaly detection result is exemplary and may be achieved in any other manner.
A map is displayed at a lower part of the main screen 901, indicating the current position of each train. The detection of the anomaly and the name of a used model are displayed in a word balloon for train B at which the anomaly is detected.
Click on a train of interest on the main screen 901 illustrated in FIG. 14 performs transition to an anomaly detailed screen. The scheme of screen transition is not limited thereto but may employ any other scheme such as a predetermined keyboard operation.
FIG. 15 illustrates an exemplary anomaly detailed screen 902 to which transition is made by clicking train B.
In FIG. 15, graphs same as those illustrated FIG. 13 are displayed at a right part of the screen. Specifically, a graph of the brake notch, a graph of the predicted value of the deceleration, a graph of the actual value of the deceleration, and a graph of the residual are displayed. Each graph is displayed for a particular duration that includes a time when an anomaly is detected. A long bar indicating that a threshold is exceeded is displayed to allow visual check of a time when an anomaly is detected.
A check box is provided at a left part of the screen illustrated in FIG. 15 to allow selection of an item, the graph of which is to be displayed. Another separate means may be provided to allow specification of the time range of the graph display. When such an interface as described above is provided, a railway operator or the like can understand details of the anomaly and take fast measures.
When the anomaly detector 110 has detected an anomaly, a check screen may be presented to a maintenance person to allow the maintenance person to check whether a result of the detection is correct. FIG. 16 illustrates an exemplary check screen 903. When having determined that the detection result is false, the maintenance person inputs an instruction to correct the result. The condition generator 230 corrects a detection result in the detection result database based on this instruction.
FIG. 17 illustrates a hardware configuration of the anomaly detection device according to the present embodiment. The anomaly detection device according to the present embodiment is achieved by a computer device 100. The computer device 100 includes a CPU 151, an input interface 152, a display device 153, a communication device 154, a main storage device 155, and an external storage device 156. These components are connected with each other through a bus 157.
The central processing unit (CPU) 151 executes an anomaly detection program as a computer program on the main storage device 155. The anomaly detection program is a computer program that achieves each above-described functional component of the anomaly detection device. Each functional component is achieved by the CPU 151 executing the anomaly detection program.
The input interface 152 is a circuit for inputting, to the anomaly detection device, an operation signal from an input device such as a keyboard, a mouse, or a touch panel.
The display device 153 displays data or information output from the anomaly detection device. The display device 153 is, for example, a liquid crystal display (LCD), a cathode-ray tube (CRT), or a plasma display (PDP), but is not limited thereto. The data or information output from the computer device 100 can be displayed by the display device 153.
The communication device 154 is a circuit that allows the anomaly detection device to communicate with an external device in a wireless or wired manner. Measurement information can be input from the external device through the communication device 154. The measurement information input from the external device can be stored in the information database 310.
The main storage device 155 stores, for example, the anomaly detection program, data necessary for executing the anomaly detection program, and data generated through execution of the anomaly detection program. The anomaly detection program is loaded onto the main storage device 155 and executed. The main storage device 155 is, for example, a RAM, a DRAM, or an SRAM, but is not limited thereto. The information database 310, the model database 320, and the detection result database 330 may be constructed on the main storage device 155.
The external storage device 156 stores, for example, the anomaly detection program, data necessary for executing the anomaly detection program, and data generated through execution of the anomaly detection program. These program and data are read onto the main storage device 155 at execution of the anomaly detection program. The external storage device 156 is, for example, a hard disk, an optical disk, a flash memory, or a magnetic tape, but is not limited thereto. The information database 310, the model database 320, and the detection result database 330 may be constructed on the external storage device 156.
The anomaly detection program may be installed on the computer device 100 in advance or may be stored in a storage medium such as a CD-ROM. Alternatively, the anomaly detection program may be uploaded on the Internet.
The computer device 100 may include one or a plurality of CPUs 151, one or a plurality of input interfaces 152, one or a plurality of display devices 153, one or a plurality of communication devices 154, and one or a plurality of main storage devices 155, and may be connected with a peripheral instrument such as a printer or a scanner.
The anomaly detection device may be achieved by the single computer device 100 or may be configured as a system including a plurality of computer devices 100 connected with each other.
