Method for configuring vehicle breakdown symptom detection device

Abstract

To appropriately adjust time interval of vehicle travel data to be referred to when detecting a breakdown symptom in order to improve detection accuracy in a configuration to detect the breakdown symptom in the vehicle travel data.SOLUTION: A vehicle breakdown symptom detection device is configured to use a discriminator to detect presence or absence of a symptom of a breakdown in any component of a vehicle. In the vehicle breakdown symptom detection device, the discriminator acquires a feature, which produces a significant difference between a value obtained from normal travel data and a value obtained from pre-breakdown travel data in a similar way to the normal travel data, from instantaneous travel data based on an algorithm of machine learning using the preliminarily prepared normal travel data and the pre-breakdown travel data before the breakdown occurs in any component as training data. The discriminator is also configured to detect the symptom of the breakdown when the difference between the feature of the instantaneous travel data and the feature of the normal travel data exceeds a predetermined threshold. Time intervals of the normal travel data and the pre-breakdown travel data used for calculating the features are adjusted so that identification accuracy of the discriminator exceeds a predetermined value.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present invention relates to a device for detecting signs of a malfunction in a vehicle such as an automobile, and more specifically, to a method for configuring a device for detecting when there is behavior (signs of a malfunction) that may lead to a malfunction in various driving data while the vehicle is running. [Background technology]

[0002] In order to detect failures in various devices, various configurations have been proposed in which a failure detection model is constructed using data prepared in advance for normal times and data prepared in abnormal times for a part of the device to be monitored, and such a failure detection model is used to detect abnormalities in data for the part to be monitored while the device is in operation or working. For example, Patent Document 1 proposes a configuration in which a distribution (normal time distribution) of detection value vectors obtained in normal times by a plurality of sensors in a part to be monitored of the device is obtained, and the presence or absence of an abnormality and a sensor having an abnormality are estimated from the deviation of the detection value vectors of the plurality of sensors obtained during the operation of the device from the normal time distribution. Patent Document 2 discloses a configuration in which the minimum fuel pressure in a fuel pipe in a fuel supply system of a vehicle or the like and data indicating the state when the minimum fuel pressure was stored are used to determine the location of a failure related to a drop in fuel pressure in the fuel pipe according to a machine learning algorithm such as a decision tree, and a system abnormality is diagnosed. Patent Document 3 proposes a configuration in which parameters such as total acid number are calculated from sensor values ​​such as the relative dielectric constant and electrical conductivity of the hydraulic oil of the device, and a symptom of failure of the device and parameters of its cause are estimated from the parameters using a decision tree structure algorithm. Patent Document 4 discloses a configuration for identifying the cause of a fault and a faulty part by using a knowledge base that stores knowledge information such as symptoms required for diagnosing faults in an automobile and artificial intelligence. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2020-98373 [Patent Document 2] Patent Publication No. 2021-195871 [Patent Document 3] Patent Publication No. 2021-167791 [Patent Document 4] JP 3-79445 Summary of the Invention [Problem to be solved by the invention]

[0004] However, when a failure occurs in any part of a vehicle, behavior leading to the failure, i.e., a failure sign, appears in various driving data of the vehicle at a stage some time before the failure occurs. Therefore, if the driving data of the vehicle is monitored and a failure sign is detected, the driver or the like can be notified of the failure. It is expected that the failure can be avoided by improving the driving behavior and inspecting the part where the failure sign appears. The failure sign in the driving data of the vehicle can be detected by a classifier configured using a machine learning algorithm. In this regard, since the driving data generally fluctuates finely, it is preferable to refer to driving data having a certain time width in order to accurately detect the failure sign from the driving data. However, since it is expected that the behavior leading to the failure will decrease the further back in time from the time when the failure occurs, even if the time width of the referenced driving data is too long, the accuracy of detecting the failure sign will decrease.

