JP2017117034A - Diagnosis device and diagnostic method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a diagnosis device that can automatically adjust a parameter suitable for a diagnosis in accordance with a condition of acquired time-series data.SOLUTION: According to the present invention, a diagnosis device diagnosing states of objects on the basis of time-series data collected from diagnosis objects comprises: pre-processing means that executes pre-processing for classifying the time-series data on the basis of a setting parameter to generate pre-processing data; classification means that generates a classification result having the pre-processing data classified in accordance with similarity of data; classification result evaluation means that generates an evaluation result having the classification result evaluated in accordance with a temporal change pattern of an amount of characteristic of the classification result; and setting value adjustment means that adjusts the setting parameter to be used in the pre-processing means so that the evaluation result is a desired value.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a diagnostic apparatus and a diagnostic method for diagnosing the state of a diagnostic target using time-series data acquired from the diagnostic target.

  For power plant such as thermal power and nuclear power, and equipment such as communication facilities, a lot of time series data is monitored for stable operation of equipment.

  For example, in a power plant, in order to grasp the state of the plant, measuring instruments for measuring pressure, temperature, flow rate, water level and the like are installed in each part. In the communication facility, a measuring instrument for measuring the amount of data communication is installed. The value of the time series data obtained from the measuring instrument is displayed to the device manager. Also, in most devices, the obtained signal value is stored in a data recording device, which is a dedicated computer, from the viewpoint of abnormality or malfunction countermeasures or maintenance.

  When there is a change in the state of the device, the administrator displays the value of the corresponding time series data on the monitoring screen in order to check whether there is a change in the value of the related time series data. Usually, the device operation control device and the monitoring device are integrated, and the device state can be monitored and controlled in real time.

  Recently, many devices and methods for detecting abnormal signs before an abnormal state occurs in a diagnosis target have been studied. Patent Document 1 discloses a diagnostic apparatus using adaptive resonance theory (ART). Here, ART is a theory for classifying multi-dimensional time series data into categories according to their similarity.

  In the technique of Patent Document 1, first, normal measurement data is classified into a plurality of categories (normal categories) using ART. Next, the current time series data is input to ART and classified into categories. When the time series data cannot be classified into normal categories, a new category (new category) is generated. The occurrence of a new category means that the state of the diagnosis target has changed. Therefore, in the diagnostic device of Patent Document 1, the occurrence of an abnormal sign is determined by the occurrence of a new category.

  When classifying time-series data using a technique such as adaptive resonance theory, it is necessary to set a resolution parameter that determines the resolution of the data according to the contents of diagnosis. If this parameter is not properly selected, there is a possibility that an abnormality cannot be found or an erroneous detection that detects an abnormality although it is in a normal state may occur. Therefore, Patent Document 1 discloses a technique for automatically adjusting resolution parameters suitable for abnormality detection using time-series data in a normal state.

  In addition, as a method for classifying and diagnosing data, not only the method described above but also a number of methods have been proposed.

JP 2011-070334 A

  When classifying time-series data using methods such as adaptive resonance theory, not only resolution parameters but also many parameters other than resolution parameters such as data items used for diagnosis and data normalization range are adjusted. There is a need.

  The acquired time series data may include not only normal data but also abnormal data. Even when the time series data includes data at the time of abnormality, various situations are assumed such as when the date and time when the abnormality occurred is known, or when the date and time when the abnormality occurred cannot be specified. In order to adjust a large number of parameters according to such a situation, a large number of man-hours are required for design, and therefore a technique for automating parameter adjustment is required.

  In view of the above situation, an object of the present invention is to provide a diagnostic apparatus capable of automatically adjusting parameters suitable for diagnosis according to the conditions of acquired time-series data.

  In order to solve the above problems, in the present invention, in the diagnostic device for diagnosing the state of a target based on time series data collected from a diagnostic target, the diagnostic device classifies the time series data based on a setting parameter. Preprocessing means for executing preprocessing for generating preprocessed data, classification means for generating a classification result obtained by classifying the preprocessed data according to data similarity, and characteristics of the classification result A classification result evaluation unit that generates an evaluation result evaluated according to a temporal change pattern of a quantity; and a set value adjustment unit that adjusts a setting parameter used in the preprocessing unit so that the evaluation result becomes a desired value. It is characterized by providing.

  By using the diagnostic apparatus of the present invention, the setting parameters can be automatically adjusted according to the conditions of the acquired time series data, and the design man-hours can be reduced.

