A method for predicting and optimizing the service life of lithium-ion battery separators

Abstract

The present invention proposes a method for predicting and optimizing the service life of lithium-ion battery diaphragms, which specifically relates to the field of battery diaphragm service life prediction. It includes collecting key parameters in battery operation, preprocessing the collected key parameter data, building a lithium-ion battery diaphragm service life prediction model, training and testing the life prediction model, and the life prediction model is mainly composed of a time-frequency mapper module, a frequency domain cross-weighted attention mechanism, and an information fusion unit; wherein the time-frequency decomposer in the frequency mapper module decomposes the time series data into the frequency domain, and the main frequency component filter retains the top M frequency components that are most important to the frequency change; the frequency domain cross-weighted attention mechanism establishes the connection between the real state and the virtual state to capture the frequency domain characteristics of the time series; the information fusion unit is used to integrate the real state and virtual state information enhanced by the frequency domain cross-weighted attention mechanism. The present application can accurately predict the remaining service life of lithium battery diaphragms in the future.

Description

Technical Field

[0001] The present invention belongs to the field of battery separator service life prediction, and in particular relates to a lithium-ion battery separator service life prediction and optimization method. Background Art

[0002] As the core technology in today's energy storage field, lithium-ion batteries are widely used in many fields such as portable electronic devices, electric vehicles and energy storage systems. However, the performance and safety of lithium-ion batteries depend to a large extent on the quality and stability of their internal diaphragms. The diaphragm not only physically separates the positive and negative electrodes, but also affects the transmission efficiency of lithium ions in the battery. With the long-term use of the battery, the diaphragm may gradually deteriorate due to electrochemical reactions, mechanical stress and environmental factors, resulting in a decrease in battery performance and even safety hazards. Therefore, accurately predicting the service life of lithium-ion battery diaphragms is of great significance to extending the battery's working cycle and improving safety.

[0003] Existing methods for predicting the life of lithium battery diaphragms face many challenges, mainly including the difficulty in effectively handling long-term dependencies, complex nonlinear problems under the influence of multiple factors, and insufficient ability to deal with data noise and incompleteness. In addition, the nonlinear attenuation and random fluctuations of battery life are difficult to accurately capture through simple models, and existing methods have weak generalization capabilities in different battery types and application scenarios, and cannot fully extract multi-scale features, resulting in insufficient prediction accuracy and robustness. This application combines multivariate time series analysis with machine learning algorithms to comprehensively monitor and analyze key parameters during battery operation, which can more accurately predict the life of the diaphragm and provide more scientific guidance for battery management and maintenance. Summary of the invention

[0004] The main purpose of the present invention is to provide a lithium-ion battery diaphragm service life prediction optimization method and construct a lithium-ion battery diaphragm service life prediction model, wherein the diaphragm service life prediction model is mainly composed of a time-frequency mapper module, a frequency domain cross-weighted attention mechanism, and an information fusion unit, wherein the time-frequency decomposer in the frequency mapper module decomposes the time series data into the frequency domain, the frequency domain real state and virtual state record the amplitude and phase information of the time series at different frequencies, and the main frequency component filter retains the most important previous frequency changes. frequency components, and enhances the real state and the virtual state in the frequency domain; the frequency domain cross-weighted attention mechanism captures the frequency domain characteristics of the time series by establishing a connection between the real state and the virtual state, and uses the frequency domain cross-weighted attention mechanism to exchange information between the real state and the virtual state to enhance the frequency domain representation; the information fusion unit is used to integrate the real state and virtual state information enhanced by the frequency domain cross-weighted attention mechanism, and perform multi-scale feature extraction; the present application effectively captures long-term dependencies and complex periodic changes through frequency domain analysis, and can identify trends and fluctuations in battery life over time. Compared with traditional time domain models, the lithium-ion battery separator service life prediction model proposed in the present application is better at processing multi-frequency components and long-term changes in battery life, thereby improving the prediction accuracy and adapting to the complex characteristics of lithium battery life affected by multiple factors.