[0005] In consideration of the above circumstances, the main object of the present invention is to appropriately adjust the time width of the driving data referenced for detecting signs of failure in a configuration that detects behavior (signs of failure) that may lead to a failure in any part of a vehicle's driving data, in order to improve the accuracy of detecting signs of failure. [Means for solving the problem]

[0006] According to the present invention, the above problem is solved by a failure sign detection device that receives input of real-time driving data, which is driving data representing the state of a vehicle, and detects the presence or absence of a failure sign in each part of the vehicle, and a classifier configured according to a machine learning algorithm that uses, as teacher data, normal driving data, which is driving data of a vehicle in a normal state that has been prepared in advance, and pre-failure driving data, which is driving data of a vehicle in which a failure has occurred in any part before the failure occurred, and obtains from the real-time driving data a feature amount in which the magnitude of the difference between a value obtained from the normal driving data and a value obtained similarly from the pre-failure driving data is significant, and a method for detecting the presence or absence of a failure sign in any part of the vehicle using the classifier that identifies that there is a failure sign when the magnitude of the difference between the feature amount of the real-time driving data and the feature amount of the normal driving data exceeds a predetermined magnitude, A step of adjusting a time width of the normal driving data and the pre-fault driving data used in calculating the feature amount so that the discrimination accuracy of the discriminator exceeds a predetermined value. This is achieved by a method comprising:

[0007] In the above configuration, the "driving data representing the state of the vehicle" may be any data representing the state of each part of the vehicle, and specifically, for example, vehicle speed, accelerator opening or required driving force, shift lever position, transmission oil temperature, engine water temperature, motor temperature, mode selection switch position, output torque (MG1, MG2), etc., which are state quantities related to driving and driving detected in each part of the vehicle, and may be time-series data (values ​​from moment to moment). The "real-time driving data" is driving data obtained in real-time while the vehicle is being driven (it may be time-series data over a certain time span). The "each part of the vehicle" may be various parts of the vehicle, such as the engine / motor or each part thereof, the power transmission mechanism or transmission or each part thereof, the braking system or each part thereof, and each part of other on-board devices, and one or more parts may be selected as appropriate. As already mentioned, the "failure sign" is the behavior (fluctuation, distribution) of values ​​that lead to a failure that appear in the driving data. The "normal driving data, which is driving data of a vehicle in a normal state prepared in advance, and the pre-failure driving data, which is driving data of a vehicle with a failure in any part before the failure occurs" may be data collected in advance by any method, for example, by experiments, simulations, etc. Furthermore, the "normal driving data" and the "pre-failure driving data" may be collected from vehicles in a state in which conditions such as the period of use or the mileage are approximately the same. The "machine learning algorithm" used in the "classifier" of the "failure sign detection means" may be any algorithm that can configure a classifier that outputs a normal or fault label corresponding to the normal driving data and the pre-failure driving data when the normal driving data and the pre-failure driving data are input, using teacher data in which the feature values ​​obtained from the normal driving data and the pre-failure driving data are explanatory variables and the corresponding normal label or fault label of any part of the vehicle is the objective variable. The "feature value" is a value calculated or determined in various ways from the driving data, and the calculation or determination method may be appropriately configured or selected through experiments, etc. The "classification accuracy of the classifier" is the rate at which a failure sign is detected when the driving data input to the classifier is pre-failure driving data, i.e., the rate at which the input driving data is correctly detected as pre-failure driving data. The "predetermined value" for the classification accuracy may be set as appropriate.The failure sign detection device is realized by a computer device operating in accordance with a program.

[0008] In the above-described configuration of the device of the present invention, a classifier configured according to the above-described machine learning algorithm is used to detect the presence or absence of behavior that may lead to a breakdown, i.e., a breakdown sign, in the vehicle's moment-to-moment running data. At this time, the time width of the normal running data and pre-failure running data used to calculate the feature amount for the breakdown sign detection is adjusted so that the classification accuracy of the classifier exceeds a predetermined value. This allows the feature amount to be calculated using running data with a time width that is just right for accurately detecting the breakdown sign, and is expected to improve the accuracy of detecting the breakdown sign.

