Battery health state estimation method based on mechanical signal fusion Transformer

Abstract

The invention discloses a battery health state estimation method based on mechanical signal fusion Transform. According to the method, on the basis of fully excavating battery mechanical and electrochemical multi-source information, high-precision SOH online estimation applicable to multiple scenes can be realized. And after the mechanical stress signal is introduced, the model not only can capture a small structure degradation throughout which is difficult to reflect by a traditional electrochemical index, but also can utilize a mechanical-electrochemical coupling relationship to early warn the aging trend in advance, so that the timeliness and reliability of estimation are further improved. According to the scheme, local interaction of sequence long-time dependence and statistical characteristics is considered structurally, two targets of regression and reconstruction are balanced during training, the accuracy of the model is ensured, and the generalization ability and the online response stability of the model to different battery cell batches, environment temperatures and use condition changes are greatly enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of battery health management, and specifically relates to a battery health state estimation method based on mechanical signal fusion Transformer. BACKGROUND

[0002] Lithium-ion batteries have become the core power source of electric vehicles, energy storage power stations and portable electronic devices due to their high energy density, long cycle life and other advantages. With the expansion of the industry, accurate estimation of the full life cycle state of the battery has become a key technical bottleneck for ensuring the safe operation of the system, optimizing energy scheduling and prolonging the service life. Among them, the health state (SOH) as a core indicator representing the remaining capacity and performance degradation of the battery (definition: the percentage of the current maximum capacity to the initial capacity), directly affects the available energy range of the battery pack, the development of charging and discharging strategies and the accuracy of safety risk warning. The current SOH estimation method mainly relies on electrochemical signals such as voltage, current, temperature, etc., and has certain limitations, and recent research has shown that there is a strong correlation between the internal mechanical stress changes and capacity degradation during the aging process of the battery.

[0003] Although existing research has shown that mechanical signals have advantages in battery SOH estimation, there are still some challenges in its practical application, such as the difference in time and space scales between mechanical signals and electrochemical signals, and the need to separate the thermal expansion effect in the charging and discharging process from the aging signal.

[0004] Therefore, there is an urgent need for a battery health state evaluation method based on mechanical behavior that can fuse multi-source time series signals of mechanical stress and electrochemical parameters to construct a high-robustness SOH dynamic prediction model. SUMMARY

[0005] The purpose of the present application is to solve the problems mentioned above, and to provide a battery health state estimation method based on mechanical signal fusion Transformer.

[0006] The technical scheme adopted by the present application is as follows: a battery health state estimation method based on mechanical signal fusion Transformer, the method comprising the following steps:

[0007] S1: multi-source data preprocessing and feature construction;

[0008] S2: dual-path network construction and joint training, after completing the preprocessing of multi-source time series and aggregated features, focusing on the specific construction of deep neural networks, parallel extraction of different modal information and final end-to-end joint training;

[0009] S3: using the trained network to estimate the battery health state;

[0010] The step S2 specifically comprises:

[0011] S2.1: Construct sequence branch;

[0012] S2.2: Construct aggregation branch;

[0013] S2.3: Feature fusion and regression head design;

[0014] S2.4: Joint training.

[0015] In a preferred embodiment, the step S1 specifically comprises:

[0016] S11. The original collected data includes pressure signal, voltage signal, current signal and temperature signal. Before subsequent operations, the data is first preprocessed, including time step alignment, filtering out invalid values and other operations.

[0017] S12. Charging segment screening. After data preprocessing, further screening is performed on the charging segment, and the charging segment with SOC range of 50% to 100% is selected and used as subsequent training and test samples.

[0018] S13. Group-level aggregation feature extraction. For each retained charging segment, the following aggregation features are extracted, including maximum / minimum single cell voltage, highest / lowest single cell temperature, trickle charging duration, and trickle charging segment average current.

[0019] S14. Single cell voltage PCA dimensionality reduction;

[0020] S15. Correlation test and redundancy elimination. The Pearson correlation coefficient is calculated for the extracted group-level aggregation features and the features after PCA dimensionality reduction of the single cell voltage, and the correlation degree is quantitatively evaluated. Features with a correlation coefficient less than a set value are screened out.

