Battery fault diagnosis method and system based on large model-lightweight model cooperation

Abstract

The present application belongs to the technical field of battery fault diagnosis, and particularly relates to a battery fault diagnosis method and system based on large model-lightweight model cooperation. The method comprises the following steps: obtaining time series data of a battery under different operating conditions; inputting the time series data of the battery under different operating conditions into a Transformer model, dynamically focusing on key time series fragments related to different fault types by using an attention mechanism, and realizing unified feature extraction across operating conditions; determining a lightweight model, performing battery fault diagnosis based on the unified features across operating conditions output by the Transformer model, and obtaining a fault type; and continuously quantifying the fault degree by using a denoising autoencoder, and quantitatively representing the severity of the battery fault. The present application decouples the functions of the large model and the lightweight model, and realizes unified feature representation across operating conditions without repeatedly constructing a feature model.

Description

Technical Field

[0001] This invention belongs to the field of battery fault diagnosis technology, and in particular relates to a battery fault diagnosis method and system based on the collaboration of large model and lightweight model. Background Technology

[0002] Lithium-ion batteries, as the dominant technology in the current energy storage field, have supported the rapid development of key industries such as electric vehicles and energy storage systems due to their high energy density and long cycle life. However, as application scenarios extend to larger scales and greater complexity, the safety of battery systems faces severe challenges: abnormal conditions such as overcharging, over-discharging, thermal runaway, and internal short circuits not only accelerate performance degradation but may also trigger significant safety risks. In the dynamic and ever-changing actual operating environment, the internal state of the battery is difficult to observe directly, and the fault evolution process is both uncertain and complex, making accurate health status assessment and fault early warning extremely difficult. Therefore, how to use precise monitoring technology to assess the health status of the battery in real time and achieve efficient and reliable fault diagnosis has become a key task to ensure the safe and efficient operation of battery systems and extend their service life.

[0003] While existing lithium-ion battery fault diagnosis methods have improved diagnostic accuracy to some extent, they still have significant limitations in complex, multi-condition, and multi-fault-type real-world operating environments. Existing methods often employ a single, fixed-structure model throughout the entire diagnostic process, resulting in insufficient adaptability and generalization ability across various operating conditions. For example, one method uses a CNN-Transformer cascaded structure for feature extraction and fault identification from battery voltage time-series data. Its feature extraction process relies on a fixed network structure unfolding layer by layer, making it difficult to adaptively adjust the feature extraction granularity according to operating conditions, and requiring repeated model training for different diagnostic objectives. The overall computational complexity of the model is high, hindering engineering deployment.

[0004] Furthermore, existing technologies typically rely on fixed sliding windows for battery fault feature extraction to achieve battery fault diagnosis. By segmenting battery time-series data using a pre-defined fixed-length sliding window, the feature extraction timescale is artificially fixed before modeling, making it difficult to dynamically adjust according to the timing and duration of fault features under different operating conditions. When operating conditions change, the fixed window segmentation method can easily lead to the weakening or fragmentation of key fault features, thereby affecting the model's generalization performance under different operating conditions.

[0005] Meanwhile, existing technologies rely heavily on manual feature construction and selection, which is highly subjective and prone to omissions or redundancy of feature information, making it difficult to form a unified and stable feature expression under multiple operating conditions and multiple fault scenarios. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this invention provides a battery fault diagnosis method and system that combines a large model and a lightweight model. The large model and the lightweight model are functionally decoupled. The large model learns common temporal features between different operating conditions and faults, while the lightweight model makes rapid decisions based on the discrimination requirements of different diagnostic tasks, forming a fault diagnosis framework with clear division of labor and collaborative operation. Simultaneously, an adaptive sliding window partitioning method is introduced to dynamically segment the battery operating sequence under different operating conditions, achieving adaptive feature extraction.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of this invention provides a battery fault diagnosis method based on the collaboration of a large model and a lightweight model.

[0009] A battery fault diagnosis method based on large-model-lightweight model collaboration includes the following steps:

[0010] Acquire timing data of the battery under different operating conditions;

[0011] The battery time-series data under different operating conditions are input into the Transformer model, and the attention mechanism is used to dynamically focus on key time-series segments related to different fault types, so as to achieve unified feature extraction across operating conditions.

[0012] A lightweight model is determined, and battery fault diagnosis is performed based on the unified features across operating conditions output by the Transformer model to obtain the fault type.

[0013] The degree of fault is continuously quantified by a noise-reducing autoencoder, which quantitatively characterizes the severity of battery faults.

[0014] The second aspect of this invention provides a battery fault diagnosis system that combines a large model and a lightweight model.

[0015] A battery fault diagnosis system that combines large-scale and lightweight models includes:

[0016] The data acquisition module is configured to acquire time-series data of the battery under different operating conditions.

