JP6453038B2 - Failure cause classification device - Google Patents

Description

  The present invention relates to an apparatus and method for analyzing time series data obtained from a sensor or the like.

  For commercial vehicles that are responsible for logistics, the occupancy rate affects the profit and loss of the business, because breakdowns are directly linked to business suspension. Therefore, we are trying to improve the maintenance efficiency by classifying the cause of failure from the operation data at the time of vehicle failure. Moreover, since the failure cause classification of the operation data manually can be performed only by an expert, it is difficult to classify in a short time so that it can be used for maintenance work.

Therefore, the cause of failure is automatically classified by the machine. Conventionally, the cause of failure has been classified by comparing the operation data at the time of machine failure with the operation data of past cases where the cause of the failure is known. For example, in Patent Document 1, the appearance time of a waveform feature is learned and classified for each cause of failure.

JP-A-8-221113

  The cause of failure is classified by comparing operation data at the time of machine failure and operation data of past cases. The problem with such a method is that the amount of operation data of past cases affects the classification performance of failure causes. This is because operation data similar to past cases needs to exist in order to appropriately classify operation data when a machine failure occurs. For this reason, it has conventionally been necessary to prepare sufficient operational data from past cases.

However, it is difficult to prepare a sufficient amount of operation data when a machine failure occurs. This is because the operation data at the time of the failure changes depending on various factors such as the progress of the failure, the load status, the external environment, and operation information. For this reason, it is necessary to represent the operation data at the time of occurrence of many failures with the operation data at the time of occurrence of a small number of failures. Patent Document 1 does not consider this point.

The failure cause classification device according to the present invention includes an input unit that inputs operation data indicating a state of the device, a storage unit that stores operation data of the state of the device, and the operation data input in the input unit in the storage unit. A classification unit that classifies each operation state using stored operation data, a region learning unit that learns a distribution region of machine operation data, and a transition probability calculation unit that learns a transition of the distribution region for each cause of failure A transition probability learning unit including: a region conversion unit that converts operation data related to a machine failure into region data based on the learned distribution region; and a transition of region data learned for each cause of failure. which has a probability calculation section that calculates a classification probability for each failure cause in accordance with the probability matrix, for each failure cause of operational data using transition distribution region of the fault for each cause and the distribution area The class probability is calculated, the learning operation part displays a behavior feature of each fault cause analyst, the analyst to compare the behavior of the operation data and the behavior characteristics of each of the failure cause, performs an operation of said distribution region Failure cause classification apparatus characterized by the above.

According to the present invention, it is possible to appropriately classify the cause of failure even when the waveform similar to the operation data at the time of the occurrence of the machine failure does not exist in the past case.

It is the figure which showed the system of a present Example. It is the figure which showed the structure of vehicle control. It is the figure which showed the example of the failure which generate | occur | produces in a vehicle. It is the figure which showed the structure of the cause classification server. It is the figure which showed the hierarchy of the example memory | storage part. It is the figure which showed the format of the sample in a case memory | storage part. It is the figure which showed the outline | summary of the failure cause classification | category method. It is the figure which showed the flow of the transition probability learning part. It is the figure which showed the structure of the transition probability learning part. It is the figure which showed the transition probability matrix. It is the figure which showed the structure of the classification | category part. It is the figure which showed the structure of the learning operation part. It is the figure which showed the example of the learning operation part. It is the figure which showed the structure of the classifier memory | storage part. It is the figure which showed the structure of the classification result display part.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a diagram for explaining the system of this embodiment. The system of this embodiment includes a vehicle 101, a communication base 102, a cause classification server 103, and a maintenance base 105, and the analyst 104 operates the cause classification server 103.

  A large number of sensors are built in the vehicle 101. Sensor data detected by the sensor at the time of failure is received by the cause classification server 103 from the vehicle 101 via the communication base 102.

  The cause classification server 103 performs cause classification of the failure of the received sensor data and transmits the classification result to the maintenance base 105. Setting for the cause classification is performed in advance by the analyst 104. The maintenance base 105 performs efficient maintenance work using the failure cause classification information transmitted.

Interfaces operated by the vehicle 101, the communication base 102, the cause classification server 103, the maintenance base 105, and the analyst 104 can communicate with each other through a standardized high-speed network. The high-speed network can include at least one of wired and wireless.

