A lithium battery residual life prediction method based on multi-feature fusion large model

Abstract

The application provides a lithium battery residual life prediction method based on a multi-feature fusion large model, relates to the field of lithium battery health management and life prediction, and comprises the following steps: segmenting original data through a sliding window technology, and converting the original data into key indicators representing the battery health state by using KAN to construct a key health indicator soft measurement module; splicing the data and the key indicators of the lithium battery to form fusion features, inputting the fusion features into a large language model LLM to construct a fusion feature prediction module, obtaining future fusion features through pre-training word embedding, a multi-head attention mechanism and a natural language prefix prompt, inputting the future fused features into KAN subjected to sparse processing to construct a residual life prediction model, establishing a regression relationship between the residual life and the battery, and realizing real-time prediction of the residual life of the lithium battery, which is suitable for state monitoring and maintenance decision of the lithium battery in scenarios such as electric vehicles and energy storage systems.

Description

Technical Field

[0001] This application relates to the field of lithium battery health management and life prediction, and in particular to a method for predicting the remaining life of lithium batteries based on a large model of multi-feature fusion. Background Technology

[0002] Existing lithium battery life prediction technologies still face numerous challenges. Traditional methods can acquire complete charge and discharge data in offline scenarios, but in actual online monitoring, they often only collect partial charge and discharge data, making it difficult to accurately assess key performance indicators. This incomplete data problem severely restricts the real-time performance and accuracy of life prediction.

[0003] While machine learning and deep learning methods have made some progress in this field, enabling the prediction of battery life by analyzing historical data, these technologies still have significant shortcomings. Traditional machine learning requires manually designed features, which is difficult to adapt to the complexity of lithium battery data; although deep learning can automatically extract features, it suffers from problems such as complex model structure, large training data requirements, and susceptibility to overfitting, and performs poorly when faced with the high-dimensional dynamic characteristics of lithium battery data. Summary of the Invention

[0004] The purpose of this invention is to address the problem that existing lithium battery life prediction technologies cannot balance prediction efficiency and prediction accuracy, and to provide a lithium battery remaining life prediction method based on a multi-feature fusion large model.

[0005] The above-mentioned objective of this application is achieved through the following technical solution:

[0006] S1: Obtain and process historical multivariate time series data of lithium batteries during the charging and discharging process to obtain a dataset. Construct a large multi-feature fusion model, including: the model of the large multi-feature fusion model, the fusion feature prediction module based on the large model, and the remaining lifetime prediction module;

[0007] S2: Through the dataset The KAN network based on cubic spline curves is trained to obtain the trained key health indicator soft measurement module and the predicted key health indicators.

[0008] S3: Transfer the dataset The key health indicators predicted are concatenated according to the feature dimensions to obtain a fused dataset. By merging datasets The LLM network is trained to obtain a trained fusion feature prediction module based on a large model;

[0009] S4: By fusing datasets Train the KAN network to obtain a trained remaining lifetime prediction module;

[0010] S5: Acquire real-time data of the lithium battery; process the real-time data through the key health indicator soft measurement module, the key health indicator soft measurement module and the remaining life prediction module to obtain the remaining life prediction result of the lithium battery.

[0011] Optionally, step S1 includes:

[0012] Suppose a historical multivariate time series dataset K represents the total length of the data, and p represents the dimension of the feature;

[0013] The dataset M is windowed using a sliding window technique with a window size of W and a sliding step of s, resulting in the dataset... ;No. The data for each window is W represents the length of the data, and .

[0014] Optionally, step S2 includes:

[0015] data As input to the KAN network, p represents the number of non-construction metrics. This represents the p-th non-construction metric;

[0016] , The vector is a column vector. The representative is the first One variable, ; This indicates that it is the t-th non-constructed index variable;

[0017] The output of the key health indicator soft measurement module is the predicted key health indicator F, with q indicators. The mathematical expression for the corresponding KAN network is:

[0018]

[0019] Where m is the m-th key indicator for prediction. ; and It is a continuous univariate function; Represents the t-th non-constructed index variable;

[0020] Parameterization of a univariate one-dimensional function is performed using B-spline curves;

[0021] spline function Based on nodes Defined as a piecewise function, in the interval The function within is in the form of:

[0022]

[0023] in It is a learnable coefficient. For spline functions, The time indicates that C-splines are used;

[0024] In the key health indicator soft measurement module, the output is represented as a combination of multiple functions, specifically the KAN formula as follows:

[0025]

[0026] in Indicates the first The activation function matrix of the layer, This indicates a composition operation of the function, where the output of the previous layer is used as the input of the next layer, and L is the total number of layers in the network.

