Lithium battery SOC prediction method and system based on multi-sensor data fusion

Abstract

The invention discloses a lithium battery SOC prediction method and system based on multi-sensor data fusion, and the method comprises the steps: collecting the historical charging and discharging use data of a lithium battery, extracting a health factor, and collecting the internal air pressure and optical fiber temperature data of the lithium battery; sequentially decomposing the collected data by using pole symmetry mode decomposition and ensemble empirical mode decomposition; meanwhile, a Spearman rank correlation coefficient is selected in the module to carry out correlation analysis on lithium battery health factors, decomposed data, original data and the health factors jointly form a matrix, and the matrix serves as input data of the SOC; then, on the basis of a Transform model, a Series Station and a De-Station Attention are introduced, a variant model NSTransforms is obtained, and at the same time, an IETO is used to carry out optimization on a parameter combination of the NSTransforms; and finally, the prediction and evaluation module is used for inputting the processed data into the optimized and improved NSTransformers model. According to the method, the SOC of the lithium battery is accurately predicted to obtain the prediction result, and the lithium battery is maintained and managed according to the prediction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of lithium batteries, and particularly relates to a lithium battery SOC prediction method and system based on multi-sensor data fusion. Background Art

[0002] The performance of lithium batteries will gradually degrade with the increase of usage time, which not only affects the normal operation of equipment but also may pose potential safety hazards. Traditional lithium battery state monitoring and life prediction methods have some limitations. First, when dealing with lithium battery data, these methods often rely on a single data source and simple feature extraction techniques, which limit the sensitivity of the model to battery state changes and the accuracy of prediction. Second, the charge and discharge data of lithium batteries are highly nonlinear and dynamic, making it difficult for traditional prediction models to capture these complex changes. In addition, the working environment of lithium batteries varies, including environmental factors such as temperature and pressure, and the impact of these factors on battery performance has not been fully considered. Summary of the Invention

[0003] Object of the Invention: Aiming at the problems existing in the prior art, the present invention provides a lithium battery SOC prediction method and system based on multi-sensor data fusion, which can improve the accuracy of the lithium battery life prediction model and more accurately predict the remaining service life of the lithium battery.

[0004] Technical Solution: A lithium battery SOC prediction method based on multi-sensor data fusion, characterized by comprising the following steps:

[0005] (1) Collect the usage data of the lithium battery, the air pressure data inside the battery, the fiber optic temperature data, and the health factor of the lithium battery;

[0006] (2) Use extreme point symmetric mode decomposition (ESMD), ensemble empirical mode decomposition (EEMD), and Spearman rank correlation coefficient to process the data, and construct a multi-modal input matrix with the processed data;

[0007] (3) On the basis of the Transformers model, add the Series Stationarization and De-stationary Attention modules to obtain the improved model NSTransformers;

[0008] (4) Improve the triangular optimization algorithm (ETO) by using a progressive storage optimization mechanism to obtain the IETO algorithm, and use the IETO algorithm to optimize the parameter combination of the NSTransformers model;

[0009] (5) Input the multi-modal input matrix into the NSTransformers model after parameter optimization for prediction, and based on the prediction results, give opinions on the daily maintenance and management of lithium batteries.

[0010] Further, the implementation process of step (1) is as follows:

[0011] Collect the usage data during the charge and discharge cycles of the lithium battery, including voltage, current, and temperature, extract the health factors during the charge and discharge processes, including constant current charge and discharge time, constant voltage charge and discharge time, and use sensors to collect the air pressure data and fiber optic temperature data of the lithium battery itself.

[0012] Further, step (2) includes using the extreme point symmetric mode decomposition ESMD to decompose the collected lithium battery usage data, the air pressure data inside the battery, and the fiber optic temperature data, decomposing the complex data into multiple components, and then using the ensemble empirical mode decomposition EEMD to further process the high-frequency components; at the same time, use the Spearman rank correlation coefficient to perform a correlation analysis on the health factors of the lithium battery and select the health factors with high correlation.

[0013] Further, the process of using the extreme point symmetric mode decomposition ESMD to decompose the collected lithium battery usage data, the air pressure data inside the battery, and the fiber optic temperature data is as follows:

[0014] (31) Calculate all the extreme points of the lithium battery data LB data and record them in sequence as JZ i (i = 1, 2, 3,..., n);

[0015] (32) Use line segments to connect adjacent extreme points JZ i , and record the midpoints of the line segments between every two extreme points in sequence as XZ i (i = 1, 2, 3,..., n - 1);

[0016] (33) Use the method of linear interpolation to record the points on the left boundary and the right boundary as BJ l and BJ r ;

[0017] (34) Use n + 1 midpoints to construct p interpolation curves L1, L2,... L p (p ≥ 1), and an average curve L * =(L1 + L2 +... + L p ) / p can be obtained;

[0018] (35) Calculate LB data -L *, and repeat steps (31) to (34) until the screening times reach the pre-set maximum value K, obtaining the first mode component IMF1;

[0019] (36) Calculate LB data -IMF n , and repeat steps (31) to (35) to obtain IMF1, IMF2,..., IMF n , until the number of extreme points of the last residual mode component CM(t) does not exceed the set number;

[0020] (37) The maximum screening times K vary within the integer interval [K min , K max , and repeat steps (31) to (36) to obtain a series of decomposition results, and then calculate the variance ratio σ / σ0, where σ is the relative standard deviation of LB data -CM(t), and σ0 is the standard deviation of the original data LB data ;

[0021] (37) Find the maximum screening times K0 corresponding to the minimum variance ratio, V = σ / σ0. At this time, the residual mode component CM(t) is the best fitting curve of the data, and repeat steps (31) to (36) to output the decomposition results.

