CN-121980278-A - Multi-level alignment non-supervision cross-domain health estimation method for lithium battery

CN121980278ACN 121980278 ACN121980278 ACN 121980278ACN-121980278-A

Abstract

The invention relates to an unsupervised cross-domain health estimation method for multi-level alignment of a lithium battery, which comprises the steps of constructing a source domain data set, a source domain capacity label and a target domain data set, carrying out feature extraction, normalization and sliding window sampling to construct the source domain sample set and the target domain sample set, constructing a curve scalar collaborative retrieval attention and gating enhanced multi-layer degradation characterization and alignment framework, constructing a total loss function containing mean square error supervision loss and joint migration alignment loss, carrying out training on the target domain, carrying out capacity prediction on the target domain, outputting a prediction result of a lithium battery capacity degradation curve, and realizing lithium battery health state estimation. The method solves the problem that the traditional method depends on the target domain label and the domain alignment is insufficient, realizes the effective alignment of the physical characteristics of the source/target domain under the condition of no target domain capacity label, inhibits the domain migration and the negative migration, and obtains stable and accurate lithium battery health state estimation under the conditions of cross-data set, cross-platform and cross-working condition.

Inventors

CHEN CHUANG
CHEN LEI
SHI JIANTAO
SHI GE
YUE DONGDONG
BAO DAN
LIU QINYUAN

Assignees

南京工业大学

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation method is characterized by comprising the following steps of: Extracting a constant-current charging voltage sequence, a constant-voltage charging current sequence and a discharging voltage sequence of the lithium battery in each charging and discharging cycle process from the disclosed lithium battery data set, constructing a source domain data set, extracting capacity attenuation data, and constructing a source domain capacity label; The method comprises the steps that a laboratory collects a constant-current charging voltage sequence, a constant-voltage charging current sequence and a discharging voltage sequence of a lithium battery in each charging and discharging cycle process, and a target domain data set is constructed; Respectively carrying out feature extraction and normalization on a source domain data set and a target domain data set to generate a curve type degradation feature set and a scalar degradation feature sequence which are arranged according to a charge-discharge cycle sequence, and constructing a source domain sample set and a target domain sample set through sliding window sampling; Constructing a cross-curve type gating mixed attention module, and converting a curve type degradation characteristic set in a sample set into a curve layer degradation characterization; Constructing a degradation guide frequency spectrum gating enhancement module, and converting scalar degradation characteristic sequences in a sample set into scalar layer degradation characterization; Constructing a scalar guide curve retrieval attention module, receiving curve layer degradation characterization and scalar layer degradation characterization of each sample from a sample set, generating global layer degradation characterization, and outputting a battery capacity predicted value of the next charge-discharge cycle after the last charge-discharge cycle in each sample; the cross-curve type gating hybrid attention module, the degradation guide spectrum gating enhancement module and the scalar guide curve retrieval attention module form a multi-layer degradation characterization and alignment framework of curve scalar collaborative retrieval attention and gating enhancement; And constructing a total loss function, training a multi-layer degradation characterization and alignment framework with enhanced attention and gating based on the collaborative retrieval of a curve scalar by a source domain sample set and a target domain sample set, carrying out capacity prediction on the target domain sample set after training is finished, outputting a prediction result of a lithium battery capacity degradation curve, and realizing the estimation of the health state of the lithium battery.
2. The method for estimating multi-level alignment non-supervision cross-domain health of a lithium battery according to claim 1, wherein the feature extraction and normalization are performed on the source domain data set and the target domain data set respectively to generate a curve type degradation feature set and a scalar degradation feature sequence which are arranged according to a charge-discharge cycle sequence, and the method comprises the following steps: constant-current charging voltage sequence of lithium battery in single charge-discharge cycle in source domain data set and target domain data set Constant voltage charging current sequence Sequence of discharge voltages Denoising, aligning and resampling to obtain constant-current charging voltage curve characteristics with consistent length Constant voltage charging current profile Characteristic of discharge voltage curve Forming a curve feature set, wherein The time index is represented as such, Representing a charge-discharge cycle index; According to constant voltage charging current