A method, apparatus, equipment and medium for classifying battery faults

By integrating the features of multi-band electrochemical impedance spectroscopy and equivalent circuit models, and combining symmetric decision trees and SHAP values ​​for battery fault classification, the problems of untimely identification and insufficient classification ability in battery fault diagnosis are solved, achieving early, accurate, and interpretable fault diagnosis.

CN123066566APending Publication Date: 2026-11-10BEIJING HANGSHENG NEW ENERGY TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610982846.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-02
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing battery fault diagnosis technologies suffer from problems such as untimely identification of internal electrochemical anomalies, insufficient ability to classify multiple faults, inadequate utilization of EIS information, and lack of interpretability of diagnostic results, making it difficult to meet the needs for early, accurate, and reliable fault diagnosis.

Method used

By fusing the statistical features of multi-band electrochemical impedance spectroscopy (EIS) with the electrochemical mechanism feature vectors extracted from the equivalent circuit model (ECM), a heterogeneous feature set is constructed. An ensemble learning model based on symmetric decision trees is then used to evaluate the feature contribution using a weighted multi-classification loss function and SHAP values. The feature fusion weights are dynamically adjusted to output interpretable diagnostic criteria.

Benefits of technology

It improves the accuracy and sensitivity of fault classification, enhances the credibility and practicality of diagnostic results, and enables early identification and interpretable fault classification of high-risk faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123066566A_ABST
    Figure CN123066566A_ABST
Patent Text Reader

Abstract

This application discloses a battery fault classification method, apparatus, device, and medium, relating to the field of battery technology. The method includes: fusing electrochemical impedance spectroscopy data of the battery into a heterogeneous feature set containing multi-band statistical features and electrochemical mechanism parameters; constructing a fault category model using an ensemble learning model based on a symmetric decision tree, with a weighted loss function as the optimization objective; during training, iteratively evaluating and providing feedback on the feature contribution based on SHAP values, and dynamically adjusting the feature fusion weights to enable the model to adaptively focus; finally, outputting the fault category, risk level, and interpretable diagnostic criteria based on SHAP values. This invention achieves early, accurate, and interpretable fault classification, improving the sensitivity and diagnostic reliability of high-risk fault identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power battery technology for new energy vehicles, and in particular to a battery fault classification method, device, equipment, and medium. Background Technology

[0002] Lithium-ion batteries are widely used due to their high energy density and long lifespan, but they can experience various malfunctions during manufacturing and use, such as overcharging, lithium plating, and internal short circuits. Failure to identify these issues early can lead to serious safety accidents.

[0003] Existing battery fault diagnosis methods mainly include threshold-based, equivalent circuit model-based, electrochemical impedance spectroscopy (EIS)-based, and data-driven model-based methods. However, these methods each have their own limitations.

[0004] Threshold-based methods rely on externally measurable parameters such as voltage and temperature, and are not sensitive to changes in the internal electrochemical state of the battery, resulting in delayed fault identification and difficulty in detecting early, hidden faults.

[0005] While the method based on the equivalent circuit model has physical significance, the identification of model parameters is easily affected by factors such as operating conditions and noise, and it is difficult to accurately distinguish various coupled fault types by relying solely on a few parameters.

[0006] Electrochemical impedance spectroscopy (EIS) methods can reflect the internal state of a battery, but existing techniques often suffer from insufficient utilization of features. For example, using only impedance at a single frequency point or a small number of equivalent circuit parameters cannot fully characterize the rich information about ohmic impedance, interfacial reactions, and ion diffusion contained in EIS at different frequency bands, resulting in limited distinguishing ability in multi-fault classification tasks.

[0007] Data-driven methods, such as deep learning models, while possessing strong nonlinear fitting capabilities, typically require large-scale, balanced labeled datasets. In the field of battery faults, high-risk fault samples are scarce and difficult to obtain, leading to models being prone to overfitting and having weak minority class identification capabilities. Furthermore, these "black box" models lack interpretability, failing to elucidate the electrochemical mechanisms underlying their fault judgments, thus limiting their credibility in safety-critical applications.

[0008] Therefore, existing technologies generally suffer from problems such as untimely identification of internal electrochemical anomalies in batteries, insufficient ability to classify multiple types of faults, inadequate utilization of EIS information, and lack of interpretability of diagnostic results, making it difficult to meet the needs for early, accurate, and reliable fault diagnosis of battery systems. Summary of the Invention

[0009] This invention provides a battery fault classification method and apparatus, which can solve the shortcomings of existing battery fault diagnosis technologies, such as untimely identification of internal electrochemical anomalies, insufficient multi-fault classification capability, inadequate utilization of EIS information, and lack of interpretability of diagnostic results. The technical solution is as follows: On the one hand, a battery fault classification method is provided, the method comprising: Multi-band electrochemical impedance spectroscopy (EIS) statistical features were extracted from the EIS data of lithium-ion batteries, and electrochemical mechanism feature vectors were extracted based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set was constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. An ensemble learning model based on symmetric decision trees is used, with a weighted multi-classification loss function as the optimization objective, to construct a fault category model for the heterogeneous feature set; During the training process of the fault category model, the feature contribution is iteratively evaluated and fed back based on the SHAP (SHapley Additive exPlanations) value. The fusion weights of high-frequency, mid-frequency, and low-frequency statistical features and electrochemical mechanism feature vectors are dynamically adjusted so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. The system outputs at least one of the following: fault category, confidence probability, risk level, and interpretability diagnostic criteria. These diagnostic criteria are based on SHAP contribution values ​​mapped to high-frequency, mid-frequency, and low-frequency impedance responses and equivalent circuit parameters, revealing the electrochemical mechanism pathway upon which fault classification decisions depend. On the other hand, a battery fault classification device is provided, comprising: The data preprocessing module is used to extract multi-band electrochemical impedance spectroscopy (EIS) statistical features from lithium-ion battery EIS data, and to extract electrochemical mechanism feature vectors based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set is constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. The fault classification model module embeds an ensemble learning model based on a symmetric decision tree, which is used to train the heterogeneous feature set with a weighted multi-class loss function as the optimization objective to construct a fault category prediction model. The dynamic weight optimization module iteratively evaluates the contribution of each feature in the heterogeneous feature set based on the SHAP value during the training process of the fault classification model module, and dynamically adjusts the fusion weights of high-frequency, mid-frequency, and low-frequency statistical features and electrochemical mechanism feature vectors, so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. An interpretable output module is provided to output at least one of the following in the fault classification model: fault category, confidence probability, risk level, and interpretability diagnostic criteria. These diagnostic criteria are mapped from SHAP contribution values ​​to high-frequency, mid-frequency, and low-frequency impedance responses and equivalent circuit parameters, revealing the electrochemical mechanism path upon which the fault classification decision depends. On the other hand, a computer device is provided, comprising a memory and a processor. The memory stores a computer program, and the processor executes the computer program stored in the memory to implement the steps of the battery fault classification method described above.

