Method for solving battery data reconstruction drift problem based on deep learnable parameters

Abstract

The invention relates to the field of battery data processing methods, in particular to a method for solving the problem of battery data reconstruction drift based on deep learnable parameters, which comprises the following steps: selecting historical target battery dimension data for preprocessing; inputting the battery dimension data into the linear layer I to obtain a mean value of data distribution, and inputting the battery dimension data into the linear layer II to obtain a variance of data distribution; standard normalization is carried out on the mean value and the variance to obtain a first matrix, the first matrix passes through a ResNet1 module and is multiplied by learnable parameters of the first matrix to obtain a second matrix, and the second matrix passes through a ResNet2 module and is added with learnable parameters of the second matrix to obtain a third matrix; inputting the finally obtained third matrix into a final task as a small module before model data processing; and S3, performing KL divergence loss constraint on the mean value and the variance obtained in the S2, and constraining the whole data distribution. According to the invention, an effective solution can be provided for a data reconstruction drift problem.

Description

Technical Field

[0001] This invention relates to the field of battery data processing methods, and more specifically to a method for solving the battery data reconstruction drift problem based on deep learnable parameters. Background Technology

[0002] In the fields of machine learning and data processing, data reconstruction drift refers to the problem of model performance degradation caused by changes in data distribution. This phenomenon is particularly common in time series data, sensor data, and recommender systems. Especially in time series forecasting, the non-stationary nature of data distribution and distribution shift are among the main challenges.

[0003] Existing methods for addressing these problems, such as global normalization (e.g., batch normalization) and traditional autoencoder methods, have limitations. Global normalization methods are poorly adaptable to variations in distribution and cannot effectively handle non-stationary data; while traditional autoencoder methods have high computational costs and complex hyperparameter settings when dealing with high-dimensional features. Furthermore, although transfer learning methods perform well in handling differences in data distribution, most methods focus on knowledge consolidation through shallow networks, failing to extract highly abstract and compact feature representations, leading to negative transfer problems, i.e., performance degradation of the model on new tasks.

[0004] Furthermore, traditional methods have high computational costs when dealing with high-dimensional features, and the hyperparameter settings are complex. In addition, using some shallow networks for learning can easily reduce the performance of the model. Summary of the Invention

[0005] The present invention aims to provide a method for solving the battery data reconstruction drift problem based on deep learnable parameters, so as to overcome the limitations when processing complex, multi-dimensional battery data.

[0006] The method for solving the battery data reconstruction drift problem based on deep learnable parameters in this scheme includes the following steps: S1, Select historical target battery dimension data for preprocessing; S2, input the battery dimension data into linear layer I to obtain the mean of the data distribution, and input the battery dimension data into linear layer II to obtain the variance of the data distribution; S3, standardize the mean and variance to obtain the first matrix, pass the first matrix through the ResNet1 module and multiply it with its learnable parameters to obtain the second matrix, pass the second matrix through the ResNet2 module and add it with its learnable parameters to obtain the third matrix, and use the final third matrix as a small module before model data processing to input into the final task. S4 applies KL divergence loss constraints to the mean and variance obtained in S2, and constrains the entire data distribution.

[0007] The beneficial effects of this plan are: With advantages such as dynamic normalization of deep learnable parameters, deep feature extraction from the ResNet module, data distribution optimization with KL divergence constraints, model independence, and broad applicability, this approach provides an effective solution to the data reconstruction drift problem. Through multiplication of the ResNet1 module with learnable parameters and addition of the ResNet2 module with learnable parameters, the model can learn richer and more robust feature representations. The deep feature extraction method not only effectively captures complex patterns in battery data but also dynamically adjusts the importance of features through learnable parameters, further enhancing the model's adaptability to changes in data distribution.

[0008] Furthermore, in S1, the target battery dimension data includes single-cell voltage sequence, current, SOC, and mileage data.

[0009] Furthermore, in S2, the mean and variance are values ​​obtained through learning the linear layer parameters, and the data distribution mean and variance values ​​are gradually learned using subsequent KL divergence loss constraints.

