A physical embedding-based power battery abnormality diagnosis method

Abstract

The application provides a kind of physical embedding-based power battery anomaly diagnosis method, the method first obtains original time series data in charging process and extracts mileage difference value;Further, the physical machine understanding is coupled as external excitation and battery internal response;Then, an anomaly diagnosis model composed of an encoder, a decoder and a regression head is constructed, the regression loss of mileage difference prediction value and true value is used to constrain the implicit variable distribution, so as to make the aging characteristics and abnormal characteristics decoupled in the hidden space, and the training process only needs normal samples;In inference, the reconstruction error of each charging segment is extracted, the recent abnormal signal is amplified after time sequence calibration, and then the error mean value of the first h is taken as the comprehensive health risk index, and the abnormal state is determined by comparing with the diagnosis threshold. The application solves the technical problem of inaccurate real vehicle battery anomaly diagnosis caused by missing single data, weakly correlated labels and aging drift.

Description

Technical Field

[0001] This invention relates to the technical field of power battery health management, and in particular to a method for diagnosing power battery anomalies based on physical embedding. Background Technology

[0002] The diagnosis of power batteries relies heavily on high-quality operational records. In real-world vehicle operation scenarios, due to limitations in data transmission bandwidth and the anonymization processing of onboard terminals, it is often difficult to obtain operational data containing the complete voltage of individual battery cells. Typically, condition assessment can only be achieved by relying on large-scale real-vehicle charging characteristics collected at charging piles.

[0003] However, existing power battery anomaly diagnosis technologies have the following significant drawbacks in real-vehicle applications: (1) Lack of single-cell observation data leads to the failure of traditional methods. Traditional power battery fault detection methods mainly capture anomalies by monitoring the voltage of individual cells inside the battery pack. However, in real-vehicle charging scenarios, the lack of single-cell level observations causes traditional methods based on single-cell voltage differences to completely fail. (2) Fault labels exhibit severe weak correlation. In real-vehicle datasets, it is extremely difficult to obtain fault labels accurate to the charging process. Existing labeling mechanisms typically mechanically extend vehicle-level anomaly labels (such as fire events) to all historical charging segments of the faulty vehicle. This labeling mechanism makes it easy for the model to generate a large number of misjudgments during training, because the early charging segments of the faulty vehicle may exhibit completely normal characteristics. (3) The interference of natural aging drift on anomaly diagnosis. As the battery capacity naturally decays, its normal charging voltage and current response curves will themselves undergo physical drift. If the anomaly diagnosis mechanism cannot identify and remove this baseline drift caused by normal aging evolution, it is very easy to misjudge a battery with severely decayed capacity but still in a healthy state as an anomaly.

[0004] Therefore, how to effectively decouple battery aging status and accurately identify potential anomalies under the objective physical constraints of missing individual data and weak label correlation is a key technical challenge that urgently needs to be solved in the field of power battery anomaly diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a power battery anomaly diagnosis method based on physical embedding, which solves at least one of the technical problems in existing power battery anomaly diagnosis schemes, namely, the failure of traditional monitoring methods due to the lack of individual cell voltage data, the difficulty in achieving accurate labeling due to the weak correlation between anomaly tags and charging segments, and the interference of response curve baseline drift caused by battery natural aging on diagnostic accuracy.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a method for diagnosing power battery anomalies based on physical embedding, comprising the following steps: S1. Obtain the original time series data of the power battery during the charging process, preprocess the original time series data, and extract the mileage difference value. S2. The preprocessed original time series data is split into input excitation quantity and battery response quantity according to the physical mechanism; S3. Construct a physical embedding anomaly diagnostic model, which includes an encoder, a decoder, and a regression head; iteratively train the model and optimize the network weights: input the input stimulus and the battery response into the encoder, and output the mean and variance of the posterior distribution of the latent variables; sample the latent variables from the posterior distribution; input the latent variables and the input stimulus into the decoder simultaneously, and output the reconstructed battery response; input the latent variables into the regression head, and output the mileage difference prediction value; use the regression loss between the mileage difference prediction value and the actual mileage difference value to constrain the distribution of the latent variables, thereby decoupling the battery's natural aging characteristics from its abnormal characteristics in the latent space; S4. Input the charging segment to be diagnosed of the target evaluation vehicle into the trained physical embedding anomaly diagnosis model, calculate the reconstruction error between the battery response and the reconstructed battery response; collect the reconstruction errors corresponding to multiple historical charging segments of the target evaluation vehicle to form an error sequence; perform time-series calibration on the error sequence to obtain the calibrated reconstruction error sequence. S5. Sort the calibrated reconstructed error sequence in descending order of numerical value, and extract the top-ranked sequences. h% The values ​​are calculated and averaged to obtain the comprehensive health risk index of the vehicle; the comprehensive health risk index is compared with a preset diagnostic threshold, and if the comprehensive health risk index is lowered... HR If the diagnostic threshold is exceeded, the power battery is determined to be in an abnormal state.

