CN-120870876-B - Lithium battery residual life prediction method based on multi-feature fusion large model

CN120870876BCN 120870876 BCN120870876 BCN 120870876BCN-120870876-B

Abstract

The application provides a lithium battery residual life prediction method based on a multi-feature fusion large model, which relates to the field of lithium battery health management and life prediction, and comprises the steps of carrying out segmentation processing on original data by a sliding window technology, constructing a key health index soft measurement module by using KAN, and converting the original data into key indexes for representing the health state of a battery; and (3) splicing data of the lithium battery and key indexes to form fusion characteristics, inputting the fusion characteristics into a large language model LLM to construct a fusion characteristic prediction module, obtaining future fusion characteristics through pre-training word embedding, a multi-head attention mechanism and natural language prefix prompting, inputting the future fused characteristics into a sparse KAN, and constructing a residual life prediction model. And establishing a regression relation with the residual life of the battery, realizing the real-time prediction of the residual life of the lithium battery, and being suitable for the state monitoring and maintenance decision of the lithium battery in the scenes of electric automobiles, energy storage systems and the like.

Inventors

WANG ZHAOJING
XU TIANWEI
YANG FEI
HU XINRONG
PENG TAO
LI LI
LI HAONAN

Assignees

武汉纺织大学

Dates

Publication Date: 20260508
Application Date: 20250627

Claims (7)

1. The lithium battery residual life prediction method based on the multi-feature fusion large model is characterized by comprising the following steps of: s1, acquiring and processing historical multivariable time series data of the lithium battery in the charging and discharging process to obtain a data set Constructing a multi-feature fusion large model, wherein the multi-feature fusion large model comprises a multi-feature fusion large model, a fusion feature prediction module based on the large model and a residual life prediction module; S2, passing the data set Training a KAN network based on a cubic spline curve to obtain a trained key health index soft measurement module and a predicted key health index; S3, collecting the data set And the predicted key health indexes are spliced according to the dimension of the characteristics to obtain a fusion data set By fusing datasets Training the LLM network to obtain a trained fusion characteristic prediction module based on a large model; s4, fusing the data set Training the KAN network to obtain a trained residual life prediction module; s5, acquiring real-time data of the lithium battery, and processing the real-time data through a key health index soft measurement module, a fusion characteristic prediction module and a residual life prediction module to obtain a residual life prediction result of the lithium battery; the step S5 comprises the following steps: S51, acquiring real-time data of the lithium battery, performing sliding window technical processing, inputting a trained key health index soft measurement module to acquire a first key health index, splicing the real-time data and the first key health index to form fusion data, inputting a trained fusion characteristic prediction module based on a large model to acquire a predicted value reflecting the state trend of the future lithium battery, inputting the predicted value into a trained residual life prediction module to acquire a predicted value of the future battery capacity of the lithium battery; and S52, repeating the step S51, and comparing the continuously obtained predicted value with 70% of rated capacity to determine a residual life prediction result of the lithium battery.
2. The method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model as set forth in claim 1, wherein the step S1 includes: setting up 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 technology is adopted to carry out the data set M according to the window size W and the sliding step length Windowing to obtain data set First, a third step The data of each window is W represents the length of the data, and 。
3. The method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model as set forth in claim 2, wherein the step S2 includes: Data As an input to the KAN network, p represents the number of non-building metrics, Represents the p-th non-building index; , The vector is a column vector which is a vector, Representative is the first The number of variables that can be used, ; The non-construction index represents the average voltage value of the constant current charging stage or the voltage variance of the constant current charging stage; the output of the key health index soft measurement module is the predicted key health index F, the index number is q, and the mathematical expression of the corresponding KAN network is as follows: Wherein m is the key index of the mth prediction, ; And Is a continuous univariate function; Represent the first A non-building index; Parameterizing the univariate one-dimensional function through a B spline curve; spline function Node-based Defined as a piecewise function, in the interval The functional form in is: Wherein the method comprises the steps of Is a coefficient that can be learned and is, As a function of the spline, When C-splines are used; in the key health indicator soft measurement module, the output is expressed as a combination of functions, and the specific KAN formula is as follows: Wherein the method comprises the steps of Represent the first The matrix of activation functions of the layer, Representing the composite operation of the function, i.e. the output of the previous layer is taken as the input of the next layer, L is the total layer number of the network; the loss function for KAN network training is as follows: wherein, the first The output of each window is , , Represent the first The sampled values of the individual health indicators are, Represent the first Predicted values of individual health indicators.
4. The method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model as set forth in claim 1, wherein the step S3 includes: s31, setting up a data set Is the first of (2) The window of the original data is Key health index Then the data sets are fused Is the fusion data of (1) ; S32, normalizing the data of each feature of the fused data through a reversible inverse instance normalization network; set the first of the fusion data The input sequences are Predicting future values as ; Input data is input through mean and variance Normalized, the data formula for normalization is as follows: Wherein, the Is the mean value; Is the variance; Is a vector of affine parameters that can be learned, Is a small positive number; S33, normalizing the first Personal data Conversion by linear intercalation ; S34, word embedding pre-trained by utilizing a main trunk large model Wherein V is vocabulary, D represents the length of a word, and the word embedding E is used for sparsification to generate a text prototype set ; S35, processing by using multi-head attention mechanism And (3) with Polymerizing to obtain feature vector ; Feature vector Obtaining time sequence feature vector through a linear mapping layer ; S36, using the prompt as prefix, inputting the context information of the time sequence feature vector as prefix in natural language form, to obtain embedded vector ; The prompt word vector and the time sequence feature vector Splicing to obtain a fused feature vector ; S37, using the fused feature vector Inputting the data into a pre-trained trunk LLM to obtain an output Y; Flattening the output Y into a one-dimensional vector S is the length after expansion; vector after flattening Mapping to target dimensions through a linear layer Let the weight matrix of the linear layer be The offset vector is The resulting predictive formula: 。
5. The method for predicting the remaining life of a lithium battery based on a multi-feature fusion large model as set forth in claim 1, wherein the step S4 includes: First, for a fused dataset De-windowing, and then according to the window size And a sliding step length of Windowing to obtain data set First, a third step The windows are ; Will fuse the data The sparse coding layer is input for mapping, the mapped characteristics are input into a KAN network, and a predicted value reflecting the battery capacity of the lithium battery is obtained ; The total loss function of the improved KAN is: wherein the first term is a prediction error term, and the prediction value is measured And true value Mean square error of (a); is the L1 regularization of spline parameters, which is the complexity of the constrained nonlinear transformation, Is a sparse coding layer weight L1 regularization of (2) is feature screening of constraint sparse coding layer, parameter And Is a regularized intensity superparameter.
6. An electronic device comprising a processor, a memory, a user interface, and a network interface, the memory for storing instructions, the user interface and the network interface for communicating to other devices, the processor for executing the instructions stored in the memory to cause the electronic device to perform the method of any of claims 1-5.
7. A computer readable storage medium storing instructions which, when executed by a computer, perform the method of any one of claims 1-5.