[0017] The large model feature extraction module is configured to: input the time series data of the battery under different operating conditions into the Transformer model, and use the attention mechanism to dynamically focus on key time series segments related to different fault types to achieve unified feature extraction across operating conditions;

[0018] The lightweight model fault diagnosis module is configured to: determine the lightweight model, perform battery fault diagnosis based on the unified features across operating conditions output by the Transformer model, and obtain the fault type;

[0019] The self-encoder quantitative characterization module is configured to continuously quantify the degree of fault through a noise-reducing self-encoder, thereby quantitatively characterizing the severity of battery faults.

[0020] The above one or more technical solutions have the following beneficial effects:

[0021] This invention provides a battery fault diagnosis method and system that combines a large model and a lightweight model. Addressing the issues of subjective manual feature selection and difficulty in adapting to various operating conditions in traditional lithium-ion battery fault diagnosis, this invention introduces an adaptive sliding window partitioning method. Based on the changing characteristics of battery voltage, current, and other time-series signals, the method dynamically segments the battery operating sequence under different operating conditions. This enables the feature extraction window to adaptively adjust with changes in battery state, automatically extracting multi-scale time-series features reflecting the fault evolution process. This avoids the limitations of manually designed features and improves the objectivity and consistency of fault feature representation.

[0022] To address the challenge of existing methods failing to balance unified feature representation with efficient multi-task decision-making, this invention constructs a diagnostic framework that combines a large model and a lightweight model. This framework is not a simple aggregation of multiple models, but rather designed based on the inventive concept of functional decoupling. The large model is used only for unified representation learning of time-series features across operating conditions and does not directly participate in specific diagnostic decisions; the lightweight model performs specific tasks such as anomaly screening and fault type identification based solely on the unified features, thus avoiding the repeated training of highly complex models for different diagnostic objectives.

[0023] This invention enables the common capability of feature extraction to be reused in different application scenarios by decoupling functions. It can achieve specific tasks such as fault diagnosis, thermal runaway risk assessment or lifetime prediction by combining different lightweight models. Compared with the method of using a large model to complete all tasks, it further improves the overall computational efficiency and engineering applicability.

[0024] Unlike existing fault location or classification methods based on autoencoders, the noise-reducing autoencoder in this invention does not directly operate on the original time-series data. Instead, it models based on a unified feature representation after fault type identification, and its output is used to characterize the severity variations within the same fault type, thereby achieving continuous quantitative assessment of fault severity. Correlation analysis between the model output and experimentally measured battery internal resistance values ​​shows a significant correlation, verifying that this quantitative method can reliably reflect the severity of battery faults.

[0025] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0027] Figure 1 This is a flowchart of the method in Example 1.

[0028] Figure 2 This is a clustering result diagram of traditional statistical features in Example 1.

[0029] Figure 3 This is a clustering result diagram of the features extracted based on the Transformer feature adaptive extraction method in Example 1.

[0030] Figure 4 This is a comparison chart of the fault identification accuracy of the large model and lightweight model under different operating conditions in Example 1.

[0031] Figure 5 The graph shows the test results for the first fault type and different fault degrees in Example 1.

[0032] Figure 6 The graph shows the test results for the second fault type and different fault degrees in Example 1. Detailed Implementation

[0033] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0035] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0036] Example 1

[0037] As mentioned earlier, existing lithium-ion battery fault diagnosis methods have limitations such as subjective feature selection and insufficient model generalization. Furthermore, they have not yet effectively solved the problem of how to achieve unified feature representation across operating conditions without repeatedly constructing feature models under complex multi-operating conditions, while also taking into account the accuracy, real-time performance, and engineering deployment efficiency of fault diagnosis.

[0038] Based on this, this embodiment discloses a battery fault diagnosis method that combines a large model and a lightweight model. By decoupling cross-operating condition feature extraction from specific diagnostic tasks, the large model learns the common temporal features between different operating conditions and different faults, while the lightweight model makes rapid decisions based on the discrimination requirements of different diagnostic tasks, thereby forming a fault diagnosis framework with clear division of labor and collaborative work, achieving high-precision and low-complexity fault identification.

[0039] like Figure 1 As shown, the battery fault diagnosis method based on the collaboration of a large model and a lightweight model includes the following steps:

[0040] Acquire timing data of the battery under different operating conditions;

[0041] The battery time-series data under different operating conditions are input into the Transformer model, and the attention mechanism is used to dynamically focus on key time-series segments related to different fault types, so as to achieve unified feature extraction across operating conditions.

[0042] A lightweight model is determined, and battery fault diagnosis is performed based on the unified features across operating conditions output by the Transformer model to obtain the fault type.

[0043] The degree of fault is continuously quantified by a noise-reducing autoencoder, which quantitatively characterizes the severity of battery faults.

[0044] Overall, this embodiment performs adaptive sliding window feature extraction on the original time-series data such as lithium-ion battery voltage and current under diverse operating conditions and multiple working conditions, so that the feature extraction window can be dynamically adjusted with the changes in battery operating status, thereby automatically obtaining multi-scale time-series features that reflect different fault evolution stages.

