Lightweight battery residual life prediction method and system based on step-by-step knowledge distillation

Abstract

The invention relates to a lightweight battery residual life prediction method and system based on step-by-step knowledge distillation, and belongs to the field of battery energy technologies. According to the method, a teacher model for local-global degradation feature reconstruction and a student model carrying working condition information monitoring are constructed, a two-stage parameter freezing distillation strategy is adopted, potential features and reconstruction features are aligned firstly, and then a soft and hard loss optimization prediction module is fused, so that high-precision and lightweight prediction of the residual life of the battery is realized; and the adaptability and the calculation efficiency under different working conditions are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of battery energy technology and relates to a lightweight battery remaining life prediction method and system based on stepwise knowledge distillation. Background Technology

[0002] Developing the new energy vehicle industry, represented by electric vehicles, is crucial for addressing the global energy crisis, solving environmental pollution, and meeting travel demands. The power battery is the energy heart of an electric vehicle, directly determining its overall performance. Power batteries are essentially complex electrochemical systems that operate under extreme conditions of high and low temperatures, humidity, alternating heat, strong vibrations, and randomized operating environments. How to effectively capture the state characteristics of degradation modes and achieve accurate remaining life prediction using external data such as voltage, current, and temperature collected under actual operating conditions, rather than internal data from electrochemical reactions such as SEI film and electrochemical concentration polarization, is essential for the safe use and predictive maintenance of power batteries and is a pressing challenge for both academia and industry.

[0003] (1) Teacher models in the distillation process typically employ high-precision models to ensure the effectiveness of knowledge transfer. Existing high-precision RUL prediction models usually improve prediction accuracy by simply reconstructing battery degradation features or capturing the dependencies between local and global degradation features. However, simply reconstructing battery degradation features may ignore the complex nonlinear relationships in the degradation process, thus limiting the accuracy of the prediction model. At the same time, if the model relies too much on local features, it may ignore the overall degradation trend or overfit, making the model unable to accurately characterize the battery degradation state. Therefore, it is necessary to comprehensively consider local-global feature extraction and feature reconstruction within a unified framework to leverage their complementary advantages in the RUL prediction task.

[0004] (2) Student model training is usually based on the assumption that the training set and the test set have consistent distributions, but this premise is often difficult to meet in practical applications. For example, battery data under different operating conditions have significant temporal dynamics and distribution differences, which makes the training of student models without operating condition discrimination model architecture face adaptive challenges. Therefore, in order to cope with the diverse distribution of battery data under complex operating conditions, it is particularly necessary to introduce a student model with an operating condition discrimination mechanism.

[0005] (3) In the existing collaborative training framework of teacher and student models, the knowledge transfer efficiency is poor and the training hardware memory overhead is increased due to training methods such as direct distillation to reconstruct knowledge and global update of student model parameters. Therefore, it is necessary to distill key layers or important features to avoid redundant information transfer, while freezing most of the parameters of the student model and updating only specific layers to improve the overall efficiency of model distillation.

[0006] Therefore, there is an urgent need to design a new knowledge distillation model for the early remaining battery life to solve the above problems. Summary of the Invention

[0007] In view of this, the purpose of the present invention is to provide a lightweight battery remaining life prediction method and system based on stepwise knowledge distillation.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A lightweight battery remaining lifetime prediction method based on stepwise knowledge distillation includes the following steps: Acquire battery time-series data and preprocess the time-series data to obtain an input feature dataset containing the variance of the discharge voltage-capacity QV curve, the area of ​​the incremental capacity IC curve, and the difference between the maximum and minimum QV values. A teacher model and a student model are provided, wherein the teacher model is a Conditional Variational Autoencoder Mamba (CVAE-Mamba) model for local-global degradation feature reconstruction, and the student model is an Operating Condition-aware Conditional Variational Autoencoder (OC-CVAE) model carrying monitoring of operating condition information. A knowledge distillation framework with phased freezing of teacher-student model parameters is adopted. Based on freezing the teacher model parameters, the student model undergoes two-stage step-by-step knowledge distillation training. The two-stage step-by-step knowledge distillation training includes: In the first distillation stage, the parameters of the non-autoencoder part of the student model are frozen, and the autoencoder parameters of the student model are updated by minimizing the differences between the latent autoencoder features of the teacher model and the student model, the differences between their reconstructed features, and the reconstruction loss of the student model itself. In the second distillation stage, the autoencoder parameters of the student model are frozen, and the parameters of the non-autoencoder part of the student model are updated by minimizing the hard loss between the student model's prediction results and the true lifetime label, and the soft loss between the student model and the teacher model's predicted lifetime. The remaining battery life is predicted using the trained student model.