Description

Lithium battery residual life prediction method based on multi-feature fusion large model Technical Field The application relates to the field of lithium battery health management and life prediction, in particular to a lithium battery residual life prediction method based on a multi-feature fusion large model. Background Existing lithium battery life prediction techniques still face a number of challenges. The traditional method can acquire complete charge and discharge data in an offline scene, but in the actual online monitoring process, only partial charge and discharge fragment data can be acquired, so that key performance indexes are difficult to evaluate accurately. This incomplete data severely constrains the real-time and accuracy of life predictions. While machine learning and deep learning approaches have advanced in this field, it is possible to predict battery life by analyzing historical data, but these techniques still suffer from significant shortcomings. The traditional machine learning needs manual design features, is difficult to adapt to the complexity of lithium battery data, and the deep learning can automatically extract the features, but has the problems of complex model structure, large training data requirement, easy overfitting and the like, and has poor performance when facing to the high-dimensional dynamic characteristics of the lithium battery data. Disclosure of Invention The invention aims to solve the problem that the existing lithium battery life prediction technology cannot achieve both prediction efficiency and prediction precision, and provides a lithium battery residual life prediction method based on a multi-feature fusion large model. The above object of the present application is achieved by the following technical solutions: s1, acquiring and processing historical multivariable time series data of the lithium battery in the charging and discharging process to obtain a data set Constructing a multi-feature fusion large model, wherein the multi-feature fusion large model comprises a multi-feature fusion large model, a fusion feature prediction module based on the large model and a residual life prediction module; S2, passing the data set Training a KAN network based on a cubic spline curve to obtain a trained key health index soft measurement module and a predicted key health index; S3, collecting the data set And the predicted key health indexes are spliced according to the dimension of the characteristics to obtain a fusion data setBy fusing datasetsTraining the LLM network to obtain a trained fusion characteristic prediction module based on a large model; s4, fusing the data set Training the KAN network to obtain a trained residual life prediction module; And S5, acquiring real-time data of the lithium battery, and processing the real-time data through the key health index soft measurement module, the key health index soft measurement module and the residual life prediction module to obtain a residual life prediction result of the lithium battery. Optionally, step S1 includes: setting up a historical multivariate time series dataset K represents the total length of the data, and p represents the dimension of the feature; windowing a data set M by adopting a sliding window technology according to a window size W and a sliding step length s to obtain the data set First, a third stepThe data of each window isW represents the length of the data, and。 Optionally, step S2 includes: Data As an input to the KAN network, p represents the number of non-building metrics,Represents the p-th non-building index; , The vector is a column vector which is a vector, Representative is the firstThe number of variables that can be used,;The representation is the t-th non-structural index variable; the output of the key health index soft measurement module is the predicted key health index F, the index number is q, and the mathematical expression of the corresponding KAN network is as follows: Wherein m is the key index of the mth prediction, ;AndIs a continuous univariate function; represents the t-th non-structural index variable; Parameterizing the univariate one-dimensional function through a B spline curve; spline function Node-basedDefined as a piecewise function, in the intervalThe functional form in is: Wherein the method comprises the steps of Is a coefficient that can be learned and is,As a function of the spline,When C-splines are used; in the key health indicator soft measurement module, the output is expressed as a combination of functions, and the specific KAN formula is as follows: Wherein the method comprises the steps of Represent the firstThe matrix of activation functions of the layer,Representing the composite operation of the function, i.e. the output of the previous layer is taken as the input of the next layer, L is the total layer number of the network; the loss function for KAN network training is as follows: wherein, the first The output of each window is,,Rep