[0010] Furthermore, in S3, the calculation formula for the first matrix is: ; Where x is the input battery dimension data, E[x] is the mean of the battery dimension data, Var[x] is the variance of the battery dimension data, and ε is a constant; The formula for calculating the second matrix is: ; Where γ is a learnable parameter; The formula for calculating the third matrix is: ; Where β is a learnable parameter.

[0011] The beneficial effects are: the dynamic normalization method based on deep learnable parameters can automatically adapt to the distribution characteristics of different battery data, avoiding the limitations of traditional normalization methods when dealing with complex, multi-dimensional battery data; in this way, the model can better handle the non-stationarity and distribution changes in battery data, significantly improving the robustness and accuracy of battery state estimation.

[0012] Furthermore, in S4, the constraint formula is: ; Where σ is the variance of the battery dimension data, and μ is the mean of the battery dimension data.

[0013] The beneficial effects are: the KL divergence constraint-based data distribution optimization method allows the model to further optimize the data distribution, ensuring that the reconstructed data and the original data maintain a high degree of consistency in statistical characteristics; the KL divergence-based constraint mechanism not only improves the accuracy of data reconstruction, but also enhances the model's robustness to changes in data distribution, enabling the model to maintain high performance when facing the non-stationarity and distribution drift of battery data.

[0014] Furthermore, in S3, the final task includes data reconstruction, data generation, or other classification and regression tasks.

[0015] The beneficial effects are: the data obtained from the third matrix processing can be used for various subsequent tasks, allowing for flexible use of different types of data and facilitating corresponding data processing. Attached Figure Description

[0016] Figure 1 This is a schematic block diagram illustrating an embodiment of the method for solving the battery data reconstruction drift problem based on deep learnable parameters according to the present invention. Detailed Implementation

[0017] The following detailed description provides further details on specific implementation methods.

[0018] Methods for solving battery data reconstruction drift problems based on deep learnable parameters, such as Figure 1 As shown, it includes the following steps: S1. Since battery data has more dimensions, is more complex, and has a different order of magnitude compared to other time-series data, historical target battery dimension data is selected for preprocessing to provide a basis and reference for subsequent data reconstruction. The target battery dimension data is selected according to the requirements, such as single cell voltage sequence, current, SOC, mileage, etc. Some abnormal characters and invalid data, such as NAN, are deleted.

[0019] S2, input the battery dimension data into linear layer I to obtain the mean of the data distribution, input the battery dimension data into linear layer II to obtain the variance of the data distribution. The mean and variance are both values ​​obtained after learning the parameters of the linear layer, and the mean and variance of the data distribution are gradually learned using the subsequent KL divergence loss constraint.

[0020] S3 standardizes the mean and variance to obtain the first matrix. The first matrix is ​​then passed through the ResNet1 module and multiplied by its learnable parameters to obtain the second matrix. The second matrix is ​​then passed through the ResNet2 module and added to its learnable parameters to obtain the third matrix. The ResNet1 module is the residual module. The final third matrix is ​​used as a small module before model data processing and input into the final task. The final task includes subsequent data reconstruction, data generation, or other classification and regression tasks. The formula for calculating the first matrix is: ; Where x is the input battery dimension data, E[x] is the mean of the battery dimension data, Var[x] is the variance of the battery dimension data, and ε is a constant; The formula for calculating the second matrix is: ; Where γ is a learnable parameter; The formula for calculating the third matrix is: ; Where β is a learnable parameter.

[0021] The two learnable parameters, γ and β, are obtained by training an existing AI model to initialize random values. The objective function is then continuously optimized by the backpropagation algorithm. The learned parameters are saved as a .pth file after the training is completed.

[0022] S4. Apply KL divergence loss constraints to the mean and variance obtained in S2, and constrain the entire data distribution to obtain standardized battery data. The constraint formula is as follows: ; Where σ is the variance of the battery dimension data, i.e., σ=Var[x], and μ is the mean of the battery dimension data, i.e., μ=E[x].