[0027] The loss function for training the KAN network is as follows:

[0028]

[0029] Among them, the The output of each window is , , Indicates the first Sample values ​​of a health indicator Indicates the first Predicted values ​​for several health indicators.

[0030] Optionally, step S3 includes:

[0031] S31: Let the dataset be... The A window for each piece of raw data Key health indicators Then merge the datasets The fused data is ;

[0032] S32: Normalize the features of the fused data using a reversible anti-instance normalization network;

[0033] Let the fused data be the first The input sequences are The predicted future value is ;

[0034] By using the mean and variance, the input data The formula for standardizing and obtaining normalized data is as follows:

[0035]

[0036] in, It is the mean; It is variance; It is a learnable affine parameter vector. It is a small positive number;

[0037] S33: The normalized first... Data Transformed through a linear embedding layer ;

[0038] S34: Utilizing pre-trained word embeddings, the main body large model pre-trains word embeddings Where V is the vocabulary size and D represents the length of a word; sparsity processing is performed through word embeddings E to generate a set of text prototypes. ;

[0039] S35: Processed using a multi-head attention mechanism and Aggregation yields feature vectors ;

[0040] Feature vector The time series feature vector is obtained after a linear mapping layer. ;

[0041] S36: Using prompts as prefixes, the contextual information of the time series feature vectors is incorporated into the input as a prefix in natural language form, resulting in an embedding vector. ;

[0042] Combine prompt word vectors and time series feature vectors The features are concatenated to obtain the fused feature vector. ;

[0043] S37: Using fused feature vectors The input is fed into the pre-trained backbone LLM to obtain the output Y;

[0044] Flatten the output Y into a one-dimensional vector. s is the length after unfolding;

[0045] Flattened vector Mapped to the target dimension through a linear layer. Let the weight matrix of the linear layer be... The bias vector is The resulting prediction formula is: .

[0046] Optionally, step S4 includes:

[0047] First, the merged dataset De-windowing, then according to window size s and sliding step size . Window the dataset to obtain the dataset. ;No. Each window is ;

[0048] Data fusion The input is mapped into a sparse coding layer, and the mapped features are then fed into a KAN network to obtain a predicted value reflecting the battery capacity of the lithium battery. ;

[0049] The improved KAN's total loss function is:

[0050]

[0051] The first term is the prediction error term, which measures the predicted value. Compared with the true value The mean square error; This is the L1 regularization of the spline parameters; this term represents the complexity of the constraint nonlinear transformation. The weights of the sparse coding layer L1 regularization is used to constrain feature selection in sparse coding layers; parameters and This is the hyperparameter for regularization intensity.

[0052] Optionally, step S5 includes:

[0053] S51: Acquire real-time data of the lithium battery and process it using sliding window technology. Input the data into the trained key health indicator soft measurement module to obtain the first key health indicator. Then, concatenate the real-time data with the first key health indicator to form fused data. Input the data into the trained fusion feature prediction module based on a large model to obtain a predicted value that reflects the future state trend of the lithium battery. Input the predicted value into the trained remaining life prediction module to obtain a predicted value of the future battery capacity of the lithium battery.

[0054] S52: Repeat step S51 and compare the continuously obtained predicted values ​​with 70% of the rated capacity to determine the remaining life prediction result of the lithium battery.

[0055] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a lithium battery remaining life prediction method based on a multi-feature fusion large model.

[0056] A computer-readable storage medium storing instructions that, when executed, perform a method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model.

[0057] The beneficial effects of the technical solution provided in this application are:

[0058] 1. This invention proposes a method for predicting the remaining life of lithium batteries based on a large multi-feature fusion model. It addresses the problem of scarce early-stage lithium battery data and the difficulty in effectively utilizing multi-dimensional features by constructing a multi-feature fusion network. This network integrates capacity data, multi-dimensional features, and text enhancement information in a hierarchical architecture, significantly improving the prediction accuracy of the remaining life of lithium batteries in small-sample scenarios.

