JP6453086B2 - Operating data 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 failure rate of commercial vehicles is directly linked to the suspension of business, so the availability of commercial vehicles affects the profit and loss of the business. Therefore, the operation rate is increased by improving the efficiency of maintenance work for commercial vehicles. In order to improve the efficiency of maintenance work, for example, the cause of failure is classified from operation data at the time of a vehicle failure, thereby trying to improve the efficiency of maintenance. However, since only specialists can classify failure causes of operating data manually, 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.

  Research has also been conducted on the selection of data used for learning. For example, in Patent Document 2, there is a description of excluding abnormal data from learning data by cross-validation. Patent Document 2 learns only data when the machine is normal and detects abnormal data.

JP-A-8-221113 JP2011-70635A

  Conventionally, failure cause classification is performed by comparing operation data at the time of a machine failure and operation data of past cases. Therefore, in order to appropriately classify operation data at the time of the occurrence of a mechanical failure, data of similar past cases needs to exist on the database. Therefore, since the problem of such a method is that the amount of operation data of past cases affects the classification performance of the cause of failure, it has conventionally been necessary to prepare sufficient operation data of 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 occurrence of the failure changes depending on various factors such as the progress of the failure, the load status, the external environment, and the operation information, and the operation data is diverse. Therefore, if a sufficient amount of operation data is prepared when a machine failure occurs, it is necessary to always acquire the data at the time of the failure and reflect it sequentially in the learning data. In this case, if all data is learned as it is, it may adversely affect the failure cause classification performance. This is because the operation data does not reflect the characteristics of the failure, and learning is performed together with the cause of the failure due to an incorrect maintenance operation.

  In patent document 1, the point which selects data is not considered. Patent document 2 describes the selection of learning data in a method generally called “unsupervised learning”, but it is simple in a problem called “supervised learning” in which operation data and its failure cause are combined and learned. Application is difficult. In “Unsupervised learning” described in Patent Document 2, it is possible to determine whether it is normal or not normal. However, if it is not normal, the reason why it is not normal cannot be determined. It is because it cannot be done.

  Therefore, in the conventional method, there is a possibility that the data at the time of abnormality is learned as normal, which has an adverse effect on the abnormality detection performance.

An apparatus for realizing the present invention is an operation data classification apparatus that classifies machine operation data based on failure cause information and outputs failure cause classification information, and the operation data classification apparatus includes an evaluation unit. When evaluating whether the cause of failure information corresponding to the input operation data is correct, the evaluation unit calculates the cause of failure information included in the case data information for each case data included in the operation data. By replacing with different failure cause information, the failure cause classification information output by the operation data classification device is changed, and stored in the case storage unit and a case storage unit for storing teacher information and case data Before and after excluding a case from a certain case, calculate the correct answer rate from the teacher information and the case data, and the degree of reliability of the case excluded according to the change in the correct answer rate There is a teaching information reliability evaluation unit for evaluating the reliability, and case selection unit for selecting a case to be used for classification of the operation data in accordance with the reliability, characterized in that it has a.

  According to the present invention, it is possible to perform accurate failure cause classification even when the true failure cause of a machine is unknown.

It is the figure which showed the system of a present Example. It is the figure which showed the structure of the cause classification server. It is the figure which showed the structure of the example memory | storage part. It is the figure which showed description of the sample. It is the figure which showed the image of the process performed in a teacher information estimation part. It is the figure which showed the structure of the teacher information estimation part. It is the figure which showed the processing flow of the teacher information estimation part. It is the figure which showed the processing flow of the case selection part. It is the figure which showed the image of the process performed by teacher information reliability. It is the figure which showed the processing flow of the teacher information reliability evaluation part. It is the figure which showed the processing flow of the teacher information candidate influentiality evaluation part. It is the figure which showed the structure of the display part. It is the figure which showed the example of the screen of a display part. It is the figure which showed the structure of the cause classification | category part. It is the figure which showed the structure of the selection example memory | storage part. It is the figure which showed the structure of the classification result display part. It is the figure which showed the structure of the maintenance work input 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. In the present embodiment, the cause classification server 103 describes a method of classifying machine operation data based on failure cause information, but the same applies to a method of classifying operation data based on maintenance information. Can be realized.

  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.

  Here, failure cause classification is generally performed by a technique called “supervised learning”. This supervised learning is generally used for classification problems. In the classification problem, a combination of input data and output data is learned. Thereafter, when only input data is given, output data is predicted. In the operation data classification device according to the present embodiment, input data corresponds to sensor data at the time of vehicle failure, and output data corresponds to a cause of failure. There are various algorithms in a method called supervised learning. For example, algorithms such as support vector machines and random forests.

  FIG. 2 is a diagram for explaining the configuration of the cause classification server 103. The cause classification server 103 includes a case storage unit 202, a teacher information estimation unit 203, a case selection process execution determination unit 204, a case selection unit 205, a teacher information reliability evaluation unit 206, a teacher information candidate influentiality evaluation unit 207, and a cause classification unit. 210.

The operation data 209 from the vehicle 101 is input to the cause classification unit 210 of the cause classification server 103. The analyst 104 can view the teacher information candidate influentiality evaluation unit 207 of the cause classification server 103 on the display screen 208 . Further, the diagnosis result classified in the cause classification unit 210 of the cause classification server is displayed on the classification result display unit 211 of the maintenance base 105.

