Method, device and medium for rapid capacity estimation of retired lithium-ion batteries

Abstract

The application discloses a retired lithium ion battery capacity rapid estimation method, equipment and medium, the method comprises the following steps: for different initial state lithium ion battery, through multi-condition aging experiment, the corresponding actual capacity label and multi-modal battery characteristic data are collected; based on the multi-modal battery characteristic data, the real-time charge capacity is determined, and the dynamic capacity increment curve is constructed; based on the dynamic capacity increment curve and the multi-modal battery characteristic data, the dynamic characteristics of multiple scales are extracted, and the effective features are extracted; a collaborative model is constructed, and the corresponding battery health state is output according to the dynamic capacity increment curve and the effective features. Through the combination of ampere-hour integration and impedance correction, and the construction of the dynamic capacity increment curve, the defect that the error of the traditional voltage and current integration method is large is overcome; the self-attention enhanced one-dimensional convolutional neural network effectively captures the time sequence characteristics, and the accuracy of the capacity estimation is improved.

Description

Technical Field

[0001] This application relates to the field of lithium battery capacity estimation, specifically to methods, equipment, and media for rapid capacity estimation of retired lithium-ion batteries. Background Technology

[0002] Lithium-ion batteries, with their advantages of high energy density, long cycle life, and environmental friendliness, have become core energy storage components in new energy vehicles, large-scale energy storage power stations, and electric bicycles. Simultaneously, with the continuous upgrading and iteration of various lithium battery products, the number of retired lithium-ion batteries is increasing. Battery State of Health (SOH), as a key indicator for measuring battery performance degradation, has its core evaluation parameter being the ratio of the battery's actual capacity to its rated capacity. Therefore, accurate capacity estimation is the core task of SOH assessment, directly affecting the operational safety, reliability, and economy of the battery system, and is particularly crucial in the monitoring of electric bicycle service life and the cascade utilization of retired batteries.

[0003] In the electric bicycle industry, lithium-ion batteries operate in complex and variable environments (e.g., outdoor high and low temperatures, frequent start-stop cycles, and large fluctuations in charge and discharge rates), and users have significantly different charging habits (e.g., overcharging and long-term storage with low charge), resulting in inconsistent battery degradation rates and making SOH estimation difficult.

[0004] Existing electric bicycle battery management systems (BMS) mostly use simple voltage and current integration to estimate capacity, which has low accuracy (the error is usually greater than 10%). This makes it difficult to predict safety risks such as sudden drop in range, swelling, and fire caused by battery degradation, affecting user experience and travel safety. At the same time, the number of retired electric bicycle batteries is huge, and their capacity degradation status directly determines whether they have value for secondary use or need to be directly recycled and dismantled. However, traditional capacity testing of retired batteries relies on complete charge and discharge tests, which are inefficient and costly, and cannot meet the needs of rapid classification of large-scale retired batteries.

[0005] In traditional approaches, lithium-ion battery capacity estimation methods are mainly divided into three categories: One approach is the model-based method based on electrochemical mechanisms, which describes the degradation process by establishing a mathematical model of the electrochemical reactions inside the battery. However, this type of method requires precise matching of battery material parameters and reaction kinetics characteristics, resulting in high model complexity. Furthermore, it is difficult to quantify the degradation effect caused by the coupling of multiple mechanisms such as lithium loss and active material failure, which leads to the estimation accuracy being significantly affected by the degree of model simplification.

[0006] The second is the semi-mechanism- and semi-data-driven method based on feature extraction. Among them, the incremental capacity (IC) curve can reflect the electrochemical characteristics of the lithium insertion / extraction process inside the battery and is used for feature mining. However, traditional schemes mostly extract static features such as peak voltage and peak height, which are difficult to capture the dynamic evolution of features with the aging process. Furthermore, it is difficult to achieve synergistic optimization between feature extraction and intelligent algorithms, and it is difficult to balance the speed and accuracy of estimation.

