Lithium battery residual life prediction method based on adaptive modal number estimation and multi-scale decomposition

Abstract

The invention belongs to the technical field of lithium battery health management, and particularly relates to a lithium battery residual life prediction method based on self-adaptive modal number estimation and multi-scale decomposition, which comprises the following steps of: firstly, acquiring a capacity attenuation sequence of a battery, and then calculating the residual life of the battery through a variational modal decomposition method combining a self-adaptive modal number determination mechanism and Bayesian modal screening. Carrying out adaptive decomposition and denoising on the sequence, and separating a low-frequency trend mode and a medium-high frequency oscillation mode; according to the long-range dependence of the trend mode, a Transform model is adopted for prediction; and for the nonlinear fluctuation characteristics of the oscillation mode, a random forest model is adopted for prediction. And finally, fusing the prediction results of the two groups of components, reconstructing a complete future capacity attenuation curve, and calculating the remaining service life based on a preset failure threshold. According to the method, through adaptive decomposition and grouping modeling, the adaptability, noise immunity and prediction precision of the prediction model to different battery individuals and complex working conditions are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery health management technology, specifically involving a method for predicting the remaining life of lithium batteries based on adaptive mode number estimation and multi-scale decomposition. Background Technology

[0002] Currently, methods for predicting the remaining useful life (RUL) of lithium batteries can be mainly divided into three categories: model-based, data-driven, and fusion-based. Model-based methods (such as electrochemical models, equivalent circuit models, or empirical models) can better reveal the degradation mechanism of batteries and have a certain degree of physical interpretation. However, their modeling process is complex, highly dependent on experimental conditions, and has poor adaptability when faced with non-stationary operating data. In contrast, data-driven methods (such as support vector regression (SVR), Gaussian process regression (GPR), and long short-term memory networks (LSTM)) can directly use historical capacity or voltage sequences for lifetime prediction. They have better versatility and automatic learning capabilities, but are often susceptible to capacity regeneration, environmental disturbances, and noise interference, resulting in insufficient stability of the prediction results.

[0003] To balance model interpretability and predictive performance, fusion methods have emerged in recent years, combining signal decomposition techniques with machine learning algorithms (such as PF-LSTM and EEMD-LGBM). These methods extract features at different time scales by decomposing the capacity sequence before performing a comprehensive modeling. However, such methods still have several shortcomings in practical applications: First, most signal decomposition methods (such as EMD, EEMD, and VMD with fixed mode number) require manual setting of the number of modes, which makes it difficult to adapt to changes in different battery models and operating conditions. Second, mode aliasing is prone to occur during the decomposition process, which leads to a decrease in the physical interpretability of the components; Third, the computational complexity is high, making it difficult to meet real-time requirements while ensuring prediction accuracy; Fourth, the handling of capacity regeneration phenomena and high-frequency disturbance signals is still insufficient.

[0004] Overall, existing technologies still have significant limitations in dealing with the nonlinearity, non-stationarity, and multi-timescale characteristics of battery capacity sequences, making it difficult to meet the requirements for high accuracy and robustness in battery life prediction under complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the remaining life of lithium batteries based on adaptive mode number estimation and multi-scale decomposition, so as to solve the problems mentioned in the background art.

[0006] The present invention achieves the above objectives through the following technical solutions: This invention proposes a method for predicting the remaining life of lithium batteries based on adaptive mode number estimation and multi-scale decomposition. The method includes: Acquire capacity degradation sequence data of lithium batteries during cyclic charge and discharge processes, and assign a cycle number to each sequence data. Generate capacity-circular sequence Wherein, the capacity degradation sequence data is the discharge capacity. The sequence of values ​​changing over time, where n is the total number of cycles accumulated from the start of battery use to the current moment; For the capacity-cycle sequence Adaptive variational mode decomposition is performed to obtain modal components in different center frequency ranges; Based on the center frequency and cumulative energy percentage of the modal components, the modal components are divided into low-frequency trend modes and mid-to-high frequency oscillation modes; A time-series prediction model based on a self-attention mechanism is used to predict the low-frequency trend mode, and the prediction results of the trend components are obtained. An integrated regression model was used to predict the mid-to-high frequency oscillation mode, and the prediction results of the oscillation components were obtained. The final battery remaining life prediction result is determined based on the trend component prediction result and the oscillation component prediction result.