[0045] Subsequently, based on the extracted multi-scale time-series features, a preliminary anomaly screening of the battery operating status was conducted to identify potential faults.

[0046] For the abnormal samples identified during screening, further refined analysis is conducted from the perspective of differences in fault characteristics to achieve differentiation and identification of different fault types;

[0047] Finally, based on the identified fault types and the characteristic distribution under normal operating conditions, the severity of the faults is continuously and quantitatively assessed to achieve a quantitative characterization of the severity of lithium-ion battery faults.

[0048] like Figure 1 As shown, the overall technical solution of this embodiment uses four dashed boxes to represent the four processes of data acquisition, adaptive window feature extraction, fault classification, and fault quantification:

[0049] During the data collection process, charging conditions and simulated urban road driving conditions were taken into account.

[0050] During the adaptive window feature extraction process, voltage and / or current signals are collected and input into the Transformer model. The Transformer model outputs a unified feature vector across operating conditions.

[0051] During the fault classification process, the unified feature vector across operating conditions output by the Transformer model is input into the XGBoost model to determine whether the battery is in a normal or abnormal condition; and then TabNet is used to further refine the fault type determination, including poor contact and short circuit faults.

[0052] In the fault quantification process, an autoencoder model, including an encoder and a decoder, is used to process the input normal samples and quantify the severity of the fault.

[0053] The technical solution of this embodiment will be explained in detail below.

[0054] (I) Data Acquisition.

[0055] This embodiment first needs to acquire timing data of the battery under different operating conditions, specifically including:

[0056] Under charging conditions and simulated urban road driving conditions, the original time-series data of the battery is acquired. The original time-series data consists of the battery's current data and / or voltage data.

[0057] The original time-series data is preprocessed to obtain the battery's time-series data under different operating conditions.

[0058] In this embodiment, voltage data is selected as the raw time-series data. After preprocessing, the time-series data of the battery under different operating conditions are obtained, denoted as the voltage sequence. U , U=[u 1 ,u 2 ,…,u n ] ,in, u 1 、u 2 、u n These represent the voltage data at the first sampling time, the voltage data at the second sampling time, and the voltage data at the nth sampling time, respectively.

[0059] (II) Feature extraction of large models.

[0060] This step is based on large model technology and uses an adaptive sliding window method to automatically extract multi-scale spatiotemporal features, overcoming the limitations of traditional manual feature engineering.

[0061] Under different operating conditions, such as charging and simulated urban driving, even if lithium-ion batteries experience the same type of fault, their manifestations in external timing signals such as voltage and current often differ significantly. For example, some faults may manifest as slow voltage shifts during charging, while under dynamic driving conditions, they may exhibit short-term fluctuations or non-stationary changes. This inconsistency in characterization caused by differences in operating conditions makes it difficult for features selected based on human experience to maintain a stable ability to identify the same fault across different operating conditions.

[0062] To achieve unified feature representation across operating conditions, an attention-based adaptive modeling method for time-series features is introduced. This method enables the model to dynamically focus on key time-series segments that are highly correlated with different fault types without the need for manual setting of time windows, thereby achieving unified feature extraction across operating conditions.

[0063] Specifically, during the training phase, the Transformer model does not use fault type labels as direct supervision signals. Instead, it models the overall structure of the time-series data, enabling the model to extract consistent feature representations under different operating conditions. After training, the Transformer outputs only intermediate layer feature vectors as a unified feature representation. These features serve as input to subsequent lightweight models, supporting various downstream tasks such as fault diagnosis, thermal runaway risk assessment, or lifetime prediction, thereby decoupling feature extraction from specific task decisions.

[0064] Unified feature extraction across operating conditions, specifically including:

[0065] The self-attention mechanism in each encoder layer of the Transformer model is computed in its own subspace, capturing multi-level information in the input data.

[0066] By concatenating the outputs of all attention heads, we obtain the multi-head output features;

[0067] A dual-modal pooling strategy is adopted to adaptively aggregate the multi-head output features along the time axis to obtain the final unified features across working conditions.

[0068] Specifically, the voltage sequence mentioned above U=[u 1 ,u 2 ,…,u n ] The input features at each time step will be mapped to the query. Q ,key K Sum V Three vectors. The self-attention mechanism, by characterizing the correlation between arbitrary time steps, assigns differentiated weights to features at different time positions, achieving adaptive selection of fault-sensitive regions. Its calculation form is as follows:

[0069] (1)

[0070] in, Indicates attention score; d represents the normalized exponential function; k is the dimension of the key matrix; T denotes transpose.

