Automobile health condition monitoring method, device, equipment and medium thereof

Abstract

The application relates to a vehicle health monitoring method, device, equipment and medium. The method comprises the following steps: after preprocessing vehicle multi-sensor time series data, a pre-trained working condition perception attention coding network is used to strip the working condition to generate a working condition normalized feature sequence; a dynamic health benchmark is constructed by a health benchmark generator, a causal fault extractor extracts sparse fault features and calculates a reconstruction residual; a dynamic early warning threshold is obtained based on the extreme value theory, after triggering the early warning, a graph neural network diagnosis model is used to diagnose the fault feature subsequence and output a diagnosis report. The method improves the detection sensitivity and diagnosis accuracy of early weak faults under complex working conditions by means of working condition stripping and construction of a life cycle adaptive dynamic health benchmark, enhances the robustness of the vehicle health management system to working condition interference, the adaptability to aging evolution and the explainability to fault root cause, and realizes the technical closed loop from abnormal early warning to maintenance decision.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to methods, devices, equipment and media for monitoring the health status of automobiles. Background Technology

[0002] With the rapid development of the automotive industry towards intelligence and connectivity, the reliability of key vehicle components directly affects driving safety and operating costs. Accurate monitoring and diagnosis of early, minor faults has become a core requirement in automotive health management. However, existing automotive fault monitoring methods are mostly based on fixed thresholds or traditional machine learning models, relying on manual feature extraction, which is difficult to adapt to complex and changing driving conditions. For example, during vehicle operation, the monitoring parameters collected by sensors are easily affected by factors such as vehicle speed, load, and ambient temperature, causing early, minor fault characteristics to be masked by operating noise, resulting in missed or false alarms. Furthermore, traditional methods often use statically set health benchmarks, which cannot reflect the aging effects of vehicles over accumulated operating time, making it difficult to adapt to the health status assessment needs of different lifecycles. In the fault feature extraction stage, existing fault feature extraction technologies lack consideration of the causal relationships between components, easily misjudging features of indirectly influencing variables as fault features, leading to insufficient fault location accuracy. In addition, existing warning thresholds are mostly set based on experience, unable to dynamically match changes in data distribution, while diagnostic models often only output fault categories, failing to provide root fault component location and quantitative assessment of fault severity, thus failing to meet the needs of precise maintenance in engineering practice. Summary of the Invention

[0003] Therefore, it is necessary to provide methods, devices, equipment and media for monitoring vehicle health status in response to the above-mentioned technical problems, with the aim of improving the ability to identify early weak fault characteristics and the anti-interference capability under complex operating conditions, and enhancing the dynamic adaptability of health benchmarks to the aging state of vehicles throughout their entire life cycle.

[0004] Firstly, this application provides a method for monitoring the health status of a vehicle, including:

[0005] The raw vehicle multi-sensor time-series data, including operating condition parameters and monitoring parameters, are preprocessed to obtain preprocessed multivariate time window data. The preprocessed multivariate time window data is then subjected to operating condition stripping processing through a pre-trained operating condition perception attention coding network to generate a normalized operating condition feature sequence.

[0006] A pre-trained health baseline generator is used to process the normalized feature sequence of operating conditions to generate a dynamic health baseline sequence. A pre-trained causal fault extractor is used to extract fault features from the normalized feature sequence of operating conditions to obtain fault features. These features are then combined with the dynamic health baseline sequence to generate a sparse fault feature vector sequence. Based on the normalized feature sequence of operating conditions and the dynamic health baseline sequence, a reconstructed residual sequence is calculated.

[0007] The absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the feature amplitude sequence of each dimension. Based on the pre-collected historical normal data, the distribution fitting process of the feature amplitude sequence of each dimension and the reconstructed residual sequence is performed using extreme value theory to calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold. The real-time feature amplitude sequence of each dimension is compared with the corresponding dynamic amplitude warning threshold to obtain the first comparison result, and the reconstructed residual sequence is compared with the corresponding dynamic residual warning threshold to obtain the second comparison result.

[0008] When the first comparison result and / or the second comparison result meet the preset triggering conditions, a warning signal containing the warning time and triggering feature dimension is generated; based on the warning time in the warning signal, the corresponding time window fault feature subsequence is extracted from the sparse fault feature vector sequence; the fault feature subsequence is processed by a pre-trained graph neural network diagnostic model to perform fault diagnosis and output a diagnostic report.

[0009] In one embodiment, the pre-trained working condition awareness attention coding network includes a working condition feature embedding layer, a monitoring feature embedding layer, a working condition attention module, a condition-gated recurrent unit, and a decoding layer.

[0010] The raw vehicle multi-sensor time-series data, including operating condition parameters and monitoring parameters, are preprocessed to obtain preprocessed multivariate time window data. A pre-trained operating condition perception attention coding network is then used to perform operating condition stripping on the preprocessed multivariate time window data, generating a normalized operating condition feature sequence, including:

[0011] The original multi-sensor time-series data of the vehicle is preprocessed to obtain preprocessed data. The preprocessed data is then divided into sliding windows according to a preset window length and a preset overlap rate to obtain initial time window data. Channel normalization is then performed on each sensor channel of the initial time window data to obtain preprocessed multivariate time window data.

[0012] The preprocessed multivariate time window data is separated into operating condition indication features and health monitoring features along the channel dimension; the operating condition indication features are extracted through the operating condition feature embedding layer to obtain the operating condition embedded features; the health monitoring features are extracted through the monitoring feature embedding layer to obtain the monitoring embedded features.

[0013] The working condition attention module calculates the attention weights of each working condition embedding feature to each monitoring embedding feature at each time step based on the working condition embedding features and the monitoring embedding features, and generates a working condition context vector.

[0014] Based on the conditionally gated recurrent unit, the monitoring embedded features and the operating condition context vector are fused, and the operating condition stripping modulation operation is applied to obtain the operating condition modulation features; the operating condition modulation features are then subjected to temporal coding to obtain the final hidden state.

[0015] The final hidden state is decoded by the decoding layer to generate a normalized feature sequence of operating conditions.

[0016] In one embodiment, the health baseline generator includes an encoding layer and a decoding layer; the causal fault extractor includes an initialization layer, a residual attention layer, and a causal encoding layer.

[0017] A pre-trained health baseline generator is used to process the normalized feature sequence of operating conditions to generate a dynamic health baseline sequence. A pre-trained causal fault extractor is then used to extract fault features from the normalized feature sequence of operating conditions to obtain fault features. These features are then combined with the dynamic health baseline sequence to generate a sparse fault feature vector sequence. Based on the normalized feature sequence of operating conditions and the dynamic health baseline sequence, a reconstructed residual sequence is calculated, including:

[0018] The normalized feature sequence of working conditions is input into the health baseline generator, and the normalized feature sequence of working conditions is compressed and encoded through the coding layer to obtain the potential health variables.

[0019] Collect vehicle cumulative running time data, perform location coding on the vehicle cumulative running time data to obtain the corresponding location coding features; fuse the location coding features with health latent variables to obtain conditional health latent variables;

[0020] By decoding and reconstructing the latent variables of conditional health through a decoding layer, a dynamic health benchmark sequence is generated.

[0021] The normalized feature sequence of operating conditions and the dynamic health baseline sequence are input into the causal fault extractor. The element-wise difference between the normalized feature sequence of operating conditions and the dynamic health baseline sequence is calculated through the initial layer to obtain the initial residual features. The weights of the initial residual features are calculated through the residual attention layer to obtain the residual attention weights. The normalized feature sequence of operating conditions is weighted and focused using the residual attention weights to obtain the weighted focused features.

[0022] Based on the causal coding layer, feature extraction is performed on the weighted focusing features to obtain fault features; the fault features are processed by a sparse activation function to generate a sparse fault feature vector sequence; the Euclidean distance between the normalized feature sequence of the working condition and the dynamic health benchmark sequence at each time step is calculated to obtain the reconstructed residual sequence.

[0023] In one embodiment, the absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the feature amplitude sequence of each dimension; based on pre-collected historical normal data, the distribution fitting processing of the feature amplitude sequence of each dimension and the reconstructed residual sequence is performed using extreme value theory to calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold, including:

[0024] The absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the initial feature amplitude of each dimension; the initial feature amplitude of each dimension is then processed by moving average to obtain the feature amplitude sequence of each dimension.

[0025] The reconstructed residual sequence is processed by moving average to obtain the residual moving average sequence;

[0026] Based on pre-collected historical normal data, extract feature amplitude samples and residual moving average samples of each dimension under historical normal conditions;

[0027] Using the over-threshold model in extreme value theory, the characteristic amplitude samples and residual moving average samples of each dimension under the historical normal state are screened to obtain the over-threshold amplitude samples and over-threshold residual samples of each dimension.

[0028] A generalized Pareto distribution is fitted to the over-threshold amplitude samples and over-threshold residual samples of each dimension, and the distribution parameters and residual distribution parameters of each dimension are calculated by maximum likelihood estimation.

[0029] Based on the preset confidence level and distribution parameters of each dimension, the dynamic amplitude warning threshold is calculated; based on the preset confidence level and residual distribution parameters, the dynamic residual warning threshold is calculated.

[0030] In one embodiment, the graph neural network diagnostic model includes an embedding layer, a graph attention layer, a classification head, and a regression head;

[0031] Based on the warning time in the warning signal, fault feature subsequences corresponding to the time window are extracted from the sparse fault feature vector sequence; the fault feature subsequences are processed for fault diagnosis using a pre-trained graph neural network diagnostic model, and a diagnostic report is output, including:

[0032] Based on the warning time, a preset diagnostic window range is determined, and feature data within the preset diagnostic window range is extracted from the sparse fault feature vector sequence to obtain a fault feature subsequence.

[0033] Based on the preset physical connection relationships of the vehicle system and the causal influence rules of components, a graph structure including component nodes and connection edges is constructed.

