Vehicle fault identification methods, devices and vehicles

Abstract

This invention relates to the field of vehicle fault identification technology, and discloses a vehicle fault identification method, device, and vehicle. The method includes: inputting runtime timing data of an electric drive system into a pre-trained reconstruction function model to obtain reconstructed time-series data, wherein the reconstruction function model is obtained through unsupervised pre-training based on normal data; determining the fault representation error between the runtime timing data and the reconstructed time-series data; and triggering an anomaly warning if the fault representation error is greater than a dynamic threshold. By applying the technical solution of this invention, direct reconstruction using an unsupervised pre-trained reconstruction function model avoids the accumulation of errors in time-series prediction, thereby improving the accuracy of vehicle fault identification.

Description

Technical Field

[0001] This invention relates to the field of vehicle fault identification technology, specifically to a vehicle fault identification method, device, and vehicle. Background Technology

[0002] In vehicle fault prediction scenarios, in order to address the problem of scarce effective fault samples that can be collected from newly launched models and newly put into use vehicles, the relevant technologies use normal state data to train models, predict normal data at future moments through models, and compare the predicted results with the actual results to determine whether a fault has occurred.

[0003] However, current processing methods require models to have strong time series prediction capabilities to accurately simulate various possible future operating conditions and states. Due to the complex and ever-changing automotive operating environment, errors are inevitable in the prediction process. These errors accumulate as the prediction time increases, ultimately affecting the accuracy of fault diagnosis. Summary of the Invention

[0004] In view of the above problems, embodiments of the present invention provide a vehicle fault identification method to solve the problem of low fault judgment accuracy caused by the accumulation of prediction errors in the prior art.

[0005] According to one aspect of the present invention, a vehicle fault identification method is provided, comprising: The runtime timing data of the electric drive system is input into a pre-trained reconstruction function model to obtain reconstructed timing data, wherein the reconstruction function model is obtained by unsupervised pre-training based on normal data; Determine the fault characterization error between the runtime timing data and the reconstructed timing data; If the fault characterization error is greater than the dynamic threshold, an anomaly warning is triggered.

[0006] According to another aspect of the present invention, a vehicle fault identification device is provided, comprising: The reconfiguration module is used to input the runtime timing data of the electric drive system into a pre-trained reconfiguration function model to obtain reconfigured timing data, wherein the reconfiguration function model is obtained by unsupervised pre-training based on normal data; The detection module is used to determine the fault characterization error between the runtime timing data and the reconstructed timing data; The execution module is used to trigger an anomaly warning if the fault characterization error exceeds a dynamic threshold. According to another aspect of the present invention, a vehicle is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; the memory is used to store at least one executable instruction, wherein the executable instruction causes the processor to perform the operation of the vehicle fault identification method as described above.

[0007] This invention, through its embodiment, inputs the runtime timing data of the electric drive system into a pre-trained reconstruction function model based on normal data in an unsupervised manner to obtain corresponding reconstructed timing data. It then determines the fault representation error between the runtime timing data and the reconstructed timing data, and triggers an anomaly warning when the fault representation error exceeds a dynamic threshold. This leverages the unsupervised learning capability of the reconstruction function model to understand the inherent characteristics and distribution patterns of the electric drive system's normal operating data, directly reconstructing the input real-time runtime timing data. This avoids the technical shortcomings of high algorithm complexity and error accumulation amplification during multi-step prediction in the timing prediction process, simplifying the technical path for identifying abnormal operating states of the electric drive system. Furthermore, by comparing the error between the reconstructed data and the original real-time operating data, it can accurately capture subtle anomalies that deviate from the normal operating state of the electric drive system, achieving efficient, accurate, and robust identification of electric drive system faults. This effectively improves the timeliness and reliability of fault warnings for vehicle electric drive systems.

[0008] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0009] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a first embodiment of the vehicle fault identification method provided by the present invention is shown. Figure 2 This invention provides a schematic diagram of the overall process of the vehicle fault identification method. Figure 3 A flowchart illustrating a second embodiment of the vehicle fault identification method provided by the present invention is shown. Figure 4 A flowchart of the fault classifier for the vehicle fault identification method provided by the present invention is shown. Figure 5 A schematic diagram of the implementation process of the fourth embodiment of the vehicle fault identification method provided by the present invention is shown; Figure 6 A schematic diagram illustrating the training process of the reconstruction function model of the vehicle fault identification method provided by the present invention is shown. Figure 7 A schematic diagram of the structure of a first embodiment of the vehicle fault identification device provided by the present invention is shown. Figure 8 A structural schematic diagram of an embodiment of the vehicle provided by the present invention is shown. Detailed Implementation

[0010] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0011] Firstly, Figure 1 A flowchart of a first embodiment of the vehicle fault identification method of the present invention is shown, the method being performed by a vehicle. Figure 1 As shown, the method includes the following steps: Step S10: Input the runtime timing data of the electric drive system into the pre-trained reconstruction function model to obtain the reconstructed timing data.

[0012] The reconstruction function model is obtained through unsupervised pre-training based on normal data.

[0013] In this embodiment, the runtime sequence data of the electric drive system is a sequence of multi-dimensional operating state parameters with synchronization timestamps, continuously collected at fixed time sampling periods during the operation of the vehicle's electric drive system. This includes time-series data collected by various sensors (at least current sensors, voltage sensors, temperature sensors, etc.) related to the electric drive system's operating state and parameters. The reconstruction function model is built on an unsupervised learning framework. Using the normal operating time-series data of the electric drive system under fault-free conditions as training samples, it learns the inherent characteristics, temporal correlations, and data distribution patterns of normal data through unsupervised pre-training. This model is a deep learning model capable of restoring and reconstructing the homologous features of the input time-series data. It achieves low-error reconstruction of input data conforming to a normal distribution and high-error reconstruction of abnormal data deviating from the normal distribution. The reconstructed time-series data is the sequence of time-series data that is completely consistent with the dimensions and temporal length of the input data after the reconstruction function model receives the real-time runtime sequence data and performs feature restoration based on the learned normal data feature patterns. Understandably, unsupervised pre-training is a training method where the model is trained using only the normal operating time-series data of an electric drive system without fault labels, without the need for manually labeled fault samples. The model learns the distribution boundaries of normal data rather than learning specific fault characteristics through its own feature learning and reconstruction error optimization. The runtime time-series data is a variable, multi-dimensional time series.

[0014] In this embodiment, feature restoration and reconstruction output of the runtime sequence data of the electric drive system can be achieved through variable-length sequence adaptation. Therefore, the reconstructed time sequence data corresponding to the currently input runtime sequence data can be determined by the reconstruction function model trained based on variable-length sequence adaptation. The inference methods of the unsupervised reconstruction model based on variable-length sequence adaptation include fixed-length preprocessing reconstruction based on stacked autoencoders and personalized padding mask reconstruction based on Transformer encoding and decoding.