[0009] In one of the above-described embodiments of the present invention, the adjustment of the time width of the traveling data may be performed by gradually increasing the time width while repeatedly executing the detection of the sign of failure by the classifier and calculating the classification accuracy using the normal traveling data and the pre-failure traveling data for verification until the classification accuracy exceeds a predetermined value that may be set appropriately. Alternatively, the time width may be gradually changed while executing the detection of the sign of failure by the classifier and calculating the classification accuracy to search for the time width at which the classification accuracy is maximized.

[0010] In the above-mentioned failure sign detection device, the classifier may be provided for detecting a failure sign in each of a plurality of parts in the vehicle. In this case, the time width of the traveling data may be adjusted for each classifier of each part. Effect of the Invention

[0011] Thus, according to the failure sign detection device to which the method of the present invention is applied, when a failure sign is detected in the vehicle's driving data, the driver or the like is notified of the fact, and this makes it possible to extend the usable period of the vehicle as much as possible by taking various measures to avoid failures even when each part of the vehicle is not actually broken.In this case, according to the method of the present invention, the time width of the driving data monitored to detect the failure sign is adjusted so that the discrimination accuracy of the discriminator, i.e., the detection accuracy of the failure sign, is sufficiently high or optimal, making it possible to more appropriately detect the failure sign.

[0012] Other objects and advantages of the present invention will become apparent from the following description of preferred embodiments of the invention. [Brief description of the drawings]

[0013] [Figure 1] FIG. 1 is a diagram illustrating the outline of the configuration of a vehicle failure sign detection device to which the method of this embodiment is applied. [Diagram 2] Fig. 2(A) shows a schematic example of driving data (time series data) monitored by a vehicle failure sign detection device to which the method of this embodiment is applied. Fig. 2(B) and (C) show a schematic frequency distribution of driving data (accelerator opening, oil temperature) monitored by a vehicle failure sign detection device to which the method of this embodiment is applied. In the figure, the distribution during normal operation and before failure are shown by dotted lines and solid lines, respectively. [Diagram 3] Fig. 3(A) is a diagram showing a typical distribution in a feature space of feature vectors of driving data monitored by a vehicle failure sign detection device to which the method of this embodiment is applied. Fig. 3(B) is a diagram showing a typical change over time in a damage rate that indicates the degree of deterioration of one part of a vehicle. "Normal" is a case where deterioration progresses in a normal state in a vehicle. [Figure 4] FIG. 4 shows, in the form of a flowchart, the process of creating a failure sign detection model used in a vehicle failure sign detection device to which the method of this embodiment is applied. [Diagram 5]FIG. 5 shows, in the form of a flowchart, the process of detecting a sign of a vehicle failure in a vehicle failure sign detection device to which the method of this embodiment is applied. [Explanation of symbols]

[0014] 10. Vehicle BEST MODE FOR CARRYING OUT THEINVENTION

[0015] The present invention will now be described in detail with reference to some preferred embodiments thereof with reference to the accompanying drawings, in which like reference numerals indicate like parts.

[0016] Device configuration Briefly, as shown in Fig. 1, the failure sign detection device to which the method of the present embodiment is applied is configured to acquire data (travel data) representing the state of each part of the vehicle 10 during driving of an arbitrary vehicle 10, compare the travel data with travel data (normal travel data) obtained (in advance) of a normal vehicle in a normal, i.e., normal or general state without any failure, collected from a network or other database, and, when there is a significant difference between the two, determine that there is a failure sign, that is, that any part of the vehicle 10 is in a state that may lead to a failure (failure sign detection), and notify the driver, etc. of the same. Therefore, the failure sign detection device of the present embodiment is basically configured with a means for acquiring various travel data (real-time travel data) from each of a group of sensors provided in each part of the vehicle 10, a failure sign detection unit having a discriminator that receives the real-time travel data and detects a failure sign in any part of the vehicle 10, and a failure sign alarm that notifies the failure sign in any part detected by the failure sign detection unit. Here, the failure sign detection unit may be provided with a classifier that detects failure signs for each part of the vehicle that is to be monitored for failure signs (i.e., a plurality of classifiers may be configured). The failure sign detection unit is also provided with a means for collecting normal driving data from a database or the like for comparison with real-time driving data in a manner described in detail below. The failure sign alarm may be configured using any display or speaker that can visually or audibly detect a failure sign and notify the part where it is detected. The operation of the failure sign detection unit and the failure sign alarm is realized by operation in accordance with a program by a computer device.

