WO2022024991A1 - Abnormality detection device, abnormality detection method, and abnormality detection program - Google Patents

Abstract

A normal model generation unit (106) generates a normal model by inputting, to a competitive neural network, data consisting of data to be learned, a first state observation value, and a first latent state value. The first state observation value and the first latent state value are obtained by inputting the data to be learned to a state observer and a qualitative soft sensor, respectively. An abnormality level calculation unit (108) calculates an abnormality level by comparing data consisting of data to be monitored, a second state observation value, and a second latent state value with the normal model. The second state observation value and the second latent state value are obtained by inputting the data to be monitored to a state observer and a qualitative soft sensor, respectively.

Description

Anomaly detection device, anomaly detection method and anomaly detection program Cross-reference of related applications

This international application claims priority based on Japanese Patent Application No. 2020-13038 filed with the Japan Patent Office on July 31, 2020, and Japanese Patent Application No. 2020-13038. The entire contents are referred to in this international application.

The present disclosure relates to anomaly detection devices, methods and programs for detecting anomalies using a competitive neural network.

In recent years, a method of detecting abnormalities in various devices using machine learning has been proposed. Patent Document 1 describes the presence or absence of abnormalities even when the normal state varies depending on the number of rotations, deterioration over time, date and time, season, etc. by learning the actual sensor observation values evenly using a competitive neural network. An abnormality monitoring device for proper diagnosis is described.

Japanese Unexamined Patent Publication No. 2007-198918

As a result of the inventor's detailed examination, when an abnormality is detected from the information and features that can be read from the observed values of the actual sensor, there is a problem that an erroneous judgment may occur in which the abnormality is determined to be normal even though it is an abnormality. Was found.

This disclosure improves the accuracy of abnormality detection.

One aspect of the present disclosure is an abnormality detection device including a data acquisition unit, a normal model generation unit, an abnormality degree calculation unit, and a determination unit.

The data acquisition unit is configured to acquire the learning target data and the monitoring target data of the abnormality detection target, which is the target for detecting the abnormality.

The normal model generator represents the learning target data, the first state observation value obtained by inputting the training target data into the state observer representing the state of the system of the abnormality detection target, and the latent state of the abnormality detection target. It is configured to generate a normal model by inputting data composed of a first latent state value obtained by inputting training target data into a qualitative soft sensor into a competitive neural network.

The anomaly degree calculation unit has the monitored data, the second state observed value obtained by inputting the monitored data into the state observer, and the second latent value obtained by inputting the monitored data into the qualitative soft sensor. By comparing the data composed of the state values with the normal model, it is configured to calculate the degree of abnormality indicating the degree of abnormality of the abnormality detection target.

The determination unit is configured to determine whether the abnormality detection target is normal or abnormal based on the abnormality degree calculated by the abnormality degree calculation unit.

The anomaly detection device of the present disclosure configured in this way adds the first latent state value to the training target data and the first state observation value to generate a normal model, and adds the second state observation value to the monitoring target data and the second state observation value. Calculate the degree of anomaly by adding the latent state value. Thereby, the abnormality detection device of the present disclosure can detect the abnormality by the normal model learned including the state of the abnormality detection target, and can improve the abnormality detection accuracy.

It is a block diagram which shows the structure of an abnormality detection apparatus. It is a functional block diagram which shows the functional configuration of an abnormality detection apparatus. It is a figure which shows the structure of the drive system state observer. It is a figure which shows the structure of the air-fuel ratio system state observer. It is a flowchart which shows the learning process of 1st Embodiment. It is a flowchart which shows the qualitative soft sensor generation processing. It is a figure explaining the generation method of a qualitative soft sensor. It is a flowchart which shows the state observer generation process. It is a figure explaining the maximization of the correlation between state observers. It is a figure which shows the normal surface of 1st Embodiment. It is a flowchart which shows the monitoring process of 1st Embodiment. It is a flowchart which shows the abnormality degree calculation process. It is a timing chart of the anomaly degree, the first latent state and the second latent state. It is a flowchart which shows the learning process of 2nd Embodiment. It is a flowchart which shows the monitoring process of 2nd Embodiment. It is a timing chart of the intake air amount and the load amount, and the graph about the intake air amount and the load amount as an axis. It is a figure which shows the normal surface of the 2nd Embodiment. It is a figure explaining the abnormality determination method of 2nd Embodiment. It is a flowchart which shows the learning process of 3rd Embodiment. It is a flowchart which shows the parameter setting process. It is a flowchart which shows the clustering coefficient setting process. It is a figure explaining the method of searching for the optimum solution. It is a figure which shows that an abnormality degree is expressed by both a positive value and a negative value. It is a figure which shows the structure of the drive system state observer including the latent state value.

[First Embodiment]
The first embodiment of the present disclosure will be described below together with the drawings.

The abnormality detection device 1 of the present embodiment is mounted on a vehicle and includes a control device 2, a communication device 3, a data storage device 4, and a display device 5, as shown in FIG.

The control device 2 is an electronic control device mainly composed of a microcomputer equipped with a CPU 11, ROM 12, RAM 13, and the like. Various functions of the microcomputer are realized by the CPU 11 executing a program stored in a non-transitional substantive recording medium. In this example, ROM 12 corresponds to a non-transitional substantive recording medium in which a program is stored. Also, by executing this program, the method corresponding to the program is executed. In addition, a part or all of the functions executed by the CPU 11 may be configured in terms of hardware by one or a plurality of ICs or the like. Further, the number of microcomputers constituting the control device 2 may be one or a plurality.

The communication device 3 is connected to a plurality of ECUs via a communication line so as to be capable of data communication, and transmits / receives data according to the CAN communication protocol. Specifically, the plurality of ECUs connected to the communication device 3 are an engine ECU that performs engine control, a brake ECU that performs brake control, a steering ECU that performs steering control, and the like. CAN is an abbreviation for Controller Area Network. CAN is a registered trademark.

The data storage device 4 is a device for storing various data, and is, for example, a hard disk drive in this embodiment.

The display device 5 displays various images on the display screen based on the instruction from the control device 2.

As shown in FIG. 2, the control device 2 has a data acquisition unit 101, a state observer generation unit 104, a normal model generation unit 106, and an abnormality degree calculation unit as a configuration of functions realized by the CPU 11 executing a program. It includes 108, a determination unit 109, a factor analysis unit 110, and a display unit 111. The method for realizing these elements constituting the control device 2 is not limited to software, and a part or all of the elements may be realized by using one or a plurality of hardware. For example, when the above function is realized by an electronic circuit which is hardware, the electronic circuit may be realized by a digital circuit including a large number of logic circuits, an analog circuit, or a combination thereof.

Further, the data storage device 4 includes a learning target data storage unit 102, a monitoring target data storage unit 103, a state observer information storage unit 105, and a normal model parameter storage unit 107.