[0034] The fault feature subsequence is input into the pre-trained graph neural network diagnostic model. The dimension mapping of the fault feature subsequence is performed through the embedding layer to obtain node embedding features. The node embedding features are assigned to the corresponding component nodes in the graph structure as the initial node features of each component node.

[0035] Multi-round message passing is performed based on the graph attention layer, the feature information of adjacent nodes of each component node is aggregated, and the feature representation of each component node is updated to obtain the updated feature representation of each component node; the failure probability distribution of each component is obtained by performing probability calculation on the updated feature representation of each component node through the classification head.

[0036] Numerical regression is performed on the updated feature representations of each component node using a regression head to obtain the initial severity score for each component.

[0037] Extract the maximum value of the residual moving average sequence within the warning time corresponding to the warning signal, calculate the final severity score of each component based on the maximum value and the initial severity score of each component, and generate a diagnostic report based on the failure probability distribution of each component and the final severity score.

[0038] In one embodiment, the causal fault extractor is trained through the following steps:

[0039] S1. Collect vehicle time-series data under historical normal conditions and generate normalized feature sequences for training; collect historical vehicle cumulative running time data corresponding to the vehicle time-series data, perform location encoding on the historical vehicle cumulative running time data, and obtain location encoding features for training.

[0040] S2. Based on the physical structure and working coupling relationship of each preset key component of the vehicle, a prior cause-effect graph is constructed to describe the direct causal relationship between components. Based on the prior cause-effect graph, a set of variables directly affected by the fault and a set of variables indirectly affected by the fault are defined for each component node.

[0041] S3. Construct an adversarial discriminator. The input of the adversarial discriminator is a joint sample of fault feature dimensions and variable features, and the output is the distribution discrimination result of the joint sample.

[0042] S4, input the normalized feature sequence of the working conditions used for training and the position-encoded feature of the training into the health benchmark generator, and output the dynamic health benchmark sequence for training.

[0043] S5. Simulated fault features of different components are injected into the normalized feature sequence of the training conditions to obtain multiple sets of fault condition feature sequences; each set of fault condition feature sequences corresponds to a single type of component fault.

[0044] S6. Initialize the network parameters of the initial layer, residual attention layer and causal coding layer of the causal fault extractor to obtain the initial causal fault extractor. Input the fault condition feature sequence of each group and the corresponding dynamic health benchmark sequence for training into the initial causal fault extractor to obtain the fault features for training. Apply a sparse activation function to the fault features for training to obtain the sparse fault feature vector sequence for training.

[0045] S7. Extract the fault feature dimension corresponding to the component with injected simulated fault features from the sparse fault feature vector sequence for training, and extract the features of the variable set of indirect impact of the fault of the corresponding component from the fault condition feature sequence to obtain the indirect impact variable features for training.

[0046] S8. Combine the fault feature dimension with the training indirect influence variable features to obtain the true joint sample; randomly shuffle the training indirect influence variable features and combine them with the fault feature dimension to obtain the false joint sample; input the true joint sample and the false joint sample into the adversarial discriminator, with the goal of maximizing the discrimination error rate of the adversarial discriminator, train the adversarial discriminator, and backpropagate to adjust the network parameters of the causal fault extractor to minimize the mutual information between the fault feature dimension and the training indirect influence variable features;

[0047] S9, iterate through S5 to S8 until the loss function of the causal fault extractor converges and the discrimination accuracy of the adversarial discriminator meets the preset threshold, then stop training and obtain the trained causal fault extractor.

[0048] In one embodiment, the dynamic amplitude warning threshold is calculated using the following formula:

[0049]

[0050] in, For the first Dynamic amplitude warning threshold for dimensional feature amplitude. Confidence level; , These are the first under normal historical conditions. Mean and standard deviation of the amplitude samples of dimensional features; For the quantile function of the generalized Pareto distribution, For shape parameters, For scale parameters; For the first Historical sample size of dimensional feature magnitude; This is the logarithmic adjustment coefficient.

[0051] Secondly, this application also provides a vehicle health monitoring device, including:

[0052] The condition perception normalization module is used to preprocess the raw vehicle multi-sensor time-series data, including condition parameters and monitoring parameters, to obtain preprocessed multivariate time window data. The preprocessed multivariate time window data is then subjected to condition stripping processing by a pre-trained condition perception attention coding network to generate a condition normalized feature sequence.

[0053] The baseline construction and fault extraction module is used to generate a dynamic health baseline sequence by processing the normalized feature sequence of operating conditions through a pre-trained health baseline generator; to extract fault features from the normalized feature sequence of operating conditions using a pre-trained causal fault extractor, and to generate a sparse fault feature vector sequence by combining it with the dynamic health baseline sequence; and to calculate the reconstructed residual sequence based on the normalized feature sequence of operating conditions and the dynamic health baseline sequence.

[0054] The dynamic early warning judgment module is used to calculate the absolute value of each dimension of the sparse fault feature vector sequence to obtain the feature amplitude sequence of each dimension; based on the pre-collected historical normal data, the extreme value theory is used to perform distribution fitting processing on the feature amplitude sequence of each dimension and the reconstructed residual sequence respectively to calculate the corresponding dynamic amplitude early warning threshold and dynamic residual early warning threshold; the real-time feature amplitude sequence of each dimension is compared with the corresponding dynamic amplitude early warning threshold to obtain the first comparison result, and the reconstructed residual sequence is compared with the corresponding dynamic residual early warning threshold to obtain the second comparison result;

[0055] The intelligent diagnostic reasoning module is used to generate a warning signal containing the warning time and the trigger feature dimension when the first comparison result and / or the second comparison result meet the preset trigger conditions; based on the warning time in the warning signal, it extracts the fault feature subsequence corresponding to the time window from the sparse fault feature vector sequence; and performs fault diagnosis processing on the fault feature subsequence through a pre-trained graph neural network diagnostic model to output a diagnostic report.

[0056] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0057] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0058] The aforementioned vehicle health monitoring method, device, equipment, and medium first preprocess the original multi-sensor time-series data of the vehicle and then perform condition stripping using a pre-trained condition-aware attention encoding network. This solves the problem of condition interference masking early, subtle faults in traditional methods, improving the accuracy and anti-interference capability of feature normalization. Secondly, a dynamic health benchmark sequence is generated through a health benchmark generator, and a sparse fault feature vector sequence and reconstructed residual sequence are generated by combining a causal fault extractor, improving the purity and specificity of fault feature extraction. Furthermore, a dynamic warning threshold is fitted using extreme value theory based on historical normal data and compared in two dimensions, improving the dynamic adaptability and accuracy of fault warnings. Finally, the extracted fault feature subsequences are used to output a diagnostic report through a pre-trained graph neural network diagnostic model, effectively solving the problems of inaccurate localization and lack of quantitative evaluation in traditional diagnostic methods, and improving the comprehensiveness and engineering practicality of fault diagnosis. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart of a vehicle health monitoring method provided as an exemplary embodiment of the present invention;

[0061] Figure 2 A flowchart of a method for obtaining corresponding dynamic amplitude warning thresholds and dynamic residual warning thresholds is provided as an exemplary embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of a vehicle health monitoring device provided as an exemplary embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] In one embodiment, such as Figure 1 As shown, a method for monitoring vehicle health status is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0065] S101: Preprocess the raw vehicle multi-sensor time-series data, including operating condition parameters and monitoring parameters, to obtain preprocessed multivariate time window data; perform operating condition stripping processing on the preprocessed multivariate time window data through a pre-trained operating condition perception attention coding network to generate a normalized operating condition feature sequence.

[0066] Specifically, in the raw data collected by multiple vehicle sensors such as vibration sensors, temperature sensors, and pressure sensors, operating parameters (such as vehicle speed, engine speed, load, and ambient temperature) and monitoring parameters (such as vibration amplitude of key components and lubricating oil temperature) are often coupled. Furthermore, the raw data suffers from issues such as asynchronous timestamps, occasional outliers, and inconsistent data dimensions. Directly using this data for fault analysis can lead to operating condition interference masking early, subtle fault characteristics, reducing the accuracy of subsequent processing. Therefore, preprocessing the raw data is crucial. This includes time alignment and outlier removal. Then, by setting an appropriate sliding window length and overlap rate, sliding window segmentation can be performed to convert continuous time-series data into fixed-length fragments, resulting in preprocessed multivariate time window data.

[0067] After data preprocessing, a pre-trained working condition-aware attention encoding network can be used to perform working condition stripping on the pre-processed multivariate time window data. This network, through an attention mechanism, can dynamically capture the correlation between working condition parameters and monitoring parameters, thereby specifically eliminating the impact of working condition changes on monitoring parameters. For example, using pre-processed multivariate time window data as input, the network's pre-training process has learned the changing patterns of monitoring parameters under different working conditions. Therefore, through the network's internal attention weight allocation mechanism, it can focus on feature components strongly correlated with working condition changes and separate these components from the monitoring parameters, retaining only feature information directly related to the vehicle's health status. After encoding processing by this network, a working condition normalized feature sequence can be generated. This process effectively eliminates the interference of working condition factors such as vehicle speed and load on monitoring parameters, enabling the generated working condition normalized feature sequence to truly reflect the actual health status of key vehicle components, providing input data for the subsequent construction of dynamic health benchmarks and accurate extraction of fault features.

[0068] S102: The normalized feature sequence of operating conditions is processed by a pre-trained health benchmark generator to generate a dynamic health benchmark sequence; the normalized feature sequence of operating conditions is processed by a pre-trained causal fault extractor to extract fault features, and combined with the dynamic health benchmark sequence to generate a sparse fault feature vector sequence; the reconstructed residual sequence is calculated based on the normalized feature sequence of operating conditions and the dynamic health benchmark sequence.

