CN-121997099-A - Lithium battery health state evaluation method and system based on multi-modal mechanism constraint attention fusion network

CN121997099ACN 121997099 ACN121997099 ACN 121997099ACN-121997099-A

Abstract

The invention provides a lithium battery health state assessment method and system based on a multi-modal mechanism constraint attention fusion network, the method comprises the following steps of 1, carrying out multi-dimensional data acquisition by using a data extraction module, 2, preprocessing an original data set through cross-modal redundancy suppression and complementary enhancement, 3, transmitting a standardized multi-modal feature set into the multi-modal mechanism constraint force fusion network, 4, constructing an improved COA optimization training model which is suitable for battery characteristics and has multi-scale element learning, 5, inputting model output into the multi-modal mechanism constraint force fusion network to obtain an SOH predicted value, and verifying a predicted result to realize online incremental update prediction of the battery health state. The invention can solve the technical pain points of low precision, poor interpretability, weak generalization capability and the like of the existing lithium battery SOH evaluation model, provides a reliable technical scheme for new energy battery health management, and has wide engineering application fields.

Inventors

CHEN YONGHAO
ZHANG CHU
WANG ZHI
Han Mingen
PENG TIAN
WANG HAO
ZHANG YAN

Assignees

淮阴工学院

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. A lithium battery health state evaluation method based on a multi-modal mechanism constraint attention fusion network is characterized by comprising the following steps: Step 1, acquiring multi-dimensional data including electric parameters, mechanical deformation parameters and environmental working condition parameters through a data extraction module to form a multi-mode original data set; Step 2, preprocessing an original data set through cross-modal redundancy suppression and complementary enhancement, firstly performing sub-modal noise suppression, sequentially smoothing electric parameters by using a variable weight self-adaptive index, smoothing deformation parameters by using a self-adaptive trend factor index, and processing working condition parameters by using a temperature gradient weighted median filter, and performing cross-modal redundancy elimination and complementary feature enhancement after processing to obtain a standardized multi-modal feature set; Step 3, transmitting the standardized multi-modal feature set into a multi-modal mechanism constraint force fusion network MMC-SOH, processing the standardized multi-modal feature set through a multi-modal feature coding layer, sequentially extracting deformation parameter local features and working condition multi-scale weighted fusion network fusion working condition features through a time sequence attention-gated twin-cycle network coding electric parameter time sequence feature and a deformation gradient perception dynamic convolution network, realizing feature depth fusion through a mechanism constraint attention fusion layer, and obtaining an SOH regression output layer through dynamic attention weight calculation and feature fusion; Step 4, pre-training the model based on multi-scale element learning pre-training through multi-scale element learning and adapting to an improved COA optimization training model of battery characteristics, optimizing super-parameters of the model by adopting an improved raccoon algorithm, and transmitting optimized data into an objective function finally comprising a prediction error term, an L2 regular term and a mechanism constraint term by combining weighted sliding filtering and dynamic attention weight calculation and dynamic adjustment training process; And 5, outputting and inputting the trained model into a multi-modal mechanism constraint force fusion network, obtaining an SOH predicted value through an SOH regression output layer, and simultaneously verifying a predicted result to realize online incremental update prediction of the battery health state.
2. The method of claim 1, wherein in step 1, the electrical parameters include voltage, current, internal resistance and capacity attenuation rate, the mechanical deformation parameters include battery case shape variable and tab deformation amplitude, the mechanical deformation parameters are collected by a laser displacement sensor and a strain gauge synchronously, and the environmental condition parameters include environmental temperature, charge-discharge multiplying power and cycle times, and the mechanical deformation parameters are collected in real time by a temperature sensor and a charge-discharge controller.
3. The method according to claim 2, wherein step 2 comprises: step 2.1, firstly, noise suppression is carried out, and variable weight self-adaptive exponential weighting sliding filtering processing electric parameters are adopted, which concretely comprises the following steps: Taking the average value of the first 3 electrical parameter sampling points as a filtering initial value: , Wherein, the Is the first A plurality of original electrical parameter sampled values, An initial value for filtering the electrical parameter; calculating the relative fluctuation coefficient of the electrical parameter at the moment k in real time : , , Wherein, the Is the first The original electrical parameter sample value at the moment in time, Is the first The electric parameter value after time filtering, and the index weight is set according to the fluctuation coefficient When (when) When the temperature is less than or equal to 0.02, =0.1, Enhancing the filter smoothness, when 0.02< When the temperature is less than or equal to 0.1, =0.3, Balance smoothing and responsiveness; in the case of >0.1 of the total number of the components, =0.8, Improving the tracking ability of the real signal; And 2.2, adopting an adaptive trend factor index smoothing process deformation parameters, introducing a smooth weight for distinguishing reversible deformation and irreversible deformation of the deformation trend factor, wherein the adaptive trend factor index smoothing formula is as follows: , Wherein the method comprises the steps of Is that The deformation amount after the moment is smooth, Is that The amount of the original deformation at the moment, In order to smooth the coefficient of the coefficient, Is a deformation trend factor; And 2.3, adopting temperature gradient weighted median filtering processing working condition parameters, setting weight coefficients based on temperature gradients, and realizing noise suppression through a special median formula, wherein the temperature gradient weighted median filtering formula is as follows: , , Wherein, the Is that The temperature after the filtering at the moment, Is a median operation; In order to provide a temperature gradient, Is a weight coefficient based on temperature gradient; And then performing cross-modal redundancy inhibition, calculating mutual information entropy between any two modal features, extracting main component features of the redundant modes through main component analysis, retaining complementary information, and eliminating repeated features.