[0007] Thirdly, there are data-driven intelligent algorithms, such as neural networks and support vector machines, which establish a mapping relationship between input and capacity through training with a large amount of data. However, traditional methods mostly rely on macroscopic external features such as voltage, current, and temperature, ignoring the essential relationship between internal electrochemical changes and capacity decay in the battery. Summary of the Invention

[0008] To address the aforementioned issues, this application proposes a rapid method for estimating the capacity of retired lithium-ion batteries, comprising: For lithium-ion batteries in different initial states, corresponding actual capacity tags and multi-mode battery characteristic data are collected through multi-condition aging experiments, and the multi-mode battery characteristic data is enhanced and preprocessed. Based on multi-mode battery characteristic data, the real-time charging capacity is determined by the ampere-hour integration algorithm and impedance correction, and the capacity increment is determined by the adaptive sliding window algorithm to construct a dynamic capacity increment curve. Based on the dynamic capacity increment curve and the multimodal battery feature data, dynamic features at multiple scales are extracted, and feature selection is performed based on nonlinear correlation analysis to extract effective features. A collaborative model is constructed to output the corresponding battery health status based on the dynamic capacity increment curve and the effective features. The collaborative model extracts the temporal features of the dynamic capacity increment curve through a self-attention enhanced one-dimensional convolutional neural network. During the training process of the collaborative model, for the adaptive maximum mean difference domain adaptation module, the maximum mean difference value between the source domain and the target domain is determined by a hybrid kernel constructed by linear kernel and Gaussian kernel, and the weight of the maximum mean difference loss corresponding to the maximum mean difference value is dynamically adjusted.

[0009] In one example, for lithium-ion batteries in different initial states, corresponding actual capacity labels and multi-mode battery characteristic data are collected through multi-condition aging experiments, specifically including: Multiple lithium-ion batteries of different models and in different initial states were selected; the initial state was determined based on the degree of aging. Based on different operating parameters, corresponding multi-condition aging experiments are determined to conduct cycle aging tests on the multiple lithium-ion batteries. For each aging cycle node, the corresponding actual capacity label is obtained through benchmark capacity testing, and the corresponding multi-mode battery characteristic data is collected; the multi-mode battery characteristic data includes voltage data, current data, temperature data, and battery impedance spectrum data.

[0010] In one example, the enhanced preprocessing of the multimodal battery feature data specifically includes: For the aforementioned multimodal battery feature data, outlier removal is performed using an improved Raida criterion; Noise suppression is achieved through wavelet packet transform. Based on the inflection point of the voltage-capacity curve during the multi-condition aging test, the characteristic data of multi-mode batteries under different cycle aging tests and different operating conditions are aligned in time and capacity dimensions through a dynamic time warping algorithm. The voltage data, current data, and temperature data are standardized by interval standardization, and the battery impedance spectrum data is standardized by Z-score standardization. For missing values ​​and local outliers in the multimodal battery feature data, multimodal attention interpolation is used to construct a multi-scale attention weight matrix based on the feature peak position, peak height, and impedance change rate of the multimodal battery feature data. This matrix assigns dual weights based on the distance between adjacent valid features and the impedance correlation for interpolation processing.

[0011] In one example, based on multimodal battery characteristic data, the real-time charging capacity is determined through an ampere-hour integration algorithm and impedance correction, and the capacity increment is determined through an adaptive sliding window algorithm, constructing a dynamic capacity increment curve, specifically including: Based on multi-mode battery characteristic data, the real-time charging capacity is determined by using an ampere-hour integral algorithm that includes charging and discharging efficiency, as well as by including the impedance change at the current moment and the impedance correction coefficient. The window size of the sliding window is adaptively and dynamically adjusted according to the voltage change rate, and the corresponding sliding step size is determined according to the window size. Based on the capacity change and voltage change corresponding to the sliding window, the corresponding capacity increment is determined, and a dynamic capacity increment curve is constructed. The dynamic capacity increment curve is smoothed by Gaussian filtering and morphological filtering, and baseline correction is performed by using the capacity increment curve and impedance baseline under preset battery health conditions. Invalid curve segments are then removed based on impedance data.