[0022] By applying the pole symmetric mode decomposition technology, the collected data of lithium battery charging and discharging, air pressure data and optical fiber temperature data are decomposed, which improves the model's recognition ability for sequences and helps to improve the prediction accuracy.

[0023] Furthermore, the process of further processing the high-frequency components using the ensemble empirical mode decomposition EEMD is as follows:

[0024] (41) On the basis of EMD, add white noise following the normal distribution;

[0025] (42) Construct the decomposition signal Feng jie ,

[0026] Feng jie =[Feng jie1 , Feng jie2 , Feng jie3 ,…Feng jiei

[0027] Feng jiei =feng jie (n)+SZ i (n)

[0028] In the formula: SZ i ​(n) is a random noise that satisfies N(0,1), feng jie (n) is the original data, and i represents the number of times of adding Gaussian white noise;

[0029] (43) Perform EEMD decomposition on Feng jie to obtain the corresponding IMFs ij ; IMF ij represents the j-th IMF obtained after decomposing Feng jiei ;

[0030] (44) Perform ensemble averaging on the IMFs ij to obtain the final IMF s and the residual component,

[0031]

[0032] where EEMD j (Feng jiei ) represents the j-th IMF generated by EEMD;

[0033] (45) The original signal can be expressed as:

[0034]

[0035] Finally, N IMF components and the residual component Feng can (n) can be obtained;

[0036] Combine the high-frequency components after EEMD decomposition with the low-frequency components after ESMD decomposition in matrix form, and add the data that has not been decomposed to form the prediction input matrix of the lithium battery SOC.

[0037] The process of using the Spearman rank correlation coefficient to perform correlation analysis on the health factors of the lithium battery and selecting the health factors with high correlation is as follows:

[0038] Use the Spearman rank correlation coefficient method to calculate the correlation coefficients of the health factors, and select the health factors with the absolute value of the correlation coefficient between 0.8 and 1.

[0039] The implementation process of the lithium battery SOC dataset in step (2) is as follows:

[0040] (61) For multiple components of various usage data of the lithium battery obtained through ESMD decomposition and EEMD decomposition, arrange different data and the corresponding multiple components to form the prediction input matrix of the lithium battery SOC;

[0041] (62) Arrange the health factors selected after Spearman rank correlation coefficient analysis and the prediction input matrix of the lithium battery SOC to form the final multi-modal input matrix of the lithium battery SOC.

[0042] Furthermore, the IETO algorithm introduces a progressive storage optimization mechanism during the development stage of the triangular optimization algorithm ETO. The progressive storage optimization mechanism includes: at the initial stage of the search, introducing Gaussian perturbation or Brownian motion to perturb the current optimal individual to find a better individual; in the middle stage of the search, using the external archive technology ((External Archive)) to dynamically store non-dominated solutions for comparison and selection in subsequent iterations; in the later stage of the search, using the covariance matrix adaptation evolution strategy (CMA-ES) to optimize the solution. If the generated solution is better than the worst individual in the current population, accept the solution and incorporate it into the population.

[0043] After the ETO algorithm enters the development stage, it is difficult to get rid of the local optimum, which affects the convergence accuracy and speed and exacerbates the difficulty of searching for the global optimum. Therefore, a progressive storage optimization mechanism strategy is introduced to solve this problem.

[0044] Furthermore, the improved NSTransformers model in step (3) is as follows: NSTransformers adopts an encoder-decoder architecture and is based on the Transformer model for time series prediction. The encoder part is responsible for extracting key information from historical data, and the decoder integrates this information to optimize future predictions. Replace the Self-Attention in Transformer with De-stationary Attention and integrate Series Stationarization in the input and output stages. The terms in the Softmax function are converted into non-stationary factors τ and Δ to re-integrate non-stationary information. These improvements enable the model to more effectively process non-stationary sequences, thereby improving the accuracy of time series prediction.

[0045] Furthermore, step (5) uses the IETO algorithm to optimize the parameter combination of the NSTransformers model. The specific steps are as follows:

[0046] Input the training set of the lithium battery into the NSTransformers model, optimize the learning rate, Top_k, D_model, D_ff, Batch_size, and Dropout parameters of the model, calculate the prediction metrics RMSE and R of the model, and finally generate a new combined objective function Obj to determine whether this set of parameters has the best prediction performance. The calculation formulas of the metrics are as follows:

[0047]

[0048] Obj = 0.1RMSE + 0.9R

[0049] where P ri (Y c ) is the predicted value of the c-th sample; O b (Y c ) is the observed value of the c-th training sample; and are the average values of the predicted value and the observed value respectively, and Θ is the total number of samples.