4. The method according to claim 3, wherein the step 2 further comprises the step of aligning the deformation parameter and the working condition parameter to the same time scale by cross-modal time alignment based on the electric parameter time stamp, and specifically comprises the following steps: constructing distance matrix by setting electric parameter time sequence as , Represents the mth data point in the sequence A, and the time sequence of the mechanical deformation parameter or the environmental working condition parameter to be aligned is , Representing the nth data point in sequence B, m.noteq.n, constructing an m n distance matrix D, wherein Representation of Euclidean distance of (c): ; Find the optimal regular path from the upper left corner of distance matrix D To the lower right corner Searching a path with the smallest accumulated distance according to the rule of only right, downward and downward right, wherein the point on the path is the optimal matching pair of the two groups of sequences; interpolation and resampling complement, namely, according to the optimal matching path, carrying out linear interpolation or resampling on the mechanical deformation parameter and the environmental working condition parameter with the sampling frequency of 50Hz, and completing to the same time scale as the electrical parameter; The pearson correlation coefficient of each feature and the health state SOH is calculated through mechanism constraint feature screening, the core feature of the |r| is reserved and is more than or equal to the threshold value, and improved maximum and minimum normalization processing feature data are adopted; the pearson correlation coefficient r has the following calculation formula: , Wherein, the A modal feature value for the first sample; Is that Is a sample mean value of (2); Is the first True SOH values corresponding to the samples; is the sample mean of SOH values; the total amount of the sample; and finally, carrying out standardization processing on the screened core features by using maximum and minimum normalization, and outputting a standardized multi-mode feature set.
5. The method according to claim 4, wherein in step 3, the multimode mechanism constraint force fusion network MMC-SOH includes a multimode feature coding layer, a mechanism constraint attention fusion layer, an SOH regression output layer, and a multi-objective function, and specifically includes the following steps: step 3.1, coding electrical parameter time sequence characteristics by using a time sequence attention gate control twin circulation network, capturing full-period bidirectional time sequence dependence of charge and discharge through a twin circulation unit, and outputting high-dimensional time sequence characteristics by combining time sequence attention gate control dynamic strengthening key aging stage characteristics: the twin circulation unit comprises a forward circulation unit and a reverse circulation unit; the forward circulation unit is expressed as: , The reverse circulation unit is expressed as: , Wherein, the Representation of Time electrical parameter characteristics; And Is a weight matrix of the forward cyclic unit, And A weight matrix for the reverse circulation unit; is the offset top of the forward circulation cell, For the offset top of the reverse circulation unit, Is that The state is hidden in the forward direction at the moment, The state is reversely hidden for the moment t, Is a hyperbolic tangent function; calculating time sequence attention gating: , calculating gating strengthening characteristics: , Wherein, the 、 In order to pay attention to the weight matrix, As a result of the bias term, Is that The moment of time attention weight is given to the person, Is that The time gating strengthens the post-characteristic, For the length of the time sequence, Is a natural exponential function; Output electrical parameter encoding features : , Wherein, the ( ) The full connection layer is used for compressing the characteristic dimension to a unified hidden layer; Extracting local characteristics of deformation parameters through a deformation gradient sensing dynamic convolution network, introducing a deformation gradient sensing mechanism, and strengthening local deformation mutation characteristic extraction caused by irreversible aging through dynamically adjusting convolution kernel weight, and simultaneously inhibiting noise interference, wherein the method specifically comprises the following steps of: calculating local gradients of deformation parameters Wherein The normalized value of the mechanical deformation parameter at the moment l-i+1; By constructing a gradient perceptual weight adjustment term: , Wherein, the For a convolution kernel weight of a fixed size, In order to convolve the offset term, For the ReLU activation function, flat is feature flattening, maxPool is max pooling, Is a high-dimensional deformation feature; and 3.3, fusing the characteristic parameters of the working conditions through a working condition multi-scale weighting fusion network: Calculating an adaptive weight: , , , Wherein the method comprises the steps of Is normalized temperature weight, Is a standardized charge-discharge multiplying power weight, The characteristic weight of the cycle times after standardization; , for the working condition of the temperature, Is used for the working condition of charge-discharge multiplying power, For the gradient of the variation of the cycle number working condition, obtaining high-dimensional working condition characteristics : , Wherein the method comprises the steps of Is a fusion bias term.
