A vehicle fault detection method and system based on ECU data

Abstract

A vehicle fault detection method and system based on ECU data includes: constructing an ECU causality graph; extracting key features from all ECU data of the vehicle; obtaining statistical and temporal features corresponding to strongly correlated parameters within a sliding window and importing them into the ECU causality graph; obtaining the normal parameter distribution of the target vehicle; performing fault detection based on the statistical and temporal features corresponding to strongly correlated parameters within the sliding window of the target vehicle according to the normal parameter distribution; marking the target vehicle as a fault state or constructing a prediction model according to the detection results; evaluating the ECU causality graph according to the prediction model; and marking the target vehicle as a critical fault state or a normal state according to the evaluation results. By jointly evaluating common and rare fault prediction models, the method captures the trend signals of parameter features evolving into fault states, and identifies the risk state of the vehicle being at the critical fault state in advance.

Description

Technical Field

[0001] This invention relates to the field of vehicle fault detection technology, specifically a vehicle fault detection method and system based on ECU data. Background Technology

[0002] With the development of automotive electronics technology, modern vehicles are equipped with electronic control units in multiple domains, such as engine ECU, transmission ECU, and battery management system (BMS). ECU data contains key information about the vehicle's operating status and has become the core data source for fault detection.

[0003] However, existing vehicle fault detection methods based on ECU data often rely on isolated parameters or fault codes (DTCs) of a single ECU for judgment, forming "parameter islands" that cannot capture the coordinated anomalies of parameters across multiple ECU domains (such as transmission faults that may cause abnormal engine ECU parameters). This results in high rates of missed and false detections, especially for latent faults without fault codes. Furthermore, the training of existing fault models depends on a large number of labeled samples, while rare faults have a low probability of occurrence and a very small sample size, making it impossible to build accurate models independently. As a result, these types of faults have always been unable to be effectively detected. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a vehicle fault detection method based on ECU data, comprising the following steps:

[0005] Step s1: Construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical features and time series features corresponding to the strongly correlated parameters within the sliding window, and import them into the ECU cause-effect graph;

[0006] Step s2: Construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, and mark the target vehicle as faulty or execute step s4 based on the detection results.

[0007] Step s3: Perform statistical analysis on the full vehicle ECU fault data in the sample database to obtain common fault types and rare fault types. Perform migration feature analysis on common fault types and rare fault types to obtain migration parameter nodes for rare fault types. Construct a prediction model for common fault types. Construct a prediction model for rare fault types based on the migration parameter nodes and the model parameters of the prediction model for common fault types.

[0008] Step s4: Construct a first prediction model based on the model parameters of the prediction models for common and rare fault types. Evaluate the ECU cause-effect graph based on the first prediction model. Mark the target vehicle as either in a critical fault state or a normal state based on the evaluation results.

[0009] Furthermore, a cloud monitoring platform is constructed, and ECU acquisition terminals and edge lightweight modules are deployed on the target vehicle. The cloud monitoring platform communicates with several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full vehicle ECU data and normal vehicle driving records of the target vehicle to the edge lightweight modules and the cloud monitoring platform. The cloud monitoring platform is used to execute step s3, and the edge lightweight modules are used to execute steps s1, s2 and s4.

[0010] The cloud-based monitoring platform has a sample database for storing all vehicle ECU fault data and all vehicle ECU normal data. The vehicle ECU fault data includes fault type, statistical features and time series features corresponding to strongly correlated parameters within the sliding window, and the vehicle ECU normal data includes vehicle normal driving records, statistical features and time series features corresponding to strongly correlated parameters within the sliding window.

[0011] Furthermore, the process of constructing an ECU causality graph, extracting key features from the received full-vehicle ECU data, obtaining the statistical and temporal features corresponding to strongly correlated parameters within a sliding window, and importing them into the ECU causality graph includes:

[0012] Predefine parameter nodes, component nodes, and fault type nodes, as well as the causal relationships among them. Based on the causal relationships, construct directed edges between parameter nodes, component nodes, and fault type nodes to generate an ECU causal graph. Then, add the ECU causal graph to the edge lightweighting module.

[0013] The received full vehicle ECU data is preprocessed, and the preprocessed full vehicle ECU data is sorted by timestamp to construct a time series matrix. Strongly correlated parameter pairs in the time series matrix are obtained through the causal relationships in the ECU causal graph. The strongly correlated parameter pairs are divided into blocks using a sliding window, and the statistical features and temporal features corresponding to the strongly correlated parameters in each sliding window are obtained. The statistical features of the strongly correlated parameter pairs in the time series matrix are imported into the corresponding parameter nodes in the ECU causal graph, and the temporal features of the strongly correlated parameter pairs in the time series matrix are imported into the directed edges between the corresponding parameter nodes in the ECU causal graph.