[0050] A lithium battery SOC prediction system based on multi-sensor data fusion, comprising:

[0051] A data acquisition module that extracts the usage data and health factors of the lithium battery during charging and discharging through multiple sensors, and simultaneously collects the internal air pressure data and fiber optic temperature data of the battery using a barometric pressure sensor and a fiber optic sensor;

[0052] A data processing module that processes the lithium battery data and health factors using extreme point symmetric mode decomposition (ESMD), ensemble empirical mode decomposition (EEMD), and Spearman rank correlation coefficient;

[0053] A model and algorithm improvement module that, based on the Transformers model, adds the SeriesStationarization and De-stationary Attention modules to obtain the NSTransformers model; improves the triangular optimization algorithm (ETO) using a progressive storage optimization mechanism to obtain the IETO algorithm; and uses the IETO algorithm to optimize the parameters of the NSTransformers model;

[0054] A prediction and evaluation module that, based on the final prediction result, evaluates the SOC of the lithium battery, takes corresponding measures according to the evaluation level, and provides support for subsequent lithium battery maintenance;

[0055] The data processing module uses the extreme point symmetric mode decomposition (ESMD) to decompose the collected lithium battery usage data, internal air pressure data, and fiber optic temperature data into multiple components, and then uses the ensemble empirical mode decomposition (EEMD) to further process the high-frequency components; at the same time, uses the Spearman rank correlation coefficient to analyze the correlation of the health factors of the lithium battery and selects the health factors with high correlation.

[0056] Furthermore, in the development stage of the triangular optimization algorithm ETO, the IETO algorithm introduces a progressive storage optimization mechanism, which includes: in the initial stage of the search, Gaussian perturbation or Brownian motion is introduced to perturb the current optimal individual to find a better individual; in the middle stage of the search, the external archive technology is used to dynamically store non-dominated solutions; in the later stage of the search, the covariance matrix adaptation evolution strategy is used to optimize the solutions; if the generated solution is better than the worst individual in the current population, then accept the solution and incorporate it into the population.

[0057] Beneficial effects: In view of the characteristics of the correlation and large amount of data of the lithium battery capacitance data, the prediction accuracy of the prediction model will be interfered. Therefore, the collected historical lithium battery data is first decomposed by the extreme point symmetric mode decomposition (ESMD), and then the ensemble empirical mode decomposition (EEMD) is introduced to perform a secondary decomposition on the decomposed high-frequency components, further reducing the component complexity and improving the model prediction efficiency. Immediately afterwards, the Spearman rank correlation coefficient (SPCC) is used to screen the health factors, and the health factors closely related to the state of charge of the lithium battery are selected to enhance the prediction accuracy of the model for the state of charge data.

[0058] Aiming at the limitations of the traditional triangular optimization algorithm, the progressive storage optimization mechanism strategy is used to improve the ETO to help find the global optimum, obtain the IETO algorithm, improve the global search ability and achieve a better convergence effect.

[0059] Based on the Transformers model, this paper introduces the Series Stationarization and De-stationary Attention modules to obtain the variant model NSTransformers, and uses the IETO algorithm to optimize the parameter combinations when the NSTransformers model predicts different lithium battery data, find the model parameter combinations most suitable for the data, improve the performance of the model, enable the model to better fit the lithium battery data, improve the generalization ability of the model, and improve the accuracy and reliability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is the flowchart of the present invention;

[0061] Figure 2 is the schematic flow chart of the improved ETO algorithm (IETO) provided by the present invention;

[0062] Figure 3 is the NSTransformers model structure diagram;

[0063] Figure 4 are the evaluation indexes of each model;

[0064] Figure 5 This is a comparison chart of the true and predicted values of each model. Specific implementation mode

[0065] The present invention will be further described in detail below with reference to the accompanying drawings.

[0066] As Figure 1 shown, this paper proposes a lithium battery SOC prediction method and system based on multi-sensor data fusion. First, while the data acquisition module collects the historical charge and discharge usage data of the lithium battery and extracts the health factors, it uses a barometric pressure sensor and an optical fiber sensor to collect the internal barometric pressure and optical fiber temperature data of the lithium battery itself. Then, the data processing module uses the extreme point symmetric mode decomposition (ESMD) to decompose the collected data in sequence, decomposing the complex data into multiple components, and continues to use the ensemble empirical mode decomposition (EEMD) to further process the complex high-frequency components to reduce the data complexity. At the same time, in this module, the Spearman rank correlation coefficient (SPCC) is selected to perform a correlation analysis on the lithium battery health factors, and data with a higher degree of correlation is selected for prediction. The decomposed data, the original data, and the health factors are jointly formed into a matrix as the input data of the SOC. Then, the model and algorithm improvement module selects to use the progressive storage optimization mechanism to improve the triangular optimization algorithm (ETO) to obtain the improved triangular optimization algorithm (IETO). Then, based on the Transformers model, SeriesStationarization and De-stationary Attention are introduced to obtain the variant model NSTransformers, and at the same time, IETO is used to optimize the parameter combination of NSTransformers. Finally, the prediction and evaluation module inputs the processed data into the optimized and improved NSTransformers model to accurately predict the lithium battery SOC to obtain the prediction result, and makes maintenance and management of the lithium battery according to the prediction result. The specific steps are as follows:

[0067] Step 1: Collect the usage data during the charge and discharge cycles of the lithium battery, such as voltage, current, temperature, etc., extract the health factors during the charge and discharge processes, such as constant current charge and discharge time, constant voltage charge and discharge time, etc., and use sensors to collect the barometric pressure data and optical fiber temperature data of the lithium battery itself.