It is a block diagram explaining the diagnostic apparatus which is 1st Example of this invention. It is an operation | movement flowchart figure of a diagnostic apparatus. It is a figure explaining the timing which operates learning mode and diagnostic mode. It is a figure explaining the aspect of the data preserve | saved at a signal database. It is a figure explaining the block diagram at the time of using ART as an Example of a classification means. It is a figure explaining the classification result of time series data. It is a figure explaining operation | movement of the learning mode when abnormality occurrence time is known. It is a figure explaining the method of associating a classification result and an event in a post-processing means. It is a block diagram explaining the diagnostic apparatus which is the 2nd Example of this invention. It is a figure explaining the evaluation criteria in a classification result evaluation means. It is a figure explaining operation | movement of an abnormality occurrence date specific part. It is a block diagram explaining the diagnostic apparatus which is the 3rd Example of this invention. It is a figure explaining the Example of a power plant as a diagnostic object provided with a maintenance data recording device. It is a figure explaining the aspect of the maintenance data preserve | saved at a maintenance data recording device and a signal database.

  FIG. 1 is a block diagram for explaining a diagnostic apparatus according to a first embodiment of the present invention. The diagnostic device 300 is connected to an external device 900 and a time-series data recording device 200 that records time-series data 1 collected by the diagnostic object 100 of the diagnostic device.

  The diagnostic apparatus 300 includes a diagnostic unit 340 and an image display information generation unit 350 as arithmetic units. The diagnosis unit 340 includes a pre-processing unit 400, a classification unit 500, a post-processing unit 600, a classification result evaluation unit 700, and a set value adjustment unit 800.

  The diagnostic apparatus 300 includes a signal database 330 as a database. In FIG. 1, the database is abbreviated as DB. The signal database 330 stores computerized information, and the information is stored in a form generally called an electronic file (electronic data).

  Further, the diagnostic apparatus 300 includes an external input interface 310 and an external output interface 320 as interfaces with the outside.

  Then, the time series data 2 recorded in the time series data recording device 200 via the external input interface 310 and the operation of the external input device 910 (keyboard 910 and mouse 920) provided in the external device 900 are created. The external input signal 3 is taken into the diagnostic device 300. Further, the image display information 9 is output to the screen display device 940 via the external output interface 320. The image display device 940 can also display the data recording device information 10.

  In the diagnostic apparatus 300 shown in FIG. 1, the time series data 2 and the external input signal 3 are taken in via the external input interface 210. The measurement signal 3 captured by the diagnostic device 300 is stored in the signal database 330.

  The diagnostic apparatus 300 has two processing modes, a learning mode and a diagnostic mode. The flowchart of the learning mode and the diagnostic mode, and the operations of the diagnostic unit 340 and the image display information generation unit 350 will be described later with reference to FIGS. In the diagnosis unit 340, the pre-processing unit 400, the classification unit 500, the post-processing unit 600, the classification result evaluation unit 700, the set value is received with respect to the input of the signal database information 5 including time series data stored in the signal database 330. The adjusting means 800 is operated. A signal obtained as a result of operating the diagnostic means 340 is transmitted to the signal database 330 as signal database information 6 and stored.

  In the diagnostic apparatus 300 according to the present embodiment, the signal database 330, the diagnostic unit 340, and the image display information generation unit 350 are provided in the diagnostic apparatus 300. However, some of these apparatuses are included in the diagnostic apparatus 300. It may be arranged outside and only data may be communicated between devices.

  Further, the signal database information 50 stored in the signal database 330 installed in the diagnostic apparatus 300 can be displayed on the screen display device 940 via the external output interface 320, and the information is stored in the external input device. It can be corrected by the external input signal 3 generated by operating 910.

  In this embodiment, the external input device 910 is composed of a keyboard 920 and a mouse 930. However, any device for inputting data such as a microphone or a touch panel for voice input may be used.

  In addition, as an embodiment of the present invention, it goes without saying that the present invention can also be implemented as an information providing service that provides information obtained by operating the diagnostic method and the diagnostic apparatus 300.

  FIG. 2 is an operation flowchart diagram of the diagnostic apparatus 300. FIG. 2A is a flowchart of the learning mode, and FIG. 2B is a flowchart of the diagnosis mode.

  First, the learning mode will be described with reference to FIG. In step 1000, the time-series data 2 stored in the time-series data recording device 200 is taken into the diagnostic device 300 and stored in the signal database 330.

