A method and system for modeling dynamic attenuation behavior of SOH of a power battery

Abstract

This invention relates to the field of power battery health management technology, specifically to a method and system for modeling dynamic SOH (State of Health) degradation behavior of power batteries. The method includes: constructing a multi-dimensional stress factor time-series matrix by real-time acquisition of battery operating data; performing adaptive operation scenario segmentation based on stress mutation points; converting each scenario segment into a behavior vector and embedding it into a nonlinear feature space; combining historical segments with weighted fusion to form an enhanced feature vector; employing a threshold-pattern joint identification mechanism to determine degradation trigger segments and perform risk classification; quantifying the segment-level degradation contribution and dynamically updating feature weights; integrating short-term, medium-term, and long-term degradation rates through multi-scale time windows to achieve SOH prediction; combining confidence interval evaluation and parameter / structure-level dual-layer adaptive correction to form a closed-loop optimization; and outputting real-time SOH indicators and early warning strategies. This invention significantly improves the modeling accuracy and adaptability under dynamic operating conditions.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power battery health management, and particularly relates to a power battery SOH dynamic attenuation behavior modeling method and system. BACKGROUND

[0002] With large-scale application of electric vehicles and energy storage systems, accurately quantifying the health state (SOH) attenuation behavior of power batteries under complex dynamic working conditions has become a key technical challenge to improve the service life and safety of the batteries.

[0003] A battery health prediction method, system, storage medium and program product are disclosed in Chinese Invention Patent No. CN119471453B. In the method, driving data of a target battery pack is obtained; a micro electrochemical stress model is established based on instantaneous working condition data, a meso stress evolution model is constructed based on single trip data, and a macro stress accumulation model is generated based on long-term use data; the micro electrochemical stress model, the meso stress evolution model and the macro stress accumulation model are integrated to form an electrochemical stress field; a stress trigger model is constructed based on the electrochemical stress field; a battery performance compensation equation is established according to the output result of the material response function; the current driving data is input into the stress trigger model to obtain a real-time stress field distribution state; the stress field distribution state is input into the battery performance compensation equation to obtain an updated driving strategy constraint condition; and the battery health state is predicted according to the updated driving strategy constraint condition combined with the historical stress accumulation data.

[0004] In actual operation, the time sequence interaction effect of multi-dimensional stress factors (current, temperature, SOC and rate) and the historical cumulative influence thereof put higher requirements on the modeling accuracy of SOH; at the same time, the industry urgently needs to establish a dynamic model with real-time recognition ability of attenuation trigger, multi-scale forward prediction and online adaptive correction mechanism to support the battery management system to make accurate life management decisions. SUMMARY

[0005] The application aims to solve the problems in the background art and provides a power battery SOH dynamic attenuation behavior modeling method and system.

[0006] The technical scheme of the application is a power battery SOH dynamic attenuation behavior modeling method, which comprises the following specific implementation steps:

[0007] S1, real-time collection of battery operation data to construct a multi-dimensional stress factor time sequence matrix, adaptive operation scene segmentation based on stress mutation points, and conversion of each scene segment into a behavior vector;

[0008] S2, embedding the behavior vector into a nonlinear feature space and combining historical segments for weighted fusion to form an enhanced feature vector; adopting a threshold-pattern combined recognition mechanism to determine the decay trigger segment and perform risk grading;

[0009] S3, quantifying the segment-level decay contribution amount, dynamically updating the feature weight for decay recognition, and mapping it to a real-time SOH indicator and early warning level;

[0010] S4, integrating short-term, medium-term and long-term decay rates through a multi-scale time window to achieve SOH prediction, combining confidence interval evaluation and parameter and structure-level double-layer adaptive correction to form a closed-loop optimization.

[0011] Preferably, the generation process of the behavior vector in step S1 includes:

[0012] Real-time acquisition of the current, temperature, state of charge and rate of the power battery to form a time series matrix;

[0013] Normalizing each factor and introducing a relative contribution mapping operator to obtain a normalized and weighted stress factor matrix;

[0014] By calculating the stress factor change amplitude and setting a threshold, the time series data is adaptively divided into several operation scenario segments;

[0015] For each scenario segment, the integral mean of the time length is calculated to obtain the behavior vector of the segment.