FIG. 18 is a flowchart of anomaly detection processing performed in the operational mode according to the embodiment of the present invention. The processing of the flowchart illustrated in FIG. 18 may be executed upon a certain operation of an anomaly detection target system, may be executed in a constant period, may be executed upon reception of an instruction from a user such as a maintenance person, or may be executed at any other timing.
At step S101, the anomaly detector 110 acquires travel information as an anomaly detection target from the information database 310.
At step S102, the anomaly detector 110 selects a prediction model corresponding to an anomaly detection target system (in this example, the brake system of a vehicle) from the model database 320. The anomaly detector 110 also selects a threshold corresponding to a travel condition satisfied by the acquired travel information among a plurality of travel conditions. For example, the prediction model is a model for predicting an objective variable representing the state (for example, the deceleration) of the vehicle from an explanatory variable representing a control command value (for example, the brake notch) to the vehicle. In other words, the prediction model is a model that associates an explanatory variable representing the control command value to the vehicle with an objective variable representing the state of the vehicle.
At step S103, the anomaly detector 110 generates a feature vector from the acquired travel information. For example, the anomaly detector 110 generates a feature vector including the control command value. The number of elements of the feature vector may be one or more. The anomaly detector 110 predicts the objective variable (in this example, the deceleration) based on the feature vector and the prediction model. In other words, the anomaly detector 110 calculates a predicted value of the state of the vehicle based on the control command value and the prediction model.
At step S104, the anomaly detector 110 calculates a residual as the difference between the predicted deceleration and a deceleration included in the travel information, and compares the calculated residual with the threshold.
When the residual is larger than the threshold (YES), the anomaly detector 110 detects an anomaly and outputs information notifying the detection of the anomaly to the screen display device 900 or the like (S105).
When the residual is equal to or smaller than the threshold (NO), the anomaly detector 110 detects no anomaly (S106). In other words, the anomaly detector 110 determines that the brake system of the vehicle is normal. When no anomaly is detected, information notifying that the brake system of the vehicle is normal may be output to the screen display device 900 or the like.
FIG. 19 is a flowchart of threshold setting processing at the anomaly detection device in the learning mode. This processing may be executed in a constant period, may be executed at a timing instructed by a maintenance person, or may be executed at any other timing. The following describes an exemplary operation when a plurality of thresholds in accordance with a travel condition is set to a prediction model. Assume that the anomaly detector 110 performs anomaly detection based on a prediction model generated in advance and one threshold and the detection result database 330 stores data related to anomaly detection.
At step S201, the condition generator 230 allocates a residual class to the residual between the predicted value and the actual value in accordance with the value of the residual based on the detection result database 330. The condition generator 230 generates a data set in which the residual class is associated with travel information (refer to FIG. 10).
At step S202, the condition generator 230 sets each item of the data set to be an explanatory variable and the residual class to be an objective variable, and performs, for example, machine learning to generate a classifier that predicts the objective variable from at least one of a plurality of explanatory variables. Specifically, the condition generator 230 generates a classifier that associates a plurality of conditions related to at least one explanatory variable with a plurality of residual classes. In this example, a decision tree (refer to FIG. 11) is generated as the classifier.
At step S203, the condition generator 230 acquires the conditions included in the classifier as a plurality of travel conditions. When the decision tree is used, a condition included in a path from each residual class node (end node) to the root node is acquired as the travel condition corresponding to a residual class.
At step S204, the threshold setter 220 sets a plurality of thresholds to the travel conditions. For example, the threshold setter determines a threshold based on the distribution of the residual classified into a residual class that satisfies each travel condition of the travel information. For example, travel information used to generate the decision tree (or travel information not used to generate the decision tree) is classified into groups that satisfy the travel conditions. Anomaly detection is performed for each group, and the residual is calculated based on a result of the detection. Then, a probability distribution (refer to FIG. 8) of the residual is generated. A value of the residual corresponding to a predetermined probability (such as a higher-level X percent point) in the probability distribution, or a value based on a value two or three times as large as the standard deviation σ is determined as a threshold.
At step S205, the threshold setter 220 stores a plurality of pairs of the thresholds and the travel conditions in the model database 320 in association with the corresponding prediction model.