[0007] Furthermore, in step S1, the original time-series data includes average voltage, charging current, maximum cell voltage, minimum cell voltage, maximum probe temperature, minimum probe temperature, state of charge, and corresponding timestamps; the formula for calculating the mileage difference value is: ; in, , Indicates the first The current cumulative mileage and its difference value at the time of the next charging event.

[0008] Furthermore, in step S2, the input excitation quantities include a charging current sequence and a state of charge sequence; the battery response quantities include an average voltage sequence, a maximum cell voltage sequence, a minimum cell voltage sequence, and a maximum probe temperature sequence.

[0009] Furthermore, in step S3, the model is iteratively trained, specifically including: iteratively training the physical embedding anomaly diagnosis model using only sample data from normal vehicles, without relying on the fault labels of charging segments.

[0010] Furthermore, in step S3, optimizing the network weights specifically includes: optimizing the network weights by minimizing the joint loss function composed of reconstruction loss, divergence constraint loss, and regression loss; the formula for calculating the joint loss function is:

[0011] in, This represents the reconstruction loss. This represents the divergence constraint loss. This represents the regression loss; To balance the hyperparameters of the two task weights, and ; is the regularization coefficient.

[0012] Furthermore, in step S3, both the encoder and the decoder adopt a long short-term memory network structure, and the regression head adopts a multilayer perceptron network.

[0013] Furthermore, in step S3, when sampling the latent variable from the posterior distribution, an independent noise vector sampled from the standard normal distribution is introduced, and continuous latent variables are obtained through reparameterization.

[0014] Furthermore, in step S4, the error sequence is time-series calibrated, specifically by introducing an exponentially decaying time-weighted function to perform time-series calibration on the error sequence, thereby obtaining a calibrated reconstructed error sequence.

[0015] Furthermore, the step of introducing an exponentially decaying time-weighted function to perform time-series calibration on the error sequence to obtain a calibrated reconstructed error sequence specifically includes: introducing a time sensitivity coefficient. The error sequence is then subjected to time-series calibration to obtain the calibrated reconstructed error sequence. :

[0016] In the formula, For the first The reconstruction error after timing calibration of each charging segment is calculated using the following formula:

[0017] In the formula, The time sensitivity coefficient, For fragment index, This represents the total number of segments.

[0018] Furthermore, in step S5, the aggregation parameters The determination process includes: On the validation set, Traverse the range of 5% to 100% with a fixed step size; For each h% Calculate the comprehensive health risk index corresponding to all validation samples; Based on the true anomaly labels of the validation samples, calculate each The corresponding AUROC value; The optimal aggregation ratio is selected as the ratio that maximizes the AUROC value. .

[0019] Compared with the prior art, the present invention has at least the following beneficial effects: First, to address the problem that traditional methods fail due to missing individual cell voltage data in real-world vehicle scenarios, this invention decomposes the preprocessed original time series data into input excitation and battery response quantities according to physical mechanisms. The battery response quantity is reconstructed using an encoder, and the reconstruction error is used as the basis for anomaly detection. This allows for effective anomaly detection without relying on the observed values ​​of individual cell voltages, overcoming the defect that traditional methods completely fail under conditions of missing individual cell data.