[0071] In this embodiment, the Transformer is selected as the large model to achieve adaptive feature extraction. The Transformer model architecture typically consists of multiple stacked encoder layers. The self-attention mechanism in each encoder layer maps the input sequence to a high-dimensional space, where each attention head performs computation in a different subspace, thereby capturing multi-level information in the data. Through parallel processing of multiple heads, the Transformer can automatically learn feature associations at different time scales, thus more comprehensively modeling complex patterns in time-series data. In the multi-head attention mechanism, the outputs of all attention heads are concatenated and a linear transformation is used to obtain the output features.

[0072] (2)

[0073] in, This represents the initial splicing feature obtained by splicing; Contact This indicates that attention is being pieced together; H i Indicates the first i The output of each attention head, H 1 indicates the output of the first attention head. H 2 represents the output of the second attention head. H 8 represents the output of the 8th attention head. i Indicates the first i indivual; It is the output linear transformation weight matrix.

[0074] By stacking multiple encoders, the model extracts deep temporal features that are highly correlated with the fault state layer by layer, forming a high-dimensional unified representation of the battery's operating state.

[0075] After passing through all encoder layers, the Transformer model will generate a multi-head output feature representing the battery characteristics. ,in T Z It is the timing length. D It is the embedded dimension.

[0076] To balance the differences in the temporal manifestations of various faults with the input requirements of subsequent diagnostic models, adaptive aggregation of multi-head output feature sequences is performed along the time axis. This embodiment employs a dual-modal pooling strategy:

[0077] On the one hand, by using max pooling to highlight the extreme value characteristics in the voltage sequence related to sudden fault anomalies, the sensitivity to local anomalies is enhanced:

[0078] (3)

[0079] in, This indicates the maximum value pooling output; This represents the multi-head output characteristics at time t; This indicates taking the maximum value across all time points; Indicates the time.

[0080] On the other hand, by combining average pooling to characterize the long-term evolution trend of faults, a synergistic expression of transient anomalies and overall degradation features is achieved, thus providing stable, compact, and physically meaningful feature inputs for multi-condition battery fault diagnosis:

[0081] (4)

[0082] in, This indicates the average pooling output.

[0083] The two are concatenated to obtain the final feature vector:

[0084] (5)

[0085] in, It represents a unified characteristic across operating conditions; This indicates element-wise addition.

[0086] (III) Fault classification of lightweight model.

[0087] Due to the large size and computational cost of the Transformer architecture, it is difficult to meet real-time requirements. To efficiently filter a large number of normal samples and reduce the computational burden later, XGBoost was chosen as the anomaly filtering module. Using the feature vectors extracted from the large model as input, XGBoost minimizes the error on the training set during training by optimizing the objective function. Specifically, the objective function can be expressed as:

[0088] (6)

[0089] in, This represents the loss function value of XGBoost. These are the model's predicted values; This represents the logistic regression loss. This represents the i-th true value of the battery fault characteristic. Let represent the i-th predicted value of the battery fault feature, and N represent the number of battery fault features; For regularization terms, This represents the base learner function (i.e., the decision tree model) in XGBoost.

[0090] By optimizing the loss function, an accurate battery anomaly identification classifier can be obtained.

[0091] After XGBoost anomaly filtering, these anomalous samples still need to undergo more precise classification to determine the specific fault type. Therefore, TabNet was chosen as the subsequent fault type identification model.

[0092] Assuming in TabNet's... p In the decision-making process, the Attentive Transformer generates weights. a p The weighted processing of different features is expressed as follows:

[0093] (7)

[0094] in, It is a feature selection matrix. This is the bias term, where De represents the dimension of the feature vector corresponding to the faulty battery. Indicates the first p The weights are generated step by step. The feature selection process for each decision step dynamically adjusts based on the importance of the current input through this mechanism.

[0095] During the decision-making phase, TabNet generates the final decision output using the key features selected in the previous phase. Assume the final decision vector is... , K This represents the number of battery fault types that need to be identified. The probability distribution for each fault type is as follows:

[0096] (8)

[0097] in, Dec k Indicates the first k The probability distribution of each fault. k Indicates the first k One fault, This indicates that the fault type of the current abnormal sample belongs to the first... k The probability of a fault.

[0098] Based on this probability distribution, a specific fault type label is assigned to each fault sample.

[0099] (iv) Quantification of the fault degree of the self-encoder.

[0100] This embodiment uses a noise-reducing autoencoder to continuously quantify the degree of fault. Its output is significantly related to the battery's internal resistance, thus achieving a deeper understanding of fault detection from qualitative to quantitative assessment.

[0101] In lithium-ion battery safety monitoring, simply identifying the fault type has significant limitations. Therefore, this embodiment employs a noise-reducing autoencoder to construct a fault severity assessment model. Unlike existing techniques that directly apply autoencoders to raw time-series data for fault location or classification, this embodiment introduces a noise-reducing autoencoder after fault type identification to model samples in a unified feature space. This allows for a fine-grained characterization of fault severity within the same fault type, supporting continuous quantitative assessment of fault severity.