[0007] Furthermore, the capacity-cycle sequence Perform adaptive variational mode decomposition, including: Given a capacity-circular sequence Within the variational mode decomposition framework, based on the joint optimization criterion, a pre-defined set of candidate modes is selected. Determining the optimal decomposition mode number ; With the optimal decomposition mode number For the capacity-cycle sequence, K is the total number of modes. Perform variational mode decomposition to obtain K modal components. and the corresponding center frequency The variational objective function of the variational mode decomposition is as follows: ; in, Let K be the center frequencies of the modal components; It is the Dirac impulse function; This is the convolution operator; The imaginary unit; For time differential operators; It is the L2 norm, used to measure frequency domain bandwidth.

[0008] Furthermore, the determination of the optimal decomposition mode number ,include: For each mode number K in the candidate mode number set, perform variational mode decomposition and calculate the corresponding mode reconstruction error E. rec (K), Modal Spectral Entropy H spec (K) and the Bayesian information criterion BIC(K); Select the optimal number of decomposition modes based on the joint optimization criterion. The joint optimization criterion is as follows: ; in, The reconstruction error is the minimum number of candidate modes. These are the weight parameters.

[0009] Furthermore, the K modal components are obtained. Subsequently, the method further includes: Determine each modal component eigenvectors , ,in, Indicates frequency stability. Indicates the percentage of energy. Represents time-domain smoothness; eigenvectors are calculated using a pre-defined evaluation model. The posterior probability of the signal mode ; Modal components with a posterior probability higher than a set threshold are identified as modal components and retained, while modal components with a posterior probability lower than the set threshold are discarded as noise modes.

[0010] Furthermore, the preset evaluation model is a Bayesian inference model based on a Gaussian mixture model, and the posterior probability... The calculation formula is: ; in, These are the prior probabilities of the signal and noise categories, respectively; The mean and covariance matrix of the Gaussian mixture model of signal and noise modes; It is the Gaussian distribution density function.

[0011] Furthermore, the step of dividing the modal components into low-frequency trend modes and mid-to-high-frequency oscillation modes based on the center frequency and cumulative energy percentage of the modal components includes: The retained modal components after filtering are then sorted according to their center frequencies. Sort by low to high, select the lowest frequency modes that have reached a preset proportion of cumulative energy as low frequency trend modes, and the rest as medium and high frequency oscillation modes.

[0012] Furthermore, the time series prediction model based on the self-attention mechanism is a Transformer model, which models the global dependencies of the input low-frequency trend mode sequence through the self-attention mechanism. The computational representation of the self-attention mechanism is as follows: ; in, These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; This is for self-attention output.

[0013] Furthermore, the ensemble regression model is a random forest regression model, which provides prediction results for mid-to-high frequency oscillation modes. The average of the predictions from M decision trees is expressed as: ; in, For the m-th tree pair of input features The predicted value.

[0014] Furthermore, determining the final predicted battery remaining life includes: The predicted value of the trend component at the corresponding time point With the predicted value of the oscillation component Add them together to obtain the predicted battery capacity at time t. ; Based on the predicted capacity The time series data forms a battery capacity prediction curve; Preset capacity threshold for battery life end ; In the battery capacity prediction curve, the first drop to the capacity threshold is determined. The corresponding prediction loop number ; Based on the current loop count Calculate the remaining useful life .

[0015] The beneficial effects of this invention are as follows: 1. This invention introduces a joint optimization criterion into the variational mode decomposition (VMD) framework to achieve adaptive determination of the optimal number of modes, and combines a Bayesian mode credibility screening mechanism to achieve effective separation of signal and noise, thereby overcoming the shortcomings of traditional decomposition methods where the number of modes depends on manual setting and mode aliasing is severe.