sequence With constant current charging voltage sequence Calculating the accumulated electric quantity and the accumulated voltage to construct an incremental capacity curve The discrete form of which is expressed as: ; Wherein, the Represent the first Secondary charge-discharge cycle at the first An incremental capacity curve at the sampling points; 、 Respectively represent the first Secondary charge-discharge cycle at the first Accumulated electric quantity and accumulated voltage at the sampling points; is a positive number for preventing denominator from being 0, and takes value ; Curve incremental capacity Resampling to constant current charging voltage curve characteristic Constant voltage charging current profile Characteristic of discharge voltage curve The same length, obtain curve type degradation characteristic Combining the curve characteristic set to obtain a curve type degradation characteristic set ; Setting arbitrary charge-discharge cycle as reference cycle, and recording as Taking a constant-current charging voltage curve and a discharging voltage curve of a reference cycle as a reference constant-current charging voltage curve And a reference discharge voltage curve In combination with incremental capacity curves Extracting scalar degenerate feature sequences The extraction method is as follows: ; ; ; ; ; ; ; ; Wherein, the The standard deviation operator is represented by a set of standard deviation operators, Representing a 25% quantile operator; Representing a vector inner product calculation; representing a binary norm; And Respectively represents the current charge-discharge cycle and the reference cycle in the first Characteristic values of discharge voltage curves at the corresponding sampling points; The total number of sampling points; is a curve type degradation characteristic Is provided for the length of (a), Index set corresponding to preset key voltage interval and ; And (3) with Sampling point sets of a discharging front section interval and a charging rear section interval in a charging and discharging cycle respectively; sampling point indexes of discharge voltage curves representing current charge-discharge cycles and reference cycles in a discharge front section; 、 Respectively representing sampling point sets meeting preset charging related conditions in the current charging and discharging cycle and the reference cycle; Representing a set radix calculation; represent the first Standard deviation characteristics of the secondary charge-discharge cycle increment capacity curve; represent the first 25% Fractional characteristic of adjacent point differential sequence of secondary charge-discharge cyclic increment capacity curve; represent the first The accumulated amplitude characteristic of the constant-current charging voltage curve of the secondary charging and discharging cycle on the charging rear section; represent the first Similarity characteristics between the secondary charge-discharge cycle and the reference cycle constant current charge voltage curve; represent the first Average deviation characteristics of the discharge voltage curve of the secondary charge-discharge cycle relative to the reference cycle; represent the first Accumulated difference characteristics of secondary charge-discharge cycles in a discharge front section; represent the first Overlapping proportion characteristics of secondary charge-discharge cycle and reference cycle on a sampling point set meeting preset charge-related conditions; represent the first Euclidean distance characteristic of the secondary charge-discharge cycle and the reference cycle on the constant voltage charge current curve characteristic.
3. The method for lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation of claim 1, wherein the sliding window sampling is constructed as a source domain sample set and a target domain sample set, and comprises the following steps: Respectively arranging each curve type degradation characteristic set and scalar degradation characteristic sequence according to the charge-discharge cycle sequence with a preset window length Sliding window sampling is carried out to form a plurality of curve type degradation characteristic samples and scalar degradation characteristic samples, and the mathematical expression is as follows: ; ; Wherein, the 、 Respectively expressed by A curve type degradation characteristic sample and a scalar degradation characteristic sample which are constructed for the window starting index on the curve type degradation characteristic set and the scalar degradation characteristic sequence; represent the first A set of curvilinear degradation characteristics for each charge-discharge cycle, Represent the first Scalar degradation feature sequences for each charge-discharge cycle, Is a charge-discharge cycle index; All the curve-type degradation characteristic samples and scalar degradation characteristic samples from the source domain form the source domain sample set according to a charge-discharge cycle sequence, and all the curve-type degradation characteristic samples and scalar degradation characteristic samples from the target domain form the target domain sample set according to the charge-discharge cycle sequence; and the source domain sample set reserves a source domain capacity label corresponding to the source domain data set.
