Lithium ion battery RUL prediction method based on multi-feature fusion dimension reduction and MESA-TCCT model

Abstract

The invention discloses a lithium ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and an MESA-TCCT model. The method comprises the steps of obtaining multi-dimensional features and capacity values of a lithium ion battery; drawing a capacity degradation curve according to the capacity value; measuring the correlation degree of each feature and the capacity degradation curve through a Pearson coefficient, and reserving the features of which the Pearson coefficients meet a preset standard; the reserved features are preprocessed, and then normalization processing is carried out on the preprocessed features; a KPCA-KLPP algorithm is used to carry out fusion dimension reduction on the normalized features; and constructing an MESA-TCCT model for predicting the remaining service life of the lithium ion battery. According to the method, the KPCA-KLPP algorithm is introduced to carry out fusion and dimension reduction on the multiple features, and the MESA-TCCT model network structure design is further combined, so that the characterization capability of the degradation features and the prediction efficiency of the model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to a lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model, belonging to the technical field of lithium-ion battery status monitoring. Background Art

[0002] Lithium-ion batteries are core energy storage devices in the new energy sector, and their remaining useful life (RUL) prediction is crucial for safe operation and maintenance decisions in systems such as electric vehicles and smart grids. Currently, data-driven RUL prediction methods mainly rely on battery capacity decay characteristics. However, a single capacity feature is susceptible to measurement noise and cannot fully reflect the complex multimodal aging mechanisms within the battery (such as active lithium loss and electrode material phase transitions). In addition, existing methods often use linear models (such as PCA) for feature fusion and dimensionality reduction. These methods cannot effectively handle the nonlinear correlations between multi-source features, resulting in the loss of key degradation information, which in turn affects the quality of model input.

[0003] In terms of model construction, while traditional time series prediction models (such as LSTM) can capture short-term dependencies, they have limited ability to model long-term degradation trends and the dynamic correlations of multi-scale features. Existing solutions combining attention mechanisms with convolutional networks (such as Informer) often struggle to balance the expression of global degradation trends and local mutation features due to excessively rapid sequence length compression or loss of local details.

[0004] To address the above problems, there is an urgent need for a dimensionality reduction method that integrates multi-dimensional degradation features and adapts to nonlinear data structures, and combines it with a prediction framework with efficient time series modeling capabilities to achieve high-precision prediction of the RUL of lithium-ion batteries. Summary of the Invention

[0005] The present invention provides a lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model. By introducing the KPCA-KLPP algorithm to fuse multiple features for dimensionality reduction, and further combining it with the MESA-TCCT model network structure design, the characterization ability of degradation features and the prediction efficiency of the model are significantly improved.

[0006] The technical solution of the present invention is:

[0007] According to a first aspect of the present invention, a method for predicting the RUL of a lithium-ion battery based on multi-feature fusion dimensionality reduction and the MESA-TCCT model is provided, comprising the following steps:

[0008] S1. Obtain multi-dimensional characteristics and capacity values of the lithium-ion battery; and draw a capacity degradation curve based on the capacity values;

[0009] S2. Measure the correlation between each feature and the capacity degradation curve using the Pearson coefficient, and retain the features whose Pearson coefficient meets the preset criteria;

[0010] S3, preprocessing the retained features, and then normalizing the preprocessed features; using the KPCA-KLPP algorithm to fuse and reduce the dimension of the normalized features;

[0011] S4. Construct the MESA-TCCT model to predict the remaining service life of lithium-ion batteries.

[0012] Furthermore, the multi-dimensional features include constant current charging time, coulombic efficiency, IC curve peak and peak position voltage.

[0013] Furthermore, the KPCA-KLPP algorithm is specifically:

[0014] S3-1, use Gaussian kernel to map the normalized features to high dimension;

[0015] S3-2, using KPCA to analyze the kernel matrix K h Centralize and centralize the kernel matrix K c Decompose to obtain the projection vector α and eigenvalue λ:

[0016] S3-3. Use KLPP to construct the adjacency graph and calculate the weight matrix and Laplace matrix;

[0017] S3-4. Simultaneous kernel matrix K h , the projection matrix A, Laplace matrix L, and projection vector eigenvalue λ formed by the projection vector are used to solve the generalized characteristic equation:

[0018] S3-5, based on the kernel matrix K h , projecting the fused features as the projection matrix into the low-dimensional space to obtain the features after dimensionality reduction.