[0012] In one example, based on the dynamic capacity increment curve and the multimodal battery characteristic data, dynamic features at multiple scales are extracted, and feature selection is performed based on nonlinear correlation analysis to extract effective features, specifically including: Based on the dynamic capacity increment curve and the multimodal battery characteristic data, dynamic features at multiple scales are extracted; the dynamic features at multiple scales include: micro-peak features describing the instantaneous morphology of a single characteristic peak, meso-trend features describing the evolution trend of a single characteristic peak, and macro-correlation features describing the correlation between different characteristic peaks. The degree of nonlinear correlation between each dynamic feature and the actual capacity label is determined by multinomial kernel correlation coefficient, and nonlinear characteristics that match the actual capacity label are selected. Based on random forest and gradient boosting tree, the feature importance score of each nonlinear characteristic is calculated by weighted fusion, and candidate features are selected. By using a dual analysis of mutual information entropy and distance correlation coefficient, redundant features are screened out from the candidate features to obtain effective features that are strongly correlated with capacity decay.

[0013] In one example, the model architecture of the collaborative model sequentially includes an input layer, a first one-dimensional convolutional layer, a first self-attention module, a first one-dimensional pooling layer, a second one-dimensional convolutional layer, a third one-dimensional convolutional layer, a second self-attention module, a fourth one-dimensional convolutional layer, a fifth one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer. The first self-attention module and the second self-attention module assign corresponding weights to different features by calculating the correlation weights between temporal features; During the training process of the collaborative model, the adaptive maximum mean difference domain adaptation module dynamically adjusts the weight of the maximum mean difference loss corresponding to the maximum mean difference value based on the distribution similarity between the source domain and the target domain, and the total loss function includes the maximum mean difference loss, mean squared error regression loss, and L2 regularization loss.

[0014] In one example, the collaborative model uses semi-supervised transfer training, which includes source domain pre-training, semi-supervised domain adaptation fine-tuning, and target domain battery health state estimation. The source domain pre-training involves training the self-attention-enhanced one-dimensional convolutional neural network using labeled data from the source domain; the optimization objectives include the mean squared error regression loss and the L2 regularization loss; the fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, and the fully connected layer are retained, while other layers are frozen for training; The semi-supervised domain adaptation fine-tuning involves using a pre-trained model obtained from the source domain pre-training for prediction, selecting training samples with prediction confidence higher than a preset confidence level as high-confidence pseudo-label samples, merging them with the target domain's real label samples to form a mixed label set, and performing mixed training; the optimization objective is the total loss function; fine-tuning is performed on the fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, the first self-attention module, the second self-attention module, and the fully connected layer; The target domain battery health state estimation is achieved by applying a semi-supervised, adaptively fine-tuned collaborative model to the sample data of the target domain. The inputs are the dynamic capacity increment curve and the effective features, and the output is an estimated value of the battery health state.

[0015] In one example, the method further includes: The contribution of each effective feature to capacity decay is determined by the Shapley addition method. Based on the contribution and electrochemical prior knowledge, each effective feature is mapped to each decay mechanism, and the first contribution percentage of each effective feature to each decay mechanism is quantified. For each operating condition, determine the proportion of the second contribution of each attenuation mechanism; Based on the first contribution ratio, the second contribution ratio, and the feature change threshold, a multi-stage battery life cycle is constructed.

[0016] On the other hand, this application also proposes a device for rapid capacity estimation of retired lithium-ion batteries, comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform a rapid capacity estimation method for retired lithium-ion batteries as described in any of the above examples.

[0017] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions configured to implement a rapid capacity estimation method for retired lithium-ion batteries as described in any of the above examples.

[0018] The method for rapid capacity estimation of retired lithium-ion batteries proposed in this application can bring the following benefits: 1. Improved estimation accuracy: By combining ampere-hour integration with impedance correction and constructing dynamic capacity increment curves, the large error of traditional voltage-current integration method is overcome; the self-attention enhanced one-dimensional convolutional neural network effectively captures timing features, improving the accuracy of capacity estimation.

[0019] 2. Strong adaptability across operating conditions and states: The hybrid kernel adaptive maximum mean difference module is adopted to dynamically adjust the domain adaptation loss weights, so that the model has good transfer ability between different aging paths, different operating conditions and different batches of retired batteries, reducing the estimation bias caused by distribution differences.

[0020] 3. High correlation between features and decay mechanism: Multi-scale dynamic features are extracted and screened based on nonlinear correlation analysis, retaining effective features that are essentially related to the internal electrochemical changes of the battery, thus enhancing the interpretability of the model and its ability to characterize the capacity decay process.

