JP2015170121A - Abnormality diagnosis device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an abnormality diagnosis device and the like that have a performance sufficient in detection capability of abnormality and factor analysis capability in abnormality diagnosis of an on-vehicle system.SOLUTION: An abnormality diagnosis device is configured to: acquire normal data of vehicle data having normality known (step S1); extract a plurality of characteristic factors from the normal data (steps S2 to S4); learn a data distribution of the characteristic factors to create a monitor model (step S5); acquire evaluation data serving as an evaluation object of an abnormality diagnosis on the basis of the monitor model; extract the characteristic factor from the evaluation data; perform an abnormality determination on the basis of the monitor model; inversely convert data on a space of the characteristic factor with only the characteristic factor determined normal into the same space as the vehicle data; create re-configuration data; and perform a factor analysis of abnormality on the basis of a difference between the evaluation data and the re-configuration data.

Description

  The present invention relates to an abnormality diagnosis device for diagnosing abnormality of a vehicle based on vehicle operation data.

In recent years, in-vehicle systems mounted on vehicles tend to be larger and more complicated. In an in-vehicle system, multiple ECUs (abbreviation of “Electronic Control Unit”) are connected to CAN (“Controller
Data is transmitted and received with each other via an in-vehicle network such as “Area Network”), and the operation is performed in a coordinated manner. In such an in-vehicle system, vehicle operation data (hereinafter abbreviated as “vehicle data”) is stored as time-series data and used for vehicle abnormality diagnosis.

  As an example of a technique for performing system abnormality diagnosis, for example, Patent Document 1 discloses a manufacturing process monitoring apparatus using a statistical quality control technique. According to the monitoring apparatus described in Patent Document 1, time-series apparatus log data input from a batch production apparatus for products is subjected to wavelet transform to determine whether each frequency component is normal or abnormal. Then, a set of adopted wavelet coefficients is generated in which the value of the wavelet coefficient of the frequency component determined to be normal is set to 0 and the value of the wavelet coefficient of the frequency component determined to be abnormal is left as it is. Next, this set of adopted wavelet coefficients is subjected to inverse wavelet transform to obtain reproduced waveform data, and this reproduced waveform data is subjected to principal component analysis to obtain two statistics of hottering T2 and residual Q, and the two statistics. An upper threshold is set for the quantity, and whether the batch manufacturing process is normal or abnormal is determined depending on whether the upper threshold is exceeded.

  Further, for example, Patent Document 2 discloses an abnormality diagnosis apparatus that diagnoses, for each monitoring target device, whether there is an abnormality in response to physical quantities from a plurality of monitoring target devices. According to the abnormality diagnosis apparatus described in Patent Document 2, the sensor includes three sensors and an independent component analysis processing apparatus that indirectly receives signals from the three sensors as signals via a preprocessing apparatus. The independent component analysis processing device separates and extracts spatially, temporally, and frequency independent components from the signal by using independent component analysis, and then outputs the separated and extracted signals to the abnormality determination device. . The abnormality determination device diagnoses whether there is an abnormality in the machine based on each signal.

  Further, for example, Patent Document 3 discloses a method of analyzing factors affecting quality in manufacturing a product. According to the method described in Patent Document 3, N types of condition data indicating various conditions in the manufacture of products are converted into P types of components that are uncorrelated with each other, and N> P. Multi-variate analysis is performed to calculate P influence indexes indicating the influence of each component of the quality on the quality, and the N influence data indicating the influence on the quality of each of the N types of condition data are calculated. Convert to

  Further, for example, Patent Literature 4 discloses a remote monitoring system that remotely monitors a plurality of facilities such as an air conditioner and a generator. According to the remote monitoring system described in Patent Document 4, measured sensor values detected by a plurality of sensors transmitted from the facility are acquired as needed, and detected by the plurality of sensors when the facility is operating normally. Based on the measured sensor values, a correlation between a plurality of measured sensor values detected by a plurality of sensors is obtained by principal component analysis, the correlation is stored as a prediction model, and the measured sensor values of the plurality of sensors are obtained. At this time, predicted sensor values of a plurality of sensors are obtained from the measured sensor values and the predicted model, and a failure sign of the facility is detected based on a difference between the measured sensor value and the predicted sensor value.

JP 2011-8735 A JP 2005-284519 A JP 2005-242818 A JP 2005-149137 A

  However, when the techniques of Patent Documents 1 to 4 are applied to abnormality diagnosis of an in-vehicle system, the following technical problems exist.

  In the vehicle data, there is a need for abnormality diagnosis based on the distribution of data in a normal time domain. However, since the technique described in Patent Document 1 targets the frequency domain, it cannot meet this need. In addition, it is highly possible that the vehicle data includes characteristics that cannot be captured by principal component analysis. Like the technique described in Patent Document 1, the vehicle data monitoring based only on principal component analysis is capable of detecting abnormalities. And the performance of factor analysis ability is insufficient.

  In the technique described in Patent Document 2, since abnormality detection is performed using feature amounts extracted by independent component analysis, the relationship between the feature amounts used for abnormality detection and the original data series is complicated, and factor analysis is performed. It is difficult. Further, in the technique described in Patent Document 2, since the number of feature amounts used for abnormality detection is not large compared to the number of original data series, the performance of abnormality detection capability is insufficient.