The data acquisition unit 101 acquires learning target data and monitoring target data from various sensors directly or indirectly connected to the abnormality detection device 1.

Various sensors include sensors mounted on vehicles and sensors connected to various vehicle electronic control devices. Examples of the sensor mounted on the vehicle include a thermometer, a hygrometer, and a GPS. Examples of the sensor connected to the vehicle electronic control device include an engine rotation sensor, a turbine rotation sensor, an oxygen sensor, an air-fuel ratio sensor, and the like.

The data acquisition unit 101 may indirectly acquire the learning target data and the monitoring target data via the network. For example, the data acquisition unit 101 may acquire the learning target data and the monitoring target data by downloading from the database via the network.

The learning target data and the monitoring target data are data related to at least one of the vehicle failure and the vehicle failure mechanism, and include data related to the vehicle and data related to the driving environment. However, the learning target data is data when the vehicle is in a normal state. The monitored data is data including both the case where the vehicle is in a normal state and the case where the vehicle is in an abnormal state.

The learning target data storage unit 102 stores the learning target data acquired by the data acquisition unit 101.

The monitoring target data storage unit 103 stores the monitoring target data acquired by the data acquisition unit 101.

The state observer generation unit 104 acquires a variable configuration and uses the variables included in the acquired variable configuration to generate a state observer for each vehicle system.

The state observer is also called an observer or a soft sensor. The variables correspond to the learning target data and the monitoring target data acquired by the data acquisition unit 101. The system is formed by grouping the control devices and sensors constituting the vehicle for each function of the vehicle. Examples of the vehicle system include a drive system, an air-fuel ratio system, an air system, an ignition system, a fuel system, and a throttle system.

The state observer generator 104 may acquire a variable configuration necessary or useful for the state observer to express the function and failure mechanism of the vehicle based on the knowledge of the setter. The variable configuration is one variable used to construct the state observer, or a combination of a plurality of variables used to construct the state observer.

The state observer generation unit 104 generates a state observer by linearly combining the variables included in the acquired variable configuration, for example, as shown in the equation (1). A _p X _p in the equation (1) is the p-th state observer. _xpi is the i-th variable in the p-th state observer. a _pi is the i-th coefficient in the p-th state observer. That is, a _pi is a coefficient of the variable x _pi . i is an integer from 1 to n. The state observer generation unit 104 sets the initial value of the coefficient a _pi to a random value and combines the variables x _pi to generate the initial state of the state observer.

It is desirable that the state observer reflects the function of the device to be detected by the abnormality detection device 1 or the failure mechanism of the device. In this embodiment, as an example of the function of the device, the driving of the vehicle and the air-fuel ratio of the vehicle are included.

The device failure mechanism is a mechanism that causes a device failure. "Reflecting the failure mechanism of the equipment" means that the state observation value output from the state observer changes according to the function of the equipment or the failure mechanism of the equipment.

It is desirable that the variables included in the variable configuration of the state observer are factors of the function of the equipment to be detected by the abnormality detection device 1 or factors of the failure mechanism of the equipment. For example, if the equipment failure mechanism is engine overheating, engine temperature is a direct factor in causing overheating. Further, the engine speed and the amount of the coolant are direct factors that cause the temperature rise of the engine, that is, indirect factors that cause overheating.

For example, as shown in FIG. 3, a state observer (hereinafter referred to as a drive system state observer) in which the drive system of a vehicle is a function of an apparatus is generated with the engine rotation speed and the turbine rotation speed as variables. Specifically, the drive system state observer a ₁ X ₁ , which is the first state observer, sets the engine rotation speed and the turbine rotation speed as variables x ₁₁ and variables x ₁₂ , respectively, as shown in the equation (2). , Generated by linear coupling with coefficients _a11 and _a12 .

Further, as shown in FIG. 4, a state observer having the air-fuel ratio system of the vehicle as a function of the device (hereinafter referred to as an air-fuel ratio system state observer) is generated with the oxygen sensor voltage and the air-fuel ratio sensor current as variables. .. Specifically, the air-fuel ratio system state observer a ₂ X ₂ , which is the second state observer, changes the oxygen sensor voltage and the air-fuel ratio sensor current into variables x ₂₁ and variables x, respectively, as shown in the equation (3). As ₂₂ , it is generated by a linear coupling using the coefficients a ₂₁ and the coefficients a ₂₂ .

The state observer may be generated by using variables other than the above-mentioned variables. Further, the number of variables to be combined is not limited to two, and may be three or more. Further, the number of state observers to be generated is not limited to two, and may be one or three or more.

In this embodiment, as an example of the state observer, an example having a two-layer structure of input and output is given, but three or more layers may be used. In the case of three or more layers, it may be described using a hierarchical neural network.

Also, as a state observer, an example of linearly connecting variables was given, but the present invention is not limited to this, and may be composed of non-linear coupling. For example, when describing using a hierarchical neural network, since a sigmoid function is used, it is composed of non-linear coupling.

By generating a state observer, it is possible to estimate state observation values that cannot be directly measured by a sensor or the like. Further, by reflecting the function of the device and the failure mechanism in the state observer, the cause analysis of the failure becomes easy.

The state observer generator 104 generates a qualitative soft sensor, which will be described later. Here, "qualitative" data refers to data that cannot be directly measured numerically, and is a type of data to which a nominal scale and an ordinal scale belong.

The state observer information storage unit 105 stores the coefficients and variable configurations of the state observer and the qualitative soft sensor generated by the state observer generation unit 104.

The abnormality detection device 1 may include a state observer generation unit 104 when generating a state observer. That is, even if the anomaly detection device 1 generates a state observer and stores the variable configuration and coefficient of the state observer in the state observer information storage unit 105, the state observer generation unit 104 is separated from the anomaly detection device 1. good.

The normal model generation unit 106 reads out the learning target data from the learning target data storage unit 102, and reads out the variable configurations and coefficients of the state observer and the qualitative soft sensor from the state observer information storage unit 105. Then, the normal model generation unit 106 calculates the first state observation value by applying the training target data to the variables constituting the state observer for each state observer. Further, the normal model generation unit 106 calculates the first latent state value by applying the learning target data to the variables constituting the qualitative soft sensor for each qualitative soft sensor.

The normal model generation unit 106, for example, substitutes the engine rotation speed and the turbine rotation speed acquired as learning target data into the equation (2), and uses the values obtained by substituting the values obtained by substituting the values obtained by substituting the engine rotation speed and the turbine rotation speed into the first state observation of the drive system state observer a ₁ X ₁ . Use as a value. Further, the normal model generation unit 106 substitutes the oxygen sensor voltage and the air fuel ratio sensor current acquired as the learning target data into the equation (3), and uses the values obtained by substituting the values obtained by substituting the values obtained by substituting the values obtained by substituting the oxygen sensor voltage and the air fuel ratio sensor current into the _{first state of the air fuel ratio system state observer a2 X 2 .} Use as an observed value.

