BPST network-based echelon utilization battery fault diagnosis method
By combining the BPST network with a pseudo-label training strategy, the problems of speed and accuracy in power battery fault diagnosis are solved, and the generalization ability of the model under cross-operating conditions is improved, making it suitable for fault monitoring and safety early warning of batteries used in cascade applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU DIANZI UNIV
- Filing Date
- 2026-01-15
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies are insufficient for rapid and accurate fault diagnosis of power batteries, especially in terms of model generalization ability across batches and operating conditions, which cannot meet the needs of rapid and accurate fault diagnosis for batteries used in cascade applications.
A fault diagnosis method based on BPST network is adopted. By mining the temporal correlation within the channel and the feature correlation between the channels, and combining the pseudo-label training strategy, cross-operating condition transfer learning is achieved to improve the generalization ability of the model.
It enables rapid and accurate fault diagnosis of batteries used in cascade applications, is applicable to batteries in different aging states, and provides a reliable fault monitoring and safety early warning solution.
Smart Images
Figure CN121935698A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy battery cascade utilization technology, specifically to a fault diagnosis method for cascade utilization batteries based on BPST networks. Background Technology
[0002] The new energy vehicle industry is experiencing exponential growth, and the power batteries that support it, after phased development, are now approaching the window of opportunity for large-scale retirement. Even after reaching retirement standards (typically when rated capacity decays to below 80%), power batteries are still sufficient to support low-power applications such as energy storage, low-speed vehicles, and backup power, enabling the tiered utilization of retired batteries. This not only fulfills the battery's full life-cycle mission but is also one of the most economical ways to reduce carbon emissions and environmental pollution. However, the uncertainty of the operating environment and conditions of power batteries poses significant challenges to their safe operation.
[0003] Therefore, rapid and accurate fault diagnosis of power battery systems is both a practical necessity and a prerequisite and bottleneck for their large-scale reuse. Traditional fault diagnosis approaches mainly rely on electrochemical impedance spectroscopy (EIS) or long-term charge-discharge tests, using manual thresholds to judge anomalies such as capacity drops and sudden increases in internal resistance. While these methods have clear physical meanings, they are time-consuming and do not meet the requirements of speed and accuracy for field-level diagnostic strategies. In recent years, machine learning and deep learning have provided new ideas for fault diagnosis. Early research used shallow models such as support vector machines and random forests, which could provide fault diagnosis labels within tens of seconds, but their accuracy was limited and their practical feasibility was not high. Subsequent deep networks (CNN, LSTM, Transformer) can significantly improve recognition accuracy, but the models perform poorly on single-channel signals and their generalization ability decreases significantly when transferring between batches and operating conditions. Summary of the Invention
[0004] The purpose of this invention is to provide a fault diagnosis method for cascaded batteries based on BPST networks. By simultaneously mining intra-channel temporal correlations and inter-channel feature correlations, the method compresses the sequence length while retaining local dynamic features, thus solving the problem of weak model performance on single-channel signals. Furthermore, a pseudo-label training strategy is introduced to fine-tune the model across domains, improving its generalization ability across batches and operating conditions. This provides technical support for rapid and accurate fault diagnosis of cascaded batteries, thereby addressing the problems mentioned in the background.
[0005] The present invention provides the following specific technical solution: a method for fault diagnosis of secondary-use batteries based on BPST networks, comprising the following operational steps: Step S1: Using the secondary-use batteries as the data collection object, collect battery fault data under multiple operating conditions to form a multi-operating-condition dataset.
[0006] Preferably, the collected data covers three classic battery operating conditions, each of which includes five actual operating states, specifically including normal state, cell inconsistency, internal short circuit, sensor noise, and sensor sticky fault. Simultaneously, 12 characteristic operating data such as voltage, current, and power are collected under the five battery operating states, and corresponding fault labels are matched to the collected data to form a multi-operating-condition dataset.
[0007] Step S2: Perform systematic data preprocessing on the multi-condition dataset and output a two-dimensional fusion feature vector.
[0008] Preferably, a bidirectional block package is constructed by performing a binary split of the multi-condition dataset in terms of time dimension and feature dimension; then, a linear transformation operation is performed on each bidirectional block package to map it into a one-dimensional feature vector; and the obtained one-dimensional feature vector is concatenated column by column to obtain a two-dimensional fused feature vector.
