Cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional network

Abstract

The application discloses a cross-condition lithium ion battery fault detection method based on a domain adversarial graph convolutional network, and the method comprises the following steps: splicing all voltage and temperature data of a lithium ion battery group at the same time to obtain a fusion feature vector; splicing the fusion feature vectors at continuous multiple times into a sample vector through a sliding window with a set length and labeling a category label to construct a fully-labeled source domain data set and a target domain data set containing a small amount of labeled samples; constructing a domain adversarial graph convolutional network, including a feature extractor, a label classifier and a domain discriminator, wherein the feature extractor comprises a convolutional neural network, a graph generation layer and a multi-receptive field graph convolutional network; constructing a target function and adopting a two-stage optimization strategy to train the model; inputting a sample vector to be detected into the trained domain adversarial graph convolutional network, extracting structural features through the feature extractor and outputting a detection result through the label classifier. The application realizes cross-condition lithium ion battery internal short circuit fault detection.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery fault detection and safety monitoring technology, specifically involving a cross-operating-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have been widely used in electric vehicles, large-scale energy storage systems, and portable electronic devices. However, their complex electrochemical and thermal characteristics make them susceptible to internal short circuits (ISC) and thermal runaway, leading to serious safety risks. Therefore, accurate and timely detection of lithium-ion battery systems, especially battery packs operating under varying conditions, is crucial for ensuring system safety and operational reliability.

[0003] Existing lithium-ion battery fault detection methods can be broadly categorized into model-based methods and data-driven methods. Model-based methods typically employ mathematical frameworks such as equivalent circuit models or electrochemical models to characterize the battery's dynamic characteristics, and combine them with extended Kalman filtering (EKF) or observer-based methods for state estimation and fault diagnosis. However, model-based methods rely on accurate physical models, and due to the complex characteristics and variable operating conditions of lithium-ion batteries, such models are difficult to obtain and maintain, limiting their performance in practical online applications.

[0004] Data-driven methods are widely used in fault detection because they can learn features directly from data without requiring explicit physical models. Multivariate statistical techniques such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Partial Least Squares (PLS) have been extensively applied in industrial process anomaly detection and have been introduced into lithium-ion battery fault diagnosis. Traditional machine learning methods such as Support Vector Machines (SVM) are also widely used for fault and anomaly detection in lithium-ion battery systems. In recent years, deep learning methods, especially Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, have been increasingly used for state monitoring and fault diagnosis of lithium-ion batteries.

[0005] Although the above methods have achieved some results, they still have many shortcomings: traditional multivariate statistical techniques (such as PCA, ICA, PLS) are essentially linear methods and have limited ability to characterize the nonlinear dynamics of battery systems; traditional machine learning methods have poor generalization and robustness under different operating conditions; deep learning methods usually require a large amount of data and are prone to performance degradation in cross-domain scenarios with distribution shifts.

[0006] Furthermore, existing methods suffer from two key drawbacks: First, most methods rely on only a single signal (such as voltage or temperature alone), failing to effectively capture complementary fault information; second, they fail to utilize structural relationships between samples for modeling, making the detection results insensitive to early, subtle faults. Simultaneously, distributional differences between different operating conditions significantly reduce cross-domain detection performance, resulting in models trained under one operating condition being unable to effectively transfer to other operating conditions.

[0007] Therefore, there is an urgent need for a lithium-ion battery fault detection method that can integrate multimodal signals, capture structural relationships between samples, and achieve reliable cross-domain detection under different operating conditions. Summary of the Invention

[0008] To address the shortcomings of existing lithium-ion battery fault detection methods, such as insufficient information from a single signal source, significant differences in distribution across operating conditions, inadequate modeling of fault sample structural correlations, scarcity of labeled samples in the target domain, and lack of optimal selection mechanisms for key hyperparameters, this invention provides a cross-operating-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks. Through multimodal data fusion, DAGCN domain adversarial feature learning, two-stage adversarial training, and Bayesian optimization of key hyperparameters, this method achieves high accuracy, low latency, and robust cross-operating-condition internal short-circuit fault detection.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a cross-condition fault detection method for lithium-ion batteries based on domain adversarial graph convolutional networks, the method comprising: The fusion feature vector is obtained by concatenating all voltage and temperature data of the lithium-ion battery pack at the same time. The fusion feature vectors of multiple consecutive time points are concatenated into a sample vector and labeled with corresponding category labels through a sliding window of a set length. This constructs a fully labeled source domain dataset and a target domain dataset containing a small number of labeled samples. A domain adversarial graph convolutional network is constructed, comprising a feature extractor, a label classifier, and a domain discriminator. The feature extractor includes a convolutional neural network, a graph generation layer, and a multi-receptive-field graph convolutional network. The convolutional neural network extracts temporal features of samples, the graph generation layer transforms the temporal features into graph structure data, and the multi-receptive-field graph convolutional network extracts structural features based on the graph structure data. The label classifier is used to predict the fault category of the structural features, and the domain discriminator is used to determine the domain to which the structural features belong. We construct the objective function of the domain adversarial graph convolutional network and use a two-stage optimization strategy for model training. The sample vector to be detected is input into a trained domain adversarial graph convolutional network, structural features are extracted by a feature extractor, and the detection results are output by a label classifier.

