Lithium battery data reconstruction and classification method based on adversarial learning driven feature distribution alignment

Abstract

A lithium battery data reconstruction and classification method driven by adversarial learning and feature distribution alignment, relating to the field of lithium battery state monitoring technology, mainly includes the following steps: constructing a feature recognizer to capture the temporal correlation features of lithium battery data throughout its entire life cycle; integrating the feature recognizer into an adversarial training framework, aligning the features of the data generated by the feature extractor with those of the real data through adversarial learning between real lithium battery data and simulated data generated by the feature extractor; constructing a feature transfer-based data reconstruction network based on the feature extractor and combining it with the Unet network to reconstruct battery data consistent with the features of the real data, forming an expanded dataset with the original data; and building a battery state classification network based on the feature recognizer, training it with the expanded dataset to complete the state classification of lithium batteries. This method ensures the effectiveness of the reconstructed data through feature distribution alignment, significantly improving the generalization ability and recognition accuracy of the classification model.

Description

Technical Field

[0001] This invention relates to the field of lithium battery state monitoring technology, and in particular to a lithium battery data reconstruction and classification method based on adversarial learning-driven feature distribution alignment. Background Technology

[0002] Monitoring the entire lifecycle status of lithium batteries is a core aspect of ensuring safe equipment operation and extending the lifespan of lithium batteries. Current mainstream lithium battery status classification methods rely on the support of a large amount of full lifecycle data. By constructing classification models, accurate judgment of the three-level status of lithium batteries—"normal state - critical failure - failure"—can be achieved.

[0003] However, in practical applications, lithium battery lifecycle data suffers from significant scarcity: on the one hand, the lithium battery lifecycle is long, and collecting complete data on a single lithium battery from activation to failure is extremely costly; on the other hand, lithium battery data at critical failure states consists of small sample data, which is random and sudden, making large-scale acquisition difficult. This data scarcity directly leads to insufficient training of classification models, resulting in overfitting, weak generalization ability, and other problems, severely affecting the accuracy of lithium battery state classification and failing to meet practical application requirements.

[0004] To address the problem of data scarcity, traditional data augmentation methods, such as random pruning, noise addition, and time series rearrangement, are essentially superficial transformations of existing data. They cannot generate effective data with new feature distributions. For example, in scenarios where lithium battery failure state data is scarce, simply adding noise to normal state data cannot simulate the core features of voltage drop and abnormal temperature rise during critical failure. The augmentation effect is limited, and it cannot solve the problems of model overfitting and weak generalization ability.

[0005] More importantly, existing technologies have not established a collaborative mechanism for "feature alignment - data generation - state classification". Feature extraction, data generation and classification networks are mostly designed independently, resulting in poor feature reusability. Each module needs to be trained from scratch, which not only leads to long training cycles and high resource consumption, but also the generated simulated data may have distribution bias, reducing the accuracy of the classification model. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, such as limited lithium battery data, unreasonable data expansion strategies, and low model training efficiency, this disclosure provides a lithium battery data reconstruction and classification method driven by adversarial learning and feature distribution alignment. By building a feature extractor and aligning it with the real data distribution through adversarial training, and then integrating it into the Unet data reconstruction model and classification network respectively through weight transfer, the method achieves full-process feature reuse and collaborative optimization of feature extraction, data reconstruction, and state classification, thereby improving the quality of data generation and classification accuracy, and solving the problem of accurate classification of lithium battery states under data scarcity.

[0007] This method first uses adversarial learning to enable the feature extractor to accurately grasp the characteristic patterns of real lithium battery data, thus fundamentally solving the problem of distribution deviation between generated data and real data. Based on this feature extractor, a data reconstruction network is constructed by integrating the Unet network. By leveraging the spatial feature mining capability of the U-shaped structure and the temporal feature capture capability of the feature extractor, effective samples that are highly consistent with the features of real data are generated, thus addressing the problem of data scarcity in a targeted manner. Finally, the weights of the pre-trained feature extractor are transferred to the classification network to achieve full-process feature reuse, thereby improving training efficiency and classification accuracy.

[0008] This technical approach aims to systematically address the core shortcomings of existing technologies, such as "insufficient feature capture, low data reconstruction quality, and poor model synergy," ultimately achieving accurate classification of lithium battery states in small-sample scenarios.

