CN-121995238-A - High-fidelity estimation method for lithium ion battery health state under complex working conditions

CN121995238ACN 121995238 ACN121995238 ACN 121995238ACN-121995238-A

Abstract

The invention discloses a high-fidelity estimation method for the health state of a lithium ion battery under a complex working condition, which belongs to the technical field of health state monitoring of the lithium ion battery and comprises the steps of extracting short-time voltage and current fragments from a constant-current constant-voltage charging stage of the battery as model input signals, mapping short-time signal data into entropy domain feature sequences to enhance degradation characterization, constructing a physical information neural network model for fusing an attention mechanism and time sequence modeling to output a battery health state baseline estimation, constructing an explicit sparse dynamics model to realize residual correction of a complex data set, wherein the degradation characteristic comprises a double-weighted data fitting term, an empirical degradation constraint term and a solid phase interface film growth mechanism constraint term as degradation loss constraint. By adopting the method, the invention can realize the cooperative improvement of the estimation precision of the battery health state and the generalization robustness, and can also consider the consistency of the degradation mechanism and the training stability.

Inventors

MA SHAOJUAN
Du haichuan
LIU WEI
Shen Juhong
DING WEIFU
WANG YITU
ZHANG MENGQING

Assignees

北方民族大学

Dates

Publication Date: 20260508
Application Date: 20260402

Claims (8)

