CN-121978528-A - Battery pack consistency assessment method based on space magnetic field imaging and deep learning

CN121978528ACN 121978528 ACN121978528 ACN 121978528ACN-121978528-A

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 assessment method based on space magnetic field imaging and deep learning, which comprises the following steps of data acquisition and preprocessing, wherein standardized input data for assessment are acquired; constructing and training a dual-path deep learning network model; the model training and online evaluation have the beneficial effects that the global magnetic field distribution change and the local magnetic field distortion feature caused by the abnormal battery can be captured simultaneously through the double-branch feature enhancement module in the magnetic field image path, so that the omission of the fine abnormal feature is avoided. Meanwhile, the working condition context information provided by the electric parameter path is focused by cross-modal attention dynamic guiding image features, so that the model can read a magnetic field image in combination with the actual working state of the battery, and the depth complementation and cooperative enhancement of the electric-magnetic physical information are realized, so that extremely high detection sensitivity and overall evaluation accuracy are shown for early and slight inconsistency.

Inventors

MAO LEI
LI YAN
CHE SHOUQUAN
ZHENG DONGQIAN
SUN YUNING
YANG RUI
WANG SIMING
SONG LIJUN

Assignees

中国科学技术大学

Dates

Publication Date: 20260505
Application Date: 20251126

Claims (10)

1. The battery pack consistency assessment method based on the space magnetic field imaging and the deep learning is characterized by comprising the following steps of: Data acquisition and pretreatment: 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 assessment method based on spatial magnetic field imaging and deep learning according to claim 1 is characterized in that the specific mode of data acquisition in data acquisition and preprocessing is that two types of data acquisition are synchronously carried out when the battery pack is in constant current charge and discharge or specific dynamic working conditions, an M multiplied by N magnetic field sensor array is used and is arranged above the battery pack or on a specific side face in a planar grid mode, an original data matrix B of spatial magnetic field distribution of the battery pack is acquired, the total terminal voltage V and the total current I of the battery pack are synchronously acquired, and M and N are both larger than 1, and M multiplied by N is not smaller than 4.
3. The battery pack consistency assessment method based on space magnetic field imaging and deep learning according to claim 1 is characterized by comprising the specific modes of pre-using normal battery packs with consistent performances of all single batteries in data acquisition and pre-processing, acquiring magnetic field data of the normal battery packs under the same standard working condition, taking the average processed magnetic field data as a reference magnetic field data matrix B_ref, carrying out matrix differential operation on the acquired magnetic field data matrix B_test and B_ref of the battery pack to be tested to obtain a differential magnetic field matrix delta B, namely delta B=B_test-B_ref, carrying out normalization processing on the delta B matrix, mapping element values of the delta B matrix into pixel gray values according to the pixel gray values, and generating a two-dimensional gray image of M multiplied by N pixels, namely a differential magnetic field image I_m, or generating a pseudo-color image with the same size according to a preset color mapping table.
4. The battery pack consistency assessment method based on space magnetic field imaging and deep learning according to claim 1, wherein the method is characterized in that a magnetic field image feature extraction path in a dual-path deep learning network model is constructed and trained by taking a generated differential magnetic field image I_m as input, a main convolution neural network is included for extracting basic visual features, a dual-branch feature enhancement module is arranged, the module works in parallel after the main network and comprises a global context branch and a local salient region branch, and multi-scale features output by the global context branch and weighted features output by the local salient region branch are spliced in a channel dimension to form a final multi-scale enhanced magnetic field image feature F_m.
5. The battery pack consistency assessment method based on spatial magnetic field imaging and deep learning according to claim 1, wherein global context branches in a magnetic field image feature extraction path adopt a spatial pyramid pooling layer, and a basic feature map output by a main network is subjected to multi-scale pooling so as to capture global context information in different ranges in an image, and overall magnetic field distribution mode change caused by abnormal batteries is perceived.
6. The method for evaluating consistency of a battery pack based on space magnetic field imaging and deep learning according to claim 1, wherein a local salient region branch in a magnetic field image feature extraction path adopts a convolution-based attention mechanism, specifically a convolution block attention module, which sequentially calculates attention weight graphs of an input feature graph in a channel dimension and a space dimension to generate a feature salient graph for identifying a key region, and then multiplies the salient graph by a basic feature graph output by a backbone network element by element, so as to adaptively weight local magnetic field distortion features corresponding to abnormal batteries in an enhanced image.
7. The battery pack consistency assessment method based on space magnetic field imaging and deep learning according to claim 1 is characterized in that the design of an electric parameter characteristic extraction path in a double-path deep learning network model is constructed and trained by taking a vector formed by total terminal voltage V and total current I of a battery pack acquired synchronously in time with magnetic field data as input, normalizing the vector [ V, I ] and then inputting the vector into an encoder network formed by at least two fully connected layers, and mapping the vector into a high-dimensional and dense electric parameter characteristic vector F_e through nonlinear transformation.
8. The battery pack consistency assessment method based on space magnetic field imaging and deep learning according to claim 1 is characterized in that the workflow of a cross-modal attention fusion module in a dual-path deep learning network model is constructed and trained by projecting an electric parameter feature vector F_e into a query vector through linear transformation, projecting a magnetic field image feature F_m into a feature sequence through linear transformation respectively, calculating scaling dot products of the query vector and all key vectors, normalizing by a Softmax function to obtain attention weight distribution, and carrying out weighted summation on the obtained attention weight value vector to finally output a joint feature representation F_fused fused with electric-magnetic physical association.
9. The battery pack consistency assessment method based on space magnetic field imaging and deep learning according to claim 1 is characterized in that the design of a classification and positioning output layer in a dual-path deep learning network model is constructed and trained to be composed of two independent sub-networks for realizing integrated output, a consistency state classification sub-network is used for outputting a probability distribution for representing the probability that a battery pack belongs to a plurality of predefined consistency state categories through a full-connection layer and Softmax activation function after a fusion feature vector F_fused is flattened, and an abnormal battery positioning sub-network is used for outputting a probability vector with the length of N_cell of a battery pack through another full-connection layer and Sigmoid activation function after the fusion feature vector F_fused is flattened, wherein each element value represents the probability that a battery cell with a corresponding serial number is abnormal, and accurate positioning of an abnormal battery is realized.
10. The method for evaluating the consistency of the battery pack based on the spatial magnetic field imaging and the deep learning according to claim 1, wherein the specific modes of model training and online evaluation are as follows: The model training comprises the steps of carrying out end-to-end training on the dual-path deep learning network by using a labeling data set containing various abnormal positions and types, optimizing all parameters of the dual-path deep learning network, carrying out online evaluation and positioning, namely, inputting a real-time differential magnetic field image I_m obtained by processing a battery pack to be tested and an electric parameter vector which is synchronously acquired and normalized into a trained model, and synchronously outputting a classification result of the overall consistency state of the battery pack and abnormal probability distribution information of each battery cell position in the battery pack by the model.