[0005] To achieve the above object, the technical solution of the present invention is: a method for predicting and optimizing the service life of a lithium-ion battery separator, the method comprising:

[0006] S1. Collect key parameters during battery operation, which are parameters that directly and indirectly affect the service life of the battery separator, including the internal temperature of the battery, the voltage at both ends of the battery, the current during discharge, the depth of discharge, the number of charge and discharge cycles, the internal resistance of the battery, the ambient humidity, and the remaining capacity of the battery, wherein the remaining capacity of the battery is used as the label value of the battery separator life;

[0007] S2. Preprocessing the collected key parameter data, wherein the preprocessing includes normalization processing and data time series segmentation operations;

[0008] S3. Constructing a lithium-ion battery separator service life prediction model, the specific method includes:

[0009] S31, performing feature conversion on the preprocessed data, uniformly converting the data of multiple variables and embedding them into the same feature space;

[0010] S32, propose a time-frequency mapper module, the processing of the time-frequency mapper module is divided into a time-frequency decomposer, an extraction of real and virtual states in the frequency domain, and a main frequency component filter. The time-frequency decomposer decomposes the time series data into the frequency domain. The real and virtual states in the frequency domain record the amplitude and phase information of the time series at different frequencies. The main frequency component filter retains the most important previous information for frequency changes. frequency components, enhancing the real state and virtual state in the frequency domain and reducing the redundancy of frequency domain features;

[0011] S33. A frequency domain cross-weighted attention mechanism is proposed. The frequency domain cross-weighted attention mechanism captures the frequency domain characteristics of the time series by establishing a connection between the real state and the virtual state, and uses the frequency domain cross-weighted attention mechanism to exchange information between the real state and the virtual state to enhance the frequency domain representation.

[0012] S34. An information fusion unit is proposed to integrate the real and virtual information obtained by frequency domain cross-weighted attention mechanism enhancement and perform multi-scale feature extraction;

[0013] S35, introducing a fully connected layer, using the output of the information fusion unit as the input of the fully connected layer, and finally the output of the fully connected layer is the output value of the final prediction model;

[0014] S4, training a lithium-ion battery separator service life prediction model, wherein the training is to set model hyperparameters and parameters in the model training process;

[0015] S5. Testing the lithium-ion battery separator service life prediction model, wherein the test is to test the accuracy and performance of the model in a real deployment environment.

[0016] Furthermore, in the steps S3 and S31, the feature conversion projects the time series data after the preprocessing operation from its high-dimensional input space to the same feature space through linear embedding, wherein the linear embedding process is expressed as:

[0017] ;

[0018] In the formula, is the output value after linear embedding, is the weight value of the linear embedding operation, is the bias term of the linear embedding operation, is the normalized value, i.e., the input data of feature transformation.

[0019] Further, in the step S3, in S32, the time-frequency decomposer decomposes the time series data into the frequency domain, and the time-frequency decomposer is obtained by short-term Fourier transform calculation, and the process is to decompose a long time series into multiple overlapping time windows, and perform short-term Fourier transform calculation in each window, and the time-frequency decomposer records the frequency component changes of the time series in different time periods to capture the instantaneous frequency characteristics; wherein, in the process of decomposing the time series data into the frequency domain, the method of decomposing the long time series into multiple short time windows is the Hanning window method, and the time series input using the Hanning window is divided into multiple time windows, and the data in each window is subjected to short-term Fourier transform calculation to calculate the frequency component, and for a given time series data , the short-term Fourier change calculation formula is as follows:

[0020] ;

[0021] In the formula, For the Time series data of variables, is the Hanning window function, which is used to segment time series data. For the The angular frequency corresponding to the frequency component, t is the time step, is a complex exponential form used to calculate the frequency component of each time step, It is the output of the short-term Fourier transform calculation; the result of the short-term Fourier transform calculation is a joint representation of the time domain and the frequency domain, capturing the frequency information of the time series in different time periods.

[0022] Furthermore, in the step S3, in S32, the frequency domain real state and the frequency domain virtual state record the amplitude and phase information of the time series at different frequencies, and the output of the short-term Fourier change includes complex exponential , expressed as:

[0023] ; Therefore, the output of the short-term Fourier change is decomposed into the real state in the frequency domain and the virtual state in the frequency domain. The calculation formula of the real state in the frequency domain is expressed as: ;

[0024] In the formula, For the The time series in windows and frequencies The real state at the frequency domain is expressed as:

[0025] ;

[0026] In the formula, For the The time series in windows and frequencies The virtual state of being.

[0027] Furthermore, in the steps S3 and S32, the main frequency component filter selects the one with the largest amplitude in each window. frequency components, amplitude The calculation formula is as follows:

[0028] ;

[0029] According to the calculated amplitude, select Important frequency components:

[0030] ;

[0031] In the formula, The calculated front There are several important frequency components, among which, is the result of real-state enhancement in the frequency domain, It is the result of frequency domain imaginary part enhancement;

[0032] Finally, the main frequency component filter selects the most important After the frequency components, The real and imaginary parts extracted from the windows are spliced ​​into and , the formulas are:

[0033] ;

[0034] ;

[0035] In the formula, For splicing operation, is the number of windows, is the number of variables in the time series, is the number of most important frequency components selected in each window, is the representation of the real state in the frequency domain, is the representation of the virtual state in the frequency domain.