[0015] As an optional implementation, before inputting the data into the reconstruction function model, the variable-length runtime sequence data can be padded and masked based on statistical features to unify it into a fixed-length sequence. Then, the reconstruction function model based on the Transformer encoder-decoder architecture performs feature extraction and reconstruction, outputting the reconstructed time-series data. Specifically, the reconstruction function model, based on this encoder-decoder architecture, compresses the input into a context vector using a Transformer encoder, and the decoder reconstructs data with the same structure as the input sequence based on this vector, thus obtaining the reconstructed time-series data.

[0016] Therefore, after obtaining variable-length runtime sequence data, the preset sequence length corresponding to the runtime sequence data can be determined. The missing sequences in the runtime sequence data are filled based on pre-stored sensor statistical features, while a mask matrix identifies the real data and the filled data. By padding all variable-length sequences to the longest sequence length and using a mask to distinguish between real and filled data, the input requirements of the Transformer are adapted, while the attention mechanism reduces its focus on the filled data. Subsequently, the filled input sequence is compressed into a context vector through the encoding layer of the reconstruction function model, and the context vector is reconstructed into the reconstructed time-series data with the same length and dimension as the input sequence based on the decoding layer. The vehicle control system can collect historical normal operating data from each sensor channel under different operating conditions (idle, constant speed, acceleration, etc.). Each sensor corresponds to one data channel. For the historical data of each sensor channel, its mean (μ), standard deviation (σ), median (M), and quantile intervals are calculated, thereby establishing a personalized feature library at the sensor dimension and storing it in the storage unit.

[0017] Specifically, during the personalized padding and mask matrix generation process for variable-length sequences, after acquiring the currently collected variable-length runtime sequence data, the longest sequence length for this inference is first determined, such as taking the longest sequence length in the current batch or a preset fixed longest length. For positions with insufficient sequence length, for each sensor channel, the current vehicle operating condition is first identified. By retrieving the mean (μ) and standard deviation (σ) of the sensor under historical operating conditions, the padding value is calculated as: padding value = historical mean of the sensor + random noise. The amplitude of the random noise is constrained by the historical standard deviation (e.g., the noise follows a normal distribution with a mean of 0 and a standard deviation of 0.3σ) to ensure that the data distribution in the padding area is consistent with the real data. Simultaneously, a mask matrix that perfectly matches the dimension and length of the input sequence needs to be generated. The real data positions are marked as 1, and the padding data positions are marked as 0, so that the Transformer attention mechanism can distinguish between the real observations and the padding content, reducing the interference of the padding data on feature learning. Finally, the reconstruction inference operation of the Transformer encoder-decoder architecture is executed in the reconstruction model. The Transformer's self-attention mechanism is used to capture long-distance temporal dependencies, and data reconstruction is completed through encoding and decoding, outputting reconstructed data with the same structure as the input sequence. During the encoding and decoding process, the padded variable-length input sequence and the corresponding mask matrix are input into the pre-trained unsupervised Transformer encoder-decoder reconstruction function model. In this process, the encoder, through a multi-head self-attention mechanism combined with the mask matrix, compresses the input sequence into a highly abstract context vector; the decoder receives this context vector and, through self-attention and encoder-decoder attention mechanisms, reconstructs reconstructed temporal data with the same length and dimension as the input sequence. The mask matrix is ​​used to mask the padded positions during attention calculation, ensuring that attention focuses only on the real data.

[0018] For example, a variable-length sequence of 250 sampling points is currently being collected, including three sensor channels such as motor d / q-axis current and bus voltage. The preset maximum sequence length is 400 sampling points, and the vehicle is currently in a climbing condition. Based on this, the average values ​​of the motor d-axis current (μ_d=20A, σ_d=3A), the average values ​​of the q-axis current (μ_q=10A, σ_q=2A), and the average values ​​of the bus voltage (μ_v=350V, σ_v=10V) under this condition are retrieved from the personalized feature library. For the missing 150 sampling points, the motor d-axis current is padded with 20A plus N(0, 0.9). 2） Random noise, q-axis current is 10A plus N(0, 0.6) 2 Random noise, with a bus voltage of 350V plus N(0, 3) 2）Random noise is introduced. A mask matrix is ​​generated, with the first 250 positions set to 1 and the last 150 positions set to 0. Finally, the padded sequence and the mask matrix are input into the Transformer encoder-decoder reconstruction function model. The encoder compresses the data into a context vector, and the decoder reconstructs 400×3 dimensional temporal data.

[0019] It should be noted that, in addition to Transformer encoder-decoder or encoder structures, the reconstruction function model can also be replaced by other deep learning models for processing temporal data, such as hybrid models of Temporal Convolutional Networks (TCNs), Long Short-Term Memory Networks (LSTMs), and attention mechanisms. Furthermore, besides directly using the original sensor signals, manually extracted frequency domain features, such as spectral energy and signal envelopes of specific frequency bands, can be added and concatenated with the original signal before being input into the model to improve model performance.

[0020] As an alternative implementation, reconstruction processing can be performed based on a fixed-length preprocessing reconstruction model using a stacked autoencoder. Similarly, variable-length runtime sequence data can first be padded and masked based on statistical features to unify it into a fixed-length sequence, and then feature extraction and restoration can be performed using a reconstruction function model based on a stacked autoencoder architecture to output reconstructed time-series data.

[0021] Specifically, the vehicle's electric drive domain controller collects historical normal operating data from each sensor channel under different operating conditions (idle, constant speed, acceleration, etc.). Each sensor corresponds to one data channel, including motor three-phase current, IGBT (Insulated Gate Bipolar Transistor) junction temperature, motor vibration acceleration, etc. For the historical data of each sensor channel, its mean (μ), standard deviation (σ), median (M), and other statistical characteristics are calculated to establish a personalized feature library for the sensor dimension and store it in the storage unit. After acquiring variable-length runtime sequence data, for the missing positions of each sensor channel, the mean (μ) and standard deviation (σ) of the sensor's historical data under the same operating conditions are retrieved from the statistical feature library, and the fill value is calculated according to the formula: fill value = historical mean of the sensor + random noise. During the filling process, a mask matrix with the same length as the sequence is generated, with the real data position marked as 1 and the filled data position marked as 0. Finally, the reconstruction inference process of the stacked autoencoder is executed. The padded fixed-length runtime sequence data and the mask matrix are input into the pre-trained unsupervised stacked autoencoder model. The encoder extracts high-dimensional features through three one-dimensional convolutional layers, and then the decoder restores the sequence through three one-dimensional transposed convolutional layers. The final output is reconstructed temporal sequence data with the same length and dimension as the input sequence. The mask matrix is ​​only used to reduce the weight of the padded data in feature learning and does not affect the forward propagation during inference.