Then, the normal model generation unit 106 combines the first state observed value, the first latent state value, and the learning target data. For example, the normal model generator 106 may include engine speed, turbine speed, first state observed value of drive system state observer a ₁ X ₁ , oxygen sensor voltage, air fuel ratio sensor current, air fuel ratio system state observer a ₂ . The seven data of the _first state observed value of X2 and the first latent state value are combined. The combination of the first state observed value, the first latent state value, and the training target data should be such that the first state observed value, the first latent state value, and the training target data can be input to the competitive neural network at the same time. Just do it.

Further, the normal model generation unit 106 learns the normal surface as a normal model by inputting the combined data into the competitive neural network. A competitive neural network is a network having two layers, an input layer and an output layer, and is composed of a plurality of input layer neurons and a plurality of output layer neurons fully connected to the input layer neurons. ing. The weight data of each output layer neuron corresponds to the normal plane.

The initial values given to the competitive neural network are, for example, evenly sampled or randomly sampled or randomly when there are multiple combinations of data attribute attributes to be input, such as vehicle type, measurement season, day and night, customized specifications, and age. It is desirable to sample. This makes it possible to accelerate the convergence of the weight vectors of neurons on the map of the competitive neural network during learning.

In the present embodiment, the degree of abnormality, which is the difference between the learning target data, the first state observed value and the first latent state value, and the neuron weight data of the winning unit, which are input to the competitive neural network, is calculated. Then, the threshold value is calculated using the set of differences. For example, a constant multiple of the 99.9% quantile of a set of differences (or absolute values of differences) is used as the threshold.

The normal model parameter storage unit 107 stores learned parameters representing the normal surface learned by the normal model generation unit 106. Further, the normal model parameter storage unit 107 stores the threshold value calculated by the normal model generation unit 106.

The abnormality degree calculation unit 108 reads the monitored data from the monitored data storage unit 103, and reads the variable configurations and coefficients of the state observer and the qualitative soft sensor from the state observer information storage unit 105. Then, the abnormality degree calculation unit 108 calculates the second state observation value by applying the monitored data to the variables constituting the state observer for each state observer. Further, the abnormality degree calculation unit 108 calculates the second latent state value by applying the monitored data to the variables constituting the qualitative soft sensor for each qualitative soft sensor.

The anomaly degree calculation unit 108, for example, substitutes the engine rotation speed and the turbine rotation speed acquired as monitoring target data into the equation (2) and obtains the values obtained by observing the second state of the drive system state observer a ₁ X ₁ . Use as a value. Further, the abnormality degree calculation unit 108 substitutes the oxygen sensor voltage and the air fuel ratio sensor current acquired as the monitored data into the equation (3), and uses the values obtained by substituting the values obtained by substituting the values obtained by substituting the values obtained by substituting the oxygen sensor voltage and the air fuel ratio sensor current into the _{second state of the air fuel ratio system state observer a2 X 2 .} Use as an observed value.

Then, the abnormality degree calculation unit 108 combines the second state observed value, the second latent state value, and the monitored data. For example, the abnormality degree calculation unit 108 includes the engine speed, the turbine speed, the second state observed value of the drive system state observer a ₁ X ₁ , the oxygen sensor voltage, the air fuel ratio sensor current, and the air fuel ratio system state observer a ₂ . The 7 data of the _2nd state observed value of X2 and the 2nd latent state value are combined. The combination of the second state observed value, the second latent state value, and the monitored data should be such that the second state observed value, the second latent state value, and the monitored data can be input to the competitive neural network at the same time. Just do it.

Further, the abnormality degree calculation unit 108 reads learned parameters such as the normal surface (that is, neuron weight data) from the normal model parameter storage unit 107. Then, the abnormality degree calculation unit 108 calculates the distance between the monitored data, the second state observed value and the second latent state value and the normal surface as the abnormality degree by inputting the combined data into the competitive neural network. ..

The determination unit 109 determines whether the vehicle is normal or abnormal based on the threshold value read from the normal model parameter storage unit 107 and the abnormality degree calculated by the abnormality degree calculation unit 108.

When the determination unit 109 determines that the vehicle is abnormal, the factor analysis unit 110 uses the second state observed value, the second state observed value, and the monitored data that caused the determination of the abnormality. Identify the cause of the anomaly.

The display unit 111 displays information for identifying the cause of the abnormality on the display screen of the display device 5.

Next, the procedure of the learning process executed by the CPU 11 of the control device 2 will be described. The learning process is a process that is repeatedly executed during the operation of the control device 2.

When the learning process is executed, the CPU 11 first acquires learning target data from various sensors in S10, as shown in FIG.

Further, in S20, the CPU 11 performs a distribution conversion to bring the learning target data acquired in S10 closer to a normal distribution. Examples of distribution transformations include the Bоx-Cоx transformation and the Johnson transformation. The distribution conversion may improve the determination accuracy in the determination unit 109.

Then, in S30, the CPU 11 stores the learning target data after the distribution conversion in the learning target data storage unit 102.

Next, the CPU 11 executes a qualitative soft sensor generation process in S40. Here, the procedure of the qualitative soft sensor generation process will be described.

When the qualitative soft sensor generation process is executed, the CPU 11 first generates a co-occurrence matrix in S210 as shown in FIG.

As shown in FIG. 7, the co-occurrence matrix is composed of a plurality of flags indicating changes in the state of the equipment mounted on the vehicle (hereinafter referred to as in-vehicle equipment). The co-occurrence matrix CM1 is composed of flag A, flag B, flag C and flag D. The state of the in-vehicle device includes on / off of the headlamp, gear of the transmission (for example, 1st speed, 2nd speed, 3rd speed, etc.), on / off of the ventilation of the car air conditioner, and the amount of ventilation of the car air conditioner (for example, weak). Medium, strong, etc.).

Flags D "1", "2", and "3" indicate "weak", "medium", and "strong" in the air volume of the car air conditioner, respectively. As shown by the co-occurrence matrix CM2, the flags D1, the flag D2, and the flag D3 may be used instead of the flag D. The flag D1 is a flag set when the air volume of the car air conditioner is "weak". The flag D2 is a flag set when the air volume of the car air conditioner is "medium". The flag D3 is a flag set when the air volume of the car air conditioner is "strong".

As the state of the in-vehicle device, instead of the value of the flag, the rising edge from 0 to 1 in the flag or the rising edge from 1 to 0 in the flag may be counted. Further, a co-occurrence matrix may be generated by using a time window or the like.

When the generation of the co-occurrence matrix is completed, the CPU 11 generates a variance-covariance matrix based on the cosine similarity for the co-occurrence matrix generated in S210 in S220 as shown in FIG. The CPU 11 may generate a variance-covariance matrix based on mutual information, or may generate a variance-covariance matrix without preprocessing.