FIG. 2 shows a vehicle control system mounted in the vehicle 101. A sensor 1 (target value) 201 represents a target value and is a target value suitable for vehicle control given to the vehicle 101. This sensor 1 is automatically determined according to the state of the vehicle. A sensor 2 (observed value) 204 represents an observed value acquired from the actual vehicle 101, and acquires information necessary for control from the vehicle system 203. For example, sensor 1 is a target value for engine speed, and sensor 2 is an observed value for engine speed. The controller 202 determines the input to the vehicle system 203 so that the difference between the sensor 1 (target value) 201 and the sensor 2 (observed value) 204 becomes zero. The vehicle system 203 controls the vehicle according to this input. For example, the controller 202 adjusts the fuel injection amount so that the difference between the target value of the engine speed represented by the sensor 1 and the observed value represented by the sensor 2 becomes zero. Then, according to the adjusted injection amount, the vehicle system 203 burns fuel and rotates the engine.

  FIG. 3 shows an example of a failure that occurs in the vehicle. Failure example 301 shows the waveform of sample 1 in failure cause A, and failure example 302 shows the waveform of sample 2 in failure cause A. When failure cause A occurs with respect to sensor 1 (301a and 302a) that is the target value, sensor 2 (301b and 302b) that is the observed value does not follow sensor 1 that is the target value as time passes. And the difference between the values indicated by the sensor 2 is large. For example, this applies when a large load is applied to the engine. Although not shown in FIG. 3, for example, sample 1 of failure example 301 and sample 2 of failure example 302 differ in the speed at which the difference between the values indicated by sensor 1 and sensor 2 increases. The difference between the samples is due to the fact that the larger load is applied in a short time because the sample 1 has a more rapid change.

  Further, failure example 303 shows the waveform of sample 1 in failure cause B, and failure example 304 shows the waveform of sample 2 in failure cause B. The sensor 1 (303a and 304a) that is the target value has a feature that if the cause of failure B occurs, the follow-up delay of the sensor 2 (303b and 304b) that is the observed value increases. For example, this is applicable when a controller failure occurs. The delay time is different between sample 1 of failure example 303 and sample 2 of failure example 304. Since sample 1 of failure example 303 has a larger delay time, this corresponds to a case where the failure of the controller has further progressed.

  The sample of failure cause A is an example in which the waveform of the sample changes depending on operation information, the external environment, and the load status. The failure cause B sample is an example in which the waveform of the sample changes depending on the progress of the failure. Depending on the cause of the failure, the waveform of the sample changes depending on one or a combination of the operation information, the external environment, the load status, and the progress of the failure.

In this embodiment, even if only a sample that changes rapidly like the failure example 301 is held as a past case, a sample that changes gradually like the failure example 302 is appropriately classified as the same cause of failure as the failure example 301. It is characterized by doing.

FIG. 4 shows the configuration of the cause classification server. The cause classification server includes a case data storage unit 401, a transition probability learning unit 402, operation data 403, a first classification unit 404, and a classification result display unit 405. The case storage unit 401 is stored in a state in which past failure causes and operation data are linked. The transition probability learning unit 402 expresses a time-series change of operation data for each cause of failure as a probability, so that even in the case where only a sample that changes suddenly as in the failure example 301 is in the case storage unit, as in the failure example 302 Learning is performed to appropriately classify slowly changing samples as failure cause A. The operation data 403 is data actually obtained from the vehicle. The first classification unit 404 classifies the operation data 403 based on the data of the transition probability learning unit 402. The classification result display unit 405 displays the result classified in the first classification unit 404.

FIG. 5 shows an example of a hierarchical structure in which past cause of failure and operation data stored in the case storage unit 401 in this embodiment are linked. In the vehicle type hierarchy 501, vehicle types M1, M2, M3... Are stored. The failure type hierarchy 502 stores failure types F1, F2,. The failure cause hierarchy 503 stores failure causes related to the failure types in the failure type hierarchy 502. For example, the cause of failure A, B,... X associated with the failure type F1 of the vehicle type M1 is stored. More specifically, when engine stop is stored as the failure type hierarchy F1, the failure cause hierarchy 503 stores failure causes such as cracks, clogging, and sensor failures. The sample hierarchy 504 stores samples associated with the vehicle type, the failure type, and the cause of the failure.