[0022] For example, a variable-length runtime timing data segment with 300 sampling points is currently acquired, including two sensor channels: motor three-phase current and IGBT junction temperature. The preset maximum sequence length is 500 sampling points. The mean value μ1=15A and standard deviation σ1=2A of the motor three-phase current under constant speed conditions are retrieved from a personalized feature library, along with the mean value μ2=75℃ and standard deviation σ2=3℃ of the IGBT junction temperature. For the missing 200 sampling points, the padding value for the motor three-phase current is 15A plus random noise in the range [-1A, 1A], and the padding value for the IGBT junction temperature is 75℃ plus random noise in the range [-1.5℃, 1.5℃]. A mask matrix is ​​generated, with the first 300 positions set to 1 and the last 200 positions set to 0. The padded 500×2-dimensional sequence is input into a stacked autoencoder, which outputs reconstructed timing data of the same dimension.

[0023] This embodiment directly reconstructs real-time runtime sequence data using an unsupervised pre-trained reconstruction function model, effectively avoiding the error accumulation problem in time series prediction. Furthermore, for variable-length sequences, a personalized padding method based on historical sensor statistical features, rather than the traditional fixed-value padding method, maintains the authenticity of the data distribution in the padding region. Simultaneously, a mask matrix helps the model distinguish between real and padding data, reducing padding interference. By outputting reconstructed time series data without error accumulation, accurate data is provided for subsequent error calculation, improving the accuracy of error detection.

[0024] Step S20: Determine the fault characterization error between the runtime timing data and the reconstructed timing data.

[0025] In this embodiment, the fault characterization error is a quantitative error index that accurately characterizes the deviation between the input runtime time-series data and the reconstructed time-series data. This error effectively distinguishes between normal data fluctuations and abnormal fault deviations, and its magnitude is positively correlated with the degree to which the data deviates from the normal pattern. The fault characterization error can be calculated based on a multi-dimensional enhanced reconstruction loss function, or it can be determined by combining time-series data deviation quantization methods using a mask matrix, such as time-domain weighted mean square error quantization based on the actual data location, or distribution deviation quantization based on feature domain and mask fusion.

[0026] As an optional implementation, a multi-dimensional enhancement of the reconstruction loss function can be used to quantify the deviation of time-series data. This quantification method determines the fault characterization error between the runtime time-series data and the reconstructed time-series data. This fault characterization error simultaneously covers global reconstruction bias, local abnormal fluctuations, and / or channel importance differences, including basic reconstruction loss, local sensitivity loss, and / or channel-weighted loss. Therefore, the basic reconstruction error, local sensitivity error, and channel-weighted loss between the runtime time-series data and the reconstructed time-series data can be determined first. Then, the basic reconstruction error, local sensitivity error, and channel-weighted loss are fused based on preset loss weights to obtain the fault characterization error. The basic reconstruction error refers to the global mean square error calculated from the positions of all real data, used to characterize the overall reconstruction deviation between the input sequence and the reconstructed sequence over the entire time-series length. The local sensitivity error is the maximum reconstruction error within a window calculated based on a sliding local window, used to capture brief and severe local abnormal fluctuations in the time-series sequence, enhancing the model's ability to identify sudden, short-term faults. The channel-weighted loss is a weighted reconstruction error calculated by combining the importance weights of different sensor channels. The channel weights are automatically learned through an attention mechanism and are used to amplify the bias contribution of fault-sensitive channels, thereby improving the discriminative power of fault characterization. This fault characterization error can also be calculated using two of the losses or just one of them.

[0027] Specifically, the fault characterization error is calculated as follows: Fault characterization error = λ1 × basic reconstruction error + λ2 × local sensitivity error + λ3 × channel weighted loss, where λ1, λ2, and λ3 are the weights corresponding to different losses, obtained through self-learning. The formula for calculating the basic reconstruction loss is as follows: , in, Given the input sequence, Reconstruct the sequence, where B is the data volume of each batch, T is the time length, C is the feature dimension, b is the sample number in the batch (from 1 to B), t is the time step in the sequence (from 1 to T), and c is the dimension of the feature vector (from 1 to C).

[0028] The formula for calculating local sensitivity error is as follows: , in, For the length of a local window (e.g.) This loss focuses on the maximum reconstruction error within a local window of the time series, enhancing its ability to capture short-term, sharp fluctuations. The formula for calculating the channel-weighted loss is as follows: , in, The importance weight of sensor c is automatically learned through an attention mechanism.

[0029] During the calculation process, the standardized errors of the three dimensions of global deviation, local anomaly, and channel sensitivity can be calculated separately. Then, combined with the prior knowledge of the fault type, hierarchical weights are set, and the weighted fusion is used to obtain the final fault characterization error.

[0030] For example, for the runtime timing data and reconstructed timing data with 400 sampling points and 3 sensor channels in the aforementioned steps, a local window length w=10 is preset, and the basic reconstruction loss L is calculated. base =0.021, Locally sensitive loss L local =0.082, channel-weighted loss L channel =0.035. The weights obtained by the model's self-learning are λ1=0.5, λ2=0.3, and λ3=0.2. According to the formula, the fault characterization error is calculated as 0.5×0.021+0.3×0.082+0.2×0.035=0.0421, that is, the fault characterization error is 0.0421.

[0031] As another alternative implementation, the three loss terms of the multi-dimensional enhanced reconstruction loss function can also be calculated completely based on the real data of the mask matrix.

[0032] Specifically, the original runtime timing data X and the reconstructed timing data output from the aforementioned steps can be obtained. And the corresponding mask matrix M. Then, according to the basic reconstruction loss formula, the calculation is performed only on the real data positions marked as 1 in the mask matrix. The corrected formula is: , Where M is the mask matrix. The basic reconstruction loss, containing only the contribution of real data, is calculated, effectively eliminating the interference of padding data on the global error calculation. Simultaneously, when calculating local sensitive errors, a local window length w can be preset (e.g., w=10). For each batch and each sensor channel's time series, a local window of length w slides along the time dimension. For each sliding window, the average squared reconstruction error within the window is calculated based only on the real data positions marked as 1 in the mask matrix, using the following formula: , Where t∈[1,T-w+1] is the starting position of the sliding window. Then, for each batch and each sensor channel, the maximum error value across all sliding windows is taken, and then the average value is calculated for all batches and all channels to obtain the local sensitivity error. , This enables the accurate detection of sudden local anomalies.

[0033] Then, channel attention weight learning and channel weighted loss calculation are performed. Through the channel attention module built into the reconstruction function model, the importance weight of each sensor channel is automatically learned based on historical normal data and real-time data. Then, combined with the mask matrix, the channel weighted loss is calculated only for the actual data location. The corrected formula is as follows: , This allows for differentiated weighting of different sensor channels.

