Battery pack consistency evaluation method based on space magnetic field imaging and deep learning

Abstract

The invention relates to the technical field of lithium ion battery pack state monitoring and fault diagnosis, in particular to a battery pack consistency evaluation method based on space magnetic field imaging and deep learning, which comprises the following steps: data acquisition and preprocessing: obtaining standardized input data for evaluation; constructing and training a dual-path deep learning network model; carrying out model training and online evaluation; the method has the beneficial effects that global magnetic field distribution change and local magnetic field distortion features caused by an abnormal battery can be captured at the same time through a dual-branch feature enhancement module in a magnetic field image path, and omission of fine abnormal features is avoided. Meanwhile, working condition context information provided by an electric parameter path dynamically guides image feature focusing through cross-modal attention, so that the model can interpret a magnetic field image in combination with the actual working state of the battery, and deep complementation and cooperative enhancement of electric-magnetic physical information are realized; therefore, the method shows extremely high detection sensitivity and overall evaluation precision for early and slight inconsistency.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery pack state monitoring and fault diagnosis, specifically a battery pack consistency evaluation method based on spatial magnetic field imaging and deep learning. Background Technology

[0002] Currently, battery management systems typically assess consistency by monitoring external electrothermal parameters such as the total voltage and current of the battery pack, as well as the voltage of individual cells. These methods are relatively simple to implement and low in cost, and are therefore widely used. In addition, some studies have attempted to assess battery consistency by analyzing the voltage sequence during battery charging and discharging, or by considering the impact of ambient temperature changes on open-circuit voltage, to determine the entropy-thermal coefficient. These methods primarily rely on indirect judgments based on the battery's external performance parameters.

[0003] In the prior art, the closest to this invention is a battery pack consistency detection method based on magnetic field imaging. Specifically, it includes the following two technical solutions: During the constant current discharge of the lithium-ion battery pack under test, the magnetic field distribution outside the battery pack is measured at equal discharge capacity intervals. Subsequently, the relative changes in the magnetic field distribution within each equal capacity interval are calculated, and statistical analysis methods (such as calculating the arithmetic mean and dividing the magnetic field map into sub-regions to extract statistical feature vectors) are used to process these changes. Finally, statistical learning methods (such as principal component analysis) are used to analyze these feature vectors, thereby achieving the evaluation of battery pack consistency and the location of batteries with abnormal performance.

[0004] In the existing technology, the core of "Lithium battery pack performance consistency detection technology and system based on in-situ magnetic field sensing" is also to analyze the uneven current changes in the battery pack by sensing the changes in the magnetic field in the external space of the battery pack, thereby evaluating the performance consistency of the battery pack and accurately locating abnormal batteries.

[0005] Compared to traditional electrical parameter monitoring, the two magnetic field detection schemes mentioned above both attempt to achieve non-destructive and non-contact detection through the physical quantity of magnetic field, and aim to solve the problem of ensuring the consistency of all individual cells inside the battery pack under online monitoring conditions.

[0006] Although the aforementioned existing magnetic field imaging-based technologies have made progress in non-invasive detection, they still have the following significant technical drawbacks compared to the proposal of this application. These drawbacks are precisely the core technical problems that this application aims to solve: 1. Poor adaptability to dynamic operating conditions and insufficient robustness: Schemes represented by the "magnetic field imaging detection method based on statistical analysis" rely on the prerequisite of "equal discharge capacity interval" for magnetic field data acquisition. This means that this method is only applicable to scenarios with constant current or slow changes in operating conditions. In practical applications, such as dynamic loads like frequency regulation in electric vehicles or energy storage systems, the current and power of the battery pack change in real time, failing to meet the measurement requirements of "equal discharge capacity interval," leading to the method's failure or a significant decrease in accuracy. Therefore, this application proposes to address the technical problem of maintaining high robustness of the battery pack consistency evaluation model under complex dynamic operating conditions.