[0010] Furthermore, the graph generation layer transforms temporal features into graph-structured data, including: The temporal features are nonlinearly mapped using a multilayer perceptron to obtain the transformed feature matrix. The initial adjacency matrix is ​​calculated based on the transformed feature matrix. The initial adjacency matrix is ​​then subjected to row-wise maximum normalization, and the Top-k strategy is used to retain the k largest values ​​in each row and their corresponding indices to obtain the sparse adjacency matrix. Finally, the transformed feature matrix is ​​used as the node features and the sparse adjacency matrix is ​​used as the edge weights to construct the graph structure data.

[0011] Furthermore, multi-receptive-field graph convolutional networks extract structural features based on graph-structured data, including: Multiple Chebyshev polynomial graph filters of different orders are set up to correspond to different receptive field scales. Based on the normalized Laplacian matrix, graph convolution operation is performed on the graph structure data at each receptive field scale and the feature matrix of the corresponding scale is output. The feature matrices output from all receptive field scales are concatenated to obtain the fused multi-scale structural features. The hierarchical structural features are then extracted layer by layer by stacking multiple multi-receptive field graph convolution layers.

[0012] Furthermore, the objective function of the domain adversarial graph convolutional network includes classification loss, domain alignment loss, and structural alignment loss. The classification loss is the cross-entropy loss of labeled samples, used to measure the difference between the predicted labels and the true labels of the label classifier. The domain alignment loss is the binary cross-entropy loss, used to measure the classification performance of the domain discriminator on the structural features of the source and target domains. The structural alignment loss is the maximum mean difference loss, used to map the structural features of the source and target domains to the regenerating kernel Hilbert space and measure the distribution difference by calculating the distance between the mapped features, thereby aligning the source and target domains at the graph structure level.

[0013] Furthermore, a two-stage optimization strategy is employed for model training to achieve domain-invariant feature learning across different operating conditions, including: Initial training phase: Using labeled samples from the source domain dataset and the target domain dataset, update the parameters of the feature extractor and label classifier based solely on the classification loss; Adversarial training phase: The domain discriminator updates its parameters with the goal of minimizing the domain adversarial loss; the feature extractor receives the inverted gradient of the domain adversarial loss through the gradient inversion layer, and updates its parameters with the goal of maximizing the classification error of the domain discriminator and combining the gradients of the classification loss and the structure alignment loss; the label classifier updates its parameters based only on the gradient of the classification loss.

[0014] Furthermore, the method also includes a key hyperparameter optimization step: Select a representative subset from all migration tasks; Using the sliding window size and the proportion of labeled data in the target domain as parameters to be optimized, a fitness function that balances the accuracy and timeliness of fault detection is constructed. The fitness function is used as the black-box objective function, and a Bayesian optimization method is employed to find the optimal value on the representative subset. Specifically, a surrogate model of the black-box objective function is constructed using a Gaussian process, and the next evaluation point is selected based on the expected improvement of the acquisition function. The process is iteratively updated until the optimal parameter combination that maximizes the fitness function value is output.

[0015] Furthermore, methods for selecting representative subsets include: Calculate the maximum mean difference between the source and target domains, and the Fisher discrimination rate between normal and faulty samples within the target domain, to construct a joint difficulty score, calculated using the following formula: ; in, The maximum mean difference between the normalized source and target domains. The Fisher discrimination rate between normal and faulty samples within the target domain after normalization. Based on the joint difficulty score, all migration tasks are divided into different difficulty levels, and a representative subset is selected from each difficulty level.

[0016] Furthermore, the fitness function is: ; Where FDR is the fault detection rate and FDD is the fault detection delay. For the maximum allowable fault detection delay, The weighting coefficients are used to balance the accuracy and timeliness of fault detection.