The case storage unit 202 stores maintenance work content data input from the maintenance work input unit 201. This is directly input by maintenance personnel at the maintenance site 105. In addition, the analyst 104 may input it, not a maintenance person. The input is performed every time maintenance work is completed at the maintenance site 105. Alternatively, it may be performed at a predetermined timing such as once a day or once a week.
The case storage unit 202 stores data related to failures that have occurred in the past. The present embodiment is an apparatus that learns failure causes and case data using the cause classification server 103 and classifies operation data based on the cause of failure, and a maintenance work input unit that inputs maintenance work to the case storage unit 202 201, a case storage unit 202 for storing the cause of failure and the case data, vehicle specifying information that is information for specifying a vehicle of each case, and a maintenance history, and the maintenance history stored in the case storage unit 202 Teacher information estimation unit 203 for estimating the cause of failure from the probability of failure cause when performing maintenance work and the probability of failure cause due to maintenance work performed within a certain period of the same vehicle with the vehicle identification information, A teacher information reliability evaluation unit 206 that evaluates the reliability representing the degree of reliability of the cause of failure linked to the case, and a case selection unit 2 that selects a case to be used for classification according to the reliability 5 will be mainly described operation data classification device characterized in that it has a classification unit 210 classifies the operation data on the basis of the failure causes and selected cases, the.

  FIG. 3 shows an example of data stored in the case storage unit 202. The failure number 301 stores a failure number assigned for each failure. The failure numbers are automatically assigned from the case storage unit 202 in the order registered in the case storage unit 202. The date and time 302 stores the date and time when the failure occurred. The location 303 stores the location where the failure occurred. The vehicle ID 304 stores the vehicle ID of the vehicle in which the failure has occurred. The vehicle ID is a string of numbers and letters assigned to each vehicle body, and there is only one vehicle body having the same number. In the vehicle information 305, vehicle information of a vehicle in which a failure has occurred is input. The vehicle information is information such as a model of a vehicle and a model of a component such as a mounted engine. The failure type 306 stores the failure type of the failure that has occurred. The failure type is information indicating a malfunction, abnormality, symptom, or the like caused by the failure. For example, there is a sudden engine stop or temperature abnormality. The failure cause 307 stores the cause of the failure that has occurred. A failure cause is a cause corresponding to a failure type. For example, failure types such as sudden engine stop include failure causes such as engine failure and fuel supply stop. The case sample 308 stores a sample when a failure occurs. This sample is data of various sensors attached to the vehicle 101, for example. In the maintenance operation 309, the maintenance operation performed for the failure is input. In this maintenance work, for example, engine parts are replaced when an engine malfunctions, and fuel is replenished and piping is checked when fuel supply is stopped.

  Failure number 301, date and time 302, location 303, vehicle ID 304, vehicle information 305, failure type 306, and case sample 308 are the cause classification when the cause classification server 103 receives the sensor data at the time of failure from the vehicle 101 through the communication base 102. Automatically stored by the server 103. The maintenance work 309 is input from the maintenance work input unit 201. The failure cause 307 is input by the teacher information estimation unit 203 at a predetermined timing such as when the maintenance history is input or after a certain period of time has elapsed after input.

  FIG. 4 shows an example of the case sample. The time column 401 stores time information when a failure occurs. Sensor data relating to the vehicle, such as an accelerator train 402 and an engine speed train 403, is stored in synchronization with time. The cause classification server 103 learns by combining the failure cause 307 and sample information, and classifies the cause of the operation data 209 when the failure occurs.

  FIG. 5 is a diagram for explaining the concept of the teacher information estimation unit 203. The horizontal axis represents the passage of time. The time has passed as you go to the right. The upper side of the broken line represents the maintenance history related to the vehicle with the vehicle ID 0002. In vehicle ID0002, a failure occurred at time t1, and maintenance H1 was performed. Thereafter, since no failure has occurred within a certain period, it is estimated that the maintenance H1 was effective. Therefore, it is estimated that there is a high possibility that the failure cause is the failure cause C1 corresponding to the maintenance H1. In the above example, the maintenance H1 corresponds to fuel replenishment and piping inspection, and the failure cause C1 corresponds to fuel supply stoppage. In such a case, the failure cause C1 is input to the case storage unit 202 as a promising candidate for the case sample of the failure that occurred at time t1. In this embodiment, only one example is described as the maintenance work. However, it is common for maintenance personnel to perform a plurality of maintenance, and the maintenance work is not limited to suggesting only one example. .

  On the other hand, the lower side of the broken line represents the maintenance history related to the vehicle ID 0001. For vehicle ID 0001, maintenance H1 was performed for the failure that occurred at time t2. However, at time t3 not so far from time t2, a failure of the same failure type occurred again, and maintenance H1 and maintenance H2 were performed. After that, the failure did not occur. In this case, it is estimated that the maintenance H1 is not effective and the maintenance H2 is effective. Therefore, it is presumed that there is a high possibility that it is the failure cause C2 corresponding to the maintenance H2. In the above-described example, the maintenance H2 corresponds to the replacement of the engine parts, and the failure cause C2 corresponds to the engine failure. In such a case, the failure cause C2 is selected as an effective candidate for the case examples of the failure that occurred at time t2 and time t3, and is input to the case storage unit 202.

  As described above, the teacher information estimation unit 203 estimates that the maintenance work that has been performed in advance is not valid when the same failure type occurs within a certain period of time, and the maintenance work that has been valid from the history. In addition, it is assumed that failures that occur within a certain period are caused by the same cause of failure. Furthermore, even if the same vehicle ID is used, if the same type of failure occurs at a certain time or longer, the same cause of failure is not necessarily caused. For example, in the vehicle ID 0002 in FIG. 5, even when the same failure type occurs at time t4 after a certain period of time after time t1, maintenance H1 remains effective for the failure that occurred at time t1. Suppose there was. Note that the maintenance history information of the failure that occurred at time t1 is not used to estimate the cause of the failure that occurred at time t4.

FIG. 6 shows the configuration of the teacher information estimation unit 203. The teacher information estimation unit 203 includes a vehicle maintenance history search unit 601, a failure period extraction unit 602, and a teacher information analogy estimation unit 603.

FIG. 7 shows a processing flow of the teacher information estimation unit 203.

  In step 701, a failure cause estimation flow is started. This is performed at a predetermined timing, such as once a day or once a month, or when the teacher information estimation unit is operated by an analyst.