[0010] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the above-described battery fault classification method.

[0011] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the above-described battery fault classification method.

[0012] The technical solution provided by this invention can bring at least the following beneficial effects: By fusing multi-band statistical features and electrochemical mechanism feature vectors, a more comprehensive heterogeneous feature set is constructed, overcoming the problem that existing technologies rely only on partial EIS information or external signals, resulting in insufficient feature expression capabilities, thereby improving the accuracy of fault classification; By adopting a weighted multi-classification loss function, the model can prioritize high-risk fault types (such as lithium plating and internal short circuits) with few samples during training, effectively improving the sensitivity and detection rate of high-risk fault identification; By iteratively evaluating and feeding back the feature contribution based on SHAP values, By dynamically adjusting the fusion weights of statistical features and electrochemical mechanism feature vectors in the high-frequency, mid-frequency, and low-frequency bands, the model can adaptively focus on the electrochemical response dimension that is most discriminative for the current classification task. This achieves synergistic optimization of multi-scale information and improves the overall performance of the model. By mapping the SHAP contribution value to the impedance response and equivalent circuit parameters in the high-frequency, mid-frequency, and low-frequency bands while outputting the classification results, interpretable diagnostic basis is provided. This maps abstract model decisions to specific frequency band responses and electrochemical parameters, revealing the physical mechanism path of fault occurrence and enhancing the credibility and practicality of the diagnostic results. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the 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 based on these drawings without creative effort.

[0014] Figure 1 This is an overall flowchart of a battery fault classification method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the feature fusion process; Figure 3 This is a schematic diagram of the interpretability analysis process; Figure 4 This is a structural block diagram of a battery fault classification device according to an embodiment of the present invention; Figure 5 This is a schematic diagram of an equivalent circuit model for fitting the electrochemical impedance spectrum of a battery according to an embodiment of the present invention; Figure 6 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0015] 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 some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] As mentioned above, the technical problem to be solved by the present invention is to provide a method and apparatus for early, accurate and interpretable battery fault classification, addressing the shortcomings of existing battery fault diagnosis technologies, such as untimely identification of internal electrochemical anomalies, insufficient ability to classify multiple faults, inadequate utilization of EIS information, and lack of interpretability of diagnostic results.

[0017] The following describes the specific implementation of the above concept.

[0018] Please refer to Figure 1 The present invention provides a battery fault classification method, the method comprising: Step 100: Extract multi-band electrochemical impedance spectroscopy (EIS) statistical features based on the EIS data of lithium-ion batteries, and extract electrochemical mechanism feature vectors based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. Construct a heterogeneous feature set based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. Step 101: Using an ensemble learning model based on symmetric decision trees, and with a weighted multi-classification loss function as the optimization objective, a fault category model is constructed for the heterogeneous feature set; Step 102: During the training process of the fault category model, the feature contribution is iteratively evaluated and fed back based on the SHAP value, and the fusion weights of the statistical features and electrochemical mechanism feature vectors of the high-frequency, mid-frequency, and low-frequency bands are dynamically adjusted so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. Step 103: Output the fault category, confidence probability, risk level, and interpretability diagnostic criteria. The diagnostic criteria are based on the SHAP contribution value mapped to the impedance response and equivalent circuit parameters of the high-frequency, mid-frequency, and low-frequency bands, revealing the electrochemical mechanism path on which the fault classification decision depends.

[0019] The following is a detailed description Figure 1 The execution method of each step is shown. This application aims to provide an advanced battery fault classification method to address the problems in existing technologies, such as delayed identification of internal electrochemical anomalies, limited ability to distinguish between multiple fault types, and lack of interpretability of diagnostic results.

[0020] The core idea of ​​this method is to achieve accurate, fast and reliable classification of various fault modes of lithium-ion batteries by deeply integrating the data-driven features and physical mechanism features of electrochemical impedance spectroscopy (EIS) and using an adaptive and interpretable machine learning framework.

[0021] First, for step 100, multi-band electrochemical impedance spectroscopy (EIS) statistical features are extracted based on the EIS data of lithium-ion batteries, and electrochemical mechanism feature vectors are extracted based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set is constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. Electrochemical impedance spectroscopy (EIS) in this embodiment refers to a technique that calculates the complex impedance of a battery at different frequencies by applying a small-amplitude, wide-bandwidth alternating current signal as excitation to the battery and simultaneously measuring the voltage response of the battery system under this excitation. EIS can non-destructively probe the electrochemical processes inside the battery. Different frequency bands of its spectral response correspond to different physicochemical processes, such as ohmic resistance in the high-frequency band, charge transfer and interfacial reactions in the mid-frequency band, and ion diffusion processes in the low-frequency band. Therefore, it is a powerful tool for assessing the internal state of a battery and diagnosing faults.

[0022] In this embodiment, before specifically fusing the heterogeneous feature set, a small-sample fault classification dataset is constructed based on the electrochemical impedance spectroscopy (EIS) data of lithium-ion batteries. In one specific implementation, EIS data of lithium-ion batteries under different health and fault states are collected, and a labeled training sample set is established: in, For the sample size, Let i be the feature vector of the i-th battery sample. Let be the fault category label for the i-th battery sample.

[0023] The set of fault categories is represented as: Where K represents the number of fault categories.

[0024] In this embodiment, the fault categories include: normal; connection abnormality; overcharge; over-discharge; lithium plating; internal short circuit; and precursor to thermal runaway. Depending on the actual application, these can be expanded to include copper plating, external short circuit, leakage, electrode wrinkling, and SEI (Solid Electrolyte Interphase) film abnormalities.

[0025] The electrochemical impedance spectroscopy (EIS) data consists of complex impedances at multiple frequency points: ;in, The real part of the impedance. is the imaginary part of the impedance, f is the frequency, and j is the imaginary unit.

[0026] The impedance amplitude is: ; The phase angle is: , Record the battery's State of Charge (SOC), State of Health (SOH), temperature, sampling time, and fault tag.

[0027] The subsequent extraction of multi-band electrochemical impedance spectroscopy (EIS) statistical features and electrochemical mechanism feature vectors based on the equivalent circuit model (ECM) were all calculated and fitted based on this original data. This ensures that all features are from the same source, experimental conditions are traceable, and model training and diagnosis are based on the same physical experimental system, thereby guaranteeing the reliability and interpretability of the fault classification results.