Next, the CPU 11 generates a coefficient (for example, a singular value) in a normal state by performing SVD on the variance-covariance matrix generated in S220 in S230. SVD is an abbreviation for Singular Value Decomposition. The CPU 11 may generate the coefficient in the normal state by using vari / promax rotation, Word2Vec, or the like instead of SVD.

Further, the CPU 11 performs dimension reduction in S240 based on the contribution rate calculated by the SVD of S230. Then, the CPU 11 generates a qualitative soft sensor in S250 using the eigenvector obtained by the dimension reduction in S240.

The qualitative soft sensor is formed by, for example, a linear combination of a plurality of flags. For example, as shown in FIG. 7, the qualitative soft sensor QS1 indicating the first latent state is formed by a linear combination of the flag A and the flag B. The qualitative soft sensor QS2 indicating the second latent state is formed by a linear combination of the flag C, the flag D1 and the flag D2. The first latent state indicates, for example, that the vehicle is before warming up, and the second latent state indicates, for example, that the vehicle is after warming up.

When the processing of S250 is completed, as shown in FIG. 6, the CPU 11 stores the coefficient and variable configuration of the qualitative soft sensor generated in S250 in the state observer information storage unit 105 in S260, and qualitatively. The soft sensor generation process is terminated.

When the qualitative soft sensor generation process is completed, the CPU 11 executes the state observer generation process in S50 as shown in FIG. Here, the procedure of the state observer generation process will be described.

When the state observer generation process is executed, the CPU 11 first acquires the variable configuration in S310 as shown in FIG.

Then, the CPU 11 generates a state observer in S320 by linearly combining the variables included in the acquired variable configuration, for example, as shown in the equation (1).

Next, the CPU 11 determines in S330 whether or not the number of state observers generated in S320 is one. Here, when the number of state observers is one, the CPU 11 shifts to S370.

On the other hand, when the number of state observers is not one, the CPU 11 determines in S340 whether or not the number of state observers generated in S320 is two. Here, when the number of state observers is two, the CPU 11 maximizes the correlation between the two generated state observers in S350, and shifts to S270. Hereinafter, of the two state observers generated in S320, one state observer is referred to as a first state observer, and the other state observer is referred to as a second state observer.

Specifically, the CPU 11 has the coefficient of the first state observer and the coefficient of the second state observer so that the correlation coefficient ρ shown in the equation (4) is maximized under the constraint conditions shown in the equation (5). And adjust. Equation (5) is a constraint of Lagrange's undetermined multiplier method. n and m are the number of the state observer, and V is the variance. The molecule of equation (4) is the sample covariance of the first state observer and the second state observer, and the denominator of equation (4) is the square root of the sample dispersion of the first state observer and the second state observer. It is the product of the square root of the sample variance.

If the number of state observers is not two in S340, the CPU 11 maximizes the correlation between the three or more state observers generated in S320 in S360 and shifts to S370. For example, when the first state observer, the second state observer, and the third state observer are generated in S320, the sum of the correlations between the two state observers is maximized as shown in FIG. It is desirable to make it. Specifically, equation (6) is used to maximize the sum of the correlations between the two state observers. N is the total number of state observers. g is a function representing the sum, and the inside of the parentheses of g is the object of the sum. Even if there are four or more state observers, maximization is performed in the same manner as in the case of three. In the present embodiment, the "sum" may include the sum in the calculation such as the sum of squares and the sum of absolute values, in addition to the simple sum.

After shifting to S370, the CPU 11 stores the coefficients and variable configurations of the state observer generated by the processes of S320 to S360 in the state observer information storage unit 105, and ends the state observer generation process.

When the state observer generation process is completed, the CPU 11 reads out the training target data, the variable configuration and the coefficient in S60 as shown in FIG. 5, and for each state observer, for the variables constituting the state observer. The first state observation value is calculated by applying the training target data.

Further, the CPU 11 reads out the training target data, the variable configuration, and the coefficient in S70, applies the training target data to the variables constituting the qualitative soft sensor for each qualitative soft sensor, and obtains the first latent state value. calculate.

Then, in S80, the CPU 11 combines the learning target data acquired in S10, the first state observed value calculated in S60, and the first latent state value calculated in S70.

Next, in S90, the CPU 11 learns the normal surface as a normal model by inputting the data combined in S80 into the competitive neural network.

As shown in FIG. 10, the initialized normal plane PL1 is a SOM in which neurons are arranged in a two-dimensional grid pattern. SOM is an abbreviation for Self-Organizing Map. Each neuron is composed of learning target data, a first state observed value, and a first latent state value. For example, each neuron has engine speed, turbine speed, first state observer of drive system state observer a ₁ X ₁ , oxygen sensor voltage, air fuel ratio sensor current, air fuel ratio system state observer a ₂ X ₂ . It is composed of seven data, the first state observed value and the first latent state value. Hereinafter, each data constituting the neuron is referred to as neuron weight data.

By learning, the weights of the neurons that make up the normal surface change from the initial value, and the learned normal surface PL2 is obtained.

When the learning of the normal surface is completed, as shown in FIG. 5, the CPU 11 stores the learned parameters representing the normal surface learned in S90 in the normal model parameter storage unit 107 in S100, and ends the learning process.

Next, the procedure of the monitoring process executed by the CPU 11 of the control device 2 will be described. The monitoring process is a process that is repeatedly executed after the learning of the normal surface is completed.

When the monitoring process is executed, the CPU 11 first acquires monitoring target data from various sensors in S410, as shown in FIG.

Further, the CPU 11 performs distribution conversion in S420 to bring the monitored data acquired in S410 closer to a normal distribution.

Then, in S430, the CPU 11 stores the monitored data after the distribution conversion in the monitored data storage unit 103.

Next, the CPU 11 reads out the monitored data, the variable configuration, and the coefficient in S440, applies the monitored data to the variables constituting the state observer for each state observer, and calculates the second state observed value. do.

Further, the CPU 11 reads out the monitored data, the variable configuration and the coefficient in S450, applies the monitored data to the variables constituting the qualitative soft sensor for each qualitative soft sensor, and sets the second latent state value. calculate.

Then, the CPU 11 combines the monitored data acquired in S410 in S460, the second state observed value calculated in S440, and the second latent state value calculated in S450. Hereinafter, the combined monitored data, the second state observed value, and the second latent state value are collectively referred to as verification data.

Further, the CPU 11 acquires the learned parameters from the normal model parameter storage unit 107 in S470.

Then, the CPU 11 executes the abnormality degree calculation process in S480. Here, the procedure of the abnormality degree calculation process will be described.

When the abnormality degree calculation process is executed, the CPU 11 first sets the value of the loop counter l to 1 in S610, as shown in FIG.