[0059] 2. When dealing with small sample data of lithium batteries, this invention employs an LLM-enhanced prediction mechanism. Leveraging the semantic reasoning capabilities of a large language model, natural language prefixes are added to a small amount of data through prompting learning, associating features with text semantics; the knowledge generalization capability of LLM is utilized to efficiently extract features from small samples.

[0060] 3. To address the issues of high feature dimensionality and redundant interference, this invention adds a sparsity layer before the KAN network, using regularization to automatically filter key features and reduce dimensionality, thereby achieving efficient processing of complex data. Attached Figure Description

[0061] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0062] Figure 1 This is a step diagram of an embodiment of this application;

[0063] Figure 2 This is a diagram illustrating the capacity degradation process of the CALCE-CS2 series battery in the embodiments of this application;

[0064] Figure 3 This is a schematic diagram of the prediction results of CS2_35 in the embodiments of this application;

[0065] Figure 4 This is a schematic diagram of the prediction results of CS2_36 in the embodiments of this application;

[0066] Figure 5 This is a schematic diagram of the prediction results of CS2_37 in the embodiments of this application;

[0067] Figure 6 This is a schematic diagram of the prediction results of CS2_38 in the embodiments of this application;

[0068] Figure 7 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0069] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0070] The embodiments of this application provide a method for predicting the remaining life of lithium batteries based on a large model of multi-feature fusion.

[0071] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a lithium battery remaining life prediction method based on a multi-feature fusion large model in an embodiment of this application, including:

[0072] S1: Obtain and process historical multivariate time series data of lithium batteries during the charging and discharging process to obtain a dataset. Construct a large multi-feature fusion model, including: the model of the large multi-feature fusion model, the fusion feature prediction module based on the large model, and the remaining lifetime prediction module;

[0073] S2: Through the dataset The KAN network based on cubic spline curves is trained to obtain the trained key health indicator soft measurement module and the predicted key health indicators.

[0074] In one embodiment of this application, data The input is a KAN network using cubic spline curves, which is processed through a multi-layer nonlinear structure to output indices. By calculating the loss between the sampled and predicted values ​​of the indicators, the model parameters are updated using the backpropagation algorithm, and optimized through multiple iterations. Finally, a well-trained soft measurement module for key health indicators is obtained.

[0075] S3: Transfer the dataset The key health indicators predicted are concatenated according to the feature dimensions to obtain a fused dataset. By merging datasets The LLM network is trained to obtain a trained fusion feature prediction module based on a large model;

[0076] In one embodiment of this application, the fused data First, the data is preprocessed using a reversible inverse instance normalization network. After standardization, it is then transformed back to the original scale by a normalization layer to obtain normalized data. Then on Transform the linear embedding layer into ; Introduce pre-trained word embeddings, and generate a set of text prototypes from the word embeddings E. Use multi-head attention mechanism to process and The feature vector Z is obtained through aggregation; the feature vector Z is aligned with the backbone model through a linear mapping layer to obtain feature O. Natural language prefixes are added using prompt-as-prefix (PaP) to convert the relevant text into embedding vectors. ; Combine feature O with embedding vector Feature A is obtained by concatenation and then input into the pre-trained LLM for mapping. The mapped output is flattened and then mapped through a linear layer to obtain the predicted value Y of the future window. The mean squared error is used as the loss function, and the parameters of the pre-trained LLM are updated through backpropagation. After multiple iterations, a well-trained future fusion feature prediction module based on a large model is obtained.

[0077] S4: By fusing datasets Train the KAN network to obtain a trained remaining lifetime prediction module;

[0078] In one embodiment of this application, the fused dataset Perform dewindowing and rewindowing to obtain the dataset. The data is processed through a sparse coding layer, and then KAN (Knowledge Anonymous) based on cubic spline curves is used to predict the remaining lifespan of lithium batteries. The model establishes a regression relationship between the input fused features and the remaining lifespan to obtain the predicted battery capacity. Based on the total loss function of the terms that measure the difference between the predicted and actual remaining lifetime and the terms that implement the sparsity constraint, the model parameters are updated using the backpropagation algorithm. After multiple iterations of training, the trained remaining lifetime prediction module is finally obtained.