[0009] The specific steps mainly include: S1, acquire the actual operating data of the lithium-ion battery as the raw data; S2, Construct a feature recognizer to capture the time-series correlation features of lithium battery data throughout its entire life cycle; S3, Construct a feature alignment network based on adversarial learning, integrate the feature recognizer into the adversarial training framework, and enable the feature recognizer to master the feature distribution pattern of the real data through adversarial learning between real lithium battery data and simulated data generated by the feature recognizer. S4, based on the feature recognizer trained by adversarial learning, combined with the Unet network, constructs a feature reconstruction network based on feature transfer. The Unet structure is responsible for mining the coupling relationship between different independent variables of battery data, and the feature recognizer is responsible for deepening the mining of the temporal dependency relationship of battery data. The combination of the two realizes feature fusion enhancement and reconstructs new lithium battery data consistent with the features of real data. The new lithium battery data and the original data form an augmented dataset to realize data enhancement. S5 constructs a battery state classification network based on feature transfer using a feature recognizer, trains the classification network using an extended dataset, and completes the state classification of lithium batteries.

[0010] Furthermore, the method for constructing the feature recognizer in step S2 includes: Starting with the input data, the distribution of the input data is first standardized through a normalization layer to improve training stability; The normalized output is fed into two parallel linear layers for dimensionality transformation. The output of the linear layer in one branch is first processed by a 1×1 convolution to achieve lightweight adjustment of the feature dimensions, and then passed through the normalization layer and the SiLU activation function for non-linear transformation. The output of the linear layer in the other branch is directly processed by the SiLU activation function. Afterwards, the SiLU outputs from these two branches are first added together, and the result is simultaneously fed into two parallel branches. In one branch, the SSM module is used for modeling to achieve efficient feature modeling of long sequences, while in the other branch, the SiLU activation function is applied. The results from the two branches are then multiplied. Finally, the result of the multiplication operation is input into a linear layer, and after processing, the output data of the feature extractor is obtained.

[0011] Furthermore, the specific method of step S3 includes: Using a random vector as input, it is fed into the feature recognizer constructed in step S2 for processing, and the feature recognizer generates simulated data. The raw battery data, as the real data source, is input into the multilayer perceptron along with the data generated by the feature recognizer. The multilayer perceptron then judges the input data to distinguish whether it belongs to real battery data or simulated data generated by the feature recognizer. The discrimination results of the multilayer perceptron are fed back to the feature recognizer through an adversarial training mechanism. When the discriminator identifies that the features of the generated data do not match the real battery data, the feature recognizer adjusts its own feature mapping strategy based on the feedback to make the features of the generated data closer to the feature statistics of the real battery data. This enables the feature recognizer to continuously improve the feature alignment of the generated data, ultimately achieving alignment with the features of real battery data.

[0012] Furthermore, the specific method of step S4 includes: S41, Based on the pre-trained feature recognizer trained through adversarial learning in step S2 and having achieved alignment with real battery features, a bidirectional feature recognizer is constructed. The internal logic of the bidirectional feature recognizer is as follows: The input data is split into two parallel processing paths. One path is directly fed into a pre-trained feature recognizer to capture the forward sequence and feature dependencies. The other path first undergoes a sequence inversion operation and then is fed into the same pre-trained feature recognizer to capture the reverse sequence and feature dependencies. The two encoding results are fused to obtain an output that covers both forward and reverse feature information. S42, Based on the bidirectional feature recognizer, a feature reconstruction network based on feature transfer is constructed. The feature reconstruction network is built on the Unet network, and the overall model structure is U-shaped, specifically including: Based on raw battery data As input to the feature reconstruction network, a 3×3 convolutional layer with n input channels and 2n output channels is first input to complete the feature dimension expansion and preliminary spatial feature extraction of the original data; The output of the convolutional layer then enters the first-layer bidirectional feature recognizer for forward and reverse feature encoding. The output of the first-layer bidirectional feature recognizer is fused with the output of the second-layer bidirectional feature recognizer across layers. At the same time, the output of the third-layer bidirectional feature recognizer is passed to the input of the second-layer bidirectional feature recognizer, and fused again with the fusion results of the first and second-layer bidirectional feature recognizers to achieve complementarity of features at different levels. The input of the third-layer bidirectional feature recognizer comes from the output of the fourth-layer bidirectional feature recognizer, thus forming a cascaded structure of multiple bidirectional feature recognizers. The features, after multi-layer cascading and cross-layer fusion, are input into the fifth-layer bidirectional feature recognizer to complete the final feature integration. Its output is then fused with the output temporarily stored in the 3×3 convolutional layer to obtain reconstructed battery data that highly matches the features of the original battery data. ; S43, New Battery Data Compared with the original data Composition of extended dataset . Furthermore, in step S5, the structure of the constructed battery state classification network specifically includes: First, the input data is fed into the linear layer to adjust its dimensions, making it suitable for the input requirements of subsequent modules; The output of the linear layer enters the bidirectional feature recognition module derived from the transfer; The output of the bidirectional feature recognizer is fed into the feedforward neural network, which dynamically allocates the fusion weights of positive and negative features according to the effectiveness of the features, and integrates the positive and negative features to form a unified feature representation. Finally, the output of the feedforward neural network is fed into the softmax classification head, and the softmax function maps the features to the probability distributions of each category, ultimately outputting the corresponding classification results.

