Lithium battery remaining service life prediction method based on data processing and neural network

Abstract

The invention discloses a lithium battery remaining service life prediction method based on data processing and a neural network, and relates to the technical field of lithium battery remaining service life prediction. Comprising the following steps: acquiring original capacity data of a lithium battery; performing smoothing processing on the lithium battery capacity historical data and converting the lithium battery capacity historical data into a low-frequency smooth sequence and a high-frequency fluctuation sequence; obtaining the weight of each sequence relative to the original data; constructing a neural network model; key parameter optimization is carried out according to loss values corresponding to different model parameters by using a particle swarm optimization algorithm; and obtaining the residual charge-discharge cycle number of the lithium battery. According to the method, the original data of the battery capacity is smoothed and converted into the low-frequency smooth data sequence and the high-frequency fluctuation data sequence according to the non-linear change trend of the lithium ion battery capacity, the weight of each sequence relative to the original data is calculated, and the weights are used for optimizing the prediction model parameters; and meanwhile, the prediction result of each sequence is reconstructed by using the weight, so that the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery remaining service life prediction, and in particular to a method for predicting the remaining service life of a lithium battery based on data processing and a neural network. Background Art

[0002] Lithium-ion batteries have become a common power source for electric vehicles due to their high energy density, low self-discharge rate, long cycle life, energy saving and environmental protection. However, as the battery is used for a longer time, the number of times the battery is used and the number of charge and discharge cycles continue to increase, the battery performance will decline, the battery capacity will usually decrease, and the safety will also be greatly reduced. Therefore, it is particularly important and practical to accurately predict the remaining useful life (RUL) of lithium-ion batteries.

[0003] In the prior art, the RUL prediction methods mainly include: (1) model-based prediction methods. The disadvantages of model-based methods are that the battery changes in a complex manner during actual use, the accuracy of the model is easily affected by variable current and temperature, and it is difficult to obtain an accurate mechanism model under the influence of different external conditions.

[0004] (2) Data-driven methods, such as the technical solution disclosed in the patent document with publication number CN 117849645 A, in which the battery operation data is collected to build a database, feature data of different working conditions are extracted to form a training set and a test set, a prediction model is generated by a machine learning algorithm, and the model is optimized by combining knowledge distillation with the LWF algorithm. The technical solution disclosed in the patent document with publication number CN118311437A, in which the battery-related monitoring data is first obtained and feature data is constructed, and then a hybrid neural network prediction model consisting of a feature approximation unit, a bidirectional gated recurrent unit and a fully connected unit is constructed, and an expert knowledge module is designed to embed expert knowledge into the model, thereby realizing the fusion of data information and expert knowledge. For example, the technical solution disclosed in the patent document with publication number CN 117849645 A, in which the features are determined by multiple operating parameters, the maximum available capacity of the battery is predicted by the first prediction module, and the remaining service life of the battery is predicted by the second prediction module. Most of the above existing data-driven methods do not consider the impact of battery capacity regeneration on the remaining service life prediction. Summary of the invention

[0005] The technical problem to be solved by the present invention is: to overcome the deficiencies of the prior art and provide a method for predicting the remaining service life of a lithium battery based on data processing and a neural network, thereby reducing the difficulty of predicting the remaining service life of a lithium battery and improving the prediction accuracy.

[0006] The technical solution adopted by the present invention to solve the technical problem is: the method for predicting the remaining service life of a lithium battery based on data processing and neural network is characterized by comprising the following steps: Step S1, obtaining battery capacity data of a lithium battery at different charge and discharge cycles, and using it as original capacity data, arranging the original capacity data sequence according to the number of charge and discharge cycles, and the data is the original data for training a neural network; Step S2, smoothing the different frequencies of the original capacity data in time according to its own characteristics, and converting the original capacity data into a low-frequency smooth sequence and a high-frequency fluctuation sequence; Step S3, taking each sequence as a characteristic variable and the original capacity data as a response variable, and performing correlation weight calculation to obtain the weight of each sequence relative to the original data; Step S4, using a long short-term memory neural network, introducing an encoder-decoder structure, and combining an attention mechanism to build a neural network model; Step S5, using a particle swarm optimization algorithm to optimize key parameters based on the corresponding loss values ​​under different model parameters; Step S6, inputting the hourly segment data of each sequence into the prediction model for prediction, obtaining the predicted values ​​of all sequences, and then weighting and reconstructing the predicted values ​​of each sequence using their respective weights to obtain the final estimated value, thereby knowing the remaining number of charge and discharge cycles.