[0014] Furthermore, the fault detection process includes:

[0015] Construct a parameter distribution model, extract normal data of all vehicle ECUs from the sample database as training data, use the training data to train the parameter distribution model, and obtain the completed parameter distribution model;

[0016] The normal driving record of the target vehicle is input into the parameter distribution model at the start timestamp of the sliding window. The normal parameter distribution of the target vehicle is output according to the parameter distribution model. The statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window of the target vehicle are compared with the normal parameter distribution to obtain the statistical feature deviation and time-series feature deviation. A preset error threshold is set. If there is a statistical feature deviation or the time-series feature deviation is greater than the error threshold, an unknown fault alarm signal is generated and the target vehicle is marked as faulty. The fault type of the unknown fault is obtained, and the fault type, the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window are uploaded to the sample database.

[0017] If both the statistical characteristic deviation and the time series characteristic deviation are less than or equal to the error threshold, then proceed to step s4.

[0018] Furthermore, the process of obtaining common and rare fault types includes:

[0019] Extract all vehicle ECU fault data from the sample database within the historical sliding window, extract the fault types from the vehicle ECU fault data for statistical analysis, obtain the occurrence ratio of each fault type, preset the sample training threshold, mark the fault types with an occurrence ratio less than the sample training threshold as rare fault types, and mark the fault types with an occurrence ratio greater than or equal to the sample training threshold as common fault types.

[0020] Furthermore, the process of performing migration feature analysis on common and rare fault types to obtain migration parameter nodes for rare fault types includes:

[0021] Extract the statistical and temporal features of strongly correlated parameter pairs of common and rare fault types in the sliding window from the fault data of the whole vehicle ECU. Obtain the parameter nodes corresponding to the strongly correlated parameter pairs in the ECU causal graph and the part nodes connected to the parameter nodes. Mark the parameter nodes corresponding to the strongly correlated parameter pairs in the ECU causal graph and the part nodes connected to the parameter nodes as feature chains.

[0022] Select a feature chain for a rare fault type, compare the parameter nodes of the feature chain for each common fault type with the parameter nodes of the feature chain for a rare fault type, and obtain the total number of times that parameter nodes of each type appear together in the feature chains of common fault types and the feature chain for a rare fault type.

[0023] A preset migration threshold is used to filter out parameter nodes whose total number of migrations exceeds the migration threshold, and these parameter nodes are marked as migration parameter nodes of a certain rare fault type.

[0024] Furthermore, the process of constructing a predictive model for common fault types based on the full vehicle ECU fault data corresponding to common fault types includes:

[0025] Statistical and temporal features of strongly correlated parameter pairs for common fault types are extracted from the vehicle ECU fault data. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of common fault types. The temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of common fault types.

[0026] The feature chains of common fault types with completed statistical feature import and temporal feature addition are used as training data to construct a prediction model for common fault types. The prediction model for common fault types is trained using the training data to obtain the prediction model for common fault types after training. The model parameters of the prediction model for common fault types after training are obtained and sent to the edge lightweight module.

[0027] Furthermore, the process of constructing a prediction model for rare fault types includes:

[0028] Construct a prediction model for rare fault types. Extract the feature chain containing the most migration parameter nodes of rare fault types from the feature chains of each common fault type. Use the model parameters of the prediction model for common fault types corresponding to the feature chain as the initial parameters of the prediction model for rare fault types.

[0029] Statistical and temporal features of strongly correlated parameter pairs of rare fault types are extracted from the ECU fault data of the whole vehicle. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of rare fault types, and the temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of rare fault types.

[0030] The feature chains of rare fault types, after completing the import of statistical features and the addition of temporal features, are used as training data. The prediction model of rare fault types is trained using the training data, and the prediction model of rare fault types is obtained. The model parameters of the prediction model of rare fault types are then sent to the edge lightweight module.

[0031] Furthermore, step s4 includes the following process:

[0032] A first prediction model is constructed based on the model parameters of the prediction models for common fault types and rare fault types. Parameter nodes with imported statistical features, part nodes connected to these parameter nodes, and directed edges with imported temporal features between these parameter nodes and the ECU causal graph are extracted and input into the first prediction model. The probability of occurrence of various fault types is output based on the first prediction model. A preset probability threshold is set. If the probability of occurrence of a fault type is greater than or equal to the probability threshold, a warning signal for the fault type is generated, and the target vehicle is marked as being in a critical fault state. The statistical features and temporal features corresponding to the fault type and strongly correlated parameters within the sliding window are uploaded to the sample database.

[0033] If the probability of occurrence of all fault types is less than the probability threshold, the target vehicle is marked as normal, and the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window are uploaded to the sample database.

[0034] A vehicle fault detection system based on ECU data includes a cloud monitoring platform. The cloud monitoring platform is connected to several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full ECU data of the target vehicle and the vehicle's normal driving records to the edge lightweight modules and the cloud monitoring platform.