[0016] 2. This invention uses the Transformer-TS model and the Random Forest (RF) model to model and predict the trend component and oscillation component obtained by decomposition, respectively. It makes full use of the time-dependent modeling capability of deep learning and the nonlinear fitting advantage of machine learning, and finally merges and outputs a complete and more accurate capacity prediction curve.

[0017] 3. This invention can achieve adaptive decomposition with dynamic adjustment of modality number according to data features, effectively avoiding deviations caused by human parameter setting; at the multi-scale modeling level, it can simultaneously capture long-term degradation trends and short-term oscillation characteristics; in terms of noise resistance and interpretability, the Bayesian screening mechanism significantly improves the credibility of signal decomposition; in terms of prediction performance, the model has high accuracy and strong generalization ability, and can be adapted to various battery types and complex working environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of a lithium-ion battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition in this invention. Figure 2 This is a flowchart of the capacity sequence grouping process of AVMD-BMS in this invention; Figure 3 This is another flowchart of the lithium-ion battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition in this invention. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0020] In predictive health management and maintenance decision-making for lithium batteries, accurately predicting their remaining useful life (RUL) is a core element for achieving condition-based maintenance, optimizing operating strategies, and improving system reliability and economy. Currently, most data-driven battery life prediction methods rely on single-dimensional time series such as capacity or voltage, employing fixed-parameter signal decomposition techniques and machine learning models for modeling. However, these methods have significant limitations: the actual degradation process of batteries is highly nonlinear, non-stationary, and exhibits individual variability. Their capacity sequences simultaneously contain multi-scale mixed characteristics, including long-term degradation trends, short-term fluctuations, operating condition disturbances, and measurement noise. Traditional methods often require manual pre-setting of the decomposition mode number, making it difficult to adapt to different battery types and operating conditions. Furthermore, their ability to handle capacity regeneration phenomena and high-frequency noise is limited, resulting in insufficient accuracy, robustness, and generalization ability of the prediction models.

[0021] To address the aforementioned problems, a preferred embodiment of the present invention provides a method for predicting the remaining life of lithium-ion batteries based on adaptive mode number estimation and multi-scale decomposition. Please refer to [link to relevant documentation]. Figures 1-3 The method described herein can be applied to battery management systems (BMS), cloud-based battery health monitoring platforms, and corresponding computer program products. The steps of the method will be explained in detail below with reference to the accompanying drawings.

[0022] S1: Data Acquisition and Capacity Sequence Construction First, key performance data during the cyclic charge-discharge process of the target lithium-ion battery are obtained from its BMS or historical test database. The core data is the discharge capacity value recorded or calculated at the end of each charge-discharge cycle. Assuming n charge-discharge cycles have been accumulated, a capacity-time series can be constructed in cyclic order. Among them, the capacity degradation sequence data represents the discharge capacity. A sequence of values ​​changing over time; Each sequence of data is assigned a cycle number; n is the total number of cycles accumulated from the start of battery use to the current moment, i.e., the sequence length.

[0023] To ensure data quality, the original sequence needs to be preprocessed, including but not limited to: smoothing random measurement noise using methods such as moving average and median filtering; using interpolation to handle outliers caused by data loss; and normalizing the sequence to eliminate the influence of dimensions, scaling the data to the [0,1] interval or converting it to a form with zero mean and unit variance to improve the numerical stability of subsequent algorithms.

[0024] S2: Adaptive Multiscale Decomposition Based on AVMD-BMS (Adaptive Variational Mode Decomposition with Bayesian Mode Selection) This step will mix the original capacity-time series data. It decomposes into a series of sub-components (modes) with clear physical meaning, and achieves signal-noise separation to distinguish physical components at different time scales, such as long-term decay, short-term fluctuations, and random disturbances; specifically for capacity-cyclic sequences. Adaptive variational mode decomposition is performed to obtain modal components in different center frequency ranges.