[0017] In a second aspect, the present invention provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0018] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention effectively enhances the ability to perceive the structured features of early, subtle faults in lithium-ion batteries by fusing multimodal data of voltage and temperature and combining it with a multi-receptive-field graph convolutional network. It significantly suppresses cross-domain distribution shifts caused by changes in operating conditions by utilizing a domain adversarial graph convolutional network and a maximum mean difference structure alignment loss. Simultaneously, a two-stage optimization strategy is employed to achieve reliable domain-invariant feature learning under a semi-supervised setting where labeled samples in the target domain are scarce. Furthermore, a Bayesian optimization method based on a representative task subset enables efficient automatic configuration of key hyperparameters such as the sliding window size and the proportion of labels in the target domain. In summary, this invention exhibits high detection accuracy, short response latency, and strong cross-domain generalization robustness under complex and variable cross-operating conditions. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a multimodal data fusion and sliding window preprocessing process according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the semi-supervised domain adaptation (SSDA) setting according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a multi-receptive-field graph convolutional network (MRF-GCN) according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the overall domain adversarial framework based on Domain Adversarial Graph Convolutional Network (DAGCN) according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a representative task subset selection process according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the internal structure (4P6S arrangement) of a battery pack according to an embodiment of the present invention; Figure 7 This is a visualization diagram of the accuracy of 30 migration tasks according to an embodiment of the present invention; Figure 8 This is a visualization diagram of the fault detection rate (FDR) of 30 migration tasks according to an embodiment of the present invention; Figure 9 This is a rain cloud map showing the distribution of fault detection delay (FDD) according to an embodiment of the present invention; Figure 10 This is a heatmap of fault detection delay (FDD) for each migration task according to an embodiment of the present invention; Figure 11 This is a schematic diagram illustrating the maximum mean difference (MMD) values ​​between different domains according to an embodiment of the present invention. Figure 12 This is a schematic diagram of the Fisher discrimination rate of normal and faulty samples in each domain according to an embodiment of the present invention; Figure 13This is a t-SNE visualization diagram of the distribution of several target domain features according to an embodiment of the present invention; wherein, Figure 13 (a) in the diagram is the feature distribution map of the target domain in domain 0. Figure 13 (b) in the diagram is the feature distribution map of the target domain in Domain 1. Figure 13 (c) in the figure is the feature distribution map of the target domain in domain 3; Figure 14 This is a schematic diagram of the confusion matrix for three migration tasks according to an embodiment of the present invention; wherein, Figure 14 (a) in the diagram is the confusion matrix for task 1→0. Figure 14 (b) in the diagram is the confusion matrix for tasks 3→5. Figure 14 (c) in the diagram is the confusion matrix for task 5→3; Figure 15 is a schematic diagram illustrating the detailed results of four ablation study configurations across five tasks according to an embodiment of the present invention; wherein... Figure 15A This is a graph comparing accuracy. Figure 15B This is a comparison chart of Fault Detection Rate (FDR). Figure 15C A comparison chart of fault detection delay (FDD); Figure 16 is a schematic diagram showing a detailed comparison of the methods of an embodiment of the present invention on each task; wherein... Figure 16A This is a graph comparing accuracy. Figure 16B This is a comparison chart of Fault Detection Rate (FDR). Figure 16C This is a comparison chart of fault detection delay (FDD). Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0022] Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this invention, modifications to design, manufacturing, or production based on the technical content disclosed in this invention are merely conventional technical means and should not be construed as insufficient disclosure of the present invention.

[0023] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention may be combined with other embodiments without conflict.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "a," "an," "an," "the," and similar words used in this invention do not indicate quantity limitation and may indicate singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms "connected," "linked," "coupled," and similar words used in this invention are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "A plurality" used in this invention refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships may exist; for example, "A and / or B" can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects have an "or" relationship. The terms "first," "second," and "third" used in this invention are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0025] This invention provides a cross-operating-condition fault detection method for lithium-ion batteries based on Domain Adversarial Graph Convolutional Networks (DAGCNs). Specifically, it is a cross-domain detection method for internal short-circuit (ISC) faults in lithium-ion battery packs under variable operating conditions. In this invention, the lithium-ion battery pack is composed of multiple battery cells connected in series and parallel, and is configured with... A voltage sensor and A temperature sensor samples multimodal time-series signals with distribution differences between the source and target domains, and only a small number of labeled samples are in the target domain. This method, under a semi-supervised domain adaptation (SSDA) framework, utilizes fully labeled source domain data and a small number of labeled target domain samples to learn domain-invariant fault characteristics, enabling cross-condition internal short-circuit fault detection.

[0026] This method aims to address the following technical problems existing in current lithium-ion battery fault detection methods: (1) A single signal source (such as voltage or temperature only) has insufficient information and cannot capture complementary fault characteristics, resulting in limited detection capability for early and weak faults; (2) There are significant distribution differences between different operating conditions (such as different discharge rates and different short-circuit resistances), making it difficult for traditional methods to generalize and severely degrading the cross-domain detection performance; (3) When labeled samples in the target domain are scarce, pure supervision methods cannot achieve reliable cross-domain adaptive detection using limited labeling information.

[0027] like Figure 4 As shown, the cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks of the present invention mainly includes: S1. Construction of a domain adversarial framework based on DAGCN; S2. Objective function design and model optimization; S3. Battery pack fault detection and key hyperparameter optimization.

[0028] In the above scheme, step S1 includes: S11. Multimodal data fusion; S12. Feature extractor; S13. Task Module.

[0029] Step S11 includes: To eliminate the dimensional differences between voltage (V) and temperature (°C) data in lithium-ion battery monitoring, Z-score normalization was performed independently for each mode. Let the original voltage data be denoted as... , indicating at time The voltage data points ( ; Record the original temperature data as , indicating at time The z-th temperature data point ( ),in and These represent the number of voltage and temperature sensors, respectively. and Perform Z-score standardization independently: ; ; in, , and , These represent the mean and standard deviation of the voltage and temperature data, respectively. After standardization, the resulting data points... and The mean is 0 and the variance is 1.