[0007] Further, step S2 includes the following steps: Step S2-1, find all the maximum points in the original capacity data, and use the interpolation method to fit them to form the upper envelope of the original capacity data, denoted as En max (t); Similarly, find all the minimum points in the original capacity data, and use the interpolation method to fit them to form the lower envelope of the original capacity data, denoted as En min (t); Step S2-2, calculating the weighted mean of the upper envelope and the lower envelope, and obtaining the function c1(t); Step S2-3, subtract function c1(t) from the original capacity data to obtain a new function without high-frequency components, and repeat the process of obtaining c1(t) until the smooth trend part of the remaining data is extracted, and the original capacity data is converted into a low-frequency smooth sequence and several high-frequency fluctuation sequences.

[0008] Furthermore, in step S2-2, the weighted mean of the upper envelope and the lower envelope is calculated, and the calculation formula is: ; in, Represents the weight of the i-th extreme point; The calculation is based on the distance between the extreme point and the local trend, and different weights are assigned according to the location information; the envelope mean m(t) is subtracted from the original data. If the sum of the number of local maxima and local minima differs from the number of zero-crossing points by two or more or the envelope mean of the local maximum value and the local minimum value is not zero, the above steps are repeated until the conditions are met to obtain the function c1(t).

[0009] Further, step S3 includes the following steps: Step S3-1: For each sequence, the importance of the component is quantified by calculating its energy. The calculation formula is: ; Among them, X i is the value of the i-th sequence at the n-th time point, and N is the total number of samples; Step S3-2, calculate the weight of each sequence relative to the original capacity data, the calculation formula is: ; Among them, W i is the weight of the ith sequence.

[0010] Further, step S4 includes the following steps: Step S4-1, using LSTM neural network instead of RNN network to implement encoder-decoder structure; Step S4-2 introduces a multi-head attention mechanism, where each attention head calculates its weight independently and concatenates them at the end.

[0011] Further, when executing step S4-2: The attention mechanism works by assigning different weights to the input data: ; in, represents the attention mechanism, d k represents the number of columns of the K matrix, i.e., the dimension of the vector, Q, K, and V represent the query, key, and value matrices respectively; The multi-head attention mechanism calculates attention in multiple subspaces in parallel: ; Enables the model to simultaneously capture and integrate contextual information from different aspects; Multihead represents the multi-head attention mechanism, W G represents the trainable weight matrix, and p represents the number of heads.

[0012] Furthermore, in step S5, the number of particles in the group is set to 4, and the maximum number of iterations of the algorithm is 16; the number of iterations of the model and the number of multi-head attention heads are selected for parameter optimization, the upper and lower bounds of the number of model iterations are set to 70 and 150 respectively, and the upper and lower bounds of the number of multi-head attention heads are set to 2 and 8 respectively, and the final neural network prediction model is obtained after optimization.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The lithium battery remaining service life prediction method based on data processing and neural network of the present application, in view of the nonlinear change trend of lithium-ion battery capacity, smoothes the original data of battery capacity and converts it into a low-frequency smooth data sequence and a high-frequency fluctuating data sequence, calculates the weight of each sequence relative to the original data, and reconstructs the prediction results of each sequence using the weight, thereby reducing the difficulty of predicting the remaining service life of the lithium battery and improving the prediction accuracy.

[0014] The original capacity data of the battery is converted into a low-frequency smooth sequence and a high-frequency fluctuation sequence after smoothing. Different sequences have different degrees of expression of the original data. The low-frequency smooth sequence can reflect the overall degradation trend of the original data, so it can represent the original data to a large extent, but the high-frequency fluctuation data also contains the local feature information of the original data and cannot be completely discarded. The weights of each sequence are obtained by correlation calculation. The low-frequency smooth sequence is a trend item, so the weight is fixed to 1. The weight of each high-frequency fluctuation sequence calculated in this way, the prediction results obtained by inputting each sequence into the prediction model are reconstructed by this weight to obtain the final prediction result.