[0025] In a preferred embodiment, for capacity-cycle sequences Perform adaptive variational mode decomposition, including: Given a capacity-circular sequence Within the variational mode decomposition framework, based on the joint optimization criterion, a pre-defined set of candidate modes is selected. Determining the optimal decomposition mode number .

[0026] Specifically, the range of the candidate set It can be set based on experience to ensure that it can cover multiple scale components from trend to noise, without causing over-decomposition or excessive computation due to too many modes.

[0027] In predicting the remaining useful life (RUL) of lithium-ion batteries, the capacity degradation sequence typically contains multiple time-scale components, including long-term degradation trends, short-term fluctuations, random disturbances, and measurement noise. When using variational mode decomposition (VMD) to decompose the capacity degradation sequence at multiple scales, the selection of the number of modes K directly affects the separation effect between trend modes and oscillating modes, thereby affecting the accuracy and stability of the subsequent lifetime prediction model.

[0028] If the number of modes K is too small, under-decomposition can easily occur, causing trend, oscillation, and noise components to become mixed. If the number of modes K is too large, over-decomposition can easily occur, resulting in redundant modes or even spurious oscillations, which not only increases computational cost but may also reduce the robustness of the prediction model. Therefore, it is necessary to adaptively determine the optimal number of decomposed modes under the constraints of multiple evaluation metrics. .

[0029] In one preferred approach, the optimal number of decomposition modes is determined. ,include: For each mode number K in the candidate mode number set, perform variational mode decomposition and calculate the corresponding mode reconstruction error E. rec (K), Modal Spectral Entropy H spec (K) and Bayesian Information Criterion BIC(K).

[0030] The modal reconstruction error is as follows: ; in: The original capacity sequence; Let K be the k-th modal component; K is the total number of modes. The smaller this error, the higher the fidelity of the decomposition. This is used to measure the fidelity of reconstructing the original capacity sequence from K modes. When K is too small, the capacity degradation trend, oscillatory components, and noise cannot be effectively separated, resulting in a significant increase in reconstruction error and consequently a decrease in RUL prediction accuracy. By introducing a reconstruction error constraint, the optimization process can be driven to select a sufficient number of modes to ensure the complete preservation of trend information and reduce the risk of under-decomposition.

[0031] Modal spectral entropy is as follows: First check each mode Perform power spectrum estimation: ; The normalized power spectrum yields the frequency probability distribution: ; Calculate the spectral entropy of a single modality: ; Finally, the multimodal average spectral entropy is obtained: ; in: This represents the number of power spectrum sampling points. For mode k at frequency The power spectrum, The normalized power spectrum represents the frequency probability distribution. Let be the spectral entropy of the k-th mode.

[0032] Modal spectral entropy This is used to characterize the concentration of modal spectral energy distribution. Low-frequency trend modes typically have a more concentrated spectral distribution and lower spectral entropy; while high-frequency noise modes have a more dispersed spectrum and higher spectral entropy. By introducing the modal spectral entropy index, the ability to identify low-frequency trend modes can be enhanced during the mode number selection process, which helps to suppress the interference of high-frequency disturbances on trend prediction, thereby improving the robustness of the RUL prediction model.

[0033] The Bayesian information criterion is as follows: ; Where: n is the data length (number of data points in the capacity sequence), and k is the number of model parameters (here, K can be taken as the number of modes). Let V be the residual variance based on the reconstruction of K modes.

[0034] Simultaneously considering model fitting error and model complexity, this approach aims to achieve a balance between fitting capability and the number of parameters. When K is too large, VMD tends to generate a large number of modes containing only noise or redundant oscillations, which not only increases computational complexity but may also introduce spurious trends. Since BIC(K) increases with the number of modes, this metric penalizes excessively large K in the objective function, thereby effectively suppressing over-decomposition.

[0035] The modal reconstruction error, modal spectral entropy, and Bayesian information criterion exhibit non-uniform trends as the number of modes changes: as the number of modes K increases, the reconstruction error typically decreases, while the model complexity term BIC(K) increases accordingly, and the modal spectral entropy Hspec(K) shows non-linear characteristics. Therefore, a single evaluation index is insufficient to simultaneously consider decomposition accuracy, trend extraction capability, and model complexity; it is necessary to construct a joint optimization criterion for a comprehensive selection of the number of modes.