Although the present embodiment describes the example in which the objective variable of a prediction model is the deceleration of a brake, a prediction model for predicting another state of the vehicle, for example, a braking distance of the brake may be used instead. The braking distance may be measured by calculating, for example, a distance since braking starts until the vehicle stops or until a desired deceleration or speed is reached. Alternatively, a prediction model for predicting both of the deceleration and the braking distance of the brake may be used. In this case, for example, Formula (1) is prepared for each of the deceleration and the braking distance. Accordingly, the number of objective variables of the prediction model is two. In this manner, a prediction model may have a plurality of objective variables instead of a single objective variable. In such a case, an anomaly may be detected when the residual of each objective variable or any one of the objective variables exceeds a threshold.
According to the present embodiment, a large number of anomaly detection models applicable to various conditions can be generated by setting a threshold in accordance with a travel condition. It is possible to set a threshold appropriate for a detailed condition of, for example, a plurality of time slots such as morning, afternoon, and night, routes in a plurality of areas such as an urban area, a suburban area, and a mountainous area, all seasons of spring, summer, fall, and winter, and a plurality of weathers such as rainy, snowy, and sunny weathers.
In the above-described first embodiment, a plurality of thresholds are set to an identical prediction model in accordance with travel conditions. However, in a second embodiment, a plurality of anomaly detection models (a plurality of pairs of a prediction model and a threshold) may be generated in accordance with travel conditions. In this case, when anomaly detection is performed, a travel condition that satisfies current travel information is specified, and an anomaly detection model (a prediction model and a threshold) corresponding to the specified travel condition is used.
The model generator 210 generates a prediction model for each of a plurality of travel conditions. The threshold setter 220 sets a threshold corresponding to each prediction model (in other words, a threshold corresponding to each travel condition).
Specifically, similarly to the first embodiment, the model generator 210 generates a plurality of travel conditions. The model generator 210 extracts, from travel information, data that satisfies each travel condition, and generates a prediction model by using the extracted data. The prediction model is generated by a method same as that in the above-described embodiment. The threshold setter 220 sets a threshold corresponding to each prediction model (in other words, a threshold corresponding to each travel condition) in a manner same as that in the above-described embodiment. The generated prediction model, the set threshold, and the corresponding travel condition are stored in the model database 320. FIG. 20 illustrates an exemplary model database 320 according to the second embodiment. Models 0001_A, 0001_B, and 0001_C are generated in place of the model 0001 illustrated in FIG. 7. In other words, three anomaly detection models are newly generated in place of one anomaly detection model. This generation of a plurality of models in place of one model is referred to as model division. The column of travel condition is additionally provided to store the travel condition corresponding to each model.
When the above-described decision tree illustrated in FIG. 11 is generated, the model 0001_A is used when the travel condition "sunny weather (no precipitation) and the boarding rate equal to or lower than 90%" is satisfied. The model 0001_B is used when the travel condition "sunny weather (no precipitation) and the boarding rate higher than 90%" is satisfied. The model 0001_C is used when the travel condition "rainy weather (no precipitation)" is satisfied.
A large number of anomaly detection models applicable to various conditions can be generated by recursively repeating the model division on the anomaly detection models generated in the present embodiment. It is possible to generate an anomaly detection model appropriate for a detailed condition of, for example, a plurality of time slots such as morning, afternoon, and night, routes in a plurality of areas such as an urban area, a suburban area, and a mountainous area, all seasons of spring, summer, fall, and winter, and a plurality of weathers such as rainy, snowy, and sunny weathers.
The present embodiment may be combined with the first embodiment. Specifically, a plurality of thresholds in accordance with travel conditions can be set to each of a plurality of anomaly detection models generated through the model division. This allows generation of an anomaly detection model corresponding to a further detailed condition.
While certain embodiments have been described, these embodiments have been presented by way of example only, and are not intended to limit the scope of the inventions. Indeed, the novel embodiments described herein may be embodied in a variety of other forms; furthermore, various omissions, substitutions and changes in the form of the embodiments described herein may be made without departing from the spirit of the inventions.