[0028] Next, multi-band electrochemical impedance spectroscopy (EIS) statistical features are extracted. In a preferred embodiment, the statistical features of the real part, imaginary part, impedance amplitude, and phase angle extracted in the high-frequency, mid-frequency, and low-frequency bands are designed to characterize the synergistic changes in the electrochemical response characteristics of the battery at different time scales in a data-driven manner. Specifically, the statistical features in the high-frequency band (e.g., 1 kHz to 10 kHz) mainly capture ohmic response features related to electrolyte ionic conductivity, electrode material electronic conductivity, and contact resistance; the statistical features in the mid-frequency band (e.g., 1 Hz to 1 kHz) focus on reflecting the charge transfer process at the electrode / electrolyte interface, SEI film characteristics, and interfacial reaction kinetics such as double-layer capacitance effect; while the statistical features in the low-frequency band (e.g., 0.01 Hz to 1 Hz) are mainly related to the solid-phase diffusion process of lithium ions within the electrode active material. By conducting a synergistic analysis of the characteristics of these three frequency bands, a snapshot of the coupled state of multiple physical processes inside the battery can be constructed, solving the problem that a single frequency point or single parameter cannot fully describe the battery state, thus providing richer criteria for distinguishing faults of different mechanisms.

[0029] Specifically, the EIS frequency range is divided into high-frequency band, mid-frequency band, and low-frequency band: in, Indicates high frequency band, Indicates mid-frequency band, Indicates low frequency band.

[0030] In one implementation, the frequency band can be divided as follows:

[0031] For each frequency band, statistical features of the real part of impedance, the imaginary part of impedance, impedance amplitude, and phase angle are extracted, including maximum value, minimum value, mean, variance, median, skewness, kurtosis, differential, and integral features.

[0032] The average real part of the impedance is: The mean of the imaginary part of the impedance is: The impedance amplitude integral characteristic is: In the discrete case, numerical integration can be used for calculation.

[0033] The phase angle change characteristics are as follows: in, and These represent the maximum and minimum frequencies within frequency band B, respectively.

[0034] This yields the high-frequency, mid-frequency, and low-frequency characteristics: ; Among them, the high-frequency band characteristics mainly reflect ohmic impedance, connection status and current collector anomalies; the mid-frequency band characteristics mainly reflect SEI film, charge transfer process and interface reaction; and the low-frequency band characteristics mainly reflect diffusion process, polarization accumulation and precursors of thermal runaway.

[0035] Next, based on the electrochemical mechanism feature vector extracted from the equivalent circuit model, in order to extract electrochemical terminal information in EIS and enhance the physical interpretability of EIS features, the equivalent circuit model (ECM) in this embodiment refers to a circuit model that uses ideal or non-ideal electronic components such as resistors, capacitors, and inductors and their topological combinations to simulate and characterize the impedance response characteristics of a complex electrochemical system (such as a lithium-ion battery).

[0036] In another preferred embodiment, the process of extracting the electrochemical mechanism feature vector of the equivalent circuit model aims to establish a causal mapping relationship between the observable EIS impedance characteristics and the electrochemical states inside the battery that cannot be directly measured.

[0037] Further, please refer to Figure 5 In one specific embodiment, the equivalent circuit model may employ a model that includes an ohmic internal resistance. 501, SEI film impedance 502 and SEI film non-ideal capacitance 503 parallel circuit, charge transfer impedance 504 and non-ideal double-layer capacitors 505 parallel circuit, and Warburg diffusion impedance A structure consisting of 506 series connectors. This model, namely... It can effectively simulate the impedance behavior of typical lithium-ion batteries.

[0038] Among them, ohmic internal resistance 501 represents the combined resistance of the current collector, electrode material, diaphragm, and electrolyte. SEI film impedance. 502 and SEI film non-ideal capacitance The parallel circuit composed of 503 electrodes simulates the impedance characteristics of the SEI film on the negative electrode surface. (Charge transfer impedance) 504 and non-ideal double-layer capacitors The parallel circuit composed of 505 elements characterizes the kinetic impedance during electrochemical reactions at the electrode / electrolyte interface. Warburg diffusion impedance. Section 506 describes the diffusion process of lithium ions within the particles of the electrode active material. By identifying these parameters, the performance state of different components within the battery can be directly quantified.

[0039] constant phase element The impedance of a (Constant Phase Element) is expressed as: in, Here, CPE is the coefficient, and n is the diffusion index. ω=2πf ω is the angular frequency.

[0040] Warburg diffusion impedance can be expressed as: in, for Warburg Diffusion coefficient.

[0041] The equivalent circuit model (ECM) total impedance model is obtained as follows: ECM parameters are identified by minimizing the error between the measured impedance and the model impedance. The objective function can be expressed as: 。

[0042] in, The set of parameters to be identified, i.e. The final electrochemical mechanism eigenvector obtained. This provides the model with deep physical insights. Next, high-frequency features, mid-frequency features, low-frequency features, and electrochemical mechanism feature vectors are fused to construct a heterogeneous feature set: .

[0043] The heterogeneous feature set in this embodiment refers to a collection that integrates data features from multiple sources and types. In this application embodiment, it specifically refers to a comprehensive feature set formed by combining features directly extracted from electrochemical impedance spectroscopy (EIS) data based on statistical distribution, such as the mean, variance, and kurtosis of the real part, imaginary part, amplitude, and phase of impedance in different frequency bands, with mechanistic parameters (e.g., ohmic internal resistance, SEI film impedance, charge transfer impedance, etc.) obtained by fitting a physical model (i.e., equivalent circuit model) that have clear electrochemical physical significance. This integration allows the feature set to contain both macroscopic distribution information of the data and microscopic physical mechanisms, thereby enabling a more comprehensive characterization of the battery's health status.

[0044] Specifically, the construction of the heterogeneous feature set involved in step 100 above comprises two parallel paths. The first is from a data-driven perspective, extracting statistical features of the real part of impedance, the imaginary part of impedance, impedance amplitude, and phase angle at high, mid, and low frequencies, respectively. The second is from a physical model perspective, extracting electrochemical mechanism feature vectors with clear physical meaning by constructing and fitting an equivalent circuit model. In this way, the original EIS curve, which is difficult to understand intuitively, is transformed into a structured, information-rich feature vector, solving the problems of high dimensionality and information redundancy in the original data, and significantly improving the characterization ability of the features.

[0045] Next, step 101 is executed, in which an ensemble learning model based on symmetric decision trees is used to construct a fault category model for the heterogeneous feature set with a weighted multi-classification loss function as the optimization objective. In one alternative implementation, the ensemble learning model of the symmetric decision tree specifically employs the CatBoost architecture. CatBoost is an advanced gradient boosting decision tree (GBDT) algorithm, unique in that it uses symmetric decision trees as base learners. At the same depth level in each decision tree, all nodes are split using the exact same features and thresholds.

[0046] In one specific implementation, the multi-band weighted EIS feature vector enhanced by ECM is used as input, and the battery fault category label is used as output to construct a CatBoost multi-classification model.

[0047] The CatBoost model consists of multiple symmetric decision trees, and its model function is expressed as follows: in, For the number of decision trees, For learning rate, For the first A symmetric decision tree.