  In step 702, a sample for estimating the cause of the failure is determined. When performing at the aforementioned predetermined timing, the cause of failure is estimated for all samples. Further, the cause of failure may be estimated only for samples whose cause of failure is unknown. When operated by the aforementioned analyst, there may be samples specified by the analyst in addition to all samples, unknown samples of the cause.

  In step 703, the vehicle maintenance history search unit 601 reads from the case storage unit 202 a failure case having the same vehicle ID as the sample whose failure cause is estimated. This reading reads all cases with the same vehicle ID. Also, only failure cases that occurred within a certain period from the failure occurrence time of the sample for estimating the cause of failure may be read.

  In step 704, an estimated case belonging to the same failure period as the sample for estimating the cause of the failure is extracted from the cases read by the failure period extraction unit 602. Here, the failure period indicates a period in which a failure has occurred due to the same cause of failure. Extraction of cases belonging to the same failure period is performed by the following method. First, it is assumed that failure cases that occurred within a certain period are extracted from a sample for estimating the cause of failure and belong to the same failure period. Furthermore, failure cases that occurred within a certain period are extracted from newly extracted failure cases that belong to the same failure period. Failure cases are extracted until there are no newly added failure cases, and the same failure period is set. The fixed period is given in advance by the analyst. This value may be the same for all failures, or may be different for each failure type and failure cause. When a different value is set for each failure cause, a different failure period is generated for each failure cause.

  In step 705, the cause of the failure is estimated from failure cases belonging to the same failure period. This is estimated by the following equation.

  Here, N indicates failure cases belonging to the same failure period. p (*) represents the probability. Ck is the cause of failure at the time of occurrence of the k-th failure, and Hk is maintenance work performed when the k-th failure occurs. p (Ck | Hk) represents the probability of the failure cause Ck when the maintenance work Hk is performed when the k-th failure occurs. Further, p (Ck | H1,..., Hk-1) represents the probability of the failure cause Ck when the maintenance operations H1 to Hk-1 are performed from the first to the (k-1) th. The left side represents the probability of each failure cause. These probabilities are given in advance by the analyst. Further, it may be obtained from past case data. The failure cause with the highest probability obtained by Equation 1 is set as the failure cause input to the case storage unit 202. Alternatively, the cause classification unit 210 may classify the failure causes N times for N cases, and the failure cause that is classified most frequently may be input to the case storage unit 202 as a failure cause.

  In the present embodiment, the cause of the failure is estimated by the teacher information estimation unit 203. However, for example, an effective maintenance work may be estimated as the teacher information. In this case as well, it is estimated in the same format as in Equation 1.

  In step 706, the same cause of failure is input to the data of failure cases belonging to the same failure period. At this time, the cause of failure is input for all the same failure periods. Further, the cause of failure may be input only to failure cases within a period shorter than the certain period used in step 704.

  In step 707, the failure cause estimation flow is terminated.

  The case selection process execution determination unit 204 determines whether to execute the process in the case selection unit 205. Processing performed by the case selection unit 205 described later has a high calculation load. Therefore, efficient operation is performed by providing the case selection process execution determination unit 204. This execution determination is performed for each unit in which failure cause classification is performed. Here, the unit in which the failure cause classification is performed is, for example, a failure type. Further, the unit in which the failure cause classification is performed may be a combination of failure type and vehicle information. The execution determination of the process in the case selection unit 205 is determined by the number of samples stored in the case storage unit 202. Here, the number of samples is the number of failure type samples to be processed by the case selection unit 205. If the number of samples is equal to or greater than a certain value, the processing in the case selection unit 205 is not executed. The threshold for the number of samples related to the case selection process execution determination is given in advance by an analyst. In addition to the case selection process execution determination based on the number of samples, determination based on the correct answer rate may be considered. In this case, a cross-validation called a single-out method is performed on the failure type sample to be subjected to the case selection process, and the correct answer rate is calculated using supervised learning. Here, the correct answer rate is a value obtained by dividing the number of samples classified into the cause of failure stored in the case storage unit by cross-validation by the number of samples to be cross-validated. If this correct answer rate is above a certain level, the process in the case selection unit 205 is not executed. The threshold of the correct answer rate regarding the case selection process execution determination is given in advance by an analyst. Further, the case selection process execution determination may be performed by the analyst. In this case, it is not executed after the time point when the analyst decides not to execute the case selection process, and is executed after the time point when the case selection process is executed.

  The case selection unit 205 selects a case used for failure cause classification.

  A technique for removing abnormal data from normal data by cross-validation as shown in Patent Document 2 is generally called outlier detection. A method of learning only normal sensor data as in Patent Document 2 is called “unsupervised learning”. Unlike unsupervised learning, unsupervised learning only learns input data. Thereafter, when new input data is given, the distance from the data learned so far is output. On the other hand, supervised learning learns input data and failure cause information or maintenance information. In Patent Document 2, sensor data when the machine is normal is given as input data for learning. Thereafter, sensor data is given when the machine is operating, and the distance from the learned sensor data when the machine is normal is calculated. If the distance is more than a certain distance, the machine is diagnosed as abnormal. Therefore, if abnormal data is mixed in the data to be learned, an appropriate abnormality diagnosis cannot be performed. This is because even if abnormal sensor data is given during machine operation, if there is data close to the learning data, the distance from the learning data will not be sufficiently large. Therefore, in unsupervised learning, outlier detection that excludes data that is estimated to be abnormal from learning data has been sufficiently studied. Data excluded by outlier detection in unsupervised learning is data with a low occurrence frequency. Therefore, in unsupervised learning for calculating the distance from learned data, data with low occurrence frequency becomes noise and adversely affects the learning result.