10 BRAKE LEVER
20 RAIL
30 WHEEL
41 BRAKE BLOCK
42 TREAD BRAKE
43 AIR CYLINDER
50 LOAD RESPONSIVE DEVICE
51 AIR SPRING
60a, 60b MAIN ELECTRIC MOTOR
70 RESISTOR
80 PANTOGRAPH
90 LINE
100 ANOMALY DIAGNOSIS DEVICE
101 VEHICLE INFORMATION COLLECTOR
102 ENVIRONMENT INFORMATION COLLECTOR
103 DATA PROCESSOR
110 ANOMALY DETECTOR
120 ALARM
130 SCREEN GENERATOR
151 CPU
152 INPUT INTERFACE
153 DISPLAY DEVICE
154 COMMUNICATION DEVICE
155 MAIN STORAGE DEVICE
156 EXTERNAL STORAGE DEVICE
157 BUS
200 ANOMALY DETECTION MODEL GENERATOR
210 MODEL GENERATOR
220 THRESHOLD SETTER
230 CONDITION GENERATOR
310 INFORMATION DB
310a, 310b, 310c TABLE
320 MODEL DB
330 DETECTION RESULT DB
500 VEHICLE SYSTEM
600 ENVIRONMENT INFORMATION SYSTEM
700 TERMINAL
800 INPUT DEVICE
900 SCREEN DISPLAY DEVICE
901 MAIN SCREEN
902 ANOMALY DETAILED SCREEN
903 CHECK SCREEN
20 RAIL
30 WHEEL
41 BRAKE BLOCK
42 TREAD BRAKE
43 AIR CYLINDER
50 LOAD RESPONSIVE DEVICE
51 AIR SPRING
60a, 60b MAIN ELECTRIC MOTOR
70 RESISTOR
80 PANTOGRAPH
90 LINE
100 ANOMALY DIAGNOSIS DEVICE
101 VEHICLE INFORMATION COLLECTOR
102 ENVIRONMENT INFORMATION COLLECTOR
103 DATA PROCESSOR
110 ANOMALY DETECTOR
120 ALARM
130 SCREEN GENERATOR
151 CPU
152 INPUT INTERFACE
153 DISPLAY DEVICE
154 COMMUNICATION DEVICE
155 MAIN STORAGE DEVICE
156 EXTERNAL STORAGE DEVICE
157 BUS
200 ANOMALY DETECTION MODEL GENERATOR
210 MODEL GENERATOR
220 THRESHOLD SETTER
230 CONDITION GENERATOR
310 INFORMATION DB
310a, 310b, 310c TABLE
320 MODEL DB
330 DETECTION RESULT DB
500 VEHICLE SYSTEM
600 ENVIRONMENT INFORMATION SYSTEM
700 TERMINAL
800 INPUT DEVICE
900 SCREEN DISPLAY DEVICE
901 MAIN SCREEN
902 ANOMALY DETAILED SCREEN
903 CHECK SCREEN
Claims (12)
- An anomaly detection device comprising:
a condition generator configured to generate a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model;
a threshold setter configured to set a plurality of thresholds for the conditions; and
an anomaly detector configured to perform anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions. - The anomaly detection device according to claim 1, wherein
the condition generator uses a set of data including a plurality of explanatory variables based on the travel information of the vehicle and a class corresponding to the difference, and generates a classifier that associates a plurality of conditions with a plurality of the classes, the conditions being related to at least one of the explanatory variables and
the conditions associated with the classifier corresponds to the conditions for classifying the difference. - The anomaly detection device according to claim 2, wherein the threshold setter determines the threshold based on a distribution of the differences classified into the class by the classifier.
- The anomaly detection device according to claim 3, wherein the threshold setter generates a probability distribution of the differences and sets the threshold to be a value based on a standard deviation of the probability distribution or a value of the difference corresponding to a predetermined probability in the probability distribution.
- The anomaly detection device according to any one of claims 1 to 4, wherein the threshold setter receives an instruction to set the plurality of thresholds through a user interface and sets the thresholds based on the setting instruction.
- The anomaly detection device according to any one of claims 1 to 5, wherein the anomaly detector calculates the predicted value of the state of the vehicle based on a control command value corresponding to a first timing and the prediction model, specifies a condition satisfied by travel information corresponding to the first timing among the conditions, and detects whether the vehicle has an anomaly by comparing the difference between the predicted value and the measured value of the state of the vehicle with the threshold corresponding to the specified condition.