[0030] Time of Each voltage point and The temperature points are sequentially concatenated to obtain a dimension of... The fused feature vector is shown in the following formula: ; in, This indicates a splicing operation.

[0031] Subsequently, a length of Step size is The sliding window will A continuous These are concatenated into a single sample vector. sample vectors Defined as: ; Each sample vector The dimension is .

[0032] Figure 1 The process of multimodal data fusion and sliding window preprocessing is presented.

[0033] Considering the differences between data sources and annotation scenarios in lithium-ion battery fault detection, and in accordance with the semi-supervised domain adaptation setting of this invention, the source domain and target domain are defined as follows: and .in, and All are sample vectors obtained through multimodal data fusion and sliding window preprocessing; and For the corresponding category label and , where 0 and 1 represent normal and fault respectively.

[0034] This invention is conducted under the semi-supervised domain adaptation (SSDA) setting. The source and target domains satisfy two common assumptions: (1) the label spaces are consistent, i.e. (2) Feature space correlation, i.e. and They follow different distributions but are essentially correlated. The goal of this invention is to learn a model capable of predicting the labels of unlabeled target domain samples using labeled source domain data and a limited number of labeled target domain samples. The overall setting of the semi-supervised domain adaptation is as follows: Figure 2 As shown.

[0035] Step S12 includes: The feature extractor aims to learn fault-sensitive discriminative features from multimodal battery data. Since changes in the raw data before a fault often occur are subtle, the architecture employs a combination of CNN and Graph Convolutional Network (GCN) to extract deep structural features for early fault detection.

[0036] The input to the feature extractor is a mini-batch sample matrix. , its origin sample vectors The matrix is ​​formed by stacking rows. The data is fed into a CNN, which extracts local temporal features through alternating convolutional and pooling layers. .

[0037] To model the structural correlations in the failure evolution process of lithium-ion batteries, the feature extractor extracts the feature matrix output by the CNN in the graph generation layer. The data is transformed into a graph structure. First, it is processed using a Multilayer Perceptron (MLP). Perform a nonlinear mapping to obtain the transformed feature matrix. : ; Then, the initial adjacency matrix is ​​calculated using the following formula: ; right Perform row-wise maximum normalization to obtain .

[0038] Next, a Top-k strategy is adopted, retaining only... The largest in each row Each value and its corresponding index are used to obtain the sparse adjacency matrix. Finally, construct the graph. ,in Representing node characteristics, The weights of the edges are defined.

[0039] Based on this graph structure, Graph Spectral Convolution (GConv) uses graph spectral filters. For input signal Smoothing: ; in For learnable parameter vectors, ; Represents the GConv operator; From the characteristic matrix Any column of represents a single feature channel. Here, Laplace matrix of symmetric normalized graph The eigenvectors are composed of, and are defined as: ; in Represents a node The degree, for The identity matrix. Corresponding to The graph Fourier transform.

[0040] To reduce computational complexity, the graph filter Restricted to a polynomial expansion based on Chebyshev polynomials, as described in the work of Defferrard et al. The graph convolution operation is extended to the entire feature matrix. We can obtain: ; in Let be the highest order of the Chebyshev polynomial. The scaled Laplace matrix is ​​defined as follows: , for The largest eigenvalue. Indicates the first The coefficients of the Chebyshev polynomial; Indicates the first Chebyshev polynomial.

[0041] To overcome the limitation of fixed receptive fields in standard GCNs, this invention employs a multi-receptive-field graph convolutional network (MRF-GCN), the structure of which is as follows: Figure 3 As shown. Specifically, the MRF-GConv layer in Different sensory fields ( Perform graph convolutions on each graph and concatenate the results as follows: ; in, These are the weight parameters of the GConv layer under different receptive fields; This represents the fused feature matrix. By stacking multiple MRF-GConv layers, the model can capture hierarchical structural features layer by layer, and the final output is denoted as... .

[0042] Step S13 includes: Structural features based on the output of the feature extractor (F) The task module comprises two core components: a label classifier (C) and a domain discriminator (D). A schematic diagram of the domain adversarial framework based on DAGCN is shown below. Figure 4 As shown.

[0043] The label classifier (C) is responsible for fault state identification. It uses a fully connected layer followed by Softmax activation to form a binary classification network. The input is structural features. The output is the predicted probability for each possible state, denoted as . .

[0044] The domain discriminator (D) is a key component for achieving cross-domain alignment. Due to domain covariate shift, a label classifier trained using only fully labeled source domain samples and a small number of labeled target domain samples may still struggle to generalize to the entire target domain. Therefore, D is used to discriminate structural features. From the source domain ( ) or target domain ( This provides an adversarial objective for the feature extractor to learn domain-invariant features. Structurally, D consists of fully connected layers with ReLU activation and Dropout, and finally estimates the probability through the Sigmoid output. .

[0045] The specific process of step S2 includes: S21. Objective function design; S22. Two-stage model optimization.