[0021] In a preferred embodiment, in the step S14, the single cell voltage PCA dimensionality reduction first performs zero mean and standardization processing on the single cell voltage at each time point, so that the mean of each dimension data is 0 and the variance is 1. A standardized data matrix X ∈ R N×K is constructed, where K represents the number of battery cells, N represents the number of samples, and the covariance is calculated based on the matrix, as shown in formula (1)

[0022]

[0023] Then the covariance matrix is decomposed to obtain K characteristic values λ k and corresponding unit eigenvectors v k . Where λ k represents the data variance size retained after projection along v k . Let λ (1)≥ λ (2) ≥...≥ λ (20) The eigenvalues are arranged in descending order, and the explained variance ratio of the ith principal component is

[0024]

[0025] If the cumulative explained variance of the first L principal components reaches a set threshold, generally 95%, the L principal components are finally retained, and the original high-dimensional vector is projected into a subspace with a dimension of L.

[0026] In a preferred embodiment, the step S2.1 specifically comprises:

[0027] S2.1-1. Input preparation and embedding:

[0028] First, the preprocessed voltage principal component sequence and pressure signal sequence are spliced in parallel according to the time sequence to construct a multi-channel time sequence data block. Then each channel data is first mapped to a high-dimensional feature space required by the model through a fully connected mapping layer, and then added with a fixed time position coding to retain the sequence order information.

[0029] S2.1-2. Multi-layer self-attention encoder:

[0030] A stacked encoder layer is used, and each layer internally includes two sub-modules: one is a traditional multi-head self-attention mechanism, which is used to automatically learn the dependency between time steps within the sequence; the other is a specially designed "force-electricity correlation attention head", which focuses on modeling the interaction between the mechanical stress channel and the electrochemical channel, thereby highlighting the guiding effect of the stress signal on the voltage change. After the attention calculation of each layer, a group of feedforward networks, residual connections and layer normalization operations are followed to ensure stable information flow and effective gradient transmission.

[0031] S2.1-3. Time sequence feature aggregation:

[0032] After the completion of the encoder stacking, the time sequence features of the entire sequence are condensed into a fixed-length vector through a simple pooling operation (such as taking the average or maximum value of all time step vectors), which is used to represent the global evolution features of the multi-source sequence.

[0033] In a preferred embodiment, the step S2.2 specifically comprises:

[0034] S2.2-1. Feature concatenation:

[0035] The battery-level seven-dimensional statistical features obtained in step 1 are spliced into a large vector with the global statistics of the principal components of the single-cell voltage (such as the mean and standard deviation of each principal component on the sequence) to represent the overall aggregated information of the charging segment.

[0036] S2.2-2. Local pattern mining:

[0037] To mine the local interaction patterns between different statistical dimensions, the vector is first reshaped into a "pseudo-temporal" or "pseudo-spatial" structure, and then input into a multi-layer one-dimensional convolutional network. Each layer of the convolutional network learns several convolution kernels that can automatically capture several combinations and local co-occurrence rules between feature dimensions.

[0038] Pooling operations are interleaved between convolution layers to compress the dimensions of the feature map while retaining the main information, thereby reducing complexity in subsequent calculations.

[0039] S2.2-3. Global representation generation:

[0040] After the last convolution layer, the features of different channels are unified and gathered into a fixed-length vector through global average pooling, which serves as a deep representation of the macro statistical information of the segment for the aggregation branch.

[0041] In a preferred embodiment, the step S2.3 specifically comprises:

[0042] S2.3-1. Splicing and fusion:

[0043] The global temporal vector output by the Transformer branch is directly spliced with the global aggregation vector output by the CNN branch in the feature dimension to form a comprehensive representation containing multiple sources of information.

[0044] S2.3-2. Regression network:

[0045] The comprehensive representation vector passes through two fully connected networks in sequence, each followed by a nonlinear activation and moderate regularization (such as Dropout or weight decay) to enhance the generalization ability of the model. The last layer is a linear regression layer that directly outputs the normalized SOH estimate.

[0046] In a preferred embodiment, the step S2.4 specifically comprises:

[0047] S2.4-1. Loss function:

[0048] A multi-task joint loss is adopted: including SOH regression error and two-way sequence self-supervised reconstruction error (voltage reconstruction and pressure reconstruction). By setting weights for different tasks, the accuracy of the main task (SOH regression) is ensured, and the auxiliary role of the sequence reconstruction task in feature expression is fully utilized.