1. The high-fidelity estimation method for the lithium ion battery health state under the complex working condition is characterized by comprising the following steps of: S1, general input signals, namely extracting short-time voltage segments and short-time current segments from a constant-current-constant-voltage charging stage of a lithium ion battery to serve as input signals of a model; s2, extracting entropy domain features, namely mapping short-time signal data into an entropy domain feature sequence to enhance degradation characterization; S3, physical loss constraint, namely constructing a physical degradation loss function comprising a double weighted data fitting term, an empirical degradation constraint term and a solid phase interface film growth mechanism constraint term; S4, establishing a model architecture, namely establishing a physical information neural network model integrating an attention mechanism and time sequence modeling to output a SOH baseline estimation of the battery health state; S5, residual adaptive correction, namely constructing an explicit sparse dynamic model to realize residual correction of the complex data set.
2. The high-fidelity estimation method for the lithium ion battery health state under the complex working condition according to claim 1, wherein the short-time voltage segment in S1 is interval data of a constant-current charging stage of the battery, the short-time current segment is interval data of a constant-voltage stage, and a uniformly input charging cycle signal is formed through time alignment and interpolation resampling.
3. The high-fidelity estimation method for the lithium ion battery health state under the complex working condition according to claim 2 is characterized in that the entropy domain feature extraction in S2 adopts a fine composite multi-scale Hilbert cumulative residual entropy algorithm, the short-time charging cycle signal is sequentially subjected to scale decomposition, fourier transformation, hilbert filter and Fourier inverse transformation according to steps to obtain an analytic signal, frequency domain normalization is carried out, the cumulative residual entropy is calculated, fine composite and average are carried out in a scale dimension, and the cumulative residual entropy is mapped into an entropy domain feature sequence, wherein the formula is as follows: ; Wherein, the Is a cumulative distribution function; a charge cycle signal for input; Is a scale-down factor; Expressed in scale Lower (th) Sub-sequence is subjected to Hilbert transform and frequency normalization and then is subjected to discrete index Probability distribution values at (i.e., power spectrum normalization values).
4. The method for estimating the health state of a lithium ion battery under the complex working conditions according to claim 3, wherein the step S3 of constructing the physical degradation loss function specifically comprises the following steps: s31, constructing double weighted data fitting loss, which is specifically expressed as: ; Wherein, the And (3) with Representing true and estimated SOH values, respectively; A time weight set to linearly increase from 0.3 to 1; 、 respectively representing batch size and sequence length, trend perception weight The definition is as follows: , and (3) with Respectively representing the variation of the real value and the estimated value at adjacent moments; S32, constructing a monotonic regularization constraint loss for punishing the rising increment of the SOH prediction track, wherein the monotonic regularization constraint loss is specifically expressed as follows: ; Wherein, the 、 Respectively representing estimated values of adjacent moments; Is a monotonicity tolerance threshold for allowing numerical fluctuations of extremely small magnitude without penalty; for enhancement factors, for adjusting penalty strength; s33, constructing a global slope constraint loss for inhibiting a long-term trend deviation overall descent rule, wherein the global slope constraint loss is specifically expressed as follows: ; Wherein, the 、 Respectively represent the first The 1 st time and the 1 st time of the sample Estimated values for each time; indicating that the maximum value is taken; S34, constructing a mechanism constraint loss for representing the influence of the recyclable lithium loss and capacity fading caused by solid phase interface film SEI growth, wherein the mechanism constraint loss is specifically expressed as follows: ; Wherein, the Respectively represent the nominal capacity of the battery Remaining capacity at the moment; an estimated attenuation value for a state of health of the battery due to solid phase interfacial film growth; The molar gas constant is given as ; Is Faraday constant and takes the value of ; The coupling coefficient is used for SEI growth, and the value is ; Generating reaction density for SEI, and taking value as ; The value is as the reaction level ; The reaction rate is proportional coefficient, which takes the value of ; Respectively representing absolute temperature, characteristic current density and SEI film overpotential parameters, and regarding the parameters as adjustable variables inverted through iterative training in a network; S35, constructing a total loss function, and introducing an adjustable weight coefficient and a dynamic scaling factor to realize self-adaptive balance aiming at the problem that the dimension and the numerical scale of a data fitting item are inconsistent with those of each physical constraint item, wherein the method comprises the following specific expressions: ; wherein each portion scales a factor Is constrained to Within the scope of this invention, In order to assist in losing the term(s), Is a smooth term, and prevents denominator from being too small; The weight coefficient can be regulated for each loss term.
5. The method for estimating the health state of the lithium ion battery under the complex working condition according to claim 4, wherein the physical information neural network model in S4 comprises a multi-head attention feature extraction module and a time sequence modeling module, and specifically comprises the following steps: s41, a multi-head attention feature extraction module is a transducer coding structure comprising an input layer, a coding layer, a multi-head attention module, a feedforward network module and an output layer, captures global dependence characteristics of a time sequence, and converts a current-voltage entropy domain feature sequence into a coded feature representation; s42, the time sequence modeling module is a bidirectional gating circulating unit BiGRU structure comprising an input layer, a double-layer propagation layer, an activation layer and an output layer, models local nonlinear time sequence characteristics through forward and backward gating recursion in parallel, and splices the bidirectional hidden states to obtain time sequence representation containing context information, wherein the bidirectional hidden states are the same as the input layer, the double-layer propagation layer, the activation layer and the output layer The expression of (2) is: ; Wherein, the 、 Respectively representing a forward GRU hidden state and a reverse GRU hidden state; s43, outputting the SOH baseline estimation value by the full connection layer after the outputs of the multi-head attention feature extraction module and the time sequence modeling module are aligned by indexes.
6. The method for estimating the health state of a lithium ion battery under complex working conditions according to claim 5, wherein in S41, the specific steps are as follows: S411, coding each position by combining sine and cosine functions, and ensuring that the model understands the relative and absolute position relation between data; S412, calculating self-attention mechanism weight, capturing the dependency relationship among all data in the input sequence, wherein the self-attention mechanism formula is as follows: ; Wherein, the Is a query matrix; is a key matrix; is a matrix of values; is the dimension of the bond; Representing the dot product between the query and the key, deriving attention weights by softmax function, and finally correlating with the matrix of values Multiplying to obtain a weighted summation result; s413, using multi-head attention mechanism Dividing the space into a plurality of subspaces, and calculating the respective attention output in parallel, wherein the calculated attention formula is as follows: ; Wherein, the Representing the number of heads by a linear transformation weight array Splicing and mapping the outputs of all the heads to the final dimension; s414, performing nonlinear conversion on the representation of each position by using a feedforward neural network, where the feedforward neural network includes two linear conversions and an activation function, and the nonlinear conversion expression is: ; Wherein, the Accepting an input feature vector for the feed-forward network; And Is a weight matrix; And Is a bias term, and ReLU is a nonlinear activation function to increase the nonlinear expression capability of the model.
7. The method for estimating the health state of a lithium ion battery under complex working conditions according to claim 6, wherein the specific steps in S43 are as follows: S431, index alignment and splicing are carried out on the multi-head attention coding output sequence and the BiGRU output sequence according to the same time step, and the fused characteristic sequence expression is obtained: ; Wherein, the Representing the Transformer Structure at the first The attention code feature vectors output by the time steps; BiGRU is shown at the first Bi-directional gating timing feature vectors output by the time steps; Representing vector concatenation operators to form fusion features containing both global dependency characterization and local timing characterization ; S432, integrating the input features extracted from the previous layer into a high-dimensional representation of the current task by the full-connection layer, and outputting an SOH baseline estimate, wherein an output calculation formula is as follows: ; Wherein, the Fusing vectors for the input features; The weight matrix is used for representing the connection weight of the input node and the output node; Is a bias vector for adjusting an activation threshold; To activate a function, to introduce non-linear capabilities, the network is enabled to express more complex feature relationships.
8. The method for estimating the health state of the lithium ion battery under the complex working condition according to claim 7, wherein the explicit sparse dynamic model in S5 performs sparse expression on the residual evolution relationship by constructing a candidate basis function library, and adopts regression optimization solution model parameters containing sparse constraint to obtain a residual prediction value and use the residual prediction value for SOH correction, and the method specifically comprises the following steps: s51, constructing a residual sequence according to the SOH baseline estimation value, wherein the expression is as follows: ; Wherein, the The cycle time is the circulation time; Is residual; is true SOH; outputting SOH for the baseline model; s52, constructing a candidate basis function library for representing residual evolution relation Combining the residual error and its combination with auxiliary variable to form candidate library vector The expression is: ; Wherein, the Is an auxiliary variable; Mapping for candidate basis function library, belonging to A dimension real number vector; comprising polynomials, cross terms 、 A nonlinear function therein for covering the residual evolution form; S53, adopt and include - The mixed regular form of the regular term constructs a sparse constraint regression objective function, and the expression is as follows: ; Wherein, the Residual error at k+1 time; The sparse coefficient vector is to be solved; 、 to control the trade-off coefficient between sparsity and smoothness; S54, representing a discrete evolution relation of residual errors by using sparse linear combination, wherein the expression is as follows: ; Wherein, the As the sparse coefficient vector after the integration, To correct the residual error; S55, the predicted residual error is added to the baseline estimation to obtain a corrected SOH, and the expression is as follows: ; Wherein, the As the estimated value after the residual correction, Is a baseline estimate.