  In step 1010, the signal processing unit 400 extracts the signal database information 5 used for learning from the signal database 330, and performs preprocessing necessary for diagnosis of the target. The preprocessing unit 400 includes at least a data item setting unit 410, a normalization range setting unit 420, and a resolution setting unit 430. The data item setting unit 410 sets items of time series data used for diagnosis (for example, sensor information such as temperature and pressure), and the preprocessing unit 400 signals the time series data of the items set by the data item setting unit 410. Extract from database 400. The normalization range setting unit 420 sets the normalization range of the extracted time series data. Based on the set normalization range, the time series data is normalized to a range of 0 to 1, for example. The preprocessing unit 400 normalizes data for each time series item. Data including data Nxi (n) obtained by normalizing time-series data and the complement CNxi (n) (= 1−Nxi (n)) of the normalized data is defined as input data Ii (n). Also, the resolution setting unit 430 sets the resolution when the data is classified by the classification unit 500.

  The preprocessing data 20 obtained by operating the preprocessing unit 400 is transmitted to the classification unit 500.

  The preprocessing unit 400 executes the above processing based on setting parameters such as data items to be extracted, normalization range, and resolution. This setting parameter is adjusted by setting value adjusting means 800 described later. The initial value of the setting parameter can be arbitrarily set by a method such as extracting a data item measured by a measuring instrument close to the position where the abnormality has occurred.

  In step 1020, the classification means 500 is operated to classify the time series data extracted from the signal database 330 by the clustering technique, and the classification result 21 is output. Known clustering techniques include ART510 (Adaptive Resonance Theory), MT520 (Mahalano-Bistactic method), VQC530 (vector quantization), and the like. In the present embodiment, the case where ART 510 is used will be mainly described, but there is no limitation on the method used for classification means 500.

  In step 1030, the classification result evaluation unit 700 is operated to evaluate the classification result 21. The evaluation result 22 of the classification result is transmitted to the set value adjustment unit 800.

  In step 1040, an end determination is performed. If the number of repetitions of steps 1010 to 1050 exceeds a threshold value, or the evaluation result 22 of the classification result exceeds a reference, the process proceeds to step 1060 if a predetermined end determination condition is satisfied, and if not satisfied, the process proceeds to step 1060. Proceed to 1050.

  In step 1050, the setting parameter necessary for operating the pre-processing unit 400 by adjusting the set value adjusting unit 800 is adjusted. The set value adjusting unit 800 is a unit that adjusts parameters set by the data item setting unit 410, the normalization range setting unit 420, and the resolution setting unit 430 so that the evaluation result 22 becomes a desired value. Apply any optimization method such as learning or reinforcement learning.

  In Step 1060, the post-processing means 600 is operated, and Steps 1010 to 1050 are repeated, and the information of the classification result 21 with the best evaluation result 22 is transmitted to the signal database 330 as the signal database information 6 and stored. In addition, not only the result that the evaluation result 22 becomes the best but also arbitrary information obtained in the process of repeating steps 1010 to 1050 may be transmitted to the signal database 330 and stored.

  In step 1070, the image display information generating means 350 converts the signal database information 7 from the signal database 330 into image display information 8 and transmits it to the external output interface 320. In the external output interface, the image display information 9 is transmitted to the image display device 940, and the information is displayed on the screen (monitor).

  Next, the diagnosis mode will be described with reference to FIG. In step 1100, the time-series data 2 stored in the time-series data recording device 200 is taken into the diagnostic device 300 and stored in the signal database 330.

  In step 1110, the preprocessing means 400 extracts the signal database information 5 used for diagnosis from the signal database 330, and performs necessary preprocessing as in the learning. The preprocessing data 20 obtained by operating the preprocessing unit 400 is transmitted to the classification unit 500.

  In step 1120, the classification means 500 is operated to classify the time series data extracted from the signal database 330 by the clustering technique, and the classification result 21 is output.

  In step 1130, the post-processing unit 600 is operated to transmit the information of the classification result 21 to the signal database 330 as the signal database information 6 and store it.

  In step 1140, the image display information generation means 350 converts the signal database information 7 from the signal database 330 into image display information 8 and transmits it to the external output interface 320. In the external output interface, the image display information 9 is transmitted to the image display device 940 to display the information on the screen (monitor).

  FIG. 3 is a diagram illustrating timings at which the learning mode and the diagnosis mode are operated.