However, in the conventional supervised learning, Study on exclude data that is estimated to abnormal from the learning data approach is not sufficiently performed. This is because the main problem handled by supervised learning is a classification problem. In the classification problem, it is important to find different features from many sensor data. For example, the problem of classifying different failure causes such as engine failure and fuel supply stop in the failure type of sudden engine stop is handled. The feature of the failure classification is that the engine does not operate even when fuel is sufficiently supplied in the event of an engine failure, and that fuel is not sufficiently supplied when fuel supply is stopped. In supervised learning, a feature effective for classification is searched from a lot of sensor data. Therefore, even if the frequency of data generation is low, it may be an effective feature for classification of failure causes. In addition, there is sufficient learning data for problems handled in normal supervised learning. In such a case, when data with low occurrence frequency is not effective for classification, it is not adopted as a feature. For this reason, verification with respect to learning data such as outlier detection is not performed in supervised learning.

  However, in failure cause classification, selection of learning data is important. This is because the failure cause, which is output data at the time of learning, has low reliability. In normal supervised learning, the reliability of the output data is sufficiently high. For example, general supervised learning deals with natural language processing and image recognition. For example, consider the problem of identifying sea images and mountain images from a large number of images in image recognition. Many people can identify an image as an ocean image or a mountain image. Therefore, the reliability of the output data (whether it is a sea image or a mountain image) learned by supervised learning is high. On the other hand, in the problem of failure cause classification, the reliability of output data (failure cause) is not high. This is because it is difficult for machine specialists to identify the cause of failure, there are not many machine specialists, and it takes a long time to identify the cause of failure per unit. For this reason, it is practically difficult for the machine specialist to generate output data. It is also difficult to generate output data by maintenance work at the maintenance site. This is because the maintenance staff does not have sufficient knowledge of maintenance due to inexperienced maintenance personnel or a new model of the failed vehicle. This is also because maintenance personnel at maintenance bases may perform maintenance more than necessary. Here, the maintenance more than necessary is, for example, to perform fuel replenishment and piping inspection in addition to replacement of engine parts for the cause of failure such as engine failure. In such a case, a machine specialist is required to determine which maintenance work was effective against the cause of the failure. Therefore, even in this case, since the maintenance that was effective is unknown, the true cause of the failure is unknown, the appropriate input cannot be performed, and the reliability of the output data is lowered. Like the teacher information estimation unit 203 in the present embodiment, it is possible to estimate the true cause of failure from the maintenance history, but in this case as well, the reliability of the output data is not sufficiently high.

  In order to deal with such a problem, in this embodiment, a case used for failure cause classification is selected. This case selection evaluates the reliability of the output data for each case, and the unreliable data is not used for learning or gives a cause of failure that is assumed to be highly reliable. Thereby, the reliability of the learning data used for failure cause classification is improved.

FIG. 8 shows a processing flow of the case selection unit 205 , teacher information reliability evaluation unit 206, and teacher information candidate influentiality evaluation unit 207 .

  In step 801, the processing flow of the case selection unit is started. This is performed at a predetermined timing such as once a day, once a month, or when a new case is input to the case storage unit. Further, the flow of the case selection unit is executed and determined for each unit in which the failure cause classification is performed. If the case selection process execution determination unit determines not to execute the case selection process, the flow is not started in step 801. In that case, all samples are used for failure cause classification. Alternatively, use a sample that is already selected at that time.

  In step 802, the teacher information reliability evaluation unit 206 evaluates the reliability of each sample. This reliability represents the degree to which the cause of failure given to each sample can be trusted. A sample given a cause of failure different from the true cause of failure is less reliable, and a sample given a cause of failure equal to the true cause of failure is more reliable. Here, the teacher information is output data in supervised learning, and is a cause of failure in this embodiment.

  In step 803, the teacher information candidate influentiality evaluation unit 207 estimates more effective teacher information for the sample whose reliability is determined to be lower than a certain threshold value in step 802. Here, influential teacher information is teacher information presumed to be a true cause of failure. A certain threshold value is given in advance by an analyst. Further, in the present embodiment, influential teacher information is estimated only for samples whose reliability is lower than a certain threshold value, but teacher information may be estimated for all samples.

  In step 804, the cause of failure of the sample with low reliability is changed. In other words, the teacher information that is determined to have a high level of power by the teacher information candidate power level evaluation unit 207 is used for analysis as a possible failure cause.

  In step 805, the correct answer rate before and after the failure cause change is calculated. That is, the correct answer rate calculated for the cause of failure at the time of step 802 and the correct answer rate after changing the cause of failure of the sample with low reliability in step 804 are calculated.

  In step 806, the correct answer rate before and after the failure cause change is compared. If the correct answer rate is high when the cause of failure is changed, the process proceeds to step 807. On the other hand, if the correct answer rate is high when the cause of failure is not changed, the process proceeds to step 808.

  In step 807, the cause of failure, which is teacher information used for failure cause classification, is determined. In this case, the cause of failure having a high intensity estimated in step 804 is transmitted to the cause classification unit 210 as a case and cause of failure used by the cause classification unit 210. Thereafter, the process proceeds to step 812 to end the case selection flow.

  In step 808, the correct answer rate is calculated before and after excluding samples with low reliability without changing the cause of failure. If the process proceeds to step 808, the correct cause rate is lowered when the cause of the failure is changed, so the failure cause is not changed. In this case, the correct answer rate in the cause classification unit 210 is improved by excluding samples with low reliability as samples that adversely affect learning.

  Therefore, the correct answer rate is calculated based on the cause of failure at the time of step 802, the correct answer rate in terms of the number of samples, and the number of samples after excluding samples with low reliability that are equal to or less than the number of samples at the time of step 802.

  In step 809, the correct answer rates before and after excluding samples with low reliability are compared. If the correct answer rate is high when samples with low reliability are not excluded, the process proceeds to step 810. If the correct answer rate is high when samples with low reliability are excluded, the process proceeds to step 811.

  In step 810, a sample used for failure cause classification is determined. In this case, the cause at the time of step 802 and the sample are transmitted to the cause classifying unit 210 as a case of using the cause classifying unit 210 and the cause of failure. Thereafter, the process proceeds to step 812 to end the case selection flow.

