A method for detecting abnormal operation of vehicle software

Abstract

The present invention discloses a method for detecting abnormal operation of vehicle software, which mainly includes three major steps: determining the input and detection indicators of the software module, establishing a prediction model for the input and detection indicators through an AI training model, and comparing the detection indicators during the operation of the software; the present invention analyzes the relationship between the CPU occupancy rate and the input data of the perception fusion program based on the input data of the CPU of the perception fusion program and the CPU occupancy rate at the corresponding moment of perception fusion, and finally predicts the CPU occupancy rate of the perception fusion program using the input data of the perception fusion program. If the deviation between the actual operation output of the program and the result of the model prediction output is too large, it can be judged as an operation abnormality.

Description

Technical Field

[0001] The present invention belongs to the technical field of detecting the operation status of automobile software or software modules, and specifically relates to a method for detecting abnormal operation of vehicle software. Background Art

[0002] In the field of vehicle autonomous driving, real-time monitoring of the operation of software or software modules, such as CPU occupancy, stack usage, key output variables and other indicators, and determining whether they are in abnormal operation, plays a very important role. It can prevent vehicle driving abnormalities caused by abnormal software operation and greatly improve the safety of vehicle operation.

[0003] In the prior art, such as the patent application named "Vehicle Domain Controller Operation Status Monitoring Method" disclosed in CN113176771A, the disclosed monitoring method includes monitoring the occurrence of relevant important events and abnormal events, and caching the event occurrence time and event information in the form of logs into the memory. By using the microcontroller unit and the microprocessor unit as the core chip of the domain controller, using the microcontroller unit to process the software functions with predetermined requirements for real-time performance and / or safety performance, and using the microprocessor unit to process the software functions with predetermined requirements for computing performance and / or communication performance, based on the reasonable allocation and redundant backup of the microcontroller unit and the microprocessor unit, the vehicle domain controller takes into account real-time performance, safety performance, computing performance and communication performance, meets the requirements of different software functions for various performances in vehicle network management, and effectively improves the experience effect of the vehicle.

[0004] However, the above descriptions are mostly about monitoring and recording the overall operation of the domain controller, not detecting a single software or software module. It is impossible to judge whether the function of a certain software or software module is operating abnormally. It is more of an abnormal event log management, and the definition of abnormal events is relatively vague. It cannot use big data analysis methods to determine whether a certain software or software module is in an abnormal operating state at a certain moment, and it cannot prevent vehicle driving abnormalities caused by abnormal software or software module operation, and cannot timely discover or eliminate safety risks caused by abnormal software operation. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for detecting abnormal operation of vehicle software, so as to determine whether a certain software or software module is in an abnormal operation state at a certain moment through a big data analysis method, and promptly discover vehicle driving abnormalities caused by abnormal operation of the software or software module, thereby promptly discovering or eliminating the safety risks caused by abnormal software operation.

[0006] The technical solution of the present invention is achieved in this way:

[0007] A method for detecting abnormal operation of vehicle software, characterized by comprising the following steps:

[0008] 1) Parse the input data and CPU occupancy of the perception fusion program from the blf file;

[0009] 2) Filter the input data of the perception fusion program to find the input values ​​of all perception fusion programs corresponding to the CPU occupancy output time;

[0010] 3) Predict the input data corresponding to the perception fusion program in all cases through classification methods

[0011] Each CPU usage of

[0012] 4) Take the input data of the perception fusion program as the feature variable and first standardize the feature variable;

[0013] 5) Regularize the CPU input items to remove items that have little impact on CPU usage;

[0014] 6) Using CPU input items as feature variables, converting these feature variables into two dimensions and images, inputting the image-formatted data into the scene prediction model MobileNet, training the model, and when the model finally begins to converge, the training is considered complete; continuously inputting image-formatted data into the model to optimize the model until the final model classification accuracy reaches the set index N, and then terminate the model training;

[0015] 7) In the model verification phase, the program and the model are run simultaneously, and the input of the perception fusion program is consistent with the input of the scene prediction model. If the difference between the CPU occupancy rate issued by the scene prediction model and the actual controller is higher than the set value M, it is determined to be abnormal. In this way, the present invention uses the big data analysis method to determine whether a certain software or software module is in an abnormal operation state at a certain moment, and promptly discovers the vehicle driving abnormality caused by the abnormal operation of the software or software module, and then promptly discovers or eliminates the safety risks caused by the abnormal operation of the software.