Description

High-fidelity estimation method for lithium ion battery health state under complex working conditions Technical Field The invention relates to the technical field of lithium battery health state monitoring, in particular to a high-fidelity estimation method for the health state of a lithium ion battery under complex working conditions. Background The lithium ion battery increases in internal impedance with increasing charge and discharge cycles, resulting in capacity and power performance degradation. Therefore, accurate assessment of battery state of health to grasp the degradation situation is of great importance for ensuring safe operation of the battery system and for extending its lifetime to the maximum. Existing health state estimation methods can be generally classified into direct measurement methods and indirect analysis methods. The direct measurement method obtains data of the battery under specific working conditions or laboratory conditions through experimental means, and deduces the state of the battery according to the data. For example, ampere-hour integration (coulomb counting) methods achieve state estimation by integrating the current to capacity, but are susceptible to accumulated errors under complex conditions. Other methods such as electrochemical impedance spectroscopy, internal resistance measurement, cycle counting, destructive testing, etc. also have the trade-off problem between accuracy, cost and deployability, limiting its engineering application in battery management systems. The indirect analysis method mainly comprises a model-based method and a data-driven method. The model-based method utilizes an electrochemical mechanism model or an equivalent circuit model to realize state estimation by combining filtering and parameter identification. The method has a certain interpretation, but usually depends on the strong nonlinear and multifactor coupling characteristics of accurate calibration and difficult adaptation to the degradation of the battery, and has higher application cost. The data driving method realizes state estimation by learning performance degradation rules from historical test data. The deep learning method can automatically learn nonlinear association in time sequence data by using a hierarchical network structure, so that estimation accuracy is improved. But the problems of strong data dependence, insufficient generalization across working conditions, unexplained mechanism and the like are generally faced, and the requirements of high precision and high reliability are difficult to meet at the same time. Under different battery types, different use conditions and different aging modes, the degradation mechanism difference is obvious, and a single data model is difficult to meet engineering requirements of high precision and high reliability at the same time. In recent years, physical information neural networks which integrate physical models and deep learning technologies are emerging research methods. It embeds physical priors or mechanism constraints into the network structure or loss function to enhance mechanism consistency and stability. However, the existing physical information neural network method still has the defects of insufficient fusion of physical characteristics and deep characteristics, limited depicting capability of complex declining process and lack of systematic improving means of generalization capability of model structures and constraint forms under different chemical systems and different working conditions. Disclosure of Invention The invention aims to provide a high-fidelity estimation method for the health state of a lithium ion battery under complex working conditions, which realizes the cooperative improvement of the estimation precision, robustness and interpretability of the health state of the battery and effectively reduces the running cost through an integrated framework of entropy domain degradation characterization, mechanism constraint and fusion modeling. In order to achieve the above purpose, the invention provides a high-fidelity estimation method for the health state of a lithium ion battery under complex working conditions, which comprises the following steps: S1, general input signals, namely extracting short-time voltage segments and short-time current segments from a constant-current-constant-voltage charging stage of a lithium ion battery to serve as input signals of a model; s2, extracting entropy domain features, namely mapping short-time signal data into an entropy domain feature sequence to enhance degradation characterization; S3, physical loss constraint, namely constructing a physical degradation loss function comprising a double weighted data fitting term, an empirical degradation constraint term and a solid phase interface film growth mechanism constraint term; S4, establishing a model architecture, namely establishing a physical information neural network model integrating an attention mechanism and ti