- The anomaly detection device according to any one of claims 1 to 6, further comprising a model generator configured to generate, for the conditions, a plurality of prediction models in each of which a control command value to the vehicle is associated with the state of the vehicle, wherein the anomaly detector performs anomaly detection on the vehicle based on the prediction models, the thresholds, and the conditions.
- The anomaly detection device according to claim 7, wherein the anomaly detector specifies a condition satisfied by travel information corresponding to a first timing among the conditions, calculates the predicted value of the state of the vehicle based on the prediction model corresponding to the specified condition and a control command value corresponding to the first timing, and detects whether the vehicle has an anomaly by comparing the difference between the predicted value and the measured value of the state of the vehicle with the threshold corresponding to the specified condition.
- The anomaly detection device according to any one of claims 1 to 8, wherein
the control command value is a command value related to the magnitude of a brake of the vehicle, and
the state includes a deceleration or an air brake pressure of the vehicle. - The anomaly detection device according to any one of claims 1 to 9, wherein the travel information includes at least one of measurement information of at least one of sensors of the vehicle and environment information of the vehicle.
- An anomaly detection method comprising:
generating a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model;
setting a plurality of thresholds for the conditions; and
performing anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions. - A computer program which causes a computer to perform processes comprising:
generating a plurality of conditions for classifying a difference between a predicted value of a state of a vehicle and a measured value of the state of the vehicle based on travel information of the vehicle, the predicted value being based on a control command value and a prediction model;
setting a plurality of thresholds for the conditions; and
performing anomaly detection on the vehicle based on the prediction model, the thresholds, and the conditions.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201880003324.7A CN109641603B (en) | 2017-07-19 | 2018-02-26 | Abnormality detection device, abnormality detection method, and computer program |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2017140204A JP6889059B2 (en) | 2017-07-19 | 2017-07-19 | Information processing equipment, information processing methods and computer programs |
JP2017-140204 | 2017-07-19 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019016996A1 true WO2019016996A1 (en) | 2019-01-24 |
Family
ID=61656284
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2018/007061 WO2019016996A1 (en) | 2017-07-19 | 2018-02-26 | Anomaly detection device, anomaly detection method, and computer program |
Country Status (4)
Country | Link |
---|---|
JP (1) | JP6889059B2 (en) |
CN (1) | CN109641603B (en) |
TW (1) | TWI688492B (en) |
WO (1) | WO2019016996A1 (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2580873A (en) * | 2018-11-02 | 2020-08-05 | Bombardier Transp Gmbh | Method and arrangement for safely operating at least one track-bound vehicle |
CN111780809A (en) * | 2020-07-21 | 2020-10-16 | 重庆大学 | Rail vehicle part temperature and vibration monitoring and early warning method and system |
US20210284209A1 (en) * | 2020-03-16 | 2021-09-16 | Kabushiki Kaisha Toshiba | Information processing apparatus and method |
CN114537361A (en) * | 2022-03-22 | 2022-05-27 | 中车制动系统有限公司 | Brake control system component failure prediction method |
Families Citing this family (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110287439A (en) * | 2019-06-27 | 2019-09-27 | 电子科技大学 | A kind of network behavior method for detecting abnormality based on LSTM |
DE102019210227A1 (en) * | 2019-07-10 | 2021-01-14 | Robert Bosch Gmbh | Device and method for anomaly detection in a communication network |