[0009] Step S3: Embed the obtained two-dimensional fusion feature vector The corresponding positional codes are matched and input into the BPST network model. The BPST network model extracts features based on a sparse attention mechanism, and the extracted output contains... The input is fed into the classifier to determine the fault type.
[0010] Preferably, to achieve global summarization of the entire diagnostic information, the two-dimensional fusion feature vector is embedded. After matching the corresponding positional encoding, it is input into the BPST network model, and subsequently the output contains... Classification can be completed by directly connecting to a classifier. The BPST network model contains n layers. Each floor All adopt a pre-layer normalization architecture, and the process is as follows: embedding The resulting two-dimensional fused feature vectors undergo linear transformation and After normalization, sparse multi-head self-attention calculation based on Top-K mask is performed, followed by residual connection and... After normalization, the output is a uniform feature vector through residual connections. Finally, it passes through a Dropout regularization layer to complete a single layer. Joint refinement of "global-local" features. Subsequently, n layers will be performed. The processed output is used to extract nonlinear features through a two-layer MLP dimensionality increase-decrease process, yielding the output of the entire BPST network. Finally, global information is extracted from this output. The input is used to classify the fault type and obtain the predicted label.
[0011] Preferably, the specific implementation includes: Step S31: Initialize a trainable embedding at the beginning position of the two-dimensional fused feature vector sequence. , to obtain embedding The subsequent two-dimensional feature vector representation; Step S32: For trainability Match the corresponding position code to extract absolute and relative position information; Step S33: Use multiple layers of stacked Transformer Blocks for feature extraction and representation, and add a Layer Norm layer within each Transformer Block; Step S34: In the BPST network model, an architecture based on sparse attention mechanism is adopted to calculate the query matrix, key matrix and value matrix of the input vector respectively.
[0012] Step S4: Optimize the model using transfer learning methods and introduce a pseudo-label training strategy to perform unsupervised cross-working condition transfer learning.
[0013] Preferably, the pseudo-label training strategy includes: utilizing unlabeled data by generating pseudo-labels for the target domain data, generating pseudo-labels based on the prediction results of the BPST network model for the target domain, and filtering high-confidence prediction results as pseudo-labels through a confidence threshold.
[0014] Preferably, the implementation of the pseudo-label training strategy includes: Step S41: Convert the output of the BPST network model for the target domain data into a probability distribution using the softmax function; Step S42: For each target domain sample, retain its maximum prediction probability and its corresponding prediction label; Step S43: Filter the maximum predicted probability of each sample. If it is higher than the set confidence threshold, then use the predicted label of the sample as the pseudo label of the sample.
[0015] Step S5: Compare the predicted labels output by the BPST network model with the true labels, and calculate the precision, recall, accuracy, and F1 score to verify the performance of the BPST network model.
[0016] Preferably, a confusion matrix is composed of true positive examples, false positive examples, true negative examples, and false negative examples; precision, recall, accuracy, and F1 score are calculated based on the above examples; and the performance of the BPST network model is validated by inputting source domain and target domain datasets.
[0017] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention addresses the problem of fault diagnosis in cascade-use batteries by collecting multi-condition fault data, including voltage and current data, and matching them with corresponding fault labels to form a multi-condition dataset, laying a high-quality sample foundation for data-driven modeling. Based on a bidirectional block strategy, the original multi-condition dataset is transformed into bidirectional block packages, and after linear transformation and vector concatenation, two-dimensional fused feature data is obtained, preserving temporal correlations while mining feature correlations; based on a linear embedding strategy, by embedding into the two-dimensional fused feature vector... The system matches positional codes to summarize global diagnostic information, which is then fed into the BPST network. Each of the n TransformerBlock layers in the BPST network employs a pre-level normalization architecture to achieve joint refinement of global and local features, extracting the final output. The input is fed into a classifier for fault identification. Utilizing unsupervised transfer learning, the BPST network model is fine-tuned and optimized using multi-condition data from the target domain. This effectively improves the model's cross-condition transfer capability without relying on the true labels in the target domain. This method is applicable to batteries in different aging states for secondary use, providing a reliable and scalable technical solution for battery screening, fault monitoring, and safety early warning, and possesses strong practical application value. Attached Figure Description
[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the operation logic of the cascaded battery fault diagnosis method based on BPST network provided in the embodiments of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on 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.