[0017] Operation of the device (1) Configuration of failure prediction detection In an apparatus to which the method of this embodiment is applied, real-time driving data measured in each part of the target vehicle (hereinafter referred to as the "host vehicle"), for example, time series data of vehicle speed, accelerator opening or required driving force, shift lever position, transmission oil temperature, engine water temperature, motor temperature, mode selection switch position, and output torque (MG1, MG2), are referenced, and if the behavior of the data or the data is significantly different from that of the corresponding (prepared in advance) normal driving data and may be linked to a malfunction, a malfunction is detected as being present.

[0018] More specifically, for example, as shown in FIG. 2(A), various driving data (in the illustrated example, the accelerator opening, vehicle speed, and oil temperature are included, but are not limited to these) change from moment to moment, and when looking at the frequency distribution of the values ​​of these data as shown in FIGS. 2(B) and (C), when any part of the vehicle (for example, the drive system device) is in a state where a failure may occur, the frequency distribution of the values ​​of the data (solid line) at the stage before the failure occurs may differ from that in the case of normal driving conditions (dotted line). Therefore, in this embodiment, a feature (which may be a vector having multiple values) is calculated or extracted appropriately from driving data (pre-failure driving data) at a stage before the failure over an appropriate time span Tp and normal driving data, and a feature that shows a significant difference or deviation between the pre-failure driving data and the normal driving data is calculated or extracted from the real-time driving data of the vehicle, and the degree of deviation of the feature of the real-time driving data from the feature of the normal driving data is evaluated.If the degree reaches a predetermined level that is set appropriately, it is determined that there is a sign of a failure.

[0019] The evaluation of the degree of deviation of the feature of the real-time driving data from the feature of the normal driving data may be performed in various ways. As one example, as shown in FIG. 3(A), a distribution of the feature of the normal driving data (normal distribution) is defined in the feature space of the driving data, and when the distance of the feature of the real-time driving data from the normal distribution exceeds a predetermined distance, more specifically, when the plot of the feature of the real-time driving data is so far away from the normal distribution that it is included in the distribution of the feature of the (pre-fault) pre-fault driving data (prepared in advance), it may be determined that a failure sign appears in the real-time driving data.

[0020] Generally, each part of a vehicle deteriorates with the increase in the period of use or the distance traveled, and the degree of deterioration is reflected in the driving data. When the degree of deterioration is expressed as a "damage rate" which is the ratio of the degree of deterioration until the part is deteriorated to the endurance limit (a state in which replacement or repair is required), as shown in Fig. 3(B) as a schematic diagram, when the vehicle is used in a normal manner, the damage rate progresses over time to reach S at time Tov, as shown by the dashed line (normal). On the other hand, if the damage rate Sh of the host vehicle at a certain time Tc is significantly higher than the normal damage rate St at the same time (if ΔS is significant), it can be determined that the part of the host vehicle is deteriorated earlier than normal and is at risk of breaking down. Therefore, in the detection of signs of failure in this embodiment, normal driving data and pre-failure driving data are prepared for each period of use or mileage of the vehicle, or for each section thereof, and feature values ​​obtained from the normal driving data and pre-failure driving data that show a significant difference in value are selected.In detecting signs of failure in the vehicle, the distance of the feature value distribution of the normal driving data from the state (point) of the real-time driving data for the selected feature values ​​for each period of use or mileage of the vehicle, or for each section thereof, may be evaluated as described above.