[0013] Furthermore, in step S5, the method for training the classification network using the augmented dataset includes: Bundle Divided into training set and verification set ; use As a training set, the parameters of the bidirectional feature recognizer are frozen, and only the linear layers of the classification network and the feedforward neural network are trained. Then freeze the parameters of the classification network, and... Input the classification network and validate the model's performance.

[0014] Compared with the prior art, the beneficial effects of this disclosure are: ① Strong feature reusability and high training efficiency: After the feature extractor is aligned with the real data distribution through adversarial training, its weights are transferred to the data reconstruction network and the classification network, avoiding the model from training from scratch, reducing the need for training data, and improving the convergence speed of the data reconstruction model and the classification model. ② Higher data reconstruction quality: The spatial correlation of data captured by the Unet network complements the temporal features aligned by the feature extractor. The distribution similarity between the generated data and the real data after fusion is significantly higher than that of traditional generative models. ③ High classification accuracy and generalization: The classification network is built based on a feature extractor that is aligned with the true distribution. Combined with the expanded and diversified data, the classification accuracy of lithium battery status can be effectively improved, and the generalization error on different types of lithium battery data is reduced. Attached Figure Description

[0015] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0016] Figure 1 Here is a flowchart of the lithium battery state time series prediction method based on a dual-domain adversarial feature distillation transfer network according to this disclosure; Figure 2 This is a diagram of the feature recognizer model structure. Figure 3 The network structure diagram is aligned with the adversarial features; Figure 4 Diagram of the feature reconstruction network and bidirectional feature recognizer structure; Figure 5 This is a diagram of the classification network structure. Detailed Implementation

[0017] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0018] This disclosure proposes an adversarial learning-driven feature distribution alignment method for lithium battery data reconstruction and classification, aiming to address the problem of insufficient battery state classification accuracy caused by the scarcity of full-lifecycle data for lithium batteries. It constructs a complete solution for battery data "adversarial alignment - feature extraction - data reconstruction - feature transfer - classification inference," with the core process including five key steps: "construction of a feature extractor based on a state-space model - adversarial training to align data distribution - data reconstruction using a feature extractor based on UNet and feature transfer - training of a feature transfer classification network - state inference." First, a feature recognizer is constructed, whose function is to capture the temporal correlation features of lithium battery data throughout its entire life cycle; The feature recognizer is integrated into an adversarial training framework. Through adversarial learning between real lithium battery data and simulated data generated by the feature extractor, the feature extractor can master the feature distribution patterns of the real data. Next, based on the feature extractor and combined with the Unet network, a data reconstruction network based on feature transfer is constructed. The Unet structure is responsible for mining the coupling relationship between different independent variables of the battery data, while the feature extractor is responsible for deepening the mining of the temporal dependency relationship of the battery data. The two are combined to achieve feature fusion enhancement and reconstruct new lithium battery data consistent with the features of real data. The new lithium battery data and the original data form an expanded dataset, realizing data enhancement. Finally, a battery state classification network was built based on the feature recognizer, and the classification network was trained using an extended dataset to achieve accurate classification of lithium batteries into "normal state - critical failure - failure".