  In the technique described in Patent Document 3, normal data and abnormal data are given as teacher data when calculating an abnormality influence index, but abnormal data is rarely obtained in vehicle data. It is difficult to apply to abnormality diagnosis. Similarly to the technique described in Patent Document 2, the number of abnormality influence indexes is not large compared to the number of original data series, and thus the performance of abnormality detection capability is insufficient.

  In the technique described in Patent Document 4, a failure sign is detected based on a difference from the model obtained by the principal component analysis, but the tendency of the difference from the model obtained only by the principal component analysis is based on each abnormal phenomenon. Therefore, irrelevant noise is likely to be introduced, and the performance of anomaly detection capability is insufficient. Similarly to the technology described in Patent Document 1, monitoring of vehicle data based only on principal component analysis has insufficient performance in detecting abnormality and analyzing factor.

  The present invention has been made in view of the above-described problems, and the object of the present invention is to provide an abnormality diagnosis device having sufficient performance in detecting abnormality and factor analysis in abnormality diagnosis of an in-vehicle system. Is to provide.

  A first invention for achieving the above object is an abnormality diagnosing device for diagnosing a vehicle abnormality based on vehicle data which is multi-sequence time-series data, and is normal among the vehicle data. Normal data acquisition means for acquiring known normal data, feature factor extraction means for extracting a plurality of feature factors from the normal data, model generation means for learning a data distribution of the feature factors and generating a monitoring model, Evaluation data acquisition means for acquiring evaluation data to be evaluated for abnormality diagnosis out of the vehicle data, and abnormality determination means for extracting the characteristic factor from the evaluation data and performing abnormality determination based on the monitoring model; The feature factor space data is inversely transformed into the same space as the vehicle data only by the feature factor determined to be normal by the abnormality determining means, An abnormality diagnosis device comprising: reconfiguration data generation means for generating composition data; and abnormality factor analysis means for performing a cause analysis of abnormality based on a difference between the evaluation data and the reconstruction data .

  According to the first aspect of the present invention, it is possible to provide an abnormality diagnosis apparatus having sufficient performance in detecting abnormality and factor analysis ability in abnormality diagnosis of an in-vehicle system. In particular, according to the first invention, the data series that directly contributes to the abnormality can be clarified. In addition, even if it is not possible to clearly determine whether or not each data sample is normal, detailed analysis can be performed by factor analysis to support precise abnormality determination.

  In the first invention, when the number of series of the vehicle data is M (M is a natural number), the feature factor extracting means extracts K feature factors that satisfy K (K is a natural number)> M. Is desirable. This makes it possible to accurately represent a complex input sequence having many states with many feature factors. For example, the feature factor extraction means performs an overcomplete independent component analysis of the normal data, extracts K independent components, and uses them as the feature factors. Further, for example, the feature factor extracting means performs a principal component analysis of the normal data, extracts P principal components satisfying P (P is a natural number) <M, and weights the independent components as weights of the principal components. Expressed as a linear sum.

  The monitoring model generation means in the first invention sets a determination threshold value for each feature factor based on the monitoring model, and the abnormality determination means determines the feature factor in the evaluation data based on the monitoring model. It is desirable that the likelihood of the feature factor is calculated, the number of the feature factors whose likelihood of the feature factor is smaller than the determination threshold is an abnormality level, and the abnormality determination is performed based on the abnormality level. As a result, the number of feature factors that behave abnormally can be used as a criterion for abnormality determination, and it is possible to detect abnormality that is resistant to noise. For example, the monitoring model generation unit learns the data distribution of the feature factors by kernel density estimation.

  Further, for example, the abnormality factor analysis means in the first invention calculates a contribution to the abnormality for each feature factor based on a difference between the evaluation data for each feature factor and the reconstructed data, The group of the largest contribution series and the second largest series is extracted as a high contribution series set, and the number of appearances of the high contribution series set in all the samples of the evaluation data is totaled, and the number of appearances is large. The high contribution series set is output in order. This makes it possible to capture the overall abnormality.

  Further, for example, the abnormality factor analysis means in the first invention calculates a contribution to the abnormality for each feature factor based on a difference between the evaluation data for each feature factor and the reconstructed data, The set of the largest contribution series and the second largest series is extracted as a high contribution series set, and a sample of the evaluation data that maximizes the total reconstruction error of the high contribution series set is output. . As a result, a sudden abnormality can be captured.

  A second invention is a program for causing a computer to function as an abnormality diagnosis device for diagnosing an abnormality of a vehicle based on vehicle data which is multi-series time series data, and is normal among the vehicle data Normal data acquiring means for acquiring known normal data, feature factor extracting means for extracting a plurality of feature factors from the normal data, and model generation means for learning a data distribution of the feature factors and generating a monitoring model And evaluation data acquisition means for acquiring evaluation data to be evaluated for abnormality diagnosis out of the vehicle data, and abnormality determination means for extracting the characteristic factor from the evaluation data and performing abnormality determination based on the monitoring model And an abnormality degree calculating means for calculating the number of characteristic factors determined to be abnormal by the abnormality determining means as an abnormality degree; and The reconstructed data generating means for reversely converting the space data of the feature factor into the same space as the vehicle data only by the feature factor determined to be normal, and generating the reconstructed data, the evaluation data, It is a program for functioning as an abnormality diagnosing device comprising abnormality factor analysis means for performing abnormality factor analysis based on a difference from the reconstructed data. By installing the second invention in a computer, the abnormality diagnosis apparatus of the first invention can be obtained.