  In the method shown in FIG. 3A, after the operation data is accumulated for a certain period, the learning mode is operated once and the diagnosis mode is operated at a certain period.

  In the method shown in FIG. 3B, the diagnosis mode is operated after the learning mode is operated at regular intervals to update the classification result.

  In the method shown in FIG. 3C, the learning mode is operated when an instruction is given from the user. After executing the learning mode at an arbitrary timing to update the classification result, the diagnostic mode is operated.

  In addition to the timing described in the present embodiment, the timing for operating the learning mode and the diagnostic mode can be arbitrarily set.

  FIG. 4 is a diagram for explaining the mode of data stored in the signal database 330.

  As shown in FIG. 4A, in the signal database 330, values of time-series data (in the figure, data items A, B, and C are described) are stored for each sampling period (time on the vertical axis). . By using scroll boxes 302 and 303 that can be moved vertically and horizontally on the display screen 301, a wide range of data can be scroll-displayed.

  As shown in FIGS. 4B and 4C, the signal database 330 stores setting values and classification results in the preprocessing unit 400.

  As shown in FIG. 4D, the signal database 330 stores the number of repetitions in the learning mode, the set value at that time, and the evaluation result.

  FIG. 5 is a diagram for explaining a block diagram when an ART 510 is used as an example of the classification unit 500.

  In ART 510, time-series data as input data is classified into a plurality of categories.

  The ART module 510 includes an F0 layer 511, an F1 layer 512, an F2 layer 513, a memory 514, and a selection subsystem 515, which are coupled to each other. The F1 layer 512 and the F2 layer 513 are connected via a weighting factor. The weighting coefficient represents a prototype (prototype) of a category into which input data is classified. Here, the prototype represents a representative value of the category.

  Next, the algorithm of ART510 will be described.

  The outline of the algorithm when input data is input to the ART 510 is as shown in the following processing 1 to processing 5.

  Process 1: The input vector is normalized by the F0 layer 511, and noise is removed.

  Process 2: A suitable category candidate is selected by comparing the input data input to the F1 layer 512 with a weighting factor.

  Process 3: The validity of the category selected by the selection subsystem 515 is evaluated by the ratio with the parameter ρ. If it is determined to be valid, the input data is classified into the category, and the process proceeds to process 4. On the other hand, if it is not determined to be valid, the category is reset, and an appropriate category candidate is selected from the other categories (repeat process 2). Increasing the value of parameter ρ makes the category classification finer. That is, the category size is reduced. Conversely, if the value of ρ is reduced, the classification becomes coarse. Category size increases. This parameter ρ is referred to as a vigilance parameter. In the resolution setting unit 430 described above, the value of the vigilance parameter is set.

  Process 4: When all the existing categories are reset in Process 2, it is determined that the input data belongs to the new category, and a new weighting factor representing the prototype of the new category is generated.

  Process 5: When input data is classified into category J, weight coefficient WJ (new) corresponding to category J uses past weight coefficient WJ (old) and input data p (or data derived from input data). Is updated by equation (1).

  Here, Kw is a learning rate parameter (0 <Kw <1), and is a value that determines the degree to which the input vector is reflected in the new weighting factor.

  The arithmetic expressions of Equation 1 and Equations 2 to 12 described later are incorporated in the ART 510.

  The feature of the data classification algorithm of the ART 510 resides in the processing 4 described above.

  In process 4, when input data different from the learned pattern is input, a new pattern can be recorded without changing the recorded pattern. Therefore, it is possible to record a new pattern while recording a pattern learned in the past.

  Thus, when the operation data given in advance is given as input data, the ART 510 learns the given pattern. Therefore, when new input data is input to the learned ART 510, it is possible to determine which pattern in the past is close by the above algorithm. If the pattern has never been experienced before, it is classified into a new category.

FIG. 5B is a block diagram showing the configuration of the F0 layer 511. In the F0 layer 511, the input data I _i is normalized again at each time to create a normalized input vector u _i ⁰ that is input to the F1 layer 512 and the selection subsystem 515.

First, W _i ⁰ is calculated from the input data I _i according to Equation 2. Here, a is a constant.

Next, X _i ⁰ obtained by normalizing W _i ⁰ is calculated using Equation 3. Here, || W ⁰ || represents the norm of W ⁰ .

Then, using equation 4, to calculate the V _i ⁰ obtained by removing noise from the X _i ^0. However, θ is a constant for removing noise. Since the minute value becomes 0 by the calculation of Equation 4, noise of the input data is removed.