[0007] 2. Limited Feature Extraction Methods and Insensitivity to Early Subtle Anomalies: Existing magnetic field detection schemes largely rely on statistical analysis methods or basic feature engineering, failing to deeply explore the rich spatial features contained in magnetic field images. They often treat magnetic field images as ordinary optical image processing, lacking mechanisms to specifically enhance subtle global distribution pattern changes and local magnetic field distortion features caused by abnormal batteries. This homogeneous feature extraction approach results in insensitivity to early, weak inconsistencies that arise during the initial service life of battery packs or when performance is slightly degraded, limiting their early warning capabilities. Therefore, this application proposes to address the technical challenge of how to extract multi-scale visual features more fully and intelligently from magnetic field images to improve the detection sensitivity for early anomalies.

[0008] 3. Failure to integrate multimodal physical information, resulting in insufficient basis for model decision-making: Existing solutions typically only process magnetic field data, failing to effectively combine real-time, easily obtainable electrical parameters such as battery pack terminal voltage and total current. This ignores the inherently close coupling between "electricity" and "magnetism." Due to the lack of contextual guidance from the electrical operating state, the model struggles to distinguish between global magnetic field fluctuations caused by normal power changes and local magnetic field distortions caused by battery anomalies, leading to misjudgments. Furthermore, single-modal data exhibits instability when faced with sensor noise or poor data quality. Therefore, this application proposes to address the technical challenge of deeply integrating electro-magnetic multimodal physical information to achieve information complementarity, thereby improving evaluation accuracy and model reliability in real-world environments.

[0009] 4. Limited or absent positioning capabilities hinder precise maintenance: While some existing technologies propose positioning intentions, their implementation methods (such as dividing areas based on statistical analysis) lack sufficient positioning accuracy and reliability. They often fail to directly output the anomaly probability of each individual cell end-to-end, failing to provide maintenance personnel with a clear and intuitive basis for decision-making regarding "where the anomaly is," making subsequent precise maintenance (such as replacing specific cells) difficult to implement. Therefore, this application proposes to address the technical problem of how to accurately locate abnormal individual cells within a group while completing a consistency status assessment. Summary of the Invention

[0010] The purpose of this invention is to provide a battery pack consistency evaluation method based on spatial magnetic field imaging and deep learning, so as to overcome the shortcomings of existing battery pack consistency evaluation technologies in terms of sensitivity, positioning capability and feature utilization mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a battery pack consistency evaluation method based on space magnetic field imaging and deep learning, comprising the following steps: Data acquisition and preprocessing: Acquiring standardized input data for evaluation; Constructing and training a dual-path deep learning network model; Model training and online evaluation.

[0012] Preferably, the specific method of data acquisition in data acquisition and preprocessing is as follows: when the battery pack is under constant current charging and discharging or a specific dynamic operating condition, two types of data acquisition are carried out simultaneously. An M×N magnetic field sensor array is used, arranged in a planar grid form above the battery pack or on a specific side to acquire the original data matrix B of its spatial magnetic field distribution; the total terminal voltage V and total current I of the battery pack are acquired simultaneously; M and N are both greater than 1, and M×N is not less than 4.

[0013] Preferably, the specific method of data preprocessing in data acquisition and preprocessing is as follows: First, using a normal battery pack with consistent individual cell performance, collect its magnetic field data under the same standard operating conditions. After averaging, this data serves as the reference magnetic field data matrix B_ref. For the battery pack under test, perform matrix difference operations on its collected magnetic field data matrix B_test and B_ref to obtain the differential magnetic field matrix ΔB, i.e., ΔB = B_test - B_ref. Normalize the ΔB matrix and map its element values ​​to pixel grayscale values ​​to generate an M×N pixel two-dimensional grayscale image, i.e., the differential magnetic field image I_m. Alternatively, generate a pseudo-color image of the same size based on a preset color mapping table.