[0021] As already described, the failure sign detection unit of the device to which the method of the present embodiment is applied may be configured with a classifier for detecting a failure sign as described above for each part. In an embodiment, when the difference between the feature amount of the real-time driving data and the feature amount of the normal driving data for the feature amount having the largest difference from the feature amount of the corresponding normal driving data, the feature amount having the largest rate of change, or the feature amount having a rate of change exceeding a predetermined value among the feature amounts acquired for each of the multiple parts from the real-time driving data exceeds a predetermined value, it may be detected that there is a failure sign in the part corresponding to that feature amount. Also, the damage rate or remaining life representing the degree of deterioration of each part of the vehicle may be estimated, and when the difference between the feature amount of the real-time driving data and the feature amount of the normal driving data for the feature amount having the largest damage rate or the smallest remaining life exceeds a predetermined value, it may be detected that there is a failure sign in the part corresponding to that feature amount.

[0022] In the configuration for detecting a fault sign using features calculated from driving data as described above, since the driving data is prone to fluctuation as already described, it is preferable that the normal driving data and pre-fault driving data used to calculate the features span a certain time span Tp (see FIG. 2(A)), but going back too far from the occurrence of the fault reduces the contribution of the fault sign in the driving data. Therefore, in this embodiment, as exemplified below, in the machine learning processing stage for constructing a fault sign detection model, the time span of the normal driving data and pre-fault driving data used to calculate the features is adjusted so that the classification accuracy of the classifier is sufficiently high or optimal.

[0023] (2) Machine learning processing (construction of a classifier) ​​that constitutes a failure sign detection model In this embodiment, the machine learning process for constructing the failure sign detection model may be executed, for example, according to the processing procedure illustrated in Fig. 4. Note that the machine learning process may be executed for each part of the vehicle for which it is desired to detect a failure sign, and a classifier for each part may be prepared.

[0024] Referring to the figure, first, the running data of a vehicle in which a failure has occurred (failure running data) is collected by any method (preferably together with information on the period of use or mileage of each vehicle) (step 1), and data before the occurrence of a failure (pre-failure running data) is extracted from the collected data (step 2). The failure running data may be collected by an experiment or the like, or may be collected through a communication network from a database in which such data is accumulated. The failure running data is preferably collected from a vehicle of the same type as the vehicle itself. The time width Tp of the pre-failure running data is initially set to an initial value Tpi that is appropriately set. In addition, normal running data is also collected by any method, similar to the failure running data (step 3). The normal running data may be data obtained from a vehicle that is in a state similar to the period of use or mileage of each vehicle in the pre-failure running data and is not broken down. A part of the pre-failure running data and the normal running data collected up to this point are used to configure a failure sign detection model, and the other part is used to verify the configured model. Here, the pre-fault driving data and the normal driving data are explanatory variables in the classifier, and the state of each driving data, whether it is normal or indicates a fault symptom, is the objective variable.

[0025] Next, various feature quantities, which may be vectors, are calculated from the normal driving data and the pre-failure driving data (step 4), and the distance between the distribution of the feature quantities of the normal driving data and the distribution of the feature quantities of the pre-failure driving data in the feature quantity space may be calculated (step 5). The distance between the distributions may be calculated using a statistical value of the distribution, such as the average value, the median value, the variance value, the standard deviation value, etc. In addition, the processing of steps 4 and 5 may calculate the distance between the distributions and the degree of separation of the distributions while automatically preparing a combination of various feature quantities using a computer device. In this regard, a combination of various feature quantities may be prepared for each state, such as the driving mode, the period of use, or the mileage of the vehicle, and the distance between the distributions and the degree of separation of the distributions may be calculated. Thereafter, a feature quantity that separates the group of normal driving data and the group of pre-failure driving data as clearly as possible is selected with reference to the calculated distance between the distributions and the degree of separation of the distributions for each state, such as the driving mode, the period of use, or the mileage of the vehicle (step 6). Such a selection process may also be automatically performed using a computer device.