FIG. 6 is a diagram illustrating a format of a sample stored in the sample hierarchy 504 stored in the case storage unit 401. Failure cause A 601 and failure cause B 602 are the same as the failure causes in the failure cause hierarchy 503. Samples 603 and 604 associated with the failure cause A 601 are stored in a lower hierarchy of the failure cause A. A data format 605 indicates a sample format. In the sample 1, as in the data format 605, sensor data at the time of failure exemplified by the accelerator opening, the engine speed, and the temperature is stored as time series data in synchronization with the time. As shown in FIG. 6, the failure cause A is associated with a plurality of samples such as samples 1, 2,.

  FIG. 7 shows an outline of the failure cause classification method of this embodiment.

  The failure distribution example 701 is a scatter diagram of the failure examples 301 and 302 of the failure cause A. The horizontal axis represents sensor 1, and the vertical axis represents sensor 2. Each point is data at each time. An ellipse 701a represents a normal state. In a normal state, the observed value normally follows the target value of the sensor 1, so that the sensor 1 and the sensor 2 have a correlation. Therefore, if there is data in the area indicated by the ellipse 701a of the failure distribution example 701, it can be viewed as a normal state. An arrow 701b represents the behavior of the sensor value over time of the failure cause A. The failure cause A has a feature of failure in which the sensor 2 as an observation value does not follow the sensor 1 as a target value. Therefore, as indicated by an arrow 701b, even if the value of the sensor 1 is high, the sensor 2 is low, and when the value of the sensor 1 is low, the behavior returns to an ellipse 701a indicating the value of the sensor in the normal state.

  The failure distribution example 702 is a scatter diagram of the failure examples 303 and 304 of the failure cause B. The horizontal axis represents sensor 1, and the vertical axis represents sensor 2. Each point is data at each time. An ellipse 702a represents a normal state. Similar to the failure distribution example 701, the observed value normally follows the target value of the sensor 1 in the normal state, so that the sensor 1 and the sensor 2 have a correlation. Therefore, when there is data in the area indicated by the ellipse 702a of the failure distribution example 702, it can be viewed as a normal state. An arrow 702b represents the behavior of the failure cause B. The failure cause B has a feature of failure in which the delay in following the sensor 2 with respect to the sensor 1 becomes large. Therefore, as indicated by an arrow 702b, the value of sensor 2 does not increase when the value of sensor 1 increases, and the value of sensor 2 increases when the value of sensor 1 maintains a large value. .

  As shown in failure distribution examples 701 and 702, failure examples 301 and 302, and failure examples 303 and 304 show the same scatter diagram and behavior on the scatter diagram. This is because, when expressed as a distribution in a scatter diagram, information on time was lost, so it became impossible to identify the speed and delay time that could not be followed when following the behavior of each failure example on the scatter diagram. It is. In this way, in this embodiment, each failure example is expressed as the behavior of data on the distribution, so that failure examples having the same failure characteristics but different sample waveforms are appropriately classified.

  The failure cause A classification feature 703 and the failure cause B classification feature 704 are detected by the cause classification server 103 to classify the failure cause A and the failure cause B by expressing each failure example as data behavior on the distribution. The characteristics to be shown.

  The failure cause A classification feature 703 indicates a feature detected by the cause classification server 103 to classify the failure cause A. A broken line 703a indicates a distribution boundary for classifying the failure cause A. Here, the distribution boundary 703a is determined as one or more threshold values for each sensor. In the failure cause A classification feature 703, two threshold values are set for the sensor 1 and the sensor 2, respectively. This threshold is set by the analyst 104.

  The distribution area 703b indicates an area delimited by a broken line 703a. This is automatically determined by the cause classification server according to the threshold value set for each sensor. Here, since two threshold values are set for each of the sensors 1 and 2, nine distribution regions 703b (1) to (9) are generated. The cause classification server probabilistically expresses in what order each sample moves in the distribution area 703b, and classifies the failure cause with the most similar feature as the feature of each failure cause.

  The behavior in the normal state is that the sensor 2 follows the sensor 1 and moves with correlation. Therefore, in the distribution region 703b, the three regions (3), (5), and (7) are moved relative to each other.

  The behavior of the failure cause A has a feature that the sensor 2 does not follow the sensor 1. In this case, the distribution region 703b is moved in the order of (3), (6), (9), (8), and (7). The cause classification server 103 classifies the sample that moves in the same distribution area as the failure cause A, with such movement of the distribution area 703b as the feature of the failure cause A.