Finally, a normalized input vector u _i ⁰ is obtained using Equation 5. u _i ⁰ is input to the F1 layer.

FIG. 5C is a block diagram showing the configuration of the F1 layer 512. In the F1 layer 512, u _i ⁰ obtained by Equation 5 is held as a short-term memory, and P _i to be input to the F2 layer 513 is calculated. Formulas for the F2 layer are collectively shown in Formulas 6 to 12. However, a and b are constants, f (·) is a function expressed by Equation 4, and T _j is a fitness calculated by the F2 layer 513.

但し、

However,

  FIG. 6 is a diagram for explaining the classification result of the time series data.

  FIG. 6A is a diagram illustrating an example of a classification result obtained by classifying time-series data into categories.

  FIG. 6A shows two items of time-series data as an example, and is represented by a two-dimensional graph. In addition, the vertical axis and the horizontal axis indicate the time series data of each item normalized.

  The time series data is divided into a plurality of categories 519 (circles shown in FIG. 6A) by the ART module 510. One circle corresponds to one category.

  In this embodiment, measurement signals are classified into four categories. Category number 1 is a group in which item A has a large value and item B has a small value, category number 2 is a group in which both items A and B have small values, category number 3 has a small value in item A, and item B A group with a large value of, category number 4 is a group with a large value for both items A and B.

  FIG. 6B is a diagram for explaining the result of classifying time-series data into categories. The horizontal axis represents time, and the vertical axis represents measurement signals and category numbers.

  As shown in FIG. 6B, the data of the learning period before the start of diagnosis was classified into categories 1 to 3. The first half of data after the start of monitoring is classified into category 2, which is the same category as the model data. In this case, since the data trends are the same, it is determined that the state has not changed. On the other hand, the latter half of data after the start of monitoring is classified into category 4, and is classified into a category different from the data of the learning period. Since the data trends are different, it is determined that the state of the diagnosis target has changed.

  As described above, the diagnosis technique using the clustering technique has a feature of detecting a change in data tendency. FIG. 6 describes the case where the learning data is normal data.

  Hereinafter, the operation in the learning mode under the condition that the time series data includes the data at the time of abnormality will be described with reference to FIG. Under this condition, the learning mode operates to search for a preprocessing parameter for detecting an abnormality.

  The user inputs an abnormality occurrence time using the external input device 910 from the screen displayed on the image display device 940. FIG. 7A shows an example of an input screen for an abnormality occurrence time. Trend information of time series data or the like may be displayed to assist in inputting an abnormality occurrence time.

  When using the ART 510 in the classification means 500, the classification result evaluation means 700 operates so that the evaluation value becomes high when a new category occurs at the time of occurrence of an abnormality. The classification result evaluation unit 700 is implemented by applying an optimization method such as a reinforcement learning method or a genetic algorithm. It should be noted that the classification result evaluation unit 700 of the present invention is not limited to the optimization method.

  A formula for calculating the evaluation value R is shown in Equation 13.

  Here, T1 is an abnormality occurrence time, T2 is a time when a new category is generated and an abnormality is detected, and α is a minute value for preventing zero percent.

  In this embodiment, the case where the classification unit 500 is an ART has been described. However, the classification unit 500 is not limited to an ART.

  In this embodiment, the classification result is evaluated based on the occurrence status of the new category. However, the classification result is based on information obtained by operating the classification means such as the degree of deviation from the normal state and the ratio of the amount of data that is different. There is no limitation on the evaluation method of the classification result as long as it is a data change pattern over time.

  FIG. 7B is a diagram for explaining an embodiment when the classification result is displayed on the image display device 940. It is possible to confirm whether or not a desired classification result has been obtained by displaying the information of FIG. 4B and the abnormality occurrence time input in FIG.

  FIG. 7C is a diagram for explaining an example when the learning mode repetition count stored in the signal database 330 and the evaluation result at that time are displayed on the image display device 940. Transition of evaluation results can be confirmed.

  FIG. 8 is a diagram for explaining a method for associating a classification result with an event in the post-processing means 600.

FIG. 8A is a diagram for explaining an embodiment for associating a classification result with an event.
Number registration is a method of registering an event and a category number in association with each other. For example, category numbers 4, 5, and 6 are registered as event A. Then, next time, when the time series data is classified into any of category numbers 4, 5, and 6, event A is displayed as an abnormal event.