  According to the present invention, it is possible to provide an abnormality diagnosis apparatus or the like having sufficient performance in detecting abnormality and factor analysis ability in abnormality diagnosis of an in-vehicle system.

Diagram showing an overview of the abnormality diagnosis system Hardware configuration diagram of abnormality diagnosis device Flow chart showing the flow of normal data learning processing Flow chart showing the flow of evaluation data diagnosis processing The figure explaining reconstruction data generation processing The figure explaining difference data generation processing The figure which shows the calculation result of the abnormality degree of the sample unit computed by the comparative example 1, the comparative example 2, and the Example Schematic diagram explaining abnormality determination processing in units of independent components with a set threshold value The figure which shows the calculation result of the number of abnormal factors calculated by the comparative example 1, the comparative example 2, and the Example The figure which shows the result of the factor analysis by the comparative example 1 The figure which shows the result of the factor analysis with execution example

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a diagram showing an outline of an abnormality diagnosis system. As shown in FIG. 1, the abnormality diagnosis system 1 includes an abnormality diagnosis device 2 and a vehicle 3. The vehicle 3 is equipped with an in-vehicle electronic control system 4 that is an abnormality diagnosis target. The in-vehicle electronic control system 4 is provided with a data logger device 5.

  The on-vehicle electronic control system 4 is a control system in which a plurality of execution paths operate in parallel. In the on-vehicle electronic control system 4, a plurality of ECUs operate in parallel, or a single ECU outputs values of a plurality of variables simultaneously.

  The data logger device 5 records operation data (vehicle data) of the vehicle 3. The data logger device 5 transmits vehicle data to the abnormality diagnosis device 2 by wireless or wired communication means.

  The abnormality diagnosis device 2 diagnoses an abnormality of the in-vehicle electronic control system 4 based on the vehicle data transmitted from the data logger device 5. The abnormality diagnosis device 2 executes normal data learning processing (see FIG. 3) and evaluation data diagnostic processing (see FIG. 4), as will be described later.

  Although only a single vehicle 3 is illustrated in FIG. 1, the abnormality diagnosis system 1 may include a plurality of vehicles 3. For example, the abnormality diagnosis device 2 may receive the plurality of vehicle data from the plurality of data logger devices 5 with the in-vehicle electronic control system 4 according to the same control program as the same evaluation target.

  Further, in FIG. 1, only a single abnormality diagnosis device 2 is illustrated, but normal data learning processing and evaluation data diagnosis processing may be executed by separate computers. That is, the abnormality diagnosis device 2 may be configured by a plurality of computers.

  FIG. 2 is a hardware configuration diagram of the abnormality diagnosis apparatus. Note that the hardware configuration in FIG. 2 is an example, and various configurations can be adopted depending on the application and purpose. As shown in FIG. 2, the abnormality diagnosis apparatus 2 includes a control unit 11, a storage unit 12, a media input / output unit 13, a communication control unit 14, an input unit 15, a display unit 16, a peripheral device I / F unit 17, and the like. Connection is made via a bus 18.

  The control unit 11 includes a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and the like. The CPU calls a program stored in the storage unit 12, ROM, recording medium, etc. to a work memory area on the RAM, executes it, controls the drive of each device connected via the bus 18, and the abnormality diagnosis device 2 The process to be described later is realized. The ROM is a non-volatile memory and permanently holds a computer boot program, a program such as BIOS, data, and the like. The RAM is a volatile memory, and temporarily stores programs, data, and the like loaded from the storage unit 12, ROM, recording medium, and the like, and includes a work area used by the control unit 11 for performing various processes.

  The storage unit 12 is an HDD (Hard Disk Drive) or the like, and stores a program executed by the control unit 11, data necessary for program execution, an OS (Operating System), and the like. As for the program, a control program corresponding to the OS and an application program for causing a computer to execute processing to be described later are stored. Each of these program codes is read by the control unit 11 as necessary, transferred to the RAM, read by the CPU, and executed as various means.

  The media input / output unit 13 (drive device) inputs / outputs data, for example, media such as a CD drive (-ROM, -R, -RW, etc.), DVD drive (-ROM, -R, -RW, etc.) Has input / output devices. The communication control unit 14 includes a communication control device, a communication port, and the like, and is a communication interface that mediates communication between a computer and a network, and performs communication control between other computers via the network. The network may be wired or wireless.

  The input unit 15 inputs data and includes, for example, a keyboard, a pointing device such as a mouse, and an input device such as a numeric keypad. An operation instruction, an operation instruction, data input, and the like can be performed on the computer via the input unit 15. The display unit 16 includes a display device such as a liquid crystal panel, and a logic circuit or the like (video adapter or the like) for realizing a video function of the computer in cooperation with the display device. The input unit 15 and the display unit 16 may be integrated like a touch panel display.

  The peripheral device I / F (Interface) unit 17 is a port for connecting a peripheral device to the computer, and the computer transmits and receives data to and from the peripheral device via the peripheral device I / F unit 17. The peripheral device I / F unit 17 is configured by USB (Universal Serial Bus), IEEE 1394, RS-232C, or the like, and usually includes a plurality of peripheral devices I / F. The connection form with the peripheral device may be wired or wireless. The bus 18 is a path that mediates transmission / reception of control signals, data signals, and the like between the devices.