[0036] Furthermore, in the steps S3 and S33, the frequency domain cross-weighted attention mechanism consists of two parts, namely, a real-state attention mechanism enhanced by a virtual state in the frequency domain and a virtual-state attention mechanism enhanced by a real state in the frequency domain;

[0037] The real state attention mechanism for frequency domain virtual state enhancement, virtual state as query , virtual state as key , the actual state as value , the formula for enhancing the real-state attention mechanism through frequency-domain virtual-state enhancement is:

[0038] ;

[0039] In the formula, The virtual state query matrix is ​​obtained by linear mapping for the virtual state. for The transpose operation of To calculate the weight of the linear transformation in the virtual state query matrix, For the virtual state, we obtain the virtual state bond matrix through linear mapping. To calculate the weight of the linear transformation in the process of virtual bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weight of the linear transformation in the process of calculating the virtual state value matrix, is the dimension of the virtual key matrix, used to scale the attention scores.

[0040] Furthermore, in the steps S3 and S33, the virtual state attention mechanism for frequency domain real state enhancement is used, and the real state is used as the query , the real state as a key , virtual state as value , the formula for virtual state attention mechanism enhancement through frequency domain real state enhancement is:

[0041] ;

[0042] In the formula, The actual state query matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weight of linear transformation in the actual query matrix, The real state bond matrix is ​​obtained by linear mapping. To calculate the weight of the linear transformation in the process of calculating the real bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weights of the linear transformation in the process of calculating the real-state value matrix, is the dimension of the real key matrix, used to scale the attention scores.

[0043] Furthermore, in the steps S3 and S33, the multi-head attention mechanism is finally used to output the real state and virtual state enhanced by the frequency domain cross-weighted attention mechanism, and the enhanced real part is expressed as:

[0044] ;

[0045] In the formula, For the enhanced reality, is the number of attention heads, To calculate the projection matrix used in the enhanced reality process, It is the matrix concatenation operation;

[0046] ;

[0047] In the formula, For the enhanced virtual state, It is the projection matrix used in calculating the enhanced virtual state.

[0048] Furthermore, in the steps S3 and S34, the information fusion unit is composed of a linear projection and a multi-scale information fuser; the linear projection transformation fuses the frequency domain information to make it suitable for the next multi-scale convolution operation. The linear projection performs a linear transformation on the input data through a weight matrix. For the real state and virtual state of the input, the linear projection is expressed in the following form:

[0049] ;

[0050] In the formula, For the enhanced reality, For the enhanced virtual state, is the weight matrix of linear projection, is the bias term of the linear projection, To concatenate the real and imaginary parts, a new joint representation is obtained;

[0051] Then the multi-scale information fusion extracts multiple feature representations from different scales. Specifically, multiple convolution kernels are used to process the input data at multiple scales. The size of the convolution kernel is , , , and concatenate the results. The formula is:

[0052] ;

[0053] In the formula, is the final output value of the multi-scale information fusion device, is the convolution operation, It is a maximum pooling operation used to highlight significant features; the results of each convolution and pooling operation are concatenated The operations are combined together to form a multi-scale joint representation while preserving both local and global feature information.

[0054] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0055] In the present invention, a lithium-ion battery diaphragm service life prediction model is constructed. The diaphragm service life prediction model is mainly composed of a time-frequency mapper module, a frequency domain cross-weighted attention mechanism, and an information fusion unit. The time-frequency decomposer in the frequency mapper module decomposes the time series data into the frequency domain, and the frequency domain real state and virtual state record the amplitude and phase information of the time series at different frequencies. The main frequency component filter retains the most important previous information for the frequency change. frequency components to enhance the real state and virtual state in the frequency domain; the frequency domain cross-weighted attention mechanism captures the frequency domain characteristics of the time series by establishing a connection between the real state and the virtual state, and uses the frequency domain cross-weighted attention mechanism to exchange information between the real state and the virtual state to enhance the frequency domain representation; the information fusion unit is used to integrate the real state and virtual state information enhanced by the frequency domain cross-weighted attention mechanism, and perform multi-scale feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 The present invention is a flowchart of the steps of a lithium-ion battery separator service life prediction and optimization method.

[0057] Figure 2 Flowchart of the steps in the separator service life prediction model for lithium-ion batteries.

[0058] Figure 3 This is the architecture diagram of the time-frequency mapper module in the lithium-ion battery separator service life prediction model.