[0069] Specifically, the health baseline generator can employ a conditional variational autoencoder or a physical information neural network-based structure to establish a statistical distribution model of the feature sequence under normal operating conditions. By introducing aging-related variables such as cumulative vehicle operating time and cycle count as conditional inputs, the generated health baseline sequence can dynamically evolve with the vehicle's lifecycle. For example, the encoder can map the current input feature sequence to the latent space to obtain the latent variable distribution parameters. The decoder samples and reconstructs the health baseline sequence based on this distribution. During training, minimizing the reconstruction error while constraining the KL divergence between the latent variable distribution and the standard normal distribution ensures the continuity and interpolability of the latent space, thus supporting smooth estimation of health status at different aging stages. Illustratively, the causal fault extractor can be constructed based on a structural causal model or a Granger causality test framework. It can identify fault features with direct causal effects from the normalized feature sequence of operating conditions, eliminating interference from indirect related variables caused by fault propagation. For example, this extractor can construct a directed acyclic graph between features, use the backdoor criterion or do-calculus to calculate the causal effect strength of each feature on the system's health status, retain only feature dimensions with significant causal relationships, and combine this with the dynamic health baseline sequence to calculate the degree of deviation, generating a sparse fault feature vector sequence. Its sparsity is reflected in activating only the feature dimensions that truly reflect the root cause of the fault, thus effectively suppressing false alarms caused by spurious correlations. Furthermore, a residual sequence can be reconstructed based on the difference between the normalized feature sequence and the dynamic health baseline sequence. This residual sequence quantifies the deviation between actual observations and the expected health status, serving as direct evidence of fault occurrence and supplementing the sparse fault feature vector sequence as the basis for judgment.

[0070] S103: Calculate the absolute value of each dimension of the sparse fault feature vector sequence to obtain the feature amplitude sequence of each dimension; based on the pre-collected historical normal data, use extreme value theory to perform distribution fitting processing on the feature amplitude sequence of each dimension and the reconstructed residual sequence respectively, and calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold; compare the real-time feature amplitude sequence of each dimension with the corresponding dynamic amplitude warning threshold to obtain the first comparison result, and compare the reconstructed residual sequence with the corresponding dynamic residual warning threshold to obtain the second comparison result.

[0071] Specifically, since the characteristic amplitudes and reconstructed residuals under normal vehicle operation follow a specific probability distribution, but the probability of extreme values ​​is extremely low, traditional threshold setting methods based on the mean or standard deviation are difficult to capture low-probability anomalies. Therefore, extreme value theory can be used to fit the distribution of the characteristic amplitude sequences and reconstructed residual sequences of historical normal data in various dimensions. For example, extreme values ​​of samples can be extracted using the peak exceeding threshold method or the block maxima method, and the distribution parameters can be solved using maximum likelihood estimation or moment estimation methods to obtain the cumulative distribution function of the fitted distribution. Subsequently, based on the preset false alarm probability or confidence level, the corresponding quantile can be calculated as a dynamic warning threshold. This threshold can be automatically adjusted according to the statistical characteristics of the data distribution. When changes in the vehicle operating environment or sensor characteristic drift cause the data distribution to evolve slowly, the threshold can be adaptively corrected through an online update mechanism to ensure the long-term stability and reliability of the warning system. In the real-time monitoring phase, the feature amplitude sequence is obtained by calculating the absolute value of each dimension of the current window sparse fault feature vector sequence. This sequence is then compared element by element with the corresponding dynamic amplitude warning threshold. Simultaneously, the comparison results between the current reconstruction residual sequence and the dynamic residual warning threshold are calculated to obtain the first comparison result and the second comparison result. Thus, the state of the fault occurrence can be characterized from the perspectives of fault feature activation intensity and overall reconstruction deviation.

[0072] S104: When the first comparison result and / or the second comparison result meet the preset triggering conditions, generate a warning signal containing the warning time and the triggering feature dimension; based on the warning time in the warning signal, extract the fault feature subsequence corresponding to the time window from the sparse fault feature vector sequence; perform fault diagnosis processing on the fault feature subsequence through a pre-trained graph neural network diagnostic model, and output a diagnostic report.

[0073] Specifically, the preset triggering conditions may include any skill dimension amplitude exceeding a threshold in the first comparison result, the mean or peak value of the residual sequence exceeding a threshold in the second comparison result, or a combination of both. When the first comparison result and / or the second comparison result meet the preset triggering conditions, a warning signal containing a precise warning timestamp and a trigger feature dimension identifier can be generated. The warning time corresponds to the moment when the fault feature first significantly deviates from the normal pattern, and the trigger feature dimension points to the feature dimension with the highest activation intensity in the causal fault extractor, which can provide initial clues for subsequent root cause localization. Subsequently, the corresponding time window can be extracted from the sparse fault feature vector sequence based on the warning time, and fault feature subsequences of its preceding and following extended intervals can also be extracted. The extended intervals need to cover the time delay of fault propagation to ensure that complete fault evolution information is included. The graph neural network diagnostic model uses this fault feature subsequence as input. Its graph structure can be a prior graph with sensors or components as nodes and physical connection relationships or causal dependencies as edges. Schematic, this graph structure can also be a data-driven graph that dynamically learns the adjacency matrix through an attention mechanism. Graph neural networks aggregate neighborhood node information through multi-layer message passing, update node representations, and then input them into fully connected layers or readout functions. They can output fault category probability distributions, root cause location results, and fault severity metrics. Fault severity is calculated based on the degree, duration, and propagation range of features deviating from a health baseline, and is presented using a normalized score or remaining service life prediction. Based on these outputs, the final diagnostic report can include fault descriptions, root cause localization, severity assessments, and maintenance recommendations, achieving a closed-loop technology from anomaly detection to actionable maintenance decisions.

[0074] The aforementioned method first performs condition-based stripping processing on multi-sensor time-series data to generate a normalized feature sequence, effectively suppressing interference from complex operating conditions and accurately extracting fault features, thus enhancing the ability to identify early, subtle faults. Second, based on the synergistic effect of a health baseline generator and a causal fault extractor, a dynamic health baseline sequence and a sparse fault feature vector sequence are constructed, improving the timeliness of health status assessment and the physical interpretability of fault location. Subsequently, extreme value theory is used to fit the distribution of the feature amplitude sequence and the reconstructed residual sequence, calculating a dynamic early warning threshold, enhancing the adaptability and reliability of anomaly detection. Finally, a graph neural network diagnostic model performs deep reasoning on the fault feature subsequences, outputting a diagnostic report, improving the accuracy and executability of maintenance decisions.

[0075] In one embodiment, the pre-trained working condition awareness attention coding network includes a working condition feature embedding layer, a monitoring feature embedding layer, a working condition attention module, a condition-gated recurrent unit, and a decoding layer.

[0076] The raw vehicle multi-sensor time-series data, including operating condition parameters and monitoring parameters, are preprocessed to obtain preprocessed multivariate time window data. A pre-trained operating condition perception attention coding network is then used to perform operating condition stripping on the preprocessed multivariate time window data, generating a normalized operating condition feature sequence, including:

[0077] The original multi-sensor time-series data of the vehicle is preprocessed to obtain preprocessed data. The preprocessed data is then divided into sliding windows according to a preset window length and a preset overlap rate to obtain initial time window data. Channel normalization is then performed on each sensor channel of the initial time window data to obtain preprocessed multivariate time window data.

[0078] The preprocessed multivariate time window data is separated into operating condition indication features and health monitoring features along the channel dimension; the operating condition indication features are extracted through the operating condition feature embedding layer to obtain the operating condition embedded features; the health monitoring features are extracted through the monitoring feature embedding layer to obtain the monitoring embedded features.

[0079] The working condition attention module calculates the attention weights of each working condition embedding feature to each monitoring embedding feature at each time step based on the working condition embedding features and the monitoring embedding features, and generates a working condition context vector.

[0080] Based on the conditionally gated recurrent unit, the monitoring embedded features and the operating condition context vector are fused and the operating condition stripping modulation operation is applied to obtain the operating condition modulation features. The operating condition modulation features are then subjected to temporal encoding to obtain the final hidden state. The final hidden state is then decoded by the decoding layer to generate the operating condition normalized feature sequence.

[0081] Specifically, based on the timestamp information of each sensor's data, linear interpolation can be used to complete and align the time-offset data, ensuring that all sensors form data samples at the same time scale. Then, for outliers in the original data that may be caused by sensor momentary malfunctions, electromagnetic interference, etc., an outlier removal method based on data statistical distribution can be used to retain data points within a reasonable distribution range. After data cleaning, the processed data can be divided into sliding windows according to a preset window length and a preset overlap rate, transforming the continuous time-series data stream into fixed-length initial time window data. The preset overlap rate ensures the continuity of data in adjacent windows, preventing the loss of critical information. Since different sensors have different measurement dimensions and numerical ranges, channel normalization can be performed on each sensor channel of the initial time window data. By mapping the data of each channel to the same numerical range, the influence of dimensional differences is eliminated, ultimately obtaining preprocessed multivariate time window data.

[0082] Furthermore, based on the functional attributes of the sensors, the preprocessed multivariate time window data is separated into operating condition indication features and health monitoring features along the channel dimension. The operating condition indication features originate from sensors characterizing the vehicle's operating conditions, while the health monitoring features originate from sensors directly reflecting the operating status of key components, such as vibration sensors, temperature sensors, and pressure sensors. Subsequently, both the operating condition feature embedding layer and the monitoring feature embedding layer can employ a multi-layer fully connected network structure, and the ReLU function can be used to introduce nonlinear mapping capabilities. The operating condition feature embedding layer receives the separated operating condition indication features as input and performs linear transformations and nonlinear activations on the high-dimensional original operating condition features using the weight matrix of the network layers, mapping them to a low-dimensional, dense operating condition embedding feature space to extract an abstract representation of the operating condition features and enhance the distinguishability of different operating condition modes. Similarly, the monitoring feature embedding layer can perform the same dimensional embedding processing on the health monitoring features to generate monitoring embedding features, thereby condensing and strengthening the key information of the health monitoring features.