[0009] Furthermore, the teacher model includes: A Mamba module with local self-attention mechanism is used to learn local degradation features from the input feature dataset; A global self-attention correction module is used to dynamically adjust weights by calculating the relationships between local features in the input sequence in order to capture the dependencies between the local degenerate features. A Multi-Layer Perceptron (MLP) Conditional Autoencoder is used for reconstructing degradation features. This encodes the local and global degradation features into latent variables and reconstructs them. The reconstructed degradation features are then fed into a predictive Conditional Autoencoder (CVAE) for lifetime prediction.

[0010] Furthermore, the student model includes: A deep separable key feature extraction module with a dual-channel attention layer is used to enhance degraded key features; A condition monitor driven by unsupervised clustering and dual-path gating mechanism is used to extract and perceive feature information under different working conditions; A conditional autoencoder OC-CVAE composed of a one-dimensional convolutional neural network (1D-CNN) and a predictor with a two-channel attention mechanism are used to characterize and model the degradation process to achieve lifetime prediction.

[0011] Furthermore, the operating condition monitor achieves adaptive calculation of category distance and capture of common patterns for different operating condition modes by updating the weights of clustering gating and common pattern gating, and drives a multi-channel structure composed of depthwise separable convolutions through the synergistic effect of the two gating weights.

[0012] Furthermore, the training in the first distillation stage minimizes the loss function. To update the autoencoder parameters of the student model, where: ;in, The potential spatial alignment loss is calculated using the following formula: ,in For teachers' potential representation, As a potential variable for students, For the input sample, The total number of samples; The feature reconstruction loss is calculated using the following formula: ,in Reconstructing characteristics for teachers Reconstructing characteristics for students and The teacher model and the student model were respectively in the first... l Feature mapping of an intermediate layer; The self-supervised reconstruction loss is calculated using the following formula: ,in For the original input, and These are the encoder and decoder for the student model, respectively.

[0013] Furthermore, the training of the second distillation stage is achieved by minimizing the loss function. To update the parameters of the non-autoencoder part of the student model, where: ;in, Hyperparameters to balance hard and soft losses; For hard loss, its calculation formula is: ,in Labeled to true lifespan. For the prediction results of the student model, The total number of samples; As a soft loss function, the Kullback-Leibler Divergence (KLD) is used to measure the difference between the predicted distributions of teachers and students. Its calculation formula is as follows: ,in For the prediction results of the teacher model, For the prediction results of the student model, This refers to the distillation temperature.

[0014] A lightweight battery remaining lifetime prediction system based on stepwise knowledge distillation includes: The data processing module is used to acquire the time-series data of the battery and preprocess the time-series data to obtain an input feature dataset containing the variance of the discharge voltage-capacity QV curve, the area of ​​the incremental capacity IC curve, and the difference between the maximum and minimum QV values. The teacher model is a Conditional Variational Autoencoder Mamba (CVAE-Mamba) model that reconstructs local-global degenerate features. The student model is an operating condition-aware conditional variational autoencoder (OC-CVAE) model that carries monitoring of operating condition information. The knowledge distillation module employs a phased freezing framework for teacher-student model parameters. Based on the frozen teacher model parameters, it performs two-stage step-by-step knowledge distillation training on the student model. This two-stage step-by-step knowledge distillation training includes: in the first distillation stage, freezing the parameters of the non-autoencoder portion of the student model and updating the student model's autoencoder parameters by minimizing the differences between the latent autoencoder features of the teacher and student models, the differences between their reconstructed features, and the student model's own reconstruction loss; in the second distillation stage, freezing the student model's autoencoder parameters and updating the parameters of the non-autoencoder portion of the student model by minimizing the hard loss between the student model's prediction results and the true lifetime label, and the soft loss between the student model's predicted lifetime and the teacher model's predicted lifetime. The prediction module is used to predict the remaining battery life using the trained student model.

[0015] Furthermore, the teacher model includes: A Mamba module with local self-attention mechanism is used to learn local degradation features from the input feature dataset; A global self-attention correction module is used to dynamically adjust weights by calculating the relationships between local features in the input sequence in order to capture the dependencies between the local degenerate features. A Multi-Layer Perceptron (MLP) Conditional Autoencoder is used for reconstructing degradation features. This encodes the local and global degradation features into latent variables and reconstructs them. The reconstructed degradation features are then fed into a predictive Conditional Autoencoder (CVAE) for lifetime prediction.