[0059] Figure 4 Structural diagram of the frequency-domain cross-weighted attention mechanism in the lithium-ion battery separator service life prediction model.

[0060] Figure 5 This is a comparison chart of the predicted value of the prediction model and the true value. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0062] See also Figure 1-Figure 5 The present invention provides a technical solution: a method for predicting and optimizing the service life of a lithium-ion battery diaphragm, the method comprising the steps of collecting key parameters in the battery operation process, preprocessing the collected key parameter data, constructing a lithium-ion battery diaphragm service life prediction model, training the lithium-ion battery diaphragm service life prediction model, and testing the lithium-ion battery diaphragm service life prediction model.

[0063] Please refer to Figure 1 As shown, a lithium-ion battery separator service life prediction and optimization method in an embodiment of the present application, the specific steps are as follows:

[0064] S1. Collect key parameters during battery operation, where the key parameters are parameters that directly and indirectly affect the service life of the battery separator.

[0065] Furthermore, in step S1, the parameters that directly and indirectly affect the service life of the battery diaphragm include: the internal temperature of the battery, the voltage across the battery, the current during discharge, the depth of discharge, the number of charge and discharge cycles, the internal resistance of the battery, the ambient humidity and the remaining capacity of the battery, wherein the remaining capacity of the battery is used as the label value of the battery diaphragm life; the collection frequency of temperature, voltage, current, internal resistance and humidity values ​​is set to once every 5 minutes, and the depth of discharge, number of cycles and capacity values ​​are collected once after each charge and discharge cycle is completed.

[0066] S2. Preprocessing the collected key parameter data, wherein the preprocessing includes normalization processing and data time series segmentation operations.

[0067] Furthermore, in step S2, the normalization process is to scale all features to the range of [0,1], wherein normalization eliminates the differences between different feature dimensions, making the model training more stable and improving the performance of the model. The maximum-minimum normalization method is used to process the data, and the maximum-minimum normalization formula is as follows:

[0068] ;

[0069] In the formula, is the normalized data, For the original data collected, is the minimum value of this feature in the data set, is the maximum value of the feature value in the data set; the time series segmentation operation is to convert the normalized data into multiple samples according to the time window, and there is a time series relationship between each sample. Each input matrix The dimension is ,in is the size of the time window, which is 10, that is, each sample contains 10 consecutive time points. is the number of variables in the data.

[0070] S3. Construct a prediction model for the service life of lithium-ion battery separators, such as Figure 2 As shown, the steps are: performing feature conversion on the preprocessed data, proposing a time-frequency mapper module, proposing a frequency domain cross-weighted attention mechanism, proposing an information fusion unit, and introducing a fully connected layer.

[0071] S31, performing feature conversion on the preprocessed data, uniformly converting the data of multiple variables and embedding them into the same feature space.

[0072] Furthermore, in step S3, the feature conversion projects the time series data after the preprocessing operation from its high-dimensional input space to the same feature space through linear embedding, wherein the linear embedding process is expressed as:

[0073] ;

[0074] In the formula, is the output value after linear embedding, is the weight value of the linear embedding operation, is the bias term of the linear embedding operation, is the normalized value, i.e., the input data of feature transformation.

[0075] S32, propose a time-frequency mapper module, the processing of the time-frequency mapper module is divided into a time-frequency decomposer, an extraction of real and virtual states in the frequency domain, and a main frequency component filter. The time-frequency decomposer decomposes the time series data into the frequency domain. The real and virtual states in the frequency domain record the amplitude and phase information of the time series at different frequencies. The main frequency component filter retains the most important previous information for frequency changes. frequency components, enhancing the real state and virtual state in the frequency domain and reducing the redundancy of frequency domain features.

[0076] Furthermore, in the step S3, in S32, the time-frequency decomposer decomposes the time series data into the frequency domain. The time-frequency decomposer is obtained by short-term Fourier transform calculation. The process is to decompose a long time series into multiple overlapping time windows, such as Figure 3 As shown, a short-term Fourier transform is calculated in each window. The time-frequency decomposer records the frequency component changes of the time series in different time periods and captures the instantaneous frequency characteristics. Among them, in the process of decomposing the time series data into the frequency domain, the method of decomposing the long time series into multiple short time windows is the Hanning window method. The time series input by the Hanning window is divided into multiple time windows, and the data in each window is subjected to a short-term Fourier transform calculation to calculate the frequency component. For a given time series data , the short-term Fourier change calculation formula is as follows:

[0077] ;

[0078] In the formula, For the Time series data of variables, is the Hanning window function, which is used to segment time series data. For the The angular frequency corresponding to the frequency component, t is the time step, is a complex exponential form used to calculate the frequency component of each time step, It is the output of the short-term Fourier transform calculation; the result of the short-term Fourier transform calculation is a joint representation of the time domain and the frequency domain, capturing the frequency information of the time series in different time periods.