[0035] The cloud-based monitoring platform is used to perform statistical analysis on the full vehicle ECU fault data in the sample database, obtain common fault types and rare fault types, perform migration feature analysis on common fault types and rare fault types, obtain migration parameter nodes for rare fault types, construct prediction models for common fault types, and construct prediction models for rare fault types based on migration parameter nodes and model parameters of prediction models for common fault types.

[0036] The edge lightweight module is used to construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window and import them into the ECU cause-effect graph, construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, mark the target vehicle as faulty or construct a first prediction model based on the model parameters of the prediction model of common fault types and rare fault types, evaluate the ECU cause-effect graph based on the first prediction model, and mark the target vehicle as faulty critical state or normal state based on the evaluation results.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. In vehicle malfunctions, most faults (such as low battery charge and insufficient engine oil) have sufficient samples, but rare faults (such as ECU chip solder joint defects and intermittent short circuits in sensor circuits) have low occurrence probabilities and extremely small sample sizes, making it difficult for traditional supervised learning models to train effectively. The common fault type prediction model of this invention is built based on structured training data of feature chains, enabling accurate matching of feature patterns for different fault types. The rare fault type prediction model reuses effective parameters from common fault models by transferring parameter nodes, eliminating the need to collect a large number of samples for each type of rare fault, thus reducing the cost and cycle of model training. Simultaneously, the sample database continuously receives fault data and normal operation data uploaded by the target vehicle, enabling dynamic iterative optimization of model parameters. This allows the model to adapt to dynamic situations such as vehicle component aging, changes in usage scenarios, and ECU firmware upgrades, ensuring that the model maintains good detection performance throughout the vehicle's entire lifecycle and is suitable for different vehicle types, including gasoline vehicles and new energy vehicles, as well as diverse usage scenarios such as urban commuting, highway driving, and extreme environments.

[0039] 2. This invention innovatively sets up a "critical fault state" judgment step, rather than simply dividing the vehicle into a binary state of "fault" and "normal." Instead, it captures the trend signals of parameter characteristics evolving towards a fault state through joint evaluation of common and rare fault prediction models, identifying the risk state of the vehicle being at the critical fault state in advance. This proactive early warning mechanism allows users to receive a warning before the fault fully occurs, giving them ample time for inspection and maintenance. It avoids the risks of breakdowns and sudden accidents caused by the traditional "fault detection after it occurs," upgrading from passive maintenance to proactive prevention, comprehensively ensuring driving safety and vehicle operational stability. For detected unknown fault types, combined with manual annotation (fault type confirmed by maintenance personnel), they are transformed into "known faults," expanding the model sample database. Attached Figure Description

[0040] Figure 1 This is a flowchart of a vehicle fault detection method based on ECU data according to an embodiment of this application;

[0041] Figure 2 This is a flowchart of a vehicle fault detection system based on ECU data, according to an embodiment of this application. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0043] like Figure 1As shown, a vehicle fault detection method based on ECU data includes the following steps:

[0044] Step s1: Construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical features and time series features corresponding to the strongly correlated parameters within the sliding window, and import them into the ECU cause-effect graph;

[0045] Step s2: Construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, and mark the target vehicle as faulty or execute step s4 based on the detection results.

[0046] Step s3: Perform statistical analysis on the full vehicle ECU fault data in the sample database to obtain common fault types and rare fault types. Perform migration feature analysis on common fault types and rare fault types to obtain migration parameter nodes for rare fault types. Construct a prediction model for common fault types. Construct a prediction model for rare fault types based on the migration parameter nodes and the model parameters of the prediction model for common fault types.

[0047] Step s4: Construct a first prediction model based on the model parameters of the prediction models for common and rare fault types. Evaluate the ECU cause-effect graph based on the first prediction model. Mark the target vehicle as either in a critical fault state or a normal state based on the evaluation results.

[0048] It should be further explained that, in the specific implementation process, a cloud monitoring platform is built, and ECU acquisition terminals and edge lightweight modules are deployed on the target vehicle. The cloud monitoring platform communicates with several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full vehicle ECU data and normal vehicle driving records of the target vehicle to the edge lightweight modules and the cloud monitoring platform. The cloud monitoring platform is used to execute step s3, and the edge lightweight modules are used to execute steps s1, s2 and s4.

[0049] The cloud-based monitoring platform has a sample database for storing vehicle ECU fault data and vehicle ECU normal data. The vehicle ECU fault data includes fault types, statistical features and time-series features corresponding to strongly correlated parameters within the sliding window, and vehicle ECU normal data includes normal vehicle driving records (including the total mileage and number of years the vehicle has driven under normal conditions), statistical features and time-series features corresponding to strongly correlated parameters within the sliding window. The sources of the vehicle ECU fault dataset include: data accumulation from existing and new users of the cloud-based monitoring platform, and cooperation with the vehicle connectivity big data platform of commercial vehicle companies. The vehicle ECU fault data and vehicle ECU normal data are obtained with the authorization of vehicle users, and the reliability and legality of the data sources are ensured in accordance with the principle of privacy protection.