  The failure cause B classification feature 704 indicates a feature detected by the cause server 103 to classify the failure cause B. A broken line 704a indicates a distribution boundary for classifying the failure cause B. Similarly to the failure cause A classification feature 703, a broken line 704a and a distribution region 704b that determine the distribution boundary are also determined by the analyst for the scatter diagram of the failure cause B.

  The behavior of the failure cause B has a feature that the delay of the follow-up of the sensor 2 with respect to the sensor 1 becomes large. In this case, the distribution area 704b is moved in the order of (7), (8), (9), (6), and (3). The cause classification server 103 classifies a sample that moves in the same distribution region as the failure cause B, with such movement of the distribution region 704b as a feature of the failure cause B.

  FIG. 7 shows an example for two sensors. However, the number of sensors is not limited as long as it is one or more. In the case of one sensor, the movement of the sensor relative to the threshold value is a feature of each failure cause. When there are three or more sensors, the movement of the distribution area on the multidimensional superspace generated by the threshold value of each sensor is a feature of each failure cause.

  In FIG. 7, the distribution area is a rectangular area. However, it is not actually necessary to be a rectangular area, and various forms of distribution areas are conceivable. For example, a method is considered in which each cluster obtained by the k-means method is used as a distribution region. In this case, the k-means method is applied to samples belonging to the same failure type stored in the case storage unit 401. As a result, the data is classified by k clusters. This cluster is classified into the i-th cluster from the area where each data exists, and therefore corresponds to the distribution areas 703b and 704b. Further, a method in which each normal distribution by mixed normal distribution approximation is used as a distribution region is conceivable. In this case, the data is expressed as a collection of k normal distributions. Since the k normal distributions can be used to classify the i-th normal distribution from the area where each data exists, similarly to the k-means method, the k normal distributions correspond to the distribution areas 703b and 704b.

When the k-means method is used, the distribution region takes a super-circular shape, and when the mixed normal distribution approximation is used, the distribution region takes a super-ellipse shape.

FIG. 8 shows a flow of the transition probability learning unit 402. FIG. 9 shows the configuration of the transition probability learning unit 402.
The transition probability learning unit 402 includes a region learning unit 901, a transition probability calculation unit 902, a second classification unit 903, a learning operation unit 904, and a classifier storage unit 905. The transition probability learning unit 402 generates a classifier using information stored in the case storage unit 401. By this classifier, the first classification unit 404 classifies the cause of failure in the operation data 403.
Hereinafter, the transition probability learning unit flow illustrated in FIG. 8 will be described using the configuration of the transition probability learning unit 402 illustrated in FIG. 9.
In step 801, the flow is started. This is executed when learning is selected by the analyst 104.
In step 802, the vehicle type and failure type to be learned are input to the learning operation unit 904 by the analyst.
In step 803, the area learning unit 901 reads all the samples corresponding to the vehicle type and the failure type input from the case storage unit 401 to the learning operation unit 904 and the cause of the failure.
In step 804, the region learning unit 901 learns a distribution region for the sample read from the case storage unit 401. In the embodiment, a rectangular area based on a threshold for each sensor and a supercircle area based on the k-means method are described. In order to learn the distribution region, it is necessary to determine a threshold value for each sensor in the case of a rectangular region, and the number k of clusters in the k-means method in the case of a supercircle region.
This is input to the learning operation unit 904 by the analyst. Also, an initial value may be set in advance. After each sensor threshold value and the number of clusters k are determined, the region learning unit learns the distribution region.

  In the case of a rectangular area, the distribution area is determined by the threshold value of each sensor. When the number of sensors is n and the number of thresholds for the i-th sensor is mi, the number of rectangular regions to be generated s is given by the following equation.

A rectangular area is generated by determining an area number from 1 to s for the area divided by each sensor.

In the case of a supercircle region, k distribution regions are determined by the k-means method. In the k-means method, first, cluster numbers 1 to k are randomly determined for each data. Then k cluster centers are calculated. The calculation of the center of the i-th cluster is obtained by the average value of the data of the i-th cluster number. Next, each data is compared with k cluster centers, and the cluster number of the cluster center closest to each data is assigned. The calculation of the cluster center and the allocation of the cluster number are performed alternately until the cluster number of each data does not change.
The hypercircle region is determined by the obtained cluster center.
In step 805, the transition probability learning unit 902 learns a transition probability matrix for each failure cause according to the region obtained by the region learning unit 901.
FIG. 10 shows the transition probability matrix when there are three distribution regions. The transition probability matrix is a probability representation of the behavior of the distribution region for each cause of failure. The vertical axis represents the distribution area to which the data at time j belongs. The horizontal axis is the distribution region to which the data at time j + 1 belongs. The bottom row of the transition probability matrix shows that the data in the distribution region (3) at time j has a probability of 0.2 in the distribution region (1) at time j + 1 and the probability of being in the distribution region (2) is 0.3. The probability of being in the distribution region (3) is 0.5.