[0079] Furthermore, in the step S3, in S32, the frequency domain real state and the frequency domain virtual state record the amplitude and phase information of the time series at different frequencies, and the output of the short-term Fourier change includes complex exponential , expressed as:

[0080] ; Therefore, the output of the short-term Fourier change is decomposed into the real state in the frequency domain and the virtual state in the frequency domain. The calculation formula of the real state in the frequency domain is expressed as: ;

[0081] In the formula, For the The time series in Windows and frequencies The real state at the frequency domain is expressed as:

[0082] ;

[0083] In the formula, For the The time series in windows and frequencies The virtual state of being.

[0084] Furthermore, in the steps S3 and S32, the main frequency component filter selects the one with the largest amplitude in each window. frequency components, reducing the amount of calculation while retaining the most important frequency information, amplitude The calculation formula is as follows:

[0085] ;

[0086] According to the calculated amplitude, select Important frequency components:

[0087] ;

[0088] In the formula, The calculated front There are several important frequency components, among which, is the result of real part enhancement in frequency domain, As a result of frequency domain imaginary part enhancement, the main frequency component filter retains the most sensitive to frequency changes. frequency components, reducing the redundancy of frequency domain features and improving computational efficiency.

[0089] Finally, if Figure 3 As shown, the main frequency component filter selects the most important After the frequency components, The real and imaginary parts extracted from the windows are spliced ​​into and , the formulas are:

[0090] ;

[0091] ;

[0092] In the formula, For splicing operation, is the number of windows, is the number of variables in the time series, is the number of most important frequency components selected in each window, is the representation of the real state in the frequency domain, It is the representation of the virtual state in the frequency domain. The representation of the real state and the virtual state in the frequency domain will be used as the input of the subsequent modules.

[0093] Furthermore, in the steps S3 and S32, the time-frequency decomposer can effectively capture complex periodic characteristics, process long-term dependencies, reduce noise and extract key features in the prediction of lithium battery diaphragm life by converting time series data from the time domain to the frequency domain, especially has better recognition ability for the nonlinear attenuation law of battery life. It can also reduce data dimensions and improve computing efficiency by retaining the most important frequency components, thereby improving the prediction accuracy and robustness of the model and making it more adaptable to the changing battery usage environment.

[0094] S33. A frequency domain cross-weighted attention mechanism is proposed. The frequency domain cross-weighted attention mechanism captures the frequency domain characteristics of the time series by establishing a connection between the real part and the imaginary part, and uses the frequency domain cross-weighted attention mechanism to exchange information between the real part and the imaginary part to enhance the frequency domain representation.

[0095] Furthermore, in the steps S3 and S33, Figure 4 As shown, the frequency domain cross-weighted attention mechanism consists of two parts, namely, the real-state attention mechanism enhanced by the virtual state in the frequency domain and the virtual-state attention mechanism enhanced by the real state in the frequency domain;

[0096] Real-state attention mechanism for frequency-domain virtual-state enhancement, such as Figure 4 As shown in the left part, the virtual state is used as a query , virtual state as key , the actual state as value , the formula for enhancing the real-state attention mechanism through frequency-domain virtual-state enhancement is:

[0097] ;

[0098] In the formula, The virtual state query matrix is ​​obtained by linear mapping for the virtual state. for The transpose operation of To calculate the weight of the linear transformation in the query matrix, For the virtual state, we obtain the virtual state bond matrix through linear mapping. is the weight of the linear transformation in the process of calculating the bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of is the weight of the linear transformation in the process of calculating the value matrix, is the dimension of the key matrix, which is used to scale the attention score; the real-state attention mechanism output of the frequency domain virtual-state enhancement performs weighted summation on the real-state values ​​to generate a virtual-state enhanced real-state representation.

[0099] Furthermore, in the steps S3 and S33, the virtual state attention mechanism for frequency domain real state enhancement is as follows: Figure 4 As shown in the right part, the actual state is used as a query , the real state as a key , virtual state as value , the formula for virtual state attention mechanism enhancement through frequency domain real state enhancement is:

[0100] ;

[0101] In the formula, The actual state query matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weight of linear transformation in the actual query matrix, The real state bond matrix is ​​obtained by linear mapping. To calculate the weight of the linear transformation in the process of calculating the real bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weights of the linear transformation in the process of calculating the real-state value matrix, is the dimension of the real-state key matrix, which is used to scale the attention score. The virtual-state attention mechanism output of the frequency-domain real-state enhancement performs weighted summation on the virtual-state values ​​to generate a real-state enhanced virtual-state representation.