[0014] Preferably, the design of the magnetic field image feature extraction path in the construction and training of the dual-path deep learning network model is as follows: the generated differential magnetic field image I_m is used as input; a backbone convolutional neural network is included to extract basic visual features; a dual-branch feature enhancement module is set up, which works in parallel after the backbone network, including a global context branch and a local salient region branch; the multi-scale features output by the global context branch and the weighted features output by the local salient region branch are concatenated in the channel dimension to form the final multi-scale enhanced magnetic field image feature F_m.

[0015] Preferably, the global context branch in the magnetic field image feature extraction path uses a spatial pyramid pooling layer to perform multi-scale pooling on the basic feature map output by the backbone network in order to capture global context information of different ranges in the image and perceive the changes in the overall magnetic field distribution pattern caused by abnormal batteries.

[0016] Preferably, the local salient region branch in the magnetic field image feature extraction path adopts a convolution-based attention mechanism, specifically a convolutional block attention module. This module sequentially calculates attention weight maps on the input feature map in the channel dimension and spatial dimension to generate a feature salient map for identifying key regions. Then, the salient map is multiplied element-wise with the basic feature map output by the backbone network, thereby adaptively weighting and strengthening the local magnetic field distortion features in the image corresponding to the abnormal battery.

[0017] Preferably, the design of the electrical parameter feature extraction path in the construction and training of the dual-path deep learning network model is as follows: the vector consisting of the total terminal voltage V and the total current I of the battery pack, which are collected synchronously with the magnetic field data in time, is used as the input; after normalizing the vector [V,I], it is input to an encoder network consisting of at least two fully connected layers, and mapped into a high-dimensional, dense electrical parameter feature vector F_e through nonlinear transformation.

[0018] Preferably, the workflow of the cross-modal attention fusion module in the construction and training of the dual-path deep learning network model is as follows: the electrical parameter feature vector F_e is projected into a query vector through a linear transformation; the magnetic field image feature F_m is reshaped into a feature sequence and then projected into a key vector and a value vector through linear transformations respectively; the scaled dot product of the query vector and all key vectors is calculated, and the attention weight distribution is obtained after normalization by the Softmax function; the value vector is weighted and summed using the attention weights obtained above, and finally a joint feature representation F_fused that integrates the electro-magnetic physical correlation is output.

[0019] Preferably, the design of the classification and localization output layers in the dual-path deep learning network model for construction and training is as follows: It consists of two independent sub-networks for integrated output; the consistency state classification sub-network flattens the fused feature vector F_fused, passes it through a fully connected layer and a Softmax activation function, and outputs a probability distribution representing the probability that the battery pack belongs to multiple predefined consistency state categories; the abnormal battery localization sub-network flattens the fused feature vector F_fused, passes it through another fully connected layer and a Sigmoid activation function, and outputs a probability vector of length N_cell, where each element represents the probability that the corresponding battery cell is abnormal, thus achieving accurate localization of abnormal batteries.

[0020] The preferred method for model training and online evaluation is as follows: Model Training: The dual-path deep learning network is trained end-to-end using a labeled dataset containing various anomaly locations and types to optimize all its parameters; Online Evaluation and Localization: The real-time differential magnetic field image I_m obtained after processing of the battery pack under test and the synchronously acquired and normalized electrical parameter vectors are input into the trained model. The model synchronously outputs the classification result of the overall consistency state of the battery pack and the anomaly probability distribution information of each battery cell location within the pack.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a battery pack consistency evaluation method based on spatial magnetic field imaging and deep learning. By introducing a dual-path architecture and a cross-modal attention fusion mechanism, it fundamentally solves the problem of insufficient traditional single feature extraction. Specifically, the "dual-branch feature enhancement module" in the magnetic field image path can simultaneously capture global magnetic field distribution changes caused by abnormal batteries (through spatial pyramid pooling) and local magnetic field distortion features (through weighted enhancement via an attention mechanism), avoiding the omission of subtle abnormal features. Simultaneously, the operating context information provided by the electrical parameter path dynamically guides image feature focusing through cross-modal attention, enabling the model to interpret the magnetic field image in conjunction with the actual operating state of the battery. This achieves deep complementarity and synergistic enhancement of electro-magnetic physical information, thus exhibiting extremely high detection sensitivity and overall evaluation accuracy for early and minor inconsistencies.