[0016] Furthermore, the student model includes: A deep separable key feature extraction module with a dual-channel attention layer is used to enhance degraded key features; A condition monitor driven by unsupervised clustering and dual-path gating mechanism is used to extract and perceive feature information under different working conditions; A conditional autoencoder OC-CVAE composed of a one-dimensional convolutional neural network (1D-CNN) and a predictor with a two-channel attention mechanism are used to characterize and model the degradation process to achieve lifetime prediction.

[0017] Furthermore, the knowledge distillation module minimizes the loss function in the first distillation stage. To update the autoencoder parameters of the student model, and in the second distillation stage, by minimizing the loss function. To update the parameters of the non-autoencoder part of the student model, wherein, For potential spatial alignment loss, For feature reconstruction loss, For the loss of self-supervised reconstruction, For hard losses, For soft loss, This is a hyperparameter.

[0018] The beneficial effects of this invention are as follows: First, this invention significantly improves the accuracy and robustness of battery remaining life prediction. The teacher model employs a conditional autoencoder Mamba model that reconstructs local-global degradation features, effectively capturing the complex nonlinear relationships and dependencies in the battery degradation process, avoiding the overall trend or overfitting problems that past models might have overlooked. The student model, through a mechanism that carries monitoring of operating conditions, adaptively handles the differences in data distribution under different operating conditions, ensuring the stability of the prediction model in real-world scenarios.

[0019] Secondly, this invention optimizes the efficiency and scalability of the knowledge distillation process. It employs a framework of phased parameter freezing for the teacher-student model. In the first distillation stage, it focuses on aligning and reconstructing the autoencoder parameters, reducing redundant information transmission. In the second distillation stage, it integrates soft and hard losses to update the prediction module, thereby reducing training hardware memory overhead, accelerating model convergence, and making the entire distillation process more efficient and resource-friendly.

[0020] Furthermore, this invention enables lightweight deployment of the model, making it suitable for resource-constrained embedded devices. The student model significantly reduces the number of parameters and computational complexity while maintaining high predictive performance through depthwise separable convolutions and a dual-channel attention mechanism. This allows the system to be easily deployed on edge computing platforms such as the NVIDIA Jetson TX2, providing real-time, low-power remaining lifetime prediction capabilities for battery management systems, supporting predictive maintenance in applications such as electric vehicles.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 Overall layout of the distillation process; Figure 2Detailed diagram of the teacher model; Figure 3 Detailed images of the student model; Figure 4 A detailed diagram of the distillation process; Figure 5 The result diagram for lightweight deployment. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0025] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0026] Please see Figures 1-4 This invention provides a stepwise knowledge distillation method carrying operating condition information for predicting the remaining useful life (RUL) of lightweight batteries. This method is used to transfer knowledge between different network architectures for RUL prediction. The steps are as follows: 1) This invention uses the MIT dataset publicly available from Severson, which contains 124 APR18650M1A lithium iron phosphate battery samples (nominal capacity 1.1Ah, provided by A123 Systems). These batteries underwent cyclic testing in a 30°C constant-temperature chamber according to different fast-charge and constant-discharge curves. During the charge-discharge cycle, the batteries were first charged to 80% capacity using a one-step or two-step fast-charge strategy, and then charged to 100% using a 1C constant-current, constant-voltage mode. During discharge, a constant current of 4A was used until the voltage dropped to 2V. The dataset contains time-series data of various batteries during charge-discharge cycles, including voltage, current, temperature, etc., as well as the final cycle life label for each battery. The capacity during the discharge process and its corresponding voltage and cycle time were selected from 30 batteries (numbered #10-22, 25-40, 42) in the first battery set as benchmark data. The selected 30 batteries have complete charge-discharge cycle records, including samples with different number of cycles and degree of degradation, which can represent the performance changes of the batteries at different stages of their lifespan. Furthermore, by skipping adjacent cell numbers (such as #23, 24, 41, etc.), systematic biases introduced by differences in production batches or experimental conditions are reduced. The test set cells (numbers #43-45, 6-8) have no overlap with the training set, avoiding data leakage. Simultaneously, the test samples contain different cell number ranges and different lifespans; the small-scale test set allows for rapid verification of the model's predictive stability with limited data.