[0083] Subsequently, the operating condition attention module can dynamically allocate weights based on the correlation strength between features, accurately capturing the influence of operating conditions on monitoring features. For example, the operating condition attention module receives operating condition embedded features and monitoring embedded features as input, and can use a dot product attention mechanism to calculate the similarity between each operating condition embedded feature and each monitoring embedded feature at each time step, with the formula as follows: , Indicates the first Time steps down The embedded features of the first working condition are used for the first... Attention weights for each monitored embedded feature, For the first The first step of time Each working condition is embedded in a feature vector. For the first The first step of time Each monitoring embedded feature vector, This represents the vector dot product operation. For the dimension of the embedded features, This is a scaling factor to avoid the problem of fading sound caused by excessively large dot product results. This represents the total number of embedded features for different operating conditions. The attention weights calculated using this formula quantify the influence of different operating condition features on the monitoring features; a larger weight value indicates a more significant interference from the corresponding operating condition feature on the monitoring features. Subsequently, based on the calculated attention weights, all embedded features for different operating conditions are weighted and summed to generate an operating condition context vector for each time step. This vector integrates the operating condition information that has the greatest impact on the current monitoring features, providing a targeted basis for subsequent accurate removal of operating condition interference.

[0084] Specifically, after obtaining the operating condition context vector, the monitoring embedded features and the operating condition context vector can be fused using a condition-gated recurrent unit (CGRU) and then subjected to operating condition stripping modulation. The CGRU introduces the operating condition context vector as a conditional variable into the traditional gated recurrent unit (GRU), dynamically adjusting the influence of operating condition information on the monitoring features through a gating mechanism. For example, the monitoring embedded features and the operating condition context vector are concatenated to obtain a fused vector, which is then input into the update and reset gates of the CGRU. Through the synergistic effect of the update and reset gates, the CGRU can adaptively retain effective information related to health status in the monitoring embedded features while suppressing interference information represented by the operating condition context vector. This process is the operating condition stripping modulation operation, ultimately outputting integrated timing information and interference-reduced operating condition modulation features. To further capture the long-term dependencies in the operating condition modulation features, the operating condition modulation features are subjected to time-series encoding. The feature sequence is then processed in depth by stacking multiple layers of conditionally gated recurrent units to extract cross-time step correlation information, ultimately obtaining the final hidden state containing complete time-series features and the removed operating condition interference.

[0085] Finally, the decoding layer can adopt a combination structure of a multi-layer fully connected network and a batch normalization layer. The batch normalization layer can accelerate network training convergence and improve feature stability. Therefore, this layer receives the final hidden state as input and can map the low-dimensional hidden state back to the feature space with the same dimension as the original monitoring features through the linear transformation of the fully connected network. Then, after nonlinear processing by the activation function, the physical meaning and distribution characteristics of the features are restored, generating a working condition normalized feature sequence. This feature completely removes the interference of working condition factors such as vehicle speed and load, and only retains feature information directly related to the health status of key vehicle components, providing input data support for subsequent stages.

[0086] In one embodiment, the health baseline generator includes an encoding layer and a decoding layer, and the causal fault extractor includes an initialization layer, a residual attention layer, and a causal encoding layer. A pre-trained health baseline generator processes the normalized feature sequence of operating conditions to generate a dynamic health baseline sequence. A pre-trained causal fault extractor extracts fault features from the normalized feature sequence of operating conditions to obtain fault features, which are then combined with the dynamic health baseline sequence to generate a sparse fault feature vector sequence. Based on the normalized feature sequence of operating conditions and the dynamic health baseline sequence, a reconstructed residual sequence is calculated, which may include:

[0087] The normalized feature sequence of operating conditions is input into the health baseline generator. The normalized feature sequence of operating conditions is compressed and encoded through the coding layer to obtain the health latent variables. The cumulative running time data of vehicles is collected and the cumulative running time data of vehicles is position encoded to obtain the corresponding position encoded features. The position encoded features are fused with the health latent variables to obtain the conditional health latent variables.

[0088] The conditional health latent variables are decoded and reconstructed through a decoding layer to generate a dynamic health baseline sequence. The normalized feature sequence of operating conditions and the dynamic health baseline sequence are input into a causal fault extractor. The element-wise difference between the normalized feature sequence of operating conditions and the dynamic health baseline sequence is calculated through an initial layer to obtain initial residual features. The initial residual features are weighted through a residual attention layer to obtain residual attention weights. The normalized feature sequence of operating conditions is weighted and focused using the residual attention weights to obtain weighted focused features.

[0089] Based on the causal coding layer, feature extraction is performed on the weighted focusing features to obtain fault features; the fault features are processed by a sparse activation function to generate a sparse fault feature vector sequence; the Euclidean distance between the normalized feature sequence of the working condition and the dynamic health benchmark sequence at each time step is calculated to obtain the reconstructed residual sequence.

[0090] Specifically, the encoding layer of the health benchmark generator can adopt a stacked structure of convolutional layers, fully connected layers, and batch normalization layers. The convolutional layers use 1D convolutional kernels to capture local temporal correlation information of the normalized feature sequence. The batch normalization layer accelerates network training convergence and suppresses gradient explosion. The fully connected layer maps the local features extracted by the convolution to a low-dimensional dense space. For example, the normalized feature sequence is input into the encoding layer step by step, and temporal features at different scales are extracted through multiple convolutional operations. Then, the feature dimension is compressed through the fully connected layer, outputting a fixed-dimensional, condensed health latent variable. This variable can characterize the core statistical characteristics and temporal evolution patterns of the vehicle's health status. Subsequently, the vehicle's cumulative running time data can be collected through the vehicle control system's OBD interface, and the cumulative running time parameter can be read from the vehicle bus. This parameter directly reflects the vehicle's aging degree and is a key input for dynamic health benchmark adaptation throughout its entire lifecycle. Since the cumulative running time is continuous numerical data, directly inputting it into the model can easily lead to model bias caused by differences in numerical ranges. Therefore, positional encoding can be performed to transform the time information into a high-dimensional dense vector to preserve its relative temporal relationship. In positional encoding, periodic vectors can be generated using sine and cosine functions. The calculation formula is as follows:

[0091]

[0092] in, Indicates cumulative running time In the Dimensional positional encoding value, The total dimension of the location-encoded features is consistent with the dimension of the health latent variables. This is the index for the encoded dimension, with values ​​ranging from 0 to... This encoding method can map different cumulative running times into high-dimensional vectors with discriminative power, and ensures the similarity of encoded vectors with similar running times through a periodic function, adapting to the continuity of vehicle aging. After location encoding is completed, a vector concatenation method is used to fuse the location encoding features with latent health variables. The fused variables simultaneously contain core health status information and aging time-series information, providing dual constraints for the generation of dynamic health benchmarks.

[0093] Specifically, the decoding layer can adopt a structure symmetrical to the encoding layer: a deconvolutional layer + a fully connected layer + a batch normalization layer. The deconvolutional layer is used to restore the compressed temporal dimension during the encoding process, while the fully connected layer maps the low-dimensional conditional health latent variables back to a feature space consistent with the dimension of the normalized feature sequence. For example, the conditional health latent variables are first expanded in dimension by the fully connected layer, and then the temporal length is gradually restored by the deconvolutional layer. The batch normalization layer ensures the stability of the feature distribution during the reconstruction process, ultimately generating a dynamic health baseline sequence. This sequence can adjust the reference standard of health status according to the changes in the vehicle's cumulative operating time. For example, as the cumulative operating time increases, the dynamic health baseline sequence will adaptively reflect the normal fluctuation range of the component after aging, avoiding misjudging normal changes caused by aging as faults and improving the accuracy of health status assessment.

[0094] Specifically, the initial layer of the causal fault extractor can be composed of element-wise subtraction units. After inputting the normalized feature sequence of the operating condition and the dynamic health baseline sequence into the causal fault extractor, subtraction operations can be performed on the normalized feature sequence of the operating condition and the dynamic health baseline sequence at the same time step and the same feature dimension. The magnitude of the final initial residual feature directly reflects the degree to which the feature of that dimension deviates from the health baseline. The larger the value, the higher the probability of a fault in that dimension. However, it still contains some meaningless small fluctuations caused by measurement noise and minor operating condition fluctuations. Therefore, a residual attention layer can be used to dynamically allocate attention weights based on the significance of the initial residual features, so that the model focuses on residual features with large deviations and weakens the influence of small fluctuations. For example, firstly, the absolute value of the initial residual feature at each time step and in each dimension is calculated to unify the influence of the deviation direction. Then, the absolute value of the residual in each dimension at each time step is normalized using the softmax function to obtain the residual attention weight. The calculation formula can be... , For the first Time step Dimensional residual attention weights, For the first Time step The absolute value of the initial residual feature of the dimension. The total dimension of the feature sequence is given. This formula amplifies the weight proportion of significant residuals through an exponential function and ensures that the sum of the weights is 1 through normalization, giving the weights probabilistic distribution characteristics. After obtaining the residual attention weights, they are multiplied element-wise with the normalized feature sequence of the operating conditions, thereby achieving weighted focusing processing. ,in For the first Time step Dimensionally weighted focusing features. This process strengthens fault-related feature components, suppresses meaningless noise, and lays the foundation for subsequent fault feature extraction.