[0048] Symmetric decision trees use the same splitting features and splitting thresholds for all nodes at the same depth, which makes the model highly efficient for inference and suitable for deployment in vehicle-mounted BMS or edge computing units.

[0049] For multi-class classification tasks, CatBoost performs sample... Output the predicted score for each fault category: The predicted score is converted into class probability using the Softmax function: in, Indicates sample Belongs to the Predicted probability of class-specific faults.

[0050] The final fault category identification result is as follows: Considering that battery fault data contains many normal samples and few high-risk fault samples, this invention uses a weighted multi-class cross-entropy loss function to train the CatBoost model. In a specific embodiment, the weighted multi-class loss function can be defined as: .

[0051] in, As an indicator function, when the sample The true category is the first The value is 1 if the condition is met, and 0 otherwise. For the first The training weights for fault classes are calculated as follows: .

[0052] Weight here It consists of two weighted parts: Part 1 It is the inverse class frequency based on the number of samples, used to solve the sample imbalance problem. When the number of samples of a certain fault class... When the number of cases is low, this factor increases, thereby increasing the model's focus on that minority class; Part Two It is a preset physical severity coefficient with values ​​between [0,1], used to reflect the danger of a fault. When a fault category has a high degree of danger, Take a larger value to improve the model's sensitivity in identifying this high-risk fault category. This is a balancing factor used to adjust the influence of these two weights. Through this design, the loss function can guide the model training towards optimization that can both identify minority classes and prioritize the detection of high-risk faults.

[0053] The design of the weighted multi-class loss function described above, by introducing the physical severity weight of the fault category, enables the model to go beyond simple data statistical distribution during training and optimize from a higher dimension of safety.

[0054] The principle behind this approach is that different battery faults pose vastly different threats to system safety. For example, a slight capacity decay poses a significantly higher risk than an internal short circuit that is about to cause thermal runaway. Traditional loss functions balance based solely on the number of samples, which may lead the model to sacrifice its ability to identify a few high-risk faults in order to improve its performance on most normal or low-risk samples. This solution assigns a higher weight coefficient to high-risk fault categories such as lithium plating, internal short circuits, and precursors to thermal runaway. This forces the model to be penalized much more severely if it misclassifies a high-risk fault sample during optimization than if it misclassifies a normal sample. This mechanism automatically enhances the model's sensitivity to high-risk faults, enabling it to issue critical safety warnings earlier and more reliably in practical applications.

[0055] Next, step 102 is executed. During the training process of the fault category model, the feature contribution is iteratively evaluated and fed back based on the SHAP value. The fusion weights of the statistical features and electrochemical mechanism feature vectors of the high-frequency, mid-frequency, and low-frequency bands are dynamically adjusted so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. In one specific implementation, during model training, the model is updated in the t-th iteration as follows: in, For the previous model, This is the symmetric decision tree obtained in the current round of training.

[0056] The negative gradient is represented as: By fitting the negative gradient direction round by round, the model continuously reduces the weighted classification loss and improves its ability to identify a few high-risk fault categories. The SHAP value in this embodiment refers to a model interpretation method derived from cooperative game theory. Its core idea is to calculate the marginal contribution of each input feature to the model's single prediction result, i.e., quantifying the impact of adding that feature on the final prediction output. In this embodiment, the SHAP value is used as a diagnostic tool, not only to evaluate the contribution strength of statistical features and electrochemical mechanism feature vectors of each frequency band to specific fault classification decisions, thereby achieving interpretability analysis of the model prediction results, but also innovatively used as a feedback signal to dynamically guide and optimize the fusion weights of different feature groups during model training.

[0057] In a preferred embodiment, the mechanism for iteratively evaluating and feeding back feature contributions based on SHAP values ​​involves systematically calculating the overall contribution strength of each feature group in the heterogeneous feature set to the current model classification result after each (or several) rounds of model training. Specifically, the initial feature vectors mentioned above are: Input the CatBoost classification model for initial training, and calculate the SHAP value of each feature based on the trained CatBoost model. S .

[0058] Next, based on the relative importance (i.e., influence) of each group of features, their weights are dynamically adjusted during feature fusion before the next round of training. Essentially, this process allows the model to guide its next learning steps through self-reflection (SHAP evaluation), gradually allocating more attention resources (i.e., higher weights) to the electrochemical response dimensions that have proven to be most effective in distinguishing different fault types, thus achieving an adaptive feature selection and optimization process.

[0059] Furthermore, dynamically adjusting the fusion weights of multi-band features and mechanistic parameters aims to avoid the dominance or bias of a single feature dimension in model decision-making, achieving synergistic optimization of multi-scale electrochemical information. In practice, the four feature groups—high-frequency, mid-frequency, low-frequency statistical features, and equivalent circuit parameters—describe the battery state from different angles and scales. For example, high-frequency features may be highly sensitive to detecting loose connections, while low-frequency features and the Warburg coefficient are crucial for diagnosing electrode aging and lithium-ion diffusion obstruction. If fixed, equal weights are used, the model may not be able to fully utilize the key information under specific faults. This approach ensures that the model can flexibly rely on the most relevant feature combinations when diagnosing different faults by grouping and evaluating these four feature groups and adjusting their weights based on their SHAP contribution values.

[0060] In one specific embodiment, the iterative update process of dynamically adjusting weights can be precisely mathematically quantified. After each round of model training, the SHAP contribution values ​​of four feature subsets—high-frequency band, mid-frequency band, low-frequency band, and equivalent circuit parameters—are calculated and denoted as follows: , , , Then, the fusion weights of each feature subset are updated according to a normalization formula, for example: 。

[0061] in, Representing a feature subset g∈{H,M,L,E} , Indicates the number of iterations. It is a small smoothing factor to prevent the denominator from being zero, and C is the normalization constant. The updated weights... The feature fusion input is used for the next round of training, and this process is repeated until the weight change between two adjacent iterations is reached. Less than a preset convergence threshold If the preset maximum number of iterations is reached, then the iteration stops. 。

[0062] In this way, the feature weights can automatically converge to an optimal or suboptimal stable state.

[0063] Finally, the SHAP dynamically weighted ECM-enhanced multi-band EIS feature vector is obtained: ; in, The high-frequency feature weights after convergence; Weights of mid-frequency features after convergence The low-frequency feature weights after convergence; These are the feature weights of the converged ECM parameters.

[0064] To further enhance the intelligence level of the model and the reliability of the diagnostic results, this embodiment of the application introduces an adaptive feedback mechanism based on SHAP values ​​during the training process of the fault category model. Specifically, the feature contribution is iteratively evaluated and fed back based on the SHAP values, and the fusion weights of the four sets of features—high-frequency, mid-frequency, low-frequency statistical features, and electrochemical mechanism feature vectors—are dynamically adjusted according to the evaluation results.