[0034] This embodiment ensures the quantification capability of global reconstruction deviation through basic reconstruction error, solves the problem of global average error masking local sudden anomalies by using local sensitive error, improves the ability to capture short-term severe faults, and achieves differentiated weighting of different sensor channels by combining channel weighted loss, amplifies the deviation contribution of fault-sensitive channels, improves the distinguishability of fault characterization, thereby ensuring the accuracy of error calculation and providing a precise data foundation for dynamic threshold comparison in subsequent steps.

[0035] Step S30: If the fault characterization error is greater than the dynamic threshold, trigger an anomaly warning.

[0036] In this embodiment, the dynamic threshold is an error threshold that is adjusted in real time based on the vehicle's current operating conditions, the electric drive system's operating status, or the recognition accuracy of the current reconstruction function model. It is used to distinguish between normal data fluctuations and abnormal fault deviations, avoiding false alarms and missed alarms. The anomaly warning is an action taken by the vehicle to alert the driver or control system when an anomaly is detected in the electric drive system. Specifically, the dynamic threshold can be updated in real time based on a sliding window statistical analysis of historical fault characterization errors. For example, a heaped buffer structure can be used to maintain the reconstruction error values ​​(i.e., the fault characterization error sequence) of the most recent N time-step samples. Each error node contains an error value ek, a timestamp t, and a corresponding data segment identifier. The p-quantile is calculated within the sliding window as a candidate threshold, and a weighted moving average smoothing process is used to obtain the dynamic threshold corresponding to the current window.

[0037] When the fault characterization error parameter is greater than the dynamic threshold, an abnormal warning is output. When the fault characterization error is less than or equal to the dynamic threshold, the runtime sequence data is used as normal data for unsupervised pre-training of the reconstructed function model.

[0038] Furthermore, when the fault characterization error is greater than the dynamic threshold, the number of consecutive time windows exceeding the threshold is counted. If the fault characterization error of multiple consecutive time windows, such as three time windows, is greater than the dynamic threshold, a local anomaly warning is triggered.

[0039] For example, to help understand the implementation process of this embodiment, please refer to Figure 2By collecting time-series data from various sensors related to the electric drive's operating status and parameters during actual operation, a large dataset of normal data is obtained. This normal data is then used for unsupervised pre-training to generate a reconstruction function model. Next, new time-series data from the operating process is input, and the error between the output of the reconstruction function model and the input data is calculated. An error threshold is then used to determine whether the data is faulty. Finally, confirmed faulty data is labeled and classified, and a classifier is connected after the model to achieve fault detection and classification.

[0040] This embodiment fills in variable-length multidimensional time series based on historical statistical features of sensors, while using a mask matrix to reduce filling interference and maintain the authenticity of the data distribution. Then, it directly reconstructs the data using an unsupervised pre-trained reconstruction function model, avoiding the error accumulation problem in time series prediction. Furthermore, it eliminates the need for fault samples to train the model, addressing the pain point of sample scarcity, achieving full coverage identification of different operating conditions and fault types, and improving fault identification accuracy.

[0041] Based on any of the above embodiments, in Embodiment 2 of this application, please refer to... Figure 3 If the fault characterization error is greater than the dynamic threshold, before triggering the anomaly warning, steps S40 to S60 are also included: Step S40: Based on a preset sliding window, obtain the fault characterization error sequence of the electric drive system under normal operating conditions.

[0042] In this embodiment, the preset sliding window is the window length, which can be set according to actual needs. For example, it can be set to include a fault characterization error sequence containing the past 1000 normal data sampling periods as the calculation benchmark. It is understood that during the operation of the electric drive system, the sliding window will continuously slide forward over time, always updating the fault characterization error sequence within the window with the latest normal operation data, thereby providing real-time and effective data support for the adaptive adjustment of the dynamic threshold.

[0043] Therefore, when determining the specific parameters of the current dynamic threshold, it is necessary to obtain the fault characterization error sequence of the electric drive system under normal operating conditions within a preset sliding window, so as to select a suitable historical error sequence for calculation. During the acquisition process, taking a window length of 1000 as an example, the reconstruction error values ​​of 1000 sets of normal electric drive system samples under full operating conditions obtained from unsupervised training can be obtained and filled into a stacked circular buffer.

[0044] Understandably, during vehicle online operation, after each time window fault characterization error calculation is completed, the error value is first pre-validated. If the fault characterization error is less than or equal to the currently effective dynamic threshold, it is determined to be a valid reconstruction error under normal operating conditions of the electric drive system. The error value, corresponding timestamp, and time window identifier are then packaged into an error node and added to a heaped circular buffer. When the number of error nodes stored in the buffer reaches the preset window length of 1000, each time a new valid error node is added, the buffer automatically removes the oldest historical error node, always maintaining a fault characterization error sequence in the buffer that only retains the 1000 most recent normal operating states of the vehicle.

[0045] Step S50: Sort the fault characterization error sequence according to the preset confidence quantile, and determine the benchmark dynamic threshold based on the sorting result.

[0046] A quantile is a numerical point that is divided into 100 equal parts according to a cumulative probability distribution of an ascending sequence of values. The p-th percentile represents the percentage of values ​​in the sequence that are less than or equal to that point, and only (100-p)% of the values ​​that are greater than that point. In this embodiment, the confidence quantile is a percentile that is superimposed with the business confidence requirement. The preset confidence quantile p% represents the p% of the reconstruction error under normal operation of the electric drive system that the current benchmark dynamic threshold must cover. Only (100-p)% of extreme normal fluctuations are allowed to exceed the threshold. This percentage is the theoretically acceptable maximum false alarm rate. This preset parameter is set based on requirements. The benchmark dynamic threshold represents the statistical boundary of the reconstruction error of the electric drive system under normal operation within the sliding window. Errors exceeding this boundary can be judged as abnormal data that exceeds the normal fluctuation range.

[0047] Therefore, when determining the baseline dynamic threshold, it is necessary to sort the fault characterization error sequence based on a preset confidence quantile. The sorted result is usually an ascending order result. Based on this result, the quantile is calculated to obtain the baseline dynamic threshold. Specifically, outliers exceeding the normal fluctuation range in the sequence can be removed using the 3σ criterion to avoid distortion of the baseline threshold caused by occasional fluctuations under extreme normal operating conditions. Then, the cleaned fault characterization error sequence is sorted in ascending order according to the numerical values ​​to obtain an ordered error sequence. Next, based on the preset confidence quantile, the value at that quantile position in the ordered error sequence is calculated, and this value is the baseline dynamic threshold at the current moment.

[0048] Understandably, the sorting can also be performed in descending order, and the sorting results can be reversed to obtain the baseline dynamic threshold.

[0049] Step S60: Smooth the baseline dynamic threshold according to the dynamic threshold corresponding to the previous moment to obtain the dynamic threshold.