[0046] Step S21 includes: The overall objective function of DAGCN consists of a weighted sum of three parts: classification loss, domain alignment loss, and structural alignment loss.

[0047] Classification loss Based on labeled samples, the difference between the predicted labels of the label classifier and the true labels is measured using the cross-entropy loss function. ; in, The number of labeled samples. For the first The true label of each sample Its corresponding structural features, This is the probability predicted by the label classifier. This loss ensures that the model learns reliable battery fault discrimination rules, laying the foundation for cross-domain transfer.

[0048] Domain alignment loss Based on binary cross-entropy, this loss is used to measure the ability of the extracted structural features to distinguish between the source and target domains. This loss directly achieves the adversarial objective described in step S13: the domain discriminator minimizes... Training, with the expectation of training on source domain samples For target domain samples The feature extractor, on the other hand, maximizes... To generate domain-invariant features that are difficult for the discriminator to distinguish, ideally making The loss is defined as follows: ; in, and These represent the number of samples in the source domain and the target domain, respectively. and These are their corresponding structural features.

[0049] Primarily used for aligning global feature distributions, structural alignment loss further aligns the source and target domains at the graph structure level using Maximum Mean Discrepancy (MMD). MMD measures distribution discrepancy by calculating the distance between features mapped to the Reproducing Kernel Hilbert Space (RKHS), defined as follows: ; in, For feature mapping to RKHS, This indicates the RKHS.

[0050] Taking into account the above three losses, the overall objective function is defined as follows: ; in, and These are the trade-off parameters between control domain alignment and structural alignment strength, respectively.

[0051] Step S22 includes: The core objective of the initial training phase is to initialize the model parameters. Only using... Update feature extractor ( ) and label classifier ( ) parameters: ; ; in, The learning rate is minimized during the initial training phase. Achieving high fault detection accuracy provides a stable feature base for subsequent domain adaptation.

[0052] During the confrontation training phase, and and Combined to form total loss The optimization process simultaneously updates the feature extractor, label classifier, and domain discriminator. The parameters of the domain discriminator are as described in steps S13 and S21. The feature extractor maximizes the gradient through a gradient reversal layer (GRL). During forward propagation, GRL behaves as an identity mapping; during backward propagation, it will originate from... The gradient is inverted and then fed back to the feature extractor. Therefore, the domain discriminator receives data from... The standard gradient, while the feature extractor receives from... The inverted gradient and from and The standard gradient; the label classifier only receives from... The gradient. This adversarial process ultimately yields domain-invariant features, enabling reliable fault detection under variable operating conditions. The parameter update rules are as follows: ; ; .

[0053] The specific process of step S3 includes: S31. Selection of a representative task subset; S32. Hyperparameter optimization based on Bayesian optimization; S33. Fault detection decision and alarm triggering.

[0054] Step S31 includes: To reduce the computational cost of parameter optimization across all migration tasks, this invention employs both Maximum Mean Difference (MMD) and Fisher's discrimination rate to select representative tasks. MMD quantifies the distribution difference between the source and target domains, as given in step S21; the Fisher's rate measures the separability between normal and fault states within the target domain, and its calculation is as follows: ; in, , and , These represent the mean and variance of normal and faulty samples in the original feature space of the target domain, respectively. For each transfer task... The joint difficulty score is defined as: ; in, and These are the values ​​after min-max normalization. Based on... Based on the ranking results, all tasks are divided into three difficulty levels: high, medium, and low. A representative subset is selected from these levels for subsequent parameter optimization. The selection of this representative task subset is illustrated in the following example. Figure 5 .

[0055] Step S32 includes: In real-time health monitoring of lithium-ion batteries, fault detection performance is measured by fault detection rate (FDR), and timeliness is evaluated by fault detection delay (FDD). FDR and FDD are primarily affected by two key parameters: the sliding window size. The proportion of labeled data in the target domain To achieve a balance between high FDR and low FDD, it is necessary to determine the optimal combination of parameters. .

[0056] To simultaneously maximize FDR and minimize FDD, a unified fitness function is constructed. To eliminate the dimensional differences between the two and maintain a consistent optimization direction: ; in, The maximum allowable fault detection delay value; This is a weighting coefficient used to weigh the relative importance of accuracy and timeliness in battery fault detection.

[0057] Given a fitness function, the parameter optimization problem is transformed into finding a fitness value that satisfies the fitness function. Maximize the optimal combination of parameters ,in In the parameters The fitness values ​​obtained from the FDR and FDD calculations of the running model. Because and There are complex mapping relationships between them. This invention employs a Bayesian optimization method, using a Gaussian Process (GP) to represent this unknown black-box function. Construct a proxy model. The proxy model is denoted as a random function. Its prior assumption is: ; in For the sake of simplicity, we usually take the prior mean function. ; The covariance function is implemented using a Gaussian kernel: ; in, This is a length-scale hyperparameter.

[0058] Given historical observation data ,in For the first A combination of parameters, For the corresponding fitness value, the new parameter combination can be estimated using the posterior prediction distribution of the Gaussian process. The degree of adaptability.