[0022] This solution achieves a breakthrough in precise location, moving beyond simply identifying "abnormalities" to pinpointing "where the abnormality is." Through end-to-end network design, it integrates consistency classification with anomaly localization. The key technical aspect lies in the localization sub-network directly using the deeply fused feature F_fused as input. This feature incorporates spatial location information enhanced by branches in local salient regions and is filtered for physical correlations of electrical parameters. Therefore, the model can not only determine whether the battery pack as a whole is abnormal but also accurately output the anomaly probability of each individual battery within the pack. This leap from macroscopic state assessment to microscopic individual battery localization provides a direct and reliable basis for subsequent precise maintenance (such as replacing specific batteries), overcoming the limitations of traditional methods that can only trigger alarms but not locate the anomaly. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall method steps of the present invention; Figure 2 This is a diagram of the dual-path network structure of the present invention; Figure 3 This is a diagram of the cross-modal attention module of the present invention; Figure 4 This is the magnetic field direction diagram of the battery of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-4 This invention provides a technical solution: a battery pack consistency evaluation method based on space magnetic field imaging and deep learning, comprising the following steps: S1: Data Acquisition and Preprocessing This step aims to obtain standardized input data for evaluation.

[0026] Data Acquisition: Two types of data are collected simultaneously when the battery pack is under constant current charging / discharging or specific dynamic operating conditions: 1. Use an M×N magnetic field sensor array (preferably a tunnel magnetoresistive sensor or a Hall effect sensor) (M and N are both greater than 1, and M×N is not less than 4) to arrange the array in a planar grid above the battery pack or on a specific side to collect the raw data matrix B of its spatial magnetic field distribution.

[0027] 2. Synchronously collect the total terminal voltage V and total current I of the battery pack.

[0028] Data preprocessing: 1. Use a normal battery pack with all individual cells of the same performance to collect its magnetic field data under the same standard operating conditions. After averaging, the data is used as the reference magnetic field data matrix B_ref.

[0029] 2. For the battery pack under test, perform matrix difference operation on the magnetic field data matrix B_test and B_ref collected to obtain the differential magnetic field matrix ΔB (i.e., ΔB=B_test-B_ref).

[0030] 3. Normalize the ΔB matrix and map its element values ​​to pixel grayscale values ​​to generate a two-dimensional grayscale image of M×N pixels, i.e., the differential magnetic field image I_m; or generate a pseudo-color image of the same size according to a preset color mapping table.

[0031] S2: Constructing and training a dual-path deep learning network model This model is the core of this invention, and its structural design is as follows: 1. Feature extraction path for magnetic field images: The differential magnetic field image I_m generated in step S1 is used as input.

[0032] This path includes a backbone convolutional neural network (such as ResNet, VGG, etc.) for extracting basic visual features.

[0033] Key innovation: Dual-branch feature enhancement module. This module operates in parallel after the backbone network. Global Context Branch: Spatial pyramid pooling layers are used to perform multi-scale (e.g., 1x1, 2x2, 4x4) pooling on the basic feature maps output by the backbone network to capture global context information of different ranges in the image and perceive changes in the overall magnetic field distribution pattern caused by abnormal batteries.

[0034] Local salient region branch: A convolution-based attention mechanism (specifically, a convolutional block attention module, which sequentially calculates attention weight maps on the input feature map in both the channel and spatial dimensions) is employed to generate a feature saliency map for identifying key regions. Subsequently, this saliency map is multiplied element-wise with the base feature map output by the backbone network, thereby adaptively weighting and strengthening the local magnetic field distortion features in the image corresponding to the anomalous battery.