Then, in S620, the CPU 11 first calculates the Euclidean distance d _{k, l, i, j} between the verification data Z _k at the monitoring time point k and the neurons _{Wi, j} constituting the normal plane. i is a number indicating the lateral position of the neuron on the normal plane, and is an integer of 1 to N. N indicates the horizontal size of the normal surface. j is a number indicating the vertical position of the neuron on the normal plane, and is an integer of 1 to M. M indicates the vertical size of the normal surface.

The verification data Z _k is composed of X data as shown in the equation (7). As shown in Eq. (8), the neurons _{Wi and j} are composed of X weight data.

Further, the CPU 11 sets the minimum Euclidean distance d _{k, l} from the Euclidean distances d _{k, l, i, j} calculated in all the neurons Wi _{, j} in S620.

Further, the CPU 11 calculates in S630 the cosine similarity cos θ _{k, l} between the verification data Z _k and the neurons W'k _{, l} closest to the verification data Z _k .

Further, the CPU 11 calculates in S640 the product of the minimum Euclidean distance d _{k, l} and the cosine similarity cos θ _{k, l} as the abnormality degree c _{k, l} .

Then, the CPU 11 determines in S650 whether or not the value of the loop counter l exceeds the preset end determination value L (for example, 10). Here, if the value of the loop counter l does not exceed the end determination value L, the CPU 11 increments (that is, adds 1) the loop counter l in S660. Further, in S670, the CPU 11 removes the neurons W'k _{, l} closest to the verification data Z _k from the neurons Wi _{, j} constituting the normal plane, and shifts to S620.

Further, in S650, when the value of the loop counter l exceeds the end determination value L, the CPU 11 calculates in S680 the sum of the L abnormality degrees c _{k and l} as the total abnormality degree c _k . Then, the abnormality degree calculation process is terminated.

When the abnormality degree calculation process is completed, the CPU 11 determines the abnormality in S490 and ends the monitoring process, as shown in FIG. Specifically, the CPU 11 reads a threshold value from the normal model parameter storage unit 107, and determines whether or not the total abnormality degree _ck calculated in S680 is equal to or less than the threshold value. When the total abnormality degree c _k is equal to or less than the threshold value, the CPU 11 determines that the vehicle is normal. When the total abnormality degree c _k exceeds the threshold value, the CPU 11 determines that the vehicle is abnormal.

The abnormality detection device 1 configured in this way includes a data acquisition unit 101, a normal model generation unit 106, an abnormality degree calculation unit 108, and a determination unit 109.

The data acquisition unit 101 acquires the learning target data and the monitoring target data of the vehicle.

The normal model generation unit 106 generates a normal model by inputting data composed of learning target data, a first state observed value, and a first latent state value into a competitive neural network. The first state observation value is obtained by inputting the learning target data into the state observer representing the state of the system possessed by the vehicle. The first latent state value is obtained by inputting the learning target data into the qualitative soft sensor representing the latent state of the vehicle.

The abnormality degree calculation unit 108 calculates the degree of abnormality indicating the degree of abnormality of the vehicle by comparing the data composed of the monitored data, the second state observed value, and the second latent state value with the normal model. do. The second state observation value is obtained by inputting the monitored data into the state observer. The second latent state value is obtained by inputting the monitored data into the qualitative soft sensor.

The determination unit 109 determines whether the vehicle is normal or abnormal based on the abnormality degree calculated by the abnormality degree calculation unit 108.

In this way, the abnormality detection device 1 adds the first latent state value to the training target data and the first state observed value to generate a normal model, and adds the second latent state value to the monitored data and the second state observed value. To calculate the degree of abnormality. As a result, the abnormality detection device 1 can detect the abnormality by the normal model learned including the state of the vehicle, and can improve the abnormality detection accuracy.

Further, as shown in FIG. 13, for example, the abnormality detection device 1 can visualize from what kind of data background the abnormality occurred at each monitoring time point and perform factor analysis. In FIG. 13, the abnormality detection device 1 can perform factor analysis in consideration of the fact that the values of the first latent state and the second latent state are not 0 at the monitoring time point t1 and the monitoring time point t2.

The qualitative soft sensor is generated by combining a plurality of flags indicating the states of the components (for example, headlamps, transmissions, car air conditioners) constituting the vehicle. As a result, the anomaly detection device 1 can generate a variable representing a latent state by dimensional compression, and can acquire a variable representing a data state robust to noise and the like.

In the embodiment described above, the vehicle corresponds to the abnormality detection target, S10 and S410 correspond to the processing as the data acquisition procedure, S90 corresponds to the processing as the normal model generation procedure, and S480 corresponds to the abnormality degree calculation procedure. Corresponds to the process of, and S490 corresponds to the process as the determination procedure.

[Second Embodiment]
The second embodiment of the present disclosure will be described below together with the drawings. In the second embodiment, a part different from the first embodiment will be described. The same reference numerals are given to common configurations.

The abnormality detection device 1 of the second embodiment is different from the first embodiment in that the learning process and the monitoring process are changed.

The learning process of the second embodiment is different from the first embodiment in that the processes of S45 and S75 are added and the processes of S85 and S95 are executed instead of S80 and S90.

That is, as shown in FIG. 14, when the processing of S40 is completed, the CPU 11 generates a Mahalanobis distance calculation parameter in S45. Specifically, the CPU 11 calculates the sample mean and the sample covariance matrix as the Mahalanobis distance calculation parameters for each state observer with the training target data corresponding to the variables constituting the state observer. For example, in the drive train state observer, the Mahalanobis distance calculation parameter is calculated with the engine speed and the turbine speed as variables. The sample average is calculated by Eq. (9). The sample covariance matrix is calculated by Eq. (10). X ⁽ⁿ⁾ in the equations (9) and (10) is a variable constituting the state observer. N in the equations (9) and (10) is the number of data to be learned.

When the processing of S45 is completed, the process proceeds to S50.

Further, when the processing of S70 is completed, the CPU 11 calculates the first index by the equation (11) in S75 using the learning target data acquired in S10. The left side of the equation (11) is the first index. The right side of the equation (11) is an equation for calculating the Mahalanobis distance.

When the processing of S75 is completed, the CPU 11 has the learning target data acquired in S10 in S85, the first state observed value calculated in S60, the first latent state value calculated in S70, and the first latent state value calculated in S75. Combine with 1 index.

Next, in S95, the CPU 11 learns the normal surface as a normal model by inputting the data combined in S85 into the competitive neural network, and shifts to S100.

The monitoring process of the second embodiment is different from the first embodiment in that the process of S455 is added and the process of S465 is executed instead of S460.

That is, as shown in FIG. 15, when the processing of S450 is completed, the CPU 11 calculates the second index by the equation (11) using the monitored data acquired in S410 in S455.

Then, the CPU 11 combines the monitored data acquired in S410 in S465, the second state observed value calculated in S440, the second latent state value calculated in S450, and the second index calculated in S455. , S470. In the second embodiment, the combined monitored data, the second state observed value, the second latent state value, and the second index are collectively referred to as verification data.