[0050] The vehicle's ECU data includes engine ECU (speed, fuel injection quantity, intake pressure, coolant temperature), transmission ECU (input shaft speed, oil temperature), and fuel system ECU (fuel pump pressure, fuel injector switch status), etc.

[0051] It should be further explained that, in the specific implementation process, the process of constructing the ECU cause-effect graph, extracting key features from the received full-vehicle ECU data, obtaining the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window, and importing them into the ECU cause-effect graph includes:

[0052] Combining vehicle dynamics principles and maintenance expert experience, parameter nodes (such as speed sensor data, fuel pump pressure data), component nodes (such as speed sensor, fuel pump, engine block), and fault type nodes (such as misfire fault, battery voltage drift fault) are predefined, as well as the causal relationships between parameter nodes, component nodes, and fault type nodes (e.g., fuel pump (component) → fuel pump pressure (parameter) → fuel injection quantity (parameter) → engine speed (parameter) → misfire fault (fault)). Based on the causal relationships, directed edges are constructed between parameter nodes, component nodes, and fault type nodes to generate an ECU causal graph, which is then added to the edge lightweighting module.

[0053] The received vehicle ECU data undergoes preprocessing, including timestamp alignment: using the high-precision clock of the vehicle ECU as a reference, linear interpolation is employed to correct the timestamps of each parameter, ensuring that all data are synchronized in the time dimension (error ≤ 1ms); outlier filtering: Kalman filtering is used to remove random noise caused by environmental factors such as road conditions and temperature, and extreme outliers (such as numerical mutations caused by sensor failures) are identified using the 3σ criterion, and the mean of adjacent time points is used for replacement; data standardization: parameters of different dimensions (such as speed unit r / min and pressure unit bar) are Z-score standardized to the [0,1] interval to eliminate the impact of dimensional differences on subsequent fusion and modeling, and the preprocessed vehicle ECU data is arranged by timestamp. A time-series matrix of "time × parameter" is constructed (e.g., 1000 time points × 50 key parameters). Strongly correlated parameter pairs (parameters corresponding to two parameter nodes with a causal relationship) are obtained from the causal relationships in the ECU causal graph. A sliding window (window size 5s, step size 1s) is used to divide the strongly correlated parameter pairs into blocks. Statistical characteristics (mean, variance, peak value, kurtosis) and temporal characteristics (specifically, causal influence coefficients, obtained using the numerical time-series sequence of the strongly correlated parameter pairs within the sliding window) and the time-series characteristics of the strongly correlated parameters within each sliding window are obtained through structural equation modeling (SEM). For example, for the causal relationship "fuel pump pressure (X) → fuel injection quantity (Y)," a SEM equation is constructed. The coefficients are solved using the least squares method. (like =0.85), indicating that for every 10% change in fuel pump pressure, the fuel injection quantity changes by 8.5%. The statistical characteristics of the strongly correlated parameter pairs in the time series matrix are imported into the corresponding parameter nodes of the ECU causal graph, and the temporal characteristics of the strongly correlated parameter pairs in the time series matrix are imported into the directed edges between the corresponding parameter nodes of the ECU causal graph.

[0054] It needs further explanation that, in the specific implementation process, the process of constructing a parameter distribution model, outputting the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, and performing fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, and marking the target vehicle as faulty or executing step s4 based on the detection results includes:

[0055] A parameter distribution model is constructed by extracting normal data from the entire vehicle ECU in the sample database as training data. The parameter distribution model is trained using the training data. An autoencoder (AE) is used to train the model using only the entire vehicle ECU data under normal conditions, learning the parameter distribution under normal conditions (such as the causal influence coefficient and statistical characteristics of speed, oil pressure, and voltage), and obtaining the completed parameter distribution model.

[0056] The normal driving record of the target vehicle is input into the parameter distribution model at the start timestamp of the sliding window. Based on the parameter distribution model, the normal parameter distribution of the target vehicle (normal statistical features and normal temporal features of strongly correlated parameter pairs) is output. The statistical features and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle are compared with the normal parameter distribution to obtain the statistical feature deviation. ) and time series characteristic deviation ( The system sets a preset error threshold (5%). If there is a statistical feature deviation or a time series feature deviation greater than the error threshold, an unknown fault alarm signal is generated, and the target vehicle is marked as faulty. The fault type of the unknown fault is obtained (obtained through manual detection), and the statistical features and time series features corresponding to the fault type and the strongly correlated parameters in the sliding window are uploaded to the sample database.