4. The method for estimating multi-level alignment of an unsupervised cross-domain health of a lithium-ion battery of claim 1, wherein said cross-curve type gating hybrid attention module comprises: convolving and pooling each curve feature in the curve type degradation feature set along the length dimension to obtain curve type time embedding ; Constructing curve type description vectors The mathematical expression is as follows: ; Curve type description vector Is the first of (2) The first sample of Individual elements Can be expressed as: ; Wherein, the Representation embedding at Curve type time Carrying out average value operation on the channel dimension and the time dimension; The channel index representing the time-embedding of the curve type, Representing the total channel number of curve type time embedding; the time index is represented as such, Representing a total time; a batch index representing a time-embedded curve type; the method comprises the steps of representing curve characteristic category indexes, wherein the curve characteristic categories comprise four categories of constant-current charging voltage curve characteristics, constant-voltage charging current curve characteristics, discharging voltage curve characteristics and curve type degradation characteristics; further obtaining curve type weight through learning linear gating mapping The mathematical expression is as follows: ; Wherein, the Representing softmax normalization over the curve feature class dimension; 、 Respectively representing a learnable weight matrix and a bias term of the learnable linear gating map; obtaining the degradation characterization of the initial curve layer by adopting gate control weighted fusion The mathematical expression is as follows: ; Wherein, the Represent the first Batch No. Weights of class curve features; representing the characteristic category number of the curve; characterization of initial Curve layer degradation Applying a lightweight time-series refining residual to obtain a curvilinear layer degradation characterization The mathematical expression is as follows: ; Wherein, the And the light time sequence refining residual error module is formed by depth separable one-dimensional convolution and point-by-point convolution.
5. The method for lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation of claim 1, wherein the degradation-oriented spectrum gating enhancement module comprises: Mapping scalar channels of each scalar degradation feature in a sequence of scalar degradation features to a hidden space by point-wise convolution to obtain scalar type temporal embedding ; Time embedding scalar types Performing dimension substitution and performing real number fast Fourier transform along the time dimension, wherein the mathematical expression is as follows: ; ; Wherein, the Representing a dimension permutation function; embedding scalar type time after dimension substitution; representing a real fast fourier transform; representing scalar type frequency domain embedding; The number of frequency bins is indicated, Representing a rounding down operation; defining frequency energy: ; Wherein, the Represent the first Batch No. The frequency energy of the frequency component, Representing scalar type frequency domain embedding Is a channel index of (2); Square computation representing complex modulus; normalizing the frequency energy with the energy median to obtain normalized frequency energy, wherein the normalized frequency energy is expressed in the following mathematical expression: ; Wherein, the Represent the first Batch No. Normalized frequency energy of the frequency components; represent the first The energy median of a batch, defined as , Representing a median operation; Is a minimum positive number for preventing denominator from being 0, takes value ; Constructing a degradation-oriented high-frequency gating mask: ; Wherein, the Representation of representation No Batch No. A high frequency gating mask of the frequency components; representing sigmoid activation; Is a learnable temperature coefficient; Is a threshold superparameter; high-frequency gating mask Frequency domain embedding for scalar types The low frequency and gating high frequency components of the process are differentially increased, and the process is expressed as follows: ; Wherein, the Representing a scalar type frequency domain representation; 、 Respectively representing low-frequency learnable complex weights and high-frequency learnable complex weights; Representing element-by-element multiplication; the scalar layer degradation characterization is obtained by inverse transformation, and the mathematical expression is as follows: ; Wherein, the Representing scalar layer degradation characterizations; for dimension permutation functions An inverse function of (2); is a real number fast Fourier transform Is an inverse transform of (a).