[0102] Furthermore, in the steps S3 and S33, the multi-head attention mechanism is finally used to output the real state and virtual state enhanced by the frequency domain cross-weighted attention mechanism, and the enhanced real part is expressed as:

[0103] ;

[0104] In the formula, For the enhanced reality, is the number of attention heads, To calculate the projection matrix used in the enhanced reality process, It is the matrix concatenation operation;

[0105] ;

[0106] In the formula, For the enhanced virtual state, It is the projection matrix used in calculating the enhanced virtual state.

[0107] Furthermore, in the steps S3 and S33, the frequency domain cross-weighted attention mechanism in the prediction of lithium battery diaphragm life enhances the complementarity of real and virtual frequency domain information, captures the interaction of multi-frequency components, improves the attention to key frequencies and the modeling ability of nonlinear relationships. This mechanism helps the model better understand the complex changes in battery life, enhances the extraction of effective frequencies, and reduces noise interference, thereby improving the prediction accuracy and robustness of the model, making it more stable and reliable in complex battery usage scenarios.

[0108] S34. An information fusion unit is proposed to integrate the real and virtual information enhanced by the frequency domain cross-weighted attention mechanism and perform multi-scale feature extraction.

[0109] Furthermore, in the steps S3 and S34, the information fusion unit is composed of a linear projection and a multi-scale information fuser; the linear projection transformation fuses the frequency domain information to make it suitable for the next multi-scale convolution operation. The linear projection performs a linear transformation on the input data through a weight matrix. For the real state and virtual state of the input, the linear projection is expressed in the following form:

[0110] ;

[0111] In the formula, For the enhanced reality, For the enhanced virtual state, is the weight matrix of linear projection, is the bias term of the linear projection, To concatenate the real and imaginary parts, a new joint representation is obtained;

[0112] Then the multi-scale information fusion extracts multiple feature representations from different scales. Specifically, multiple convolution kernels are used to process the input data at multiple scales. The size of the convolution kernel is , , , and concatenate the results. The formula is:

[0113] ;

[0114] In the formula, is the final output value of the multi-scale information fusion device, is the convolution operation, It is a maximum pooling operation used to highlight significant features; the results of each convolution and pooling operation are concatenated The operations are combined together to form a multi-scale joint representation that preserves both local and global feature information.

[0115] S35. Introduce a fully connected layer, use the output of the information fusion unit as the input of the fully connected layer, and finally the output of the fully connected layer is the output value of the final prediction model.

[0116] Furthermore, in the steps S3 and S35, the output of the information fusion unit is After full connection, the final prediction value is output. The calculation formula of the full connection layer is as follows:

[0117] ;

[0118] In the formula, is the final prediction value of the prediction model, and also the result of the output layer. is the weight matrix of the fully connected layer, is the bias term of the fully connected layer.

[0119] S4. Training a lithium-ion battery separator service life prediction model, wherein the training is to set model hyperparameters and parameters during the model training process.

[0120] Furthermore, in step S4, in order to prevent the model from overfitting, The layer is added, and the dropout rate is set to 0.5 to prevent the model from overfitting. In terms of optimizing the model hyperparameters, the learning rate LR is set to 0.001, and the learning rate decay decreases with the decay of the step size. The optimizer uses the Adam optimizer, and the batch size Batch_size is set to 128. The training-related hyperparameters and the number of iterations Epochs are set to 1000. During the training process, the loss function uses the mean square error loss function, and the formula of the mean square error loss function is:

[0121] ;

[0122] In the formula, is the loss value of the mean square error loss function, is the sample size, is the true label value of the jth sample, is the output value of the prediction model for the jth sample; in the process of training the prediction model, as the number of iterations increases, The loss value is getting smaller and smaller, so the prediction accuracy of the prediction model is higher.

[0123] S5. Testing the lithium-ion battery separator service life prediction model, wherein the test is to test the accuracy and performance of the model in a real deployment environment.