[0095] Furthermore, the causal coding layer can adopt a structure of deep convolutional layers + gated activation units. The deep convolutional layers can capture long-term temporal dependencies and fault mode association information in the weighted focusing features through multiple convolutional kernels. The gated activation units (such as gated linear units, GLUs) can adaptively filter effective features and suppress redundant information through gating mechanisms. For example, after the weighted focusing features are processed by deep convolutional layers to extract multi-scale fault-related features, the features are filtered by gated activation units to retain the core features directly related to the occurrence and development of faults. Finally, the output can be a fault feature with reduced dimensions and dense information, which has eliminated most of the noise and irrelevant interference, retaining only the key information related to faults.

[0096] Subsequently, sparse activation function processing is applied to the fault features to further refine them, activating only the feature dimensions directly related to the root cause of the fault. The sparse activation function can be the Softshrink function, which sets feature elements with absolute values ​​less than a threshold to zero, retaining only significant feature elements with absolute values ​​greater than the threshold. This results in a sparse distribution of the generated sparse fault feature vector sequence, where only the feature dimension corresponding to the root cause of the fault remains non-zero, while other irrelevant dimensions are suppressed to zero. This sparsity effectively avoids false fault identification caused by spurious relevance features, significantly improving the purity and specificity of the fault features.

[0097] Finally, Euclidean distance is calculated between the normalized feature vector of the operating condition and the dynamic health baseline vector at each time step to obtain the reconstruction residuals for each time step. Arranging these residuals in chronological order yields the reconstruction residual sequence. This sequence quantifies the degree of deviation between the actual vehicle operating state and the health baseline state at each time step. The larger the value, the more significant the overall deviation. This sequence can serve as a supplementary judgment basis for the sparse fault feature vector sequence. The two sequences work together to comprehensively reflect the fault occurrence state and provide dual reliable support for subsequent fault warnings.

[0098] In one embodiment, the causal fault extractor can be trained through the following steps:

[0099] S111: Collect vehicle time-series data under historical normal conditions and generate normalized feature sequences for training; collect historical vehicle cumulative running time data corresponding to the vehicle time-series data, perform location encoding on the historical vehicle cumulative running time data, and obtain location-encoded features for training.

[0100] Specifically, the historical normal vehicle time-series data can be processed in the same way as in the above embodiments, such as data alignment and outlier removal, to generate a normalized feature sequence for training conditions, ensuring the consistency of the distribution between training and inference data. Similarly, the location encoding processing method for historical vehicle cumulative running time data is consistent with the location encoding logic in the dynamic health benchmark generation process. Continuous time values ​​are mapped to high-dimensional dense vectors using sine and cosine periodic functions to obtain location encoding features for training. These features can provide vehicle aging time-series information for the training process.

[0101] S112. Based on the physical structure and working coupling relationship of each preset key component of the vehicle, a priori causal graph is constructed to describe the direct causal relationship between components. Based on the priori causal graph, a set of variables directly affected by a fault and a set of variables indirectly affected by a fault are defined for each component node.

[0102] Intuitively, a prior causal graph can be constructed based on the mechanical structure, electrical connections, and working coupling principles of key vehicle components (such as the transmission relationship between the engine and transmission, and the hydraulic transmission relationship between the master cylinder and wheel cylinders). Nodes in the graph are defined as preset key vehicle components (such as the engine and transmission) and their corresponding monitoring sensors. Edges are defined as direct causal relationships between components, i.e., directed edges pointing from the causal source node to the affected node. For example, an engine failure directly affects the speed sensor data; therefore, a directed edge can be constructed from the engine node to the speed sensor node. Based on this prior causal graph, a set of variables is defined for each component node. The set of variables directly affected includes sensor feature dimensions that have a direct physical connection or direct monitoring relationship with the component, while the set of variables indirectly affected includes sensor feature dimensions that are affected after being transmitted through other components.

[0103] S113, Construct an adversarial discriminator. The input of the adversarial discriminator is a joint sample of fault feature dimensions and variable features, and the output is the distribution discrimination result of the joint sample.

[0104] Specifically, the adversarial discriminator can employ a multi-layer fully connected network structure with batch normalization layers. The network has 3-5 layers, and the ReLU activation function is used in the hidden layers to introduce non-linear discriminative capability. The output layer uses the Sigmoid activation function, mapping the output value to the [0,1] interval to represent the probability that the input joint sample belongs to the "true joint distribution." The discriminator's input is a joint sample of fault feature dimensions and variable features. These dimensions are mapped to the same dimension through the fully connected layers and then concatenated to form a joint sample vector. The output is a single-dimensional distribution discrimination result; a value close to 1 indicates the discriminator classifies the input as a true joint sample, while a value close to 0 indicates a false joint sample. Furthermore, during network initialization, the weight parameters can be initialized using the Xavier initialization method, and the bias parameters are initialized to 0 to ensure stable network output during the initial training phase.

[0105] S114 inputs the normalized feature sequence of the working conditions used for training and the position-encoded feature of the training into the health benchmark generator and outputs the dynamic health benchmark sequence for training.

[0106] Specifically, after inputting the normalized feature sequence of training conditions and the position-encoded features of training conditions into the generator, a dynamic health benchmark sequence for training can be output through the process of generating a dynamic health benchmark sequence as described in the above embodiment. This sequence can accurately reflect the distribution of health features of the corresponding training data under normal conditions.

[0107] S115, Inject simulated fault features of different components into the normalized feature sequence of training conditions to obtain multiple sets of fault condition feature sequences; wherein, each set of fault condition feature sequences corresponds to a single type of component fault.

[0108] Specifically, simulated fault features can include typical fault mechanisms of key vehicle components, such as wear, fatigue, loosening, and short circuits. Using an additive injection method, disturbance signals conforming to the fault evolution law can be superimposed on specific feature dimensions (the set of directly influencing variables corresponding to the target faulty component) of the normalized feature sequence of training conditions. The amplitude of the disturbance signal is set according to the fault severity level, and the temporal variation trend is derived from the statistical laws or physical models of actual fault data; for example, linear increase represents increased wear. Each set of normalized feature sequences of training conditions is injected with only one type of simulated fault feature, resulting in a set of fault condition feature sequences. This ensures that each set of data corresponds to a single fault root cause, ultimately forming multiple sets of fault condition feature sequences covering major key components and various fault types.

[0109] S116. Initialize the network parameters of the initial layer, residual attention layer and causal coding layer of the causal fault extractor to obtain the initial causal fault extractor. Input the fault condition feature sequence of each group and the corresponding dynamic health benchmark sequence for training into the initial causal fault extractor to obtain the fault features for training. Apply a sparse activation function to the fault features for training to obtain the sparse fault feature vector sequence for training.

[0110] For example, the Xavier initialization strategy can be used to initialize the network parameters of the causal fault extractor. The weights of the initial layer, residual attention layer, and causal coding layer are randomly generated according to this strategy, and the bias parameters are initialized to 0 to ensure consistent variance in the output data of each layer, avoiding gradient vanishing or exploding in the early stages of training. Each set of fault condition feature sequences and the corresponding dynamic health baseline sequence for training are synchronously input into the initial causal fault extractor. The initial layer calculates the element-wise difference between the two to obtain the initial residual features for training. The residual attention layer calculates attention weights based on the absolute value of the initial residual features, and performs weighted focusing on the fault condition feature sequences to obtain the weighted focused features for training. The causal coding layer extracts the fault features for training through depthwise convolution and gated activation units. Finally, the Softshrink sparse activation function is used to set feature elements smaller than a preset threshold to zero, retaining significant fault features to obtain the sparse fault feature vector sequence for training. This process is consistent with the inference process of the causal fault extractor in the above embodiment, ensuring that the training effect can be directly transferred to practical applications.

[0111] S117, extract the fault feature dimension corresponding to the component with injected simulated fault features from the sparse fault feature vector sequence for training, extract the features of the variable set of indirect influence of the corresponding component from the fault condition feature sequence, and obtain the indirect influence variable features for training.

[0112] Specifically, based on the component identifier of the injected simulated fault, a non-zero feature dimension can be extracted from the training sparse fault feature vector sequence. This dimension is the fault feature dimension corresponding to the target faulty component (since the sparse activation function has suppressed irrelevant dimensions, the non-zero dimension directly corresponds to the root cause of the fault). Based on the set of variables indirectly affected by the component fault as defined in S112, feature data of the corresponding dimension can be extracted from the fault condition feature sequence to obtain the training indirect influence variable feature. This feature represents the indirect interference information caused by fault propagation, and further training is needed to decouple the fault feature dimension from this feature.

[0113] S118: Combine the fault feature dimension with the indirect influence variable features used in training to obtain real joint samples; randomly shuffle the indirect influence variable features used in training and combine them with the fault feature dimension to obtain false joint samples; input the real joint samples and false joint samples into the adversarial discriminator, with the goal of maximizing the discrimination error rate of the adversarial discriminator, train the adversarial discriminator, and backpropagate to adjust the network parameters of the causal fault extractor to minimize the mutual information between the fault feature dimension and the indirect influence variable features used in training.

[0114] Specifically, two types of joint samples can be constructed first. The true joint samples are obtained by concatenating the fault feature dimension with the indirect influence variable features used in training, representing the true correlation distribution between the fault features and the indirect influence variable features. The false joint samples are obtained by randomly shuffling the elements of the indirect influence variable features used in training (disrupting their original correlation with the fault feature dimension) and then concatenating them with the fault feature dimension, representing an uncorrelated random distribution. The two types of joint samples are mixed in a 1:1 ratio, with true joint samples labeled as 1 and false joint samples as 0. This mixture is then input into an adversarial discriminator for training, using a binary cross-entropy loss function. During discriminator training, backpropagation can be used to adjust the discriminator parameters and minimize the loss function, enabling the discriminator to accurately distinguish between true and false joint samples. After a certain number of training iterations of the discriminator, the discriminator parameters can be fixed, and the causal fault extractor can be optimized instead. For example, the goal can be to maximize the discriminator's error rate, i.e., to make the discriminator unable to distinguish whether the input joint sample is real or fake. This goal is equivalent to minimizing the mutual information between the fault feature dimension and the indirect influence variable features used in training. Illustratively, the total loss function of the causal fault extractor can be a weighted sum of sparse loss and adversarial loss, where the adversarial loss is the discriminator's prediction loss on the joint sample constructed from the extractor's output features. By backpropagating the gradient of the adversarial loss back to the causal fault extractor and adjusting the network parameters of its initial layer, residual attention layer, and causal coding layer, the correlation between the fault feature dimension and the indirect influence variable features used in training can be gradually reduced, ultimately minimizing the mutual information.