[0036] ① Construction of the comprehensive objective function Based on the three complementary evaluation metrics mentioned above, and considering the dimensions of fitting accuracy, trend concentration, and model complexity, the modality number selection problem is transformed into a single-objective optimization problem under multiple metric constraints. By normalizing each metric and employing a linear weighting method, a joint optimization objective function is constructed: ; Among them, by using the minimum number of modes Reconstruction error As a normalization benchmark, it can reduce the impact of differences in capacity scale and noise level of different battery samples on the numerical range of the objective function, thereby improving the stability and comparability of the optimal mode number selection results under different operating conditions. is a weighting parameter used to adjust the relative importance of each indicator in the objective function.

[0037] Used to constrain model complexity and prevent K from becoming too large; for real-time online computing or resource-constrained systems, increasing... The K-value can be more strictly limited to reduce computational overhead. For scenarios requiring extremely high prediction accuracy but with ample computational resources, the K-value can be appropriately reduced. .

[0038] To ensure accurate decomposition and reduce the risk of under-decomposition, and for applications with high RUL accuracy requirements, such as battery state prediction and lifespan management, the RUL can be appropriately increased. This ensures the integrity of trend information. For offline analysis with lower real-time requirements, decomposition accuracy can be prioritized.

[0039] This is used to reward low-frequency trend modes with high spectral concentration, thereby improving trend extraction capabilities. For scenarios requiring strong trend extraction, such as trend recognition and anomaly detection, the value can be appropriately increased. .

[0040] ②Comprehensive adjustment strategy for weight parameters Before constructing the joint optimization criterion, it is necessary to first determine the relative weights of each evaluation index in the objective function. To this end, a comprehensive adjustment strategy for weight parameters based on historical battery capacity cycling data is adopted.

[0041] Specifically, on the historical battery capacity cycling dataset, the number of candidate modes... Perform traversal decomposition and combine different weights ( Under these conditions, the corresponding capacity prediction error index (such as RMSE) and computational complexity index are calculated respectively. Through cross-validation, the weight combination that achieves the optimal trade-off between prediction accuracy and computational cost is selected, and this set of weight parameters is fixed for the subsequent mode number selection process.

[0042] In different application scenarios, the strategy for determining the value of the weight parameter can vary, for example: High-precision lifetime prediction scenarios: increase This ensures that the reconstruction error term dominates the joint objective function, thereby guaranteeing the complete preservation of trend mode information; Online high-speed calculation scenarios: increase Strengthen the constraints on model complexity to avoid the computational burden caused by an excessive number of modes; Anomaly detection or early degradation recognition scenarios: Appropriately increase This highlights the spectral concentration of low-frequency trend modes and improves the identifiability of anomalous changes.

[0043] ③ Solving for the optimal decomposition mode number In weight parameters Once determined, the objective function value J(K) corresponding to each K is calculated within the preset range of candidate modes, and the mode number that minimizes the objective function is selected as the optimal decomposition mode number.

[0044] Select the optimal decomposition mode number according to the joint optimization criterion. The joint optimization criterion is as follows: ; in, The Bayesian information criterion is used to balance model complexity and fitting accuracy. This refers to the modal reconstruction error. The reconstruction error represents the minimum number of candidate modes. Modal spectral entropy measures the concentration of modal spectra. The weighting parameters (e.g., 0.6, 0.3, 0.1) can be further adjusted through cross-validation. This mechanism enables the number of decomposed modes to adapt to different battery models and operating conditions while ensuring reconstruction accuracy.

[0045] In this embodiment, the goal of the joint optimization criterion is to find a K value that ensures the model is neither too complex nor too simple, while maintaining high reconstruction accuracy and spectral concentration. The K value that minimizes the above equation is selected as the optimal number of decomposition modes K*.