  Frequency registration is a method of registering an event and a category appearance frequency in association with each other. For example, it is registered as event B that the appearance frequencies of certain 60 consecutive category numbers 4, 5, and 6 are in the range of 30 to 40%, 30 to 50%, and 10 to 40%, respectively. Then, when the frequency of appearance of certain 60 consecutive category numbers falls within the above range next time, event B is displayed as an abnormal event.

  Serial number registration is a method of registering an event and a category appearance number in association with each other. For example, the appearance of category numbers in the order of 4 → 5 → 4 → 6 is registered as event C. When the category number appears in the order of 4 → 5 → 4 → 6, event C is displayed as an abnormal event.

  FIG. 8B is a diagram for explaining a second embodiment for associating a classification result with an event.

  Detection group registration is a method in which an event and a group in which an abnormality is detected (diagnostic group in which a new category has occurred) are associated and registered. For example, a case where an abnormality is detected in the monitoring group 3 and the monitoring group 9 is registered as an event F. Next, when an abnormality is detected in monitoring group D and monitoring group E, event F is displayed as an abnormal event.

  FIG. 8C is an example of a screen that displays an abnormal event. In the diagnosis mode, when the pattern is classified into the registered pattern, the estimated factor is displayed on the screen as shown in FIG.

  As described above, in the first embodiment of the present invention, the setting parameter for operating the preprocessing unit 400 can be determined by the setting adjustment unit 800 so that a new category is generated when an abnormality occurs.

  By registering the new category and the abnormal event in association with each other, it is possible to provide information to the user by using the abnormal event registered when the same data tendency becomes the next time as an estimation factor.

  In the present embodiment, the time series data includes data at the time of abnormality, but when the abnormality occurrence date and time cannot be specified, the setting parameter of the preprocessing unit 400 is adjusted using the setting value adjustment unit 800. A diagnostic device for identifying the date and time of occurrence of an abnormality will be described.

  FIG. 9 is a block diagram for explaining a diagnostic apparatus according to a second embodiment of the present invention.

  This embodiment differs from the first embodiment in that the post-processing means 600 is provided with an abnormality occurrence date / time specifying unit 610, and the abnormality occurrence date / time specifying unit 610 operates in step 1060 of the learning mode shown in FIG. Further, the evaluation criteria in the classification result evaluation means 700 are different from those in the first embodiment.

  Hereinafter, the evaluation criteria in the classification result evaluation unit 700 will be described with reference to FIG. 10, and the operation of the abnormality occurrence date and time specifying unit 610 will be described with reference to FIG.

  FIG. 10 is a diagram for explaining an example of evaluation criteria in the classification result evaluation means 700.

  As shown in NO1 of FIG. 10, when the number of categories continues to increase with the passage of time, the evaluation is low. This is because abnormalities cannot be detected for reasons such as selection of data items with large noise and fluctuations and fine resolution.

  As shown in NO2 of FIG. 10, when the number of categories first increases and then maintains a constant number, the evaluation is low. This is because the number of categories increases during learning in a normal state, but an abnormality cannot be detected for reasons such as poor resolution.

  As shown in NO3 of FIG. 10, when the number of categories is increased for the first time, a constant number is maintained, and when it rises again, it is evaluated as high. This is because it can be determined that an abnormality has occurred at the time when the increase in the number of categories increases during learning in the normal state and then the number of categories increases again. In this way, the classification result evaluation unit 700 evaluates the classification result based on the increase pattern of the total number of categories.

  In the present embodiment, the evaluation method has been described focusing on the number of categories. However, any method that evaluates the classification result based on the temporal change pattern of the feature value of the classification result is within the scope of the present embodiment.

  FIG. 11 is a diagram for explaining the operation of the abnormality occurrence date and time identification unit 610.

  In the abnormality occurrence date and time identification unit 610, the timing at which the number of categories increases for the second time is set as the abnormality occurrence date and time. The identified result is displayed on the image display device 940.

  As described above, by using the diagnostic apparatus 300 according to the second embodiment of the present invention, it is possible to determine a setting parameter for preprocessing suitable for abnormality detection when the abnormality occurrence date / time is unknown. In addition to the effect described in the first embodiment, the effect of specifying the date and time of occurrence of an abnormality can be obtained.

  In this example, by associating the diagnosis result with the maintenance data that was performed when an abnormality occurred, when the same abnormal event as in the past occurred, the associated maintenance data is displayed on the screen to create a candidate proposal for how to handle the abnormality. A diagnostic device for guidance will be described.