[0020] Second, to address the problem of weak correlation between abnormal labels and charging segments, which makes model training difficult, this invention uses sample data from normal vehicles to iteratively train a physically embedded anomaly diagnosis model. This eliminates the need for fault labels that are precise enough to indicate the charging process. At the same time, by introducing a mileage differential regression task as a weak supervision signal to constrain the latent space distribution, the model can still effectively learn normal charging patterns even under conditions of weak label correlation. Furthermore, in the inference stage, accurate vehicle-level anomaly determination is achieved through temporal calibration and high-risk feature aggregation, significantly reducing the dependence on precise segment-level labels and enabling high-accuracy diagnosis even in the absence of segment-level annotations.

[0021] Third, to address the issue of baseline drift in response curves caused by natural battery aging interfering with diagnostic accuracy, this invention utilizes mileage difference as a physical prior label as a weak supervision signal during the training phase. By guiding the model through a regression task, it decouples the natural aging characteristics of the battery from potential abnormal characteristics in the latent space, ensuring that the reconstruction error primarily reflects abnormal deviations rather than aging drift. Simultaneously, during the inference phase, an exponential decay time-weighted function is introduced to perform temporal calibration on the reconstruction error sequence, amplifying recent abnormal signals and further suppressing the interference of aging baseline drift on the judgment results, thereby improving the accuracy and reliability of the diagnosis. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the physical embedding-based power battery anomaly diagnosis method provided by the present invention. Figure 2 This is a schematic diagram of the decoupling architecture and training process of the physical embedding anomaly diagnosis model provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] I. Specific Implementation The dataset used in this embodiment is the EVBattery real-vehicle charging dataset, which is collected from electric vehicle charging stations and aggregates charging records from 464 real electric vehicles produced by three different car manufacturers, containing a total of 1,278,277 charging segments. The raw data uses a sliding window to divide the continuous charging data into fixed-length charging segments, with the time series length of each charging segment set to 128 and the sampling frequency set to 0.1Hz.

[0026] The complete execution process of the power battery anomaly diagnosis method based on physical embedding in this embodiment is as follows: 1. Regarding step S1 S1. Acquisition and Preprocessing of Multidimensional Charging Segment Data: First, the original time-series data of the charging segment is acquired, covering eight physical dimensions. A systematic preprocessing procedure is then performed on the acquired original time-series data, specifically including two sub-steps: S1.1, Abnormal Data Cleaning and Sequence Alignment: Based on the actual electrochemical operating characteristics of the power battery, upper and lower limits of safe voltage and allowable extreme temperatures are set as perturbation range thresholds. Outlier charging segments exceeding these perturbation range thresholds are directly removed. For charging segments whose time series length does not conform to the standard length of 128, truncation or forward padding strategies are used for unified alignment to meet the requirements of the network's fixed input dimension. S1.2 Mileage Difference Processing: The current cumulative mileage in the original static metadata is an absolute value that monotonically increases over time, including the absolute baseline deviation introduced when the vehicle was first put into use. To eliminate this baseline deviation and accurately reflect the current state of the battery, this embodiment converts the absolute mileage into a mileage difference value between two adjacent valid charging events. The calculation formula is as follows:

[0027] In the formula, Indicates the first Current cumulative mileage at the time of the second charging event; The cumulative driving mileage since the last effective charge; For the calculated first The range difference value corresponding to the second charging event. This will serve as a physical prior label for characterizing battery capacity degradation in subsequent steps.

[0028] Step S1, through mileage differential processing, solves the problem of initial usage benchmark deviation in the original mileage data and accurately reflects the battery load status between two charging events.

[0029] 2. Regarding step S2 S2. Decoupling of physical features between input excitation and battery response: To guide the model in learning the causal dynamic mapping of the battery system's internal state to external excitation, this embodiment abandons the traditional approach of jointly reconstructing multidimensional time series of all dimensions. Instead, it decouples and partitions the feature inputs based on the physical mechanism of the battery system. S2.1. Divide the charging current sequence and the state-of-charge sequence into input excitation quantities. This refers to the external excitation applied to the battery system by the charging pile. S2.2, The average voltage sequence, maximum cell voltage sequence, minimum cell voltage sequence, and maximum probe temperature sequence are divided into battery response quantities. This refers to the physical response of the battery's internal state to the external stimulus.