  Describe learning of transition probability matrix.

  First, a distribution area to which each data of each sample of each cause of failure belongs is obtained. In the case of a rectangular area, the area number assigned to the rectangular area to which the data belongs becomes the distribution area. In the case of a supercircle region, the cluster number of the cluster center closest to the data is the distribution region.

  Next, the number of movements of the distribution area of each data is counted and used as a transition matrix. For example, when a sample has l pieces of data, the movement of the distribution region occurs l−1 times. When there are p movements from the distribution region (2) to (3), the second row and the third column of the transition matrix are p.

Finally, the transition matrix is normalized so that the sum of each row of the transition matrix becomes 1 to obtain a transition probability matrix.
In step 806, the second classification unit 903 classifies the cause of failure of each sample read from the case storage unit using the transition probability matrix obtained by the transition probability calculation unit 902.
The configuration of the second classification unit is shown in FIG. The second classification unit includes operation data 1100, region conversion unit 1101, distribution region 1102, probability calculation unit 1103, transition probability matrix 1104, and classification result 1105.
The operation data 1100 indicates operation data for classifying the cause of failure. In step 806, the sample belongs to the vehicle type and failure type to be learned.
The area conversion unit 1101 converts the operation data 1100 into area data using the distribution area 1102. The conversion to the area data is the same as the calculation for obtaining the distribution area to which each data belongs performed by the transition probability calculation unit. Therefore, in the case of a rectangular area, the area number assigned to the rectangular area to which the data belongs becomes the distribution area. In the case of a supercircle region, the cluster number of the cluster center closest to the data is the distribution region. As a result, operation data such as the data format 605 is converted into a distribution area to which each time belongs, and is set as area data.
The probability calculation unit 1103 calculates the classification probability for each failure cause in the operation data 1100 using the transition probability matrix 1104. Here, the transition probability matrix 1104 is a transition probability matrix obtained by the transition probability learning unit 902.
First, the probability calculation unit calculates the classification probability for each failure cause. The classification probability for failure cause X is qX. The initial value of the classification probability of each failure cause is 1. The classification probability is calculated by integrating qX with a probability transition matrix based on the movement of the distribution region at each time.

  Here, description will be made using an example in which the area data obtained by the area conversion unit 1101 is (1), (1), (2), (3),. The transition probability matrix of the failure cause A is shown in FIG. The area data starts from the distribution area (1) to the distribution area (1). From FIG. 10, in the failure cause A, the probability that the data in the distribution region (1) is also in the distribution region (1) at the next time is 0.3. At this time, qA is 1 * 0.3 = 0.3. At the next time, the distribution area (1) moves to the distribution area (2). At this time, qA becomes 1 * 0.3 * 0.7 = 0.21. In this way, qA = 1 * 0.3 * 0.7 * 0.2 * is calculated. QA, qB,..., QX are obtained for each cause of failure.

  Finally, normalization is performed so that the sum of the classification probabilities for each cause of failure becomes 1. For example, the classification probability qX to failure cause X is expressed by the following equation.