[0019] Furthermore, the MESA-TCCT model constructed in step S4 is specifically as follows: based on the informer model, an information entropy sparse attention mechanism is introduced into the knowledge distillation block of the encoder instead of the probability sparse attention mechanism. The information entropy sparse attention mechanism completes pruning by selecting query vectors with high information entropy and optimizes the distribution of attention weights; and an extended causal convolution is introduced into the knowledge distillation block of the encoder instead of the standard convolution to expand the receptive field of the convolution kernel on the time series; a sequentially connected passthrough mechanism and a linear connection layer are introduced into the output of the knowledge distillation block of each scale, and the feature maps of different scales are spliced through the passthrough mechanism to obtain finer-grained information; a mask-information entropy sparse attention mechanism is introduced into the decoder instead of the mask-probability sparse attention mechanism.

[0020] Furthermore, the information entropy sparse attention mechanism is specifically:

[0021] S4-1-1. For the input data, a linear transformation layer is used to obtain the corresponding query matrix Q, key matrix K and value matrix V;

[0022] S4-1-2, for the i-th query vector Q in the query matrix i , randomly sample m keys from the key matrix set, and get γ={K1,…,K m}, calculate the query vector Q i Scaled dot product with the sampling key

[0023]

[0024] Where, d k is the query vector Q i The sequence length, K j is the jth sampling key, j = 1, 2, ..., m;

[0025] S4-1-3. Calculate the scaled dot product through the softmax function The probability distribution of , and calculate the information entropy of each query vector:

[0026]

[0027] Where H i is the i-th query vector Q i Information entropy of

[0028] S4-1-4. Select the first q query vectors with the largest information entropy to construct a sparse query matrix Q M ; Based on the sparse query matrix Q M , the softmax function is used to calculate the sparse attention score matrix, and the sparse attention score matrix is aggregated with the value matrix to obtain a weighted sum vector, which is the final attention score output:

[0029]

[0030] Where, is the sparse attention score matrix, T is the transpose; Output is the final attention score of the output.

[0031] According to a second aspect of the present invention, a lithium-ion battery RUL prediction system based on multi-feature fusion dimensionality reduction and MESA-TCCT model is provided, including a module of any one of the lithium-ion battery RUL prediction methods based on multi-feature fusion dimensionality reduction and MESA-TCCT model described above.

[0032] According to a third aspect of the present invention, a processor is provided, which is used to run a program, wherein when the program is running, any of the above-mentioned lithium-ion battery RUL prediction methods based on multi-feature fusion dimensionality reduction and MESA-TCCT model is executed.

[0033] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, which includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned lithium-ion battery RUL prediction methods based on multi-feature fusion dimensionality reduction and MESA-TCCT model.

[0034] The beneficial effects of the present invention are:

[0035] First, the present invention proposes a KPCA-KLPP dimensionality reduction method, which combines the global feature extraction of kernel principal component analysis (KPCA) with the local structure preservation ability of kernel locality preserving projection (KLPP) to achieve nonlinear feature fusion and redundancy elimination, and output low-dimensional and high-discriminative features.

[0036] Second, the proposed MESA-TCCT model introduces the information entropy sparse attention mechanism (MESA), which dynamically filters high-contribution queries based on information entropy, constructs a sparse attention matrix, and optimizes the weight distribution of key features.

[0037] Third, the proposed MESA-TCCT model introduces an extended causal convolution mechanism, which compresses the sequence length through causal convolution layers with increasing dilation rates and maximum pooling layers, expands the temporal receptive field, and enhances the gradient propagation capability.

[0038] Fourth, the MESA-TCCT model proposed in this paper introduces a multi-scale passthrough mechanism, which takes into account both fine-grained local information and global degradation trends by fusing feature maps of self-attention blocks at different levels.

[0039] In summary, the present invention solves the problems of feature redundancy and large computational complexity in RUL prediction by introducing the KPCA-KLPP algorithm to fuse multiple features and combine it with the MESA-TCCT model network structure design, enhances the feature characterization ability and the network model's ability to capture key features, and improves the prediction accuracy of battery RUL. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the overall flow chart of the present invention.

[0041] Figure 2 This is the result of multi-feature fusion dimensionality reduction based on the KPCA-KLPP method in the present invention.

[0042] Figure 3 This is the overall structure diagram of the MESA-TCCT model of the present invention.

[0043] Figure 4 This is a comparison chart of the prediction results of the method of the present invention and other methods. DETAILED DESCRIPTION

[0044] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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. It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other in any way.