[0016] Preferably, the embedding process of embedding the behavior vector into a nonlinear feature space in step S2 is specifically:

[0017] The behavior vector is mapped to a high-dimensional feature space through a nonlinear function to form a high-dimensional embedded feature vector; the nonlinear function includes element-wise sine mapping, element-wise exponential mapping of the vector, and their combinations, i.e.

[0018] ;

[0019] Where F i represents the high-dimensional embedded feature vector of the i-th operation scenario segment; represents the element-wise multiplication operation of the vector; sin(B i ) represents the element-wise sine mapping of the behavior vector; represents the element-wise exponential mapping of the behavior vector; a represents the exponential mapping adjustment coefficient; B i represents the behavior vector of the i-th operation scenario.

[0020] Preferably, the fusion process of combining historical segments for weighted fusion to form an enhanced feature vector in step S2 is specifically:

[0021] The embedding feature vector of the current segment is fused with the embedding feature vectors of the previous k historical segments to obtain an enhanced feature vector with fused historical information:

[0022] ;

[0023] wherein, represents the enhanced feature vector of the i-th segment; β represents a historical information fusion coefficient; k represents a historical window length; w i,j represents a historical segment decay weight, reflecting the influence size of the past j-th segment on the current i-th segment.

[0024] Preferably, the threshold-mode joint recognition mechanism in step S2 is specifically:

[0025] The enhanced feature vector is input into a decay-sensitive function to calculate a comprehensive output value thereof:

[0026] ;

[0027] wherein, represents the decay-sensitive function; represents the m-th component of the enhanced feature vector; p m represents a nonlinear index; γ m represents a component weight; and M represents the total number of dimensions of the enhanced feature vector.

[0028] The output value is compared with a preset decay trigger threshold value, and if the output value exceeds the threshold value, it is determined that the segment triggers a decay event.

[0029] According to the size range of the output value, the triggered segment is classified into a risk level including a slight trigger, a moderate trigger, or a strong trigger.

[0030] Preferably, the quantification of the segment-level decay contribution in step S3 is specifically:

[0031] For the segment determined to trigger a decay event, the SOH decay amount of the segment is calculated according to the enhanced feature vector and the decay sensitivity coefficient λ i .

[0032] The decay amount of the current segment is combined with the historical decay cumulative amount, and a historical influence factor is introduced to simulate the cumulative acceleration effect, so as to calculate the cumulative SOH decay amount up to the current segment.

[0033] Preferably, the dynamic updating of the feature weight for decay recognition in step S3 is specifically: based on the deviation between the model-predicted segment decay amount and the actual observation value, the feature weight in the decay-sensitive function is dynamically updated by the gradient descent method: ;

[0034] wherein, denotes the weight of the mth enhanced feature component in the decay sensitive function after update; γ denotes the learning rate of weight update; m denotes the weight of the mth enhanced feature component in the decay sensitive function before update; η denotes the learning rate of weight update; denotes the decay amount predicted by the model for the ith segment before update, i.e. the value calculated based on the old weight and the old parameter; denotes the actual decay amount of SOH caused by the ith operation segment; denotes the partial derivative of the decay sensitive function with respect to the mth enhanced feature component.

[0035] Preferably, the implementation process of SOH prediction through the multi-scale time window in step S4 is specifically as follows:

[0036] Build three time scale windows of short-term, medium-term and long-term;

[0037] Calculate the decay rate in each window respectively:

[0038] ;

[0039] Combine the decay rates of the three scales by weighting to predict the future SOH value:

[0040] ;

[0041] wherein, r x (t i ) denotes the decay rate of the window, including the short-term decay rate r s (t i ), the medium-term decay rate r m (t i ) and the long-term decay rate r l (t i ); W x denotes the length of the time window, including the short-term window W s , the medium-term window W m and the long-term window W l ; denotes the decay amount of the jth segment; denotes the predicted future SOH; α s , α m and α l respectively denote the short-term, medium-term and long-term time window weights; denotes the prediction time step.

[0042] Preferably, the double-layer adaptive correction in step S4 specifically includes:

[0043] Parameter-level correction: when the prediction deviation exceeds the threshold, fine-tune the decay sensitivity coefficient λ i ;​

[0044] Structural level correction: when the prediction deviation exceeds the threshold value for consecutive times, the weight of the multi-scale time window is adjusted s , alpha m , alpha l .