[0115] S119, iterate through S115 to S118 until the loss function of the causal fault extractor converges and the discrimination accuracy of the adversarial discriminator meets the preset threshold, then stop training and obtain the trained causal fault extractor.

[0116] Specifically, during the iteration process, after each step from S115 to S118, the total loss function value (sparse loss + adversarial loss) of the causal fault extractor can be calculated, and the discrimination accuracy of the adversarial discriminator on the joint samples of the test set can be statistically analyzed. Training can then be stopped when the following two conditions are met: first, when the fluctuation range of the total loss function value of the causal fault extractor is less than a preset fluctuation threshold for a preset number of rounds (e.g., 20 rounds), it indicates that the model parameters have stabilized; second, when the discrimination accuracy of the adversarial discriminator approaches a random guessing level of around 50%, it indicates that the fault feature dimension and the indirect influencing variable features have been effectively decoupled, and mutual information has been minimized. By saving the network parameters (weight matrix, bias vector) of the corresponding causal fault extractor, a causal fault extractor with causal separation capabilities and the ability to accurately extract the root cause features of faults can be obtained.

[0117] In one embodiment, such as Figure 2 As shown, the absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the feature amplitude sequence of each dimension. Based on pre-collected historical normal data, extreme value theory is used to perform distribution fitting processing on the feature amplitude sequence of each dimension and the reconstructed residual sequence, respectively, to calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold, including:

[0118] S201: Calculate the absolute value of each dimension of the sparse fault feature vector sequence to obtain the initial feature amplitude of each dimension; perform moving average processing on the initial feature amplitude of each dimension to obtain the feature amplitude sequence of each dimension; perform moving average processing on the reconstructed residual sequence to obtain the residual moving average sequence.

[0119] S202: Based on pre-collected historical normal data, extract feature amplitude samples and residual moving average samples of each dimension under historical normal conditions;

[0120] S203: Using the over-threshold model in extreme value theory, the characteristic amplitude samples and residual moving average samples of each dimension under the historical normal state are screened to obtain the over-threshold amplitude samples and over-threshold residual samples of each dimension.

[0121] S204: Fit a generalized Pareto distribution to the over-threshold amplitude samples and over-threshold residual samples of each dimension respectively, and calculate the distribution parameters and residual distribution parameters of each dimension through maximum likelihood estimation.

[0122] S205: Calculate the dynamic amplitude warning threshold based on the preset confidence level and distribution parameters of each dimension; calculate the dynamic residual warning threshold based on the preset confidence level and residual distribution parameters.

[0123] Specifically, the dimensions of the sparse fault feature vector sequence correspond to the feature components of different components or fault modes. Their positive and negative values ​​only reflect the direction of feature fluctuations, while the absolute value characterizes the activation intensity of the fault feature. Therefore, the absolute value can be used to unify the quantification standard of fault intensity. Subsequently, the initial feature amplitudes and reconstructed residual sequences for each dimension are processed using moving averages to smooth out instantaneous noise and random fluctuations, preserving the evolution trend of fault features. For example, the length of the sliding window is consistent with the time window length of the previous data preprocessing. Smoothing can be achieved by calculating the arithmetic mean of the data within the window, avoiding false threshold triggering due to noise at a single moment, resulting in the feature amplitude sequences and residual moving average sequences for each dimension. Historical normal data, collected from vehicles under fault-free and abnormal operating conditions and processed using the same preprocessing and feature extraction procedures, covers different operating conditions and aging stages of the vehicle. Therefore, the feature amplitude samples and residual moving average samples extracted from this dataset provide the basic data for subsequent extreme value distribution fitting.

[0124] Indicatively, the core principle of the over-threshold model is to focus on extreme values ​​in the sample that exceed a certain high threshold. However, because early, minor faults are low-probability extreme events, their characteristic amplitudes or residuals usually exhibit extreme values, which cannot be effectively captured by simply using the mean and standard deviation of the overall sample. Therefore, a heuristic method of "mean plus a fixed multiple of standard deviation" can be used, combined with QQ plot verification to ensure that the over-threshold samples conform to the assumption of a generalized Pareto distribution. All samples exceeding the threshold are selected as over-threshold samples. These samples directly reflect the distribution pattern of extreme fluctuations under normal conditions, providing extreme value data for subsequent fitting.

[0125] For example, a generalized Pareto distribution can be fitted to the over-threshold amplitude samples and over-threshold residual samples in each dimension, respectively. The generalized Pareto distribution is a core model in extreme value theory used to describe the distribution of over-threshold extreme values, and its probability density function is: , The shape parameter reflects the tail characteristics of the distribution; The scale parameter reflects the fluctuation range of extreme values. Illustratively, by constructing the likelihood function of over-threshold samples, the shape parameter can be... With scale parameters The partial derivatives are calculated and set to zero, and the optimal estimates of the parameters are obtained through numerical iteration. Furthermore, for automotive fault monitoring scenarios, the shape parameters... The value is usually greater than 0, indicating that the extreme values ​​follow a heavy-tailed distribution, which is consistent with the characteristics of low probability and high impact of fault events.

[0126] Specifically, based on the preset confidence level and distribution parameters of each dimension, the dynamic amplitude warning threshold can be calculated, and its calculation formula is as follows:

[0127]

[0128] in, Indicates the first Dimensional feature magnitude at confidence level The dynamic early warning threshold is below; For the first time under normal historical conditions The mean of the feature amplitude samples of a dimension represents the average activation intensity of that dimension in its normal state. For the first time under normal historical conditions The standard deviation of the amplitude samples of a dimension characteristic represents the fluctuation range of the normal state of that dimension; The quantile function of the generalized Pareto distribution, i.e., given a confidence level. Shape parameters Scale parameters At that time, the corresponding extreme value quantiles reflect the critical level of extreme fluctuations under normal conditions; For the first Historical sample size of dimensional feature magnitude; This is the logarithmic adjustment coefficient, used to balance the influence of sample size and distribution tails. The larger the sample size, the higher the weight of the logarithmic term, to avoid excessive threshold fluctuations under small sample sizes.

[0129] In the specific calculation, the extreme value quantiles corresponding to the confidence level can be obtained first through the quantile function of the generalized Pareto distribution, and then divided by the square root of the sample size. The effect of standardized sample size on quantiles, plus a logarithmic adjustment term. To balance the stability of the tail estimate, the result is then multiplied by the standard deviation of that dimension. Scale to the original feature scale and finally add the mean. The final dynamic amplitude warning threshold can be obtained. The calculation logic of the dynamic residual warning threshold is exactly the same as that of the dynamic amplitude warning threshold; only the input parameters need to be replaced with the mean, standard deviation, distribution parameters, and sample size of the residual moving average sample. Furthermore, this dynamic threshold has adaptive characteristics, meaning that when historical normal data is updated, the mean... Standard deviation Shape parameters Scale parameters The threshold will be re-estimated and dynamically adjusted accordingly to ensure that the threshold always matches the distribution characteristics of the current data. This effectively solves the problems of poor adaptability and susceptibility to operating conditions and aging of traditional fixed thresholds, and significantly improves the accuracy and reliability of fault warning.

[0130] In one embodiment, the graph neural network diagnostic model includes an embedding layer, a graph attention layer, a classification head, and a regression head;

[0131] Based on the warning time in the warning signal, fault feature subsequences corresponding to the time window are extracted from the sparse fault feature vector sequence; the fault feature subsequences are processed for fault diagnosis using a pre-trained graph neural network diagnostic model, and a diagnostic report is output, including:

[0132] Based on the warning time, a preset diagnostic window range is determined, and feature data within the preset diagnostic window range is extracted from the sparse fault feature vector sequence to obtain a fault feature subsequence; a graph structure including component nodes and connection edges is constructed based on the preset physical connection relationship of the vehicle system and the causal influence law of components.

[0133] The fault feature subsequence is input into the pre-trained graph neural network diagnostic model. The dimension mapping of the fault feature subsequence is performed through the embedding layer to obtain node embedding features. The node embedding features are assigned to the corresponding component nodes in the graph structure as the initial node features of each component node.

[0134] Multi-round message passing is performed based on the graph attention layer, the feature information of adjacent nodes of each component node is aggregated, and the feature representation of each component node is updated to obtain the updated feature representation of each component node; the failure probability distribution of each component is obtained by performing probability calculation on the updated feature representation of each component node through the classification head.

[0135] The initial severity score of each component is obtained by numerically regressing the feature representation of each component node after the update using a regression head; the maximum value of the residual moving average sequence within the warning time corresponding to the warning signal is extracted, and the final severity score of each component is calculated based on the maximum value and the initial severity score of each component; a diagnostic report is generated based on the failure probability distribution of each component and the final severity score.