[0045] Example 1: Figure 1-Figure 4 As shown in FIG, a lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model includes the following steps:

[0046] S1. Obtain multi-dimensional characteristics and capacity values of lithium-ion batteries through experimental testing; draw a capacity degradation curve based on the capacity values; the multi-dimensional characteristics include constant current charging time, coulombic efficiency, IC curve peak value and peak position voltage;

[0047] Specifically, during the charge and discharge process, the constant current charging time, charge and discharge energy ratio, voltage, and capacity value of the lithium-ion battery are obtained; further, the coulombic efficiency in step S1 is the charge and discharge energy ratio, the charging energy is the sum of the constant current charging energy and the constant voltage charging energy, and the discharge energy is the constant current discharge energy; the IC curve is obtained by differentiating the capacity and voltage data of the charging process.

[0048] S2. Measure the correlation between each feature and the capacity degradation curve using the Pearson coefficient, and retain the features whose Pearson coefficient meets the preset criteria;

[0049] Furthermore, in step S2, the Pearson coefficient range is [-1, 1], where a coefficient close to -1 indicates a complete negative correlation, close to +1 indicates a complete positive correlation, and 0 indicates no correlation. In this embodiment of the present invention, features with an absolute value of the Pearson coefficient greater than 0.95 are retained as features.

[0050] S3, preprocessing the retained features, and then normalizing the preprocessed features; using the KPCA-KLPP algorithm to fuse and reduce the dimension of the normalized features;

[0051] Furthermore, the KPCA-KLPP algorithm is specifically:

[0052] S3-1, use Gaussian kernel to map the normalized features to high dimension;

[0053] K h =exp(-φ||XY|| 2 )

[0054] Where K h is the kernel matrix, φ is the Gaussian kernel width, X and Y are sample vectors of two different features.

[0055] S3-2, using KPCA to analyze the kernel matrix K h Centralize and centralize the kernel matrix K c Decompose to obtain the projection vector α and eigenvalue λ:

[0056] (1) Kernel matrix K h Centralization:

[0057]

[0058] Where n is the number of samples, 1 n is an n×n matrix of all 1s, K c is the centralized kernel matrix;

[0059] (2) Decomposition of the central kernel matrix K c Get the projection vector:

[0060] K c α=λα

[0061] Where α is the projection vector obtained after decomposition, and λ is the eigenvalue corresponding to α.

[0062] S3-3. Use KLPP to construct the adjacency graph and calculate the weight matrix and Laplace matrix;

[0063] (1) Construct an adjacency graph and calculate the weight matrix:

[0064]

[0065] Where W XY is the adjacency weight matrix, and β is the weight parameter.

[0066] (2) Calculate the Laplace matrix:

[0067] L=D XY -W XY ,

[0068] Where L is the Laplace matrix, D XYis the degree matrix, W XY is the adjacency weight matrix, and y represents the sample number of the Y feature.

[0069] S3-4. Simultaneous kernel matrix K h , the projection matrix A, Laplace matrix L, and projection vector eigenvalue λ formed by the projection vector are used to solve the generalized characteristic equation:

[0070] K c A=λ(K h LK h )A

[0071] Where A is the projection matrix composed of the projection vector α, that is, the fusion feature.

[0072] S3-5, based on the kernel matrix K h , projection matrix A, projecting the fused features into the low-dimensional space to obtain the features after dimensionality reduction:

[0073] Z=A T K h

[0074] Where Z is the low-dimensional feature representation after dimensionality reduction.

[0075] Through the KPCA-KLPP algorithm, multi-dimensional features can be integrated and high-contribution principal components can be extracted to achieve data dimensionality reduction. While retaining the core characteristics of battery capacity degradation, a low-dimensional representation system is constructed to provide more discriminative data input for subsequent training models.

[0076] S4. Construct the MESA-TCCT model to predict the remaining service life of lithium-ion batteries.

[0077] Furthermore, the MESA-TCCT model is constructed in step S4, specifically: based on the informer model, the information entropy sparse attention mechanism is introduced into the knowledge distillation block of the encoder instead of the probability sparse attention mechanism. The information entropy sparse attention mechanism completes pruning by selecting query vectors with high information entropy and optimizes the distribution of attention weights; and the extended causal convolution is introduced into the knowledge distillation block of the encoder instead of the standard convolution, and the receptive field of the convolution kernel on the time series is expanded to more effectively transmit gradient information; the output of the knowledge distillation block of each scale introduces a sequentially connected passthrough mechanism and a linear connection layer, and the feature maps of different scales are spliced through the passthrough mechanism to obtain finer-grained information and improve the prediction ability of the model; the mask-information entropy sparse attention mechanism is introduced into the decoder instead of the mask-probability sparse attention mechanism.