Here, the denominator on the right side represents the sum of the classification probabilities of all failure causes A, B,..., C,.
The classification result 1105 is the failure cause with the highest probability among the classification probabilities for each failure cause in the operation data. This is automatically calculated by the second classification unit 903.
In step 807, the cause classification server 103 displays the sample read from the case storage unit 401, the distribution region obtained by the region learning unit, and the classification result obtained by the second classification unit on the learning operation unit 904.
In step 808, the analyst uses the learning operation unit 904 to confirm the classification result. Here, confirmation of the classification result is performed by confirming the correct answer rate. Here, the correct answer rate is an index indicating how many of the learning target samples are classified as a correct cause of failure, and is a value obtained by dividing the number of correctly answered samples by the number of learning target samples.
In step 809, the end of learning is determined by the analyst. For example, the end of learning is determined when a predetermined correct answer rate is achieved. If the end of learning is determined, the process proceeds to step 812. If the learning is not terminated, the process proceeds to step 810.
In step 810, the analyst operates the distribution region using the learning operation unit.
In step 811, the analyst decides to re-execute learning, and proceeds to step 804.
FIG. 12 shows the configuration of the learning operation unit 904.
The learning operation unit 904 includes a learning target display unit 1201, a distribution display unit 1202, a region display unit 1203, a time-series data display unit 1204, a classification result display unit 1205, a representative example display unit 1206, a region operation unit 1207, and a learning confirmation unit 1208. It is constituted by.
The learning target display unit 1201 displays the learning target vehicle type and failure type. In step 802, the learning object display unit 1201 is operated by the analyst, and the vehicle type and failure type to be learned are determined.
The distribution display unit 1202 displays a scatter diagram of data to be learned.
The area display unit 1203 displays the distribution area obtained as a result of learning.
The time series data display unit 1204 displays learning target data in time series order.
The classification result display unit 1205 displays the classification result 1105 obtained by the second classification unit 903.
A representative example display unit 1206 displays a representative example of each cause of failure. Here, the representative example of each failure cause is a sample having the highest classification probability to the corresponding failure cause. For example, a representative example of the failure cause A is a sample having the highest classification probability to the failure cause A among the learned failure cause A samples. The representative example may be determined by an analyst.
In the region operation unit 1207, the analyst performs an operation on the region learning unit. The operation for the area learning unit is a change of a threshold value for each sensor in the case of a rectangular area, and a change in the number of clusters and the cluster center in the case of a supercircle area.
The learning determination unit 1208 is selected when the end of learning is determined by the analyst. When the learning confirmation unit 1208 is selected by the analyst, the process proceeds from step 809 to step 812.

FIG. 13 shows an example of the learning operation unit.
The learning target display unit 1301 displays the learning target vehicle type and failure type. Further, as shown in FIG. 13, the total number of samples corresponding to the selected vehicle type, failure type, and the number of samples for each cause of failure may be displayed.
The distribution display unit 1302 displays a scatter diagram of the learning target sample. The samples to be displayed here can be selected from all learning targets, only a certain cause of failure, and only certain samples. In addition, a specific sample and all other samples can be displayed in different colors, and a specific sample and all samples belonging to a specific cause of failure can be displayed in different colors. It is also possible to display all samples of a specific failure cause and different samples of other failure causes in different colors. For example, the analyst can identify the behavior on the distribution causing the wrong answer by comparing the sample of the wrong cause X with the distribution of the entire cause X.
The distribution display unit 1302 can display a distribution area. As described above, the distribution display unit 1202 and the region display unit 1203 may be on the same graph. The region display unit displays the distribution region obtained by the region learning unit. FIG. 13 shows an example of a rectangular area. In the case of a supercircle region, the region is circular. In addition, when mixed normal distribution approximation is used for region learning, the region is elliptical.
In this way, by displaying the scatter diagram of each data and the distribution area at the same time, the analysis conditions of the analyst are supported.
For example, the analyst analyzes the reason why the cause of failure is misclassified by displaying the scatter chart with different colors and other samples in a different scatter diagram and displaying the distribution area in an overlapping manner. Factors of misclassification include the size, number, and position of distribution areas. Therefore, the analyst adjusts the area by the area operation unit 1207 with reference to these pieces of information. In addition, the area displayed on the area display unit 1202 can be directly operated.
For example, in the case of a rectangular area, the threshold value for each sensor is manipulated. If there is a broken line with a threshold between the failure behavior characteristic of the misclassified sample and the failure behavior characteristic of the same failure cause sample that has been correctly answered, a misclassification occurs. This is because the distribution region based on the sensor threshold is too finely divided. Therefore, the analyst manipulates the sensor threshold so that the misclassified sample and the correct sample of the same cause of failure move in the same distribution region. This is performed by changing or deleting the threshold value. In the case of a supercircle region, the distribution region is manipulated by moving the cluster center or increasing / decreasing the number of clusters.
The area display unit can also display the behavior for each cause of failure. This is done by displaying an arrow from the center of the i-th distribution area toward the center of the distribution area where the i-th distribution area is likely to transition next. A distribution region having a high probability that the i-th distribution region will transition next is a distribution region that gives a high transition probability matrix in the i-th row of the transition probability matrix.
In addition to arrows, a distribution region with a high probability of transition from the center of the i-th distribution region to the next may be displayed using a radar chart.
In the case of a rectangular area, the center of the distribution area is obtained by the average value of the threshold values of the sensors forming the rectangular area. In the case of a supercircle region, it is obtained as the center of each cluster.
The analyst analyzes the cause of each sample answering correctly and the cause of error by comparing the behavior for each failure cause with the behavior of each sample. For example, if the behavior of the sample is closer to the wrong cause of failure than the correct cause of failure, it becomes a cause of an incorrect answer. In such a case, the region is manipulated so that the distribution region is separated from the sample and the distribution region of the cause of failure.
The time series data display unit 1303 displays one or more time series data of each sample. The time series data display unit 1303 can simultaneously display the time series data and the distribution area. In the case of a rectangular area, the threshold value of each sensor is displayed perpendicular to the vertical axis. In the case of a supercircle region, the cluster number to which the data at each time belongs is displayed.
The time-series data display unit 1303 has a function of clearly indicating a portion in which the classified failure cause behavior features appear in each sample. In FIG. 13, the colored part in the time-series data display unit 1303 is a part where the behavior characteristics of the cause of failure appearing. This behavior feature is determined by the similarity rj at time j obtained by comparing the behavior of each failure cause with the behavior of the sample. The similarity rj at time j is obtained by the following equation.