[0020] Combination Figure 1 As shown, the present invention provides a technical solution: a method for fault diagnosis of secondary-use batteries based on BPST networks, comprising the following operational steps: Step S1: Using the secondary-use batteries as the data acquisition target, collect battery fault data under multiple operating conditions.
[0021] In this embodiment, the data collected by the present invention covers three classic battery operating conditions. Each type of operating condition includes five actual operating states, specifically including normal state, cell inconsistency, internal short circuit, sensor noise, and sensor stickiness fault. Twelve characteristic operating data such as voltage, current, and power are collected simultaneously under the five battery operating states, and corresponding fault labels are matched to the collected data to form a multi-operating condition dataset.
[0022] Step S2: Based on the collected multi-condition dataset, perform systematic data preprocessing operations.
[0023] In this embodiment, a bidirectional segmented package is constructed by performing a binary split on the multi-condition dataset using both time and feature dimensions. Then, a linear transformation operation is performed on each bidirectional segmented package to map it into a one-dimensional feature vector. Finally, the resulting one-dimensional feature vectors are concatenated column-wise to obtain a two-dimensional fused feature vector. The specific implementation process includes: For example, the multi-channel signals contained in the multi-condition dataset are split in terms of time and feature channels. The multi-channel signals are presented in a two-dimensional matrix structure, where the matrix row width is... Represents the characteristic number, column length To represent time points and preserve the temporal and feature correlations of characteristic data, multi-channel signals are... By splitting the data along the time and feature channels, we obtain: Among them, a total Each of the 2-D bidirectional packet segments is represented as: ,in, It is the number of lines in each binary block. It is the number of columns in each binary block packet.
[0024] For example, based on the obtained bidirectional block packet, it is mapped to a low-dimensional space by a linear transformation operation to obtain a one-dimensional feature vector. Then, the one-dimensional feature vector is concatenated column-wise to obtain a two-dimensional fused feature vector. The specific process is as follows: ,in , The one-dimensional feature sequence obtained by dimensionality reduction for each bidirectional block packet has a length of... The one-dimensional feature vector is obtained by concatenating it column by column: ,in It is a two-dimensional fused feature vector.
[0025] Step S3: Embed the obtained two-dimensional fusion feature vector It then matches the corresponding positional encoding and inputs it into the BPST network model.
[0026] For example, to achieve global summarization of the entire diagnostic information, the two-dimensional fusion feature vector is embedded. After matching the corresponding positional encoding, it is input into the BPST network model, and subsequently the output contains... Classification can be completed by directly connecting to a classifier. The BPST network model contains n layers. Each floor All adopt a pre-layer normalization architecture, and the process is as follows: embedding The resulting two-dimensional fused feature vectors undergo linear transformation and After normalization, sparse multi-head self-attention calculation based on Top-K mask is performed, followed by residual connection and... After normalization, the output is a uniform feature vector through residual connections. Finally, it passes through a Dropout regularization layer to complete a single layer. Joint refinement of "global-local" features. Subsequently, n layers will be performed. The processed output is used to extract nonlinear features through a two-layer MLP dimensionality increase-decrease process, yielding the output of the entire BPST network. Finally, global information is extracted from this output. The input is used to classify the fault type and obtain the predicted label.
[0027] For example, the specific implementation process includes: Step S31: Initialize a trainable embedding at the beginning position of the two-dimensional fused feature vector sequence. , to obtain embedding The resulting two-dimensional feature vector representation.
[0028] For example, embedding The resulting two-dimensional feature vector is represented as: Embedded trainable Subsequently, it is used as the feature output of the BPST network model and directly used for classification.
[0029] Step S32: For trainability Match the corresponding position code to extract absolute and relative position information.
[0030] For example, the specific representation of the two-dimensional feature vector after matching the corresponding position encoding is as follows: ,in, , is a randomly initialized learnable matrix that can encode positional information within and between tokens, compensating for the model's inability to perceive element positional relationships during parallel data processing.
[0031] Step S33: Use multi-layer Transformer Block stacking for feature extraction and representation. Add a Layer Norm layer within each Transformer Block to reduce model complexity and prevent overfitting.
[0032] For example, LN is calculated by the following formula: in, The mean of the feature data. For the variance of the feature data, The value is , These are normalized feature data. and For learnable parameters, For addition and bitwise operations, the effect of multiplication can be approximated, thereby reducing computational complexity.