[0050] In this embodiment, the smoothing process is an average weighted smoothing process. During the smoothing process, it is necessary to set a preset smoothing factor, value range, etc., so as to balance the weight ratio of the current benchmark dynamic threshold and the historical effective threshold, avoid threshold abrupt changes caused by data fluctuations within the sliding window, and reduce misjudgments.

[0051] Therefore, the vehicle control system retrieves the dynamic threshold that took effect at the previous moment from its local storage, and smooths it using a weighted moving average formula, with the base dynamic threshold and the dynamic threshold that took effect at the previous moment as follows: , in, The dynamic threshold that takes effect at the current moment. As a preset smoothing factor, The dynamic threshold of the previous moment. This is the current benchmark dynamic threshold.

[0052] For example, during the vehicle power-on initialization phase, the electric drive system calls 1000 pre-stored offline normal reconstruction error values ​​and fills them into a stacked circular buffer of preset length N=1000, with an initial effective dynamic threshold of 0.09. During normal vehicle operation, fault characterization error calculation is completed every 2 seconds for a time window. When the current fault characterization error is 0.06, which is less than the currently effective dynamic threshold of 0.09, it is determined to be a normal valid error, and the error node is added to the buffer. The buffer automatically removes the oldest historical error node, always maintaining 1000 valid error nodes. When determining the dynamic threshold at the current moment, 1000 error values ​​need to be extracted from the buffer to form a fault characterization error sequence. This sequence is then 3σ-cleaned and sorted in ascending order. Based on a preset 99.7% confidence quantile, the value of the 997th position after sorting is calculated to be 0.092, which is the current baseline dynamic threshold of 0.092. Meanwhile, the dynamic threshold that took effect at the previous moment was 0.09. Based on the preset smoothing factor of 0.3, the dynamic threshold that takes effect at the current moment is calculated to be: 0.3×0.09 + 0.7×0.09² = 0.0914. This parameter is used for comparison and judgment of fault characterization errors.

[0053] This embodiment uses an adaptive dynamic threshold system based on sliding window quantiles to replace the fixed threshold scheme in traditional electric drive fault identification. It forms a closed loop with the unsupervised reconstruction architecture, solving the industry pain points of fixed thresholds being unable to adapt to changes throughout the vehicle's entire life cycle, having poor adaptability to operating condition fluctuations, and having high false alarm and false negative rates. The entire calculation logic is completed locally on the vehicle and does not rely on cloud communication. Even in scenarios without a network, it can continuously complete the adaptive update of the threshold, ensuring the full-scenario availability of the fault identification function.

[0054] Based on any of the above embodiments, in Embodiment 3 of this application, after determining the fault characterization error between the runtime timing data and the reconstructed timing data, steps S70~S80 are further included: Step S70: Obtain the fault judgment result determined by the cloud based on runtime sequence data.

[0055] Step S80: If the fault judgment result indicates that a fault exists and an abnormal warning is triggered, the electric drive system is determined to be in a fault state.

[0056] In this embodiment, to improve the accuracy of the identification results, multi-dimensional cross-validation is performed based on the comparison between the current error and the dynamic threshold, combined with vehicle cloud data, to further improve the accuracy of fault identification.

[0057] Specifically, runtime sequence data can be uploaded to the cloud and compared with unified vehicle historical data and data of the same vehicle model group stored in the cloud. When comparing the fault characterization error of the current time sequence window with the currently effective dynamic threshold, it is necessary to count the number of time sequence windows that consecutively exceed the threshold. If the fault characterization error of three consecutive time sequence windows is greater than the dynamic threshold, a local anomaly warning is triggered. At the same time, through the vehicle networking module, the runtime sequence data corresponding to the three consecutive anomaly windows, the fault characterization error of each window, the local anomaly warning trigger status, and the current vehicle operating condition information are encrypted and uploaded to the car manufacturer's cloud big data platform.

[0058] After receiving the data uploaded by the vehicle, the cloud platform compares the received runtime sequence data with the distribution characteristics of the vehicle's historical normal operating data under the same conditions stored in the cloud. This confirms whether the data deviates from the vehicle's historical normal operating boundary. Specifically, the data is cross-compared with the distribution of normal operating data under the same conditions of a large group of vehicles of the same model and the labeled electric drive system fault feature library to eliminate interference caused by operating condition fluctuations and normal sensor drift. At the same time, it can call the high-precision electric drive system fault identification model pre-trained in the cloud to perform deep fault feature extraction and identification on the runtime sequence data, and finally output a clear binary fault judgment result, i.e., fault exists / fault does not exist.

[0059] When an anomaly warning is triggered and the fault assessment result indicates the presence of a fault, it means that the local identification result is normal, and the electric drive system can be determined to be in a faulty state. Upon confirming the electric drive system is in a faulty state, fault tags can be added to the current runtime sequence data. This includes simultaneously storing the runtime sequence data corresponding to this fault, the fault characterization error, and the cloud-based fault tag to both the vehicle's local and cloud-based fault sample libraries. These samples will serve as labeled samples for subsequent reconstruction function model iteration optimization and fault feature library updates. In addition to triggering anomaly warnings, the warning level can be upgraded based on the existing local warnings. Fault alerts can be issued to the driver through multiple channels, including the instrument panel, central control screen, and voice broadcast. Simultaneously, based on the fault type feedback from the cloud, corresponding safety control actions can be executed, such as limiting the electric drive output power for minor faults and controlling the vehicle to enter limp mode for serious faults. Further anomaly warning prompts can be output, and SMS reminders can be sent.

[0060] For further details, please refer to... Figure 4 After collecting the labeled data, an additional classifier connected to the encoder processes it. Through supervised fine-tuning, further fault classification is achieved. For example, in the supervised fine-tuning stage, a strategy is selected based on the scale of the labeled data: when there are very few labels, the encoder is frozen and only the classification head is trained to avoid destroying unsupervised features. When there are relatively sufficient labels, the top layer of the encoder is unfrozen for end-to-end fine-tuning, using cross-entropy loss as the optimization objective, iteratively updating the model parameters until the classification accuracy on the validation set reaches the target or the loss converges. The finally trained model can perform fault classification.

[0061] This embodiment uses cloud-based cross-validation to filter out false alarms caused by fluctuations in vehicle operating conditions and sensor drift, while also supplementing early faults that were missed by the vehicle, thereby further improving the accuracy of fault identification.

[0062] Based on the third embodiment described above, in the fourth embodiment of this application, the dynamic threshold is updated when the dynamic threshold update conditions are met. These dynamic threshold update conditions include triggering an anomaly warning and determining that no fault exists, or not triggering an anomaly warning and determining that a fault exists.

[0063] Specifically, when the fault assessment result in the cloud indicates that there is no fault, but an anomaly warning is triggered locally, it means that the original dynamic threshold may no longer be able to accurately capture the potential fault characteristics of the current electric drive system. In this case, the dynamic threshold needs to be updated. Similarly, if the fault assessment result indicates that there is a fault, but no anomaly warning is triggered locally, it means that there are cases of missed detection locally. In this case, the dynamic threshold needs to be lowered to improve the recognition accuracy. The dynamic threshold can be updated by updating the confidence quantile, or by updating the current dynamic threshold based on the difference between the two results.