6. The method for lithium battery multi-level alignment based unsupervised cross-domain health estimation of claim 1, wherein the scalar guide curve retrieval attention module comprises: characterizing curve layer degradation from each sample in a sample set from a cross-curve type gated mixed attention module And scalar layer degradation characterization from each sample in the sample set from the degradation-oriented spectrum gating enhancement module Performing dimensional substitution to form a sequence representation: ; ; ; Wherein, the Representing a dimension permutation function; A sequence of queries is represented and, The sequence of the keys is represented by a sequence of keys, Representing a sequence of values; the multi-head attention mechanism is utilized for cross-modal alignment fusion, and the process is expressed as follows mathematically: ; Wherein the method comprises the steps of Representation is based on query sequences Key sequence Sequence of values Is a multi-headed attention operator of (c), Representing an attention fusion feature; Fusion door with leading-in gate control enhancement layer structure The mathematical expression is as follows: ; Wherein the method comprises the steps of Representing a sequence of queries Attention fusion feature Splicing; representing sigmoid activation; 、 respectively representing a learnable weight matrix and a bias term of a gating enhancement layer; fusion door Performing gating fusion to obtain global layer degradation characterization The mathematical expression is as follows: ; Wherein the method comprises the steps of Representing element-by-element multiplication; Characterization of global layer degradation Performing inverse dimension displacement, and performing regression prediction by taking the end time step part to obtain the battery capacity predicted value of the next charge-discharge cycle after the last charge-discharge cycle in each sample This process is mathematically expressed as follows: ; Wherein the method comprises the steps of Representing a regression header network; for dimension permutation functions An inverse function of (2); Representation pair Slicing and retaining only Characterization after the time step, Representing the end time step index.
7. The lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation method of claim 1, wherein the constructing the total loss function trains a multi-layer degradation characterization and alignment framework of curve scalar collaborative retrieval attention and gating enhancement based on a source domain sample set and a target domain sample set, and comprises the following steps: the total loss function comprises a mean square error supervision loss and a joint migration alignment loss; the method comprises the steps of sending a curve type degradation characteristic set and a scalar degradation characteristic sequence of a current charge-discharge cycle in a source domain sample set and a target domain sample set into a curve scalar collaborative retrieval attention and gating enhanced multi-layer degradation characterization and alignment framework together, and respectively obtaining each degradation characterization of a source domain and a target domain and a battery capacity prediction value of a next charge-discharge cycle; calculating a mean square error supervision loss by using a battery capacity predicted value of a source domain and a capacity label of the source domain ; Introducing joint migration alignment loss consisting of a maximum mean difference loss term and a covariance alignment loss term between degradation characterizations from a source domain sample set, from a target domain sample set, expressed mathematically as follows: ; Wherein, the Alignment loss for joint migration; for the maximum mean value difference loss term, Representing a maximum mean difference loss calculation, 、 Representing degradation characterizations from the source domain sample set, from the target domain sample set, respectively, including a curvilinear layer degradation characterization Scalar layer degradation characterization Global layer degradation characterization , For the index of the three degradation characterizations described above, Index sets characterizing the three degradation; 、 Respectively the first Seed degradation characterizes pairs Ji Quan weights in the maximum mean difference loss term, in the covariance alignment loss term; the loss terms are aligned for the covariance, Representing covariance alignment loss calculation; And (3) with The weight coefficients of the maximum mean difference loss term and the covariance alignment loss term are respectively; monitoring losses with mean square error Loss of alignment with joint migration The sum is the total loss function The mathematical expression is as follows: 。
8. The method for estimating multi-level alignment non-supervision cross-domain health of a lithium battery according to claim 1, wherein the method for estimating the capacity of the target domain sample set after training is completed and outputting the prediction result of the capacity degradation curve of the lithium battery comprises the following steps: sending each curve type degradation characteristic sample and scalar degradation characteristic sample in the target domain sample set into a multi-layer degradation characterization and alignment frame with enhanced attention and gating through the trained curve scalar collaborative retrieval, and predicting a battery capacity predicted value of the next charge and discharge cycle after the last charge and discharge cycle in each sample; And forming a lithium battery capacity degradation curve according to the predicted battery capacity values of all the charge and discharge cycles according to the charge and discharge cycle sequence.
9. An electronic device comprising at least one processor and a memory communicatively coupled to the at least one processor, wherein the memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation method of any one of claims 1-8.
10. A computer readable storage medium storing computer instructions for causing a processor to implement the lithium battery multi-level alignment-oriented unsupervised cross-domain health estimation method of any one of claims 1-8 when executed.