[0045] The technical scheme of the application: a power battery SOH dynamic attenuation behavior modeling system is used to execute the above-mentioned power battery SOH dynamic attenuation behavior modeling method, comprising:

[0046] The time sequence analysis and operation scenario segmentation module is used to collect battery operation data and perform time sequence analysis and operation scenario segmentation;

[0047] The behavior feature embedding and attenuation trigger mode recognition module is used to embed the segmented data into a high-dimensional feature space and recognize the attenuation trigger mode;

[0048] The attenuation quantification and dynamic weight update module is used to calculate the segment-level attenuation amount and dynamically update the model weight;

[0049] The attenuation prediction and adaptive correction closed loop module is used to perform multi-scale SOH prediction, confidence interval evaluation and trigger model correction.

[0050] Compared with the prior art, the above technical scheme of the application has the following beneficial technical effects:

[0051] The application designs a power battery SOH dynamic attenuation behavior modeling method and system, which significantly improves the modeling accuracy under complex dynamic working conditions through adaptive segmentation of operation scenarios and nonlinear feature embedding, effectively captures the time sequence interaction effects and cumulative effects of multi-dimensional stress factors such as current, temperature, SOC and rate, etc.; threshold-mode joint recognition and multi-scale attenuation rate prediction are used to enhance the robustness and foresight of the model, combined with confidence interval evaluation and parameter / structure level double closed loop correction mechanism, to realize continuous adaptive optimization of the prediction result; based on segment-level attenuation quantification and dynamic weight update, the battery management system is provided with real-time SOH indicators and graded early warning, supporting accurate life management decision; through behavior vector compression expression and selective embedding of high-dimensional features, the accuracy is guaranteed while the algorithm power consumption is reduced, improving the engineering applicability of embedded platforms; the application completely covers the whole process of data collection, feature extraction, attenuation recognition, quantitative prediction and closed loop control, significantly improving the modeling accuracy and adaptability under dynamic working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A method flowchart of a power battery SOH dynamic attenuation behavior modeling method proposed by the application;

[0053] Figure 2This is a system architecture diagram of a power battery SOH dynamic degradation behavior modeling system proposed in this invention. Detailed Implementation

[0054] Example 1, as Figure 1 As shown, the present invention proposes a method for modeling the dynamic degradation behavior of a power battery at state of equilibrium (SOH), which includes the following specific implementation steps:

[0055] S1. By collecting battery operating data, a multi-dimensional stress factor time series matrix is ​​constructed, and dynamic features are extracted in segments based on the operating scenario to provide structured input for subsequent degradation behavior modeling, thus characterizing the battery stress state under complex operating conditions. The specific implementation process is as follows:

[0056] S11. During battery operation, the original state parameters of the power battery are collected in real time, including but not limited to current I(t), temperature T(t), state of charge SOC(t), and rate C(t).

[0057] To preserve the dynamic fluctuation characteristics of these factors over time, a multidimensional stress factor time series matrix is ​​constructed: M(t)=[I(t),T(t),SOC(t),C(t)];

[0058] Where I(t) represents the instantaneous current of the battery at time t; T(t) represents the temperature of the cell or module at time t; SOC(t) represents the state of charge of the battery at time t; C(t) represents the rate, which is the ratio of the instantaneous current to the rated capacity of the battery; and M(t) represents the multidimensional stress factor time series matrix.

[0059] S12. Normalize each factor in the multidimensional stress factor time series matrix M(t) and introduce a relative contribution mapping operator:

[0060] ;

[0061] Among them, I max T max and C max These represent the rated upper limits of current, temperature, and rate of increase, respectively, used for normalization; λ I , λ T , λ SOC and λ C The stress factor weight represents the relative contribution of each stress factor to the attenuation. Its initial value is obtained through accelerated aging experiments and can be dynamically adjusted in subsequent steps. This represents the normalized and weighted stress factor matrix;

[0062] S13, construct an operation scenario segmentation method based on time sequence mutation points, based on the adaptive division of stress factor mutation points, which can automatically identify potential stress mode turning points, specifically:

[0063] Define the segmentation criterion function: ;

[0064] When δ(t) exceeds the set threshold θ, it is determined as the switching point of the operation scenario, thereby dividing the time sequence data into several segments: ;

[0065] Where δ(t) represents the stress factor change amplitude at time t; represents the sampling interval of consecutive time points; S represents the segment set; s i represents the time segment of the i-th operation scenario, containing the start time and the end time ; n is the number of segments divided;

[0066] It should be noted that each segment s i corresponds to a relatively stable operation scenario, such as "constant current discharge in medium SOC interval" or "fast charging under high temperature conditions";

[0067] S14, after completing the scenario segmentation, convert each scenario segment into a behavior vector for compactly expressing the stress characteristics within the segment:

[0068] ;

[0069] Where B i represents the behavior vector of the i-th operation scenario; |s i | represents the time length of segment s i .