[0030] It is important to note that the relative timestamps and minimum probe temperatures in the original time-series data are not used in the subsequent construction of the physical embedding anomaly diagnostic model. This is because the time-series model itself has the ability to handle temporal relationships, and adding relative timestamps would introduce feature redundancy and have a negative impact.

[0031] Step S2 addresses the causal physical mechanism of the power battery system by using an input-output decoupling design to guide the model to learn the internal response law driven by external stimuli, thereby reducing feature redundancy and variable coupling.

[0032] 3. Regarding step S3 S3. Training of a decoupled variational autoencoder model based on weakly supervised constraints: such as... Figure 2 As shown, the physical embedding anomaly diagnostic model constructed in this embodiment has an internal topology consisting of an encoder, a decoder, and a lightweight regression head (MLP).

[0033] S3.1 Both the encoder and the decoder use a Long Short-Term Memory (LSTM) network structure to process time series data, with a hidden layer dimension of 64; the encoder's hidden space size is set to 32; and the regression head uses a Multilayer Perceptron (MLP) network topology. S3.2, the input excitation quantity and the battery response quantity The encoder is input to the common input, and it fits the posterior distribution of the latent variables and outputs their mean vector. Sum of variance vectors The formula is as follows:

[0034] in, Indicates encoder operations, and These represent the mean and variance, respectively. S3.3 To ensure the smooth execution of the backpropagation algorithm, a reparameterization technique is introduced, that is, an independent noise vector sampled from the standard normal distribution is introduced. Continuous latent space vectors are obtained through reparameterization. :

[0035] in, Represents hidden variables; This represents the element-wise multiplication operator.

[0036] S3.4, Input excitation amount of the external input With the sampled latent space vector These are used together as input to the decoder, which reconstructs the sequence of response observations, i.e., reconstructs the battery response. The formula is as follows:

[0037] in, This indicates decoder operations. This represents the reconstructed value of the battery response; S3.5, the latent space vector The regression head is input, and the predicted mileage difference value corresponding to the charging sequence is output. The formula is as follows:

[0038] in, MLP This indicates the regression head operation. This represents the predicted value of the mileage difference; S3.6. To enable the model to sensitively detect potential battery anomalies and isolate baseline drift caused by natural aging, this embodiment adopts a joint training strategy. First, a training set is established following a vehicle-level independent partitioning strategy. All normal vehicles are randomly divided into a normal training set, a normal validation set, and a normal test set in a 6:2:2 ratio; all abnormal vehicles are divided into an abnormal validation set and an abnormal test set in a 5:5 ratio.

[0039] S3.7 During the training phase, the model uses only the normal training set to learn the charging distribution manifold in a purely healthy state. Model parameter optimization is achieved by minimizing the joint loss function. Implement the joint loss function The mathematical formula for calculation is as follows:

[0040] In the formula, The reconstruction loss is the reconstructed battery response. Battery response to actual input The mean square error between them; For introduction KL The divergence regularization constraint term is used to constrain the distribution of the encoder output to approximate the standard normal distribution in order to prevent mode collapse in the latent space. The regression loss is the predicted mileage difference. Difference from actual mileage The residuals between; To balance the hyperparameters of the two task weights, in this embodiment, the regression task is defined as an auxiliary weakly supervised constraint. To ensure that the total loss is dominated by the unsupervised reconstruction loss, the hyperparameters... Set to 0.95; is the divergence constraint regularization coefficient.

[0041] S3.8 The model uses the Adam optimizer for backpropagation and embeds the cosine annealing algorithm. It uses the normal validation set to monitor the loss to trigger an early stopping mechanism. After 10 iterations of training, the model fully converges. At this point, all network weights of the physical embedding anomaly diagnosis model are fixed.

[0042] Step S3 employs a physical prior weak supervision mechanism, which introduces a mileage differential regression task to constrain the latent space distribution, forcing the model to decouple the battery aging mode in the absence of precise labels.