[0068] Step 2: Decompose the usage data, barometric pressure data, and optical fiber temperature data collected in step (1). The specific steps are as follows:

[0069] (21) Calculate all the extreme points of the lithium battery data LB data to be decomposed, and record them in sequence as JZ i (i = 1, 2, 3,..., n);

[0070] (22) Connect adjacent extreme points JZ with line segments i , and successively denote the midpoints of the line segments between every two extreme points as XZ i (i = 1, 2, 3,..., n - 1);

[0071] (23) Use the method of linear interpolation to denote the midpoints of the left and right boundaries as BJ l and BJ r ;

[0072] (24) Use n + 1 midpoints to construct p interpolation curves L1, L2,... L p (p ≥ 1), and an average curve L * =(L1 + L2 +... + L p ) / p;

[0073] (25) Calculate LB data -L * , and repeat steps (31) - (34) until the screening times reach the pre - set maximum value K to obtain the first modal component IMF1;

[0074] (26) Calculate LB data -IMF n , and repeat steps (31) - (35) to successively obtain IMF1, IMF2,..., IMF n , until the number of extreme points of the last residual modal component CM(t) does not exceed a certain number;

[0075] (27) The maximum screening times K vary within the integer interval [K min , K max , and repeat steps (31) - (36) to obtain a series of decomposition results, and then calculate the variance ratio σ / σ0, where σ is the relative standard deviation of LB data -CM(t), and σ0 is the standard deviation of the original data LB data ;

[0076] (28) Find the maximum screening times K0 corresponding to the minimum variance ratio, V = σ / σ0. At this time, the residual modal component CM(t) is the best - fitting curve of the data. Repeat steps (31) - (36) and output the decomposition results.

[0077] By applying the extreme - point symmetric modal decomposition technique, the collected charging, discharging usage data, air pressure data, and optical fiber temperature data of lithium batteries are decomposed, which improves the model's recognition ability for sequences and helps to improve the prediction accuracy.

[0078] Step 3: Select EEMD decomposition to perform secondary decomposition on the high-frequency components obtained after the decomposition in Step 2. The specific steps are as follows:

[0079] (31) On the basis of EMD, add white noise following a normal distribution;

[0080] (32) Construct the decomposition signal Feng jie .

[0081] Feng jie = [Feng jie1 , Feng jie2 , Feng jie3 , … Feng jiei

[0082] Feng jiei = feng jie (n) + SZ i (n)

[0083] where: SZ i (n) is the random noise satisfying N(0,1), feng jie (n) is the original data, and i represents the number of times of adding Gaussian white noise;

[0084] (33) Perform EEMD decomposition on Feng jie to obtain the corresponding IMF ij ; IMF ij represents the jth IMF obtained after decomposing Feng jiei .

[0085] (34) Perform ensemble averaging on IMF ij to obtain the final IMF s and the residual component.

[0086]

[0087] where, EEMD j (Feng jiei ) represents the jth IMF generated by EEMD;

[0088] (35) The original signal can be expressed as:

[0089]

[0090] Finally, N IMF components and the residual component Feng can (n) can be obtained;

[0091] The components obtained after EEMD decomposition are combined with the low-frequency components in a matrix, and the data that has not been decomposed is added to form the prediction input matrix of the lithium battery SOC.

[0092] Step 4: The implementation process of selecting the SPCC to analyze the correlation between the health factor and the lithium battery charge state is as follows:

[0093] When performing Spearman rank correlation analysis, multiple sets of observation data of two variables SPCC X and SPCC Y are usually collected. These data can be expressed as (SPCC Xi -SPCC Yi ), where the range of i is from 1 to n;

[0094] The Pearson correlation coefficient is actually obtained by calculating the ratio of the sample covariance to the sample standard deviation, and its mathematical formula is expressed as follows:

[0095]

[0096] In the formula, and respectively represent the average values of the variables in n experiments;

[0097] The value range of the correlation coefficient |SPCC| is between -1 and 1. If |SPCC| is greater than 0, it means there is a positive correlation between the two variables; if |SPCC| is less than 0, it means there is a negative correlation between them. The larger the absolute value of |SPCC|, the stronger the correlation between the two variables; if |SPCC| is equal to 0, it means there is no linear correlation between the two variables;

[0098] Using the Pearson correlation coefficient method to calculate the correlation coefficient of the health factor and selecting the health factors with the absolute value of the correlation coefficient between 0.8 and 1 can further improve the prediction accuracy of the lithium battery SOC.