[0049] S2.4-2. Optimization strategy:

[0050] The adaptive learning rate optimizer is selected, and a learning rate scheduling strategy is combined to dynamically adjust the step size to accelerate convergence. At the same time, an early stopping mechanism is set: when the main task loss on the validation set does not decrease significantly within a few training rounds, the learning rate is reduced or terminated in advance to prevent overfitting.

[0051] S2.4-3. Batch and regularization:

[0052] According to the data volume, select an appropriate batch size to balance training efficiency and memory usage; introduce moderate regularization methods (such as Dropout layer, weight decay) in convolutional and fully connected layers to further improve the robustness of the model in unseen conditions.

[0053] S2.4-4. Distributed and parallelization:

[0054] For large-scale data sets, it is recommended to use data parallelism or model parallelism strategies to copy the model to a multi-card or multi-machine training environment, while synchronizing the gradient, thereby greatly shortening the training time.

[0055] S2.4-5. Model saving and version management:

[0056] At the end of each training cycle, record and save the model parameters, optimizer state, and related hyperparameters; compare the performance on the validation set to select the optimal model snapshot, making it easier to deploy or further fine-tune later.

[0057] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present application are:

[0058] In the present application, high-precision, multi-scenario SOH online estimation can be achieved based on the full exploitation of battery mechanical and electrochemical multi-source information. After introducing mechanical stress signals, the model not only captures the early signs of minor structural degradation that traditional electrochemical indicators cannot reflect, but also uses the mechanical-electrochemical coupling relationship to provide early warning of aging trends, thereby further improving the timeliness and reliability of the estimation. This scheme balances the local interaction of sequence long dependence and statistical characteristics in structure, and balances the regression and reconstruction goals during training, ensuring the accuracy of the model while greatly enhancing its generalization ability and online response stability to different battery batches, environmental temperatures, and usage conditions. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The flowchart of the present application is shown. DETAILED DESCRIPTION

[0060] In order to make the purposes, technical solutions and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0061] Referring to Figure 1 ,

[0062] A battery health state estimation method based on mechanical signal fusion Transformer, the method comprising the following steps:

[0063] S1: multi-source data preprocessing and feature construction;

[0064] S2: dual-path network construction and joint training, after completing the preprocessing of multi-source time series and aggregated features, focusing on the specific construction of deep neural network, parallel extraction of different modal information and final end-to-end joint training;

[0065] S3: using the trained network to estimate the battery health state;

[0066] Step S2 specifically includes:

[0067] S2.1: constructing a sequence branch;

[0068] S2.2: constructing an aggregation branch;

[0069] S2.3: feature fusion and regression head design;

[0070] S2.4: joint training.

[0071] Step S1 specifically includes:

[0072] S11. The original collected data includes pressure signal, voltage signal, current signal and temperature signal. Before performing subsequent operations, the data is first preprocessed, including time step alignment, filtering out invalid values and other operations.

[0073] S12. Charging segment screening. After data preprocessing, the charging segment is further screened, and the charging segment with SOC range of 50% to 100% is selected and used as the subsequent training sample and test sample.

[0074] S13. Group-level aggregated feature extraction, for each retained charging segment, the following aggregated features are extracted, including maximum / minimum single cell voltage, highest / lowest single cell temperature, trickle charging time length, trickle charging segment average current.

[0075] S14. Single cell voltage PCA dimension reduction;

[0076] S15. Correlation test and redundancy elimination, the pearson correlation coefficient of the extracted group level aggregation features and the features of the single cell voltage after PCA dimension reduction is calculated, and the correlation degree is quantitatively evaluated. And the features with correlation coefficient less than the set value are screened out.