JP7206165B2 (en) * | 2019-07-11 | 2023-01-17 | 東芝インフラシステムズ株式会社 | Remote monitoring device |
CN112241511B (en) * | 2019-07-19 | 2024-05-14 | 上海杰之能软件科技有限公司 | Performance detection method and device for auxiliary electrical system of vehicle and terminal |
JP7269822B2 (en) * | 2019-08-06 | 2023-05-09 | 株式会社日立製作所 | Communication monitoring device and communication monitoring method |
CN110395242B (en) * | 2019-08-13 | 2020-06-16 | 中车青岛四方车辆研究所有限公司 | Air brake pressure control system and brake control method |
JP7363567B2 (en) * | 2019-09-20 | 2023-10-18 | 富士電機株式会社 | Model generation device, power conversion device and power control system |
CN111210604B (en) * | 2019-10-22 | 2021-06-15 | 高新兴创联科技有限公司 | Abnormal event early warning method based on railcar operation event correlation model |
JP7332152B2 (en) * | 2019-10-24 | 2023-08-23 | 日本電気株式会社 | LEARNING MODEL GENERATING METHOD, LEARNING MODEL GENERATING DEVICE, AND LEARNING PROGRAM |
JP7382791B2 (en) * | 2019-10-30 | 2023-11-17 | 株式会社日立製作所 | Abnormality determination device, vehicle support system |
KR20210054096A (en) * | 2019-11-04 | 2021-05-13 | 한국철도기술연구원 | a pantograph measuring device |
CN111708354B (en) * | 2020-07-06 | 2021-11-05 | 四川创客知佳科技有限公司 | Smart hotel terminal fault detection method based on big data |
JP7046148B1 (en) * | 2020-11-26 | 2022-04-01 | 三菱電機株式会社 | Diagnostic system, diagnostic program and diagnostic method |
CN112729099B (en) * | 2020-12-26 | 2022-05-17 | 北京磁浮交通发展有限公司 | Gap probe detection method and device, electronic equipment and computer storage medium |
CN114323706B (en) * | 2021-11-22 | 2024-04-12 | 卡斯柯信号有限公司 | Train ATO control fault detection method, device, equipment and medium |
WO2023181232A1 (en) * | 2022-03-24 | 2023-09-28 | 日本電気株式会社 | Model analysis device, model analysis method, and recording medium |
CN114972824B (en) * | 2022-06-24 | 2023-07-14 | 小米汽车科技有限公司 | Rod detection method, device, vehicle and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130325762A1 (en) * | 2010-09-28 | 2013-12-05 | Siemens Aktiengesellschaft | Adaptive remote maintenance of rolling stocks |
US20170178426A1 (en) * | 2015-12-17 | 2017-06-22 | Kabushiki Kaisha Toshiba | Abnormality diagnostic device and method therefor |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4592235B2 (en) * | 2001-08-31 | 2010-12-01 | 株式会社東芝 | Fault diagnosis method for production equipment and fault diagnosis system for production equipment |
JP2007026134A (en) * | 2005-07-19 | 2007-02-01 | Matsushita Electric Works Ltd | Abnormality decision device |
JP5685175B2 (en) * | 2011-10-26 | 2015-03-18 | 株式会社日立製作所 | Wireless train signal system and communication device |
JP2013100111A (en) * | 2013-03-07 | 2013-05-23 | Mitsubishi Electric Corp | Abnormal cause specifying device, abnormal cause specifying system, and abnormal cause specifying method |
JP2015162032A (en) * | 2014-02-27 | 2015-09-07 | 株式会社日立製作所 | Diagnostic device for traveling object |
SE538909C2 (en) * | 2014-04-15 | 2017-02-07 | Eber Dynamics Ab | Method and apparatus for determining structural parameters of a railway track |
EP2937241B1 (en) * | 2014-04-24 | 2017-03-08 | Hitachi, Ltd. | Railway vehicle damage estimation |
JP6293579B2 (en) * | 2014-06-02 | 2018-03-14 | 日本信号株式会社 | Orbit inspection device |
JP6509548B2 (en) * | 2014-12-15 | 2019-05-08 | 日本製鉄株式会社 | Rail vehicle condition monitoring device |
JP2017022853A (en) * | 2015-07-09 | 2017-01-26 | 株式会社東芝 | Train control device |
KR101832189B1 (en) * | 2015-07-29 | 2018-02-26 | 야마하하쓰도키 가부시키가이샤 | Abnormal image detecting apparatus, image processing system with abnormal image detecting apparatus and vehicle mounted with image processing system |
CN106203626A (en) * | 2016-06-30 | 2016-12-07 | 北京奇虎科技有限公司 | Car steering behavioral value method and device, automobile |
CN106652465B (en) * | 2016-11-15 | 2020-04-07 | 成都通甲优博科技有限责任公司 | Method and system for identifying abnormal driving behaviors on road |
CN106596123B (en) * | 2016-11-22 | 2020-03-27 | 东软集团股份有限公司 | Method, device and system for diagnosing equipment fault |
-
2017
- 2017-07-19 JP JP2017140204A patent/JP6889059B2/en active Active
-
2018
- 2018-02-26 WO PCT/JP2018/007061 patent/WO2019016996A1/en active Application Filing