[0099] Step 5: The process of constructing the lithium battery SOC prediction data set is as follows:

[0100] (51) Multiple components of various usage data of the lithium battery obtained through ESMD decomposition and EEMD decomposition are arranged with different data and the corresponding multiple components to form a new data set;

[0101] (52) The health factors selected after SPCC analysis are arranged with the data set in the previous step to form the final lithium battery SOC prediction data set.

[0102] Step 6: Select the triangular optimization algorithm and improve it. The specific steps are as follows:

[0103] (61) A bounded exploration technique is introduced in the ETO algorithm. This method helps optimize the search process, reduce the consumption of time and computing resources, and ensure that the optimal solution is not overlooked. The formula is as follows:

[0104] CETO i+1 = floor[2 - 2×t×(Max ter - CETO i ×TZ a )] + CETO i

[0105]

[0106] In the formula, t represents the current iteration number, Max ter corresponds to the total number of iterations, floor is a function for rounding operations in MATLAB, TZ a and TZ b are two adjustment coefficients; CETO i+1 represents the iteration count for starting the current and subsequent constraint exploration methods.

[0107] After adopting the constraint search strategy, this method simultaneously updates the upper and lower bounds of the search space. The formula is as follows:

[0108]

[0109] Among them, Up i and Low i represent the upper and lower bounds of the expected search space respectively. rand1 and rand2 are random numbers between 0 and 1; represents the j - th position of the optimal solution obtained so far; X j represents the position of the sub - optimal solution at the j - th index.

[0110] (62) Initialization phase

[0111] The initialization formula is as follows:

[0112] I i,j = (Up j - Low j )×rand() + Low j

[0113] Among them, rand() represents a random value in the interval [0, 1]. The j - th dimension is bounded by Up j and Low j which represent the upper and lower bounds respectively, i = 1, 2,..., N and j = 1, 2,..., d.

[0114] (63) Exploration phase

[0115] During the optimization process, the exploration stage is divided into two stages, and the transition formula between these two stages is as follows:

[0116]

[0117] In the first exploration stage, the update formula for the individual position is as follows:

[0118]

[0119] The j-th position of the optimal solution is represented by , and are used to represent the j-th position of the i-th solution in the current iteration and the subsequent iteration, respectively;

[0120] In the second exploration stage, the update formula for the individual position is as follows:

[0121]

[0122] In the formula, q2 represents a random number within the range of [0, 1].

[0123] (64) Development stage

[0124] The development stage is divided into two stages, and the specific stages are as follows:

[0125] In the first development stage, the update formula for the individual position is as follows:

[0126]

[0127]

[0128] In the formula, q3 and q4 represent random numbers within the range of [0, 1];

[0129] In the second development stage, the update formula for the individual position is as follows:

[0130]

[0131]

[0132] In the formula, the coefficient c is determined by a clever combination of the exponential function and the trigonometric function, which is a unique and effective mechanism;

[0133] (65) Progressive storage optimization mechanism

[0134] After the ETO algorithm enters the development stage, it is difficult to get rid of the local optimum, which affects the convergence accuracy and speed and makes it more difficult to search for the global optimum value. Therefore, a progressive storage optimization mechanism strategy is introduced to solve this problem.

[0135] By repeatedly executing the optimization mechanism, the ability to escape from local optima is enhanced, and the overall solution efficiency and accuracy are improved.

[0136] Step 7: Introduce two components to improve the Transformers model to obtain the NSTransformers model. The specific steps are as follows:

[0137] (71) Normalization module: To weaken the non-stationarity of each input sequence, normalize the time dimension through a sliding window. For each input sequence perform translation and scaling operations to obtain where SL and NV represent the sequence length and the number of variables respectively. The formula of the normalization module is as follows:

[0138]

[0139] where, is the element-wise division operation, and ⊙ is the element-wise multiplication operation.

[0140] (72) Inverse normalization module: After the basic model GC predicts the future values of length n, the output sequence will be obtained. Then, this output is de-normalized with σ Is and μ Is to obtain the final prediction result The two-point expression of the inverse normalization process is:

[0141]

[0142] (73) In the analysis of the flat model, over-stationarization will eliminate the inherent non-stationary information in the data, resulting in the model being unable to effectively capture the complex dependencies of the time series. Therefore, the original non-stationary series can be used for learning. The formula is as follows:

[0143]

[0144] In the above formula, Q, K, and V are all vectors, their lengths are all Le, and their dimensions are Di. The Softmax operation is performed row by row on these vectors to capture the time dependence.

[0145] After normalization, the input received by the model is where μ Is and σ Is are the mean and standard deviation of the sequence Is respectively. Based on the linear assumption, the attention layer outputs the transformed query vector The vectors K' and V' are also transformed accordingly.

[0146] In the case where sequence stationarization is not performed, the input to Softmax in self-attention should be In the de-stationarized attention mechanism, the input to Softmax is calculated based on Q′ and K′, and the formula is as follows:

[0147]

[0148] Repeating the operation, since Softmax is invariant to the same translation in the row dimension of the input, the equation is as follows:

[0149]

[0150] Starting from the original sequence Is, the direct expression of the Softmax function in the self-attention mechanism can be derived, that is This expression depends on Q and K obtained from the stationary sequence Is, and also involves the non-stationary information removed through the sequence stationary process.