  In step 811, a sample used for failure cause classification is determined. In this case, the sample after excluding the sample with low reliability at the time of the step 802 is transmitted to the cause classifying unit 210 as an example of using the cause classifying unit 210 and the cause of the failure. Thereafter, the process proceeds to step 812 to end the case selection flow.

  FIG. 8 shows an example of changing the cause of failure for all samples with low reliability, but the cause of failure may be changed for only a part of samples with low reliability. In this case, the failure cause of the sample with the lowest reliability is changed, and the correct answer rate is calculated. If the correct answer rate has improved from before the change of the cause of failure, the cause of failure of the sample with the lowest reliability is changed. If the correct answer rate does not improve before the change of the cause of failure, the failure cause of the sample with the lowest reliability is not changed. Next, the teacher information of the second least reliable sample is changed, and the correct answer rate is calculated. If the correct answer rate has improved from before the change of the cause of failure, the cause of failure of the second least reliable sample is changed. If the correct answer rate does not improve before the change of the cause of failure, the failure cause of the second least reliable sample is not changed. In this way, the correct answer rate may be calculated in order from the sample with the lowest reliability, and the cause of failure of the sample with the improved correct answer rate may be changed. In addition, if there is a sample whose failure cause has been changed after considering the change of the failure cause of all the samples once, calculate the reliability again, and change the failure cause in order from the sample with the lowest reliability. You may consider it. In addition, the change of the cause of failure may be examined in order from the sample with the highest reliability, or the change of the cause of failure may be examined in an order irrelevant to the reliability, for example, the date of occurrence of the case.

  Further, although FIG. 8 shows an example in which all samples with low reliability are excluded from learning at once, only a part of samples may be excluded from learning as in the failure cause changing procedure. Also in this case, the correct answer rate is examined in order from the sample with the lowest reliability, and the sample with the lower reliability is excluded when the correct answer rate is improved from before the exclusion. Further, the order of exclusion is not limited as in the case of changing the cause of failure.

  Moreover, the change of the cause of failure and the removal of samples may be repeated alternately.

  In addition, the evaluation may be based on the number of correctly answered samples instead of the correct answer rate, and the change of the cause of failure and the exclusion of samples may be applied only when the correct answer rate is improved by a certain level or more.

  In this way, the correct answer rate of the cause classification unit is improved by selecting the cause of failure in the use, nonuse or cause classification unit of each case based on the correct answer rate.

  FIG. 9 shows an image of processing in the teacher information reliability evaluation unit 206. Here, for the sake of simplicity, a case will be described in which only one sample is given a failure cause different from the true cause of failure among the N samples.

  All samples 901 indicate that all samples are given as learning data. A correct failure cause sample 1 surrounded by a solid line indicates that the true failure cause is given to the sample 1. The same applies to the correct failure cause sample 2 to the correct failure cause sample N-1. An erroneous failure cause sample N surrounded by a broken line indicates that the failure cause different from the true failure cause is given to the sample N. That is, N-1 samples from Sample 1 to Sample N-1 are given a true cause of failure, and only one sample N is given a cause of failure different from the true cause of failure. Assume that the correct answer rate is A when the correct answer rate is calculated in this state.

N-1 sample 902 represents N-1 samples. That is, in 902a, there are N-1 samples of samples 2 to N excluding the sample 1 of the correct cause of failure. In 902b, there are N-1 samples of samples 1, 3 to N excluding sample 2 of the correct cause of failure. In 902c, there are N-1 samples of Samples 1 to N-1, excluding the sample N that caused the wrong failure.

  Reference numeral 903 shows the correct answer rate and reliability in the N-1 sample. Reference numeral 903a shows the correct answer rate obtained from the N-1 samples of samples 2 to N and the reliability of the sample 1. At this time, it is assumed that the correct answer rate is equal to the correct answer rate A in all samples. In such a situation, sample 1 is not a sample that has a large influence on the cause of failure classification with at least N samples. Therefore, the reliability of sample 1 is medium. Reference numeral 903b shows the correct answer rate obtained from N-1 samples of samples 1 and 3 to N and the reliability of sample 2. At this time, the correct answer rate is lower than A. In this case, sample 2 is an important sample for the correct answer rate in the failure cause classification with N samples. Therefore, the reliability of this sample is high. Reference numeral 903c shows the N-1 correct answer rate and the reliability of the sample N after removing the sample N to which the cause of failure different from the true cause of failure is removed. In such a case, the correct answer rate is improved because samples having erroneous teacher information that adversely affects the failure cause classification are excluded. Therefore, the reliability of sample N is low.

As described above, the teacher information reliability evaluation unit 206 calculates a correct answer rate by excluding samples from all samples one by one. If the excluded sample is a sample given the correct cause of failure, the correct answer rate may not change or may decrease. Therefore, when the correct answer rate does not change or decreases, a medium or high reliability is given to the excluded sample. When the excluded sample is given a cause of failure different from the true cause of failure, it is assumed that the correct answer rate is improved. Therefore, when the correct answer rate is improved, the reliability of the excluded sample is lowered. In this way, the reliability of each sample is calculated.

  FIG. 10 shows a processing flow of the teacher information reliability evaluation unit 206.

  In step 1001, the processing flow is started. This is started at the timing when the process flow step 802 of the case selection unit is executed.

  In step 1002, the correct answer rate for all N samples is calculated.

  In step 1003, 1 is substituted for the number of repetitions i.

  In step 1004, the i-th sample is excluded from all the samples, and N−1 samples are generated.

  In step 1005, the correct answer rate is calculated using the N-1 samples generated in step 1004.

  In step 1006, the reliability is calculated from the correct answer rate for all N samples calculated in step 1002 and the correct answer rate for N-1 samples calculated in step 1005. The reliability is calculated by Equation 2.

Here, Ri represents the reliability of the i-th sample. A represents the correct answer rate in N samples. A _-i indicates the correct answer rate calculated for N-1 samples excluding the i-th sample.