[0057] If both the statistical characteristic deviation and the time series characteristic deviation are less than or equal to the error threshold, then proceed to step s4.

[0058] It should be further explained that, in the specific implementation process, the statistical analysis of fault types in the full vehicle ECU fault data of the sample database to obtain common and rare fault types includes:

[0059] Extract all vehicle ECU fault data from the sample database within the historical sliding window, extract the fault types from the vehicle ECU fault data for statistical analysis, obtain the occurrence percentage of each fault type (the number of occurrences of a certain fault type / the total number of occurrences of all fault types), preset the sample training threshold (3%), mark the fault types with an occurrence percentage less than the sample training threshold as rare fault types, and mark the fault types with an occurrence percentage greater than or equal to the sample training threshold as common fault types.

[0060] It should be further explained that, in the specific implementation process, the process of performing migration feature analysis on common and rare fault types to obtain migration parameter nodes for rare fault types includes:

[0061] Extract the statistical and temporal features of strongly correlated parameter pairs of common and rare fault types in the ECU fault data of the whole vehicle within a sliding window. Obtain the parameter nodes (e.g., parameter nodes fuel pump pressure (parameter) → fuel injection quantity (parameter)) and the part nodes connected to the parameter nodes in the ECU causal graph of the strongly correlated parameter pairs. Mark the parameter nodes and the part nodes connected to the parameter nodes in the ECU causal graph of the strongly correlated parameter pairs as feature chains.

[0062] Select a feature chain for a rare fault type, compare the parameter nodes of the feature chain for each common fault type with the parameter nodes of the feature chain for a rare fault type, and obtain the total number of times that parameter nodes of each type appear together in the feature chains of common fault types and the feature chain for a rare fault type.

[0063] For example: Common fault types F1: Insufficient fuel pump pressure (sample size 1000); F2: Clogged fuel injectors (sample size 800); F3: Aging spark plugs (sample size 600); Rare fault types (Category 1): Fx: Clogged fuel pump filter + slight fuel injector sticking (sample size 30, percentage ≤3%).

[0064] Extract the parameter node set Sx = {fuel pump pressure, fuel injection quantity, injector current, intake air pressure, engine speed} for rare faults Fx. Extract the parameter node sets for each common fault: S1(F1) = {fuel pump pressure, fuel pump voltage, fuel injection quantity, engine speed}; S2(F2) = {fuel injection quantity, injector current, intake air pressure, engine speed}; S3(F3) = {ignition coil voltage, ignition energy, intake air pressure, engine speed}. Determine whether each parameter node in Sx exists in S1(F1) / S2. In (F2) / S3 (F3), the number of times they occur together is counted. The total number of times "engine speed" occurs together is 3. This indicates that this parameter is the "common core parameter" of all common faults (F1 / F2 / F3) and rare faults Fx. No matter whether the root cause of the fault is fuel supply (F1 / F2), ignition system (F3) or rare fault (Fx), it will eventually be reflected through "abnormal engine speed". This is consistent with the actual fault logic (speed is a direct representation of engine power output. Any combustion / power-related fault will cause speed fluctuations).

[0065] A preset migration threshold (N×80%, where N is the total number of comparisons between the parameter nodes of the feature chain of each common fault type and the parameter nodes of the feature chain of a rare fault type) is used to filter out parameter nodes whose total number of comparisons exceeds the migration threshold, and these parameter nodes are marked as migration parameter nodes of a rare fault type.

[0066] It should be further explained that, in the specific implementation process, the process of constructing a predictive model for common fault types based on the full vehicle ECU fault data corresponding to common fault types includes:

[0067] Statistical and temporal features of strongly correlated parameter pairs for common fault types are extracted from the vehicle ECU fault data. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of common fault types. The temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of common fault types.

[0068] The feature chains of common fault types with completed statistical feature import and temporal feature addition are used as training data to construct a prediction model for common fault types. The prediction model for common fault types is trained using the training data to obtain the prediction model for common fault types after training. The model parameters of the prediction model for common fault types after training are obtained and sent to the edge lightweight module.

[0069] It should be further explained that, in the specific implementation process, the prediction model for common and rare fault types includes an input layer: the input data consists of feature chains that have completed the import of statistical features and the addition of temporal features, organized into three types of tensors according to a graph structure, as shown in Table 1 below:

[0070] Table 1

[0071]

[0072] Node feature encoding layer: compresses the 4-dimensional statistical features of parameter nodes into low-dimensional embedding vectors to reduce computation, while assigning fixed embeddings to part nodes (to maintain topological integrity).

[0073] Structure: 2-layer fully connected network (FC) + ReLU activation function, lightweight design (no bias term, reduced parameters): First layer: Input dimension 4 (statistical feature count) → Output dimension 8 (intermediate embedding); Second layer: Input dimension 8 → Output dimension 4 (final node embedding vector); Part node processing: Assign a fixed 4-dimensional embedding vector to each part node (e.g., fuel pump → [0.1, 0.2, 0.3, 0.4], fuel injector → [0.2, 0.3, 0.4, 0.5]), avoiding interference from part nodes on feature calculation, only retaining the topological connection function.