[0016] Further: the classification method described in step 3) sets parameter d as the interval of CPU occupancy rate, manually divides CPU occupancy rate into 100 / d categories, uses the input items of the perception fusion program as features, classifies the input data of the corresponding output into the corresponding CPU occupancy rate category, and gradually reduces parameter d according to the classification effect until parameter d is reduced to 1. Then, the input items of the perception fusion program corresponding to the output items can be classified into every 1% of CPU occupancy rate through the classification method, that is, each CPU occupancy rate corresponding to the input data of the perception fusion program in all cases can be predicted, and the prediction of CPU occupancy rate is realized through the classification method. In this way, since the CPU occupancy rate of the present invention is indirectly related to the program input items, the general regression method cannot achieve good results, and the regression problem needs to be converted into a classification problem, and the prediction of CPU occupancy rate is realized through the classification method.

[0017] Further: the standardization of feature variables described in step 4) is standard feature scaling or maximum and minimum feature scaling.

[0018] Further: the N in step 6) is 90%. This data N is set according to the specific situation, so as to facilitate adjustment according to different models.

[0019] Further: the M in step 7) is 10%. This data M is set according to the specific situation, so as to facilitate adjustment according to different models.

[0020] Further: the abnormal result is saved or output to the vehicle computer for display, which is convenient for subsequent playback or monitoring.

[0021] Further: the scene prediction model adopts the MobileNET lightweight classification network.

[0022] Further: The input data source of the perception fusion program is sensor data, including at least target data and lane line data.

[0023] In summary, the present invention has the following beneficial effects:

[0024] 1. Use big data analysis methods to determine whether a certain software or software module is in an abnormal operating state at a certain moment, and promptly discover vehicle driving abnormalities caused by abnormal software or software module operation, thereby promptly discovering or eliminating safety risks caused by abnormal software operation.

[0025] 2. The present invention can guide developers to locate the cause of the problem and can play the role of warning of abnormal software operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a processing flow chart of this embodiment;

[0027] Figure 2 is a flow chart of the prediction model system of this embodiment;

[0028] Figure 3 This is a verification flow chart in this embodiment. DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] Abnormalities described in the present invention: The present invention defines abnormalities by means of scenario simulation, rather than simply giving a standard. If it exceeds or falls below a certain standard, it is considered abnormal. A software function module (here we use the perception module running on the MPU in the L3 project as an example) has different scenarios for different inputs, and different scenarios correspond to different indicators. The indicator may be the CPU occupancy rate or the length, curvature and other characteristics of a lane line under the input or scenario; if the detection indicator deviation is too large in similar scenarios, it means that the software is running abnormally.

[0031] Prediction model: The prediction model uses the MobileNET lightweight classification network. The present invention takes the L3 perception fusion program as an example. The input source of the perception fusion program is sensor data, including at least target data and lane line data; the detection target is the CPU occupancy rate of the perception fusion program on the MPU side of the controller. The present invention selects the CPU occupancy rate parameter as the detection indicator.

[0032] See also Figure 1 As shown in FIG. 3 , a method for detecting abnormal operation of vehicle software according to the present invention comprises the following steps:

[0033] 1. From the blf file (blf is the abbreviation of binary logging format, that is, binary data file, which is a format for storing data packets by the on-board data acquisition device. The blf file parsed here contains the relevant data when the on-board controller is running), that is, the CAN message and the Ethernet intermediate variable, parse out the input data and detection indicators of the perception fusion program, that is, the CPU occupancy rate;

[0034] 2. Filter the input data of the perception fusion program to find the input values ​​of all perception fusion programs corresponding to the CPU occupancy output time;

[0035] 3. Predict the CPU occupancy rate of each item corresponding to the input data of the perception fusion program in all cases by classification method; In traditional regression prediction problems, the explained quantity is generally regarded as a predictable value, but in this project, the correlation between CPU occupancy rate and program input items is not direct, or it is indirectly related, so the general regression method cannot achieve good results. In order to solve the problem, it is necessary to convert the regression problem into a classification problem, set the parameter d as the interval of CPU occupancy rate, and manually divide the CPU occupancy rate into 100 / d categories. Using the input items of the perception fusion program as features, the input data corresponding to the output is classified into the corresponding CPU occupancy rate category, and the parameter d is gradually reduced according to the classification effect. If the classification effect is ideal, the d parameter can be reduced to 1 in the end. Then, the input items of the perception fusion program corresponding to the output items can be classified into every 1% of CPU occupancy rate by classification method, that is, the CPU occupancy rate of each item corresponding to the input data of the perception fusion program in all cases can be predicted, and the prediction of CPU occupancy rate is realized by classification method.

[0036] 4. Take the input data of the perception fusion program as feature variables and first standardize the feature variables; the current common scheme is standard feature scaling or maximum and minimum feature scaling.