[0059] The posterior prediction distribution is: ; Among them, the predicted mean The expected fitness level is represented by the standard deviation of the prediction. To represent uncertainty, the calculation is as follows: ; ; in, Let be the covariance matrix, and its elements ; for The covariance vector between each observation point ; To observe the noise variance; This is the observed fitness value vector.

[0060] After updating the surrogate model, this invention employs an Expected Improvement (EI) acquisition function to intelligently select the next evaluation point, thereby achieving a balance between exploring high-uncertainty regions and utilizing the current high-fitness regions. To maximize the overall performance of battery fault detection, the EI acquisition function is defined as follows: ; in, This is the current historical best fitness value; and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. This function tends to be utilized in regions of high fitness while encouraging exploration in regions of high uncertainty. By maximizing... It can efficiently find the optimal parameter combination with a limited number of model evaluations. This significantly improves the configuration efficiency of battery fault detection models.

[0061] The optimization process is performed on the representative task subset selected in step S31. For each set of candidate parameters... The fitness of the agent model is averaged across all representative tasks to update the agent model. The resulting optimal combination It will be directly applied to battery fault detection in all scenarios.

[0062] Step S33 includes: In real-time health monitoring of lithium-ion batteries, the fault detection performance of DAGCN is measured by the fault detection rate (FDR), and its timeliness is evaluated by the fault detection delay (FDD), defined as follows: ; ; Where TP represents the number of correctly identified fault samples, and FN represents the number of fault samples misclassified as normal. This represents the actual moment the failure occurred. This is the moment when the model first detects a fault.

[0063] In real-time detection, the system operates at a sampling frequency of 1 Hz. As described in step S11, the data comes from... A voltage sensor and The raw data from each temperature sensor is normalized by Z-score and fused using multimodal methods to generate feature vectors. Use the optimal sliding window size. ,Will consecutive eigenvectors Concatenate to form sample vector The data is then input into the trained DAGCN model to obtain the predicted class labels. : ; in, This is the structural feature matrix obtained by the feature extractor. Indicates the first One sample was predicted as faulty. This indicates that it is predicted to be normal. This represents the fault prediction probability given by the label classifier. When two consecutive predictions are both faults (i.e.,...), the probability of a fault is determined by the fact that... and When this occurs, a fault alarm is triggered.

[0064] The following detailed description of the implementation of this invention is based on specific experiments. The experimental platform consists of a temperature chamber, a battery testing system (BTS), a battery management system (BMS), a host computer for data acquisition and monitoring, and the battery under test. Considering the high risk, high cost, and poor repeatability of real ISC experiments, this invention employs an alternative experiment: and The equivalent resistances are connected to the middle position of the faulty battery cell along the longitudinal axis to simulate ISC faults of different severity.

[0065] like Figure 6 As shown, the battery pack consists of 24 cylindrical battery cells arranged in a 4-parallel, 6-series (4P6S) configuration. Based on this setup, six domains are constructed to represent different fault scenarios, as shown in Table 1.

[0066] Table 1 Description of the six domains

[0067] The main model parameters are listed in Table 2. According to the Bayesian optimization process of this invention, the sliding window size... Set to 30, the proportion of labeled data in the target domain. Set to 15%, sliding step size to 1, number of voltage / temperature sensors The sampling frequency is 1Hz and the batch size is 64.

[0068] Table 2 Parameter Setting Table

[0069] Each domain contains 2000 seconds of continuous data: the first 1000 seconds represent normal operation, and the last 1000 seconds represent an ISC (Independent System Capacity) fault. The fault occurrence time is specified in the table. Each domain yields 971 completely normal samples and 971 completely faulty samples. 29 transitional samples are generated near the fault location.

[0070] A total of 30 cross-domain migration tasks were constructed across six domains (removing source-target pairs). The proposed method was evaluated on the unlabeled target domain test set for all 30 tasks. Table 3 summarizes the overall performance metrics.

[0071] Table 330 Overall Performance Indicators for Migration Tasks

[0072] The accuracy and FDR results for all 30 transfer tasks are as follows: Figure 7 and Figure 8 As shown. For most tasks, both FDR and accuracy are close to 1, and many tasks even reach the ideal value of 1. The lowest FDR (0.982) appears on task [missing information]. and The lowest accuracy rate (0.970) occurred during the task. These results demonstrate that DAGCN exhibits excellent fault detection performance under variable operating conditions.

[0073] Figure 9 The rain cloud plot shows the distribution of FDD values; Figure 10 The heatmap shows the FDD values ​​for each task. Most tasks have low detection latency; 25 out of 30 tasks have FDD values ​​between 0 and 31 s. The maximum latency of 52 s occurs in task [unclear - likely a task name]. In the middle, the next step is the task. (51 s); Task The FDD is 0 s, meaning it was detected within the first fault window. The heatmap further shows that tasks targeting domain 0 typically have a larger delay, indicating that fault detection in this domain is more difficult.