Here, cj is a failure cause behavior vector at time j. This is a normalized vector from the center of the distribution region at time j to the center of the distribution region with a high probability of being at time j + 1. Dj is a sample change vector at time j. This is a normalized vector from the data value at time j to the data value at time j + 1.

The part where rj obtained in this way is equal to or greater than the threshold value is the part where the behavior characteristics of the classified cause of failure appear. By clarifying to the analyst the part where the failure cause behavior characteristics classified in this way appear, the analyst can adjust the area by paying attention to the data related to the corresponding time, and the analysis work Can be made more efficient.
The classification result display unit 1304 displays the sample name, the cause of failure stored in the case storage unit 401, the classification result obtained by the second classification unit 903, and the classification probability for each learning target sample.
The classification result display unit 1304 has a sorting function. This displays sample names, failure causes, classification results, and classification probabilities in ascending and descending order. It also has a color-coding function. For example, it is possible to display a different color for each cause of failure, display a different color for each classification result, or change the color shade according to the value of the classification result.
When an analyst selects a specific sample in the classification result display unit 1304, the corresponding sample is displayed in the distribution display unit 1302 and the time-series data display unit 1303.
A representative example display unit 1305 displays a representative sample for each cause of failure. A representative sample is a sample with the highest classification probability. For example, as a representative example of the failure cause A, a sample having the highest classification probability for the failure cause A is selected.
When a specific sample is selected on the representative example display unit 1305 by the analyst, the corresponding sample is displayed on the distribution display unit 1302 and the time-series data display unit 1303.
The analyst can efficiently learn the characteristic behavior for each cause of failure using the sorting function and the color coding function of the representative example 1305 and the classification result display unit 1304.
The analyst operates the region so that the characteristic behavior region for each cause of failure matches the distribution region displayed on the region display unit.
In the area operation unit 1306, the analyst operates the area. In the case of a rectangular area, this is done by changing the threshold value, creating a new area, or deleting it. In the case of a supercircle region, the distribution region is manipulated by moving the cluster center or increasing / decreasing the number of clusters. Further, the region operation unit 1306 is constantly updated by the distribution region directly operated on the region display unit.
If the reverification execution unit 1307 is selected by the analyst, the process proceeds to step 804. For example, in the case of a rectangular area, the distribution area is determined by the threshold value of each sensor set by the area operation unit. In the case of a supercircle region, the k-means method is executed again with the cluster center set by the region operation unit 1306 as an initial value.
When it is determined by the analyst that a predetermined correct answer rate has been achieved, the learning determination unit 1308 is selected by the analyst. Thereafter, the process proceeds to step 812.
As described above, the analyst repeatedly confirms the behavior characteristics for each failure cause, compares the behavior characteristics for each failure cause with the behavior characteristics of each sample, manipulates the distribution area, and updates the classification result by re-learning. Study to improve the correct answer rate.
In step 812, the classifier storage unit 905 stores the region obtained by the region learning unit and the transition probability matrix obtained by the transition probability learning unit.