[0043] 4. Regarding step S4 S4. Segment-level reconstruction error extraction and temporal weighted calibration: After the network weights are fixed, the anomaly diagnosis enters the test feature evaluation stage. The historical charging segments of the target evaluation vehicle are sequentially input into the physical embedding anomaly diagnosis model to extract segment-level reconstruction errors and perform temporal calibration.

[0044] S4.1, the first Input a segment into the trained model to obtain the reconstructed response, and calculate its reconstruction error. :

[0045] in, This indicates the calculation of mean square error; Collect all of the target evaluation vehicles The reconstruction errors corresponding to the historical charging segments ordered chronologically constitute an error sequence. :

[0046] S4.2 Considering the significant irreversibility and time-cumulative effect of power battery fault evolution, the closer the reconstruction error is to the final time point, the higher its implied anomaly probability. Therefore, a time sensitivity coefficient is introduced, and an exponential decay time weighting function is used to weight the error sequence. Calibration is performed to obtain the calibrated reconstruction error sequence. :

[0047] In the formula, For the first The reconstruction error after timing calibration of each charging segment is calculated using the following formula:

[0048] In the formula, The time sensitivity coefficient is preferred in this embodiment. ; This represents the total number of segments; the reconstructed error after calibration will be used as a segment-level anomaly feature.

[0049] Step S4 introduces an exponential decay time-weighted mechanism, which solves the problem of unreliable diagnosis of a single segment and amplifies the irreversible abnormal characteristics of the power battery over time.

[0050] 5. Regarding step S5 S5. High-risk feature aggregation and vehicle-level abnormal state determination: In order to eliminate false alarms and missed alarms caused by occasional operating condition noise in a single charging segment, this embodiment uses the calibrated reconstructed error sequence. Sort by numerical value in descending order. Extract the top-ranked items. The timing calibration error is calculated and its mean is defined as the comprehensive health risk index of the target vehicle. The calculation formula is as follows:

[0051] In the formula, the feature aggregation parameter The optimal value is adaptively determined through a grid search on a hybrid validation set consisting of a normal validation set and an abnormal validation set. Let... Within the range of 5% to 100%, a fixed step size is used for traversal. Combined with real vehicle labels, the Area Under the Receiver Operating Characteristic (AUROC) for different parameters is calculated. Through a grid search on the validation set, the parameter that maximizes the AUROC value is selected as the optimal aggregation ratio. .

[0052] Finally, the calculated Compare with the high-risk diagnostic threshold, if If the threshold is exceeded, the diagnostic assessment determines that the power battery of the target vehicle is in an abnormal fault state.

[0053] Step S5 employs an adaptive aggregation parameter optimization strategy, determining the optimal aggregation ratio through grid search. It balances the ability to detect local anomalies with the ability to resist global noise.

[0054] II. Model Comparison To objectively evaluate the comprehensive diagnostic performance of the method of this invention, this embodiment introduces traditional standard autoencoders (AE), single-class support vector data description (SVDD), graph deviation networks (GDN), multivariate temporal feature anomaly diagnosis algorithm (MTAD-GAT), and dynamic autoencoder models (DyAD) as benchmark models for comparison. The specific parameter settings for each baseline model are shown in Table 1. Table 1. Model Parameter Settings

[0055] The final vehicle-level fault diagnosis performance (AUROC) of this embodiment under different manufacturer datasets (EVBattery B1, EVBattery B3) is shown in Table 2.