[0079] S5: Acquire real-time data of the lithium battery; process the real-time data through the key health indicator soft measurement module, the key health indicator soft measurement module and the remaining life prediction module to obtain the remaining life prediction result of the lithium battery.

[0080] Step S1 includes:

[0081] Suppose a historical multivariate time series dataset K represents the total length of the data, and p represents the dimension of the feature;

[0082] The dataset M is windowed using a sliding window technique with a window size of W and a sliding step of s, resulting in the dataset... ;No. The data for each window is W represents the length of the data, and .

[0083] As one example, data from the charging and discharging process of a lithium battery is acquired and compiled into a dataset. Where K represents the number of samples and n represents the number of sensors, which is also the number of variable features; the dataset M is divided into data samples using a sliding window to obtain a windowed dataset. ;No. Each window is Specifically: For the lithium battery charging and discharging dataset M, set the window size W (the number of consecutive samples for a single sample) and the sliding step size s (the window movement interval). Starting from the initial position, extract the first sample segment according to W as the first sample, and then repeatedly extract the first sample segment with the sliding window of s until the entire length W cannot be extracted. Combine the obtained consecutive sample segments in order to obtain the windowed dataset X.

[0084] Step S2 includes:

[0085] data As input to the KAN network, p represents the number of non-construction metrics. This represents the p-th non-construction metric;

[0086] , The vector is a column vector. The representative is the first One variable, ; This indicates that it is the t-th non-constructed index variable;

[0087] The output of the key health indicator soft measurement module is the predicted key health indicator F, with q indicators. The mathematical expression for the corresponding KAN network is:

[0088]

[0089] Where m is the m-th key indicator for prediction. ; and It is a continuous univariate function; Represents the t-th non-constructed index variable;

[0090] In one embodiment of this application, the non-constructed index represents the average voltage and voltage variance during the constant current charging phase. Since the functions to be learned are all univariate functions, B-spline curves are used to parameterize their one-dimensional functions. This method possesses learnable local B-spline basis function coefficients. B-splines include various orders; cubic splines (C-splines), as a special form of basis splines, are commonly used due to their good balance between smoothness and computational complexity.

[0091] Parameterization of a univariate one-dimensional function is performed using B-spline curves;

[0092] spline function Based on nodes Defined as a piecewise function, in the interval The function within is in the form of:

[0093]

[0094] in It is a learnable coefficient. For spline functions, The time indicates that C-splines are used;

[0095] In the key health indicator soft measurement module, the output is represented as a combination of multiple functions, specifically the KAN formula as follows:

[0096]

[0097] in Indicates the first The activation function matrix of the layer, This indicates a composition operation of the function, where the output of the previous layer is used as the input of the next layer, and L is the total number of layers in the network.

[0098] The loss function for training the KAN network is as follows:

[0099]

[0100] Among them, the The output of each window is , , Indicates the first Sample values ​​of a health indicator Indicates the first Predicted values ​​for several health indicators.

[0101] Step S3 includes:

[0102] S31: Let the dataset be... The A window for each piece of raw data Key health indicators Then merge the datasets The fused data is ;

[0103] S32: Normalize the features of the fused data using a reversible anti-instance normalization network;

[0104] In one embodiment of this application, considering the possibility of data drift, a reversible anti-instance normalization network is used, which consists of two parts: a normalization layer with a symmetric structure and an anti-normalization layer.

[0105] Let the fused data be the first The input sequences are The predicted future value is ;

[0106] By using the mean and variance, the input data The formula for standardizing and obtaining normalized data is as follows:

[0107]

[0108] in, It is the mean; It is variance; It is a learnable affine parameter vector. It is a small positive number;

[0109] S33: The normalized first... Data Transformed through a linear embedding layer ;

[0110] S34: Utilizing pre-trained word embeddings, the main body large model pre-trains word embeddings Where V is the vocabulary size and D represents the length of a word; sparsity processing is performed through word embeddings E to generate a set of text prototypes. ;

[0111] In one embodiment of this application, smaller text prototypes are generated through sparsification, reducing the computational burden. A smaller set of text prototypes is generated through sparsification. ,in .