[0151] The calculation formula of de-stationarized attention is as follows:

[0152] logτ = MLP(σ Is , Is), Δ = MLP(μ Is , Is)

[0153] (74) NSTransformers adopts an encoder-decoder architecture and performs time series prediction based on the Transformer model. The encoder part is responsible for extracting key information from historical data, while the decoder integrates this information to optimize future predictions. To improve the model's ability to process non-stationary sequences, the Self-Attention in the classical non-stationary Transformer is replaced by the proposed De-stationary Attention, and SeriesStationarization is integrated at the input and output stages. For different variants of Transformer, the terms in the Softmax function are converted into non-stationary factors τ and Δ to re-integrate non-stationary information. These improvements enable the model to process non-stationary sequences more effectively, thereby improving the accuracy of time series prediction.

[0154] Step 8: Select the IETO algorithm to optimize the parameter combination of the NSTransformers model. The specific steps are as follows

[0155] Input the training set of lithium batteries into the NSTransformers model, optimize the key prediction parameters of the model such as learning rate, Top_k, D_model, D_ff, Batch_size, Dropout, etc., calculate the prediction metrics RMSE and R of the model, and finally generate a new combined objective function Obj to determine whether this set of parameters has the best prediction performance. The calculation formulas of the metrics are as follows:

[0156]

[0157] Obj = 0.1RMSE + 0.9R

[0158] Among them, P ri (Y c ) is the predicted value of the c-th sample; O b (Y c ) is the observed value of the c-th training sample; and are the average values of the predicted values and the observed values respectively, and Θ is the total number of samples.

[0159] Step 8: The specific process of the prediction and evaluation module is as follows:

[0160] According to the predicted value of the lithium battery state of charge obtained, divide the safety of the lithium battery into four levels:

[0161] Level 1 safety: SOC ∈ [0.8, 1]: Good performance, long remaining service life, no immediate maintenance required.

[0162] Level 2 safety: SOC ∈ [0.4, 0.8]: Stable performance, medium remaining service life, regular inspection recommended.

[0163] Level 3 safety: SOC ∈ [0.2, 0.4]: Performance begins to degrade, short remaining service life, planned maintenance recommended.

[0164] Level 4 safety: SOC ∈ [0, 0.2]: Performance significantly degraded, extremely short remaining service life, immediate maintenance required.

[0165] Maintenance time point and content suggestions:

[0166] Level 1 safety: Monitor performance, no maintenance required.

[0167] Level 2 safety: Arrange for inspection within the next operating cycle, including visual inspection and basic functional tests.

[0168] Level 3 safety: Develop a detailed maintenance plan, including cleaning, component replacement and performance testing.

[0169] Level 4 Safety: Immediately implement emergency maintenance measures to avoid potential failures and downtimes.

[0170] This module automatically determines and alerts the need for necessary preventive measures using these safety assessment levels. Meanwhile, the system provides a detailed assessment report, recording performance changes and recommended maintenance activities, providing data support for operation and maintenance decisions.

[0171] Step 9: The modules of the system are introduced as follows:

[0172] Data processing module, which processes lithium battery data and health factors using Ensemble Empirical Mode Decomposition (EEMD), Extreme Symmetric Mode Decomposition (ESMD), and Spearman's rank correlation coefficient;

[0173] Model and algorithm improvement module, which improves the Enhanced Triangular Optimization (ETO) algorithm using the reverse learning initialization strategy and progressive storage optimization mechanism. Meanwhile, based on the Transformers model, the Series Stationarization and De-stationary Attention modules are added, and after two improvements, the Improved ETO (IETO) is used to optimize the parameters of the NSTransformers model;

[0174] Prediction and evaluation module, the system evaluates the state of charge of the lithium battery according to the final prediction results, and takes corresponding measures according to the evaluation level, providing support for subsequent lithium battery maintenance.

[0175] The historical charge and discharge usage data, internal air pressure, and fiber optic temperature data of a certain type of lithium battery are collected, and the hybrid model is compared and analyzed with other models. Table 1 shows the evaluation indicators of each model.

[0176] Table 1 Statistical table of the result performance indicators of the model of the present invention and the control group models

[0177]

[0178] To more intuitively display the error indicator values, Figure 4 The evaluation indicators of 4 models corresponding to the collected lithium battery data are shown.

[0179] From Table 1 and Figure 4 it can be concluded that the proposed ESMD-EEMD-SPCC-IETO-NSTransformers model performs best compared to other basic models and has superior non-linear fitting ability. Through the comparison between models, it shows that the decomposition technology and algorithm optimization can effectively optimize the parameter selection of the NSTransformers model, thereby improving the prediction accuracy.

[0180] To more intuitively display the prediction results of the prediction model of the present invention and further verify the effectiveness of the proposed model, Figure 5 The line graph comparison of the SOC prediction results and the true values of the ESMD-EEMD-SPCC-IETO-NSTransformers model and the basic model is given. It can be clearly seen from the figure that the predicted values are basically in line with the actual values, fully proving the effectiveness of the model proposed by the present invention.