[0033] Step S34: Using an architecture based on sparse attention mechanism, calculate the query matrix of the input vector. Key matrix Sum matrix .
[0034] For example, in the BPST network model, an architecture based on a sparse attention mechanism is adopted to achieve efficient feature interaction and information extraction. To achieve this goal, the query matrix of the input vector is calculated separately. Key matrix Sum matrix : Wherein, the learning parameters are , These are the dimensions of three matrices. Next, local sparsification is applied to the query matrix and key matrix, and sparse attention scores are calculated. : Calculate sparse attention weights : Finally, the weighted value is calculated to obtain the output: in, These are the query matrix, key matrix, and value matrix after local sparsification, respectively. Key matrix Feature dimensions, As a sparsity factor, it is used to maintain numerical stability. This is the output result.
[0035] Step S4: Use pseudo-labels to optimize the model. Based on the BPST network model, fine-tune the model using target domain data under different working conditions. Introduce a pseudo-label training model fine-tuning strategy to perform unsupervised cross-working-condition transfer learning.
[0036] In this embodiment, to improve the generalization performance of the model in cross-batch and cross-condition transfer tasks, the present invention adopts a pseudo-label method to optimize the model: based on the obtained BPST network model, the model is fine-tuned using target domain data of different conditions, and a pseudo-label training model fine-tuning strategy is introduced, which does not need to rely on the real labels of the target domain throughout the process, thereby realizing unsupervised cross-condition transfer learning.
[0037] For example, the pseudo-labeling method utilizes unlabeled data by generating pseudo-labels for the target domain data, thereby improving the model's cross-domain accuracy. Pseudo-label generation is based on the model's predicted probability distribution of samples in the target domain; the category with the highest predicted probability is selected as the predicted label. Then, only samples with predicted probabilities higher than a set confidence threshold are retained, and their predicted labels are adopted as the final pseudo-labels. The specific implementation process includes: Step S41: The output is converted into a probability distribution by the softmax function. .
[0038] in, It is the amount of data in the source domain. It is the prediction result of the target domain data.
[0039] Step S42: For each target domain sample Retain its maximum prediction probability and their corresponding predicted labels : in, It is a sample The maximum predicted probability (confidence level). It is a sample Pseudo-tags.
[0040] Step S43: Filter the maximum predicted probability of each sample. If it is higher than the set confidence threshold, then use the predicted label of the sample as the pseudo label of the sample.
[0041] For example, the filtering expression is: in, The set confidence threshold.
[0042] Step S5: During the testing phase, the predicted labels are compared with the real labels to calculate the precision (Prec), recall, accuracy, and F1 score (F1) to verify the performance of the BPST network model. Based on the error feedback, the network parameters of the model are iteratively optimized to further improve the model's accuracy and cross-domain recognition capabilities.
[0043] In this embodiment, the specific implementation process includes: The confusion matrix consists of true positive examples (TP), false positive examples (FP), true negative examples (TN), and false negative examples (FN). TP represents the number of samples correctly predicted as positive by the model, FP represents the number of samples incorrectly predicted as positive by the model, TN represents the number of samples correctly predicted as negative by the model, and FN represents the number of samples incorrectly predicted as negative by the model.
[0044] Performance Indicator Calculation: Based on the above factors, the following performance indicators can be calculated using the following formulas: In this embodiment, the present invention verifies network performance by inputting source domain and target domain datasets, and the test results are shown in Table 1 below: Table 1
[0045] Among them, Inc (inconsistent state fault), ISC (internal short circuit fault), Noi (sensor noise fault), Nor (normal state), and Sti (sensor viscosity fault) represent five battery operating states.
[0046] The performance validation results show the model's accuracy in identifying different fault types. The Sti class model performed best, with precision, recall, and F1 score all approaching or exceeding 0.98, demonstrating extremely high classification accuracy and stability. The Noi and Nor class models also performed relatively well, while the ISC class had the lowest scores across all metrics, indicating that the model's ability to identify faults in this category needs improvement. Although the accuracy for all classes was high, approaching or reaching 0.938, accuracy does not fully reflect the model's performance on imbalanced datasets. Therefore, precision, recall, and F1 score provide a more comprehensive evaluation. These data provide clear directions for further model optimization, especially regarding the need for more attention and improvement in the lower-performing classes.