  FIG. 12 is a block diagram for explaining a diagnostic apparatus according to a third embodiment of the present invention.

  The first and second embodiments differ from the first and second embodiments in that the post-processing unit 600 includes an association unit 620 for classification results and maintenance data, and that data can be transmitted and received between the maintenance data recording device 210 and the diagnostic device 300. In the maintenance data recording device 210, the user inputs the maintenance data 30 using the external input device 910 and stores it. The maintenance data 31 stored in the maintenance data recording device 210 can be taken into the signal database 330 of the diagnostic device 300 via the external input interface 310.

  FIG. 13 is a diagram for explaining an embodiment of a power plant as a diagnosis target including the maintenance data recording device 210.

  As illustrated in FIG. 13A, the diagnosis target is configured by a plant 110 and a control device 120. The control device 120 receives a measurement signal 61 measured by a sensor provided in the plant 110 and transmits an operation signal 62 for operating an operation end provided in the plant 110.

  The values of the measurement signal 61 and the operation signal 62 are transmitted to the time series data recording device 200 as time series data 1.

  FIG. 13B is a diagram illustrating a device configuration of a C / C plant that is an example of the plant 110. The gas turbine 2080 includes a compressor 2010, an expander 2020, and a combustor 2030. In the gas turbine 2080, the compressor 2010 takes in and compresses air, then the combustor 2030 takes in the compressed air and fuel to generate combustion gas, and the expander 2020 takes in the combustion gas and obtains power. The output of the gas turbine 2080 is the difference between the power output from the expander 2020 and the power used by the compressor 2010. The exhaust heat recovery boiler 2050 is provided with a heat exchanger 2060 and generates high temperature steam using high temperature exhaust gas from the gas turbine 2080. The steam turbine 2070 takes in the high-temperature steam generated by the exhaust heat recovery boiler 2050 and obtains power. In the condenser 2090, the steam from the steam turbine 2070 is taken in and exchanged with the cooling water to condense the steam into water. The generator 2040 generates power using the outputs of the gas turbine 2080 and the steam turbine 2070.

  FIG. 14 is a diagram for explaining a mode of maintenance data stored in the maintenance data recording device 210 and the signal database 330.

  Further, as shown in FIG. 14 (a), maintenance data recording device 210 and signal database 330 have maintenance such as an abnormal event that has occurred, maintenance contents, cost required for maintenance, the number of days required for maintenance, and the amount of lost opportunity due to maintenance. Information is saved.

  When the event associated with the method described in the first embodiment and the abnormal event of the maintenance data are the same, the classification result and the maintenance data associating unit 620 according to the present invention associates them. That is, the classification result, event, and maintenance data are associated with each other. As a result, the associated event and maintenance data can be displayed on the screen display device 940, as shown in FIG. Based on this display result, the plant operator can acquire information for determining what kind of maintenance is to be performed.

  The present invention can be applied to abnormality detection of various plants such as thermal power plants and nuclear power plants, and abnormality detection of communication devices as a diagnostic device for diagnosing the state of a diagnosis target based on time-series data. Further, it can be applied to various uses such as processing time series data such as an electrocardiogram to diagnose a health condition.

1, 2 Time series data 3 External input signal 4 Input data 5, 6, 7, 50 Signal DB information 8, 9 Image display information 10 Data recording device information 20 Preprocessing data 21, 25 Classification result 22 Evaluation result 23, 24 Classification Setting value 100 Diagnosis target 200 Time series data recording device 300 Diagnosis device 310 Data input interface 320 Data output interface 330 Signal database 340 Diagnosis means 400 Preprocessing means 410 Data item setting section 420 Normalization range setting section 430 Resolution setting section 500 Classification means 510 ART
520 MT
530 VQC
600 Post-processing means 700 Classification result evaluation means 800 Set value adjustment means 900 External device 910 External input device 920 Keyboard 930 Mouse 940 Image display device

Claims (23)