  FIG. 3 is a flowchart showing the flow of normal data learning processing. As shown in FIG. 3, the control unit 11 of the abnormality diagnosis device 2 inputs vehicle data from the data logger device 4 and acquires normal data known to be normal among the vehicle data (step S <b> 1).

  The vehicle data is multi-series (also referred to as “multivariable” or “multivariate”) time-series data. Hereinafter, the number of vehicle data series is M (M is a natural number). The normal data is, for example, vehicle data that is determined to be normal in driving the vehicle 3 by the test driver before the vehicle 3 is put on the market.

  Next, the control unit 11 standardizes each series so that the normal data acquired in step S1 has an average of 0 and a variance of 1 (step S2). The control unit 11 stores the average and standard deviation values of normal data used for standardization in the storage unit 12 together with the standardized normal data obtained at this time.

  Next, the control unit 11 performs principal component analysis of normal data standardized in step S2 (step S3). Principal component analysis is a technique for converting an original observed value having a correlation between variables into a value called a principal component having no correlation using orthogonal rotation. The control unit 11 stores the conversion model (transformation expression matrix) into the principal component obtained at this time in the storage unit 12 as a principal component analysis model.

  The principal components extracted by principal component analysis represent the main correlation from which unnecessary noise has been removed, and are uncorrelated with each other. Assuming that the number of principal components is P (P is a natural number), the control unit 11 extracts P principal components satisfying P <M, for example, based on an extraction criterion based on eigenvalue values, cumulative contribution ratios, and the like of the correlation matrix. To do.

  Next, the control part 11 performs an overcomplete independent component analysis using the main component extracted in step S3 (step S4). The independent component extracted here is a feature factor indicating the characteristics of normal data. The control unit 11 stores the conversion model (transformation expression matrix) into independent components obtained at this time in the storage unit 12 as an overcomplete independent component analysis model.

  In general, two components being independent of each other means that there is no correlation between the two components, that is, when one value fluctuates, the other does not interlock according to a certain rule. Using this definition, in the independent component analysis in the present embodiment, the independent components of K (K is a natural number) sequences to be extracted are converted into linear components (weighted linear sums) of P principal components. Express.

  However, in an actual problem, it is difficult to extract an independent component that completely satisfies the independent definition. Therefore, the control unit 11 performs optimization calculation, and calculates the linear transformation of the principal components of P sequences that satisfy the independent definition as much as possible by iterative calculation. Since it is an exploratory method based on iterative calculation, the independent components finally obtained are not strictly independent. Therefore, in the independent component analysis in the present embodiment, the control unit 11 calculates a component that is approximately close to being independent as an independent component.

  Also, overcomplete means that the number of independent components is larger than the original number of sequences, that is, K> P is satisfied, unlike ordinary independent component analysis. In particular, in this embodiment, it is desirable to extract K independent components that satisfy K> M. For example, the control unit 11 extracts independent components such that the number of independent components (= K) is several hundreds with respect to several tens of time series (= M) of vehicle data. As a result, the expressive power of the monitoring model based on independent components described later is improved, and the performance of abnormality detection capability and factor analysis capability is improved.

  The merit of overcompleteness is its expressive ability. In the present embodiment, the independent component is calculated as a weighted linear sum with respect to the input sequence. In other words, this means that the input sequence is represented by a weighted linear sum of independent components. Now, considering that an input sequence is represented by a weighted linear sum of independent components, if the number of independent components is small, the expression capability is limited. This is particularly problematic when the input sequence has a complicated relationship. Therefore, in the present embodiment, many independent components are extracted by overcomplete, and the expression ability is improved. Since each independent component imposes a constraint of approximate independence, it has information that is almost independent of each other, and therefore, has various local information that the input sequence has separately. In other words, when expressing an instantaneous local state of an input sequence, it can be expressed using only a few independent components corresponding to that state, and for complex input sequences with many states. By overcompleted, it can be accurately expressed with many independent components.

  The control unit 11 may extract feature factors by a method different from the principal component analysis and the independent component analysis, but it is desirable to extract K feature factors that satisfy K> M.

  Next, the control part 11 learns the data distribution of an independent component (feature factor), and produces | generates a monitoring model (step S5). The control unit 11 learns the data distribution for each independent component and generates a monitoring model for each independent component (feature factor). The control unit 11 stores the monitoring model for each independent component (feature factor) obtained at this time in the storage unit 12.

  For example, the control unit 11 learns the data distribution of independent components (feature factors) by kernel density estimation, and generates a monitoring model (probability density function). Kernel density estimation is one of methods for estimating the probability density function of a random variable, and can estimate the entire distribution from a finite sample.

  Next, the control unit 11 calculates the likelihood of each independent component (feature factor) obtained in step S4 using the monitoring model generated in step S5 (step S6). Here, the likelihood of each independent component (feature factor) is a value calculated by the probability density on the independent component (feature factor) of each data by the monitoring model, and how likely the data is from the model It is a value indicating whether or not it conforms. It can be considered that the lower the likelihood, the greater the degree of abnormality.

  Next, the control unit 11 sets the minimum value of the likelihood of each independent component (feature factor) calculated in step S6 as a determination threshold value for each independent component (feature factor) (step S7).