[0070] S2, convert the operation behavior vector into enhanced features through nonlinear feature embedding and historical segment weighting, determine the decay trigger segment through threshold-mode joint recognition, and perform risk classification labeling to provide interpretable input for SOH (State of Health) decay quantification, the specific implementation process is as follows:

[0071] S21, in order to capture more complex nonlinear interaction effects within the segment, map the behavior vector B i to a high-dimensional nonlinear feature space: ;

[0072] Where F i represents the high-dimensional embedding feature vector of the i-th operation scenario segment; represents the element-wise multiplication operation of the vector; sin(B i) represents the element-wise sine mapping of the behavior vector; α represents the element-wise exponential mapping of the behavior vector; α represents the exponential mapping adjustment coefficient.

[0073] S22. The degradation behavior of a battery exhibits a significant cumulative effect, meaning that past high-stress behavior influences the current state. To reflect this temporal continuity, this embodiment proposes embedding the feature sequence {F1, F2, ..., F...} of continuous segments... n Perform weighted fusion of historical information:

[0074] ;

[0075] in, represents the enhanced feature vector of the i-th segment, which incorporates information from historical segments; β represents the historical information fusion coefficient; k represents the historical window length, i.e., the number of segments considered forward, which is set based on battery usage characteristics or experimental experience; w i,j This indicates the attenuation weight of historical segments, reflecting the magnitude of the influence of past segments on the current segment;

[0076] S23, In obtaining enhanced features Then, a threshold-pattern joint identification mechanism is used to determine whether each operation segment is likely to trigger a decay event:

[0077] ;

[0078] ;

[0079] Among them, Trigger i Indicates whether the i-th segment triggers the decay event (1 = triggered, 0 = not triggered); This represents the attenuation sensitivity function, which integrates the contributions of various features to attenuation; p represents the m-th component of the enhanced feature vector; m γ represents the nonlinear exponent, used to amplify the effects of high stress states; m This represents the component weight, which characterizes the relative contribution of the feature to the decay triggering. This represents the degradation trigger threshold, defining the battery degradation sensitivity level; M represents the total number of dimensions of the enhanced feature vector, including the original behavior vector dimension and all non-linear extended features;

[0080] S24. After identification, all fragment trigger states are formed into a decay trigger mode sequence Trigger_Seq:

[0081] Trigger_Seq=[Trigger1,…,Trigger i Trigger n ];

[0082] Further, each trigger segment is subdivided into annotations:

[0083] Minor trigger: Accumulated stress is small, short-term impact is limited;

[0084] Moderate trigger: May cause measurable degradation, need attention;

[0085] Strong trigger: High-risk segment, may cause significant SOH decline immediately or in the short term.

[0086] S3, quantifying the degradation trigger segments identified in step S2, by calculating the degradation amount of each trigger segment, accumulating the nonlinear effect, dynamically updating the feature weight, and mapping the results to real-time SOH indicators and warning levels, the specific implementation process is as follows:

[0087] S31, in step S2, each operation segment i has been determined whether to trigger degradation event Trigger i , calculate the segment degradation contribution, convert these trigger events into quantifiable SOH loss:

[0088] ;

[0089] Where, represents the SOH degradation amount of the i-th operation segment, representing the immediate contribution of the segment to the overall battery health status; λ i represents the degradation sensitivity coefficient of the i-th segment, used to adjust the actual contribution of the same mode under different environments / conditions;

[0090] S32, the segment degradation amount is accumulated to form the overall degradation curve:

[0091] ;

[0092] ;

[0093] Where, SOH cum (t i ) represents the cumulative SOH loss up to segment i; SOH cum (t i -1) represents the total amount of SOH degradation accumulated by the power battery before the i-1 time, i.e. represents the health status value lost by the battery before reaching the i-th operation segment; represents the historical influence (amplification) factor of the i-th segment, used to simulate the latent / cumulative acceleration effect; μ represents the historical sensitivity coefficient, which determines the sensitivity of the amplification effect of the historical degradation amount on the current segment; L represents the length of the historical window, i.e. the number of recent segments considered when calculating the historical influence; w hist(i,j) represents a historical segment weight function, i.e. the degree of influence of the decay of the jth segment on the ith segment;