[0124] Furthermore, in step S5, the deployed lithium-ion battery separator service life prediction model is tested using test data, such as Figure 5As shown in the figure, the data of the first 60 minutes in the data set is used to predict the remaining life of the battery separator in the next 20 minutes. Figure 5 As shown in Figure 2, the remaining life of the battery separator is gradually decreasing over time, and the prediction model accurately captures this feature. Figure 5 It can be seen from the second figure that the predicted value of the prediction model is roughly the same as the true value, which leads to the conclusion that the lithium-ion battery diaphragm service life prediction model can accurately predict the remaining life of the battery diaphragm in the future through continuous iterative learning; in addition, the algorithm of the present application has the smallest error and higher accuracy compared with the algorithms of the other two models, so the lithium-ion battery diaphragm service life prediction model proposed in the present application can meet actual prediction needs.

[0125] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting and optimizing the service life of a lithium-ion battery separator, characterized in that: The following steps are involved: S1. Collect key parameters during battery operation, which are parameters that directly and indirectly affect the service life of the battery separator, including the internal temperature of the battery, the voltage at both ends of the battery, the current during discharge, the depth of discharge, the number of charge and discharge cycles, the internal resistance of the battery, the ambient humidity, and the remaining capacity of the battery, wherein the remaining capacity of the battery is used as the label value of the battery separator life; S2. Preprocessing the collected key parameter data, wherein the preprocessing includes normalization processing and data time series segmentation operations; S3. Constructing a lithium-ion battery separator service life prediction model, the specific method includes: S31, performing feature conversion on the preprocessed data, uniformly converting the data of multiple variables and embedding them into the same feature space; S32, propose a time-frequency mapper module, the processing of the time-frequency mapper module is divided into a time-frequency decomposer, an extraction of real and virtual states in the frequency domain, and a main frequency component filter. The time-frequency decomposer decomposes the time series data into the frequency domain. The real and virtual states in the frequency domain record the amplitude and phase information of the time series at different frequencies. The main frequency component filter retains the most important previous information for frequency changes. frequency components, enhancing the real state and virtual state in the frequency domain and reducing the redundancy of frequency domain features; S33. A frequency domain cross-weighted attention mechanism is proposed. The frequency domain cross-weighted attention mechanism captures the frequency domain characteristics of the time series by establishing a connection between the real state and the virtual state, and uses the frequency domain cross-weighted attention mechanism to exchange information between the real state and the virtual state to enhance the frequency domain representation; the frequency domain cross-weighted attention mechanism consists of two parts, namely, a real state attention mechanism enhanced by the frequency domain virtual state and a virtual state attention mechanism enhanced by the frequency domain real state; The real state attention mechanism for frequency domain virtual state enhancement, virtual state as query , virtual state as key , the actual state as value , the formula for enhancing the real-state attention mechanism through frequency-domain virtual-state enhancement is: ； In the formula, The virtual state query matrix is ​​obtained by linear mapping for the virtual state. for The transpose operation of To calculate the weight of the linear transformation in the virtual state query matrix, For the virtual state, we obtain the virtual state bond matrix through linear mapping. To calculate the weight of the linear transformation in the process of virtual bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weight of the linear transformation in the process of calculating the virtual state value matrix, is the dimension of the virtual key matrix, used to scale the attention scores; Virtual state attention mechanism for frequency domain real state enhancement, real state as query , the real state as a key , virtual state as value , the formula for virtual state attention mechanism enhancement through frequency domain real state enhancement is: ； In the formula, The actual state query matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weight of linear transformation in the actual query matrix, The real state bond matrix is ​​obtained by linear mapping. To calculate the weight of the linear transformation in the process of calculating the real bond matrix, The real state value matrix is ​​obtained by linear mapping. for The transpose operation of To calculate the weights of the linear transformation in the process of calculating the real-state value matrix, is the dimension of the real key matrix, used to scale the attention scores; Finally, the multi-head attention mechanism is used to output the real state and virtual state enhanced by the frequency domain cross-weighted attention mechanism. The enhanced real part is expressed as: ； In the formula, For the enhanced reality, is the number of attention heads, To calculate the projection matrix used in the enhanced reality process, It is the matrix concatenation operation; ； In the formula, For the enhanced virtual state, The projection matrix used in calculating the enhanced virtual state; S34. An information fusion unit is proposed to integrate the real and virtual information obtained by frequency domain cross-weighted attention mechanism enhancement and perform multi-scale feature extraction; S35, introducing a fully connected layer, using the output of the information fusion unit as the input of the fully connected layer, and finally the output of the fully connected layer is the output value of the final prediction model; S4, training a lithium-ion battery separator service life prediction model, wherein the training is to set model hyperparameters and parameters in the model training process; S5. Testing the lithium-ion battery separator service life prediction model, wherein the test is to test the accuracy and performance of the model in a real deployment environment.