[0112] S35: Processed using a multi-head attention mechanism and Aggregation yields feature vectors ;

[0113] Feature vector The time series feature vector is obtained after a linear mapping layer. ;

[0114] S36: Using prompts as prefixes, the contextual information of the time series feature vectors is incorporated into the input as a prefix in natural language form, resulting in an embedding vector. ;

[0115] In one embodiment of this application, prompts are used as prefixes to incorporate contextual information of the time-series feature vectors into the input in natural language form. This information, encompassing dataset descriptions, task instructions, and input statistics, assists the LLM in understanding and reasoning. The prompt words are represented as text and input into a pre-trained LLM for mapping, resulting in prompt word vectors. For each window of data, the prompt word vector is... .

[0116] Combine prompt word vectors and time series feature vectors The features are concatenated to obtain the fused feature vector. ;

[0117] S37: Using fused feature vectors The input is fed into the pre-trained backbone LLM to obtain the output Y;

[0118] The main LLM backbone is composed of Transformers; prefixes are added to the input to enhance inference capabilities, while the prefixes are discarded and the parts relevant to the time series data are retained in the output stage.

[0119] Flatten the output Y into a one-dimensional vector. s is the length after unfolding;

[0120] Flattened vector Mapped to the target dimension through a linear layer. Let the weight matrix of the linear layer be... The bias vector is The resulting prediction formula is: .

[0121] Step S4 includes:

[0122] First, the merged dataset De-windowing, then according to window size s and sliding step size . Window the dataset to obtain the dataset. ;No. Each window is ;

[0123] Data fusion The input is mapped into a sparse coding layer, and the mapped features are then fed into a KAN network to obtain a predicted value reflecting the battery capacity of the lithium battery. ;

[0124] The improved KAN's total loss function is:

[0125]

[0126] The first term is the prediction error term, which measures the predicted value. Compared with the true value The mean square error; This is the L1 regularization of the spline parameters; this term represents the complexity of the constraint nonlinear transformation. The weights of the sparse coding layer L1 regularization is used to constrain feature selection in sparse coding layers; parameters and This is the hyperparameter for regularization intensity.

[0127] Step S5 includes:

[0128] S51: Acquire real-time data of the lithium battery and process it using sliding window technology. Input the data into a trained key health indicator soft measurement module to obtain the first key health indicator. Then, concatenate the real-time data with the first key health indicator to form fused data. Input the data into a trained fusion feature prediction module based on a large model to obtain a predicted value that reflects the future state trend of the lithium battery. Input the predicted value into a trained remaining life prediction module to obtain a predicted value of the future battery capacity of the lithium battery.

[0129] S52: Repeat step S51 and compare the continuously obtained predicted values ​​with 70% of the rated capacity to determine the remaining life prediction result of the lithium battery.

[0130] In one embodiment of this application, the remaining lifespan during the online phase is determined as follows: First, the raw data collected in real time is processed using a sliding window technique to form windowed data with a fixed time step. This data is then input into a key health indicator soft measurement module, where the module's internal algorithm model extracts and outputs key health indicators for the lithium battery. Next, the raw real-time data and the predicted key health indicators are concatenated along their feature dimensions to generate fused data containing multi-source information. This fused data is then input into a fusion feature prediction module based on a large model. Utilizing the deep semantic understanding and feature interaction capabilities of the large model, a predicted value reflecting the lithium battery's state trend is output. This predicted value is then passed to a remaining lifespan prediction module, where a regression model or physical model is used to calculate the predicted future battery capacity. To achieve dynamic monitoring, the system continuously repeats the above process, comparing each predicted battery capacity value with 70% of the lithium battery's rated capacity (a commonly used industry lifespan threshold). Based on the comparison results, the remaining lifespan status of the lithium battery is determined in real time, thus providing timely health warnings and maintenance decision-making basis for the battery management system.

[0131] In one embodiment, Figure 3 , Figure 4 , Figure 5 and Figure 6 This diagram illustrates the prediction results of various methods for predicting the remaining life of lithium batteries with different serial numbers using the CALCE-CS2 dataset, compared to the method used in this experiment.