  In step 1007, it is checked whether the reliability has been calculated for all samples. If i = N, it is determined that the calculation of the reliability of all N samples is completed, and the process proceeds to Step 1009. If i = N, the process proceeds to step 1008.

  In step 1008, i is incremented by one.

  In step 1009, the processing flow of the teacher information reliability evaluation unit is terminated.

  Further, here, a method for evaluating the reliability by excluding only one sample has been described. At this time, except for one sample to be excluded, the failure cause given first is used, but the calculation of the reliability and the change of the failure cause may be performed in parallel. For example, first, a sample with the lowest reliability is specified, and a possible failure cause is estimated by the teacher information candidate influence evaluation unit 207 for the sample with the lowest reliability, and the failure cause is changed. Thereafter, the reliability is calculated for all the samples once more, and the teacher information inefficiency evaluation unit 207 again estimates a possible cause of failure for the sample with the lowest reliability, and changes the cause of failure. In this way, calculation of reliability, estimation of influential teacher information of samples with low reliability, and change of cause of failure may be repeated. In this case, the iterative calculation is terminated when the number of repeated calculations reaches a certain value or when the cause of failure is not changed.

  In addition, here, a method of evaluating reliability by excluding only one sample has been described. Conversely, a method of adding samples one by one and evaluating the reliability is also conceivable. In this case, the correct answer rate is calculated with N samples whose reliability is already known. After that, a sample for calculating the reliability is added, the correct answer rate is calculated with N + 1 samples, and the reliability is calculated from the correct answer rate before and after adding the sample. When there are a plurality of samples whose reliability is to be calculated, the samples are sequentially added to N samples whose reliability is known, and the calculation of the correct answer rate and the calculation of the reliability are repeated with N + 1 samples.

The reliability calculated in this way represents the contribution of each sample to the correct answer rate. For this reason, if the sample with high reliability is used in the cause classification unit 210, an improvement in the correct answer rate can be expected. On the other hand, if a sample whose reliability is evaluated to be low is used in the cause classification unit 210, there is a high risk of a decrease in the correct answer rate. Based on this reliability, the case selection unit 205 selects cases so that the correct answer rate in the cause classification unit 210 is high. The reliability is output to the display screen 208 and confirmed by the analyst 104. Thereafter, the analyst can determine whether each sample is used or not used in the cause classification unit 210 according to the reliability.

  FIG. 11 shows a processing flow of the teacher information candidate influentiality evaluation unit 207.

  In step 1101, a processing flow is started. This starts at the timing when step 803 of the processing flow of the case selection unit 205 is executed.

  In step 1102, a sample for calculating the influential degree of the teacher information candidate is determined. This is automatically executed by the teacher information candidate power calculator.

  In step 1103, the correct answer rate is calculated. Here, when calculating the correct answer rate, the cause of failure, which is the teacher information, is changed each time the steps 1104 and 1105 are repeated for the sample for calculating the influential degree. For the samples other than the sample for calculating the influential degree, the failure cause given at step 1101 is fixed as teacher information, and the correct answer rate is calculated.

  In step 1104, it is checked whether or not the sample for calculating the proficiency has calculated the correct answer rate when all failure causes are used as teacher information. If all failure causes are used as teacher information, the process proceeds to step 1106 if the calculation of the correct answer rate is completed for all, and the process proceeds to step 1105 if not completed.

  In step 1105, the cause of failure of the sample for which the degree of influence is calculated is changed. The failure cause to be changed is a failure cause that has not yet been calculated in step 1103.

  In step 1106, the influential degree of each failure cause is calculated from the correct answer rate for each failure cause. The calculation of the power level is performed by Equation 3.

  Here, Wc represents the influential degree of failure cause c. Ac represents the correct answer rate when the cause of failure is c. The denominator on the right side represents the sum of correct answer rates for all causes of failure.

  When changing the cause of failure of a sample with low reliability in the case selection unit or the like, the cause is changed to the cause of failure having the highest degree of influence. In this example, only the correct answer rate is used for the calculation of the influential degree. However, weighting may be performed using the occurrence probability of each failure cause obtained by Equation 1. In this case, the influential degree is obtained by Equation 4.

  Here, p (c) is the probability of occurrence of the failure cause c obtained by Equation 1.

  In step 1107, it is checked whether the influential degree has been calculated for all the samples. If the influential degree has not been calculated for all the samples, the process proceeds to step 1102 to calculate the influential degree of the sample for which the influential degree has not been calculated. If the calculation of potency is completed for all samples, the process proceeds to step 1108, and the processing flow of the teacher information candidate potency evaluation unit is terminated.

The influential degree calculated in this way represents a change in the correct answer rate when the cause of failure of the sample is changed. That is, when the cause is changed to a cause with a high degree of influentialness, an improvement in the correct answer rate in the cause classification unit 210 is expected. On the other hand, when the failure cause is changed to a low-potential cause of failure , a reduction in the correct answer rate in the cause classification unit 210 is expected. The case selection unit 205 determines the cause of failure in each case by using the influential degree of each sample calculated in this way. The influential degree is output to the display screen 208 and confirmed by the analyst 104. After that, the analyst can select the cause of failure when using each sample in the cause classification unit 210 according to the degree of proficiency.

FIG. 12 is a diagram showing the configuration of the display screen 208.

  The selection target display unit 1201 displays the target of the case selection unit. Here, the object is, for example, vehicle information or a failure type. Further, the vehicle ID may be displayed.

  The reliability display unit 1202 displays the reliability of each case calculated by the teacher information reliability evaluation unit 206.

  The power display unit 1203 displays the power of each case calculated by the teacher information candidate power evaluation unit 207.

  The maintenance history display unit 1204 displays the maintenance history of each case stored in the case storage unit 202.

  The data display unit 1205 displays sensor data such as time series data and scatter diagrams for each case.

  The representative example display unit 1206 displays representative time series data of each cause of failure and sensor data such as a scatter diagram.