[0023] Compared to existing technologies, the method in this embodiment forms a small module during the initial data processing to help concentrate the data distribution in the early stages of the task. Throughout the entire task, whether it's data reconstruction or data generation, the addition of KL divergence loss constraints helps the model learn a more discriminative distribution space, making the distribution constraints of the entire data space stronger and solving the data drift problem. Through a dynamic normalization method using deep learnable parameters, it can automatically adapt to the distribution characteristics of different battery data, avoiding the limitations of traditional normalization methods when dealing with complex, multi-dimensional battery data. In this way, the model can better handle the non-stationarity and distribution variations in battery data, significantly improving the robustness and accuracy of battery state estimation. Using deep feature extraction from the ResNet module and multiplication with learnable parameters, as well as addition operations between the ResNet2 module and learnable parameters, the model can learn richer and more robust feature representations. This deep feature extraction method not only effectively captures complex patterns in battery data but also dynamically adjusts the importance of features through learnable parameters, further enhancing the model's adaptability to changes in data distribution. By employing a KL divergence-based data distribution optimization approach, the model can further optimize the data distribution, ensuring that the reconstructed data maintains a high degree of consistency with the original data in terms of statistical properties. This KL divergence-based constraint mechanism not only improves the accuracy of data reconstruction but also enhances the model's robustness to changes in data distribution, enabling it to maintain high performance even when facing the non-stationarity and distribution drift of battery data. This embodiment's method is model-independent and widely applicable, serving as a trainable layer that can be flexibly integrated into any deep neural network, whether it's a Transformer, LSTM, or CNN. It seamlessly integrates with different types of neural networks, providing a unified data processing solution. This flexibility allows the method to be applied not only to battery data processing but also to other fields such as sensor data and time series forecasting.

[0024] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for solving the battery data reconstruction drift problem based on deep learnable parameters, characterized in that, Includes the following steps: S1, Select historical target battery dimension data for preprocessing; S2, input the battery dimension data into linear layer I to obtain the mean of the data distribution, and input the battery dimension data into linear layer II to obtain the variance of the data distribution; S3, standardize the mean and variance to obtain the first matrix, pass the first matrix through the ResNet1 module and multiply it with its learnable parameters to obtain the second matrix, pass the second matrix through the ResNet2 module and add it with its learnable parameters to obtain the third matrix, and use the final third matrix as a small module before model data processing to input into the final task. S4 applies KL divergence loss constraints to the mean and variance obtained in S2, and constrains the entire data distribution.

2. The method for solving the battery data reconstruction drift problem based on deep learnable parameters according to claim 1, characterized in that: In S1, the target battery dimension data includes single cell voltage sequence, current, SOC, and mileage data.

3. The method for solving the battery data reconstruction drift problem based on deep learnable parameters according to claim 1, characterized in that: In S2, the mean and variance are values ​​obtained by learning the linear layer parameters and are gradually learned by using the subsequent KL divergence loss constraint to obtain the mean and variance of the data distribution.

4. The method for solving the battery data reconstruction drift problem based on deep learnable parameters according to claim 4, characterized in that: In S3, the formula for calculating the first matrix is: ； Where x is the input battery dimension data, E[x] is the mean of the battery dimension data, Var[x] is the variance of the battery dimension data, and ε is a constant; The formula for calculating the second matrix is: ； Where γ is a learnable parameter; The formula for calculating the third matrix is: ； Where β is a learnable parameter.

5. The method for solving the battery data reconstruction drift problem based on deep learnable parameters according to claim 4, characterized in that: In S4, the constraint formula for the loss constraint is: ； Where σ is the variance of the battery dimension data, and μ is the mean of the battery dimension data.

6. The method for solving the battery data reconstruction drift problem based on deep learnable parameters according to claim 4, characterized in that: In S3, the final task includes data reconstruction, data generation, or other classification and regression tasks.

Images