[0019] In one exemplary implementation: According to one embodiment of this disclosure, the overall flow is as follows: Figure 1 As shown, the original data Starting with feature alignment, a feature recognizer is first trained; then, through feature transfer, the feature recognizer is introduced into a feature reconstruction network based on feature transfer to obtain reconstructed data. Next, Compared with the original data Data is concatenated according to data dimensions to form training data. Then, through feature transfer, the feature recognizer is introduced into the feature transfer-based classification network, and... The training set was divided into two parts at an 8:2 ratio. and verification set , Used to train a classification network. This method was used to verify its performance. The entire process relies on multiple transfers and reuses of the feature recognizer, realizing a complete task from data feature alignment and reconstruction to classification. This expands the training data and improves the efficiency of training the model, achieving high-precision classification of lithium battery states with small sample sizes.

[0020] The detailed steps are explained below: 1. Battery data collection.

[0021] Operating data for a specific lithium-ion battery was collected using sensors and related equipment. Each sample contained 50 parameters, including voltage, current, temperature, and other operational characteristics, for a total of n samples. This dataset was preprocessed and manually labeled, serving as the raw data. .

[0022] 2. Construct a feature recognizer.

[0023] Feature recognizer structure as follows Figure 2 As shown, the core feature processing network in the entire process follows this process: Starting with the input data, the distribution of the input data is first standardized through a normalization layer to improve training stability. The normalized output is fed into two parallel linear layers for dimensionality transformation. The output of the linear layer of one branch is first processed by a 1×1 convolution to achieve lightweight adjustment of the feature dimension, and then passed through the normalization layer and the SiLU activation function for non-linear transformation. The output of the linear layer of the other branch is directly processed by the SiLU activation function. Afterwards, the SiLU outputs of these two branches are first added together, and the result is fed into two parallel branches. One branch enters the SSM module (state-space model) to achieve efficient feature modeling of long sequences, and the other branch is processed by SiLU. The outputs of the two branches are then multiplied. Finally, the result of the multiplication operation is input into a linear layer, and after processing, the output data of the feature extractor is obtained.

[0024] 3. Construct a feature alignment network based on adversarial learning.

[0025] The network structure diagram is as follows: Figure 3 As shown. First, a random vector is taken as input and fed into a feature recognizer for processing, which simulates the generated data. This is the initial step of feature alignment: allowing the generated data to be anchored to the distribution range of real battery data in terms of feature dimensions.

[0026] Raw battery data As the source of real data, it is input into the multilayer perceptron along with the generated data mentioned above. The core task of the multilayer perceptron is to discriminate the input data, distinguishing whether it belongs to real battery data or simulated data generated by the feature recognizer. This discrimination process essentially quantifies the feature differences between the generated data and real battery data. The greater the difference, the easier it is for the discriminator to distinguish between the two, which also means that the degree of feature alignment is lower.

[0027] Finally, the discrimination results of the multilayer perceptron are fed back to the feature recognizer through an adversarial training mechanism. When the discriminator identifies that the features of the generated data do not match the real battery data, the feature recognizer will adjust its feature mapping strategy based on this feedback, optimize the underlying features of the generated data, and make them closer to the feature statistics of the real battery data. Meanwhile, the discriminator will also optimize its own discrimination ability, forcing the feature recognizer to continuously improve the feature alignment of the generated data, so that the generated data produced is increasingly close to the real battery data at the feature level, and finally achieve the goal of high alignment of battery data features.

[0028] 4. Construct a feature reconstruction network based on feature transfer.

[0029] As shown in Figure 4, this feature reconstruction network based on feature transfer uses the original battery data. As input, its core bidirectional feature recognizer is built upon a feature recognizer that has been previously trained through adversarial learning and has achieved alignment with real battery features. The internal logic of this bidirectional feature recognizer is as follows: The input data is split into two parallel processing paths. One path is directly fed into a pre-trained feature recognizer to capture the forward sequence and feature dependencies. The other path first undergoes a sequence inversion operation before being input into the same pre-trained feature recognizer to capture the reverse sequence and feature dependencies. This bidirectional modeling approach enables bidirectional scanning of the data sequence, fully capturing the forward and backward feature dependencies of the data, avoiding information omissions during unidirectional processing. The two encoding results are fused to obtain an output that covers both forward and reverse feature information, providing more comprehensive feature support for subsequent data reconstruction and thus improving reconstruction accuracy.