[0035] Finally, the multi-scale features output by the global context branch and the weighted features output by the local salient region branch are concatenated along the channel dimension to form the final multi-scale enhanced magnetic field image feature F_m.

[0036] 2. Electrical parameter feature extraction path: The vector consisting of the total terminal voltage V and total current I of the battery pack, which are acquired synchronously with the magnetic field data in time, is used as the input.

[0037] After normalizing the vector `[V,I]`, it is input into an encoder network consisting of at least two fully connected layers, and mapped to a high-dimensional, dense electrical parameter feature vector F_e through nonlinear transformation.

[0038] 3. Cross-modal attention fusion module: This module is used for deep fusion of features from different modalities (images and electrical signals). Its workflow is as follows: The electrical parameter feature vector F_e is projected into a query vector through a linear transformation.

[0039] After reshaping the magnetic field image features F_m into a feature sequence, they are projected into key vectors and value vectors through linear transformations.

[0040] The scaled dot product of the query vector and all key vectors is calculated, and the attention weight distribution is obtained after normalization using the Softmax function. This process enables the network to dynamically focus on the most relevant region in the magnetic field image features (F_m) based on the current electrical operating state (F_e).

[0041] The value vector is weighted and summed using the attention weights obtained above, and finally a joint feature representation F_fused that incorporates the electro-magnetic physical correlation is output.

[0042] 4. Classification and Location Output Layer: This layer consists of two independent sub-networks, used to achieve unified output: Consistency State Classification Subnetwork: After flattening the fused feature vector F_fused, it is passed through a fully connected layer and a Softmax activation function to output a probability distribution, which represents the probability that the battery pack belongs to multiple predefined consistency state categories (e.g., "normal", "slightly inconsistent", "severely inconsistent").

[0043] Abnormal Battery Localization Subnetwork: After flattening the fused feature vector F_fused, it is passed through another fully connected layer and a Sigmoid activation function to output a probability vector of length N_cell, which is the total number of individual cells in the battery pack. Each element represents the probability that the corresponding cell is abnormal, thereby achieving accurate localization of abnormal batteries.

[0044] S3: Model Training and Online Evaluation Model training: The dual-path deep learning network was trained end-to-end using a labeled dataset containing various anomaly locations and types (obtained by introducing known anomalous batteries at different locations) to optimize all its parameters.

[0045] Online assessment and location: The real-time differential magnetic field image I_m obtained from step S1 of the battery pack under test, along with the synchronously acquired and normalized electrical parameter vectors, are input into the trained model. The model will simultaneously output two results: one is the classification result of the overall consistency status of the battery pack, and the other is the anomaly probability distribution information of the location of each individual battery cell within the pack, for maintenance personnel to make decisions.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A battery pack consistency evaluation method based on space magnetic field imaging and deep learning, characterized in that: Includes the following steps: Data acquisition and preprocessing: Acquiring standardized input data for evaluation; Constructing and training a dual-path deep learning network model; Model training and online evaluation.

2. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The specific data acquisition method in data acquisition and preprocessing is as follows: When the battery pack is under constant current charging and discharging or a specific dynamic operating condition, two types of data acquisition are carried out simultaneously. An M×N magnetic field sensor array is used, which is arranged in a planar grid above the battery pack or on a specific side to acquire the original data matrix B of its spatial magnetic field distribution; the total terminal voltage V and total current I of the battery pack are acquired simultaneously; M and N are both greater than 1, and M×N is not less than 4.

3. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The specific method of data preprocessing in data acquisition and preprocessing is as follows: First, using a normal battery pack with consistent individual cell performance, collect its magnetic field data under the same standard operating conditions. After averaging, this data serves as the reference magnetic field data matrix B_ref. For the battery pack under test, perform matrix difference operations on its collected magnetic field data matrix B_test and B_ref to obtain the differential magnetic field matrix ΔB, i.e., ΔB = B_test - B_ref. Normalize the ΔB matrix and map its element values ​​to pixel grayscale values ​​to generate an M×N pixel two-dimensional grayscale image, i.e., the differential magnetic field image I_m; or generate a pseudo-color image of the same size based on a preset color mapping table.

4. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The design of the magnetic field image feature extraction path in the construction and training of the dual-path deep learning network model is as follows: the generated differential magnetic field image I_m is used as input; a backbone convolutional neural network is included to extract basic visual features. A dual-branch feature enhancement module is set up, which works in parallel after the backbone network and includes a global context branch and a local salient region branch. The multi-scale features output from the global context branch and the weighted features output from the local salient region branch are concatenated along the channel dimension to form the final multi-scale enhanced magnetic field image feature F_m.

5. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: In the magnetic field image feature extraction path, the global context branch uses a spatial pyramid pooling layer to perform multi-scale pooling on the basic feature map output by the backbone network in order to capture global context information of different ranges in the image and perceive the changes in the overall magnetic field distribution pattern caused by abnormal batteries.

6. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The local salient region branch in the magnetic field image feature extraction path adopts a convolution-based attention mechanism, specifically a convolutional block attention module. This module sequentially calculates attention weight maps on the input feature map in the channel dimension and spatial dimension to generate a feature salient map for identifying key regions. Then, the salient map is multiplied element-wise with the basic feature map output by the backbone network, thereby adaptively weighting and strengthening the local magnetic field distortion features in the image corresponding to the abnormal battery.

7. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The design of the electrical parameter feature extraction path in the construction and training of the dual-path deep learning network model is as follows: the vector composed of the total terminal voltage V and the total current I of the battery pack, which are collected synchronously with the magnetic field data in time, is used as the input; after normalizing the vector [V,I], it is input into an encoder network composed of at least two fully connected layers, and mapped into a high-dimensional, dense electrical parameter feature vector F_e through nonlinear transformation.

8. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The workflow of the cross-modal attention fusion module in the construction and training of the dual-path deep learning network model is as follows: the electrical parameter feature vector F_e is projected into a query vector through a linear transformation; the magnetic field image feature F_m is reshaped into a feature sequence and then projected into a key vector and a value vector through linear transformations respectively; the scaled dot product of the query vector and all key vectors is calculated, and the attention weight distribution is obtained after normalization by the Softmax function; the value vector is weighted and summed using the attention weights obtained above, and finally a joint feature representation F_fused that integrates the electro-magnetic physical correlation is output.

9. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The design of the classification and localization output layers in the construction and training of the dual-path deep learning network model is as follows: The output layers consist of two independent sub-networks for integrated output. The consistent state classification sub-network flattens the fused feature vector F_fused and passes it through a fully connected layer and a Softmax activation function to output a probability distribution representing the probability that the battery pack belongs to multiple predefined consistent state categories. The abnormal battery localization sub-network flattens the fused feature vector F_fused and passes it through another fully connected layer and a Sigmoid activation function to output a probability vector of length N_cell, where each element represents the probability that the corresponding battery cell is abnormal, thus achieving precise localization of abnormal batteries.

10. The battery pack consistency evaluation method based on space magnetic field imaging and deep learning according to claim 1, characterized in that: The specific methods for model training and online evaluation are as follows: Model Training: The dual-path deep learning network is trained end-to-end using a labeled dataset containing various anomaly locations and types to optimize all its parameters; Online Evaluation and Localization: The real-time differential magnetic field image I_m obtained after processing of the battery pack under test and the synchronously acquired and normalized electrical parameter vectors are input into the trained model. The model synchronously outputs the classification result of the overall consistency state of the battery pack and the anomaly probability distribution information of each battery cell location within the pack.

Images