[0064] As an optional implementation, the dynamic threshold can be updated by updating the quantiles. When updating the dynamic threshold, a target quantile can be calculated based on the fault judgment result and the triggering state of the abnormal warning. This target quantile is then used as a candidate threshold and input into a preset smoothing model to obtain the target threshold, which is the updated dynamic threshold. The calculation process of the preset smoothing model is as described in the second embodiment above. It can be understood that when calculating the dynamic threshold under normal conditions, a baseline dynamic threshold is first calculated, and then a weighted moving average smoothing process is applied between the baseline dynamic threshold and the dynamic threshold at the previous moment. Similarly, when updating the dynamic threshold, the target quantile is used as the new baseline dynamic threshold to calculate the new dynamic threshold.

[0065] The formula for calculating the target quantile is as follows: .

[0066] in, For the target quantile, This is the previous quantile (usually the preset quantile). The learning rate is used to calculate accuracy. accuracy .

[0067] Wherein, TP represents true positives, which are the number of samples whose actual results are abnormal and are also judged as abnormal locally; TN represents true negatives, which are the number of samples whose actual results are abnormal and are judged as normal locally; FP represents false positives, which are the number of samples whose actual results are normal and are judged as abnormal locally; and FN represents false negatives, which are the number of samples whose actual results are abnormal and are judged as normal locally.

[0068] Therefore, when calculating the target quantile, it is necessary to determine the decision statistics based on the triggering state of the anomaly warning and the fault judgment result. These decision statistics include true positives, true negatives, false positives, and false negatives. Next, the current accuracy of the reconstruction function model is determined using the decision statistics. Then, the product of the preset learning rate and the accuracy is calculated, and the sum of this product and the current quantile is used as the target quantile.

[0069] As an alternative implementation, when updating the dynamic threshold, if the fault judgment result output by the cloud indicates that there is no fault, and an abnormal warning is triggered locally on the vehicle, the fault characterization errors of the three time-series windows corresponding to this false alarm can be marked as normal valid errors and added to the stacked circular buffer corresponding to the local preset sliding window. The buffer automatically removes the three historical error nodes with the earliest timestamps, completing the update of the fault characterization error sequence within the sliding window. Subsequently, based on the updated fault characterization error sequence, the baseline dynamic threshold corresponding to the preset confidence quantile is recalculated, and then the new baseline dynamic threshold and the currently effective dynamic threshold are subjected to a weighted moving average smoothing process to complete the update of the dynamic threshold and correct the threshold boundary that caused the false alarm.

[0070] Furthermore, if the fault assessment result output by the cloud indicates the presence of a fault, and no abnormal warning is triggered locally, the fault characterization error of the runtime sequence data corresponding to this missed report is added to the fault characterization error sequence of the local sliding window. The baseline dynamic threshold is then recalculated, tightening the threshold boundary. Next, based on the confidence level of the fault assessment by the cloud, the locally preset confidence quantile p-value is lowered, for example, from the initial 99.7% to 99.5%, reducing the proportion of normal data coverage, improving the sensitivity to early fault identification, and avoiding similar missed detections in the future.

[0071] For example, when a vehicle is traveling at a constant speed on an urban road, if the fault characterization errors calculated for three consecutive time windows are 0.10, 0.11, and 0.12, respectively, all exceeding the current effective dynamic threshold of 0.09, a local Level 1 anomaly warning is triggered. Simultaneously, the runtime sequence data, fault characterization errors, and warning status for these three windows are uploaded to the cloud platform. Upon receiving the data, the cloud platform cross-validates the data using historical data from the vehicle and data from other vehicles of the same model, and uses a large cloud model for identification. It confirms that the data is a fluctuation caused by normal sensor drift under constant speed conditions and outputs a judgment that no fault exists. Upon receiving the result, the vehicle determines it to be a local false alarm, immediately cancels the local warning, marks the fault characterization errors of the three windows as normal valid errors, and adds them to a stacked circular buffer of length 1000. The buffer automatically removes the three oldest historical error nodes, completing the error sequence update. Based on the updated error sequence, the 99.7th percentile is recalculated to obtain a new baseline dynamic threshold of 0.095. After weighted smoothing, the current effective dynamic threshold is updated to 0.0915.

[0072] Furthermore, to aid in understanding the implementation process of this embodiment, please refer to... Figure 5 , Figure 5The control process combining local and cloud-based methods is illustrated. Specifically, a large amount of unlabeled normal data from the electric drive system is used as input. A pre-trained, unsupervised reconstruction function model is input, and through the model's reconstruction inference, the corresponding reconstruction error is output. Then, the error sequence is updated; that is, the reconstruction error output by the model is updated to a fixed-length error sequence {en}. This error sequence is the effective error queue corresponding to the sliding window, only including verified normal operating condition reconstruction errors, providing a statistical data source for subsequent dynamic threshold calculation. Simultaneously, it continuously receives normal error data from subsequent stages of the process for iterative updates. Finally, the dynamic threshold T is calculated using a sliding window. t =f(p), which means that the dynamic threshold T at the current time is calculated by combining the preset confidence quantile p. t The real-time reconstruction error is then compared with the calculated dynamic threshold. If the result indicates an anomaly warning has been triggered, the subsequent vehicle-to-cloud cross-validation process is initiated. If the result indicates no anomaly, the current data is classified as normal data, and the process proceeds to another branch of the vehicle-to-cloud cross-validation process.

[0073] For branches triggering abnormal warnings, during the vehicle-cloud verification process, the cloud platform uses historical data of the vehicle, group data of the same model, and a high-precision fault identification model to determine the fault outcome of the current data within a 10-minute verification window. If the determination result is positive, meaning the cloud and local determinations are consistent, confirming a real fault in the current data, a labeling action is performed, recording the corresponding operational data, reconstruction error, and fault status as labeled fault samples for subsequent model iteration and optimization. If the determination result is negative, meaning the cloud confirms no fault, the local warning is determined to be a false alarm, and the confidence quantile p is adjusted to correct the threshold boundary and avoid similar false alarms in the future.

[0074] In the other branch, if the judgment result is yes, it indicates that the local judgment is a missed report. The value of the confidence quantile p is adjusted to tighten the threshold boundary, improve the fault identification sensitivity, and avoid similar missed reports in the future. If the judgment result is no, it confirms that the current data is normal data. The process loop returns to the error sequence update stage, adds the reconstruction error corresponding to the normal data to the error sequence, completes the queue update, and provides the latest valid data source for the next round of dynamic threshold calculation.