[0015] The parameter optimization method of the neural network model constructed for multiple sequences uses the particle swarm optimization algorithm to compare the low-frequency smooth sequence with a fixed weight of 1, and optimizes the model parameters according to the error indicators corresponding to different model parameters. In the neural network model, the parameters to be optimized are the number of iterations of the neural network model and the number of heads of multi-head attention. The appropriate combination of the two parameters can improve the accuracy of model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the principle of the method for predicting the remaining service life of lithium batteries based on data processing and neural network; Figure 2 The flowchart of the method for predicting the remaining service life of lithium batteries based on data processing and neural network is as follows; Figure 3 Schematic diagram of decoder-encoder long short-term memory network structure; Figure 4 Schematic diagram of the multi-head attention layer structure; Figure 5This is the effect diagram of sequence 1 data; Figure 6 This is the effect diagram of sequence 2 data; Figure 7 This is the effect diagram of sequence 3 data; Figure 8 This is the data effect diagram of sequence 4; Fig. 9 It is a comparison curve diagram of the fitting effect of predicted data and actual data. DETAILED DESCRIPTION

[0017] Figure 1 to Figure 9 The best embodiment of the present invention is shown below in conjunction with the attached Figure 1 to Figure 9 The present invention is further described.

[0018] 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.

[0019] like Figure 1 As shown, in the method for predicting the remaining service life of a lithium battery based on data processing and neural network of the present application, the original capacity data sequence is converted into a low-frequency smooth data sequence and a high-frequency fluctuating data sequence, and the weight of each sequence relative to the original data is calculated; a long short-term memory neural network is used, an encoder-decoder structure is introduced, and a neural network model is constructed in combination with an attention mechanism, and the parameters of the model are optimized to obtain a different neural network prediction model for each sequence; each sequence data is input into the prediction model to obtain a predicted value, and the weight of each sequence is used for weighted reconstruction to obtain the final predicted value of the battery capacity, thereby knowing the remaining number of charge and discharge cycles before the battery reaches the end of its life.

[0020] The lithium battery remaining service life prediction method based on data processing and neural network of the present application, in view of the nonlinear change trend of lithium-ion battery capacity, smoothes the original data of battery capacity and converts it into a low-frequency smooth data sequence and a high-frequency fluctuating data sequence, calculates the weight of each sequence relative to the original data, and reconstructs the prediction results of each sequence using the weight, thereby improving the accuracy of the remaining service life prediction of the lithium battery.

[0021] Combination Figure 2 The method for predicting the remaining service life of a lithium battery based on data processing and neural network in this application is described in detail: Step S1, obtaining the original capacity data of the lithium battery; The battery capacity data of the lithium battery in different charge and discharge cycles are obtained and used as the original capacity data. The original capacity data sequence is arranged according to the number of charge and discharge cycles. This data is the original data for training the neural network.

[0022] Step S2, smoothing the historical data of lithium battery capacity and converting it into a low-frequency smoothing sequence and a high-frequency fluctuation sequence; The original capacity data is smoothed according to its own characteristics at different frequencies in time, and the original capacity data is converted into a low-frequency smooth sequence and a high-frequency fluctuation sequence, so as to reduce the impact of the high-frequency dynamic and nonlinear change trend of the battery capacity on the prediction, including the following steps: Step S2-1, find all the maximum points in the original capacity data, and use the interpolation method to fit them to form the upper envelope of the original capacity data, denoted as En max (t); Similarly, find all the minimum points in the original capacity data, and use the interpolation method to fit them to form the lower envelope of the original capacity data, denoted as En min (t).

[0023] Step S2-2, calculate the weighted mean of the upper envelope and the lower envelope, the calculation formula is: ; in, Represents the weight of the i-th extreme point.

[0024] The calculation is based on the distance between the extreme point and the local trend, and different weights are assigned according to the location information. The envelope mean m(t) is subtracted from the original data. If the sum of the number of local maxima and local minima differs from the number of zero crossings by two or more or the envelope mean of the local maximum and local minimum is not zero, the above steps are repeated until the conditions are met to obtain the function c1(t).