[0102] When faulty input data is presented, the model exhibits significant reconstruction errors due to deviations from the normal distribution during the reconstruction process. Firstly, by injecting Gaussian noise into the input layer and forcing the model to recover the original data, robustness to local perturbations is significantly enhanced, equivalent to learning a smoother data representation.

[0103] (9)

[0104] in, This represents the set of parameters for the denoising autoencoder model, including the network weights and bias parameters in both the encoder and decoder. E represents the expectation. Indicates input sample x Follows the true data distribution p data , This indicates that the injected noise term follows a Gaussian distribution. x represents the input sample. This represents the encoder function used to map noisy inputs to a low-dimensional latent representation space. This represents the decoder function, which is used to reconstruct the input features based on the latent representation. Indicates injected noise; This indicates a regular term.

[0105] The model uses raw time-series voltage data and fault labels as input based on a multi-source feature fusion strategy. The autoencoder model essentially consists of two parts: an encoder and a decoder. First, the encoder maps the input data to a low-dimensional latent space:

[0106] (10)

[0107] in, and These represent the weights and biases of the first layer of the encoder, respectively. and represents the weights and biases of the second layer of the encoder, respectively. tanh represents the hyperbolic tangent activation function, used to introduce non-linear mapping capabilities, mapping features to a finite interval. This represents the low-dimensional latent feature representation output by the encoder, which is used as input to the decoder.

[0108] The decoder then attempts to reconstruct the input data by reconstructing the original input through inverse mapping:

[0109] (11)

[0110] in, This represents the reconstructed result output by the decoder. Wherein, and These represent the weights and biases of the first layer of the decoder, respectively. and These represent the weights and biases of the second layer of the decoder, respectively.

[0111] Since the training objective of this model is to learn the latent representation of the data by minimizing the reconstruction error, the training process is optimized by minimizing the mean square error of the noisy input and the original input:

[0112] (12)

[0113] in, This represents the training objective function (loss function) value of the denoising autoencoder (DAE). Num This indicates the number of normal battery samples used in the training. The square of the L2 norm represents the reconstruction error between the input features and their reconstruction results, and is used to measure the reconstruction accuracy of the model. Represents the j-th input feature. This represents the reconstruction result of the j-th input feature; j represents the j-th training sample. This represents the regularization coefficient, used to balance the weights of the reconstruction error term and the regularization term in the objective function. Let be the weight matrix for each layer, and l represent the l-th layer of the network.

[0114] When calculating the severity of different battery faults, the reconstruction error of the sample is calculated first. :

[0115] (13)

[0116] Where D is the total number of samples; The fault value is the mapping of the true value of the fault resistance during the experiment. To predict fault values.

[0117] To map the reconstruction error to a uniform, standardized range, a normalization method is used to convert the reconstruction error values ​​into fault severity scores:

[0118] (14)

[0119] in, This represents the minimum reconstruction error. This indicates the maximum reconstruction error. Let t be the reconstruction error value. Let be the fault severity score of the t-th battery sample.

[0120] (V) Specific experimental applications.

[0121] The experiment used ternary lithium batteries and pouch batteries, with a nominal capacity of 2000 mAh and a nominal voltage of 3.6 V. The charging cutoff voltage was set to 4.2 V, and the discharging cutoff voltage was set to 2.0 V. The Arbin battery testing system was responsible for simulating the battery's operating state and acquiring key signals. This system simulated the UDDS (Underlying Device Discharge) conditions of urban road driving and specific charging cycle conditions according to preset experimental conditions. Real-time acquisition of key parameters such as current and voltage was then performed. All data was uploaded to a central server for storage and subsequent analysis and processing.

[0122] The Arbin battery testing system is a battery performance testing platform that can perform charge and discharge experiments on batteries according to preset operating conditions and monitor and record key parameters such as voltage and current generated during the experiment in real time.

[0123] The UDDS (Urban Dynamometer Driving Schedule) is a driving schedule that simulates the driving characteristics of a vehicle in an urban road environment, such as frequent starts and stops, low-speed driving, and short-term acceleration and deceleration, through a series of time-varying speed curves.

[0124] For the UDDS (Underlying Variable Discharge) test, to simulate the common structural characteristics of parallel-to-series battery packs in electric vehicles, the test setup considers the equivalent single-cell modeling characteristics of the parallel units. Therefore, the current values ​​in this embodiment represent the series branch currents, and the voltage values ​​represent the individual battery cell voltages. The charging cycle test uses six individual batteries directly connected in series. The experiments were conducted in temperature chambers at 0°C, -5°C, 25°C, and 40°C to simulate different operating temperature environments for the batteries.