  In the case use selection unit 1207, the analyst 104 can select use or non-use of each case in the cause classification unit 210.

  In the cause change unit 1208, the analyst 104 changes the cause of each case.

  The classification performance display unit 1209 displays the selected case and the classification performance at the cause of failure. Here, the classification performance is, for example, a correct answer rate. In addition to the correct answer rate, statistics such as intra-class variance, inter-class variance, and average value of each fault cause when the cause of failure is regarded as a class may be used.

  The case selection verification unit 1210 calculates the classification performance such as the correct answer rate when using, not using, changing the cause of failure, etc. of each case selected by the analyst 104 and displays it on the classification performance display unit 1209. This is executed at the timing when the execution of the case selection verification unit 1210 is selected by the analyst.

  In the processing flow of the case selection unit shown in FIG. 8, an example is shown in which the cause classification server automatically selects a case used by the cause classification unit 210. However, the analyst 104 may use, not use, and change the cause of each case in the cause classification unit 210 with reference to the reliability, the influential degree, the maintenance history, the sensor data, and the like on the display screen 208.

FIG. 13 is a diagram showing a screen example of the display screen 208.

  The selection target display unit 1301a displays the target of the case selection unit. In 131a, in addition to the vehicle information and failure type, the number of samples of each failure cause is displayed. In 1301b, the sample name and vehicle ID of the sample selected by the reliability display unit 1202 are displayed.

  The reliability display unit 1302 displays the sample name, failure cause, reliability, case selection result, and selection cause of each case. Here, the failure cause is a failure cause stored in the case storage unit 202. The reliability is the reliability of each case calculated by the teacher information reliability evaluation unit 206. The case selection result indicates that each case is used or not used in the cause classification unit 210, the case that is used is used in the cause classification unit 210, and the case that is not used is the cause classification. Not used in section 210. The cause of selection displays teacher information when each case is used in the cause classification unit 210, that is, a cause of failure. The case selection result and the selection cause can be used, not used, or the selection cause can be changed by the analyst 104. These functions are the functions of the case use selection unit 1207 and the cause change unit 1208 in FIG. As described above, the case use selection unit 1207 and the cause change unit 1208 may be arranged on the same screen as the reliability display unit 1202. Here, the background of the sample 3 being gray indicates that the sample 3 is selected by the analyst 104, and the influential degree of the selected case is displayed on the influential degree display unit 1303. The maintenance history is displayed on the maintenance history display unit 1304, and the sensor data of the selected case is displayed on the data display unit 1305.

  Further, the reliability display unit can be rearranged in ascending order, descending order, etc. according to the value of reliability. Furthermore, cases where the reliability is more than a certain level or less than a certain level are displayed in different colors.

  The potency display unit 1303 displays the potency of the selected case. Here, the potency display unit 1303 can be rearranged in ascending order, descending order, or the like according to the value of the potency. In addition, it is also possible to display cases where the influential degree is more than a certain level or less than a certain level by color coding.

  The maintenance history display unit 1304 displays a maintenance history of a vehicle having the same vehicle ID as the selected example. Here, the horizontal axis represents the date on which maintenance was performed, and displays the maintenance performed on each date.

  The data display unit 1305 displays sensor data of the selected case. The data displayed here can be selected from time-series data, a scatter diagram between sensors, data after processing such as FFT and moving average, and the like.

The representative example display unit 1306 is a part that displays data of a case selected as a typical example for each cause of failure. Similar to the data display unit 1305, the data to be displayed here can be selected from time series data, a scatter diagram between sensors, data after processing such as FFT and moving average, and the like.

  The classification performance display unit 1307 displays the current classification performance. Here, the case where the correct answer rate is used as the classification performance is illustrated.

  When selected by the analyst 104, the case selection verification unit 1308 performs cross verification using the current case selection result and the selection cause, calculates classification performance, and updates the value of the classification performance display unit 1307.

  The confirmation unit 1309, when selected by the analyst 104, transmits the current case selection result and the selection cause to the cause classification unit 210 as a case and cause of failure used by the cause classification unit 210.

  The analyst 104 efficiently uses the display unit to select a case where the correct answer rate is most improved.

  For example, a case with low reliability is selected using the reliability display unit 1302, and the influential degree and maintenance history of the case with low reliability are confirmed by the effectiveness display unit 1303 and the maintenance history display unit 1304. In the potency display unit 1303, the analyst 104 can confirm the cause of the failure that is expected to improve the correct answer rate most. In addition, by comparing the data of each case with the representative data of each cause of failure in the data display unit 1305 and the representative example display unit 1306, the data of each case and the data of potential failure cause candidates are compared, The cause that seems to show similar characteristics such as waveform and distribution is considered to be the dominant cause. In the maintenance history display section, if maintenance corresponding to a possible failure cause has not been performed and the failure has not recurred after a certain period of time has elapsed after maintenance, It is possible to infer that a probable cause of failure due to similarities in is likely to be different from the true cause. In this way, the analyst 104 comprehensively considers reliability, potency, data, maintenance history, etc., estimates the most likely failure cause, and changes the selection cause. In addition, if it is determined that any cause is not credible, the sample selection result of the sample is not used, thereby minimizing noise mixing in the data used in the cause classification unit 210 and improving the correct answer rate. There is expected.

  In addition, after changing the case selection result and the selection cause of each case, the change of the correct answer rate is confirmed by selecting the case selection verification unit 1308. When the correct answer rate is improved from before verification, it is assumed that a more correct cause of failure can be estimated, and the determination unit 1309 is selected to end the case selection. On the other hand, if the correct answer rate decreases, it is determined that there is an error in the case selection, and the change is rejected, or the case selection result that increases the correct answer rate and the selection cause are searched.

  In this way, the analyst 104 selects cases efficiently using information such as reliability, influentialness, data, and maintenance history.