[0065] The principle behind this design is to allow the model to autonomously learn and discover the most discriminative information dimensions for the current classification task during training iterations. For example, when distinguishing certain faults, mid-frequency interface response information may be more important than low-frequency diffusion information. Through this feedback mechanism, the model can adaptively focus on these key information, thereby optimizing the feature fusion strategy and improving the final classification performance. This solves the problem in existing methods where feature fusion weights are usually manually set or fixed, making them unable to adapt to different fault diagnosis tasks.

[0066] Step 103: Output at least one of the following: fault category, confidence probability, risk level, and interpretability diagnostic criteria. The diagnostic criteria are based on the SHAP contribution value mapped to the impedance response and equivalent circuit parameters of the high-frequency, mid-frequency, and low-frequency bands, revealing the electrochemical mechanism path on which the fault classification decision depends.

[0067] When the model is trained and applied to new battery samples, it outputs a comprehensive diagnostic report. This report includes at least one of the following: fault category, confidence probability, risk level, and interpretability diagnostic criteria. The fault category specifies the most likely specific problem the battery will encounter; the confidence probability quantifies the certainty of the judgment; and the risk level combines the inherent danger of the fault type with the predicted probability, providing an overall threat assessment to system safety. Crucially, the interpretability diagnostic criteria are based on mapping SHAP contribution values ​​to impedance responses and specific equivalent circuit parameters in the high-frequency, mid-frequency, and low-frequency bands. This clearly reveals the key electrochemical mechanisms underlying the model's fault classification decision. For example, the report might indicate that the judgment of this lithium plating fault is primarily based on a significant increase in mid-frequency impedance and an abnormal charge transfer impedance (Rct) parameter. This provides clear and reliable physical guidance for subsequent battery maintenance, repair, or failure analysis, addressing the pain points of traditional machine learning models operating as black boxes and their results being difficult for engineers to trust and adopt.

[0068] In one alternative implementation, interpretable diagnostic criteria are generated by establishing mapping rules between the distribution patterns of SHAP values ​​across different frequency bands and electrochemical parameters and specific electrochemical degradation mechanisms. This makes the fault classification result no longer a label, but traceable to the actual physical degradation path within the battery. SHAP analysis not only tells us which features are important, but also reveals how the values ​​of these features affect the final decision (e.g., whether they positively promote or negatively inhibit it). By systematically analyzing the SHAP value distribution patterns of a large number of known fault samples, an expert rule base can be summarized and solidified. For example, when the model identifies a lithium plating fault, if the SHAP analysis shows that the contribution values ​​of mid-frequency features and the contribution values ​​of charge transfer impedance Rct are both significantly positive, this forms a strong chain of evidence, indicating that the model's decision logic is consistent with the understanding in electrochemical theory that lithium plating leads to the deterioration of interfacial reaction kinetics. This mechanism greatly enhances the credibility of the diagnostic results.

[0069] In one specific implementation, in order to improve the interpretability of fault classification results, this invention calculates SHAP values ​​while the CatBoost model outputs fault categories and probabilities, to explain the contribution of each EIS feature and ECM parameter to the classification results.

[0070] For the For class-specific faults, the output of the CatBoost model can be represented as: ; The features are divided into high-frequency feature sets according to their source. Mid-frequency feature set Low-frequency feature set and ECM parameter feature set And calculate the contribution strength of different feature groups to the target fault category: ; ; ; ; in, , , , These represent the effects of high-frequency, mid-frequency, low-frequency, and ECM parameters on the first... The contribution intensity of fault identification.

[0071] Furthermore, these mapping rules can be specified as a series of IF-THEN logical judgments. For example, a diagnostic rule base can be established as follows: If the high-frequency band SHAP contribution value High, and the average real part of the impedance in the high-frequency band is... Ohmic resistance If the SHAP contribution value of the real part of the high-frequency impedance is significantly higher than other characteristics, the system will determine that the fault is related to abnormal connection, increased ohmic impedance or abnormal collector. If the mid-frequency band SHAP contribution value and ECM parameter SHAP contribution value And SEI film impedance Charge transfer impedance or SEI membrane CPE parameters Q seihe n sei If the SHAP contribution values ​​all show a significant positive correlation, then the fault is determined to be related to excessive SEI film thickening or lithium plating on the negative electrode. If the contribution value of low-frequency band SAP The Warburg diffusion coefficient is relatively high. If the SHAP contribution values ​​are all significant, the fault is determined to be related to the obstruction of lithium-ion diffusion in the solid phase or an early precursor to thermal runaway. Using this set of rules, the final diagnostic report can be presented to the user in a clear and physically intuitive way.

[0072] In another preferred embodiment, the risk level is calculated by weighting and aggregating the predicted confidence levels of each fault category output by the model with their corresponding preset physical severity, thereby obtaining a quantitative and comprehensive risk value. The purpose of this design is to enable the system not only to answer the question of battery fault classification, but also to answer the question of risk assessment of the fault on the system. For example, one battery is diagnosed as "minor over-discharge" with a 90% confidence level, while another battery is diagnosed as "internal short circuit" with a 60% confidence level. Although the latter has a lower confidence level, its final calculated comprehensive risk value may be higher because the physical severity of "internal short circuit" is much higher than that of "minor over-discharge." This method makes risk assessment more scientific and comprehensive, helping management systems or maintenance personnel to more accurately determine the priority of handling, and achieving refined quantitative management of the degree of battery safety threat.

[0073] In one specific embodiment, for the battery sample to be identified The CatBoost model outputs the probability of it belonging to each type of fault: The fault category output is:

[0074] Furthermore, the comprehensive risk value can be calculated based on the probability of each fault category and the fault severity coefficient using the formula... The calculation yielded the following result. It is the model's prediction of the first The probability of a failure type (i.e., confidence level), and This is the first Pre-defined, standardized physical severity coefficients for each type of fault (e.g., normal = 0, minor over-discharge = 0.2, lithium plating = 0.7, internal short circuit = 0.9, pre-existing thermal runaway = 1.0). The overall risk value is then calculated. Then, based on a series of preset thresholds Map it to different risk levels, for example: normal ( Mild risk ), moderate risk ), severe risk ( ) and danger ( These thresholds can be flexibly set according to specific battery safety standards, application scenarios, or user risk preferences. In this way, the system can output an intuitive, graded risk status, facilitating rapid decision-making.

[0075] Please see Figure 1This application provides a specific implementation of a battery fault classification method, which aims to solve the problems of inaccurate, untimely, and difficult-to-interpret battery fault diagnosis in the prior art. The method first acquires the raw electrochemical impedance spectroscopy (EIS) data of the lithium-ion battery under test using a data acquisition device. This data forms the basis for all subsequent analyses. Please refer to... Figure 2 The acquired raw EIS data 201 is then fed into the feature extraction process.