[0027] 2) Voltage curves are an effective indicator of battery capacity degradation. As batteries age, the discharge voltage-capacity curve (QV curve) gradually shifts to the left (i.e., at the same voltage, the discharge capacity decreases). Literature indicates that the QV curve changes more significantly in short-life batteries than in long-life batteries, demonstrating a strong correlation between the dynamic changes in the QV curve and battery life. By quantifying the fluctuation of the QV curve, the variance of the QV curve is obtained. Short-life batteries exhibit greater variation in the QV curve across different cycles (high variance), while long-life batteries show a more gradual variation (low variance). Calculating the derivative of capacity with respect to voltage reflects the phase transition process of the internal electrochemical reactions of the battery. The IC peak area of ​​aged batteries decreases, and the degree of IC peak area reduction varies among batteries of different lifespans during aging. The difference between the final capacity value and the initial capacity value in each cycle is a direct indicator of SOH (State of Health). Battery life termination is typically defined as capacity decay to 80% of the rated capacity, highlighting the crucial role of this characteristic in predicting RUL (Rated Usage Limit). Therefore, this paper uses the variance of the QV curve for each battery per cycle, the area enclosed by the IC curve, and the difference between the last capacity value and the first capacity value for each cycle as input features of battery degradation.

[0028] 3) Mamba combines the advantages of fast training of Transformer models and fast inference of CNN models, enabling the model to capture long-distance dependencies better while occupying less space.

[0029] In the Mamba (CVAE-Mamba) teacher prediction model of the conditional autoencoder for local-global degradation feature reconstruction, the extraction of original temporal features is achieved through the synergistic effect of three modules, the specific structure of which is as follows: Figure 1 As shown in the blue section, in the first Mamba module with local self-attention, the input sequence... ,in Indicates sequence length The feature dimension is represented by the input embedding layer and then processed by the local feature extraction module. and input self-attention layer Capture initial features ,in This represents the dimension of the input embedding layer. Then, a state-space model is used. Long-term dependencies in modeling In the second global self-attention correction module, Through self-attention mechanism Calculate the relationships between local features in the input sequence, capture the complex dependencies between degenerate local features, and obtain... In the third degenerate feature reconstruction MLP conditional autoencoder module, A conditional self-developed encoder for reconstructing and predicting tasks. The local and global degradation features are compressed and then restored layer by layer to obtain... Finally, a loss function carrying reconstruction error is used to minimize the difference between the reconstructed features and the true RUL, achieving high-precision prediction of the RUL of the power battery. The overall process can be summarized as follows:

[0030] 4) Depthwise Separable Convolution (DConvolution) significantly reduces the number of parameters and computational complexity of the model by decomposing the standard convolution into two steps: depthwise convolution (DC) and pointwise convolution (PC). In the Conditional Autoencoder Student Prediction Model (OC-CVAE) carrying monitoring of working condition information, the extraction of original temporal features is achieved through the synergistic effect of three modules, the specific structure of which is as follows: Figure 1 As shown in the green section, in the deep separable key feature extraction module of the first dual-channel attention layer, the input sequence... After transposition, the data is fed into a key feature extraction module consisting of two depthwise separable convolutional layers and a dual-channel attention layer. From In the condition monitoring module driven by the second unsupervised clustering and dual-path gating mechanism, key features Through two gating The synergistic effect of weights drives the multichannel structure composed of depthwise separable convolutions. get In the third conditional autoencoder module, which consists of 1D-CNN... In China, through strengthening Deep representation and contextual association modeling yielded Ultimately, a predictor using a dual-channel attention mechanism is used to predict the RUL (Relative Usage Limit) of the power battery. The overall process can be summarized as follows:

[0031] 5) The Teacher-Student Model Parameter Staged Freezing Knowledge Distillation Framework (OC-PKD) helps the student model, based on the frozen teacher model, to conduct two-stage staged knowledge distillation using the latent features, reconstructed features, and high-precision multi-step prediction results provided by the teacher model. The main knowledge distillation framework is as follows: Figure 1 As shown in the purple section, the framework design of the conditional autoencoder includes two sub-task modules with clear physical meaning and functional differences: feature reconstruction and prediction. A two-stage, step-by-step knowledge distillation is performed on each sub-module. The first stage is the distillation learning of the student autoencoder. In this stage, the parameters of the key feature extraction module, condition monitor, and lifetime predictor in the student model are frozen, minimizing the loss due to the difference in latent features between the teacher and student autoencoders. The difference loss between the reconstructed features of the two And the reconstruction loss between the reconstructed features and the original features of the student autoencoder. Update the student autoencoder:

[0032] Potential spatial alignment loss The main calculation of teacher latent representations and student latent variables Mean square error between:

[0033] Feature Reconstruction Loss The perceptron loss is mainly used to calculate the teacher's reconstructed features in multiple intermediate layers. Reconstructing features with students Between distance:

[0034] Self-supervised reconstruction loss Mainly used Norm measures the pixel-by-pixel difference between the student model's reconstructed output and the original input:

[0035] The second stage involves distillation learning of the non-autoencoder portion. In this stage, the autoencoder of the student model is frozen, and the soft loss between the lifespan predictions of the student and teacher models is used. Hard loss between student model predictions and actual lifespan labels To implement updates for non-autocode modules:

[0036] Hard label loss The distance between the true value and the student model's prediction is calculated using cross-entropy loss with label smoothing:

[0037] Soft target loss KL divergence is used to measure the difference between the predicted distributions of teachers and students, where Indicates the distillation temperature:

[0038] 6) In order to effectively verify and evaluate the method proposed in this invention, this invention has been compared with some benchmark methods, including standard TCN, LSTM, Transformer, Standard-KD, L1-KD, L2-KD, and KDnet, to verify the effectiveness of the proposed framework.

[0039] The trained student model was converted to ONNX format and deployed on NVIDIA Jetson TX2 hardware. Taking dataset Bs#43 as an example, its performance in terms of computational resources and inference time was compared. The comparison models included TCN, LSTM, Transformer, teacher, and student models. The results are as follows: Figure 5 As shown. The circle size represents the space occupied, labeled with computational resource usage and inference time on the TX2. These results highlight the effectiveness of the system's data processing and analysis, providing a solid foundation for further optimization of battery performance.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A lightweight battery remaining life prediction method based on stepwise knowledge distillation, characterized in that: Includes the following steps: Acquire battery time-series data and preprocess the time-series data to obtain an input feature dataset containing the variance of the discharge voltage-capacity QV curve, the area of ​​the incremental capacity IC curve, and the difference between the maximum and minimum QV values. A teacher model and a student model are provided, wherein the teacher model is a Conditional Variational Autoencoder Mamba (CVAE-Mamba) model for local-global degradation feature reconstruction, and the student model is an Operating Condition-aware Conditional Variational Autoencoder (OC-CVAE) model carrying monitoring of operating condition information. A knowledge distillation framework with phased freezing of teacher-student model parameters is adopted. Based on freezing the teacher model parameters, the student model undergoes two-stage step-by-step knowledge distillation training. The two-stage step-by-step knowledge distillation training includes: In the first distillation stage, the parameters of the non-autoencoder part of the student model are frozen, and the autoencoder parameters of the student model are updated by minimizing the differences between the latent autoencoder features of the teacher model and the student model, the differences between their reconstructed features, and the reconstruction loss of the student model itself. In the second distillation stage, the autoencoder parameters of the student model are frozen, and the parameters of the non-autoencoder part of the student model are updated by minimizing the hard loss between the student model's prediction results and the true lifetime label, and the soft loss between the student model and the teacher model's predicted lifetime. The remaining battery life is predicted using the trained student model.

2. The lightweight battery remaining life prediction method based on stepwise knowledge distillation according to claim 1, characterized in that: The teacher model includes: A Mamba module with local self-attention mechanism is used to learn local degradation features from the input feature dataset; A global self-attention correction module is used to dynamically adjust weights by calculating the relationships between local features in the input sequence in order to capture the dependencies between the local degenerate features. A Multi-Layer Perceptron (MLP) Conditional Autoencoder is used for reconstructing degradation features. This encodes the local and global degradation features into latent variables and reconstructs them. The reconstructed degradation features are then fed into a predictive Conditional Autoencoder (CVAE) for lifetime prediction.

3. The lightweight battery remaining life prediction method based on stepwise knowledge distillation according to claim 1, characterized in that: The student model includes: A deep separable key feature extraction module with a dual-channel attention layer is used to enhance degraded key features; A condition monitor driven by unsupervised clustering and dual-path gating mechanism is used to extract and perceive feature information under different working conditions; A conditional autoencoder OC-CVAE composed of a one-dimensional convolutional neural network (1D-CNN) and a predictor with a two-channel attention mechanism are used to characterize and model the degradation process to achieve lifetime prediction.

4. The lightweight battery remaining life prediction method based on stepwise knowledge distillation according to claim 3, characterized in that: The operating condition monitor achieves adaptive calculation of category distance and capture of common patterns for different operating condition modes by updating the weights of clustering gating and common pattern gating, and drives a multi-channel structure composed of depthwise separable convolutions through the synergistic effect of the two gating weights.