[0037] 5. Regularize the CPU input items and remove the items that have little impact on the CPU usage. Because there are many CPU input items, it is necessary to regularize the CPU input items and remove the items that have little impact on the CPU usage.

[0038] 6. Use the CPU input items as feature variables, convert these feature variables into two dimensions, and input the image format data into the scene prediction model (a common model in the prior art, named scene prediction model because the input data is scene-related feature variables) MobileNet, train the model, and consider the training completed when the model finally begins to converge. Continuously input data into the model and optimize the model until the final model classification accuracy reaches the set indicator N, such as 90% or above, and terminate the model training.

[0039] 7. In the scene prediction model verification phase, the program and the model are run simultaneously, and the input of the perception fusion program is consistent with the input of the scene prediction model. If the difference between the CPU occupancy rate issued by the prediction model and the actual controller is higher than the set value M, such as 10%, 12%, etc., it is judged to be abnormal. The result of the abnormal judgment is saved or output and displayed to the vehicle end, and provided to the developer to locate the problem.

[0040] The present invention provides a software operation anomaly detection method, which is a software anomaly detection method based on big data; it can be summarized into three main steps: determining the input and detection indicators of the software module, establishing a prediction model for the input and detection indicators through an AI training model, and comparing the detection indicators during software operation.

[0041] The present invention analyzes the relationship between the CPU occupancy rate and the input data of the perception fusion program based on the input data of the CPU of the perception fusion program and the CPU occupancy rate at the corresponding moment of perception fusion, and finally predicts the CPU occupancy rate of the perception fusion program using the input data of the perception fusion program. If the deviation between the actual running output of the program and the result of the model prediction output is too large, it can be judged as an operation abnormality.

[0042] Finally, it should be noted that the above examples of the present invention are merely examples for illustrating the present invention, and are not intended to limit the embodiments of the present invention. Although the applicant has described the present invention in detail with reference to the preferred embodiments, for those of ordinary skill in the art, other different forms of changes and modifications can be made based on the above description. It is impossible to list all the embodiments here. Any obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting abnormal operation of vehicle software, characterized in that The steps include: 1) Parse the input data and CPU occupancy of the perception fusion program from the blf file; 2) Filter the input data of the perception fusion program to find the input values ​​of all perception fusion programs corresponding to the CPU occupancy output time; 3) Predict the input data corresponding to the perception fusion program in all cases through classification methods Each CPU usage of The classification method sets parameter d as the interval of CPU occupancy rate, manually divides CPU occupancy rate into 100 / d categories, uses the input item of the perception fusion program as a feature, classifies the input data of the corresponding output into the corresponding CPU occupancy rate category, and gradually reduces parameter d according to the classification effect until parameter d is reduced to 1. Then, the input item of the perception fusion program corresponding to the output item can be classified into every 1% of CPU occupancy rate through the classification method, that is, each CPU occupancy rate corresponding to the input data of the perception fusion program in all cases can be predicted, and the prediction of CPU occupancy rate is realized through the classification method; 4) Taking the input data of the perception fusion program as feature variables, the feature variables are first standardized; 5) Regularize the CPU input items to remove items that have little impact on CPU usage; 6) Using CPU input items as feature variables, these feature variables are two-dimensionalized and converted into images. The image-formatted data is input into the scene prediction model MobileNet to train the model. When the model finally begins to converge, the training is considered complete. The image-formatted data is continuously input into the model to optimize the model until the final model classification accuracy reaches the set index N, and the model training is terminated. 7) In the model verification phase, the program and the model run simultaneously, and the input of the perception fusion program is consistent with the input of the scene prediction model. If the difference between the CPU occupancy rate issued by the scene prediction model and the actual controller is higher than the set value M, it is judged to be abnormal.

2. A vehicle software operation abnormality detection method according to claim 1, characterized in that: The standardization of feature variables described in step 4) is standard feature scaling or maximum and minimum feature scaling.

3. A vehicle software operation abnormality detection method according to claim 1 or 2, characterized in that: The N in step 6) is 90%.

4. A vehicle software operation abnormality detection method according to claim 1 or 2, characterized in that: In step 7), M is 10%.

5. The vehicle software operation abnormality detection method according to claim 3 is characterized by: The abnormality determination result is saved or output and displayed to the vehicle computer.

6. A vehicle software operation abnormality detection method according to claim 4, characterized in that: The abnormality determination result is saved or output and displayed to the vehicle computer.

7. A vehicle software operation abnormality detection method according to claim 1 or 2, characterized in that: The scenario prediction model uses the MobileNET lightweight classification network.

8. A vehicle software operation abnormality detection method according to claim 7, characterized in that: The input data source of the perception fusion program is sensor data, which at least includes target data and lane line data.

Images