[0136] Specifically, the warning time corresponds to the moment when the fault characteristics first significantly deviate from the normal pattern. The diagnostic window can be centered on the warning time, extending by the same time step before and after it. The window length can be set according to the dynamic response characteristics of the vehicle system. For example, for the powertrain system, the window length needs to cover the propagation time of the fault from the component to the associated sensor to avoid missing key fault evolution information. The extracted fault feature subsequence can focus on the core features of the fault occurrence period and eliminate redundant information in non-fault periods. When constructing the graph structure, the nodes of the graph structure can be defined as key components of the vehicle (such as the engine, transmission, brake master cylinder, various monitoring sensors, etc.), and the connecting edges are defined as physical connections or direct causal dependencies between components. The directed edges are used to represent the causal direction. For example, an engine fault will directly affect the output of the cylinder pressure sensor. Therefore, a directed edge can be constructed from the engine node to the cylinder pressure sensor node. The weight of the edge is preset based on domain knowledge to represent the strength of the influence. For example, the weight of a strong causal relationship is set to a higher value, and the weight of a weak relationship is set to a lower value. The graph structure is constructed in accordance with the physical design and system working principle of the vehicle, which can ensure that the direction and weight of the edges can truly reflect the causal transmission relationship between components, providing a topological foundation that conforms to physical laws for the message transmission of the subsequent graph neural network.

[0137] Subsequently, the fault feature subsequence can be input into a pre-trained graph neural network diagnostic model. Its embedding layer can employ a combination of 1D convolutional layers and fully connected layers. The 1D convolutional layers capture the local temporal correlation information of the fault feature subsequence, while the fully connected layers map the convolutional output to an embedding space consistent with the feature dimension of the graph nodes. The activation function can be ReLU to introduce non-linear mapping capabilities. For example, the fault feature subsequence is input into the embedding layer step-by-step. After multi-scale temporal features are extracted through convolutional operations, they are compressed into a dense vector of fixed dimensions by the fully connected layer. This vector is the node embedding feature, and its dimension matches the feature dimension of the component nodes in the graph. Furthermore, the node embedding features can be assigned to the corresponding component nodes in the graph structure as the initial node features for each component node. For example, fault feature subsequences related to the engine are mapped to the engine node, allowing the initial features of each node to directly associate with its corresponding fault information. Multi-round message passing can be performed based on the graph attention layer. That is, by calculating the attention weights between nodes, the features of neighboring nodes are dynamically aggregated to capture the propagation path of the fault between components. The formula for calculating the attention weights is:

[0138]

[0139]

[0140] in, For nodes with neighboring nodes Attention score For learnable attention vectors, The feature transformation weight matrix, , They are nodes , The current feature vector, This is a vector concatenation operation. The attention weights are normalized. For nodes The set of neighboring nodes. During multiple rounds of message passing, each node can aggregate the features of its neighboring nodes based on attention weights and update its own feature representation. For example, when the engine node has a fault, its neighboring nodes (such as vibration sensor and cylinder pressure sensor nodes) will aggregate the fault features of the engine node through attention weights, so that their own features contain fault propagation information. After 2-3 rounds of message passing, the feature representation of each node will simultaneously contain its own fault information and the propagation information of its neighbors, providing a key basis for root fault localization.

[0141] Specifically, the classification head can adopt a structure of fully connected layer + Softmax activation function. The fully connected layer maps node features to the fault category dimension, and the Softmax function normalizes the output value into a probability distribution. The final output fault probability distribution directly reflects the probability that each component node is the root fault, with the node with the highest probability value being the most likely root fault component. Furthermore, a regression head can be used to perform numerical regression on the updated feature representations of each component node. The regression head can adopt a structure of fully connected layer + Sigmoid activation function. The fully connected layer maps node features to a single dimension, and the Sigmoid function compresses the output value to the [0,1] interval, obtaining the initial severity score for the corresponding component node. This initial severity score quantifies the fault severity based on the fault correlation strength of node features; the closer the value is to 1, the more severe the fault. Subsequently, the maximum value of the residual moving average sequence within the warning time corresponding to the warning signal can be extracted. This value reflects the degree to which the overall system deviates from the healthy baseline during the fault occurrence period. This value is then linearly fused with the initial severity score to calculate the final severity score. This fusion calculation combines the fault intensity of local nodes with the overall deviation of the system, making the severity assessment more comprehensive and accurate.

[0142] Specifically, diagnostic reports can be generated based on the failure probability distribution and final severity score of each component. The report content can include the root fault component (the node with the highest failure probability), the failure probability value, the final severity score, the failure propagation path (a sequence of highly correlated nodes based on graph attention weights), and repair recommendations. For example, when the engine node has the highest failure probability and a high severity score, the diagnostic report can clearly state that "the engine injector is the root fault component, with a failure probability of 0.92 and a severity score of 0.78. The fault has propagated to the vibration sensor and cylinder pressure sensor. It is recommended to prioritize checking the injector's blockage." This report provides a basis for vehicle repair decisions, achieving a closed loop from anomaly detection to actionable repair recommendations.

[0143] Based on the same inventive concept, this application also provides a vehicle health status monitoring device for implementing the vehicle health status monitoring method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of the one or more vehicle health status monitoring device embodiments provided below can be found in the limitations of the vehicle health status monitoring method described above, and will not be repeated here. Figure 3 As shown, the device 300 includes:

[0144] The working condition perception normalization module 301 is used to preprocess the raw vehicle multi-sensor time-series data, including working condition parameters and monitoring parameters, to obtain preprocessed multivariate time window data; and to perform working condition stripping processing on the preprocessed multivariate time window data through a pre-trained working condition perception attention coding network to generate a working condition normalized feature sequence.

[0145] The baseline construction and fault extraction module 302 is used to generate a dynamic health benchmark sequence by performing health benchmark generation processing on the normalized feature sequence of the operating conditions through a pre-trained health benchmark generator; to perform fault feature extraction processing on the normalized feature sequence of the operating conditions through a pre-trained causal fault extractor to obtain fault features, and to generate a sparse fault feature vector sequence by combining the dynamic health benchmark sequence; and to calculate the reconstructed residual sequence based on the normalized feature sequence of the operating conditions and the dynamic health benchmark sequence.

[0146] The dynamic early warning judgment module 303 is used to calculate the absolute value of each dimension of the sparse fault feature vector sequence to obtain the feature amplitude sequence of each dimension; based on the pre-collected historical normal data, the extreme value theory is used to perform distribution fitting processing on the feature amplitude sequence of each dimension and the reconstructed residual sequence respectively to calculate the corresponding dynamic amplitude early warning threshold and dynamic residual early warning threshold; the real-time feature amplitude sequence of each dimension is compared with the corresponding dynamic amplitude early warning threshold to obtain the first comparison result, and the reconstructed residual sequence is compared with the corresponding dynamic residual early warning threshold to obtain the second comparison result;

[0147] The intelligent diagnostic reasoning module 304 is used to generate a warning signal containing the warning time and the trigger feature dimension when the first comparison result and / or the second comparison result meet the preset trigger conditions; based on the warning time in the warning signal, extract the fault feature subsequence corresponding to the time window from the sparse fault feature vector sequence; perform fault diagnosis processing on the fault feature subsequence through a pre-trained graph neural network diagnostic model, and output a diagnostic report.

[0148] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the vehicle health monitoring method of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0149] In one exemplary embodiment, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the vehicle health monitoring method of the present application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0150] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for monitoring the health status of a vehicle, characterized in that, The method includes: The original vehicle multi-sensor time-series data, including operating condition parameters and monitoring parameters, are preprocessed to obtain preprocessed multivariate time window data. The preprocessed multivariate time window data is then subjected to operating condition stripping processing through a pre-trained operating condition perception attention coding network to generate an operating condition normalized feature sequence. A pre-trained health benchmark generator is used to process the normalized feature sequence of the operating conditions to generate a dynamic health benchmark sequence. A pre-trained causal fault extractor is used to extract fault features from the normalized feature sequence of the operating conditions to obtain fault features. These features are then combined with the dynamic health benchmark sequence to generate a sparse fault feature vector sequence. Based on the normalized feature sequence of the operating conditions and the dynamic health benchmark sequence, a reconstructed residual sequence is calculated. The absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the feature amplitude sequence of each dimension. Based on pre-collected historical normal data, the distribution fitting process of the feature amplitude sequence of each dimension and the reconstructed residual sequence is performed using extreme value theory to calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold. The real-time obtained feature amplitude sequence of each dimension is compared with the corresponding dynamic amplitude warning threshold to obtain a first comparison result, and the reconstructed residual sequence is compared with the corresponding dynamic residual warning threshold to obtain a second comparison result. When the first comparison result and / or the second comparison result meet the preset triggering conditions, a warning signal containing the warning time and the triggering feature dimension is generated; based on the warning time in the warning signal, the corresponding time window of the fault feature subsequence is extracted from the sparse fault feature vector sequence; the fault feature subsequence is processed by a pre-trained graph neural network diagnostic model to perform fault diagnosis and output a diagnostic report.

2. The method according to claim 1, characterized in that, The pre-trained working condition awareness attention coding network includes a working condition feature embedding layer, a monitoring feature embedding layer, a working condition attention module, a condition-gated recurrent unit, and a decoding layer. The raw vehicle multi-sensor time-series data, including operating parameters and monitoring parameters, are preprocessed to obtain preprocessed multivariate time window data. The pre-trained working condition-aware attention encoding network is used to perform working condition stripping processing on the preprocessed multivariate time window data to generate a working condition normalized feature sequence, including: The original vehicle multi-sensor time-series data is preprocessed to obtain preprocessed data. The preprocessed data is then divided into sliding window segments according to a preset window length and a preset overlap rate to obtain initial time window data. Channel normalization is then performed on each sensor channel of the initial time window data to obtain the preprocessed multivariate time window data. The preprocessed multivariate time window data is separated into operating condition indication features and health monitoring features in the channel dimension; the operating condition indication features are extracted through the operating condition feature embedding layer to obtain operating condition embedded features; the health monitoring features are extracted through the monitoring feature embedding layer to obtain monitoring embedded features. The working condition attention module calculates the attention weight of each working condition embedding feature to each monitoring embedding feature at each time step based on the working condition embedding feature and the monitoring embedding feature, and generates a working condition context vector. Based on the condition-gated loop unit, the monitoring embedded features and the operating condition context vector are fused, and an operating condition stripping modulation operation is applied to obtain the operating condition modulation features; the operating condition modulation features are then subjected to time-series encoding to obtain the final hidden state. The final hidden state is decoded by the decoding layer to generate the normalized feature sequence of the working condition.