  FIG. 4 is a flowchart showing the flow of the evaluation data diagnosis process. As illustrated in FIG. 4, the control unit 11 of the abnormality diagnosis device 2 inputs vehicle data from the data logger device 4 and acquires evaluation data to be evaluated for abnormality diagnosis (step S <b> 11). The evaluation data is, for example, vehicle data that is unknown as to whether the user is operating normally after the vehicle 3 is put on the market.

  Next, the control unit 11 standardizes each series of evaluation data acquired in step S11 using the average and standard deviation values of normal data used for standardization in step S2 (step S12).

  Next, the control unit 11 converts the evaluation data standardized in step S12 into principal components using the principal component analysis model obtained in step S3 (step S13).

  Next, based on the overcomplete independent component analysis model obtained in step S4, the control unit 11 extracts the linear transformation of the component transformed in step S13 as an independent component (feature factor) (step S14). In the normal data learning process, when the feature factor is extracted by a method different from the principal component analysis and the independent component analysis, the control unit 11 also extracts the feature factor by the same method in step S13 and step S14.

  Next, the control unit 11 calculates the likelihood of each independent component (feature factor) obtained in step S14 using the monitoring model generated in step S5 (step S15).

  Next, the control unit 11 performs abnormality determination based on the number of independent components (feature factors) in which the likelihood of the independent components (feature factors) calculated in step S15 is smaller than the determination threshold set in step S7. This is performed (step S16). For example, the control unit 11 displays the number of independent components (feature factors) whose likelihood of independent components (feature factors) is smaller than the determination threshold on the display unit 16 as the degree of abnormality, or the degree of abnormality exceeds a predetermined value. It is determined that there is an abnormality. The control unit 11 may calculate the degree of abnormality for each sample of the evaluation data, or may calculate the degree of abnormality for each sample of the evaluation data for each independent component.

  Next, the control part 11 produces | generates reconstruction data using only the independent component (feature factor) whose likelihood is more than a determination threshold value (step S17). Specifically, the control unit 11 uses the inverse transformation of the principal component analysis model and the inverse transformation of the overcomplete independent component analysis model with only the independent component having the likelihood equal to or greater than the determination threshold to determine the independent component (feature factor). The space data is inversely transformed into the same space as the vehicle data, and reconstructed data is generated.

  FIG. 5 is a diagram for explaining the reconstruction data generation process. X (low dimension) for the spatial representation matrix of the original data (vehicle data), W for the transformation matrix by the principal component analysis model and the overcomplete independent component analysis model, and Y (high dimension) for the spatial matrix of the independent component And At this time, Y = WX holds. The matrix W is obtained by W = BA, where A is the principal component analysis model expression matrix and B is the overcomplete independent component analysis model expression matrix.

In step S17, the control unit 11 uses the data Y ′ in which the value of each independent component having a likelihood equal to or greater than the determination threshold is used for the data Y of the independent component space, and is the same as the original data (vehicle data). Reconstruction into space (representation matrix of reconstruction as W ′) generates reconstructed data X ′. At this time, X ′ = W′Y ′ holds. As for the reconstruction matrix W ′, W ′ = A ^T B ′, where B ′ is a pseudo-inverse matrix for the expression matrix B of the overcomplete independent component analysis model and ^AT is a transpose matrix for the expression matrix A of the principal component analysis model. Is required.

  Returning to the description of FIG. Next, the control unit 11 calculates a difference between the evaluation data standardized in step S12 and the reconstruction data generated in step S17, and generates difference data (step S18). In the present embodiment, it is considered that the contribution degree to the abnormality (hereinafter abbreviated as “contribution degree”) is higher in descending order of the value of the difference data. A series with a higher contribution is interpreted as having a stronger relation to the true cause of the abnormality. For example, the control unit 11 displays a list of series having a high contribution degree on the display unit 16.

  FIG. 6 is a diagram illustrating the difference data generation process. FIG. 6A shows the relationship between reconstructed data and original data (evaluation data standardized in step S12). For simplification, the number of series of original data is 2, the space of the original data is two-dimensional, and the space spanned by normal independent components is indicated by a straight line (one-dimensional). The reconstruction data is located on a space spanned by normal independent components. On the other hand, the original data is located at a location deviated from the reconstructed data (normal value).

  FIG. 6B shows an example of the contribution of each series to the reconstruction error. In the example of FIG. 6B, the contributions of five sequences are shown in descending order of contribution. The reconstruction error is, for example, the absolute value of the difference between the reconstruction data and the original data. In this case, the control unit 11 sets a value obtained by dividing the absolute value of the difference for each series by the total value of the absolute values of the differences for all series as the contribution degree for each series. The reconstruction error may be the sum of squares of the difference between the reconstruction data and the original data. In this case, the control unit 11 sets a value obtained by dividing the sum of squares of the differences for each series by the sum of the sums of the squares of the differences for all the series as the contribution degree for each series.

  Returning to the description of FIG. Next, the control unit 11 performs an abnormality factor analysis (step S19). Examples of the factor analysis method include the following two methods.

  First, the first factor analysis method will be described. The control unit 11 sets, for each series of evaluation data samples, a set of the series having the first largest contribution and the series having the second largest obtained in step S18 as a “high contribution series set”, respectively. The total number of occurrences in all samples is counted. Then, the control unit 11 displays the “high contribution series group” on the display unit 16 in descending order of the number of appearances.