[0078] Furthermore, the information entropy sparse attention mechanism is specifically:

[0079] S4-1-1. For the input data, a linear transformation layer is used to obtain the corresponding query matrix Q, key matrix K and value matrix V;

[0080] S4-1-2, for the i-th query vector Q in the query matrix i , randomly sample m keys from the key matrix set, and get γ={K1,…,K m}, calculate the query vector Q i Scaled dot product with the sampling key

[0081]

[0082] Where, d k is the query vector Q i The sequence length, K j is the jth sampling key, j = 1, 2, ..., m.

[0083] S4-1-3. Calculate the scaled dot product through the softmax function The probability distribution of , and calculate the information entropy of each query vector:

[0084]

[0085] Where H i is the i-th query vector Q i Information entropy.

[0086] S4-1-4. Select the first q query vectors with the largest information entropy to construct a sparse query matrix Q M (For example, first sort the information entropy from large to small, select the query vectors corresponding to the first q information entropies to construct the sparse query matrix Q M ); Based on the sparse query matrix Q M , the softmax function is used to calculate the sparse attention score matrix, and the sparse attention score matrix is aggregated with the value matrix to obtain a weighted sum vector, which is the final attention score output:

[0087]

[0088] Where, is the sparse attention score matrix, Q M is the sparse query matrix, K is the key matrix composed of key vectors, T is the transpose; V is the value matrix, and Output is the final attention score of the output.

[0089] Furthermore, the formula of the extended causal convolution mechanism is as follows:

[0090]

[0091] Where t is the current time step, F is the size of the convolution kernel, d is the expansion rate, ω(f) is the f-th weight of the convolution sum, and x is the input of the extended causal convolution mechanism.

[0092] The extended causal convolution mechanism mentioned above performs downsampling through a 1-dimensional extended causal convolution layer, halving the length of the input sequence, highlighting key attention features, and expanding the receptive field of the convolution kernel on the time series, enabling the model to more effectively transmit gradient information.

[0093] S4-3. Construct a passthrough mechanism to stitch together the feature maps of self-attention blocks of different scales by expanding the feature map, thereby obtaining more fine-grained information and improving the prediction ability of the model.

[0094] Furthermore, the Passthrough mechanism is as follows:

[0095] S4-3-1, stitching feature maps of different scales;

[0096]

[0097] Where, is the feature map at each scale, E fuse is the concatenated feature map.

[0098] S4-3-2, the concatenated feature map E fuse Perform 1D convolution Conv1d to further fuse multi-scale information and then output the final features:

[0099] E out =GELU(Conv1d(E fuse ,kernel=1))

[0100] In the formula, kernel=1 means the convolution kernel is 1, GELU is the activation function, E out is the final output feature.

[0101] The MESA-TCCT model proposed in this paper uses sparse attention based on information entropy to filter redundant queries and generate streamlined key features. Extended causal convolution utilizes a hole design to expand the receptive field on the streamlined features, capturing long-range dependencies while maintaining causal constraints. At the same time, the model's ability to model global patterns is enhanced through efficient gradient transfer. The passthrough mechanism uses feature splicing to compensate for local details that may be lost due to downsampling or pruning caused by sparse attention and extended convolution. The three enhance the model's ability to predict battery RUL through a pipelined collaboration of "screening-enhancement-fusion."

[0102] The following is a further description of the optional embodiment of the present invention in conjunction with simulation data:

[0103] S1. Obtain battery test data through the experimental platform. This article uses the battery testing equipment model CTE-8008-5V300A of Xinwei Company, selects the lithium battery monomer model 18650 of Delip brand as the experimental object, and uses a constant temperature box to provide a stable working environment.

[0104] The battery has a rated capacity of 2A·h, and the charge and discharge process is carried out at 24°C: first, it is charged at a constant current of 2A until the cut-off voltage is 4.2V, and then it is charged at a constant voltage of 4.2V until the charging current is less than 0.1A; the discharge process is discharged at a constant current of 2A until the cut-off voltage reaches 2.75V. One charging process and one discharging process constitute one cycle, and the experiment is repeated for a total of 800 cycles. The curve connecting the capacity data of 800 cycles is the capacity degradation curve; at the same time, the battery constant current charging time, Coulombic efficiency, IC curve peak and peak position voltage characteristic data of 800 cycles are obtained.