Next, learning of the normal surface using the Mahalanobis distance and abnormality detection using the Mahalanobis distance will be explained.

The terming chart TC1 of FIG. 16 shows the time change of the intake air amount and the load value, which are variables constituting the state observer of the vehicle air system. The terming chart TC1 shows the time change of the intake air amount and the load value acquired as the learning target data in the period T1, and shows the time change of the intake air amount and the load value acquired as the monitoring target data in the period T2. show.

In regions R1 and R2, there is a negative correlation between the intake air amount and the load value. This is normal, only the latent state has changed.

In region R3, the load value is constant while the intake air amount is low. This is a sticking abnormality of the load value. In the regions other than the regions R1, R2 and R3, there is a positive correlation between the intake air amount and the load value. This is normal.

Graphs GR1 and GR2 show the distribution of the intake air amount and the load value in the terming chart TC1 with the vertical axis as the intake air amount and the horizontal axis as the intake air amount.

A plurality of points in the region R4 in the graphs GR1 and GR2 correspond to the data in the regions R1 and R2 in the terming chart TC1. The plurality of points in the region R5 in the graphs GR1 and GR2 correspond to the data in the region R3 in the terming chart TC1. The plurality of points other than the regions R4 and R5 in the graphs GR1 and GR2 correspond to the data other than the regions R1, R2 and R3 in the terming chart TC1.

The plurality of ellipses in Graph GR1 are equidistant lines of Mahalanobis distance. The plurality of circles in the graph GR2 are equidistant lines of the Euclidean distance.

As shown in FIG. 17, the initialized normal plane PL3 is a SOM in which neurons are arranged in a two-dimensional grid pattern. Each neuron is composed of learning target data, a first state observed value, a first latent state value, and a first index. For example, each neuron has engine speed, turbine speed, first state observer of drive system state observer a ₁ X ₁ , oxygen sensor voltage, air fuel ratio sensor current, air fuel ratio system state observer a ₂ X ₂ . It is composed of eight data, a first state observed value, a first latent state value, and a first index. By learning, the weights of the neurons constituting the normal plane are changed from the initial values, and the trained normal plane PL4 is obtained.

By learning the index calculated using the Mahalanobis distance together with the latent state value, as shown in FIG. 18, it is determined that a plurality of points in the region R4 are normal and a plurality of points in the region R5 are abnormal. Is possible. The region R6 in FIG. 18 is a region determined to be normal. The relationship between the intake air amount and the load value is checked by the index, but the exception condition is set by SOM.

In the abnormality detection device 1 configured in this way, the normal model generation unit 106 calculates the first index, and uses the learning target data, the first state observation value, the first latent state value, and the first index. A normal model is generated by inputting the constructed data into the competitive neural network. The first index represents a change in the relationship between a plurality of variables constituting the state observer and is obtained by inputting learning target data.

The anomaly degree calculation unit 108 calculates the second index, and compares the data composed of the monitored data, the second state observed value, the second latent state value, and the second index with the normal model. Calculate the degree of anomaly. The second index is obtained by inputting monitored data representing changes in the relationships of a plurality of variables constituting the state observer.

As a result, the anomaly detection device 1 can detect anomalies using a normal model learned including the relationships of a plurality of variables constituting the state observer, and can improve the anomaly detection accuracy.

The first and second indicators are Mahalanobis distances. As a result, the anomaly detection device 1 can detect the relationship change that was difficult to detect by the Euclidean distance standard.

In the Mahalanobis distance, when the two variables that make up the state observer rarely collapse from the linear relationship, the index may become large even if it is normal. However, by letting the neuron learn the index when the linear relationship collapses at normal times together with other values using a competitive neural network, it is possible to judge including other values even if the index is high. Species error can be deterred.

[Third Embodiment]
The third embodiment of the present disclosure will be described below together with the drawings. In the third embodiment, a part different from the first embodiment will be described. The same reference numerals are given to common configurations.

The abnormality detection device 1 of the third embodiment is different from the first embodiment in that the learning process is changed.

The learning process of the third embodiment is different from the first embodiment in that the process of S35 is added.

That is, as shown in FIG. 19, when the processing of S30 is completed, the CPU 11 executes the parameter setting processing in S35. Here, the procedure of the parameter setting process will be described.

When the parameter setting process is executed, the CPU 11 first executes the clustering coefficient setting process described later in S710, as shown in FIG. 20.

Then, the CPU 11 executes the regularization coefficient setting process described later in S720. Further, the CPU 11 executes the map size setting process described later in S730 to end the parameter setting process.

Next, the procedure of the clustering coefficient setting process executed in S710 will be described.

When the clustering coefficient setting process is executed, the CPU 11 first performs stratified classification by external factors in S810 as shown in FIG. 21. The stratified classification based on external factors is to set the first search range, which will be described later, according to the conditions of the vehicle. As a condition of the vehicle, for example, a vehicle speed can be mentioned.

Further, the CPU 11 defines the range of normal data of the clustering coefficient in S820.

Then, the CPU 11 sets a first search range for searching the clustering coefficient of DBSCAN in S830. DBSCAN is an abbreviation for Density-Based Spatial Clustering of Applications with Noise and is a clustering algorithm. DBSCAN is performed to remove outliers in teacher data. There are two DBSCAN clustering coefficients, the neighborhood search radius ε and the minimum number of neighborhood points minpts. The first search range is set by the neighborhood search radius ε and the minimum number of neighborhood points minpts, respectively.

Next, in S840, the CPU 11 searches for the optimum solution having the neighborhood search radius ε and the minimum number of neighborhood points minpts within the first search range set in S830. Specifically, the CPU 11 uses the learning target data acquired in S10 to train the normal model while changing the neighborhood search radius ε and the minimum neighborhood score minpts within the first search range, and calculates the total abnormality degree. Then, the neighborhood search radius ε and the minimum number of neighborhood points minpts that minimize the total anomaly are determined as the optimum solution. Cross-validation is used in learning the normal model and calculating the degree of abnormality. Cross-validation can suppress overfitting and ensure generalization accuracy. Moreover, the grid search is used in the search of the clustering coefficient.

In the search for the optimum solution in S840, for example, as shown in the graph GR3 of FIG. 22, the CPU 11 first sets a plurality of set values x0, x1, x2, x3, x4 in the first search range SR1. Then, the CPU 11 calculates the total abnormality degree for each of the set values x0, x1, x2, x3, x4, and sets the set value at which the total abnormality degree is the minimum as the optimum solution. In the graph GR3, the set value x2 is the optimum solution.

When the processing of S840 is completed, as shown in FIG. 21, the CPU 11 sets a second search range narrower than the first search range in the vicinity of the optimum solution searched in S840 in S850.