[0074] Figure 11 The MMD values ​​for different domains are given. and The equal-domain pairs exhibited significant distributional differences. However, these tasks still achieved excellent fault detection performance: accuracy and FDR reached 1, and FDD were 26 s and 28 s, respectively. This indicates that DAGCN can effectively handle significant domain shifts.

[0075] To further investigate the changes in fault detection performance, Fisher's discriminant ratio was used to analyze the separability between normal and faulty samples in each domain, such as... Figure 12 As shown, the Fisher values ​​for domains 0 and 1 are relatively small (180.4 and 199.6 respectively), indicating poor separability between normal and faulty samples. In contrast, the Fisher values ​​for domains 3 and 4 are larger, making faults easier to distinguish. This is consistent with the FDD results: tasks using domain 0 or 1 as the target domain tend to have larger FDD values, indicating that poor separability increases the difficulty of fault detection.

[0076] To more intuitively demonstrate the separability between normal and faulty samples, Figure 13 Several t-SNE visualizations of target domain feature distributions are presented. For example... Figure 13As shown in (c), domain 3 exhibits clear separability; while domains 0 and 1 show significant sample overlap, as shown in (c). Figure 13 As shown in (a) and (b) in the figure. This visualization result is consistent with Fisher's analysis and further explains why the FDD is large when these domains are used as target domains.

[0077] Due to the difficulty of displaying all 30 migration tasks one by one, Figure 14 Three case studies were selected for presentation. For example... Figure 14 As shown in (a), the task This represents a challenging scenario: a subset of faulty samples are misclassified as normal, resulting in the lowest FDR (0.982) and the highest FDD (52 s). Conversely, for and ,like Figure 14 As shown in (b) and (c), the confusion matrix provides perfect classification results for both normal and faulty samples. Although the MMD values ​​are the largest for these two tasks, the FDR and accuracy both reach 1, indicating that the proposed method still has strong robustness under large domain differences.

[0078] To verify the domain discriminator (D) and the structural alignment loss ( The contribution of the proposed method was compared with four configurations on a task with domain 0 as the target domain: the complete DAGCN model, removing the domain discriminator (w / o discriminator), removing the structural alignment loss (w / o MMD), and removing both (w / o both). The results are shown in Table 4.

[0079] Table 4 Ablation Study Results (with domain 0 as the target domain)

[0080] Figure 15A , Figure 15B and Figure 15C This indicates that the domain discriminator and the structure alignment loss play complementary roles: the former provides effective domain invariance, while the latter guarantees the consistency of the graph structure. Their combined use enables fault detection of lithium-ion batteries with high FDR and low FDD under variable operating conditions.

[0081] The proposed DAGCN was compared with two traditional machine learning methods (SVM, PCA) and two unsupervised domain adaptation methods (JAN, CORAL), and the average results are summarized in Table 5.

[0082] Table 5. Comparison of different methods (with domain 0 as the target domain)

[0083] Compared to SVM and PCA, DAGCN exhibits significantly better robustness under cross-domain conditions, such as... Figure 16A , Figure 16B as well as Figure 16C As shown. In a challenging In the task, the FDD of SVM increased to 435 s, while the accuracy and FDR of PCA decreased to 0.523 and 0.042, respectively, and the FDD reached 696 s.

[0084] Compared to JAN and CORAL, DAGCN achieved the best overall performance across all migration tasks. JAN's accuracy was... The task accuracy dropped to 0.921; CORAL maintained a relatively high accuracy and FDR, but its detection latency was longer than DAGCN in all tasks.

[0085] Based on the above experimental results, in experiments involving six working conditions and 30 cross-domain tasks, the proposed method achieved an average accuracy of 99.7%, an average fault detection rate of 99.8%, and an average fault detection delay of 24.9 s. It outperforms conventional methods in terms of detection accuracy and robustness, demonstrating robust cross-domain fault detection performance under conditions of limited labeled data.

[0086] In summary, this invention significantly enhances the detection capability of early-stage weak faults in lithium-ion batteries by fusing multimodal data of voltage and temperature and introducing a multi-receptive-field graph convolutional network (MRF-GCN). The DAGCN framework based on domain adversarial training effectively alleviates the distribution offset problem across operating conditions, and the MMD structure alignment loss further ensures feature consistency in the graph domain. Under the semi-supervised domain adaptation setting where labeled samples in the target domain are scarce, the average fault detection rate reaches 99.8%, and the average fault detection latency is as low as 24.9 s. In a comprehensive experiment involving six operating conditions and 30 cross-domain tasks, this invention outperforms baseline methods such as SVM, PCA, JAN, and CORAL in terms of accuracy, fault detection rate, and latency, verifying its robust cross-domain fault detection performance under limited labeled data conditions.

[0087] Furthermore, the present invention also provides a computer device / apparatus / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.

[0088] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or block diagrams.

[0090] These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, thereby providing steps for implementing the functions specified in one or more flowcharts and / or one or more blocks in a block diagram.