[0074] Temporal feature extraction layer: Structure: 1-layer lightweight 1D-CNN (no pooling layer, reducing information loss); Kernel size: 1 (covering local associations within the temporal window);

[0075] Graph Feature Fusion Layer (Core Layer: Topology + Feature Dual Fusion): Function: Fuses "node embedding features", "edge temporal embedding features" and "feature chain topology" to output a global feature vector (reflecting the overall feature pattern of the fault); Structure: Lightweight Directed Graph Convolutional Network (DGCN) adapted to the directed causal relationship of the feature chain; Graph convolution calculation logic (fitting the fault propagation direction): For each node i, its updated embedding vector = its own node embedding + the embeddings of all predecessor nodes pointing to i; Output: The updated embedding vectors of all nodes are concatenated to obtain the global feature vector (dimension 8×8=64 dimensions).

[0076] Fault Prediction Layer (Classification + Lightweight Optimization): Function: Maps global feature vectors to probability distributions of common fault types; Structure: 2-layer fully connected network + Dropout (to prevent overfitting): First layer: Input dimension 64 (global features) → Output dimension 32 (intermediate layer), with Dropout added (probability 0.2, can be turned off on edge devices to improve speed); Second layer: Input dimension 32 → Output dimension K (K is the number of common fault types, in this example K=3: F1, F2, F3); Activation function: The output layer uses Softmax, outputting the probability of occurrence of each common fault type (e.g., [0.92, 0.05, 0.03] indicates a 92% probability of predicting an F1 fault).

[0077] It should be further explained that, in the specific implementation process, the process of constructing a rare fault type prediction model based on the migration parameter nodes of rare fault types, the model parameters of the prediction model for common fault types, and the full vehicle ECU fault data corresponding to the rare fault type includes:

[0078] Construct a prediction model for rare fault types. Extract the feature chain containing the most migration parameter nodes of rare fault types from the feature chains of each common fault type. Use the model parameters of the prediction model for common fault types corresponding to the feature chain as the initial parameters of the prediction model for rare fault types.

[0079] Statistical and temporal features of strongly correlated parameter pairs of rare fault types are extracted from the ECU fault data of the whole vehicle. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of rare fault types, and the temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of rare fault types.

[0080] The feature chains of rare fault types, after completing the import of statistical features and the addition of temporal features, are used as training data. The prediction model of rare fault types is trained using the training data, and the prediction model of rare fault types is obtained. The model parameters of the prediction model of rare fault types are then sent to the edge lightweight module.

[0081] It should be further explained that, in the specific implementation process, the process of constructing a prediction model based on the model parameters of the prediction model for common fault types and the prediction model for rare fault types, evaluating the ECU cause-effect graph based on the prediction model, and marking the target vehicle as a critical fault state or a normal state based on the evaluation results includes:

[0082] A first prediction model (including prediction models for common and rare fault types) is constructed based on the model parameters of the prediction models for common and rare fault types received. Parameter nodes with imported statistical features, part nodes connected to these parameter nodes, and directed edges with imported temporal features between these parameter nodes and the ECU causal graph are extracted and input into the first prediction model. The probability of occurrence of each fault type is output based on the first prediction model, with a preset probability threshold Q (90%). If the probability of occurrence of a fault type is greater than or equal to the probability threshold, a warning signal for the fault type is generated, and the target vehicle is marked as being in a critical fault state. The statistical features and temporal features corresponding to the fault type and strongly correlated parameters within the sliding window are uploaded to the sample database (using an immediate upload method).

[0083] If the probability of occurrence of all fault types is less than the probability threshold, the target vehicle is marked as normal, and the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window are uploaded to the sample database (using a segmented upload method, packaging and uploading the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window to the sample database every 10 minutes).

[0084] like Figure 2 As shown, a vehicle fault detection system based on ECU data includes a cloud monitoring platform. The cloud monitoring platform is connected to several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full ECU data of the target vehicle and the vehicle's normal driving records to the edge lightweight modules and the cloud monitoring platform.

[0085] The cloud-based monitoring platform is used to perform statistical analysis on the full vehicle ECU fault data in the sample database, obtain common fault types and rare fault types, perform migration feature analysis on common fault types and rare fault types, obtain migration parameter nodes for rare fault types, construct prediction models for common fault types, and construct prediction models for rare fault types based on migration parameter nodes and model parameters of prediction models for common fault types.