[0026] Thereafter, in order to verify whether a fault sign can be detected using the feature quantities selected by the above process, the feature quantities selected in the above step 6 are calculated from the pre-fault driving data and normal driving data for verification, and then the distance of the state points represented by these feature quantities in the feature quantity space from the distribution of the normal driving data selected in the above step 6 is calculated, and it is confirmed whether each point of the feature quantities of the normal driving data and the pre-fault driving data for verification is included in the distribution of the normal driving data (step 6) and the distribution of the pre-fault driving data, respectively (step 7). Then, if the proportion of cases in which the data for verification is correctly determined (estimation accuracy) exceeds a predetermined value (A%) that may be set appropriately, a model that uses the feature quantities selected in step 6 to discriminate the driving data is adopted as a fault sign detection model (steps 8 and 9). By the above-mentioned process of the machine learning algorithm, a classifier is configured that returns a result that the normal driving data is normal and returns a result that a fault sign is present for the pre-fault driving data. On the other hand, if the estimation accuracy does not meet the predetermined value (steps 8 and 10), the time width Tp of the driving data used to obtain the feature is changed to Tp+ΔT, and the above steps 1 to 8 are repeated. That is, the machine learning process is repeated while increasing the time width Tp of the driving data used to obtain the feature until the estimation accuracy exceeds the predetermined value, thereby forming a classifier that exhibits sufficient estimation accuracy.

[0027] As another aspect, the above steps 1 to 8 may be repeatedly executed while changing the time width Tp of the travel data, to search for the time width Tp that maximizes the estimation accuracy.

[0028] (3) Failure prediction detection process The process of detecting a fault sign in the vehicle in the device to which the present embodiment is applied may be executed, for example, according to the procedure illustrated in FIG. 5. More specifically, while the vehicle is being driven, real-time travel data is acquired (step 11), and the feature of the fault sign detection model adopted in step 9 of FIG. 4 is calculated for the real-time travel data (step 12). The real-time travel data is acquired with the same time width as the normal travel data and the pre-fault travel data when the above-mentioned classifier is created. Then, in the feature space, the distance of the state point of the feature of the real-time travel data from the feature distribution of the normal travel data is calculated (step 13), and when the distance exceeds the distance (predetermined distance) from the feature distribution of the normal travel data to the tail of the distribution of the pre-fault travel data (step 14), it is determined that a fault sign appears in the real-time travel data, and the fault sign or even the driving mode at that time is pointed out as being peculiar (step 15). Such a point may be conveyed to the driver or the like through a fault sign alarm as already described. When the distance of the state point of the feature of the real-time driving data from the feature distribution of the normal driving data is less than the predetermined distance, no indication (notification) is executed (step 16).

[0029] The above-described failure sign detection process may be executed for each part of the vehicle, and when a failure sign is detected in any part, the failure sign may be notified. At that time, it may also be notified in which part the failure sign was detected.

[0030] Thus, according to the configuration of the present embodiment, before a breakdown occurs in a vehicle, a fault sign in any part can be detected in the driving data, and in that case, it is pointed out to the driver that the current driving behavior is a specific driving behavior that may lead to a breakdown. This is expected to encourage the driver to improve the driving behavior or inspect the part where the fault sign was detected, and the service life of the vehicle is expected to be extended.

[0031] The above description has been given in relation to the embodiment of the present invention, but it will be apparent to those skilled in the art that many modifications and changes can be easily made thereto, and that the present invention is not limited to the embodiment exemplified above, but can be applied to various devices without departing from the concept of the present invention.

Claims

[Claim 1] A failure sign detection device receives real-time driving data, which is driving data representing the state of a vehicle, and detects the presence or absence of a failure sign in each part of the vehicle, and a classifier is configured according to a machine learning algorithm that uses, as training data, normal driving data, which is driving data of a vehicle in a normal state that has been prepared in advance, and pre-failure driving data, which is driving data of a vehicle in which a failure has occurred in any part before the failure occurred, and obtains from the real-time driving data a feature value in which the magnitude of the difference between a value obtained from the normal driving data and a value obtained similarly from the pre-failure driving data is significant, and a method for configuring a device configured to detect the presence or absence of a failure sign in any part of the vehicle using the classifier that identifies that there is a failure sign when the magnitude of the difference between the feature value of the real-time driving data and the feature value of the normal driving data exceeds a predetermined magnitude, A step of adjusting a time width of the normal driving data and the pre-fault driving data used in calculating the feature amount so that the discrimination accuracy of the discriminator exceeds a predetermined value. The method includes:

Images