[0132] In one embodiment of this application, the dataset was collected from the Center for Advanced Life Cycle Engineering (CALCE) at the University of Maryland. A 2Ah lithium battery was fully charged and discharged at approximately 25 degrees Celsius using an Arbin battery tester, with sampling intervals of 30 seconds. The capacity data used herein represents the charge or discharge value of a fully charged battery under specific discharge conditions. Figure 2 The capacity degradation process of the CALCE-CS2 series batteries, numbered CS2-35, CS2-36, CS2-37, and CS2-38, is shown. The batteries were tested under a standard constant current protocol at a current rate of 0.5C until the voltage reached 4.2V, and then charged at a constant voltage of 4.2V until the charging current dropped below 0.05A. For the CALCE-CS2 dataset, their rated capacity is assumed to be 1.10Ah, and the RUL predicted failure threshold is set to 0.77Ah (70% of the capacity class). The constructed metrics used in the dataset are the average voltage, voltage variance, and voltage variance during the constant current charging and discharging phases, calculated from the voltage data during charging and discharging. The unconstructed metrics are the peak value of the IC-V curve during charging, the trough value of the IC-V curve during discharging, and the amount of charge discharged when the voltage drops from 3.8V to 3.4V during discharging. The first two metrics are calculated from the relationship between the current and voltage during charging and discharging, while the last metric is calculated by recording the time period of voltage change and the current.

[0133] For result evaluation, absolute error (AE), relative error (RE), mean absolute error (MAE), and root mean square error (RMSE) are used. The performance of the method proposed in this invention is evaluated using the following formulas:

[0134]

[0135]

[0136]

[0137]

[0138] Where AE is Predicted values ​​and The absolute difference between the actual values ​​directly reflects the magnitude of the error in the RUL prediction and is the most intuitive way to measure error. RE is... Predicted values ​​and The ratio of the absolute value of the difference between the predicted and actual values ​​to the actual value measures the proportion of error between the predicted and actual values. It reflects the relative magnitude of the prediction error, is unaffected by data units, and is easy to compare across different datasets. MAE is the average of the absolute errors between all predicted and actual SOH values, comprehensively reflecting the model's average error level across the entire dataset. It is insensitive to outliers and stably reflects the overall performance of the model; in the formula... Representing the Battery capacity after one cycle Representing the The predicted battery capacity after [number] cycles. RMSE is the square root of the average of the squared prediction errors. It takes into account the magnitude of all prediction errors and amplifies the impact of large errors by squaring them. It is sensitive to large errors and can highlight the model's performance under extreme conditions.

[0139] The experimental platform is a server running Ubuntu, equipped with two 4th generation Intel Xeon 4210r CPUs (dual-socket), 32GB of RAM (4GB), and two NVIDIA 3090 GPUs with 24GB of VRAM each. This platform uses dual GPUs; to fully utilize resources, the DeepSpeed ​​tool is used for multi-GPU training to accelerate training speed. The code in this paper is written in Python, the model architecture uses the PyTorch framework, and the training runs 10 times with a batch size of 24.

[0140] To ensure the encoder output can be concatenated with word embeddings from different LLMs, the dimensionality of the encoder output needs to be adjusted according to the type of LLM to match the dimensionality of the word embeddings. For example, the word embedding dimension of the BERT model is 768, that of Llama 3.21B is 2048, and that of Llama 3.23B is 3072. Furthermore, considering that the number of experimental predictions will not exceed 1000, the size of the vocabulary mapped to the time series prediction is set to 1000 to ensure the accuracy and efficiency of the prediction results.

[0141] Offline Phase: Offline modeling comprises three sub-models: a key health indicator soft measurement model, a future fusion feature prediction model based on a large model, and a remaining life prediction model. The key health indicator soft measurement model takes directly measured indicators as input, using a portion of battery data as the training set and the other portion as the test set to validate its performance. The model's input and output are then fused to provide features for the next step. The future fusion feature prediction model based on the large model takes the fused features as input to obtain the fused features for the future. Finally, the remaining life prediction model takes the predicted values ​​of the fused features from the previous step as input, labeled as the capacity based on the predicted cycle number. It also uses the training and test sets for modeling and validation.

[0142] Online Phase: After offline modeling, the three models are merged to form an online prediction model, and online prediction begins. The model's output is sampled using the Monte Carlo method. The sampled results are then averaged with a 99% confidence level to obtain the final prediction result.