[0105] S2. The Pearson coefficient is used to measure the correlation between the features and the capacity degradation curve, and the features with high correlation are retained. The Pearson coefficients of the four features are shown in Table 1. They all have a high correlation with capacity degradation, so they are all retained.

[0106] Table 1 Correlation coefficients

[0107] feature Constant current charging time Coulomb efficiency IC curve peak Peak position voltage coefficient 0.9975 0.9934 0.9969 -0.9867

[0108] S3. Preprocess the retained features using local weighted regression to achieve filtering and smoothing of feature data, and then normalize the preprocessed features; use the KPCA-KLPP algorithm to fuse and reduce the dimension of the normalized features. The first three principal components are as follows: Figure 2 The contribution rate and cumulative contribution rate are shown in Table 2, among which the contribution rate of the first principal component reaches 99.436%, which can more completely characterize the characteristics of the entire capacity degradation process of the battery. Therefore, the first principal component is selected as the model training data after dimensionality reduction.

[0109] Table 2 Contribution rate of each principal component

[0110] Serial number Contribution rate (%) Cumulative contribution rate% 1 99.436 99.436 2 0.403 99.839 3 0.093 99.932

[0111] S4. Construct the MESA-TCCT model. The overall structure is as follows Figure 3 As shown in the figure, it mainly includes two parts: encoder (Encoder) and decoder (Decoder).

[0112] The encoder includes:

[0113] (1) Position encoding, which indicates the position information of the data in the sequence. The process is as follows:

[0114] PE (pos,2t) =sin(w t pos)

[0115] PE (pos,2t+1) =cos(w t ·pos)

[0116] Where, pos represents the position of each data vector, d model represents the vector dimension and t represents the time step.

[0117] (2) Information entropy sparse attention layer: optimizes the distribution of attention weights by pruning low-entropy query vectors;

[0118] (3) Extended Causal Convolution Layer (DConv1d+Maxpool): The attention blocks are connected through a 1-dimensional extended causal convolution and a maximum pooling layer to reduce the length of the input sequence and highlight the key attention features. At the same time, the receptive field of the convolution kernel on the time series is expanded, so that the model can more effectively transmit gradient information.

[0119] (4) Passthrough layer: By expanding the feature map, the feature maps of self-attention blocks of different scales are spliced together to obtain finer-grained information and improve the prediction ability of the model.

[0120] The decoder includes: a masked information entropy sparse attention layer and a global attention layer;

[0121] The masked information entropy sparse attention layer controls access to sequence information during the information entropy sparse attention layer calculation process, ensuring that the model follows the autoregressive characteristics during the decoding phase, while improving the efficiency and accuracy of long sequence prediction.

[0122] Furthermore, the grid search method was used to optimize the model parameters, and the optimized parameters were as follows: the number of attention heads of the masked information entropy sparse attention layer and the information entropy sparse attention layer h = 8, the convolution kernel F = 1 of the extended causal convolution layer, the dilation rate d = 2, the window length of 4, the data input batch size of 8, the number of training iterations of 20, and the learning rate of 0.0001.

[0123] Error evaluation indicators: Mean absolute error (MAE), root mean square error (RMSE) and absolute error (AE):

[0124]

[0125] AE=|R ULtrue -R ULpred |

[0126] Where: Z represents the predicted length, R ULtrue and R ULpred They represent the actual number of cycles and the predicted number of cycles when the failure threshold is reached, respectively, true (a) represents the true value, y pred (a) indicates the predicted value.

[0127] The feature data obtained from the experiment is processed through steps S1-S3 and then divided. The first 50% of the feature data after division is used as a training set to train the MESA-TCCT model. The capacity degradation data obtained from the experiment is divided, and the last 50% of the capacity degradation data is used as a validation set. The trained MESA-TCCT model is used to gradually predict the validation set through a sliding window mechanism (when using the trained model for prediction, the data of the sliding window is used as input, and the model outputs the predicted value of the next time step. This predicted value can be used as part of the window input data of the next time step to continue predicting the next time point). The trained model of the present invention is compared with other methods such as CNN-LSTM and Informer for battery RUL prediction. Figure 4 As shown in (True represents the actual capacity value at each cycle number), the prediction results are shown in Table 3.