[0075] This embodiment reverse-corrects the local sliding window error sequence, dynamic threshold, and confidence quantile parameters in response to false positives and false negatives confirmed by the cloud. At the same time, it uses misjudged cases as samples for incremental optimization of the model reconstruction, thereby preventing automatic correction after false positives and false negatives and improving the subsequent recognition accuracy.

[0076] Based on any of the above embodiments, in Embodiment 5 of this application, please refer to... Figure 6In the unsupervised training process of the reconstruction function model, historical runtime sequence data corresponding to the electric drive system, i.e., multi-sensor time-variable sequences, can be obtained, including historical runtime sequence data of the current vehicle and historical runtime sequence data of signals from the same vehicle. This historical runtime sequence data is considered normal data. Then, by padding and masking, missing sequences in the historical runtime sequence data are filled based on the mean, standard deviation, and median of each sensor parameter, resulting in a padded sequence. Subsequently, the padded sequence is compressed into a context vector using the Transformer encoder of the preset model, and then decoded into a training structure sequence, i.e., the reconstruction sequence, using the Transformer decoder. Finally, the reconstruction loss between the training structure sequence and the current input sequence is determined. When the reconstruction loss is less than the preset loss, the reconstruction function model is obtained.

[0077] Specifically, during the training process, the same model of electric drive system as the target vehicle is first collected, including historical running sequence data under fault-free conditions throughout the entire life cycle collected in whole vehicle bench tests and real vehicle road tests. This data covers all operating conditions such as idling, constant speed, acceleration, climbing, and regenerative braking, as well as normal operating scenarios across the entire ambient temperature range and speed range. Subsequently, the collected raw historical data is initially screened and cleaned to remove invalid samples with sensor hardware failures, interrupted acquisition links, or severely distorted data. Only the normal operating data under fault-free conditions is retained to form the basic dataset for model training. Next, the historical runtime sequence data within the training set is categorized by sensor channel and operating condition. For each sensor channel's historical normal data under the corresponding operating condition, the statistical characteristics of its mean μ, standard deviation σ, median M, and quartile intervals are calculated to establish a personalized statistical feature library for the sensor-operating condition dimension. Simultaneously, all variable-length historical runtime sequence data within the training set are uniformly padded to sampling points of a preset longest sequence length. For missing sequence positions in each sequence, the corresponding operating condition is first identified, and the mean μ and standard deviation σ of the corresponding sensor channel and operating condition are retrieved from the personalized statistical feature library. The padding value is calculated according to the rule: padding value = historical mean of the same operating condition for that sensor + random noise following a normal distribution, ensuring that the data distribution in the padding area is completely consistent with the actual data distribution under the same operating condition for that sensor. Finally, a binary mask matrix that perfectly matches the dimension and length of the padding sequence is generated synchronously, where the actual collected data positions are marked as 1, and the padded data positions are marked as 0. This is used to mask the loss calculation of the padding positions during subsequent training, reducing the interference of the padding data on the model's feature learning.

[0078] Next, the generated padding sequences and corresponding mask matrices are input batch-wise into a baseline model of a pre-defined Transformer encoder-decoder architecture. The encoder first performs linear embedding and learnable position encoding on the input padding sequences, preserving the temporal order information of the temporal data. Then, through a multi-head self-attention layer and a feedforward neural network, combined with a mask matrix, attention calculations at the padding positions are masked, allowing the model to focus solely on feature learning of the real data. Finally, the input padding sequences are compressed into highly abstract context vectors containing the temporal correlations and distribution features of normal data. Subsequently, the context vectors output by the encoder are input into the decoder module of the model. The decoder uses a masked multi-head self-attention layer, an encoder-decoder cross-attention layer, and a feedforward neural network to decode and reconstruct the context vector point by point. Finally, it outputs a reconstructed structure sequence that is completely identical to the input padding sequence in terms of dimension, temporal length, and number of sampling points, ensuring point-to-point comparability between the reconstructed sequence and the input sequence.

[0079] During model convergence, based on the generated mask matrix, only the real data locations marked as 1 are extracted. The mean squared error (MSE) between the reconstructed structure sequence and the input padding sequence at the real data locations is calculated and used as the reconstruction loss for this iteration. Using the AdamW optimizer, the weight parameters of the baseline model are iteratively optimized through backpropagation based on the calculated reconstruction loss. After each batch of training is completed, the validation reconstruction loss is simultaneously calculated on the validation set to monitor the model's generalization ability.

[0080] Secondly, Figure 7 A schematic diagram of an embodiment of the vehicle fault identification device of the present invention is shown. Figure 7 As shown, the vehicle fault identification device 300 includes: a reconstruction module 310, a detection module 320, and an execution module 330.

[0081] The reconstruction module 310 is used to input the runtime timing data of the electric drive system into a pre-trained reconstruction function model to obtain reconstructed timing data. The reconstruction function model is obtained by unsupervised pre-training based on normal data. Detection module 320 is used to determine the fault characterization error between the runtime timing data and the reconstructed timing data; The execution module 330 is used to trigger an abnormal warning if the fault characterization error is greater than a dynamic threshold.

[0082] In one optional embodiment, the vehicle fault identification device 300 further includes a threshold update module 340, which is used to acquire a fault characterization error sequence under normal operating conditions of the electric drive system based on a preset sliding window; sort the fault characterization error sequence according to a preset confidence quantile, and determine a baseline dynamic threshold based on the sorting result; and smooth the baseline dynamic threshold according to the dynamic threshold corresponding to the previous moment to obtain the dynamic threshold.

[0083] In one optional embodiment, the vehicle fault identification device 300 further includes a judgment module 350, which is used to determine that the electric drive system is in a fault state if the fault judgment result indicates that a fault exists and the abnormal warning is triggered; the threshold update module 340 is further used to update the dynamic threshold if the dynamic threshold update conditions are met; wherein, the dynamic threshold update conditions include: triggering the abnormal warning and the fault judgment result indicating that a fault does not exist; or not triggering the abnormal warning and the fault judgment result indicating that a fault exists.

[0084] In an optional manner, the threshold update module 340 is further configured to calculate a target quantile based on the fault judgment result and the triggering state of the abnormal warning; and input the target quantile as a candidate threshold into a preset smoothing model to obtain a target dynamic threshold, wherein the target dynamic threshold is the updated dynamic threshold.

[0085] In an optional manner, the threshold update module 340 is further configured to determine a decision statistic based on the triggering state of the abnormal warning and the fault judgment result, wherein the decision statistic includes true positives, true negatives, false positives, and false negatives; The accuracy of the reconstructed function model is determined based on the decision statistics. Calculate the product of the preset learning rate and the accuracy, and determine the sum of the product and the current quantile as the target quantile.