[0025] Step S2-3, subtract c1(t) from the original capacity data to obtain a new function that removes the high-frequency components, and repeat the process of obtaining c1(t) until the smooth trend part of the remaining data is extracted, and the original capacity data is converted into a low-frequency smooth sequence and several high-frequency fluctuation sequences.

[0026] Step S3, taking each sequence as a characteristic variable and the original capacity data as a response variable, and performing correlation weight calculation to obtain the weight of each sequence relative to the original data; The specific steps are as follows: Step S3-1: For each sequence, the importance of the component is quantified by calculating its energy. The calculation formula is: ; Among them, X iis the value of the i-th sequence at the n-th time point, and N is the total number of samples.

[0027] Step S3-2, calculate the weight of each sequence relative to the original capacity data, the calculation formula is: ; Among them, W i is the weight of the ith sequence.

[0028] In step S3, the original battery capacity data is converted into a low-frequency smooth sequence and several high-frequency fluctuation sequences according to the different frequencies exhibited in time by the characteristics of the original data itself. The low-frequency smooth sequence reflects the overall degradation trend of the original data, while the high-frequency fluctuation sequence represents the local oscillation characteristics of the original data.

[0029] The original capacity data of the battery is converted into a low-frequency smooth sequence and a high-frequency fluctuation sequence after smoothing. Different sequences have different degrees of expression of the original data. The low-frequency smooth sequence can reflect the overall degradation trend of the original data, so it can represent the original data to a large extent, but the high-frequency fluctuation data also contains the local feature information of the original data and cannot be completely discarded. The weights of each sequence are obtained by correlation calculation. The low-frequency smooth sequence is a trend item, so the weight is fixed to 1. The weight of each high-frequency fluctuation sequence calculated in this way, the prediction results obtained by inputting each sequence into the prediction model are reconstructed by this weight to obtain the final prediction result.

[0030] Step S4: Use the long short-term memory neural network, introduce the encoder-decoder structure, and combine the attention mechanism to build a neural network model. The specific steps are as follows: Step S4-1, use LSTM neural network instead of RNN network to realize encoder-decoder structure. ED-LSTM structure is shown in the attached figure. Figure 3 shown.

[0031] The encoder-decoder structure is a deep learning model structure that significantly improves the ability of neural networks to process sequence data by combining global context and dynamic information selection. The basic idea is to use a recursive neural network RNN ​​as an encoder to encode the input sequence into a fixed-length vector, which can be regarded as an abstract representation of the entire input sequence. This vector is then used as the initial input of another RNN acting as a decoder, and outputs a target sequence of arbitrary length. However, in time series prediction tasks, the input sequence is too long and contains more information, and the RNN cannot make good use of the characteristics of the data, which will have a greater impact on the final prediction results.

[0032] Step S4-2 introduces a multi-head attention mechanism, where each attention head calculates its weight independently and concatenates them at the end.

[0033] In order to more effectively capture the global features of the sequence, a multi-head attention mechanism is introduced. Each attention head calculates its weight independently and concatenates them at the end. The attention mechanism assigns different weights to the input data: ; in, represents the attention mechanism, d k represents the number of columns of the K matrix, i.e., the dimension of the vector, Q, K, and V represent the query, key, and value matrices, respectively.

[0034] The multi-head attention mechanism calculates attention in multiple subspaces in parallel: ; The model can simultaneously capture and integrate different aspects of context information. Figure 4 As shown. Multihead represents the multi-head attention mechanism, W G represents the trainable weight matrix, and p represents the number of heads.

[0035] Step S5, using a particle swarm optimization algorithm to optimize key parameters based on the corresponding loss values ​​under different model parameters; Each component is used as input to train the neural network model, and the particle swarm optimization algorithm is used to compare the loss values ​​corresponding to the components with the highest importance weight under different model parameters to optimize the key parameters. The core mechanism of the particle swarm optimization algorithm is to gradually approach the global optimal solution by iteratively updating the speed and position of the particles: The speed update formula is: ; The position update formula is: ; In the PSO optimizer, considering the computational cost, the number of particles in the group is set to 4, and the maximum number of iterations of the algorithm is set to 16. The number of model iterations and the number of multi-head attention heads are selected for parameter optimization. The upper and lower bounds of the number of model iterations are set to 70 and 150 respectively, and the upper and lower bounds of the number of multi-head attention heads are set to 2 and 8 respectively. After optimization, the final neural network prediction model is obtained.