[0030] Raw battery data As input to the feature reconstruction network, a 3×3 convolutional layer with n input channels and 2n output channels is first input to complete the feature dimension expansion and preliminary spatial feature extraction of the original data. The output of the convolutional layer then enters the first-layer bidirectional feature recognizer for forward and reverse feature encoding through fusion. The output of the first-layer bidirectional feature recognizer is fused with the output of the second-layer bidirectional feature recognizer across layers. Simultaneously, the output of the third-layer bidirectional feature recognizer is passed to the input of the second-layer bidirectional feature recognizer, where it is fused again with the former, achieving complementarity of features at different levels. The input of the third-layer bidirectional feature recognizer comes from the output of the fourth-layer bidirectional feature recognizer, thus forming a cascaded structure of multiple bidirectional feature recognizers. In this process, each layer of bidirectional feature recognizer maintains the consistency between features and the distribution of real battery data by relying on pre-trained feature alignment capabilities, and also enhances the comprehensiveness of feature representation by allowing forward and reverse features at multiple scales to complement each other through cross-layer fusion.

[0031] Next, the features obtained after multi-layer cascading and cross-layer fusion are input into the last bidirectional feature recognizer to complete the final feature integration. Its output is then fused with the output temporarily stored in the 3×3 convolutional layer to obtain reconstructed battery data that highly matches the features of the original battery data. .

[0032] The entire process leverages the bidirectional modeling capabilities of the transfer bidirectional feature recognizer and the multi-scale feature extraction capabilities of Unet, combined with feature complementarity through multi-layer cascading and cross-layer fusion, to achieve higher-precision battery data reconstruction.

[0033] 5. Build a classification network based on feature transfer.

[0034] The network structure is shown in Figure 5. First, the input data is fed into a linear layer to adjust its dimensions, making it suitable for the input requirements of subsequent modules. The output of the linear layer then enters the bidirectional feature recognition module, whose specific structure is described in Structure 4. This module uses the pre-trained feature recognizer that was previously transferred, and fully covers the forward and backward information of the data through bidirectional modeling. At the same time, it realizes the high reuse of the module, which not only improves the training efficiency of the model, but also enhances the accuracy of feature representation. The output of the bidirectional feature recognizer is fed into the feedforward neural network, which dynamically allocates fusion weights according to the effectiveness of the features, and efficiently integrates positive and negative features to form a unified feature representation. Finally, the output of the feedforward neural network is fed into the softmax classification head, and the softmax function maps the features to the probability distributions of each category, ultimately outputting the corresponding classification results.

[0035] When training the classification network, use As a training set, the parameters of the bidirectional feature recognizer are frozen, and only the linear layers and feedforward neural network of the classification network are trained, reducing the training difficulty of the network. Then, the parameters of the classification network are frozen, and... Input the classification network and validate the model's performance.

[0036] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are merely preferred and not restrictive.

Claims

1. A method for lithium battery data reconstruction and classification driven by adversarial learning and feature distribution alignment, characterized in that, Includes the following steps: S1, acquire the actual operating data of the lithium-ion battery as the raw data; S2, Construct a feature recognizer to capture the time-series correlation features of lithium battery data throughout its entire life cycle; S3, Construct a feature alignment network based on adversarial learning, integrate the feature recognizer into the adversarial training framework, and enable the feature recognizer to master the feature distribution pattern of the real data through adversarial learning between real lithium battery data and simulated data generated by the feature recognizer. S4, based on the feature recognizer trained by adversarial learning, combined with the Unet network, constructs a feature reconstruction network based on feature transfer. The Unet structure is responsible for mining the coupling relationship between different independent variables of battery data, and the feature recognizer is responsible for deepening the mining of the temporal dependency relationship of battery data. The combination of the two realizes feature fusion enhancement and reconstructs new lithium battery data consistent with the features of real data. The new lithium battery data and the original data form an augmented dataset to realize data enhancement. S5 constructs a battery state classification network based on feature transfer using a feature recognizer, trains the classification network using an extended dataset, and completes the state classification of lithium batteries.