3. The method according to claim 1, characterized in that, The health baseline generator includes an encoding layer and a decoding layer; the causal fault extractor includes an initialization layer, a residual attention layer, and a causal encoding layer. The pre-trained health benchmark generator performs health benchmark generation processing on the normalized feature sequence of the working conditions to generate a dynamic health benchmark sequence. A pre-trained causal fault extractor is used to extract fault features from the normalized feature sequence of the operating conditions to obtain fault features, and combined with the dynamic health benchmark sequence to generate a sparse fault feature vector sequence. Based on the normalized feature sequence of the operating conditions and the dynamic health benchmark sequence, a reconstructed residual sequence is calculated, including: The normalized feature sequence of the working condition is input into the health benchmark generator, and the normalized feature sequence of the working condition is compressed and encoded through the coding layer to obtain the potential health variables. Collect vehicle cumulative running time data, perform location encoding on the vehicle cumulative running time data to obtain corresponding location encoding features; fuse the location encoding features with the health latent variables to obtain conditional health latent variables; The conditional health latent variables are decoded and reconstructed through the decoding layer to generate the dynamic health benchmark sequence. The normalized feature sequence of operating conditions and the dynamic health baseline sequence are input into the causal fault extractor. The element-wise difference between the normalized feature sequence of operating conditions and the dynamic health baseline sequence is calculated through the initial layer to obtain the initial residual features. The initial residual features are weighted through the residual attention layer to obtain the residual attention weights. The normalized feature sequence of operating conditions is weighted and focused using the residual attention weights to obtain the weighted focused features. Based on the causal coding layer, feature extraction is performed on the weighted focusing features to obtain the fault features; the fault features are processed by a sparse activation function to generate the sparse fault feature vector sequence; the Euclidean distance between the normalized feature sequence and the dynamic health benchmark sequence at each time step is calculated to obtain the reconstructed residual sequence.

4. The method according to claim 1, characterized in that, The absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the feature amplitude sequence of each dimension. Based on pre-collected historical normal data, extreme value theory is used to perform distribution fitting processing on the feature amplitude sequences of each dimension and the reconstructed residual sequence to calculate the corresponding dynamic amplitude warning threshold and dynamic residual warning threshold, including: The absolute values ​​of each dimension of the sparse fault feature vector sequence are calculated to obtain the initial feature magnitude of each dimension; the initial feature magnitudes of each dimension are then subjected to moving average processing to obtain the feature magnitude sequence of each dimension. The reconstructed residual sequence is processed by moving average to obtain a residual moving average sequence; Based on the previously collected historical normal data, extract the feature amplitude samples and residual moving average samples of each dimension under the historical normal state. Using the over-threshold model in the extreme value theory, the feature amplitude samples and residual moving average samples of each dimension under the historical normal state are screened to obtain over-threshold amplitude samples and over-threshold residual samples of each dimension. A generalized Pareto distribution is fitted to the over-threshold amplitude samples and over-threshold residual samples of each dimension, and the distribution parameters and residual distribution parameters of each dimension are calculated by maximum likelihood estimation. Based on the preset confidence level and the distribution parameters of each dimension, the dynamic amplitude warning threshold is calculated; based on the preset confidence level and the residual distribution parameters, the dynamic residual warning threshold is calculated.

5. The method according to claim 1, characterized in that, The graph neural network diagnostic model includes an embedding layer, a graph attention layer, a classification head, and a regression head; Based on the warning time in the warning signal, the fault feature subsequence corresponding to the time window is extracted from the sparse fault feature vector sequence; The fault feature subsequence is processed for fault diagnosis using a pre-trained graph neural network diagnostic model, and a diagnostic report is output, including: Based on the warning time, a preset diagnostic window range is determined, and feature data within the preset diagnostic window range is extracted from the sparse fault feature vector sequence to obtain the fault feature subsequence; Based on the preset physical connection relationships of the vehicle system and the causal influence rules of components, a graph structure including component nodes and connection edges is constructed. The fault feature subsequence is input into the pre-trained graph neural network diagnostic model, and the dimension mapping process is performed on the fault feature subsequence through the embedding layer to obtain node embedding features; the node embedding features are assigned to the corresponding component nodes in the graph structure as the initial node features of each component node; Based on the graph attention layer, multiple rounds of message passing are performed, the feature information of adjacent nodes of each component node is aggregated, and the feature representation of each component node is updated to obtain the updated feature representation of each component node; the classification head is used to perform probability calculation on the updated feature representation of each component node to obtain the failure probability distribution of each component. The initial severity score of each component is obtained by performing numerical regression on the feature representation of each updated component node using the regression head. The maximum value of the residual moving average sequence within the warning time corresponding to the warning signal is extracted. Based on the maximum value and the initial severity score of each component, the final severity score of each component is calculated. Based on the failure probability distribution of each component and the final severity score, the diagnostic report is generated.

6. The method according to claim 3, characterized in that, The causal fault extractor is trained through the following steps: S1. Collect vehicle time-series data under historical normal conditions and generate a normalized feature sequence for training conditions; collect historical vehicle cumulative running time data corresponding to the vehicle time-series data, perform position encoding on the historical vehicle cumulative running time data, and obtain position encoding features for training. S2, Based on the physical structure and working coupling relationship of each preset key component of the vehicle, construct a priori causal graph to describe the direct causal relationship between components, and define a set of variables directly affected by the fault and a set of variables indirectly affected by the fault for each component node based on the priori causal graph; S3, Construct an adversarial discriminator, the input of which is a joint sample of fault feature dimension and variable feature, and the output is the distribution discrimination result of the joint sample; S4, input the training condition normalized feature sequence and the training position encoding feature into the health benchmark generator, and output the training dynamic health benchmark sequence; S5, inject simulated fault features of different components into the normalized feature sequence of the training conditions to obtain multiple sets of fault condition feature sequences; wherein, each set of fault condition feature sequences corresponds to a single type of component fault. S6. Initialize the network parameters of the initial layer, residual attention layer and causal coding layer of the causal fault extractor to obtain the initial causal fault extractor. Input the fault condition feature sequence of each group and the corresponding training dynamic health benchmark sequence into the initial causal fault extractor to obtain training fault features. Apply sparse activation function processing to the training fault features to obtain the training sparse fault feature vector sequence. S7. Extract the fault feature dimension corresponding to the component injected with the simulated fault feature from the training sparse fault feature vector sequence, and extract the features of the variable set of indirect influence of the fault of the corresponding component from the fault condition feature sequence to obtain the training indirect influence variable features. S8, combine the fault feature dimension with the training indirect influence variable feature to obtain a true joint sample; randomly shuffle the training indirect influence variable feature and combine it with the fault feature dimension to obtain a false joint sample; input the true joint sample and the false joint sample into the adversarial discriminator, with the goal of maximizing the discrimination error rate of the adversarial discriminator, train the adversarial discriminator, and backpropagate to adjust the network parameters of the causal fault extractor to minimize the mutual information between the fault feature dimension and the training indirect influence variable feature; S9, iterate through S5 to S8 until the loss function of the causal fault extractor converges and the discrimination accuracy of the adversarial discriminator meets the preset threshold, then stop training and obtain the trained causal fault extractor.

7. The method according to claim 4, characterized in that, The dynamic amplitude warning threshold is calculated using the following formula: in, For the first The dynamic amplitude warning threshold of the dimensional feature amplitude, Confidence level; , These are the first under normal historical conditions. Mean and standard deviation of the amplitude samples of dimensional features; Let be the quantile function of the generalized Pareto distribution. For shape parameters, For scale parameters; For the first Historical sample size of dimensional feature magnitude; This is the logarithmic adjustment coefficient.

8. A vehicle health status monitoring device, characterized in that, The device includes: The condition perception normalization module is used to preprocess the raw vehicle multi-sensor time-series data, including condition parameters and monitoring parameters, to obtain preprocessed multivariate time window data; and to perform condition stripping processing on the preprocessed multivariate time window data through a pre-trained condition perception attention coding network to generate a condition normalized feature sequence. The baseline construction and fault extraction module is used to generate a dynamic health benchmark sequence by performing health benchmark generation processing on the normalized feature sequence of the operating conditions using a pre-trained health benchmark generator; to perform fault feature extraction processing on the normalized feature sequence of the operating conditions using a pre-trained causal fault extractor to obtain fault features, and to generate a sparse fault feature vector sequence by combining the dynamic health benchmark sequence; and to calculate the reconstructed residual sequence based on the normalized feature sequence of the operating conditions and the dynamic health benchmark sequence. The dynamic early warning determination module is used to calculate the absolute value of each dimension of the sparse fault feature vector sequence to obtain the feature amplitude sequence of each dimension; based on pre-collected historical normal data, the module uses extreme value theory to perform distribution fitting processing on the feature amplitude sequence of each dimension and the reconstructed residual sequence to calculate the corresponding dynamic amplitude early warning threshold and dynamic residual early warning threshold; the module compares the real-time obtained feature amplitude sequence of each dimension with the corresponding dynamic amplitude early warning threshold to obtain a first comparison result, and compares the reconstructed residual sequence with the corresponding dynamic residual early warning threshold to obtain a second comparison result; The intelligent diagnostic reasoning module is used to generate a warning signal containing a warning time and a trigger feature dimension when the first comparison result and / or the second comparison result meet a preset trigger condition; based on the warning time in the warning signal, extract the fault feature subsequence corresponding to the time window from the sparse fault feature vector sequence; perform fault diagnosis processing on the fault feature subsequence through a pre-trained graph neural network diagnostic model, and output a diagnostic report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images