[0076] The first step of this method is to fuse the electrochemical impedance spectroscopy (EIS) data of lithium-ion batteries into a heterogeneous feature set. For example... Figure 1 As shown in steps S1 and S2, this process is performed in parallel. On one hand, the multi-band feature extraction module 202 divides the original EIS data according to a preset frequency range (e.g., high frequency, mid frequency, low frequency), and calculates a series of statistical quantities, such as mean, variance, kurtosis, and skewness, for the real part, imaginary part, amplitude, and phase angle of the impedance in each frequency band, thereby obtaining high-frequency band features 202a, mid-frequency band features 202b, and low-frequency band features 202c. These statistical features describe the shape and distribution characteristics of the impedance spectrum at different scales from a data-driven perspective.

[0077] On the other hand, such as Figure 1 Step S2 and Figure 2 As shown in the ECM parameter fitting module 203, the same original EIS data 201 is used to fit a predefined equivalent circuit model (ECM). The model parameters are adjusted using optimization algorithms (such as nonlinear least squares) to achieve the best match between the theoretical impedance spectrum and the measured spectrum, thereby retrieving a set of mechanistic parameters with clear electrochemical physical significance, forming the ECM parameter features 203a. These parameters, such as ohmic internal resistance and charge transfer impedance, reveal the internal state of the battery from a physical mechanism perspective.

[0078] Subsequently, as Figure 1 Step S3 and Figure 2 As shown in the SHAP dynamic weighted fusion module 204, the extracted high-frequency features 202a, mid-frequency features 202b, low-frequency features 202c, and ECM parameter features 203a are weighted and fused to form a unified, information-rich CatBoost input feature vector 205. This fusion process is not a simple concatenation, but is guided by a dynamic weight adjustment mechanism to ensure that features from different sources can work together in an optimal way.

[0079] The second step of this method is to use an ensemble learning model based on symmetric decision trees to construct a fault category model for the heterogeneous feature set. For example... Figure 1As shown in steps S4 and S5, the constructed feature vectors are input into an ensemble learning model for training. This model uses a specially designed weighted multi-classification loss function as its optimization objective. This loss function is designed to comprehensively consider the imbalance in the number of samples for different fault categories in practical applications, as well as the differences in the physical hazard levels of different fault types. In this way, the model focuses more on fault types with fewer samples but higher hazard levels during training, thereby improving its ability to detect critical faults and overcoming the drawback of traditional models that only pursue overall accuracy and may ignore high-risk areas.

[0080] The third step of this method involves iteratively evaluating and providing feedback on the feature contribution based on the SHAP value during the training of the fault category model, dynamically adjusting the feature fusion weights. For example... Figure 1 The SHAP contribution feedback from step S5 to step S3 forms a closed-loop optimization system. After each training round, the SHAP method is used to analyze the decision logic within the model, quantifying the contribution of high, medium, and low frequency statistical feature groups and electrochemical mechanism feature vector groups to the classification results. Then, based on the magnitude of this contribution, their weights are adjusted during feature fusion before the next training round. This process enables the model to autonomously and adaptively focus on the electrochemical response dimension that is most discriminative for the current fault discrimination task, achieving automation and intelligence in feature engineering and solving the subjectivity and suboptimal problems caused by manually setting weights.

[0081] The fourth step of this method is to finally output a multi-dimensional, interpretable diagnostic result. For example... Figure 1 As shown in step S6, when the trained model is applied to a new test sample, it not only outputs the most likely fault category and its corresponding confidence probability, but also further calculates a comprehensive risk level. More importantly, it provides an interpretable diagnostic basis. Please refer to... Figure 3 After receiving the sample to be detected 302, the trained CatBoost model 301 analyzes the high-frequency contribution through the SHAP value calculation module 303. 304. Mid-frequency contribution 305. Low-frequency contribution 306 and ECM contributions 307. The explanation output module 308, based on these contribution values, maps them to specific high, medium, and low frequency impedance responses and equivalent circuit parameters, thereby clearly revealing the key electrochemical mechanism pathways upon which this fault classification decision relies. For example, the diagnostic report can indicate a lithium plating fault, primarily based on changes in the mid-frequency impedance spectrum morphology (…). (Major contribution) and the charge transfer impedance Rct parameter is significantly increased ( (The RCT made a significant contribution). This diagnostic approach, which presents the reasons behind the findings, greatly enhances the credibility and practical value of the results, providing a solid scientific basis for subsequent maintenance decisions.

[0082] Please see Figure 4 This application also provides a battery fault classification device 400, which serves as the physical carrier of the aforementioned method. This device can be a standalone hardware unit or a functional system integrated into a battery management system (BMS), power conversion system (PCS), or cloud server. In a typical implementation, the battery fault classification device 400 includes a hardware unit and functional modules running thereon.

[0083] The hardware unit may include at least one processor 401, a memory 402, and a data interface 403. The processor 401, such as a high-performance microcontroller (MCU) or digital signal processor (DSP), is the computing core of the entire device, responsible for executing instructions stored in the memory 402. The memory 402 may include flash memory and random access memory (RAM) for storing the operating system, application programs (i.e., the functional modules described below), and temporary data and model parameters generated during operation. The data interface 403, such as a CAN bus interface, Ethernet interface, or serial interface, is responsible for communicating with external devices (such as a battery sampling unit, a host computer, or a cloud platform), receiving raw EIS data, and sending the final diagnostic results.

[0084] The functional modules running on processor 401 implement the method proposed in this invention. These modules include: a data preprocessing module 410, a fault classification model module 420, a dynamic weight optimization module 430, and an interpretable output module 440.

[0085] Data preprocessing module 410 for electrochemical impedance spectroscopy based on lithium-ion batteries EIS The data extracts multi-band electrochemical impedance spectroscopy (EIS) statistical features and extracts electrochemical mechanism feature vectors based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set is constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. The fault classification model module 420 has an embedded ensemble learning model based on a symmetric decision tree, which is used to train the heterogeneous feature set with a weighted multi-classification loss function as the optimization objective to build a fault category prediction model. The dynamic weight optimization module 430, during the training process of the fault classification model module, iteratively evaluates the contribution of each feature in the heterogeneous feature set based on the SHAP value, and dynamically adjusts the fusion weight of high-frequency, mid-frequency, and low-frequency statistical features and electrochemical mechanism feature vectors, so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. The interpretable output module 440 is used to output at least one of the following in the fault classification model: fault category, confidence probability, risk level, and interpretable diagnostic criteria. The diagnostic criteria are based on the SHAP contribution value and mapped to the impedance response and equivalent circuit parameters of the high-frequency, mid-frequency, and low-frequency bands, revealing the electrochemical mechanism path on which the fault classification decision depends.