2. The method according to claim 1, characterized in that, The method for constructing the feature recognizer in step S2 includes: Starting with the input data, the distribution of the input data is first standardized through a normalization layer to improve training stability; The normalized output is fed into two parallel linear layers for dimensionality transformation. The output of the linear layer in one branch is first processed by a 1×1 convolution to achieve lightweight adjustment of the feature dimensions, and then passed through the normalization layer and the SiLU activation function for non-linear transformation. The output of the linear layer in the other branch is directly processed by the SiLU activation function. Afterwards, the SiLU outputs from these two branches are first added together, and the result is simultaneously fed into two parallel branches. In one branch, the SSM module is used for modeling to achieve efficient feature modeling of long sequences, while in the other branch, the SiLU activation function is applied. The results from the two branches are then multiplied. Finally, the result of the multiplication operation is input into a linear layer, and after processing, the output data of the feature extractor is obtained.

3. The method according to claim 1 or 2, characterized in that, The specific method of step S3 includes: Using a random vector as input, it is fed into the feature recognizer constructed in step S2 for processing, and the feature recognizer generates simulated data. The raw battery data, as the real data source, is input into the multilayer perceptron along with the data generated by the feature recognizer. The multilayer perceptron then judges the input data to distinguish whether it belongs to real battery data or simulated data generated by the feature recognizer. The discrimination results of the multilayer perceptron are fed back to the feature recognizer through an adversarial training mechanism. When the discriminator identifies that the features of the generated data do not match the real battery data, the feature recognizer adjusts its own feature mapping strategy based on the feedback to make the features of the generated data closer to the feature statistics of the real battery data. This enables the feature recognizer to continuously improve the feature alignment of the generated data, ultimately achieving alignment with the features of real battery data.

4. The method according to claim 1, characterized in that, The specific method of step S4 includes: S41, Based on the pre-trained feature recognizer trained through adversarial learning in step S2 and having achieved alignment with real battery features, a bidirectional feature recognizer is constructed. The internal logic of the bidirectional feature recognizer is as follows: The input data is split into two parallel processing paths. One path is directly fed into a pre-trained feature recognizer to capture the forward sequence and feature dependencies. The other path first undergoes a sequence inversion operation and then is fed into the same pre-trained feature recognizer to capture the reverse sequence and feature dependencies. The two encoding results are fused to obtain an output that covers both forward and reverse feature information. S42, Based on the bidirectional feature recognizer, a feature reconstruction network based on feature transfer is constructed. The feature reconstruction network is built on the Unet network, and the overall model structure is U-shaped, specifically including: Based on raw battery data As input to the feature reconstruction network, a 3×3 convolutional layer with n input channels and 2n output channels is first input to complete the feature dimension expansion and preliminary spatial feature extraction of the original data; The output of the convolutional layer then enters the first-layer bidirectional feature recognizer for forward and reverse feature encoding. The output of the first-layer bidirectional feature recognizer is fused with the output of the second-layer bidirectional feature recognizer across layers. At the same time, the output of the third-layer bidirectional feature recognizer is passed to the input of the second-layer bidirectional feature recognizer, and fused again with the fusion results of the first and second-layer bidirectional feature recognizers to achieve complementarity of features at different levels. The input of the third-layer bidirectional feature recognizer comes from the output of the fourth-layer bidirectional feature recognizer, thus forming a cascaded structure of multiple bidirectional feature recognizers. The features, after multi-layer cascading and cross-layer fusion, are input into the fifth-layer bidirectional feature recognizer to complete the final feature integration. Its output is then fused with the output temporarily stored in the 3×3 convolutional layer to obtain reconstructed battery data that highly matches the features of the original battery data. ; S43, New Battery Data Compared with the original data Composition of extended dataset .

5. The method according to claim 4, characterized in that, In step S5, the structure of the constructed battery state classification network specifically includes: First, the input data is fed into the linear layer to adjust its dimensions, making it suitable for the input requirements of subsequent modules; The output of the linear layer enters the bidirectional feature recognition module derived from the transfer; The output of the bidirectional feature recognizer is fed into the feedforward neural network, which dynamically allocates the fusion weights of positive and negative features according to the effectiveness of the features, and integrates the positive and negative features to form a unified feature representation. Finally, the output of the feedforward neural network is fed into the softmax classification head, and the softmax function maps the features to the probability distributions of each category, ultimately outputting the corresponding classification results.

6. The method according to claim 5, characterized in that, In step S5, the method for training the classification network using the augmented dataset includes: Bundle Divided into training set and verification set ; use As a training set, the parameters of the bidirectional feature recognizer are frozen, and only the linear layers of the classification network and the feedforward neural network are trained. Then freeze the parameters of the classification network, and... Input the classification network and validate the model's performance.

Images