[0143] To verify the effectiveness of the proposed method, comparative experiments were conducted with three other methods. These methods included: CNN-LSTM, which combines a convolutional neural network (CNN) for feature extraction and a long short-term memory network (LSTM) for time series prediction; PCA-GWO-GRU, which utilizes principal component analysis (PCA) for dimensionality reduction, the grey wolf optimization algorithm (GWO) for parameter optimization, and gated recurrent units (GRUs) for prediction; and PCA-Transformer, which uses PCA for feature extraction and a Transformer network for time series prediction. All methods employed the same input and output method: direct measurement metrics as input and battery capacity values ​​as output. The same hyperparameter settings were used in the experiments to ensure fairness. Evaluation metrics included AE, RE, MAE, and RMSE. The battery capacity prediction curves for the four models are shown below. Figure 3 As shown in Table 1, the evaluation indexes for the prediction results are compared. The results in bold in the table show that the method proposed in this invention outperforms other methods in terms of performance.

[0144] Table 1: Results of RUL on CS2 of the CALCE dataset

[0145]

[0146] Comparative experiments using different methods show that the MAE and RMSE of the method proposed in this chapter are lower than the other three methods across different battery numbers. On the CS2_36 battery, the method proposed in this chapter has a higher MAE. On the CS2_35, CS2_37, and CS2_38 batteries, the prediction results of this method are better than most of the comparative methods. However, on the CS2_38 battery, PCA-GWO-GRU and PCA-GWO-Transformer are unable to predict below 70% of the battery capacity. Overall, the method proposed in this chapter generally has lower error values, more stable prediction results, and outperforms other methods in overall performance.

[0147] Furthermore, to verify the impact of multi-feature fusion on the results in this invention, an ablation experiment was conducted to remove the soft measurement model of key health indicators for comparison. Instead of predicting indirect indicators, direct measurement indicators were used for time-series prediction of LLM, and the prediction results were input into the remaining life prediction model for battery capacity prediction. The results of the ablation experiment are shown in Table 2.

[0148] Table 2 Evaluation Indicators for Ablation Comparison Experiment

[0149]

[0150] The ablation comparison experiments show that removing multi-feature fusion degrades the model's predictive stability, increases its fluctuation range, and introduces significant prediction bias. Due to the lack of feature extraction, the correlation between directly measured indicators and capacitance values ​​is insufficient, leading to deviations in predicted values ​​in future iterations. In contrast, the method proposed in this chapter utilizes the high correlation between indirect indicators and capacitance values, employing KAN for nonlinear prediction to obtain feature fusion data. This significantly improves the accuracy of predictions in future iterations. Therefore, the multi-feature fusion method presented in this chapter is highly necessary.

[0151] This application also discloses an electronic device. (See reference...) Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0152] The communication bus 502 is used to enable communication between these components.

[0153] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0154] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0155] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the above-described method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model.

[0156] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0157] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for predicting the remaining life of lithium batteries based on a large-scale multi-feature fusion model, characterized in that, The method includes the following steps: S1: Obtain and process historical multivariate time series data of lithium batteries during the charging and discharging process to obtain a dataset. Construct a large multi-feature fusion model, including: the model of the large multi-feature fusion model, the fusion feature prediction module based on the large model, and the remaining lifetime prediction module; S2: Through the dataset The KAN network based on cubic spline curves is trained to obtain the trained key health indicator soft measurement module and the predicted key health indicators. S3: Transfer the dataset The key health indicators predicted are concatenated according to the feature dimensions to obtain a fused dataset. By merging datasets The LLM network is trained to obtain a trained fusion feature prediction module based on a large model; S4: By fusing datasets Train the KAN network to obtain a trained remaining lifetime prediction module; S5: Acquire real-time data of lithium batteries; process the real-time data through the key health indicator soft measurement module, the fusion feature prediction module, and the remaining life prediction module to obtain the remaining life prediction results of lithium batteries. Step S5 includes: S51: Acquire real-time data of the lithium battery and process it using sliding window technology. Input the data into the trained key health indicator soft measurement module to obtain the first key health indicator. Then, concatenate the real-time data with the first key health indicator to form fused data. Input the data into the trained fusion feature prediction module based on a large model to obtain a predicted value that reflects the future state trend of the lithium battery. Input the predicted value into the trained remaining life prediction module to obtain a predicted value of the future battery capacity of the lithium battery. S52: Repeat step S51 and compare the continuously obtained predicted values ​​with 70% of the rated capacity to determine the remaining life prediction result of the lithium battery.