6. The method according to claim 5, wherein step 3 further comprises the steps of inputting the characteristics of the three modes into a mechanism constraint attention fusion layer, introducing a battery aging mechanism constraint coefficient, dynamically calculating the attention weight of each mode, realizing the weighted fusion of the characteristics of the multiple modes, setting the constraint coefficient based on the battery aging mechanism, distributing a working condition self-adaptive coefficient for each mode, and carrying out high-temperature working condition: =1.4; fast charge regime: =1.3; the attention weight calculation formula is: , the multi-modal feature fusion is: , Wherein, the Is the first Attention weights of the individual modes satisfy , Is the first Feature importance scores for the individual modality features, Is the first The mechanism constraint coefficients of the individual modes, Is the first The feature importance scores of the individual modalities are presented, Is the first Mechanism constraint coefficients of individual modes; as a result of the global features after the fusion, As the attention weight of the electrical parameter modality, As the attention weight of the mechanical deformation mode, Is the attention weight of the environmental working condition mode.
7. The method of claim 6, wherein step 3 further comprises inputting the fused features into an SOH regression output layer to achieve a nonlinear mapping from high-dimensional features to SOH values while adapting the embedded deployment; And (3) outputting an SOH predicted value by constructing an objective function comprising a prediction error term, an L2 regular term and a mechanism constraint term, so as to obtain a final objective function L: , wherein the prediction error term L2 regular term , All weight matrixes of the network, mechanism constraint items , A mechanism mapping value for the fusion feature; is a health index theoretical value based on aging mechanism, and attention weight rationality item , For the number of samples in a batch, Is the first SOH predictions for the individual samples, Is the first SOH true values for the individual samples, Is a regular weight of the L2, For a set of network ownership weight parameters, The weights of the terms are constrained for the mechanism, ( ) As a mechanism mapping function of the fusion features, A function is calculated for the health index theory, To pay attention to the weight rationality term weight, Is the first Sample number 1 The attention weight of the modality, Is the first Theoretical weighting of the modality.
8. The method of claim 7, wherein step 4 comprises: Step 4-1, initializing, namely setting a population scale, a maximum iteration number and a search space boundary, randomly generating individual positions of an initial raccoon population, calculating the fitness value of each individual, and determining an initial global optimal solution; and 4-2, repeating the following steps until the termination condition is met: The exploration stage adopts the following improved formula: , Wherein, the Is the first Candidate updated positions of individuals in the exploration phase; Is the first The current position of the individual before updating; In order to explore the inertial weight coefficients of the phase, The maximum value is searched for the global of the variable, The minimum value is searched for the global of the variable, Generating a uniform random number between 0 and 1, Is the global optimal individual position of the t-1 generation; The peeling stage adopts the following improved formula: , Wherein, the Is the first Candidate locations updated by individuals during the peeling stage, And Respectively represent the first Local search lower and upper bounds of the dimension variable; Is an exponential function of the number of times, Is the rate of decay adjustment factor, Is the maximum number of iterations; step 4-3, terminating the judgment, namely stopping iteration if the maximum iteration number is reached or the global optimal solution is continuously lifted for two generations without convergence to the preset precision; and 4-4, outputting a final global optimal solution and an adaptability value.
9. The method of claim 8, wherein step 5 includes inputting the preprocessed features into the optimized network, triggering the corresponding pre-warning level, and simultaneously calculating RMSE, MAE, R2 as the accuracy evaluation index.
10. A lithium battery health state evaluation system based on a multi-modal mechanism constraint attention fusion network, which is realized by adopting the method according to any one of claims 1-9, and is characterized by comprising a data extraction module, a preprocessing module, a multi-modal mechanism constraint attention fusion network module, a training module and an accuracy evaluation module; the data extraction module is used for collecting multidimensional data comprising electric parameters, mechanical deformation parameters and environmental working condition parameters to form a multimodal original data set; the preprocessing module preprocesses the original data set through cross-modal redundancy suppression and complementary enhancement, firstly carries out sub-modal noise suppression, sequentially smoothing the electrical parameters by using a variable weight self-adaptive index, smoothing the deformation parameters by using a self-adaptive trend factor index, and carrying out temperature gradient weighted median filtering processing on the working condition parameters, and carrying out cross-modal redundancy elimination and complementary feature reinforcement after processing to obtain a standardized multi-modal feature set; The multimode mechanism constraint force fusion network module transmits a standardized multimode feature set into a multimode mechanism constraint force fusion network, the standardized multimode feature set is processed through a multimode feature coding layer, deformation parameter local features and working condition multi-scale weighting fusion network fusion working condition features are sequentially extracted through a time sequence attention gating twin circulation network coding electric parameter time sequence feature and a deformation gradient perception dynamic convolution network, feature depth fusion is realized through a mechanism constraint attention fusion layer, and an SOH regression output layer is obtained through dynamic attention weight calculation and feature fusion; The training module optimizes the training model through multi-scale element learning and improved COA (chip on array) of adapting battery characteristics, pre-trains the model based on multi-scale element learning pre-training, optimizes the super-parameters of the model by adopting an improved raccoon algorithm, combines weighted sliding filtering and dynamic attention weight calculation and dynamic adjustment training process, and transmits the optimized data into an objective function finally comprising a prediction error item, an L2 regular item and a mechanism constraint item; And the precision evaluation module outputs and inputs the trained model into a multi-modal mechanism constraint force fusion network, obtains an SOH predicted value through an SOH regression output layer, verifies a predicted result and realizes online incremental update prediction of the battery health state.