[0046] With optimal decomposition mode number For the capacity-cyclic sequence, K is the total number of modes. Perform variational mode decomposition to obtain K modal components. and the corresponding center frequency The variational objective function of the variational mode decomposition is as follows:

[0047] in, Let K be the center frequencies of the modal components; It is the Dirac impulse function; This is the convolution operator; The imaginary unit; For time differential operators; The L2 norm is used to measure the bandwidth in the frequency domain. Solving this optimization problem aims to obtain a set of modes with a defined bandwidth, such that the sum of all modes can accurately reconstruct the original signal.

[0048] The optimal decomposition mode number adaptive selection method proposed in this invention systematically balances the fitting accuracy, trend extraction capability, and model complexity of the decomposition by constructing a multi-index joint optimization objective function, fundamentally overcoming the limitations of traditional methods that rely on empirically set fixed mode numbers. Its adaptability ensures that the algorithm can automatically adjust the optimal decomposition parameters according to the dynamic changes in the chemical characteristics, operating conditions, and degradation modes of different batteries, thereby achieving the optimal solution between avoiding information loss due to under-decomposition and noise interference introduced by over-decomposition. This not only significantly improves the physical interpretability and signal purity of capacity sequence decomposition but also provides a stable and reliable input foundation for subsequent high-precision lifetime prediction models, enhancing the robustness and generalization ability of the entire prediction system in different application scenarios.

[0049] Not all decomposed modes contribute to the prediction of capacity decay trends; some high-frequency modes may primarily contain measurement noise or irregular random perturbations. This invention proposes a mode selection mechanism based on Bayesian inference: After obtaining K modal components Subsequently, the method also includes: Determine each modal component eigenvectors , ,in, Indicates frequency stability. Indicates the percentage of energy. Represents time-domain smoothness; eigenvectors are calculated using a pre-defined evaluation model. The posterior probability of the signal mode ; Modal components with posterior probabilities higher than a set threshold are identified as modal components and retained, while modal components with posterior probabilities lower than a set threshold are discarded as noise modes.

[0050] In a preferred embodiment, the preset evaluation model is a Bayesian inference model based on a Gaussian mixture model, with posterior probability... The calculation formula is: ; in, These are the prior probabilities of the signal and noise categories, respectively; The mean and covariance matrix of the Gaussian mixture model of signal and noise modes; It is the Gaussian distribution density function.

[0051] It should be noted that in traditional variational mode decomposition methods, the number of modes K usually needs to be manually preset based on expert experience. This fixed parameter selection mode has significant deviation risks in practical applications. First, the capacity decay mode of lithium-ion batteries is highly dependent on their chemical system, operating conditions, and individual differences. A single preset K value cannot adapt to diverse decay characteristics, easily leading to over-decomposition or under-decomposition. Over-decomposition artificially splits the intrinsic trend or fluctuation into multiple redundant modes, introducing false information and reducing the efficiency of subsequent modeling; while under-decomposition fails to fully remove noise and aliasing frequencies, making the trend component mixed with disturbances and the oscillation component unclear, directly affecting the prediction accuracy. Second, even if a relatively reasonable K is selected through trial and error for a specific battery model, this method lacks adaptability to the slow changes in degradation dynamics of the same battery throughout its entire life cycle. The adaptive mode number estimation mechanism proposed in this invention fundamentally abandons the dependence on a fixed K value. By dynamically matching the intrinsic characteristics of the current sequence through joint optimization criteria, it ensures that the decomposition result always maintains optimal alignment with the real physical structure of the data, thereby eliminating the systematic bias introduced by parameter fixation.

[0052] S3: Based on the center frequency and cumulative energy ratio of the modal components, the modal components are divided into low-frequency trend modes and mid-to-high frequency oscillation modes.

[0053] In a preferred embodiment, modal components are divided into low-frequency trend modes and mid-to-high-frequency oscillatory modes based on their center frequencies and cumulative energy percentages. This includes: dividing the retained modal components after screening according to their center frequencies. Sort by low to high, select the lowest frequency modes that have reached a preset proportion of cumulative energy as low frequency trend modes, and the rest as medium and high frequency oscillation modes.