[0125] Mean, maximum, minimum, standard deviation, kurtosis, and skewness are selected as traditional statistical features. Both traditional statistical features and features extracted using the Transformer-based adaptive feature extraction method are dimensionality-reduced to three dimensions, then simple clustering is performed, and the comparison is visualized. Figure 2 and Figure 3 As shown. Among them, Figure 2 Clustering results based on traditional statistical features Figure 3 This is the clustering result of the features extracted based on the Transformer feature adaptive extraction method in this embodiment.

[0126] Figure 2 and Figure 3 In the diagram, the X-axis coordinate PCA1, Y-axis coordinate PCA2, and Z-axis coordinate PCA3 represent the first three principal components extracted by principal component analysis, used to characterize the relative distribution of high-dimensional features in low-dimensional space. Since the principal components are linear combinations of the original features, their coordinate values ​​only reflect the relative position of the sample in the feature space and do not possess direct physical meaning. "Normal" indicates a normal battery label, represented by a dot in the diagram; "PCC" indicates a poor contact fault, represented by a triangle; and "ISC" indicates a short circuit fault, represented by a square.

[0127] Figure 2 While the traditional feature extraction methods presented achieve type differentiation to some extent, they exhibit significant class overlap, indicating limited discriminative ability. Figure 3 For samples of different fault types, the features extracted by the feature extraction method based on the Transformer architecture exhibit a clear distribution pattern in the feature space, with significant separation between categories. This result demonstrates that the Transformer model, with its powerful sequence modeling capabilities and feature adaptation mechanism, can more effectively represent and separate multi-category fault features.

[0128] To improve the model's generalization ability and robustness under different operating conditions, this embodiment comprehensively considers the vehicle's operational characteristics during both the UDDS and charging phases in the training data. Specifically, 20% of the charging data and 80% of the UDDS data were selected as the training set to ensure that the model can learn both the vehicle's operational patterns under dynamic conditions and the key state change characteristics during the charging process. The remaining data was used as the test set to evaluate the model's performance. Experimental results are as follows: Figure 4 As shown, the vertical axis represents accuracy, and the horizontal axis represents the two stages of UDDS and charging.

[0129] The results show that the model achieves an overall accuracy of 98.2% under charging conditions, with an output variance of only 0.00005. It exhibits high stability during the charging phase, with minimal fluctuations in classification results.

[0130] In the more dynamic UDDS scenario, the accuracy reached 96.5% with a variance of 0.00022. Although the variance was slightly higher than that in the charging state, it remained at a low level, demonstrating the model's generalization ability under different conditions.

[0131] In addition, different fault types and severity levels were tested in the experiment, and operating environments at different temperatures were simulated to evaluate the quantitative results of fault severity. Battery data was processed using a noise-reducing autoencoder to calculate the reconstruction error of each fault sample, which was then converted into a unified fault severity score. The specific reconstruction results are shown below. Figure 5 and Figure 6 As shown.

[0132] exist Figure 5 and Figure 6 In the diagram, the horizontal axis represents resistance, and the vertical axis represents different fault conditions, including ISC and PCC. ISC indicates a short circuit fault, and PCC indicates a poor contact fault. Different shades of gray are used to represent the degree of battery failure under different fault conditions, with darker colors indicating a higher degree of failure.

[0133] exist Figure 5 middle:

[0134] When the resistance is 0.005Ω (ohms), the failure severity of the PCC is 23.9%.

[0135] When the resistance is 0.05Ω (ohms), the failure severity of the PCC is 47.2%.

[0136] When the resistance is 0.1Ω (ohms), the failure severity of the PCC is 70.8%.

[0137] When the resistance is 0.5Ω (ohms), the failure severity of the PCC is 91.4%.

[0138] When the resistance is 20Ω (ohms), the failure severity of the ISC is 96.8%.

[0139] When the resistance is 39Ω (ohms), the failure severity of the ISC is 81.5%;

[0140] When the resistance is 57Ω, the failure severity of the ISC is 66.3%.

[0141] When the resistance is 80Ω (ohms), the failure severity of the ISC is 52.6%.

[0142] exist Figure 6 middle:

[0143] When the resistance is 0.03Ω (ohms), the failure severity of the PCC is 9.5%.

[0144] When the resistance is 0.33Ω (ohms), the failure severity of the PCC is 22.8%.

[0145] When the resistance is 0.5Ω (ohms), the failure severity of the PCC is 33.7%.

[0146] When the resistance is 1.8Ω (ohms), the failure severity of the PCC is 76.2%.

[0147] When the resistance is 0.33Ω (ohms), the ISC's fault severity is 97.0%;

[0148] When the resistance is 0.5Ω (ohms), the failure severity of the ISC is 89.5%.

[0149] When the resistance is 1.8Ω (ohms), the fault severity of the ISC is 73.0%;

[0150] When the resistance is 2Ω (ohms), the fault severity of the ISC is 70.0%.