- 2018-02-26 CN CN201880003324.7A patent/CN109641603B/en active Active
- 2018-03-13 TW TW107108389A patent/TWI688492B/en active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130325762A1 (en) * | 2010-09-28 | 2013-12-05 | Siemens Aktiengesellschaft | Adaptive remote maintenance of rolling stocks |
US20170178426A1 (en) * | 2015-12-17 | 2017-06-22 | Kabushiki Kaisha Toshiba | Abnormality diagnostic device and method therefor |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2580873A (en) * | 2018-11-02 | 2020-08-05 | Bombardier Transp Gmbh | Method and arrangement for safely operating at least one track-bound vehicle |
US20210284209A1 (en) * | 2020-03-16 | 2021-09-16 | Kabushiki Kaisha Toshiba | Information processing apparatus and method |
CN111780809A (en) * | 2020-07-21 | 2020-10-16 | 重庆大学 | Rail vehicle part temperature and vibration monitoring and early warning method and system |
CN111780809B (en) * | 2020-07-21 | 2023-09-08 | 重庆大学 | Rail vehicle part temperature and vibration monitoring and early warning method and system thereof |
CN114537361A (en) * | 2022-03-22 | 2022-05-27 | 中车制动系统有限公司 | Brake control system component failure prediction method |
Also Published As
Publication number | Publication date |
---|---|
JP2019018755A (en) | 2019-02-07 |
TW201908163A (en) | 2019-03-01 |
JP6889059B2 (en) | 2021-06-18 |
CN109641603A (en) | 2019-04-16 |
TWI688492B (en) | 2020-03-21 |
CN109641603B (en) | 2021-10-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109641603B (en) | Abnormality detection device, abnormality detection method, and computer program | |
CN109689470B (en) | Abnormality diagnosis device, abnormality diagnosis method, and computer program | |
US10571908B2 (en) | Autonomous vehicle failure mode management | |
CN109641602B (en) | Abnormality detection apparatus, abnormality detection method, and non-transitory computer-readable medium | |
WO2015139709A1 (en) | Method and system for monitoring a vehicle | |
TW202040444A (en) | Wear prediction apparatus, wear prediction method, and computer readable medium | |
KR101944070B1 (en) | Platform screen door management system and method with big data | |
CN111465972A (en) | System for calculating error probability of vehicle sensor data | |
US20200156680A1 (en) | Railway vehicle major component and system diagnosis apparatus | |
KR20220039531A (en) | Driving pattern classifying method of autonomous vehicles | |
EP3246223B1 (en) | State monitoring device, state monitoring system, and train | |
JP6382732B2 (en) | Condition monitoring device, condition monitoring system, and train train | |
CN112884233B (en) | Multimode fusion late prediction method for high-speed railway system | |
Wang et al. | Forecasting urban rail transit vehicle interior noise and its applications in railway alignment design | |
JPWO2017168940A1 (en) | Data integration analysis system | |
CN110414855A (en) | A kind of railcar safety evaluation method based on classification | |
KR102334220B1 (en) | Method and system of diagnosing defect and calculating remaining life of components of railway vehicle | |
CN117885753B (en) | Safety guarantee system of trackless rubber-tyred vehicle | |
KR20100056597A (en) | Apparatus, method and system for collecting information and information analysis apparatus using same | |
CN216118812U (en) | Cloud access port unified operation and maintenance system for supervision places | |
WO2021010074A1 (en) | Equipment diagnostic apparatus, equipment diagnostic system, and equipment diagnostic method | |
CN117131957A (en) | Machine learning-based engine oil low pressure fault prediction method and system | |
JP2021124854A (en) | Driving mode detection device, driving mode detection system, and driving mode detection method | |
CN117985026A (en) | Driving style analysis method and device and vehicle | |
CN118182256A (en) | Vehicle energy consumption management method, device, equipment and medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18711419 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 18711419 Country of ref document: EP Kind code of ref document: A1 |