[0086] The data preprocessing module 410 corresponds to the first step in the method. It receives raw electrochemical impedance spectroscopy (EIS) data from the battery measurement unit via data interface 403. After receiving the data, the module performs feature extraction and fusion tasks, namely, extracting statistical features of impedance in the high, medium, and low frequency bands respectively, and simultaneously extracting electrochemical mechanism feature vectors by fitting an equivalent circuit model. Finally, it fuses these two types of features into a heterogeneous feature set for use by subsequent modules.

[0087] The fault classification model module 420 embeds a pre-trained or online-trainable ensemble learning model based on a symmetric decision tree. This module receives a heterogeneous feature set generated by the data preprocessing module 410 and uses its internal model to calculate and predict the battery fault category. The model's training process uses a weighted multi-classification loss function that simultaneously considers sample imbalance and the physical risk of the fault as its optimization objective, ensuring sensitivity in identifying high-risk faults.

[0088] The dynamic weight optimization module 430 works closely with the fault classification model module 420. During the model training or online update phase, this module is responsible for executing the third step of the method. After the fault classification model module 420 completes one round of training, it calls the SHAP algorithm to evaluate the contribution of each feature group (high-frequency, mid-frequency, low-frequency, and ECM parameters) in the heterogeneous feature set. Based on the evaluation results, this module dynamically calculates and updates the weights used for feature fusion and feeds the new weights back to the data preprocessing module 410 for the next round of feature fusion. This closed-loop feedback mechanism enables the entire device to have adaptive learning and optimization capabilities.

[0089] The interpretable output module 440 is responsible for generating the final diagnostic report. After the fault classification model module 420 obtains the preliminary classification results and confidence probabilities, this module further invokes SHAP analysis to map the abstract feature contribution values ​​to specific electrochemical mechanisms, generating a clear and easy-to-understand diagnostic basis. Simultaneously, it calculates a comprehensive risk level based on the predicted fault probability and the preset fault severity. Finally, this module integrates the fault category, confidence probability, risk level, and interpretable diagnostic basis, and sends it to the user or upper-level management system through data interface 403.

[0090] In a specific application scenario, this application provides a specific embodiment integrating all preferred technical solutions for diagnosing a lithium-ion battery cell suspected of having a "lithium plating" fault. The entire process will be combined with the attached... Figures 1 to 5 Please provide an explanation.

[0091] First, the vehicle's battery management system (BMS) or offline testing equipment applies an AC excitation signal with a frequency range from 10kHz to 0.01Hz to the target battery and acquires its impedance response to obtain raw EIS data. This process corresponds to... Figure 1 The data input step (101) in the middle.

[0092] Next, the data is sent to the processor 401 of the battery fault classification device 400. The data preprocessing module 410 begins operation. On one hand, it divides the EIS data into high-frequency bands (1kHz-10kHz), mid-frequency bands (1Hz-1kHz), and low-frequency bands (0.01Hz-1Hz), and calculates a total of 48 statistical characteristics for each frequency band, including the real and imaginary parts of impedance, the mean, variance, and kurtosis of amplitude and phase. Figure 2 The multi-band feature extraction module 202 in the middle. On the other hand, it calls the ECM parameter fitting module 203, using Figure 5 shown The equivalent circuit model is fitted to the EIS data by executing the formula. The optimization process identifies factors including ohmic internal resistance. 501, SEI film impedance 502. Charge transfer impedance Eight electrochemical mechanism feature vectors, including 504.

[0093] Then, the SHAP dynamic weighted fusion module 204 (whose weights have converged through iterative optimization during offline training) weights and combines these 48 statistical features and 8 mechanistic parameters according to the optimal weights to generate a final 56-dimensional feature vector, namely the CatBoost input feature vector 205. This process corresponds to... Figure 1Steps S2 (102), S3 (103), and S4 (104) in the process.

[0094] The fault classification model module 420 receives this 56-dimensional feature vector. Its built-in CatBoost model (a symmetric decision tree ensemble model) is carefully trained. During training, it uses the formula... Defined weighted loss function, Among them, weight From the formula The calculations specifically increased the weighting of high-risk faults such as lithium plating. Simultaneously, feature fusion weighting... Through formula It is being iteratively optimized.

[0095] The model infers from the input feature vector and outputs the probability of each fault category, for example: {Normal: 0.05, Connection Abnormality: 0.02, Lithium Plating: 0.85, Internal Short Circuit: 0.08}. Based on this, the fault category is initially determined to be "Lithium Plating" with a confidence level of 85%. This corresponds to... Figure 1 Steps S5 (105) and S6 (106) in the process.

[0096] Finally, the interpretable output module 440 is activated by outputting interpretable diagnostic criteria based on the SHAP value. It first calls the SHAP value calculation module 303 to analyze the attribution of this "lithium plating" judgment. For example... Figure 3 As shown, the analysis results (304, 305, 306, 307) may indicate the contribution values ​​of the mid-frequency band characteristics. and ECM parameter contribution value Much higher than high and low frequencies. Internally, the highest SHAP values ​​were found for charge transfer impedance Rct and SEI film impedance Rsei. Based on the preset interpretation rules (as described in claim 13), the system generated a diagnostic criterion: "Lithium plating is determined, primarily due to abnormal impedance characteristics in the mid-frequency range, specifically a significant increase in charge transfer impedance Rct and SEI film impedance Rsei, consistent with the mechanism of lithium plating leading to deterioration of electrochemical reaction kinetics." Then, the module used the formula... Calculate the overall risk value, taking into account the high severity coefficient of "lithium plating". Even with an internal short circuit, there's an 8% probability, and the final risk value R might fall into the "severe risk" range. Ultimately, the device outputs a complete report via data interface 403, including "Fault Category: Lithium Plating," "Confidence Level: 85%," "Risk Level: Severe Risk," and the detailed diagnostic basis mentioned above. This corresponds to... Figure 1 Steps S7 (107) and S8 (108) in the process.

[0097] Embodiments of this application also provide a computer device, please refer to... Figure 6 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, at least one program, code set or instruction set being loaded and executed by the processor to implement a battery fault classification method provided in the above-described method embodiments.

[0098] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement a battery fault classification method provided in the above-described method embodiments.

[0099] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform a battery fault classification method as described in any of the above embodiments.