[0151] Based on the above observations, to further verify the accuracy and reliability of the proposed battery fault quantification model, this embodiment experimentally verified the correlation between resistance value and fault severity score. According to the calculated correlation coefficient, the Spearman correlation coefficient for ISC faults in the UDDS state is -0.988, indicating a significant negative correlation between resistance value and ISC fault severity; that is, as the resistance value increases, the severity of ISC faults decreases significantly. This trend was also verified in the charging state, with a Pearson correlation coefficient of -0.997 for ISC faults, further confirming that higher resistance values ​​correspond to lower severity scores.

[0152] Conversely, PCC faults showed a strong positive correlation in the UDDS state, with a Spearman correlation coefficient of 0.995. In the charging state, although the correlation was slightly weaker, with a Spearman correlation coefficient of 0.856, it still indicated a positive correlation between the increase in resistance value and the increase in the degree of PCC faults.

[0153] The above results further validate the positive correlation between PCC faults and resistance, and the negative correlation between ISC faults and resistance. Furthermore, the correlation coefficient being close to 1 further demonstrates the high accuracy and reliability of the fault severity score calculated based on the noise-reducing autoencoder.

[0154] In summary, the novel large-scale model-lightweight model collaborative framework proposed in this embodiment enables fault classification and quantification of lithium-ion batteries under different fault conditions. This method effectively improves the generalization ability of lithium-ion battery fault diagnosis, and enhances the reliability and universality of fault diagnosis.

[0155] Example 2

[0156] This embodiment discloses a battery fault diagnosis system that combines a large model and a lightweight model.

[0157] A battery fault diagnosis system that combines large-scale and lightweight models includes:

[0158] The data acquisition module is configured to acquire time-series data of the battery under different operating conditions.

[0159] The large model feature extraction module is configured to: input the time series data of the battery under different operating conditions into the Transformer model, and use the attention mechanism to dynamically focus on key time series segments related to different fault types to achieve unified feature extraction across operating conditions;

[0160] The lightweight model fault diagnosis module is configured to: determine the lightweight model, perform battery fault diagnosis based on the unified features across operating conditions output by the Transformer model, and obtain the fault type;

[0161] The self-encoder quantitative characterization module is configured to continuously quantify the degree of fault through a noise-reducing self-encoder, thereby quantitatively characterizing the severity of battery faults.

[0162] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0163] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A battery fault diagnosis method based on a large-model-lightweight model synergy, characterized in that, Includes the following steps: Acquire timing data of the battery under different operating conditions; The battery time-series data under different operating conditions are input into the Transformer model, and the attention mechanism is used to dynamically focus on key time-series segments related to different fault types, so as to achieve unified feature extraction across operating conditions. A lightweight model is determined, and battery fault diagnosis is performed based on the unified features across operating conditions output by the Transformer model to obtain the fault type. The unified feature extraction across working conditions specifically includes: the self-attention mechanism in each encoder layer of the Transformer model is calculated in its own subspace to capture multi-level information in the input data; The outputs of all attention heads are concatenated to obtain the initial concatenated features; after passing through all encoder layers, a multi-head output feature is generated; the self-attention mechanism in each encoder layer of the Transformer model is computed in its own subspace, capturing multi-level information in the input data, specifically including: The input features at each time step will be mapped to the query. Q ,key K Sum V Three vectors; The self-attention mechanism, by characterizing the correlation between arbitrary time steps, assigns differentiated weights to features at different time positions, thereby achieving adaptive selection of fault-sensitive regions. A dual-modal pooling strategy is adopted to adaptively aggregate the multi-head output features along the time axis to obtain the final unified features across working conditions. On the one hand, by using maximum pooling, the extreme value characteristics of the voltage sequence related to sudden fault anomalies are highlighted, thereby enhancing the sensitivity to local anomalies; On the other hand, by combining average pooling to characterize the long-term evolution trend of faults, a synergistic expression of transient anomalies and overall degradation features is achieved, providing stable, compact and physically meaningful feature inputs for multi-condition battery fault diagnosis. By continuously quantifying the degree of fault through a noise-reducing autoencoder, the severity of battery faults can be quantitatively characterized. After the fault type identification is completed, a noise-reducing autoencoder is introduced to model the samples in a unified feature space, thereby realizing the characterization of the severity of faults within the same fault type and supporting continuous quantitative assessment of the degree of fault.

2. The battery fault diagnosis method based on large-model-lightweight model collaboration as described in claim 1, characterized in that, Acquire timing data of the battery under different operating conditions, specifically including: In charging conditions and simulated urban road driving conditions, the original time-series data of the battery is acquired, wherein the original time-series data is the battery's current data and / or voltage data; The original time-series data is preprocessed to obtain the battery's time-series data under different operating conditions.