  FIG. 14 shows the configuration of the cause classification unit. It comprises a classification unit 1401 and a selection case storage unit 1402. The classification unit 1401 performs classification by supervised learning using each vehicle and failure type sample stored in the selection example storage unit 1402 as input data, and the failure cause for each vehicle and failure type as teacher information. In this embodiment, an algorithm called SVM is used. However, this embodiment can also be applied to supervised learning methods other than SVM. Examples of supervised learning methods other than SVM include random forests and neural networks.

The selection case storage unit 1402 stores cases used in the cause classification unit. The cases stored here are a part or all of the cases stored in the case storage unit 202. The sample and the cause of failure transmitted from the case selection unit 205 are stored. Alternatively, the case selection result and the selection cause transmitted from the analyst 104 on the display screen 208 are stored.

FIG. 15 shows an example of a hierarchical structure in which past cause of failure and operation data stored in the selection example storage unit 1402 in this embodiment are linked. In the vehicle hierarchy 1501, vehicles M1, M2, M3... Are stored. The failure type hierarchy 1502 stores failure types F1, F2,. In the failure cause hierarchy 1503, the failure cause related to the failure type in the failure type hierarchy 1502 is stored. For example, for the failure type F1 of the vehicle M1, failure causes A, B. More specifically, when engine stop is stored as the failure type F1, the failure cause hierarchy 1503 stores failure causes such as engine component failures and fuel supply stoppage. In the sample hierarchy 1504 , samples associated with the vehicle type, failure type, and failure cause are stored.

FIG. 16 is a diagram showing the configuration of the classification result display unit 211. The classification result display unit 211 includes a vehicle information display unit 1600, a failure type display unit 1601, a data information display unit 1602, a classification cause display unit 1603, and a treatment method display unit 1604.
The vehicle information display unit 1600 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 1601 displays the failure type. The data information display unit 1602 displays information on the operation data 209. This is a characteristic of the value of each sensor and time series data of each sensor. The classification cause display unit 1603 displays each failure cause output from the cause classification unit 210. The treatment method display unit 1604 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.

  FIG. 17 is a diagram showing the configuration of the maintenance work input unit 201. The maintenance staff at the maintenance site 105 performs maintenance work with reference to the failure classification result displayed on the maintenance result display section. Thereafter, the maintenance work actually performed is input to the maintenance work input unit 201.

  The vehicle information display unit 1700 displays information related to the vehicle on which maintenance work has been performed. Here, the vehicle-related information includes a vehicle ID, a failure occurrence time, a model of a vehicle body, an engine, and the like.

  The failure type display unit 1701 is a type of failure that is a maintenance target, and includes, for example, an engine failure.

  The classification cause display unit 1702 displays the cause of failure displayed on the classification cause display unit 1603 of the classification result display unit 211.

  The maintenance staff refers to information displayed on the vehicle information display unit 1700, the failure type display unit 1701, and the classification cause display unit 1702, and inputs a combination of an actual failure and maintenance.

  The maintenance date input unit 1703 inputs the date and time when maintenance was performed by maintenance personnel.

  The maintenance work input unit 1704 inputs maintenance work actually performed by the maintenance staff.

  As described above, in this embodiment, the amount of data is obtained by repeatedly inputting maintenance work by maintenance personnel, estimating the cause of failure from the maintenance work, evaluating the reliability of the estimated cause of failure, and estimating the leading cause of failure. Provide a mechanism that improves the rate of correct answers as the number increases.

  In this embodiment, the cause of failure is handled as teacher information. However, for example, effective maintenance work may be handled as teacher information. When valid maintenance work is handled as teacher information, the maintenance history is given as one or more teacher information candidates. The conventional supervised learning framework is used when there is a single maintenance operation that is a teacher information candidate, but the conventional supervised learning framework cannot be used when a plurality of teacher information candidates are given. In the present embodiment, it is possible to guess and classify teacher information that is presumed to be promising for such cases.

  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)

An operation data classification device that classifies machine operation data based on failure cause information and outputs failure cause classification information ,
The operating data classification device has an evaluation unit,
The evaluation unit,
When evaluating whether the failure cause information corresponding to the input operation data is correct, the failure cause information included in the case data information is replaced with different failure cause information for each case data included in the operation data. By changing the failure cause classification information output by the operating data classification device,
A case storage unit for storing teacher information and case data;
Before and after excluding a case from the cases stored in the case storage unit, the correct answer rate is calculated from the teacher information and the case data, and the trust is the degree of trust of the case excluded according to the change in the correct answer rate A teacher information reliability evaluation unit that evaluates the degree,
A case selection unit that selects a case to be used for classification of operation data according to the reliability;
An operation data classification device characterized by comprising:   The operation data classification device according to claim 1,
  Furthermore, the operation data classification device characterized by having a selection case storage part which memorizes selected case data and teacher information. The operation data classification device according to claim 1 ,
Change the teacher information of a case, calculate the correct answer rate for each teacher information, and evaluate the influential degree indicating the degree of influentialness for each candidate of teacher information from the correct answer rate for each teacher information An operation data classification device characterized by having an evaluation unit . In the operation data classification device according to claim 3 ,
The teacher information reliability evaluation unit according to any one or more of the number of cases in the case storage unit, the correct answer rate calculated from the case data stored in the case storage unit and the teacher information, and an operation by an analyst And a case selection process execution determination unit that determines whether to execute the teacher information candidate influentiality evaluation unit. The operation data classification device according to claim 4 ,
In addition to the teacher information and the case data, a case storage unit that stores vehicle identification information and maintenance history that is information for identifying a vehicle of each case;
A maintenance work input unit that inputs maintenance work performed on each case into the case storage unit;
A teacher information estimation unit that extracts cases within a certain period of time in the same vehicle as a certain case as the same failure period, and estimates a probable cause of failure or effective maintenance work from the maintenance work performed in the case of the failure period; ,
An operation data classification device characterized by having The operation data classification device according to claim 5 ,
An operation data classification device comprising: a display unit that displays the reliability, the influential degree, and the maintenance history of each case together.