[0056] Table 2. Comparison Results of Baseline Models

[0057] As can be clearly seen from the experimental data in Table 2, the method proposed in this invention achieves optimal anomaly diagnosis performance under all test conditions. In particular, on the EVBattery B1 real-vehicle dataset, the AUROC index of this invention reaches 82.6%, significantly outperforming other baseline models.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing abnormalities in a power battery based on physical embedding, characterized in that, Includes the following steps: S1. Obtain the original time series data of the power battery during the charging process, preprocess the original time series data, and extract the mileage difference value. S2. The preprocessed original time series data is split into input excitation quantity and battery response quantity according to the physical mechanism; S3. Construct a physical embedding anomaly diagnostic model, which includes an encoder, a decoder, and a regression head; iteratively train the model and optimize the network weights: input the input stimulus and the battery response into the encoder, and output the posterior distribution mean and variance of the latent variables; Latent variables are sampled from the posterior distribution; the latent variables and the input stimulus are simultaneously input into the decoder to output the reconstructed battery response; the latent variables are input into the regression head to output the mileage difference prediction value. By using the regression loss between the predicted mileage difference and the actual mileage difference to constrain the distribution of latent variables, the natural aging characteristics and abnormal characteristics of the battery are decoupled in the latent space. S4. Input the charging segment to be diagnosed of the target vehicle into the trained physical embedding anomaly diagnosis model, and calculate the reconstruction error between the battery response and the reconstructed battery response. The reconstruction errors corresponding to multiple historical charging segments of the target evaluation vehicle are collected to form an error sequence; the error sequence is time-series calibrated to obtain a calibrated reconstruction error sequence. S5. Sort the calibrated reconstructed error sequence in descending order of numerical value, and extract the top-ranked sequences. h% The values ​​are calculated and the average value is used as the comprehensive health risk index of the vehicle. The comprehensive health risk index is compared with a preset diagnostic threshold. If the comprehensive health risk index... HR If the diagnostic threshold is exceeded, the power battery is determined to be in an abnormal state.

2. The power battery anomaly diagnosis method based on physical embedding according to claim 1, characterized in that, In step S1, the original time-series data includes average voltage, charging current, maximum cell voltage, minimum cell voltage, maximum probe temperature, minimum probe temperature, state of charge, and relative timestamp; the formula for calculating the mileage difference is: ； in, , Indicates the first The current cumulative mileage and its difference value at the time of the next charging event.

3. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S2, the input excitation quantities include a charging current sequence and a state of charge sequence; the battery response quantities include an average voltage sequence, a maximum cell voltage sequence, a minimum cell voltage sequence, and a maximum probe temperature sequence.

4. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S3, the model is iteratively trained, specifically including: iteratively training the physical embedding anomaly diagnosis model using only sample data from normal vehicles, without relying on the fault labels of charging segments.

5. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S3, optimizing the network weights specifically includes: optimizing the network weights by minimizing the joint loss function composed of reconstruction loss, divergence constraint loss, and regression loss; the formula for calculating the joint loss function is: ； in, This represents the reconstruction loss. This represents the divergence constraint loss. This represents the regression loss; To balance the hyperparameters of the two task weights, and ; is the regularization coefficient.

6. The power battery anomaly diagnosis method based on physical embedding according to claim 1, characterized in that, In step S3, both the encoder and the decoder adopt a long short-term memory network structure, and the regression head adopts a multilayer perceptron network.

7. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S3, when sampling the latent variable from the posterior distribution, an independent noise vector sampled from the standard normal distribution is introduced, and continuous latent variables are obtained through reparameterization.

8. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S4, the error sequence is time-series calibrated, specifically by introducing an exponentially decaying time-weighted function to perform time-series calibration on the error sequence, thereby obtaining a calibrated reconstructed error sequence.

9. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 8, characterized in that, The introduction of an exponentially decaying time-weighted function to perform time-series calibration on the error sequence, resulting in a calibrated reconstructed error sequence, specifically includes: introducing a time sensitivity coefficient. The error sequence is then subjected to time-series calibration to obtain the calibrated reconstructed error sequence. : ； In the formula, For the first The reconstruction error after timing calibration of each charging segment is calculated using the following formula: ； In the formula, The time sensitivity coefficient, For fragment index, This represents the total number of segments.

10. The method for abnormal diagnosis of power batteries based on physical embedding according to claim 1, characterized in that, In step S5, the aggregation parameters The determination process includes: On the validation set, Traverse the range of 5% to 100% with a fixed step size; For each h% Calculate the comprehensive health risk index corresponding to all validation samples; Based on the true anomaly labels of the validation samples, calculate each The corresponding AUROC value; The optimal aggregation ratio is selected as the ratio that maximizes the AUROC value. .

Images