2. The lithium battery remaining life prediction method based on a multi-feature fusion large model as described in claim 1, characterized in that, Step S1 includes: Suppose a historical multivariate time series dataset K represents the total length of the data, and p represents the dimension of the feature; The sliding window technique is used to process the dataset M according to a window size of W and a sliding step size of W. Window the dataset to obtain the dataset. ;No. The data for each window is W represents the length of the data, and .

3. The lithium battery remaining life prediction method based on a multi-feature fusion large model as described in claim 2, characterized in that, Step S2 includes: data As input to the KAN network, p represents the number of non-construction metrics. This represents the p-th non-construction metric; , The vector is a column vector. The representative is the first One variable, ; This indicates that it is the t-th non-construction index; the non-construction index represents the average voltage or the voltage variance during the constant current charging phase. The output of the key health indicator soft measurement module is the predicted key health indicator F, with q indicators. The mathematical expression for the corresponding KAN network is: Where m is the m-th key indicator for prediction. ; and It is a continuous univariate function; Indicates the first One non-construction metric; Parameterization of a univariate one-dimensional function is performed using B-spline curves; spline function Based on nodes Defined as a piecewise function, in the interval The function within is in the form of: in It is a learnable coefficient. For spline functions, The time indicates that C-splines are used; In the key health indicator soft measurement module, the output is represented as a combination of multiple functions, specifically the KAN formula as follows: in Indicates the first The activation function matrix of the layer, This indicates a composition operation of the function, where the output of the previous layer is used as the input of the next layer, and L is the total number of layers in the network. The loss function for training the KAN network is as follows: Among them, the The output of each window is , , Indicates the first Sample values ​​of a health indicator Indicates the first Predicted values ​​for several health indicators.

4. The lithium battery remaining life prediction method based on a multi-feature fusion large model as described in claim 1, characterized in that, Step S3 includes: S31: Let the dataset be... The A window for each piece of raw data Key health indicators Then merge the datasets The fused data is ; S32: Normalize the features of the fused data using a reversible anti-instance normalization network; Let the fused data be the first The input sequences are The predicted future value is ; By using the mean and variance, the input data The formula for standardizing and obtaining normalized data is as follows: in, It is the mean; It is variance; It is a learnable affine parameter vector. It is a small positive number; S33: The normalized first... Data Transformed through a linear embedding layer ; S34: Utilizing pre-trained word embeddings, the main body large model pre-trains word embeddings Where V is the vocabulary size and D represents the length of a word; sparsity processing is performed through word embeddings E to generate a set of text prototypes. ; S35: Processed using a multi-head attention mechanism and Aggregation yields feature vectors ; Feature vector The time series feature vector is obtained after a linear mapping layer. ; S36: Using prompts as prefixes, the contextual information of the time series feature vectors is incorporated into the input as a prefix in natural language form, resulting in an embedded vector. ; Combine prompt word vectors and time series feature vectors The features are concatenated to obtain the fused feature vector. ; S37: Using fused feature vectors The input is fed into the pre-trained backbone LLM to obtain the output Y; Flatten the output Y into a one-dimensional vector. s is the length after unfolding; Flattened vector Mapped to the target dimension through a linear layer. Let the weight matrix of the linear layer be... The bias vector is The resulting prediction formula is: .

5. The method for predicting the remaining life of lithium batteries based on a multi-feature fusion large model as described in claim 1, characterized in that, Step S4 includes: First, the merged dataset De-windowing, then adjust according to window size. and sliding step size Window the dataset to obtain the dataset. ;No. Each window is ; Data fusion The input is mapped into a sparse coding layer, and the mapped features are then fed into a KAN network to obtain a predicted value reflecting the battery capacity of the lithium battery. ; The improved KAN's total loss function is: The first term is the prediction error term, which measures the predicted value. Compared with the true value The mean square error; This is the L1 regularization of the spline parameters; this term represents the complexity of the constraint nonlinear transformation. The weights of the sparse coding layer L1 regularization is used to constrain feature selection in sparse coding layers; parameters and This is the hyperparameter for regularization intensity.

6. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the method as described in any one of claims 1-5.

Images