[0092] It should be noted that, depending on the implementation needs, the various steps / components described in this invention can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0093] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault detection method for lithium-ion batteries under various operating conditions based on domain adversarial graph convolutional networks, characterized in that, The method includes: The fusion feature vector is obtained by concatenating all voltage and temperature data of the lithium-ion battery pack at the same time. The fusion feature vectors of multiple consecutive time points are concatenated into a sample vector and labeled with corresponding category labels through a sliding window of a set length. This constructs a fully labeled source domain dataset and a target domain dataset containing a small number of labeled samples. A domain adversarial graph convolutional network is constructed, comprising a feature extractor, a label classifier, and a domain discriminator. The feature extractor includes a convolutional neural network, a graph generation layer, and a multi-receptive-field graph convolutional network. The convolutional neural network extracts temporal features of samples, the graph generation layer transforms the temporal features into graph structure data, and the multi-receptive-field graph convolutional network extracts structural features based on the graph structure data. The label classifier is used to predict the fault category of the structural features, and the domain discriminator is used to determine the domain to which the structural features belong. We construct the objective function of the domain adversarial graph convolutional network and use a two-stage optimization strategy for model training. The sample vector to be detected is input into a trained domain adversarial graph convolutional network, structural features are extracted by a feature extractor, and the detection results are output by a label classifier.

2. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 1, characterized in that, The graph generation layer transforms temporal features into graph-structured data, including: The temporal features are nonlinearly mapped using a multilayer perceptron to obtain the transformed feature matrix. The initial adjacency matrix is ​​calculated based on the transformed feature matrix. The initial adjacency matrix is ​​then subjected to row-wise maximum normalization, and the Top-k strategy is used to retain the k largest values ​​in each row and their corresponding indices to obtain the sparse adjacency matrix. Finally, the transformed feature matrix is ​​used as the node features and the sparse adjacency matrix is ​​used as the edge weights to construct the graph structure data.

3. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 1, characterized in that, Multi-receptive-field graph convolutional networks extract structural features from graph-structured data, including: Multiple Chebyshev polynomial graph filters of different orders are set up to correspond to different receptive field scales. Based on the normalized Laplacian matrix, graph convolution operation is performed on the graph structure data at each receptive field scale and the feature matrix of the corresponding scale is output. The feature matrices output from all receptive field scales are concatenated to obtain the fused multi-scale structural features. The hierarchical structural features are then extracted layer by layer by stacking multiple multi-receptive field graph convolution layers.

4. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 1, characterized in that, The objective function of a domain adversarial graph convolutional network includes classification loss, domain alignment loss, and structural alignment loss. The classification loss is the cross-entropy loss of labeled samples, used to measure the difference between the predicted labels and the true labels of the label classifier. The domain alignment loss is the binary cross-entropy loss, used to measure the classification performance of the domain discriminator on the structural features of the source and target domains. The structural alignment loss is the maximum mean difference loss, used to map the structural features of the source and target domains to the regenerating kernel Hilbert space and measure the distribution difference by calculating the distance between the mapped features, thereby aligning the source and target domains at the graph structure level.

5. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 4, characterized in that, A two-stage optimization strategy is employed for model training to achieve domain-invariant feature learning across different operating conditions, including: Initial training phase: Using labeled samples from the source domain dataset and the target domain dataset, update the parameters of the feature extractor and label classifier based solely on the classification loss; Adversarial training phase: The domain discriminator updates its parameters with the goal of minimizing the domain adversarial loss; the feature extractor receives the inverted gradient of the domain adversarial loss through the gradient inversion layer, and updates its parameters with the goal of maximizing the classification error of the domain discriminator and combining the gradients of the classification loss and the structure alignment loss; the label classifier updates its parameters based only on the gradient of the classification loss.

6. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 1, characterized in that, This method also includes a key hyperparameter optimization step: Select a representative subset from all migration tasks; Using the sliding window size and the proportion of labeled data in the target domain as parameters to be optimized, a fitness function that balances the accuracy and timeliness of fault detection is constructed. The fitness function is used as the black-box objective function, and a Bayesian optimization method is employed to find the optimal value on the representative subset. Specifically, a surrogate model of the black-box objective function is constructed using a Gaussian process, and the next evaluation point is selected based on the expected improvement of the acquisition function. The process is iteratively updated until the optimal parameter combination that maximizes the fitness function value is output.

7. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 6, characterized in that, Methods for selecting representative subsets include: Calculate the maximum mean difference between the source and target domains, and the Fisher discrimination rate between normal and faulty samples within the target domain, to construct a joint difficulty score, calculated using the following formula: ； in, The maximum mean difference between the normalized source and target domains. The Fisher discrimination rate between normal and faulty samples within the target domain after normalization. Based on the joint difficulty score, all migration tasks are divided into different difficulty levels, and a representative subset is selected from each difficulty level.

8. The cross-condition lithium-ion battery fault detection method based on domain adversarial graph convolutional networks according to claim 6, characterized in that, The fitness function is: ； Where FDR is the fault detection rate and FDD is the fault detection delay. For the maximum allowable fault detection delay, The weighting coefficients are used to balance the accuracy and timeliness of fault detection.

9. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 8.

Images