Description

Lithium battery health state evaluation method and system based on multi-modal mechanism constraint attention fusion network Technical Field The invention relates to the technical field of new energy battery health management, machine learning and multi-mode data fusion intersection, in particular to a lithium battery health state assessment method and system based on a multi-mode mechanism constraint attention fusion network. Background The lithium battery is a core energy storage element in the field of new energy, the health state of the lithium battery, namely SOH, directly determines the running reliability and safety of equipment, and the definition of SOH is the ratio of the current usable capacity of the battery to the initial rated capacity, and is a core index for battery life prediction and fault early warning. Three key pain points exist in the existing SOH evaluation technology: The characteristic dimension is single, the traditional method depends on electric parameters such as voltage, current, internal resistance and the like, and omits the coupling influence of mechanical deformation (shell expansion and tab deformation) and environmental working conditions (temperature and charge-discharge multiplying power) generated in the aging process of the battery, so that characteristic information is on the one hand, and the evaluation accuracy is insufficient under the complex working conditions. The mainstream deep learning models like LSTM and CNN are poor in physical interpretability, and focus on fitting by means of data driving, but physical mechanisms of battery aging are not integrated into the models, such as capacity fading rules caused by SEI film growth and lithium dendrite deposition, and the like, so that the models have 'black box' characteristics, prediction results lack support in physical logic, and reliability in engineering application is low. The dynamic working condition generalization capability is weaker, namely, in practical application, the battery can suffer from dynamic working conditions such as temperature fluctuation, charge-discharge multiplying power conversion, cycle number accumulation and the like, most of the existing models are developed under a single working condition when training is performed, and the adaptive optimization is not performed for multiple working conditions, so that the prediction error is greatly increased when the working conditions are migrated. Disclosure of Invention The invention aims to: the invention hopefully overcomes the technical defects of the existing lithium battery SOH evaluation model, provides a lithium battery health state evaluation method and system based on a multi-modal mechanism constraint attention fusion network, and provides a multi-modal network based on mechanism constraint attention fusion, which relates to links such as multi-dimensional data acquisition, self-adaptive preprocessing, mechanism data fusion coding, double-stage optimization training and the like, so that the purposes of high-precision quantitative evaluation of SOH under dynamic working conditions are achieved, and the physical interpretability and multi-working condition adaptation capability of the model are improved. The method comprises the following steps: Step 1, acquiring multi-dimensional data including electric parameters, mechanical deformation parameters and environmental working condition parameters through a data extraction module to form a multi-mode original data set; Step 2, preprocessing an original data set through cross-modal redundancy suppression and complementary enhancement, firstly performing sub-modal noise suppression, sequentially smoothing electric parameters by using a variable weight self-adaptive index, smoothing deformation parameters by using a self-adaptive trend factor index, and processing working condition parameters by using a temperature gradient weighted median filter, and performing cross-modal redundancy elimination and complementary feature enhancement after processing to obtain a standardized multi-modal feature set; Step 3, transmitting the standardized multi-modal feature set into a multi-modal mechanism constraint force fusion network MMC-SOH, processing the standardized multi-modal feature set through a multi-modal feature coding layer, sequentially extracting deformation parameter local features and working condition multi-scale weighted fusion network fusion working condition features through a time sequence attention-gated twin-cycle network coding electric parameter time sequence feature and a deformation gradient perception dynamic convolution network, realizing feature depth fusion through a mechanism constraint attention fusion layer, and obtaining an SOH regression output layer through dynamic attention weight calculation and feature fusion; And 4, optimizing a training model by means of multi-scale element learning and improved COA (chip on array) matching with battery characteristics, firstl