[0077] In step S14, the single cell voltage PCA dimension reduction first performs zero mean and standardization processing on the single cell voltage at each time, so that the mean value of each dimension data is 0 and the variance is 1, and a standardized data matrix X ∈ R N×K is constructed, where K represents the number of battery monomers, N represents the number of samples, and the covariance is calculated based on the matrix, as shown in formula (1)

[0078]

[0079] Then the covariance matrix is decomposed to obtain K pairs of characteristic values λ k and corresponding unit eigenvectors v k . Where λ k represents the data variance size retained after projection along the direction of v k . Let λ (1) ≥ λ (2) ≥... ≥ λ (20) be the eigenvalues in descending order, then the explanation variance ratio of the i-th principal component is

[0080]

[0081] If the cumulative explanation variance of the first L principal components reaches the set threshold, generally 95%, then the final L principal components are retained, and the original high-dimensional vector is projected into a subspace with dimension L.

[0082] In step S2.1, specifically comprising:

[0083] S2.1-1. Input preparation and embedding:

[0084] First, the voltage principal component sequence and the pressure signal sequence obtained by preprocessing are parallelly spliced according to time sequence to construct a multi-channel time sequence data block. Then each channel data is first mapped to a high-dimensional feature space required by the model through a fully connected mapping layer, and then added with a fixed time position coding to retain the sequence order information.

[0085] S2.1-2. Multi-layer self-attention encoder:

[0086] The stacked encoder layers each contain two sub-modules: one is a traditional multi-head self-attention mechanism, which automatically learns the dependencies between time steps within the sequence; the other is a specially designed "force-electricity correlation attention head", which focuses on modeling the interaction between the mechanical stress channel and the electrochemical channel, highlighting the guiding effect of stress signals on voltage changes. After the attention calculation at each layer, a set of feedforward networks, residual connections, and layer normalization operations are followed to ensure stable information flow and effective gradient transmission.

[0087] S2.1-3. Time series feature aggregation:

[0088] After the completion of the encoder stack, the time series features of the entire sequence are condensed into a fixed-length vector through a simple pooling operation (such as taking the average or maximum value of all time step vectors), which represents the global evolution characteristics of the multi-source sequence.

[0089] In step S2.2, specifically includes:

[0090] S2.2-1. Feature concatenation:

[0091] The battery pack-level seven-dimensional statistical features obtained in step 1 are concatenated with the global statistics of the single cell voltage principal components (such as the mean and standard deviation of each principal component over the sequence) to form a large vector representing the overall aggregated information of the charging segment.

[0092] S2.2-2. Local pattern mining:

[0093] To explore the local interaction patterns between different statistical dimensions, the vector is first reshaped into a "pseudo-time" or "pseudo-space" structure, and then input into a multi-layer one-dimensional convolutional network. Each convolutional network layer learns several convolution kernels that can automatically capture a number of combinations and local co-occurrence rules between feature dimensions.

[0094] Pooling operations are interleaved between convolutional layers to compress the dimensions of the feature maps while preserving the main information, thereby reducing complexity in subsequent calculations.

[0095] S2.2-3. Global representation generation:

[0096] After the last convolutional layer, global average pooling is used to unify the features from different channels into a fixed-length vector, which serves as a deep representation of the macroscopic statistical information of the segment by the aggregation branch.

[0097] In step S2.3, specifically includes:

[0098] S2.3-1. Concatenation and fusion:

[0099] The global temporal vector output by the Transformer branch and the global aggregate vector output by the CNN branch are directly concatenated along the feature dimension to form a comprehensive representation containing multi-source information.

[0100] S2.3-2. Regression Network:

[0101] The comprehensive representation vector is passed through two fully connected layers, each followed by nonlinear activation and appropriate regularization (such as Dropout or weight decay) to enhance the model's generalization ability. The final layer is a linear regression layer that directly outputs the normalized SOH estimate.

[0102] Step S2.4 specifically includes:

[0103] S2.4-1. Loss Function:

[0104] A multi-task joint loss is employed, comprising SOH regression error and two-way sequence self-supervised reconstruction errors (voltage reconstruction and pressure reconstruction). By assigning weights to different tasks, the accuracy of the main task (SOH regression) is ensured while fully utilizing the auxiliary role of the sequence reconstruction task in feature representation.

[0105] S2.4-2. Optimization Strategy:

[0106] An adaptive learning rate optimizer is selected, combined with a learning rate scheduling strategy, to dynamically adjust the step size to accelerate convergence. An early stopping mechanism is also implemented: if the loss on the main task on the validation set does not decrease significantly within several training epochs, the learning rate is reduced or the training terminates early to prevent overfitting.