  In the first factor analysis method, a set of sequences that frequently deviate from normal values is extracted throughout the evaluation data. This makes it possible to capture the overall abnormality.

  Next, the second factor analysis method will be described. For each series of evaluation data samples, the control unit 11 sets a series of the first largest contribution and the second largest series obtained in step S18 as a “high contribution series", and a high contribution series. Calculate the total reconstruction error of the set. The total reconstruction error may be the sum of absolute values of differences or the sum of square sums of differences. Then, the control unit 11 displays on the display unit 16 the sample having the maximum total reconstruction error of the high contribution series set and the high contribution series set.

  In the second factor analysis method, a set of samples and sequences that are most greatly shifted is extracted through the entire evaluation data. As a result, a sudden abnormality can be captured.

  Hereinafter, examples of the abnormality diagnosis apparatus 2 according to the present embodiment will be described. In this embodiment, vehicle data under two driving conditions of “normal driving” and “slow driving” are used. Further, with “normal travel” as normal data and “slow travel” as evaluation data, the abnormality diagnosis apparatus 2 executes the normal data learning process shown in FIG. 3 and the evaluation data diagnosis process shown in FIG. And it analyzed about how much the change of the driving conditions in evaluation data was detectable with the learned monitoring model.

  On the other hand, in Comparative Example 1, “normal travel” is normal data and “slow travel” is evaluation data, and normal data learning processing and evaluation data diagnosis processing based only on principal component analysis are executed. In Comparative Example 2, “normal driving” is normal data, “slow driving” is evaluation data, and normal data learning processing based only on independent component analysis (however, the number of independent components K ≦ the number M of vehicle data sequences) Evaluation data diagnosis processing was executed.

  FIG. 7 is a diagram illustrating a calculation result of the total abnormality degree of the sample unit calculated by the comparative example 1, the comparative example 2, and the example. The total degree of abnormality means not the degree of abnormality for each independent component but the degree of abnormality taking all independent components into consideration. The overall abnormality degree shown in FIG. 7 is a product of the likelihood of each independent component and is expressed by a negative logarithm (negative log likelihood). FIG. 7A shows a comparative example 1, FIG. 7B shows a comparative example 2, and FIG. 7C shows an example. In either case, the horizontal axis represents the sample identification number, the vertical axis represents the overall abnormality degree, the left side of the normal data and evaluation data dividing line is the normal data sample group, and the right side is the evaluation data sample group. The maximum value of the scale on the vertical axis is “60” in FIG. 7A, “30” in FIG. 7B, and “12” in FIG. 7C.

  Referring to FIGS. 7A and 7B, in Comparative Example 1 and Comparative Example 2, the difference between the running conditions of the normal data and the evaluation data is not clearly detected by the value of the overall abnormality level. Originally, the overall abnormality degree of the sample of the evaluation data should be high, but there are also some samples of normal data whose overall abnormality degree is higher than that of the evaluation data sample. That is, in Comparative Example 1 and Comparative Example 2, it is shown that there are many false detections and detection omissions, and the performance of abnormality detection capability is low.

  On the other hand, referring to FIG. 7C, in the example, the difference in the running condition between the normal data and the evaluation data is clearly detected by the value of the overall abnormality degree. In the normal data, there are almost no samples having an overall abnormality degree of “5” or more, whereas in the evaluation data, there are many samples having an overall abnormality degree of “5” or more. This indicates that the monitoring model of the embodiment learned based on the normal data (“normal driving”) has detected many differences from the evaluation data with different driving conditions (“gentle driving”). That is, the overall abnormality degree calculated by the example indicates that there are few false detections and detection omissions and the performance of the abnormality detection capability is high.

  FIG. 8 is a schematic diagram for explaining the abnormality determination process in units of independent components based on the set threshold value. The horizontal axis represents each independent component, and the vertical axis represents the degree of abnormality (negative log likelihood) of each independent component. In this example, five independent components exceeded the abnormality level threshold. The abnormality degree threshold is set based on normal data. As shown in FIG. 8, in the data determined to be abnormal, many independent components behave abnormally. Note that the abnormality threshold shown in FIG. 8 is the same value for all the independent components, but actually differs for each independent component.

  FIG. 9 is a diagram illustrating a calculation result of the number of abnormal factors calculated by Comparative Example 1, Comparative Example 2, and Example. The abnormal factor means a characteristic factor (independent component) whose abnormality exceeds the abnormality threshold. In this case, the degree of abnormality is calculated in units of feature factors (independent components). 9A is Comparative Example 1, FIG. 9B is Comparative Example 2, and FIG. 9C is an example. In both cases, the horizontal axis represents the sample identification number, the vertical axis represents the number of abnormal factors, the left side of the normal data and evaluation data dividing line is the normal data sample group, and the right side is the evaluation data sample group. The maximum value of the scale on the vertical axis is “2” in FIG. 9A, “8” in FIG. 9B, and “120” in FIG. 9C.

  Referring to FIGS. 9A and 9B, in Comparative Example 1 and Comparative Example 2, the difference between the running conditions of normal data and evaluation data is not clearly detected by the number of abnormal factors. That is, in Comparative Example 1 and Comparative Example 2, it is shown that there are many false detections and detection omissions, and the performance of abnormality detection capability is low.