[0054] S4: A time-series prediction model based on a self-attention mechanism is used to predict low-frequency trend modes, and the prediction results of trend components are obtained.

[0055] In a preferred embodiment, the time series prediction model based on the self-attention mechanism is a Transformer model. The self-attention mechanism models the global dependencies of the input low-frequency trend mode sequence. The computational representation of the self-attention mechanism is as follows: ; in, These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; This is for self-attention output.

[0056] Using the trend components of the first m cycle points of historical data as input, the trend component value of the (m+1)th cycle point is predicted. The model parameters are trained by minimizing the mean squared error (MSE) between the predicted and actual values. During the prediction phase, a rolling prediction method can be used to progressively predict the trend values ​​at multiple future time steps, thus obtaining the predicted trend component values. .

[0057] S5: An integrated regression model is used to predict the mid-to-high frequency oscillation modes, and the prediction results of the oscillation components are obtained.

[0058] In a preferred embodiment, the ensemble regression model is a random forest regression model, which predicts the mid-to-high frequency oscillation modes. The average of the predictions from M decision trees is expressed as: ; in, For the m-th tree pair of input features The predicted value.

[0059] S6: Determine the final battery remaining life prediction result based on the trend component prediction result and the oscillation component prediction result. This step aims to fuse the separately predicted trend component and oscillation component to reconstruct the complete future capacity decay trajectory, and accurately calculate the battery's remaining life (RUL) accordingly.

[0060] In a preferred embodiment, determining the final predicted battery remaining life includes: The predicted value of the trend component at the corresponding time point With the predicted value of the oscillation component Add them together to obtain the predicted battery capacity at time t. ; .

[0061] Based on capacity forecasts The time series data forms the battery capacity prediction curve; specifically, it refers to the capacity prediction value calculated at each future time point (number of cycles). The data are connected and organized into an ordered sequence according to their chronological order. Mathematically, this sequence constitutes a complete description of the future capacity decay trajectory, which can be represented as a curve that changes over time.

[0062] Preset capacity threshold for battery life end Assume the current battery capacity is Rated capacity is The capacity threshold is (in A typical value is 0.8.

[0063] In the battery capacity prediction curve, determine the first drop to the capacity threshold. The corresponding prediction loop number That is, on the obtained battery capacity prediction curve, a search is performed along the time axis (in the direction of increasing cycle count) to determine when the prediction curve first crosses (or equals) the capacity threshold from above. The number of future cycles corresponding to the time is denoted as . .

[0064] Based on the current loop count Calculate the remaining useful life (in cycles). RUL is measured in cycles, which directly reflects the number of charge-discharge cycles a battery is expected to safely operate under its current degradation trend.

[0065] To address the slow changes in battery aging kinetics, the method of this invention can integrate an online learning mechanism. The system continuously monitors prediction errors (such as the root mean square error (RMSE) between the predicted capacity and the actual capacity over the most recent N cycles). When the monitored error exceeds a preset alarm threshold, the system can automatically trigger a retraining process for part or all of the models, retraining the decomposition parameters, filtering models, and prediction models using historical sequences containing the latest data. This allows the entire prediction system to dynamically adapt to the evolution of battery performance and maintain long-term prediction accuracy.

[0066] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated.

[0067] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0068] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0069] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0070] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the remaining life of lithium batteries based on adaptive mode number estimation and multi-scale decomposition, characterized in that, The method includes: Acquire capacity degradation sequence data of lithium batteries during cyclic charge and discharge processes, and assign a cycle number to each sequence data. Generate capacity-circular sequence Wherein, the capacity degradation sequence data is the discharge capacity. The sequence of values ​​changing over time, where n is the total number of cycles accumulated from the start of battery use to the current moment; For the capacity-cycle sequence Adaptive variational mode decomposition is performed to obtain modal components in different center frequency ranges; Based on the center frequency and cumulative energy percentage of the modal components, the modal components are divided into low-frequency trend modes and mid-to-high frequency oscillation modes; A time-series prediction model based on a self-attention mechanism is used to predict the low-frequency trend mode, and the prediction results of the trend components are obtained. An integrated regression model was used to predict the mid-to-high frequency oscillation mode, and the prediction results of the oscillation components were obtained. The final battery remaining life prediction result is determined based on the trend component prediction result and the oscillation component prediction result.

2. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 1, characterized in that, The capacity-cycle sequence Perform adaptive variational mode decomposition, including: Given a capacity-circular sequence Within the variational mode decomposition framework, based on the joint optimization criterion, a pre-defined set of candidate modes is selected. Determining the optimal decomposition mode number ; With the optimal decomposition mode number For the capacity-cycle sequence, K is the total number of modes. Perform variational mode decomposition to obtain K modal components. and the corresponding center frequency The variational objective function of the variational mode decomposition is as follows: ； in, Let K be the center frequencies of the modal components; It is the Dirac impulse function; This is the convolution operator; The imaginary unit; For time differential operators; It is the L2 norm, used to measure frequency domain bandwidth.

3. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 2, characterized in that, The determination of the optimal decomposition mode number ,include: For each mode number K in the candidate mode number set, perform variational mode decomposition and calculate the corresponding mode reconstruction error E. rec (K), Modal Spectral Entropy H spec (K) and the Bayesian information criterion BIC(K); Select the optimal number of decomposition modes based on the joint optimization criterion. The joint optimization criterion is as follows: ； in, The reconstruction error is the minimum number of candidate modes. These are the weight parameters.

4. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 2, characterized in that, The K modal components are obtained. Subsequently, the method further includes: Determine each modal component eigenvectors , ,in, Indicates frequency stability. Indicates the percentage of energy. Represents time-domain smoothness; eigenvectors are calculated using a pre-defined evaluation model. The posterior probability of the signal mode ; Modal components with a posterior probability higher than a set threshold are identified as modal components and retained, while modal components with a posterior probability lower than the set threshold are discarded as noise modes.

5. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 4, characterized in that, The preset evaluation model is a Bayesian inference model based on a Gaussian mixture model, and the posterior probability... The calculation formula is: ； in, These are the prior probabilities of the signal and noise categories, respectively; The mean and covariance matrix of the Gaussian mixture model of signal and noise modes; It is the Gaussian distribution density function.

6. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 4, characterized in that, The modal components are divided into low-frequency trend modes and mid-to-high-frequency oscillation modes based on their center frequency and cumulative energy percentage, including: The retained modal components after filtering are then sorted according to their center frequencies. Sort by low to high, select the lowest frequency modes that have reached a preset proportion of cumulative energy as low frequency trend modes, and the rest as medium and high frequency oscillation modes.

7. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 1, characterized in that, The time series prediction model based on the self-attention mechanism is a Transformer model. It models the global dependencies of the input low-frequency trend mode sequence through the self-attention mechanism. The computation of the self-attention mechanism is expressed as follows: ； in, These are the query matrix, key matrix, and value matrix, respectively. The dimension of the key vector; This is for self-attention output.

8. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 7, characterized in that, The ensemble regression model is a random forest regression model, and its prediction results for mid-to-high frequency oscillation modes are as follows. The average of the predictions from M decision trees is expressed as: ； in, For the m-th tree pair of input features The predicted value.

9. The lithium battery remaining life prediction method based on adaptive mode number estimation and multi-scale decomposition according to claim 8, characterized in that, The determination of the final predicted battery remaining life includes: The predicted value of the trend component at the corresponding time point With the predicted value of the oscillation component Add them together to obtain the predicted battery capacity at time t. ; Based on the predicted capacity The time series data forms a battery capacity prediction curve; Preset capacity threshold for battery life end ; In the battery capacity prediction curve, the first drop to the capacity threshold is determined. The corresponding prediction loop number ; Based on the current loop count Calculate the remaining useful life .

Images