[0094] S33, the characteristic weight γ for attenuating the sensitive function in step S2 m Dynamic update is performed:

[0095] ;

[0096] wherein, represents the weight (relative importance) of the mth enhanced feature component in the decay sensitive function ; η represents the learning rate (step size coefficient) of weight update; represents the partial derivative (sensitivity gradient) of the decay sensitive function to the enhanced feature component m; represents the decay amount predicted by the model for the ith segment before updating, i.e. the value calculated based on the old weight and old parameters;

[0097] S34, after the cumulative decay amount and dynamic weight update are completed, the results are mapped into SOH indicators and warning levels that can be used for BMS (battery management system) management:

[0098] ;

[0099] ;

[0100] wherein, SOH0 represents the initial health benchmark (SOH at the time of launch / online); θ1 and θ2 represent the warning threshold, used to distinguish normal / warning / danger levels; represents the warning level (discrete value), which is output based on the mapping of SOH and threshold.

[0101] S4, through multi-scale prediction, confidence interval evaluation and double-layer adaptive correction, SOH decay trend forward prediction and closed-loop strategy triggering are performed to ensure prediction accuracy and application reliability, and the specific implementation process is as follows:

[0102] S41, the prediction method of short-term, medium-term and long-term time windows is used at the same time to integrate decay rate information of different scales to generate more accurate and stable future SOH prediction, specifically:

[0103] For the ith time segment, three time scale windows are constructed:

[0104] Time scale window one (i.e. short-term window W s ): attention to local rapid changes, such as instantaneous over-discharge, rapid charging segments, and the length is set to 10-50 segments in this embodiment;

[0105] Time scale window two (i.e. medium-term window W m) : focus on cumulative impact in daily operation mode, this embodiment is set to length 50-200 segments;

[0106] Time scale window three (i.e. long-term window W l ) : focus on historical long-term trends or seasonal effects, this embodiment is set to length 200-1000 segments;

[0107] Calculate the decay rate of each window:

[0108] ;

[0109] Combine the decay rates of the three scales to form a prediction:

[0110] ;

[0111] Where, r x (t i ) represents the decay rate of the window, including the short-term decay rate r s (t i ), the medium-term decay rate r m (t i ) and the long-term decay rate r l (t i ); W x represents the length of the time window, including the short-term window W s , the medium-term window W m and the long-term window W l ; represents the decay amount of the jth segment; represents the predicted future SOH; α s , α m and α l represent the short, medium and long-term time window weights, respectively; represents the prediction time step;

[0112] S42, by statistically analyzing the prediction residual sequence, a confidence interval of the future SOH is constructed, and the credibility of the prediction result is quantified, specifically:

[0113] Calculate the historical prediction residual:

[0114] ;

[0115] Calculate the residual standard deviation σ res and the mean μ res :

[0116] ; ;

[0117] Build confidence interval:

[0118] ;

[0119] wherein e j represents the prediction residual; κ represents the confidence interval control coefficient, which is set according to the safety policy; N represents the sample number of the statistical residual; represents the actual observed attenuation of the jth segment; represents the confidence interval of the predicted SOH;

[0120] S43, when the actual observation deviates from the prediction by more than the threshold value ε, dynamically correct the model parameters, specifically:

[0121] The triggering condition is:

[0122] ;

[0123] Parameter-level correction: fine-tune the sensitivity factor in step S3:

[0124] ;

[0125] Structure-level correction: if the deviation exceeds the threshold value for multiple consecutive times, adjust the time window weight:

[0126] ;

[0127] wherein represents the corrected sensitivity factor; represents the sensitivity factor before correction; represents the sensitivity correction coefficient; represents the corrected time window weight; represents the time window weight before correction; represents the window weight adjustment factor; represents the sign function, which preserves the deviation direction and ensures that the correction is in the correct direction;

[0128] It should be noted that the cumulative error is the sum of consecutive deviations;

[0129] S44, form a closed loop with the prediction result, the confidence interval, and the corrected model parameters, so that prediction, correction, and policy execution form a dynamic iteration, specifically:

[0130] Input the updated SOH prediction and confidence interval into the BMS (Battery Management System);

[0131] Trigger the policy according to the threshold values θ1 and θ2, for example: limit the peak power output, adjust the charge and discharge rate, issue a maintenance or replacement prompt;

[0132] Continue to record new attenuation data and enter the next round of prediction and correction cycle to achieve adaptive closed-loop control.