2. A lithium-ion battery separator service life prediction and optimization method according to claim 1, characterized in that: In the step S1, the parameters that directly and indirectly affect the service life of the battery diaphragm include: the internal temperature of the battery, the voltage at both ends of the battery, the current during the discharge process, the depth of discharge, the number of charge and discharge cycles, the internal resistance of the battery, the ambient humidity and the remaining capacity of the battery, wherein the remaining capacity of the battery is used as the label value of the battery diaphragm life; the collection frequency of the temperature, voltage, current, internal resistance and humidity values ​​is set to once every 5 minutes, and the depth of discharge, number of cycles and capacity values ​​are collected once after each charge and discharge cycle is completed.

3. A lithium-ion battery separator service life prediction and optimization method according to claim 2, characterized in that: In the steps S3 and S31, the feature conversion projects the time series data after the preprocessing operation from its high-dimensional input space to the same feature space through linear embedding, wherein the linear embedding process is expressed as: ； In the formula, is the output value after linear embedding, is the weight value of the linear embedding operation, is the bias term of the linear embedding operation, is the normalized value, i.e., the input data of feature transformation.

4. A lithium-ion battery separator service life prediction and optimization method according to claim 3, characterized in that: In step S32, the time-frequency decomposer decomposes the time series data into the frequency domain. The time-frequency decomposer is obtained by short-term Fourier transform calculation. The process is to decompose a long time series into multiple overlapping time windows and perform short-term Fourier transform calculation in each window. The time-frequency decomposer records the frequency component changes of the time series in different time periods and captures the instantaneous frequency characteristics; The method of decomposing the long time series into multiple short time windows is the Hanning window method. The time series input using the Hanning window is divided into multiple time windows, and the data in each window is subjected to short-term Fourier transform calculation to calculate the frequency component. For a given time series data , the short-term Fourier change calculation formula is as follows: ； In the formula, For the Time series data of variables, is the Hanning window function, which is used to segment time series data. For the The angular frequency corresponding to the frequency component, t is the time step, is a complex exponential form used to calculate the frequency component of each time step, It is the output of the short-term Fourier transform calculation; the result of the short-term Fourier transform calculation is a joint representation of the time domain and the frequency domain, capturing the frequency information of the time series in different time periods.

5. A lithium-ion battery separator service life prediction and optimization method according to claim 4, characterized in that: In step S32, the frequency domain real state and frequency domain virtual state record the amplitude and phase information of the time series at different frequencies, and the output of the short-term Fourier transform includes complex exponential , expressed as: ; Therefore, the output of the short-term Fourier change is decomposed into the real state in the frequency domain and the virtual state in the frequency domain. The calculation formula of the real state in the frequency domain is expressed as: ； In the formula, For the The time series in Windows and frequencies The real state at the frequency domain is expressed as: ； In the formula, For the The time series in windows and frequencies The virtual state of being.

6. A lithium-ion battery separator service life prediction and optimization method according to claim 5, characterized in that: In step S32, the main frequency component filter selects the one with the largest amplitude in each window. frequency components, amplitude The calculation formula is as follows: ； According to the calculated amplitude, select Important frequency components: ； In the formula, The calculated front There are several important frequency components, among which, is the result of real part enhancement in frequency domain, It is the result of enhancing the imaginary part in the frequency domain; Finally, in selecting the most important After the frequency components, The real and imaginary parts extracted from the windows are spliced ​​into and , the formulas are: ； ； In the formula, For splicing operation, is the number of windows, is the number of variables in the time series, is the number of most important frequency components selected in each window, is the representation of the real state in the frequency domain, is the representation of the virtual state in the frequency domain.

7. A lithium-ion battery separator service life prediction and optimization method according to claim 6, characterized in that: In step S34, the information fusion unit is composed of a linear projection and a multi-scale information fuser; the linear projection transformation fuses the frequency domain information to make it suitable for the next multi-scale convolution operation. The linear projection performs a linear transformation on the input data through a weight matrix. For the real state and virtual state of the input, the linear projection is expressed in the following form: ； In the formula, For the enhanced reality, For the enhanced virtual state, is the weight matrix of linear projection, is the bias term of the linear projection, To concatenate the real and imaginary parts, a new joint representation is obtained; Then the multi-scale information fusion extracts multiple feature representations from different scales. Specifically, multiple convolution kernels are used to process the input data at multiple scales. The size of the convolution kernel is , , , and concatenate the results. The formula is: ； In the formula, is the final output value of the multi-scale information fusion device, is the convolution operation, It is a maximum pooling operation used to highlight significant features; the results of each convolution and pooling operation are concatenated The operations are combined together to form a multi-scale joint representation while preserving both local and global feature information.

Images