  On the other hand, referring to FIG. 9C, in the example, the difference in running conditions between normal data and evaluation data is clearly detected by the number of abnormal factors. In normal data, there are almost no samples with an abnormal factor number of “30” or more, whereas in evaluation data, there are many samples with an abnormality degree of “30” or more. This indicates that the monitoring model of the embodiment learned based on the normal data (“normal driving”) has detected many differences from the evaluation data with different driving conditions (“gentle driving”). That is, the number of abnormal factors calculated according to the example indicates that there are few false detections and detection omissions and the performance of the abnormality detection capability is high.

  FIG. 10 is a diagram illustrating a result of factor analysis according to Comparative Example 1. FIG. 11 is a diagram illustrating a result of factor analysis according to the example. In the embodiment, the factor analysis is executed in step S19 in FIG. The examples shown in FIGS. 10 and 11 are analysis results obtained by the first factor analysis technique.

  Referring to FIG. 10, in the case of Comparative Example 1, there are many sets of sequences having a high number of appearances that have no correlation. On the other hand, referring to FIG. 11, in the case of the embodiment, there are many sets of sequences that have a large number of appearances. For example, in the scatter diagram “number of appearances 1st place” shown in FIG. 11, changes due to gentle running can be accurately expressed.

  When actually performing vehicle abnormality diagnosis, there is a need to observe how the correlation is broken in a set of correlated series. This is because the collapse of the correlation can cause abnormalities. Since the independent component represents a local feature (correlation) of the original data, finding an abnormal independent component means finding a collapse of the correlation. Therefore, in this embodiment, it can be said that an analysis result that satisfies the need to observe how the correlation is broken can be presented. On the other hand, in the principal component analysis, it is impossible to observe how the correlation is broken.

  Here, the effect by the abnormality diagnosis device 2 of the present embodiment and the reason for the effect will be described. The first effect is that a data series that directly contributes to the abnormality can be clarified. The second effect is that even if it is not possible to determine whether each data sample is normal, the factor analysis can be performed. A third effect is that abnormality detection that is resistant to noise can be performed by determining abnormality according to the degree of abnormality, with the number of abnormal feature factors being the degree of abnormality. The fourth effect is that by extracting more characteristic factors than the number of series of vehicle data, the performance of abnormality detection capability and factor analysis capability is improved.

  First, the reason why the first effect occurs will be described. When monitoring a characteristic factor instead of the original data series, even if it is determined that the characteristic factor is abnormal, the relationship with the original data series is complicated, so that it is difficult to perform the factor analysis with the conventional method. Therefore, in the abnormality diagnosis device 2 according to the present embodiment, inverse transformation is performed only with normal feature factors, reconstructed data is generated, and the original evaluation data and reconstructed data formed by the feature factors of abnormal behavior are calculated. Generate difference data and perform factor analysis using the difference data. As a result, the analysis can be performed on the original data series, not on the feature factor, and the factor analysis becomes easy. Further, only the directly contributed portion can be referred to in the form of difference data without being influenced by unrelated noise. Therefore, the data series that directly contributes to the abnormality can be clarified.

  Next, the reason why the second effect occurs will be described. In the conventional method, abnormality determination is performed in the abnormality detection process, and after detecting a small amount of abnormality candidates, factor analysis is performed only on those abnormality candidates. However, at present, it is difficult to correctly extract the abnormal candidates, and there are many false detections and omissions. In particular, this tendency becomes prominent in the early stage abnormality detection in which the abnormality is not clearly manifested. Therefore, in the abnormality diagnosis device 2 of the present embodiment, the factor analysis is performed without extremely limiting the data to be analyzed by abnormality determination. Thus, the diagnostician can designate a data range to be analyzed based on the degree of abnormality, prior knowledge, etc., and can perform factor analysis even if it is not possible to determine whether each data sample is normal.

  In each sample, paying attention to only the group with the first largest contribution and the group with the second largest contribution (high contribution series), and performing statistical processing in terms of the number of appearances and the magnitude of error A comprehensive factor analysis can be performed. The diagnostician can easily know a set of series that contributes to the error through the entire designated data or a set of series that has the largest error.

  Further, the extracted high contribution series set can be easily analyzed using a scatter diagram. Scatter charts are highly useful for diagnosticians who are not familiar with statistics because they are intuitive and easy to understand. When there is a set of series that you want to focus on, the diagnostician can analyze the detailed difference by plotting normal data and evaluation data for each set of series and confirming the difference, for example. .

  In addition, it is desirable to perform factor analysis using a statistical method under various conditions. For example, analysis accuracy can be improved by limiting the range or series of data to be analyzed based on the knowledge obtained from the first analysis result and performing factor analysis again.

  Next, the reason why the third effect occurs will be described. In the conventional method, any data whose feature factor exceeds a threshold value is often regarded as abnormal. On the other hand, in the abnormality diagnosis device 2 of the present embodiment, the number of feature factors determined to be abnormal behavior is set as the total abnormality degree, and when the abnormality degree value is large, it is determined that the vehicle is abnormal. A small amount of feature factors may be determined to be abnormal behavior due to fine noise, etc., but if the number of such feature factors is small, the overall abnormality level will not increase, so avoid false detection of abnormalities. This makes it possible to detect abnormalities that are resistant to noise.