[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0048] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for fault diagnosis of secondary-use batteries based on BPST networks, characterized in that: The following steps are included: Step S1: Using the secondary-use batteries as the data collection object, collect battery fault data under multiple operating conditions to form a multi-operating-condition dataset; Step S2: Perform systematic data preprocessing on the multi-condition dataset and output a two-dimensional fusion feature vector; Step S3: Embed the obtained two-dimensional fusion feature vector The corresponding positional codes are matched and input into the BPST network model. The BPST network model extracts features based on a sparse attention mechanism, and the extracted output contains... The input is fed into a classifier to determine the fault type. Step S4: Optimize the model using transfer learning methods and introduce a pseudo-label training strategy to perform unsupervised cross-working condition transfer learning; Step S5: Compare the predicted labels output by the BPST network model with the true labels, and calculate the precision, recall, accuracy, and F1 score to verify the performance of the BPST network model.
2. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 1, characterized in that: The formation of the multi-condition dataset in step S1 includes: collecting battery fault data covering three classic battery conditions. Each condition includes five actual operating states: normal state, cell inconsistency, internal short circuit, sensor noise, and sensor sticky fault. Feature operating data of the five battery operating states are collected simultaneously, and corresponding fault labels are matched for the collected data.
3. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 2, characterized in that: The implementation of step S2 includes: The multi-condition dataset is split into two binary components based on feature dimension and time dimension to construct a bidirectional block package; Based on the obtained bidirectional block packet, it is mapped to a low-dimensional space by a linear transformation operation to obtain a one-dimensional feature vector. A two-dimensional fused feature vector is obtained by concatenating the one-dimensional feature vector column by column.
4. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 3, characterized in that: The implementation of step S3 includes: Step S31: Initialize a trainable embedding at the beginning position of the two-dimensional fused feature vector sequence. , to obtain embedding The subsequent two-dimensional feature vector representation; Step S32: For trainability Match the corresponding position code to extract absolute and relative position information; Step S33: Use multiple layers of stacked Transformer Blocks for feature extraction and representation, and add a Layer Norm layer within each Transformer Block; Step S34: In the BPST network model, an architecture based on sparse attention mechanism is adopted to calculate the query matrix, key matrix and value matrix of the input vector respectively.
5. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 4, characterized in that: The trainable embedded in step S31 As a feature output of the subsequent BPST network model, it is directly used for classification.
6. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 5, characterized in that: In step S34, after calculating the query matrix, key matrix, and value matrix of the input vector, the query matrix and key matrix are locally sparsified, the sparse attention score and sparse attention weight are calculated, and the weighted value is calculated to obtain the final output result.
7. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 6, characterized in that: The BPST network model in step S3 contains n layers. Each floor All adopt a pre-layer normalization architecture, and the operation process includes: Embedded The resulting two-dimensional fused feature vectors undergo linear transformation and After normalization, sparse multi-head self-attention computation based on Top-K mask is performed; via residual connection and After normalization, the output is a uniform feature vector through residual connections. Finally, it passes through a Dropout regularization layer to complete a single layer. "Global-local" feature joint refinement; Will go through n layers The processed output is used to extract nonlinear features through two layers of MLP dimensionality upscaling and downscaling to obtain the output of the entire BPST network; Extract summary global information from the output. The input is used to classify the fault type and obtain the predicted label.
8. The method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 7, characterized in that: The pseudo-label training strategy in step S4 includes: utilizing unlabeled data by generating pseudo-labels for the target domain data; generating pseudo-labels based on the prediction results of the target domain using the BPST network model; and filtering high-confidence prediction results as pseudo-labels using a confidence threshold.
9. A method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 8, characterized in that: The implementation of the pseudo-label training strategy includes: Step S41: Convert the output of the BPST network model for the target domain data into a probability distribution using the softmax function; Step S42: For each target domain sample, retain its maximum prediction probability and its corresponding prediction label; Step S43: Filter the maximum predicted probability of each sample. If it is higher than the set confidence threshold, then use the predicted label of the sample as the pseudo label of the sample.
10. A method for fault diagnosis of secondary-use batteries based on BPST networks according to claim 9, characterized in that: Step S5 also includes: The confusion matrix consists of true positive examples, false positive examples, true negative examples, and false negative examples; Calculate precision, recall, accuracy, and F1 score based on the above examples; The performance of the BPST network model was validated by inputting source and target domain datasets.