5. The lightweight battery remaining life prediction method based on stepwise knowledge distillation according to claim 1, characterized in that: The training in the first distillation stage minimizes the loss function. To update the autoencoder parameters of the student model, where: ;in, The potential spatial alignment loss is calculated using the following formula: ,in For teachers' potential representation, As a potential variable for students, For the input sample, The total number of samples; The feature reconstruction loss is calculated using the following formula: ,in Reconstructing characteristics for teachers Reconstructing characteristics for students and The teacher model and the student model were respectively in the first... l Feature mapping of an intermediate layer; The self-supervised reconstruction loss is calculated using the following formula: ,in For the original input, and These are the encoder and decoder for the student model, respectively.

6. The lightweight battery remaining life prediction method based on stepwise knowledge distillation according to claim 1, characterized in that: The training in the second distillation stage is achieved by minimizing the loss function. To update the parameters of the non-autoencoder part of the student model, where: ;in, Hyperparameters to balance hard and soft losses; For hard loss, its calculation formula is: ,in Labeled to true lifespan. For the prediction results of the student model, The total number of samples; As a soft loss function, the Kullback-Leibler Divergence (KLD) is used to measure the difference between the predicted distributions of teachers and students. Its calculation formula is as follows: ,in For the prediction results of the teacher model, For the prediction results of the student model, This refers to the distillation temperature.

7. A lightweight battery remaining life prediction system based on stepwise knowledge distillation, characterized in that: include: The data processing module is used to acquire the time-series data of the battery and preprocess the time-series data to obtain an input feature dataset containing the variance of the discharge voltage-capacity QV curve, the area of ​​the incremental capacity IC curve, and the difference between the maximum and minimum QV values. The teacher model is a Conditional Variational Autoencoder Mamba (CVAE-Mamba) model that reconstructs local-global degenerate features. The student model is an operating condition-aware conditional variational autoencoder (OC-CVAE) model that carries monitoring of operating condition information. The knowledge distillation module employs a phased freezing framework for teacher-student model parameters. Based on the frozen teacher model parameters, it performs two-stage step-by-step knowledge distillation training on the student model. This two-stage step-by-step knowledge distillation training includes: in the first distillation stage, freezing the parameters of the non-autoencoder portion of the student model and updating the student model's autoencoder parameters by minimizing the differences between the latent autoencoder features of the teacher and student models, the differences between their reconstructed features, and the student model's own reconstruction loss; in the second distillation stage, freezing the student model's autoencoder parameters and updating the parameters of the non-autoencoder portion of the student model by minimizing the hard loss between the student model's prediction results and the true lifetime label, and the soft loss between the student model's predicted lifetime and the teacher model's predicted lifetime. The prediction module is used to predict the remaining battery life using the trained student model.

8. The lightweight battery remaining life prediction system based on stepwise knowledge distillation according to claim 7, characterized in that: The teacher model includes: A Mamba module with local self-attention mechanism is used to learn local degradation features from the input feature dataset; A global self-attention correction module is used to dynamically adjust weights by calculating the relationships between local features in the input sequence in order to capture the dependencies between the local degenerate features. A Multi-Layer Perceptron (MLP) Conditional Autoencoder is used for reconstructing degradation features. This encodes the local and global degradation features into latent variables and reconstructs them. The reconstructed degradation features are then fed into a predictive Conditional Autoencoder (CVAE) for lifetime prediction.

9. The lightweight battery remaining life prediction system based on stepwise knowledge distillation according to claim 7, characterized in that: The student model includes: A deep separable key feature extraction module with a dual-channel attention layer is used to enhance degraded key features; A condition monitor driven by unsupervised clustering and dual-path gating mechanism is used to extract and perceive feature information under different working conditions; A conditional autoencoder OC-CVAE composed of a one-dimensional convolutional neural network (1D-CNN) and a predictor with a two-channel attention mechanism are used to characterize and model the degradation process to achieve lifetime prediction.

10. The lightweight battery remaining life prediction system based on stepwise knowledge distillation according to claim 7, characterized in that: The knowledge distillation module minimizes the loss function in the first distillation stage. To update the autoencoder parameters of the student model, and in the second distillation stage, by minimizing the loss function. To update the parameters of the non-autoencoder part of the student model, wherein, For potential spatial alignment loss, For feature reconstruction loss, For the loss of self-supervised reconstruction, For hard losses, For soft loss, This is a hyperparameter.

Images