Description

Battery pack consistency assessment method based on space magnetic field imaging and deep learning Technical Field The invention relates to the field of state monitoring and fault diagnosis of lithium ion batteries, in particular to a battery consistency assessment method based on space magnetic field imaging and deep learning. Background Currently, battery management systems typically evaluate consistency by monitoring external electrical heating parameters such as total voltage, total current, and voltage of each cell of the battery pack. These methods are relatively simple and inexpensive to implement and are therefore widely used. In addition, some studies have attempted to evaluate the consistency of batteries by analyzing the voltage sequence during battery charging and discharging, or determining the entropy coefficient in combination with the effect of ambient temperature changes on open circuit voltage. These methods rely primarily on external performance parameters of the battery for indirect judgment. In the prior art, the method is closest to the method for detecting the consistency of the battery pack based on magnetic field imaging. Specifically, the method comprises the following two technical schemes: And in the constant-current discharge process of the lithium battery pack to be tested, measuring the magnetic field distribution outside the battery pack at equal discharge capacity intervals. The relative changes in the magnetic field distribution over each isovolumetric interval are then calculated and processed using statistical analysis methods (e.g., computing an arithmetic average, dividing the magnetic field map into sub-regions, extracting statistical eigenvectors). Finally, the feature vectors are analyzed by a statistical learning method (such as principal component analysis), so that the consistency of the battery pack is estimated and the abnormal-performance battery is positioned. In the prior art, the core of the technology and the system for detecting the performance consistency of the lithium battery pack based on in-situ magnetic field sensing is to analyze unbalanced current variation in the battery pack by sensing the variation of the external space induction magnetic field of the battery pack, so as to evaluate the performance consistency of the battery pack and accurately position the abnormal battery. Compared with the traditional electric parameter monitoring, the two magnetic field detection schemes have the common point that nondestructive and non-contact detection is realized through the physical quantity of a magnetic field, and the problem of consistency of all single batteries in the battery pack in an on-line monitoring service state is solved. Although the prior art solutions based on magnetic field imaging described above have advanced in terms of non-invasive detection, they still present the following significant technical drawbacks compared to the present proposals, which are also the core technical problems the present application aims to solve: 1. The adaptability to dynamic working conditions is poor, and the detection robustness is insufficient, namely, the scheme represented by a 'magnetic field imaging detection method based on statistical analysis' is adopted, and the magnetic field data acquisition depends on the precondition of 'equal discharge capacity interval'. This means that the method is only applicable to constant current or slow changing conditions. In practical application, under dynamic loads such as electric vehicles or energy storage system frequency modulation, the current and power of the battery pack change in real time, and the measurement requirement of the equal discharge capacity interval cannot be met, so that the method is invalid or the accuracy is seriously reduced. Therefore, the application proposes to solve the technical problem that the battery pack consistency evaluation model can still maintain high robustness under complex dynamic working conditions. 2. The feature extraction mode is single and insensitive to early weak abnormality, and the existing magnetic field detection scheme is mostly dependent on a statistical analysis method or basic feature engineering and cannot deeply excavate rich spatial features contained in a magnetic field image. They often treat the magnetic field image as a common optical image process, lacking a mechanism to purposefully enhance subtle global distribution pattern variations and local magnetic field distortion characteristics caused by abnormal cells. The homogeneous characteristic extraction mode leads to insensitivity to early and weak inconsistent signals generated in the early service stage or when the performance of the battery pack slightly declines, and has limited early warning capability. Therefore, the application proposes a technical problem of how to extract multi-scale visual features more fully and intelligently from a magnetic field image so as to improve the detection se