[0036] The parameter optimization method of the neural network model constructed for multiple sequences uses the particle swarm optimization algorithm to compare the low-frequency smooth sequence with a fixed weight of 1, and optimizes the model parameters according to the error indicators corresponding to different model parameters. In the neural network model, the parameters to be optimized are the number of iterations of the neural network model and the number of heads of multi-head attention. The appropriate combination of the two parameters can improve the accuracy of model prediction.

[0037] Step S6, obtaining the remaining number of charge and discharge cycles of the lithium battery; The hourly segment data of each sequence is input into the prediction model for prediction to obtain the predicted values ​​of all sequences. Then the predicted values ​​of each sequence are weighted and reconstructed using their respective weights to obtain the final estimated value, thereby knowing the remaining number of charge and discharge cycles.

[0038] In the method for predicting the remaining service life of lithium batteries based on data processing and neural networks of the present application, the original capacity data is smoothed to obtain multiple sequence data, and different neural network models are trained; small sequence data is taken and input into the trained neural network model to obtain the capacity prediction value under each subsequent charge and discharge cycle number of the battery; small sequence data is a small part of the sequence sequence data taken as the input of the prediction model. According to the end-of-life capacity threshold of the lithium battery, the number of charge and discharge cycles at the end of the life of the battery to be predicted is obtained, thereby obtaining the remaining charge and discharge times of the battery.

[0039] The following is an explanation of the implementation process of this solution in combination with NASA's public battery data set: The battery reaches the end of its life when the rated capacity drops to 70% of the initial capacity, that is, when the rated capacity of the four batteries drops from 2 Ah to 1.4 Ah, it can be considered that its life has ended. Taking battery B0006 as an example, this data set contains the battery capacity data of 168 charge and discharge cycles of battery B0006; some historical data of the battery capacity are shown in Table 1: Table 1 Some historical data of battery capacity

[0040] During the implementation of this solution, the hourly segment data is taken from each sequence of the first 8 cycles of battery capacity conversion.

[0041] During the experiment, the original capacity data was smoothed according to its own characteristics and converted into low-frequency smooth sequences and high-frequency fluctuation sequences; the battery capacity data was converted into four function sequences, of which sequence 4 was a low-frequency smooth sequence, and the other sequences 1 to 3 were high-frequency fluctuation sequences. The results are shown in the attached figure. Figure 5~Figure 8 The data of each sequence are shown in Table 2.

[0042] Table 2 Data of each sequence

[0043] Each sequence data 1-4 is used as a characteristic variable, and the original data is used as the response variable to calculate the importance weight. Sequence 4 is a low-frequency trend item, so the weight is fixed to 1. The weight coefficient of each high-frequency fluctuation sequence calculated in this way is shown in Table 3. And the fit degree of the four battery capacity data conversion sequences relative to the original data is all above 99%.

[0044] Table 3 Calculation results of each sequence weight

[0045] Using long short-term memory neural network, the encoder-decoder structure is introduced, and the neural network model is constructed in combination with the multi-head attention mechanism. Each component is used as input to train the neural network model.

[0046] Parameter optimization of the neural network model: The particle swarm optimization algorithm is used to optimize the model parameters by comparing the loss values ​​of the training process corresponding to the component IMF4 under different model parameters.

[0047] Within the given parameter search range, the optimal parameter combinations of each battery data set searched by PSO are shown in Table 4.

[0048] Table 4 Optimal parameter combinations for each battery data set

[0049] In the prediction stage, the hourly segment data of each sequence is input into the prediction model for prediction, and the predicted values ​​of 168 charge and discharge cycles of each sequence are obtained. Then, the predicted values ​​of each sequence are weighted and reconstructed to obtain the final predicted value. Some actual data and predicted data are shown in Table 5, so as to know the remaining number of charge and discharge cycles.

[0050] Table 5 Some actual and predicted data of battery capacity

[0051] The predicted data and actual data were evaluated using three different error evaluation indicators. The error evaluation results are shown in Table 6. The fitting effect diagram of the predicted data and actual data is shown in Fig. 9 shown.