[0086] The edge lightweight module is used to construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window and import them into the ECU cause-effect graph, construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, mark the target vehicle as faulty or construct a first prediction model based on the model parameters of the prediction model of common fault types and rare fault types, evaluate the ECU cause-effect graph based on the first prediction model, and mark the target vehicle as faulty critical state or normal state based on the evaluation results.

[0087] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A vehicle fault detection method based on ECU data, characterized by, Includes the following steps: Step s1: Construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical features and time series features corresponding to the strongly correlated parameters within the sliding window, and import them into the ECU cause-effect graph; Step s2: Construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical features and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, and mark the target vehicle as faulty or execute step s4 based on the detection results. Step s3: Perform statistical analysis on the full vehicle ECU fault data in the sample database to obtain common fault types and rare fault types. Perform migration feature analysis on common fault types and rare fault types to obtain migration parameter nodes for rare fault types. Construct a prediction model for common fault types. Construct a prediction model for rare fault types based on the migration parameter nodes and the model parameters of the prediction model for common fault types. Step s4: Construct a first prediction model based on the model parameters of the prediction models for common and rare fault types. Evaluate the ECU cause-effect graph based on the first prediction model. Mark the target vehicle as either in a critical fault state or a normal state based on the evaluation results.

2. The vehicle fault detection method based on ECU data according to claim 1, characterized in that, A cloud-based monitoring platform is constructed, and ECU acquisition terminals and edge lightweight modules are deployed on the target vehicle. The cloud-based monitoring platform communicates with several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full vehicle ECU data and normal vehicle driving records of the target vehicle to the edge lightweight modules and the cloud-based monitoring platform. The cloud-based monitoring platform is used to execute step s3, and the edge lightweight modules are used to execute steps s1, s2 and s4. The cloud-based monitoring platform has a sample database for storing all vehicle ECU fault data and all vehicle ECU normal data. The vehicle ECU fault data includes fault type, statistical features and time series features corresponding to strongly correlated parameters within the sliding window, and the vehicle ECU normal data includes vehicle normal driving records, statistical features and time series features corresponding to strongly correlated parameters within the sliding window.

3. The method of claim 2, wherein, The process of constructing an ECU cause-effect graph, extracting key features from all vehicle ECU data, obtaining statistical and temporal features corresponding to strongly correlated parameters within a sliding window, and importing them into the ECU cause-effect graph includes: Predefine parameter nodes, component nodes, and fault type nodes, as well as the causal relationships among them. Based on these causal relationships, construct directed edges between the parameter nodes, component nodes, and fault type nodes to generate an ECU causal graph. Then, add the ECU causal graph to the edge lightweighting module. The entire vehicle ECU data is preprocessed, and the preprocessed ECU data is arranged by timestamp to construct a time series matrix. Strongly correlated parameter pairs in the time series matrix are obtained through the causal relationships in the ECU cause-effect graph. The strongly correlated parameter pairs are divided into blocks using a sliding window, and the statistical and temporal features corresponding to the strongly correlated parameters in each sliding window are obtained. The statistical features of the strongly correlated parameter pairs in the time series matrix are imported into the corresponding parameter nodes of the ECU cause-effect graph, and the temporal features of the strongly correlated parameter pairs in the time series matrix are imported into the directed edges between the corresponding parameter nodes of the ECU cause-effect graph.

4. The method for detecting vehicle failure based on ECU data according to claim 3, characterized in that, The process of fault detection includes: Construct a parameter distribution model, extract normal data of all vehicle ECUs from the sample database as training data, use the training data to train the parameter distribution model, and obtain the completed parameter distribution model; The normal driving record of the target vehicle is input into the parameter distribution model at the start timestamp of the sliding window. The normal parameter distribution of the target vehicle is output according to the parameter distribution model. The statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window of the target vehicle are compared with the normal parameter distribution to obtain the statistical feature deviation and time-series feature deviation. A preset error threshold is set. If there is a statistical feature deviation or the time-series feature deviation is greater than the error threshold, an unknown fault alarm signal is generated and the target vehicle is marked as faulty. The fault type of the unknown fault is obtained, and the fault type, the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window are uploaded to the sample database. If both the statistical characteristic deviation and the time series characteristic deviation are less than or equal to the error threshold, then proceed to step s4.

5. The method for detecting vehicle failure based on ECU data according to claim 4, characterized in that, The process of obtaining common and rare fault types includes: Extract all vehicle ECU fault data from the sample database within the historical sliding window, extract the fault types from the vehicle ECU fault data for statistical analysis, obtain the occurrence ratio of each fault type, preset the sample training threshold, mark the fault types with an occurrence ratio less than the sample training threshold as rare fault types, and mark the fault types with an occurrence ratio greater than or equal to the sample training threshold as common fault types.