[0100] This embodiment achieves significant comprehensive technical effects by superimposing all the aforementioned preferred technical features. By fusing multi-band statistical features and electrochemical mechanism feature vectors, a more comprehensive heterogeneous feature set is constructed, overcoming the problem of insufficient feature representation ability caused by existing technologies relying only on partial EIS information or external signals, thereby improving the accuracy of fault classification. By adopting a weighted multi-classification loss function, the model can prioritize high-risk fault types (such as lithium plating and internal short circuits) with few samples during training, effectively improving the sensitivity and detection rate of high-risk faults. By iteratively evaluating and feeding back the feature contribution based on SHAP values, the high-frequency band, mid-frequency band, and... The fusion weights of low-frequency statistical features and electrochemical mechanism feature vectors enable the model to adaptively focus on the electrochemical response dimension that is most discriminative for the current classification task, achieving synergistic optimization of multi-scale information and improving the overall performance of the model. By mapping the SHAP contribution value to high-frequency, mid-frequency, and low-frequency impedance responses and equivalent circuit parameters while outputting classification results, interpretable diagnostic basis is provided. This maps abstract model decisions to specific frequency band responses and electrochemical parameters, revealing the physical mechanism path of fault occurrence and enhancing the credibility and practicality of the diagnostic results.

[0101] It should be noted that the battery fault classification device provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the battery fault classification device and the battery fault classification method embodiment provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiment, which will not be repeated here.

[0102] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0103] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0104] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0105] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A battery fault classification method, characterized in that, include: Multi-band electrochemical impedance spectroscopy (EIS) statistical features were extracted from the EIS data of lithium-ion batteries, and electrochemical mechanism feature vectors were extracted based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set was constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. An ensemble learning model based on symmetric decision trees is used, with a weighted multi-classification loss function as the optimization objective, to construct a fault category model for the heterogeneous feature set; During the training process of the fault category model, the feature contribution is iteratively evaluated and fed back based on the SHAP value, and the fusion weight of the statistical features and electrochemical mechanism feature vectors of high frequency, mid frequency and low frequency bands is dynamically adjusted so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. The system outputs at least one of the following: fault category, confidence probability, risk level, and interpretability diagnostic criteria. The diagnostic criteria are based on the SHAP contribution value mapped to the impedance response and equivalent circuit parameters in the high-frequency, mid-frequency, and low-frequency bands, revealing the electrochemical mechanism path on which the fault classification decision depends.

2. The method as described in claim 1, characterized in that, The extracted multi-band electrochemical impedance spectroscopy (EIS) statistical features include: Statistical features of the real part of impedance, imaginary part of impedance, impedance magnitude, and phase angle are extracted for each frequency band: The average real part of the impedance is: ; The mean of the imaginary part of the impedance is: ; The impedance amplitude integral characteristic is: ; The phase angle change characteristics are as follows: in, and These represent the maximum and minimum frequencies within frequency band B, respectively. This yields the high-frequency, mid-frequency, and low-frequency characteristics: .

3. The method as described in claim 1, characterized in that, The electrochemical mechanism feature vector extracted based on the equivalent circuit model (ECM) includes: The equivalent circuit model ECM is: ; in, Indicates the internal resistance of the ohm. Indicates the SEI film impedance. This indicates that the SEI film is a non-ideal capacitor element. Indicates charge transfer impedance, This represents a non-ideal double-layer capacitor. express Warburg Diffusion impedance; constant phase element The impedance is expressed as: ;in, for CPE coefficient, n The diffusion index, ω=2πf Angular frequency; Warburg The diffusion impedance can be expressed as: ;in, for Warburg Diffusion coefficient; The equivalent circuit model (ECM) total impedance model is obtained as follows: ; By minimizing the error between the measured impedance and the model impedance, the ECM parameters of the equivalent circuit model are identified: ; in, Θ The set of parameters to be identified: ; Finally, the characteristic vector of the electrochemical mechanism is obtained: 。 4. The method as described in claim 1, characterized in that, Based on the statistical characteristics of the multi-band electrochemical impedance spectroscopy (EIS) and the electrochemical mechanism feature vector, a heterogeneous feature set is constructed, including: The heterogeneous feature set is constructed by using high-frequency features, mid-frequency features, low-frequency features, and electrochemical mechanism feature vectors: ; .

5. The method as described in claim 1, characterized in that, The ensemble learning model of the symmetric decision tree is a CatBoost model, and the CatBoost model function is expressed as follows: in, For the number of decision trees, For learning rate, For the first A symmetric decision tree.

6. The method as described in claim 1, characterized in that, The weighted multi-class loss function is defined as follows: in, As an indicator function, when the sample The true category is the first The value is 1 if the condition is met, and 0 otherwise. For the first Training weights for different types of faults; ; in, k For the first k Number of samples of this type of fault N The total number of samples, K This represents the total number of fault categories. ∈[0,1] is the first k The physical severity coefficient of the fault type. β ∈[0,1] is the balance factor between sample distribution weights and severity weights; When the number of samples for a certain fault category is small Increase the value to make the fault category model pay more attention to the minority class during training; when a certain fault category is considered highly dangerous... Increase the value to improve the sensitivity of the fault category model in identifying the high-risk fault category.

7. The method as described in claim 1, characterized in that, The dynamic adjustment of the fusion weights of statistical characteristics and electrochemical mechanism feature vectors in the high-frequency, mid-frequency, and low-frequency bands includes: After each round of model training, the SHAP contribution values ​​of four feature subsets—high-frequency band, mid-frequency band, low-frequency band, and equivalent circuit parameters—are calculated and denoted as follows: S H , S M , S L , S E ; Update the fusion weights of each feature subset according to the normalization formula: in, g ∈ H , M , L , E Represents a feature subset. r Indicates the first r iteration ε As a smoothing factor, C This is the normalization constant; The updated weights are used as input for the next round of feature fusion, up to the weight change between two consecutive iterations. Δλ(r) Less than the preset weight convergence threshold δ The iteration stops when the maximum number of iterations is reached; among them, 。 8. A battery fault classification device, characterized in that, include: The data preprocessing module is used to extract multi-band electrochemical impedance spectroscopy (EIS) statistical features from lithium-ion battery-based EIS data, and to extract electrochemical mechanism feature vectors based on the equivalent circuit model (ECM). The multi-band features include high-frequency, mid-frequency, and low-frequency bands. A heterogeneous feature set is constructed based on the multi-band EIS statistical features and the electrochemical mechanism feature vectors. The fault classification model module embeds an ensemble learning model based on a symmetric decision tree, which is used to train the heterogeneous feature set with a weighted multi-class loss function as the optimization objective to construct a fault category prediction model. The dynamic weight optimization module iteratively evaluates the contribution of each feature in the heterogeneous feature set based on the SHAP value during the training process of the fault classification model module, and dynamically adjusts the fusion weights of high-frequency, mid-frequency, and low-frequency statistical features and electrochemical mechanism feature vectors, so that the fault category model adaptively focuses on the electrochemical response dimension that is most discriminative for fault discrimination. An interpretable output module is used to output at least one of the following in the fault classification model: fault category, confidence probability, risk level, and interpretable diagnostic criteria. The diagnostic criteria are based on the SHAP contribution value and mapped to the impedance response and equivalent circuit parameters in the high-frequency, mid-frequency, and low-frequency bands, revealing the electrochemical mechanism path on which the fault classification decision depends.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-7.