[0181] This application proposes a method and system for predicting the State of Charge (SOC) of a lithium battery based on multi-sensor data fusion. This method improves the prediction accuracy and robustness by integrating data collected by multiple sensors, including the air pressure data and fiber optic temperature data inside the battery, as well as the usage data and health factors of the battery. In addition, this method uses the Extreme Symmetric Mode Decomposition (ESMD) and Ensemble Empirical Mode Decomposition (EEMD) techniques to deeply analyze and process the collected data to reduce the data complexity and extract key features. At the same time, the Spearman Rank Correlation Coefficient (SPCC) is introduced to perform a correlation analysis on the health factors, and the factors with a higher correlation with the battery state are selected to further improve the performance of the prediction model.

[0182] This application also proposes an improved triangular optimization algorithm (IETO), and a variant model NSTransformers based on the Transformers model. By introducing the Series Stationarization and De-stationaryAttention modules, the ability of the model to process non-stationary sequences is enhanced. These innovative points enable this method to not only accurately predict the SOC of the lithium battery, but also provide suggestions for daily maintenance and management based on the prediction results, providing an effective tool for the health management of lithium batteries.

Claims

1. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion, characterized in that, It includes the following steps: (1) Collect the usage data of the lithium battery, the air pressure data inside the battery, the fiber optic temperature data, and the health factor of the lithium battery; (2) Use the Extreme Symmetric Mode Decomposition (ESMD), Ensemble Empirical Mode Decomposition (EEMD), and Spearman's rank correlation coefficient to process the data, and construct a multi-modal input matrix with the processed data; (3) Based on the Transformers model, add the Series Stationarization and De-stationary Attention modules to obtain the improved model NSTransformers; (4) Improve the Triangular Optimization Algorithm (ETO) using the Progressive Storage Optimization Mechanism to obtain the IETO algorithm, and use the IETO algorithm to optimize the parameter combination of the NSTransformers model; (5) Input the multi-modal input matrix into the NSTransformers model with optimized parameters for prediction, and based on the prediction results, give opinions on the daily maintenance and management of the lithium battery.

2. The method for predicting the SOC of a lithium battery based on multi-sensor data fusion according to claim 1, wherein, The implementation process of step (1) is as follows: Collect the usage data during the charge and discharge cycles of the lithium battery, including voltage, current, and temperature, extract the health factors during the charge and discharge processes, including constant current charge and discharge time, constant voltage charge and discharge time, and use sensors to collect the air pressure data and fiber optic temperature data of the lithium battery itself.

3. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 1, characterized in that, Step (2) includes using the Extreme Symmetric Mode Decomposition (ESMD) to decompose the collected usage data of the lithium battery, the air pressure data inside the battery, and the fiber optic temperature data, decomposing the complex data into multiple components, and then using the Ensemble Empirical Mode Decomposition (EEMD) to further process the high-frequency components among them; at the same time, use the Spearman's rank correlation coefficient to perform a correlation analysis on the health factors of the lithium battery and select the health factors with high correlation.

4. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 3, characterized in that, The process of using the Extreme Symmetric Mode Decomposition (ESMD) to decompose the collected usage data of the lithium battery, the air pressure data inside the battery, and the fiber optic temperature data is as follows: (31) Calculate the lithium battery data LB to be decomposed data for all extreme points, and record them in sequence as JZ i (i = 1, 2, 3,..., n); (32) Connect adjacent extreme points JZ with line segments i , and successively denote the midpoints of the line segments between every two extreme points as XZ i (i = 1, 2, 3,..., n - 1); (33) Denote the points on the left boundary and the points on the right boundary as BJ by using the method of linear interpolation l and BJ r ; (34) Construct \(p\) interpolation curves \(L_1, L_2, \cdots, L_p\) (\(p\geq1\)) using \(n + 1\) midpoints, and an average curve \(L\) can be obtained p \(L=\frac{L_1 + L_2+\cdots+L_p}{p}\); * p ​ (35) Calculate LB data -L * , and repeat steps (31) to (34) until the screening times reach the pre-set maximum value K to obtain the first modal component IMF1; (36) Calculate LB data -IMF n , and repeat steps (31) to (35) to obtain IMF1, IMF2,..., IMF n in sequence until the number of extreme points of the last residual modal component CM(t) does not exceed the set number; (37) The maximum number of screening times K varies within the integer interval [K min , K max , and steps (31) to (36) are repeated to obtain a series of decomposition results, and then the variance ratio σ / σ0 is calculated, where σ is the relative standard deviation of LB data -CM(t), and σ0 is the standard deviation of the original data LB data ; (37) Find the maximum screening times K0 when the variance ratio is the smallest, V = σ / σ0. At this time, the residual modal component CM(t) is the best fitting curve of the data. Repeat steps (31) to (36) to output the decomposition result.

5. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 3, characterized in that, The process of using the Ensemble Empirical Mode Decomposition (EEMD) to further process the high-frequency components among them is as follows: (41) Based on the EMD, add white noise that follows a normal distribution; (42) Structural decomposition signal Feng jie , Feng jie = [Feng jie1 , Feng jie2 , Feng jie3 , … Feng jiei ​ Feng jiei = feng jie (n) + SZ i (n) Where: SZ i (n) is a random noise that satisfies N(0,1), feng jie (n) is the original data, and i represents the number of times of adding Gaussian white noise; (43)Perform EEMD decomposition on Feng jie to obtain the corresponding IMFs ij ; IMF ij represents the j-th IMF obtained after decomposing Feng jiei ​ (44) Ensemble average of the IMF ij to obtain the final IMF s and the residual component where EEMD j (Feng jiei ) represents the j-th IMF generated by EEMD; (45) The original signal can be expressed as: Finally, N IMF components and the residual component Feng can be obtained. can (n); Combine the high-frequency components after EEMD decomposition with the low-frequency components after ESMD decomposition, and add the data that has not been decomposed to form the prediction input matrix of the lithium battery's State of Charge (SOC).

6. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 1, characterized in that, The IETO algorithm introduces a progressive storage optimization mechanism during the development stage of the triangular optimization algorithm ETO. The progressive storage optimization mechanism includes: at the initial stage of the search, Gaussian perturbation or Brownian motion is introduced to perturb the current optimal individual to find a better individual; at the middle stage of the search, an external archive technique is used to dynamically store non-dominated solutions; at the later stage of the search, the covariance matrix adaptation evolution strategy is used to optimize the solutions; if the generated solution is better than the worst individual in the current population, then accept the solution and incorporate it into the population.

7. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 1, characterized in that, The NSTransformers model improved based on Transformers in step (3) is as follows: NSTransformers adopts an encoder-decoder architecture and is based on the Transformer model for time series prediction. The encoder part is responsible for extracting key information from historical data, while the decoder integrates this information to optimize future predictions; Self-Attention in Transformer is replaced by De-stationary Attention, and SeriesStationarization is integrated at the input and output stages; the terms in the Softmax function are converted into non-stationary factors τ and Δ.

8. A method for predicting the state of charge (SOC) of a lithium battery based on multi-sensor data fusion according to claim 1, characterized in that, Step (5) uses the IETO algorithm to optimize the parameter combination of the NSTransformers model. The specific steps are as follows: Input the training set of lithium batteries into the NSTransformers model, optimize the learning rate, Top_k, D_model, D_ff, Batch_size, and Dropout parameters of the model, calculate the prediction metrics RMSE and R of the model, and finally generate a new combined objective function Obj to determine whether this set of parameters has the best prediction performance. The calculation formulas of the metrics are as follows: Obj = 0.1RMSE + 0.9R Among them, P ri (Y c ) is the predicted value of the c-th sample; O b (Y c ) is the observed value of the c-th training sample; and are the average values of the predicted value and the observed value respectively, and Θ is the total number of samples.

9. A lithium battery SOC prediction system based on multi-sensor data fusion, characterized in that Including: A data acquisition module that extracts the usage data and health factors of lithium batteries during charging and discharging through various sensors, and at the same time uses a barometric pressure sensor and an optical fiber sensor to collect the barometric pressure data and optical fiber temperature data inside the battery; A data processing module that processes lithium battery data and health factors using extreme point symmetric mode decomposition ESMD, ensemble empirical mode decomposition EEMD, and Spearman rank correlation coefficient; A model and algorithm improvement module that, based on the Transformers model, adds Series Stationarization and De-stationary Attention modules to obtain the NSTransformers model; improves the triangular optimization algorithm ETO using a progressive storage optimization mechanism to obtain the IETO algorithm; uses the IETO algorithm to optimize the parameters of the NSTransformers model; A prediction and evaluation module. The system evaluates the SOC of lithium batteries according to the final prediction results, and takes corresponding measures according to the evaluation level to provide support for the subsequent maintenance of lithium batteries; The data processing module uses Ensemble Symmetric Mode Decomposition (ESMD) to decompose the collected data on the usage of lithium batteries, the internal air pressure data of the batteries, and the optical fiber temperature data, decomposing the complex data into multiple components, and then uses Ensemble Empirical Mode Decomposition (EEMD) to further process the high-frequency components among them; at the same time, the Spearman rank correlation coefficient is used to perform a correlation analysis on the health factors of the lithium batteries, and the health factors with high correlation are selected.

10. A lithium battery SOC prediction system based on multi-sensor data fusion according to claim 9, characterized in that, The IETO algorithm introduces a progressive storage optimization mechanism in the development stage of the Triangular Optimization Algorithm (ETO). The progressive storage optimization mechanism includes: in the initial stage of the search, Gaussian perturbation or Brownian motion is introduced to perturb the current optimal individual to find a better individual; in the middle stage of the search, the external archive technology is used to dynamically store non-dominated solutions; in the later stage of the search, the covariance matrix adaptation evolution strategy is used to optimize the solutions; if the generated solution is better than the worst individual in the current population, then accept the solution and incorporate it into the population.

Images