[0133] In the embodiment two, as shown, the application proposes a power battery SOH dynamic attenuation behavior modeling system, which is used to execute the power battery SOH dynamic attenuation behavior modeling method proposed in the embodiment one, and includes a time sequence analysis and operation scenario segmentation module, a behavior feature embedding and attenuation trigger mode recognition module, an attenuation quantification and dynamic weight updating module, and an attenuation prediction and adaptive correction closed loop module. Figure 2

[0134] The time sequence analysis and operation scenario segmentation module is used to collect battery operation data (including but not limited to voltage, current, temperature, SOC, and environmental parameters) in real time, perform time sequence analysis and operation scenario segmentation on the original data, and convert the complex operation process into a standardized multi-dimensional stress factor sequence, thereby providing basic features for subsequent attenuation mode recognition.

[0135] The behavior feature embedding and attenuation trigger mode recognition module performs high-dimensional feature embedding and mode extraction on the obtained multi-dimensional stress factor sequence, recognizes the attenuation trigger mode of the battery, including rapid attenuation, chronic attenuation, and burst anomaly mode, and simultaneously outputs feature weights and mode labels, thereby providing inputs for attenuation quantification.

[0136] The attenuation quantification and dynamic weight updating module calculates the segment-level attenuation amount in combination with the output attenuation mode and historical attenuation data, dynamically updates the feature weights and multi-scale window parameters, forms the current battery SOH state and multi-scale attenuation rate, and provides a quantitative basis for subsequent prediction and adaptive correction.

[0137] The attenuation prediction and adaptive correction closed loop module constructs a multi-scale time window prediction model, predicts the future SOH (State of Health), quantifies the uncertainty in combination with residual statistics and confidence interval evaluation, triggers parameter-level and structure-level adaptive correction when the deviation exceeds the threshold, realizes online closed loop optimization, and simultaneously uses the prediction results and the corrected model parameters to trigger strategies such as power adjustment, charge and discharge limitation, or maintenance prompt, thereby realizing battery life management and operation optimization.

[0138] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.​

Claims

1. A method of modeling the dynamic attenuation behavior of the SOH of a power battery, characterized in that, The specific implementation steps include the following: S1, real-time acquisition of battery operation data to construct a multi-dimensional stress factor time sequence matrix, adaptive operation scenario segmentation based on stress mutation points, and conversion of each scenario segment into a behavior vector; S2, embedding the behavior vector in a nonlinear feature space, combining historical segments for weighted fusion to form an enhanced feature vector; adopting a threshold-pattern joint recognition mechanism to determine the decay trigger segment and perform risk grading; The fusion process of combining historical segments for weighted fusion to form an enhanced feature vector is specifically as follows: The embedded feature vector of the current segment is weighted and fused with the embedded feature vectors of the previous k historical segments to obtain an enhanced feature vector that fuses historical information: ； wherein, represents the enhanced feature vector of the i-th segment; β represents the historical information fusion coefficient; k represents the length of the historical window; w i,j represents the historical segment decay weight, reflecting the influence size of the past j-th segment on the current i-th segment; The threshold-pattern joint recognition mechanism is specifically as follows: The enhanced feature vector is input into a decay sensitivity function to calculate its comprehensive output value: ； wherein, denotes the attenuation sensitivity function; denotes the mth component of the enhanced feature vector; p m denotes the non-linear exponent; γ m denotes the component weight; M denotes the total number of dimensions of the enhanced feature vector; The output value is compared with a preset decay trigger threshold value, and if it exceeds the threshold value, it is determined that the segment triggers a decay event; According to the size range of the output value, the trigger segment is risk graded, including slight trigger, moderate trigger, or strong trigger; S3, quantifying the segment-level decay contribution, dynamically updating the feature weight for decay recognition, and mapping it to a real-time SOH indicator and warning level; S4, integrating short-term, medium-term, and long-term decay rates through a multi-scale time window to achieve SOH prediction, combining confidence interval evaluation and parameter and structure-level double-layer adaptive correction to form a closed-loop optimization.