FIG. 14 shows a hierarchical structure of the classifier storage unit.
The classifier storage unit includes a vehicle type hierarchy 1401, a failure type hierarchy 1402, a classification information hierarchy 1403, and a transition probability matrix hierarchy 1404. The vehicle type hierarchy 1401 stores vehicle types M1, M2,. The failure type hierarchy 1402 stores failure types F1, F2,. In the classification information layer 1403, a transition probability matrix and a distribution area are stored. The transition probability matrix has a transition probability matrix hierarchy 1404 further below, and a transition probability matrix is stored for each failure cause A, B,. In the distribution area, a distribution area 1102 used in the second classification unit 903 is stored. Each data of the sample is converted into area data by this distribution area.
In step 813, the flow ends.

  The first classification unit 404 has the same configuration as the second classification unit 903. Each time the cause classification server 103 receives data at the time of failure, the cause classification of the received operation data 403 is performed and output to the classification result display unit 405. In the first classification unit 404, the distribution region 1102 and the transition probability matrix 1104 read the corresponding vehicle type, failure type distribution region, and transition probability matrix from the classifier storage unit 905.

  The classification result display unit 405 displays the failure cause classification result obtained by the first classification unit 404. This information is written by the first classification unit 404. The classification result display unit 405 is executed every time a classification result is output from the first classification unit 404. When the operation data 403 is data obtained from the vehicle 101, the same information as the classification result display unit 405 is displayed on the maintenance base 105, so that it can be used for efficient maintenance work.

An overview of the classification result display unit 405 is shown in FIG. FIG. 15 is a diagram showing the configuration of the classification result display unit 405. The classification result display unit 405 includes a vehicle type information display unit 1500, a failure type display unit 1501, a data information display unit 1502, a classification cause display unit 1503, and a treatment method display unit 1504.
The vehicle information display unit 1500 displays information about the vehicle in which the operation data that classifies the cause of failure has occurred. Here, the vehicle information includes a vehicle ID, a travel distance, a model such as a vehicle body and an engine.
The failure type display unit 1501 displays the failure type. A data information display unit 1502 displays information on the operation data 403. This is a characteristic of the value of each sensor and time series data of each sensor. The classification cause display unit 1503 displays the probability corresponding to each failure cause output from the first classification unit 404. The treatment method display unit 1504 displays a treatment method for each cause of failure registered in advance. At the maintenance base 105, efficient maintenance work is performed with reference to the displayed probability of failure cause and its treatment method.

As mentioned above, this invention is not limited to an Example, Various modifications are included. The vehicle does not need to be a commercial vehicle, and the idea of the present invention may be applied to maintenance of a ship other than the vehicle, a moving body exemplified by an aircraft, and a machine other than the moving body. The above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

101 ... vehicle 102 ... communication base 103 ... cause classification server 104 ... analyzer 105 ... maintenance base

Claims (6)

In the failure cause classification device,
An input unit for inputting operation data indicating the state of the device;
A storage unit for storing operational data of the state of the device;
A classification unit that classifies the operation data input in the input unit for each state of the apparatus using the operation data stored in the storage unit ;
A transition probability learning unit having a region learning unit that learns a distribution region of machine operation data and a transition probability calculation unit that learns a transition of the distribution region for each cause of failure;
Have
The classification unit includes a region conversion unit that converts operation data related to machine failure into region data based on the learned distribution region, and a classification probability for each failure cause according to a transition probability matrix of the region data learned for each cause of failure. While having a probability calculation unit to calculate , using the transition of the distribution area and the distribution area for each failure cause, calculate the classification probability to each failure cause of the operation data,
A learning operation unit displays behavior characteristics for each cause of failure to an analyst, and the analyst compares the behavior characteristics for each cause of failure with the behavior of operation data, and operates the distribution region. Cause classification device. In the failure cause classification apparatus according to claim 1 ,
A failure cause classification apparatus comprising: a display unit characterized by displaying a characteristic portion for each failure cause of operating data. In the failure cause classification apparatus according to claim 2 ,
The failure cause classification apparatus, wherein the display unit displays a second distribution region having a high probability of transition from the first distribution region for each failure cause. In the failure cause classification apparatus according to claim 2 ,
The failure cause classification apparatus, wherein the display unit displays a portion in which the change direction of the operation data is similar to the change direction at each time of the distribution area for each failure cause. In the failure cause classification apparatus according to claim 2 ,
The failure cause classification apparatus, wherein the display unit displays the distribution area together with a scatter diagram of data used for learning. In the failure cause classification apparatus according to claim 2 ,
The display unit displays a case with a high classification probability for each cause of failure or operation data selected by an analyst as a representative example.