3. The battery fault diagnosis method based on large-model-lightweight model collaboration as described in claim 1, characterized in that, A dual-modal pooling strategy is employed to adaptively aggregate multi-head output features along the time axis, resulting in a final unified feature across operating conditions, specifically including: On the one hand, by using max pooling, extreme value features related to sudden fault anomalies in the multi-head output characteristics are highlighted, thereby enhancing the sensitivity to local anomalies: ； in, This indicates the maximum value pooling output; This represents the multi-head output characteristics at time t; This indicates taking the maximum value across all time points; Indicates time; T Z It is the timing length; On the other hand, by combining average pooling to characterize the long-term evolution trend of faults, a synergistic expression of transient anomalies and overall degradation features is achieved: ； in, This indicates the average pooling output; By combining the two, we obtain the final unified characteristics across operating conditions: ； in, It represents a unified characteristic across operating conditions; This indicates element-wise addition.

4. The battery fault diagnosis method based on large model-lightweight model collaboration as described in claim 3, characterized in that, A lightweight model is determined, and battery fault diagnosis is performed based on the unified features across operating conditions output by the Transformer model to obtain fault types, specifically including: First, the unified features across working conditions output by the Transformer model are input into the pre-trained XGBoost model to perform preliminary anomaly screening and obtain abnormal samples. Subsequently, the unified features corresponding to the abnormal samples across different operating conditions are input into the TabNet model for refined analysis from the perspective of fault feature differences, thereby enabling the differentiation and identification of different fault types.

5. The battery fault diagnosis method based on large model-lightweight model collaboration as described in claim 4, characterized in that, During training, the loss function of XGBoost is: ； in, This represents the loss function value of XGBoost. These are the model's predicted values; This represents the logistic regression loss. This represents the i-th true value of the battery fault characteristic. Let represent the i-th predicted value of the battery fault feature, and N represent the number of battery fault features; For regularization terms, This represents the base learner function in XGBoost.

6. The battery fault diagnosis method based on large model-lightweight model collaboration as described in claim 4, characterized in that, In the TabNet model: The feature selection for each decision step is dynamically adjusted based on the importance of the current input; During the decision-making phase, TabNet generates the final decision output using the key features selected in the previous phase, obtaining the probability distribution for each fault type. Based on the probability distribution of each fault type, a specific fault type label is assigned to each abnormal sample.

7. The battery fault diagnosis method based on large model-lightweight model collaboration as described in claim 6, characterized in that, The degree of fault is continuously quantified through a noise-reducing autoencoder, which quantitatively characterizes the severity of battery faults. Specifically, this includes: The timing data and fault type labels of the batteries in the abnormal samples under different operating conditions are used as inputs to the noise reduction autoencoder; The encoder maps the input data to a low-dimensional latent space, and then the decoder reconstructs the input data, reconstructing the original input through inverse mapping. Calculate the reconstruction error of outlier samples; A normalization method is used to convert the reconstruction error value into a fault severity score, and the severity of battery faults is quantitatively characterized based on the fault severity score.

8. A battery fault diagnosis system based on a large-model-lightweight model collaboration, characterized in that: include: The data acquisition module is configured to acquire time-series data of the battery under different operating conditions. The large model feature extraction module is configured to: input the time series data of the battery under different operating conditions into the Transformer model, and use the attention mechanism to dynamically focus on key time series segments related to different fault types to achieve unified feature extraction across operating conditions; The lightweight model fault diagnosis module is configured to: determine the lightweight model, perform battery fault diagnosis based on the unified features across operating conditions output by the Transformer model, and obtain the fault type; The unified feature extraction across working conditions specifically includes: the self-attention mechanism in each encoder layer of the Transformer model is calculated in its own subspace to capture multi-level information in the input data; The outputs of all attention heads are concatenated to obtain the initial concatenated features; after passing through all encoder layers, a multi-head output feature is generated; the self-attention mechanism in each encoder layer of the Transformer model is computed in its own subspace, capturing multi-level information in the input data, specifically including: The input features at each time step will be mapped to the query. Q ,key K Sum V Three vectors; The self-attention mechanism, by characterizing the correlation between arbitrary time steps, assigns differentiated weights to features at different time positions, thereby achieving adaptive selection of fault-sensitive regions. A dual-modal pooling strategy is adopted to adaptively aggregate the multi-head output features along the time axis to obtain the final unified features across working conditions. On the one hand, by using maximum pooling, the extreme value characteristics of the voltage sequence related to sudden fault anomalies are highlighted, thereby enhancing the sensitivity to local anomalies; On the other hand, by combining average pooling to characterize the long-term evolution trend of faults, a synergistic expression of transient anomalies and overall degradation features is achieved, providing stable, compact and physically meaningful feature inputs for multi-condition battery fault diagnosis. The autoencoder quantitative characterization module is configured to: continuously quantify the degree of fault through a noise-reducing autoencoder, quantitatively characterize the severity of battery faults, introduce a noise-reducing autoencoder after completing fault type identification, model samples in a unified feature space, achieve fine characterization of fault severity within the same fault type, and support continuous quantitative evaluation of fault degree.

Images