Next, in S860, the CPU 11 searches for the optimum solution having the neighborhood search radius ε and the minimum number of neighborhood points minpts within the second search range set in S850. Specifically, the CPU 11 uses the learning target data acquired in S10 to learn the normal model while changing the neighborhood search radius ε and the minimum neighborhood score minpts within the second search range, and calculates the total abnormality degree. Then, the neighborhood search radius ε and the minimum number of neighborhood points minpts that minimize the total anomaly are determined as the optimum solution.

In the search for the optimum solution in S860, for example, as shown in the graph GR4 of FIG. 22, the CPU 11 first has a plurality of set values (x2-2a), (x2-a), x2 in the second search range SR2. Set (x2 + a) and (x2 + 2a). Then, the CPU 11 calculates the total abnormality degree for each of the set values (x2-2a), (x2-a), x2, (x2 + a), and (x2 + 2a), and sets the setting value that minimizes the total abnormality degree as the optimum solution. do. In the graph GR4, the set value (x2-2a) is the optimum solution.

When the processing of S860 is completed, as shown in FIG. 21, the CPU 11 sets the optimum solution searched in S860 in S870 to the clustering coefficient (that is, the neighborhood search radius ε and the minimum number of neighborhood points minpts), and clusters. The coefficient setting process ends.

When the clustering coefficient setting process is completed, the CPU 11 executes the regularization coefficient setting process in S720 as shown in FIG. 20.

The regularization coefficient setting process is the same as the clustering coefficient setting process except that the optimum solution of the regularization coefficient is searched instead of the clustering coefficient, so detailed explanation is omitted. That is, in the regularization coefficient setting process, the CPU 11 first sets the first search range and searches for the optimum solution of the regularization coefficient within the first search range. Further, the CPU 11 sets a second search range narrower than the first search range in the vicinity of the optimum solution, and searches for the optimum solution of the regularization coefficient within the second search range. Then, the CPU 11 sets the optimum solution searched within the second search range as the regularization coefficient.

The regularization coefficient is the coefficient of the regularization term used in RGCCA to calculate the coefficient of the state observer. RGCCA is an abbreviation for Regularized Generalized Canonical Correlation Analysis.

When the regularization coefficient setting process is completed, the CPU 11 executes the map size setting process in S730 and ends the parameter setting process.

The map size setting process is the same as the clustering coefficient setting process except that it searches for the optimum solution for the horizontal size and vertical size of the normal surface instead of the clustering coefficient, so detailed explanation is omitted. That is, in the map size setting process, the CPU 11 first sets the first search range and searches for the optimum solution of the horizontal size and the vertical size of the normal surface within the first search range. Further, the CPU 11 sets a second search range narrower than the first search range in the vicinity of the optimum solution, and searches for the optimum solution of the horizontal size and the vertical size of the normal surface within the second search range. Then, the CPU 11 sets the optimum solution searched within the second search range to the horizontal size and the vertical size of the normal surface.

When the parameter setting process is completed, the CPU 11 shifts to S40 as shown in FIG.

The anomaly detection device 1 configured in this way has an optimum solution (hereinafter, parameter) within a preset first search range for the clustering coefficient, regularization coefficient, and map size (hereinafter referred to as parameters) used to generate a normal model. Hereinafter, the first optimum solution) is searched. Then, the anomaly detection device 1 searches for the optimum solution of the parameter (hereinafter referred to as the second optimum solution) within the second search range including the searched first optimum solution and set to be narrower than the first search range. ..

As a result, the anomaly detection device 1 roughly searches for the first optimum solution within the first search range, and then finely searches for the second optimum solution within the second search range to determine the optimum solution for the parameter. Therefore, the amount of calculation for searching for the optimum solution can be reduced.

The anomaly detection device 1 searches for the optimum solution in the order of the clustering coefficient, the regularization coefficient, and the map size. The order of clustering coefficient, regularization coefficient, and map size has the greatest effect on the accuracy of the normal model. As a result, the abnormality detection device 1 can suppress a decrease in the accuracy of the normal model.

In the embodiment described above, S830 and S840 correspond to the processing as the first search unit and the first search procedure, and S850 and S860 correspond to the processing as the second search unit and the second search procedure.

Although one embodiment of the present disclosure has been described above, the present disclosure is not limited to the above embodiment, and can be variously modified and implemented.

[Modification 1]
In the above embodiment, the abnormality detection device 1 shows a mode of detecting an abnormality of a vehicle. However, the anomaly detection device 1 may detect anomalies in transportation equipment, farm tool equipment, construction equipment, and the like.

[Modification 2]
In the above embodiment, the mode in which the abnormality detection device 1 is mounted on the vehicle is shown. However, the abnormality detection device 1 does not have to be mounted on the vehicle. For example, the abnormality detection device 1 may be installed outside the vehicle and may be connected to the connected ECU of the vehicle by wire or wirelessly.

[Modification 3]
In the above embodiment, the mode in which the degree of abnormality is calculated by the difference between the monitored data and the neuron is shown. However, as shown in FIG. 23, the degree of abnormality of the negative value may be expressed by adding the magnitude of the neuron of each dimension and the monitored data. This makes it possible to determine whether the value is larger or smaller than the normal value of each dimension.

[Modification 4]
In the above embodiment, a mode is shown in which the normal surface is learned as a normal model by combining the training target data, the first state observed value, and the first latent state value and inputting them into the competitive neural network. However, as shown in FIG. 24, the latent state value may be included in the variables constituting the state observer. In this case, a normal model is generated by combining the training target data and the first state observation value calculated including the latent state value and inputting them into the competitive neural network.

The control device 2 and its method described in the present disclosure are provided by a dedicated computer provided by configuring a processor and memory programmed to perform one or more functions embodied by a computer program. It may be realized. Alternatively, the control device 2 and its method described in the present disclosure may be realized by a dedicated computer provided by configuring a processor with one or more dedicated hardware logic circuits. Alternatively, the control device 2 and its method described in the present disclosure are a combination of a processor and memory programmed to perform one or more functions and a processor configured by one or more hardware logic circuits. It may be realized by one or more dedicated computers configured by. The computer program may also be stored on a computer-readable non-transitional tangible recording medium as an instruction executed by the computer. The method for realizing the functions of each part included in the control device 2 does not necessarily include software, and all the functions may be realized by using one or a plurality of hardware.

A plurality of functions possessed by one component in the above embodiment may be realized by a plurality of components, or one function possessed by one component may be realized by a plurality of components. Further, a plurality of functions possessed by the plurality of components may be realized by one component, or one function realized by the plurality of components may be realized by one component. Further, a part of the configuration of the above embodiment may be omitted. Further, at least a part of the configuration of the above embodiment may be added or replaced with the configuration of the other above embodiment.