[0128] Table 3 Prediction results

[0129] Model MAE % RMSE% AE / cycles CNN-LSTM 2.05 2.72 21 Informer 1.65 2.12 17 MESA-TCCT 0.58 0.74 5

[0130] Experimental results demonstrate that the present invention has a good predictive effect on lithium-ion batteries, verifying the accuracy of the present invention's predictions. It should be noted that capacity decreases with increasing usage. It is generally believed that a battery reaches the end of its life when its capacity decays to 70% to 80% of its rated capacity. In this example, the battery has a rated capacity of 2A·h, so the failure threshold is 70% * 2 = 1.4A·h.

[0131] Furthermore, historical data of other batteries under a preset window length and cycle number can be collected, and the trained model of the present invention can be used to achieve the purpose of RUL prediction.

[0132] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model, characterized in that: The following steps are involved: S1. Obtain multi-dimensional characteristics and capacity values of lithium-ion batteries; Draw the capacity degradation curve based on the capacity value; S2. Measure the correlation between each feature and the capacity degradation curve using the Pearson coefficient, and retain the features whose Pearson coefficient meets the preset criteria; S3, preprocessing the retained features, and then normalizing the preprocessed features; using the KPCA-KLPP algorithm to fuse and reduce the dimension of the normalized features; S4. Construct the MESA-TCCT model to predict the remaining service life of lithium-ion batteries.

2. The lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to claim 1 is characterized in that: The multi-dimensional features include constant current charging time, coulombic efficiency, IC curve peak and peak position voltage.

3. The lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to claim 1 is characterized in that: The KPCA-KLPP algorithm is specifically: S3-1, use Gaussian kernel to map the normalized features to high dimension; S3-2, using KPCA to analyze the kernel matrix K h Centralize and centralize the kernel matrix K c Decompose to obtain the projection vector α and eigenvalue λ: S3-3. Use KLPP to construct the adjacency graph and calculate the weight matrix and Laplace matrix; S3-4. Simultaneous kernel matrix K h , the projection matrix A, Laplace matrix L, and projection vector eigenvalue λ formed by the projection vector are used to solve the generalized characteristic equation: S3-5, based on the kernel matrix K h , projecting the fused features as the projection matrix into the low-dimensional space to obtain the features after dimensionality reduction.

4. The lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to claim 1 is characterized in that: The MESA-TCCT model constructed in the step S4 is specifically as follows: based on the informer model, an information entropy sparse attention mechanism is introduced into the knowledge distillation block of the encoder instead of the probability sparse attention mechanism. The information entropy sparse attention mechanism completes pruning by selecting query vectors with high information entropy and optimizes the distribution of attention weights; and an extended causal convolution is introduced into the knowledge distillation block of the encoder instead of the standard convolution to expand the receptive field of the convolution kernel on the time series; a sequentially connected passthrough mechanism and a linear connection layer are introduced into the output of the knowledge distillation block of each scale, and the feature maps of different scales are spliced through the passthrough mechanism to obtain finer-grained information; a mask-information entropy sparse attention mechanism is introduced into the decoder instead of the mask-probability sparse attention mechanism.

5. The lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to claim 4 is characterized in that: The information entropy sparse attention mechanism is specifically: S4-1-1. For the input data, a linear transformation layer is used to obtain the corresponding query matrix Q, key matrix K and value matrix V; S4-1-2, for the i-th query vector Q in the query matrix i , randomly sample m keys from the key matrix set, and get γ={K1,…,K m }, calculate the query vector Q i Scaled dot product with the sampling key Where, d k is the query vector Q i The sequence length, K j is the jth sampling key, j = 1, 2, ..., m; S4-1-3. Calculate the scaled dot product through the softmax function The probability distribution of , and calculate the information entropy of each query vector: Where H i is the i-th query vector Q i Information entropy of S4-1-4. Select the first q query vectors with the largest information entropy to construct a sparse query matrix Q M ; According to the sparse query matrix Q M , the softmax function is used to calculate the sparse attention score matrix, and the sparse attention score matrix is aggregated with the value matrix to obtain a weighted sum vector, which is the final attention score output: Where, is the sparse attention score matrix, T is the transpose; Output is the final attention score of the output.

6. A lithium-ion battery RUL prediction system based on multi-feature fusion dimensionality reduction and MESA-TCCT model is characterized by: A module comprising the lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model as described in any one of claims 1-5.

7. A processor, characterized in that: The processor is used to run a program, wherein the program executes the lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to any one of claims 1 to 5 when running.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored program, wherein, when the program is running, the device where the computer-readable storage medium is located is controlled to execute the lithium-ion battery RUL prediction method based on multi-feature fusion dimensionality reduction and MESA-TCCT model according to any one of claims 1 to 5.

Images