In the diagnostic device for diagnosing the state of the object based on the time series data collected from the diagnosis object,
The diagnostic device performs preprocessing for classifying the time-series data based on setting parameters, and generates preprocessing data, and preprocessing means,
Classification means for generating a classification result obtained by classifying the preprocessed data according to data similarity;
A classification result evaluation means for generating an evaluation result obtained by evaluating the classification result according to a temporal change pattern of the characteristic amount of the classification result;
And a setting value adjusting means for adjusting a setting parameter used in the preprocessing means so that the evaluation result becomes a desired value. The diagnostic device according to claim 1,
The setting parameter includes at least one of a data item parameter of time series data used for diagnosis, a normalization range parameter of the data item of time series data, and a parameter for determining data resolution. . The diagnostic device according to claim 1 comprises:
Post-processing means for outputting the classification result as signal database information;
A diagnostic apparatus, further comprising image display information generation means for generating image display information based on the signal database information. The diagnostic device according to claim 3,
The post-processing means performs a process of associating an event registered by the user of the diagnostic apparatus with a feature value of the classification result,
A diagnostic device characterized in that when a registered classification result is obtained, the registered event is displayed on the image display device. The diagnostic device according to claim 1,
The diagnostic apparatus according to claim 1, wherein the classification result evaluation unit makes a high evaluation when a change in the characteristic amount of the classification result with time is a desired characteristic at an abnormality occurrence time input by a user of the diagnostic apparatus. The diagnostic device according to claim 1,
The preprocessing means extracts a data item measured by a measuring instrument close to a position where an abnormality has occurred as an initial value of a parameter of a data item of time-series data used for diagnosis. The diagnostic device according to claim 3,
The post-processing means includes an abnormality occurrence date and time identifying unit that identifies an abnormality occurrence time. The diagnostic device according to claim 3,
The post-processing means includes a classification result and maintenance data associating unit. The diagnostic apparatus according to any one of claims 1 to 8,
The set value adjusting means is implemented using a genetic algorithm or a reinforcement learning method. The diagnostic device according to claim 7,
The abnormality occurrence date and time identification unit identifies the time when the feature amount starts to change for the second time as the abnormality occurrence time. The diagnostic device according to claim 3,
A diagnostic apparatus characterized in that an abnormal event associated with a classification result having the same tendency during diagnosis is displayed on an image display apparatus. The diagnostic apparatus according to claim 8, wherein
A diagnostic apparatus characterized in that an abnormal event and maintenance data associated with classification results having the same tendency during diagnosis are displayed on an image display device. The diagnostic device according to claim 1,
The classification means generates a classification result obtained by classifying the preprocessed data into categories according to data similarity,
The classification result evaluation unit generates an evaluation result obtained by evaluating the classification result as a temporal change pattern of the feature amount according to an increase pattern of a category number. In a diagnostic method for diagnosing the state of an object based on time-series data collected from the diagnosis object,
The diagnostic method includes a preprocessing step of generating preprocessing data by performing preprocessing for classifying the time-series data based on setting parameters;
A classification step for generating a classification result obtained by classifying the preprocessed data according to data similarity;
A classification result evaluation step for generating an evaluation result obtained by evaluating the classification result according to a temporal change pattern of the characteristic amount of the classification result;
And a setting value adjusting step for adjusting setting parameters used in the preprocessing step so that the evaluation result becomes a desired value. The diagnostic method according to claim 14, wherein
The setting parameter includes at least one of a data item parameter of time series data used for diagnosis, a normalization range parameter of the data item of time series data, and a parameter for determining data resolution. . The diagnostic method according to claim 14 comprises:
A post-processing step of outputting the classification result as signal database information;
An image display information generation step of generating image display information based on the signal database information. The diagnostic method according to claim 16, wherein
The post-processing step performs a process of associating an event registered by a user of the diagnostic method with a feature value of a classification result,
A diagnostic method characterized by displaying a registered event on an image display device when a registered classification result is obtained. The diagnostic method according to claim 14, wherein
The diagnostic method characterized in that the classification result evaluation step makes a high evaluation when the change over time in the characteristic amount of the classification result is a desired characteristic at the abnormality occurrence time inputted by the user of the diagnostic method. The diagnostic method according to claim 14, wherein
In the preprocessing step, a data item measured by a measuring instrument close to a position where an abnormality has occurred is extracted as an initial value of a parameter of a data item of time-series data used for diagnosis. The diagnostic method according to claim 16, wherein
The post-processing step includes an abnormality occurrence date and time identifying step for identifying an abnormality occurrence time. The diagnostic method according to claim 16, wherein
The post-processing step includes a step of associating a classification result with maintenance data. The diagnostic method according to any one of claims 14 to 21,
A genetic algorithm or reinforcement learning method is used for the set value adjustment step. The diagnostic method according to claim 20,
In the abnormality occurrence date and time specifying step, the time when the feature quantity starts to change for the second time is specified as the abnormality occurrence time.

Images