[0107] S2.4-3. Batch and Regularization:

[0108] Choose an appropriate batch size based on the amount of data to balance training efficiency and memory usage; introduce appropriate regularization techniques (such as Dropout layers and weight decay) into convolutional and fully connected layers to further improve the robustness of the model under unseen conditions.

[0109] S2.4-4. Distributed and Parallelized Systems:

[0110] For large-scale datasets, it is recommended to adopt data parallelism or model parallelism strategies, copy the model to a multi-GPU or multi-machine training environment, and synchronize gradients at the same time, thereby significantly shortening the training time.

[0111] S2.4-5. Model Saving and Version Management:

[0112] At the end of each training cycle, the model parameters, optimizer state, and related hyperparameters are recorded and saved; by comparing the performance on the validation set, the best model snapshot is selected to facilitate subsequent deployment or further fine-tuning.

[0113] Through the design of the above-mentioned double-path parallel and combined training, the application can realize high-precision, multi-scene applicable SOH online estimation on the basis of fully mining the mechanical and electrochemical multi-source information of the battery. After introducing the mechanical stress signal, the model can not only capture the early signs of small structural degradation that traditional electrochemical indicators cannot reflect, but also use the mechanical-electrochemical coupling relationship to give an early warning of aging trends, thereby further improving the timeliness and reliability of the estimation. The scheme takes into account the long-time dependence and local interaction of statistical characteristics in structure, balances the regression and reconstruction two goals during training, and not only guarantees the accuracy of the model, but also greatly enhances the generalization ability and online response stability of the model to different battery batches, environmental temperatures and use conditions.

[0114] It should be noted that, in this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", "includes", "including" or any other variation thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0115] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A battery health state estimation method based on mechanical signal fusion Transformer, characterized in that: The method includes the following steps: S1: Multi-source data preprocessing and feature construction; S2: Dual-path network construction and joint training. After completing the preprocessing of multi-source temporal and aggregated features, the focus is on the specific construction of deep neural networks, the parallel extraction of information from different modalities, and the final end-to-end joint training. S3: Use the trained network to estimate battery health status; Step S2 specifically includes: S2.1: Construct sequence branches; S2.2: Construct aggregate branches; S2.3: Feature fusion and regression head design; S2.4: Conduct joint training.

2. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 1, characterized in that: Step S1 specifically includes: S11. The raw data collected includes pressure signals, voltage signals, current signals, and temperature signals. Before proceeding with subsequent operations, the data must be preprocessed, including time step alignment and filtering out invalid values. S12. Charging segment screening: After data preprocessing, the charging segments are further screened to select those with a SOC range of 50% to 100% and used as subsequent training and testing samples. S13. Group-level aggregation feature extraction: For each retained charging segment, extract the following aggregation features, including the maximum / minimum single-cell voltage, the highest / lowest single-cell temperature, the trickle charging duration, and the average current of the trickle charging segment. S14. Single-unit voltage PCA dimensionality reduction; S15. Correlation test and redundancy removal: Calculate the Pearson correlation coefficient for the extracted group-level aggregation features and the individual voltage features after PCA dimensionality reduction, and quantitatively evaluate the correlation; and screen out features with correlation coefficients less than the set value.

3. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 2, characterized in that: In step S14, the PCA dimensionality reduction of the individual unit voltage first performs zero-mean and standardization processing on the individual unit voltage at each time step, so that the mean of each dimension of data is 0 and the variance is 1, and constructs a standardized data matrix X∈R. N×K Where K represents the number of battery cells and N represents the sample size, the covariance is calculated based on this matrix, as shown in the formula. Then, eigenvalue decomposition is performed on the covariance matrix to obtain K pairs of eigenvalues ​​λ. k and the corresponding unit eigenvector v k ;where λ k Indicates along v k The magnitude of the variance of the data retained after directional projection; let λ (1) ≥λ (2) ≥...≥λ (20) If the eigenvalues ​​are arranged in descending order, then the variance ratio explained by the i-th principal component is: If the cumulative explained variance of the first L principal components Once the set threshold is reached, typically 95%, L principal components are retained, and the original high-dimensional vector is projected onto a subspace of dimension L.

4. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 1, characterized in that: Step S2.1 specifically includes: S2.1-1. Input Preparation and Embedding: First, the preprocessed voltage principal component sequence and pressure signal sequence are spliced ​​in parallel according to time order to construct a multi-channel time series data block. Then, each channel's data first passes through a fully connected mapping layer to map the original numerical space to the high-dimensional feature space required by the model, and then is added to a fixed time position code to preserve the sequence order information; S2.1-2. Multilayer Self-Attention Encoder: The encoder employs stacked layers, each containing two main sub-modules: one is a traditional multi-head self-attention mechanism, used to automatically learn the dependencies between time steps within the sequence; the other is a specially designed "force-electric correlation attention head", which focuses on modeling the interaction between mechanical stress channels and electrochemical channels, thereby highlighting the guiding role of stress signals on voltage changes. Furthermore, after the attention calculation of each layer, a set of feedforward networks, residual connections, and layer normalization operations are immediately followed to ensure the stability of the information flow and the effective transmission of gradients. S2.1-3. Convergence of Temporal Features: After the encoders are stacked, a simple pooling operation is used to condense the temporal features of the entire sequence into a fixed-length vector to characterize the global evolutionary features of the multi-source sequence.

5. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 1, characterized in that: Step S2.2 specifically includes: S2.2-1. Feature splicing: The seven-dimensional statistical features of the battery pack level obtained in step 1 are concatenated with the global statistics of the principal components of the single cell voltage to form a large vector, which represents the overall aggregate information of the charging segment. S2.2-2. Local Pattern Mining: To uncover local interaction patterns between different statistical dimensions, the vector is first reshaped into a "pseudo-temporal" or "pseudo-spatial" structure and then input into a multi-layer one-dimensional convolutional network. Each layer of the convolutional network learns several convolutional kernels that can automatically capture several combinations and local co-occurrence patterns between feature dimensions. Interleaving pooling operations between convolutional layers allows the network to retain key information while compressing the dimension of feature maps, thereby reducing complexity in subsequent computations. S2.2-3. Global Representation Generation: After the last convolutional layer, global average pooling is used to unify the features of different channels into a fixed-length vector, which serves as the aggregation branch to represent the deep macroscopic statistical information of this segment.

6. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 1, characterized in that: Step S2.3 specifically includes: S2.3-1. Splicing and fusion: The global temporal vector output by the Transformer branch and the global aggregate vector output by the CNN branch are directly concatenated along the feature dimension to form a comprehensive representation containing multi-source information. S2.3-2. Regression Network: The comprehensive representation vector passes through two fully connected layers, each followed by nonlinear activation and appropriate regularization to enhance the model's generalization ability; the last layer is a linear regression layer that directly outputs the normalized SOH estimate.

7. The battery health state estimation method based on mechanical signal fusion Transformer as described in claim 1, characterized in that: Step S2.4 specifically includes: S2.4-1. Loss Function: A multi-task joint loss is adopted, including SOH regression error and two-way sequence self-supervised reconstruction error. By setting weights for different tasks, the accuracy of the main task is ensured, while making full use of the auxiliary role of the sequence reconstruction task in feature representation. S2.4-2. Optimization Strategy: An adaptive learning rate optimizer is selected and combined with a learning rate scheduling strategy to dynamically adjust the step size to accelerate convergence. At the same time, an early stopping mechanism is set: when the loss of the main task on the validation set does not decrease significantly within several training rounds, the learning rate is reduced or the training is terminated early to prevent overfitting. S2.4-3. Batch and Regularization: Choose an appropriate batch size based on the amount of data to balance training efficiency and memory usage; introduce appropriate regularization methods in convolutional and fully connected layers to further improve the robustness of the model under unseen conditions. S2.4-4. Distributed and Parallelized Systems: For large-scale datasets, it is recommended to adopt data parallelism or model parallelism strategies, copy the model to a multi-GPU or multi-machine training environment, and synchronize gradients at the same time, thereby significantly shortening the training time. S2.4-5. Model Saving and Version Management: At the end of each training cycle, the model parameters, optimizer state, and related hyperparameters are recorded and saved; by comparing the performance on the validation set, the best model snapshot is selected to facilitate subsequent deployment or further fine-tuning.

Images