[0052] Table 6 Error evaluation index results

[0053] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0054] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting the remaining service life of lithium batteries based on data processing and neural networks, characterized in that: The steps include: Step S1, obtaining battery capacity data of a lithium battery at different charge and discharge cycles, and using it as original capacity data, arranging the original capacity data sequence according to the number of charge and discharge cycles, and the data is the original data for training a neural network; Step S2, smoothing the different frequencies of the original capacity data in time according to its own characteristics, and converting the original capacity data into a low-frequency smooth sequence and a high-frequency fluctuation sequence; Step S3, taking each sequence as a characteristic variable and the original capacity data as a response variable, and performing correlation weight calculation to obtain the weight of each sequence relative to the original data; Step S4, using a long short-term memory neural network, introducing an encoder-decoder structure, and combining an attention mechanism to build a neural network model; Step S5, using a particle swarm optimization algorithm to optimize key parameters based on the corresponding loss values ​​under different model parameters; Step S6, inputting the hourly segment data of each sequence into the prediction model for prediction, obtaining the predicted values ​​of all sequences, and then weighting and reconstructing the predicted values ​​of each sequence using their respective weights to obtain the final estimated value, thereby knowing the remaining number of charge and discharge cycles.

2. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 1, characterized in that: Step S2 includes the following steps: Step S2-1, find all the maximum points in the original capacity data, and use the interpolation method to fit them to form the upper envelope of the original capacity data, denoted as En max (t); Similarly, find all the minimum points in the original capacity data, and use the interpolation method to fit them to form the lower envelope of the original capacity data, denoted as En min (t); Step S2-2, calculating the weighted mean of the upper envelope and the lower envelope, and obtaining the function c1(t); Step S2-3, subtract function c1(t) from the original capacity data to obtain a new function without high-frequency components, and repeat the process of obtaining c1(t) until the smooth trend part of the remaining data is extracted, and the original capacity data is converted into a low-frequency smooth sequence and several high-frequency fluctuation sequences.

3. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 2, characterized in that: In step S2-2, the weighted mean of the upper envelope and the lower envelope is calculated using the following formula: ； in, Represents the weight of the i-th extreme point; The calculation is based on the distance between the extreme point and the local trend, and different weights are assigned according to the location information; the envelope mean m(t) is subtracted from the original data. If the sum of the number of local maxima and local minima differs from the number of zero-crossing points by two or more or the envelope mean of the local maximum value and the local minimum value is not zero, the above steps are repeated until the conditions are met to obtain the function c1(t).

4. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 2, characterized in that: Step S3 includes the following steps: Step S3-1: For each sequence, the importance of the component is quantified by calculating its energy. The calculation formula is: ； Among them, X i is the value of the i-th sequence at the n-th time point, and N is the total number of samples; Step S3-2, calculate the weight of each sequence relative to the original capacity data, the calculation formula is: ； Among them, W i is the weight of the ith sequence.

5. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 2, characterized in that: Step S4 includes the following steps: Step S4-1, using LSTM neural network instead of RNN network to implement encoder-decoder structure; Step S4-2 introduces a multi-head attention mechanism, where each attention head calculates its weight independently and concatenates them at the end.

6. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 5, characterized in that: When executing step S4-2: The attention mechanism works by assigning different weights to the input data: ； in, represents the attention mechanism, d k represents the number of columns of the K matrix, i.e., the dimension of the vector, Q, K, and V represent the query, key, and value matrices respectively; The multi-head attention mechanism calculates attention in multiple subspaces in parallel: ； Enables the model to simultaneously capture and integrate contextual information from different aspects; Multihead represents the multi-head attention mechanism, W G represents the trainable weight matrix, and p represents the number of heads.

7. The method for predicting the remaining service life of a lithium battery based on data processing and neural network according to claim 5, characterized in that: In step S5, the number of particles in the group is set to 4, and the maximum number of iterations of the algorithm is 16; the number of iterations of the model and the number of multi-head attention heads are selected for parameter optimization, the upper and lower bounds of the number of model iterations are set to 70 and 150 respectively, and the upper and lower bounds of the number of multi-head attention heads are set to 2 and 8 respectively. After optimization, the final neural network prediction model is obtained.

Images