In addition to the above-mentioned abnormality detection device 1, a system having the abnormality detection device 1 as a component, a program for operating a computer as the abnormality detection device 1, a non-transitional substantive record of a semiconductor memory or the like in which this program is recorded, etc. The present disclosure can also be realized in various forms such as a medium and an abnormality detection method.

Claims (12)

1. A data acquisition unit (101) configured to acquire learning target data and monitoring target data of an abnormality detection target, which is a target for detecting an abnormality, and a data acquisition unit (101).
   The learning target data, the first state observation value obtained by inputting the learning target data into the state observer representing the state of the system of the abnormality detection target, and the qualitative state representing the latent state of the abnormality detection target. A normal model generator configured to generate a normal model by inputting data composed of a first latent state value obtained by inputting the training target data into a soft sensor into a competitive neural network (a normal model generator). 106) and
   The monitored data, the second state observed value obtained by inputting the monitored data into the state observer, and the second latent state obtained by inputting the monitored data into the qualitative soft sensor. An abnormality degree calculation unit (108) configured to calculate an abnormality degree indicating the degree of abnormality of the abnormality detection target by comparing the data composed of the values with the normal model.
   An abnormality detection device (1) including a determination unit (109) configured to determine whether the abnormality detection target is normal or abnormal based on the abnormality degree calculated by the abnormality degree calculation unit. ).
2. The abnormality detection device according to claim 1.
   The qualitative soft sensor is an abnormality detection device generated by combining a plurality of flags representing the states of the components constituting the abnormality detection target.
3. The abnormality detection device according to claim 1 or 2.
   The normal model generation unit represents a change in the relationship between a plurality of variables constituting the state observer, calculates a first index obtained by inputting the training target data, and calculates the training target data and the first index. The normal model is generated by inputting data composed of one state observed value, the first latent state value, and the first index into the competitive neural network.
   The anomaly degree calculation unit calculates a second index obtained by inputting the monitoring target data by representing the change in the relationship between the plurality of variables constituting the state observer, and the monitoring target data and the monitoring target data are described. An abnormality detection device that calculates the degree of anomaly by comparing data composed of a second state observed value, the second latent state value, and the second index with the normal model.
4. The abnormality detection device according to claim 3.
   The first index and the second index are anomaly detection devices in which the Mahalanobis distance is used.
5. The abnormality detection device according to any one of claims 1 to 4.
   A first search configured such that the normal model generation unit searches for the optimum solution as the first optimal solution within a preset first search range for at least one parameter used to generate the normal model. Part (S830, S840) and
   The optimum solution of the parameter is searched as the second optimum solution within the second search range including the first optimum solution searched by the first search unit and set to be narrower than the first search range. Anomaly detection device including a second search unit (S850, S860) configured in.
6. The abnormality detection device according to claim 5.
   The first search unit and the second search unit are abnormality detection devices that search for the parameters in descending order of influence on the accuracy of the normal model when a plurality of the parameters are present.
7. The data acquisition procedure (S10, S410) for acquiring the learning target data and the monitoring target data of the abnormality detection target, which is the target for detecting the abnormality, and
   The learning target data, the first state observation value obtained by inputting the learning target data into the state observer representing the state of the system of the abnormality detection target, and the qualitative state representing the latent state of the abnormality detection target. A normal model generation procedure (S90) for generating a normal model by inputting data composed of a first latent state value obtained by inputting the training target data into a soft sensor into a competitive neural network, and a normal model generation procedure (S90).
   The monitored data, the second state observed value obtained by inputting the monitored data into the state observer, and the second latent state obtained by inputting the monitored data into the qualitative soft sensor. An abnormality degree calculation procedure (S480) for calculating an abnormality degree indicating the degree of abnormality of the abnormality detection target by comparing the data composed of the values with the normal model.
   An abnormality detection method including a determination procedure (S490) for determining whether the abnormality detection target is normal or abnormal based on the abnormality degree.
8. The abnormality detection method according to claim 7.
   The normal model generation procedure represents a change in the relationship between a plurality of variables constituting the state observer, calculates a first index obtained by inputting the training target data, and calculates the training target data and the first index. The normal model is generated by inputting data composed of one state observed value, the first latent state value, and the first index into the competitive neural network.
   The abnormality degree calculation procedure represents a change in the relationship between a plurality of variables constituting the state observer, calculates a second index obtained by inputting the monitored data, and calculates the monitored data and the first. An abnormality detection method for calculating the degree of anomaly by inputting data composed of a two-state observed value, the second latent state value, and the second index into the competitive neural network.
9. The abnormality detection method according to claim 7 or 8.
   The first search procedure (S830, S840) for searching the optimum solution as the first optimum solution within the preset first search range for at least one parameter used to generate the normal model.
   With the second search procedure (S850, S860) in which the optimum solution of the parameter is searched as the second optimum solution within the second search range including the first optimum solution and set to be narrower than the first search range. Anomaly detection method.
10. Computer,
    A data acquisition unit configured to acquire learning target data and monitoring target data for anomaly detection targets, which are targets for detecting anomalies.
    The learning target data, the first state observation value obtained by inputting the learning target data into the state observer representing the state of the system of the abnormality detection target, and the qualitative state representing the latent state of the abnormality detection target. A normal model generator configured to generate a normal model by inputting data composed of a first latent state value obtained by inputting the training target data into a soft sensor into a competitive neural network.
    The monitored data, the second state observed value obtained by inputting the monitored data into the state observer, and the second latent state obtained by inputting the monitored data into the qualitative soft sensor. An abnormality degree calculation unit configured to calculate an abnormality degree indicating the degree of abnormality of the abnormality detection target by comparing the data composed of the values with the normal model, and an abnormality degree calculation unit.
    An abnormality detection program for functioning as a determination unit configured to determine whether the abnormality detection target is normal or abnormal based on the abnormality degree calculated by the abnormality degree calculation unit.
11. The abnormality detection program according to claim 10.
    The normal model generation unit calculates a first index obtained by inputting the training target data, representing changes in the relationships of a plurality of variables constituting the state observer, and calculates the training target data and the training target data. The normal model is generated by inputting the data composed of the first state observed value, the first latent state value, and the first index into the competitive neural network.
    The anomaly degree calculation unit calculates a second index obtained by inputting the monitoring target data by representing the change in the relationship between the plurality of variables constituting the state observer, and the monitoring target data and the monitoring target data are described. An abnormality detection program that calculates the degree of anomaly by comparing data composed of a second state observed value, the second latent state value, and the second index with the normal model.
12. The abnormality detection program according to claim 10 or 11.
    Computer, more
    A first search configured such that the normal model generation unit searches for the optimum solution as the first optimal solution within a preset first search range for at least one parameter used to generate the normal model. Department and
    The optimum solution of the parameter is searched as the second optimum solution within the second search range including the first optimum solution searched by the first search unit and set to be narrower than the first search range. Anomaly detection program to function as the second search unit configured in.

Images