6. The method for detecting vehicle failure based on ECU data according to claim 5, wherein, The process of performing migration feature analysis on common and rare fault types to obtain migration parameter nodes for rare fault types includes: Extract the statistical and temporal features of strongly correlated parameter pairs of common and rare fault types in the sliding window from the fault data of the whole vehicle ECU. Obtain the parameter nodes corresponding to the strongly correlated parameter pairs in the ECU causal graph and the part nodes connected to the parameter nodes. Mark the parameter nodes corresponding to the strongly correlated parameter pairs in the ECU causal graph and the part nodes connected to the parameter nodes as feature chains. Select a feature chain for a rare fault type, compare the parameter nodes of the feature chain for each common fault type with the parameter nodes of the feature chain for a rare fault type, and obtain the total number of times that parameter nodes of each type appear together in the feature chains of common fault types and the feature chain for a rare fault type. A preset migration threshold is used to filter out parameter nodes whose total number of migrations exceeds the migration threshold, and these parameter nodes are marked as migration parameter nodes of a certain rare fault type.

7. The method of claim 6, wherein, The process of building predictive models for common fault types includes: Statistical and temporal features of strongly correlated parameter pairs for common fault types are extracted from the vehicle ECU fault data. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of common fault types. The temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of common fault types. The feature chains of common fault types with completed statistical feature import and temporal feature addition are used as training data to construct a prediction model for common fault types. The prediction model for common fault types is trained using the training data to obtain the prediction model for common fault types after training. The model parameters of the prediction model for common fault types after training are obtained and sent to the edge lightweight module.

8. The method for detecting vehicle failure based on ECU data according to claim 7, characterized in that, The process of building a prediction model for rare fault types includes: Construct a prediction model for rare fault types. Extract the feature chain containing the most migration parameter nodes of rare fault types from the feature chains of each common fault type. Use the model parameters of the prediction model for common fault types corresponding to the feature chain as the initial parameters of the prediction model for rare fault types. Statistical and temporal features of strongly correlated parameter pairs of rare fault types are extracted from the ECU fault data of the whole vehicle. The statistical features of strongly correlated parameter pairs are imported into the corresponding parameter nodes in the feature chain of rare fault types, and the temporal features of strongly correlated parameter pairs are added to the directed edges between the corresponding parameter nodes in the feature chain of rare fault types. The feature chains of rare fault types, after completing the import of statistical features and the addition of temporal features, are used as training data. The prediction model of rare fault types is trained using the training data, and the prediction model of rare fault types is obtained. The model parameters of the prediction model of rare fault types are then sent to the edge lightweight module.

9. The method for detecting vehicle failure based on ECU data according to claim 8, characterized in that, Step s4 includes the following process: A first prediction model is constructed based on the model parameters of the prediction models for common fault types and rare fault types. Parameter nodes with imported statistical features, part nodes connected to these parameter nodes, and directed edges with imported temporal features between these parameter nodes and the ECU causal graph are extracted and input into the first prediction model. The probability of occurrence of various fault types is output based on the first prediction model. A preset probability threshold is set. If the probability of occurrence of a fault type is greater than or equal to the probability threshold, a warning signal for the fault type is generated, and the target vehicle is marked as being in a critical fault state. The statistical features and temporal features corresponding to the fault type and strongly correlated parameters within the sliding window are uploaded to the sample database. If the probability of occurrence of all fault types is less than the probability threshold, the target vehicle is marked as normal, and the statistical features and time-series features corresponding to the strongly correlated parameters within the sliding window are uploaded to the sample database.

10. A vehicle fault detection system based on ECU data, specifically applied to the vehicle fault detection method based on ECU data as described in any one of claims 1 to 9, characterized in that, It includes a cloud monitoring platform, which communicates with several ECU acquisition terminals and lightweight modules within a preset range. The ECU acquisition terminals are used to upload the full ECU data of the target vehicle and the vehicle's normal driving records to the edge lightweight modules and the cloud monitoring platform. The cloud-based monitoring platform is used to perform statistical analysis on the full vehicle ECU fault data in the sample database, obtain common fault types and rare fault types, perform migration feature analysis on common fault types and rare fault types, obtain migration parameter nodes for rare fault types, construct prediction models for common fault types, and construct prediction models for rare fault types based on migration parameter nodes and model parameters of prediction models for common fault types. The edge lightweight module is used to construct the ECU cause-effect graph, extract key features from the ECU data of the whole vehicle, obtain the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window and import them into the ECU cause-effect graph, construct a parameter distribution model, output the normal parameter distribution of the target vehicle based on the parameter distribution model and the normal driving record of the target vehicle, perform fault detection on the statistical and temporal features corresponding to the strongly correlated parameters within the sliding window of the target vehicle based on the normal parameter distribution, mark the target vehicle as faulty or construct a first prediction model based on the model parameters of the prediction model of common fault types and rare fault types, evaluate the ECU cause-effect graph based on the first prediction model, and mark the target vehicle as faulty critical state or normal state based on the evaluation results.

Images