  Next, the reason why the fourth effect occurs will be described. When the number of feature factors is small, the flexibility of expressing reconstructed data is low, and it is not possible to express the fine features inherent in the original data. Therefore, as in the present embodiment, when generating reconstructed data, as many feature factors as possible are extracted to improve the expressive power of the reconstructed data. In particular, by extracting feature factors using overcomplete independent component analysis, the feature factors are almost independent of each other and do not have the same features. Accordingly, as the number of extracted feature factors increases, a variety of features can be expressed, and the performance of abnormality detection capability and factor analysis capability improves.

  The preferred embodiments of the abnormality diagnosis device and the like according to the present invention have been described above with reference to the accompanying drawings, but the present invention is not limited to such examples. It will be apparent to those skilled in the art that various changes or modifications can be conceived within the scope of the technical idea disclosed in the present application, and these naturally belong to the technical scope of the present invention. Understood.

DESCRIPTION OF SYMBOLS 1 ......... Abnormality diagnosis system 2 ......... Abnormality diagnosis device 3 ......... Vehicle 4 ......... In-vehicle electronic control system 5 ......... Data logger device 11 ......... Control unit 12 ......... Storage unit 13 ......... Media input / output unit 14 ... Communication control unit 15 ... Input unit 16 Display unit 17 Peripheral device I / F unit 18 Bus

Claims (9)

An abnormality diagnosis device for diagnosing vehicle abnormality based on vehicle data that is multi-series time series data,
Normal data acquisition means for acquiring normal data known to be normal among the vehicle data;
A feature factor extracting means for extracting a plurality of feature factors from the normal data;
Model generation means for learning a data distribution of the characteristic factors and generating a monitoring model;
Of the vehicle data, evaluation data acquisition means for acquiring evaluation data to be evaluated for abnormality diagnosis;
Extracting the feature factor from the evaluation data, and determining abnormality based on the monitoring model;
Reconstruction data generation means for reversely converting the space data of the feature factor into the same space as the vehicle data only by the feature factor determined to be normal by the abnormality determination means,
An abnormal factor analysis means for performing an abnormal factor analysis based on a difference between the evaluation data and the reconstructed data;
An abnormality diagnosis apparatus comprising: When the number of series of vehicle data is M (M is a natural number),
The abnormality diagnosis apparatus according to claim 1, wherein the feature factor extraction unit extracts K feature factors that satisfy K (K is a natural number)> M. 3. The abnormality diagnosis apparatus according to claim 2, wherein the feature factor extraction means performs an overcomplete independent component analysis of the normal data, extracts K independent components, and uses them as the feature factors. The feature factor extraction means performs a principal component analysis of the normal data, extracts P principal components satisfying P (P is a natural number) <M, and expresses the independent component by a weighted linear sum of the principal components. The abnormality diagnosis device according to claim 3, wherein: The monitoring model generation means sets a determination threshold for each feature factor based on the monitoring model,
The abnormality determination means calculates the likelihood of the feature factor in the evaluation data based on the monitoring model, sets the number of the feature factor that is smaller than the determination threshold as the abnormality factor, and sets the abnormality as the abnormality The abnormality diagnosis apparatus according to claim 1, wherein abnormality determination is performed based on the degree. 6. The abnormality diagnosis apparatus according to claim 1, wherein the monitoring model generation unit learns a data distribution of the characteristic factor by kernel density estimation. The abnormality factor analysis means calculates a contribution degree to the abnormality for each feature factor based on a difference between the evaluation data for each feature factor and the reconstructed data, and a series having the largest contribution degree The second largest series set is extracted as a high contribution series set, the number of appearances of the high contribution series set in all samples of the evaluation data is totaled, and the high contribution series set is output in descending order of appearance counts. An abnormality diagnosis apparatus according to any one of claims 1 to 6, wherein the abnormality diagnosis apparatus comprises: The abnormality factor analysis means calculates a contribution degree to the abnormality for each feature factor based on a difference between the evaluation data for each feature factor and the reconstructed data, and a series having the largest contribution degree The second largest sequence set is extracted as a high contribution sequence set, and a sample of the evaluation data that maximizes the total reconstruction error of the high contribution sequence set is output. The abnormality diagnosis apparatus according to claim 6. A program for causing a computer to function as an abnormality diagnosis device for diagnosing an abnormality of a vehicle based on vehicle data that is multi-series time series data,
Normal data acquisition means for acquiring normal data known to be normal among the vehicle data;
A feature factor extracting means for extracting a plurality of feature factors from the normal data;
Model generation means for learning a data distribution of the characteristic factors and generating a monitoring model;
Of the vehicle data, evaluation data acquisition means for acquiring evaluation data to be evaluated for abnormality diagnosis;
Extracting the feature factor from the evaluation data, and determining abnormality based on the monitoring model;
An abnormality degree calculating means for calculating the number of the feature factors determined to be abnormal by the abnormality determining means as an abnormality degree;
Reconstruction data generation means for reversely converting the space data of the feature factor into the same space as the vehicle data only by the feature factor determined to be normal by the abnormality determination means,
An abnormal factor analysis means for performing an abnormal factor analysis based on a difference between the evaluation data and the reconstructed data;
A program for causing an abnormality diagnosis apparatus to function.

Images