2. The method of claim 1, wherein the method further comprises: The generation process of the behavior vector in step S1 includes: Real-time acquisition of the current, temperature, state of charge, and rate of the power battery to form a time sequence matrix; Normalizing each factor and introducing a relative contribution mapping operator to obtain a normalized and weighted stress factor matrix; Adaptively dividing the time sequence data into several operation scenario segments by calculating the stress factor change amplitude and setting a threshold value; For each scenario segment, calculate its integral mean over the time length to obtain the behavior vector of the segment.

3. The method of claim 2, wherein the method further comprises: The embedding process of embedding the behavior vector in a nonlinear feature space in step S2 is specifically as follows: Map the behavior vector to a high-dimensional feature space through a nonlinear function to form a high-dimensional embedded feature vector; the nonlinear function includes element-wise sine mapping, element-wise exponential mapping, and their combinations, i.e.: ； where F i represents the high-dimensional embedding feature vector of the i-th operational scenario segment; represents the vector element-wise multiplication operation; sin(B i ) represents the element-wise sine mapping of the behavior vector; represents the element-wise exponential mapping of the behavior vector; a represents the exponential mapping adjustment coefficient; B i represents the behavior vector of the i-th operational scenario.

4. The method of claim 3, wherein the method further comprises: The quantification of the segment-level decay contribution in step S3 is specifically as follows: For the segment determined as triggering the attenuation event, the SOH attenuation amount of the segment is calculated according to the enhanced feature vector of the segment and the attenuation sensitivity coefficient λ i . Combine the decay amount of the current segment with the historical decay cumulative amount, introduce a historical influence factor to simulate the cumulative acceleration effect, and calculate the cumulative SOH decay amount up to the current segment.

5. The method of claim 4, wherein the method further comprises: The dynamic updating of the feature weight for the attenuation identification in step S3 is specifically: based on the deviation between the model predicted segment attenuation and the actual observation value, the feature weight in the attenuation sensitive function is dynamically updated by the gradient descent method: ; wherein, denotes the weight of the updated mth enhanced feature component in the decay sensitivity function ; γ m denotes the weight of the mth enhanced feature component before updating in the decay sensitivity function; η denotes the learning rate of weight updating; denotes the decay amount predicted by the model for the ith segment before updating, i.e., the value calculated based on the old weight and the old parameter; denotes the actual decay amount of SOH caused by the ith operation segment; denotes the partial derivative of the decay sensitivity function with respect to the enhanced feature component m.

6. The method of claim 5, wherein the method further comprises: The implementation process of achieving SOH prediction through a multi-scale time window in step S4 is specifically as follows: Build short-term, medium-term, and long-term time scale windows; Calculate the decay rate within each window separately: ; Weight and combine the decay rates of the three scales to predict the future SOH value: ； where r x (t i ) represents the attenuation rate of the window, including short-term attenuation rate r s (t i ), medium-term attenuation rate r m (t i ), and long-term attenuation rate r l (t i ); W x represents the length of the time window, including short-term window W s , medium-term window W m , and long-term window W l ; represents the attenuation amount of the jth segment; represents the predicted future SOH; α s , α m , and α l represent the short, medium, and long-term time window weights, respectively; represents the prediction time step.

7. The method of claim 6, wherein the method further comprises: The double-layer adaptive correction in step S4 specifically includes: Parameter level correction: fine-tune the decay sensitivity coefficient λ when the prediction bias exceeds a threshold i ; Structure level correction: adjust the weight of multi-scale time window, a, when consecutive multiple prediction bias exceeds threshold s , a m , a l .

8. A system for modeling the dynamic attenuation behavior of the SOH of a power battery, for performing a method for modeling the dynamic attenuation behavior of the SOH of a power battery according to any one of claims 1 to 7, characterized in that, It includes: A time sequence analysis and operation scenario segmentation module for acquiring battery operation data and performing time sequence analysis and operation scenario segmentation; A behavior feature embedding and decay trigger pattern recognition module for embedding the segmented data in a high-dimensional feature space and recognizing the decay trigger pattern; The attenuation quantization and dynamic weight updating module is configured to calculate a segment-level attenuation and dynamically update model weights. The attenuation prediction and adaptive correction closed-loop module is configured to perform multi-scale SOH prediction, confidence interval evaluation, and trigger model correction.

Images