[0086] In an optional embodiment, the vehicle fault identification device 300 further includes a reconstruction function module 360, which is used to acquire historical runtime sequence data corresponding to the electric drive system, wherein the historical runtime sequence data is normal data; fill in the missing sequences in the historical runtime sequence data based on the mean, standard deviation, and median of each sensor parameter in the historical runtime sequence data to obtain a filled sequence; compress the filled sequence into a context vector according to the encoder of the preset model, and decode the context vector into a structure sequence corresponding to the subset based on the decoder; determine the reconstruction loss between the structure sequence and the current input sequence, and obtain the reconstruction function model when the reconstruction loss is less than the preset loss.

[0087] In an optional manner, the detection module 320 is further configured to determine the basic reconstruction error, local sensitive error, and channel weighted loss between the runtime timing data and the reconstructed timing data; and to obtain the fault characterization error by weighting and fusing the basic reconstruction error, the local sensitive error, and the channel weighted loss based on a preset loss weight.

[0088] In an alternative embodiment, the reconstruction module 310 is further configured to reconstruct the context vector into reconstructed temporal data with the same length and dimension as the input sequence based on the decoding layer of the reconstruction function model.

[0089] Figure 8 A structural schematic diagram of an embodiment of the vehicle of the present invention is provided. The specific embodiments of the present invention do not limit the specific implementation of the vehicle.

[0090] like Figure 8 As shown, the vehicle may include: a processor 402, a communication interface 404, a memory 406, and a communication bus 408.

[0091] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408. Communication interface 404 is used to communicate with other network elements such as clients or other servers. The processor 402 executes program 410, specifically performing the relevant steps described in the embodiment of the vehicle fault identification method.

[0092] Specifically, program 410 may include program code, which includes computer-executable instructions.

[0093] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The vehicle may include one or more processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0094] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0095] Thirdly, embodiments of the present invention provide a computer-readable storage medium storing at least one executable instruction that, when executed on a vehicle / vehicle control device, causes the vehicle / vehicle fault identification device to perform the vehicle fault identification method in any of the above method embodiments.

[0096] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments of this invention are not directed to any particular programming language.

[0097] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. Similarly, for the sake of brevity and to aid in understanding one or more aspects of the invention, in the description of exemplary embodiments of the invention above, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.

[0098] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

[0099] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.

Claims

1. A vehicle fault identification method, characterized in that, include: The runtime timing data of the electric drive system is input into a pre-trained reconstruction function model to obtain reconstructed timing data, wherein the reconstruction function model is obtained by unsupervised pre-training based on normal data; Determine the basic reconstruction error, local sensitive error, and channel weighted loss between the runtime timing data and the reconstructed timing data; Based on the preset loss weight, the basic reconstruction error, the local sensitive error, and the channel weighted loss are weighted and fused to obtain the fault characterization error; Based on a preset sliding window, obtain the fault characterization error sequence of the electric drive system under normal operating conditions; The fault characterization error sequence is sorted according to a preset confidence quantile, and a benchmark dynamic threshold is determined based on the sorting result. The benchmark dynamic threshold is smoothed based on the threshold corresponding to the previous moment to obtain the dynamic threshold. If the fault characterization error is greater than the dynamic threshold, an anomaly warning is triggered.

2. The vehicle fault identification method as described in claim 1, characterized in that, The process of obtaining the fault characterization error by weighting and fusing the basic reconstruction error, the local sensitivity error, and the channel weighted loss based on preset loss weights includes: Obtain the fault judgment result determined by the cloud based on the runtime sequence data; If the fault determination result indicates that a fault exists and the abnormal warning is triggered, the electric drive system is determined to be in a fault state. If the dynamic threshold update conditions are met, the dynamic threshold is updated; wherein, the dynamic threshold update conditions include: The abnormal warning is triggered, and the fault determination result is that there is no fault; or The aforementioned abnormal warning was not triggered, and the fault determination result indicates that a fault exists.

3. The vehicle fault identification method as described in claim 2, characterized in that, Updating the dynamic threshold includes: Based on the fault judgment result and the triggering status of the abnormal warning, the target quantile is calculated; The target quantile is used as a candidate threshold and input into a preset smoothing model to obtain the target dynamic threshold, which is the updated dynamic threshold.

4. The vehicle fault identification method as described in claim 3, characterized in that, The calculation of the target quantile based on the fault judgment result and the triggering status of the anomaly warning includes: Based on the triggering state of the abnormal warning and the fault judgment result, the judgment statistics items are determined, including true positives, true negatives, false positives, and false negatives. The accuracy of the reconstructed function model is determined based on the decision statistics. Calculate the product of the preset learning rate and the accuracy, and determine the sum of the product and the current quantile as the target quantile.

5. The vehicle fault identification method as described in claim 1, characterized in that, Before inputting the runtime timing data of the electric drive system into the pre-trained reconstruction function model to obtain the reconstructed timing data, the following steps are included: Acquire historical runtime sequence data corresponding to the electric drive system, wherein the historical runtime sequence data is normal data; Based on the mean, standard deviation, and median of each sensor parameter in the historical runtime sequence data, the missing sequences in the historical runtime sequence data are filled to obtain the filled sequences; The encoder of the preset model compresses the padding sequence into a context vector, and the decoder decodes the context vector into a training structure sequence. Determine the reconstruction loss between the training structure sequence and the current input sequence, and obtain the reconstruction function model when the reconstruction loss is less than a preset loss.

6. The vehicle fault identification method as described in claim 1, characterized in that, The step of inputting the runtime timing data of the electric drive system into a pre-trained reconstruction function model to obtain reconstructed timing data includes: The preset sequence length of the runtime sequence data is determined, the missing sequence of the runtime sequence data is filled based on the pre-stored sensor statistical features, and the real data and the filled data are identified based on the mask matrix; The encoding layer based on the reconstruction function model compresses the padded input sequence into a context vector; The decoding layer based on the reconstruction function model reconstructs the context vector into reconstructed temporal data with the same length and dimension as the input sequence.

7. A vehicle fault identification device, characterized in that, The device includes: The reconfiguration module is used to input the runtime timing data of the electric drive system into a pre-trained reconfiguration function model to obtain reconfigured timing data, wherein the reconfiguration function model is obtained by unsupervised pre-training based on normal data; The detection module is used to determine the basic reconstruction error, local sensitive error, and channel weighted loss between the runtime timing data and the reconstructed timing data; and to obtain the fault characterization error by weighting and fusing the basic reconstruction error, the local sensitive error, and the channel weighted loss based on a preset loss weight. The threshold update module is used to obtain the fault characterization error sequence of the electric drive system under normal operating conditions based on a preset sliding window; sort the fault characterization error sequence according to a preset confidence quantile, and determine a benchmark dynamic threshold based on the sorting result; and smooth the benchmark dynamic threshold according to the threshold corresponding to the previous moment to obtain a dynamic threshold. The execution module is used to trigger an anomaly warning if the fault characterization error is greater than a dynamic threshold.

8. A vehicle, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation of the vehicle fault identification method as described in any one of claims 1-6.

Images