Description

Multi-level alignment non-supervision cross-domain health estimation method for lithium battery Technical Field The invention relates to the technical field of lithium battery monitoring state estimation, in particular to an unsupervised cross-domain health estimation method for multi-level alignment of a lithium battery. Background Along with the rapid expansion of electric traffic and energy storage, lithium ion batteries become key energy units of electric automobiles, energy storage power stations and portable equipment. Battery state of health (SOH) determines safety boundaries, operational reliability, and overall life costs, so accurate assessment and trend prediction are the basis for BMS and operation and maintenance decisions. Existing methods include mechanism models, experience/statistics models, and data driven models. The mechanism model depends on a complex electrochemical equation, has multiple parameters and is difficult to identify, the working condition adaptability is limited, the experience/statistics method depends on artificial features and is easily influenced by feature selection and scene change, and the data driving method has strong fitting capacity, but the training/testing is generally assumed to be distributed in the same way, and domain deviation and generalization drop can be caused by different batches, platforms and temperature multiplying power differences. To mitigate domain offset, transfer learning is introduced into SOH estimation, and source domain labeling is used to improve target domain performance. Most methods are supervised/semi-supervised, still require target domain labels, are difficult to meet for 'label-free' deployment, and may be negatively migrated due to insufficient alignment. In practice, the capacity of the target domain is marked with difficult continuity and difficult quality, and signals such as voltage, current, time and the like are easier to acquire and have lower cost. Therefore, the method can be used for realizing the cross-domain estimation without labels in the target domain by only relying on physical characteristics such as constant-current charging voltage, constant-voltage charging current, discharging voltage, derivative increment capacity and the like. The method has the advantages that the characteristics are obviously disturbed by noise and working conditions, the charging and discharging strategies in different domains change curve forms and statistical structures, simple migration is difficult to realize, degradation simultaneously comprises local change and global drift of a key area, multi-scale coupling is realized, and multilayer characterization is required to be stably aligned and fused. Therefore, how to realize effective alignment of physical features of a source/target domain under the condition of no target domain capacity label, inhibit domain offset and negative migration, and obtain stable and accurate SOH estimation under the conditions of cross-data set, cross-platform and cross-working conditions is a key problem to be broken through. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an unsupervised cross-domain health estimation method for multi-level alignment of a lithium battery, which solves the problem that the traditional cross-domain lithium battery health estimation method based on transfer learning depends on a target domain label and is insufficient in domain alignment, realizes effective alignment of source/target domain physical characteristics under the condition of no target domain capacity label, inhibits domain offset and negative migration, and obtains stable and accurate SOH estimation under the conditions of cross-data set, cross-platform and cross-working condition. In order to achieve the technical aim, the invention provides a lithium battery multi-level alignment-oriented non-supervision cross-domain health estimation method, which comprises the following steps: Extracting a constant-current charging voltage sequence, a constant-voltage charging current sequence and a discharging voltage sequence of the lithium battery in each charging and discharging cycle process from the disclosed lithium battery data set, constructing a source domain data set, extracting capacity attenuation data, and constructing a source domain capacity label; The method comprises the steps that a laboratory collects a constant-current charging voltage sequence, a constant-voltage charging current sequence and a discharging voltage sequence of a lithium battery in each charging and discharging cycle process, and a target domain data set is constructed; Respectively carrying out feature extraction and normalization on a source domain data set and a target domain data set to generate a curve type degradation feature set and a scalar degradation feature sequence which are arranged according to a charge-discharge cycle sequence, and constructing a source domain sample set and a target