[0021] 4. High detection efficiency, suitable for rapid classification: It does not rely on complete charge and discharge tests, but can achieve rapid estimation based solely on dynamic capacity increment curves and multimodal characteristics, meeting the production needs of real-time monitoring of electric bicycle BMS and rapid sorting of large-scale retired batteries.

[0022] 5. Improved data utilization efficiency: By enhancing preprocessing and the adaptive sliding window algorithm, data information is fully mined under limited sample conditions, reducing the dependence on complete aging experiment data and improving the generalization ability in small sample scenarios. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the rapid capacity estimation method for retired lithium-ion batteries in this application embodiment; Figure 2 This is a schematic diagram of the collaborative model architecture in one scenario of this application embodiment; Figure 3 This is a schematic diagram of a device for rapidly estimating the capacity of decommissioned lithium-ion batteries in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0026] To address the aforementioned shortcomings, there is an urgent need for a technical solution that integrates the essential characteristics of multimodal electrochemistry with the advantages of advanced deep learning and semi-supervised transfer learning. This solution can be achieved by mining dynamic time-series features strongly correlated with aging, optimizing cross-domain adaptation mechanisms, improving the utilization rate of retired battery data, and constructing a collaborative model to enable rapid and accurate estimation of lithium-ion battery capacity.

[0027] like Figure 1As shown in the embodiments of this application, a method for rapid estimation of the capacity of retired lithium-ion batteries is provided, including: S101: For lithium-ion batteries in different initial states, collect corresponding actual capacity tags and multi-mode battery characteristic data through multi-condition aging experiments, and perform enhanced preprocessing on the multi-mode battery characteristic data.

[0028] Specifically, multiple lithium-ion batteries of different models and in different initial states are selected; the initial state is determined based on the degree of aging. The material system used in the battery may include ternary lithium (NCM523, NCM811) and lithium iron phosphate (LFP) systems, and the initial state may include fresh, mid-aged, and deep-aged.

[0029] Based on different operating parameters, corresponding multi-condition aging experiments are determined to conduct cycle aging tests on multiple lithium-ion batteries. These experiments may include cycle aging tests with different charge / discharge rates (e.g., 0.5C~2C), different ambient temperatures (e.g., -20℃~60℃), and different depths of discharge (e.g., 20%~100%).

[0030] For each aging cycle node, the corresponding actual capacity label is obtained through a baseline capacity test, and corresponding multi-mode battery characteristic data is collected. This multi-mode battery characteristic data includes voltage data, current data, temperature data, and battery impedance spectrum data. The baseline capacity test can be performed using a 0.2C constant current charge-discharge cycle to the cutoff voltage to obtain the actual capacity label. The multi-mode battery characteristic data refers to real-time voltage, current, and temperature data (sampling frequency 1Hz~10Hz) and battery impedance spectrum data (frequency range 1kHz~1MHz) during the charge-discharge process of the multi-condition aging experiment, ensuring comprehensive capture of the electrochemical process and its correlation with degradation characteristics.

[0031] Furthermore, for multimodal battery feature data, an improved Laida criterion is used to remove outliers. This improved Laida criterion, also known as the improved 3σ criterion, combines local density clustering to remove sudden outliers and isolated noise points in voltage, current, and impedance data, avoiding the erroneous deletion of valid data due to a single threshold. In the calculation process, the mean and standard deviation of each feature data are first calculated to determine the initial 3σ threshold range. Then, for data points falling outside the threshold range, their neighborhood density is calculated using a local density clustering algorithm. If the local density of a data point is significantly lower than the average density of its surrounding data points, it is identified as an outlier and removed. This effectively removes isolated outliers caused by sensor noise, equipment failure, or transient interference while retaining normal fluctuation data, thus improving data reliability.

[0032] Noise suppression is achieved through wavelet packet transform. The wavelet packet transform utilizes the db6 wavelet basis and decomposes the data into 4-6 layers, enabling layered denoising of voltage and impedance data. Thresholds are adaptively adjusted for noise in different frequency bands, preserving details of electrochemical characteristic peaks. During calculation, the wavelet packet transform decomposes the multimodal battery characteristic data into sub-signals of different frequency bands. Each sub-band signal undergoes thresholding using an adaptive threshold function, strongly suppressing high-frequency noise sub-bands and weakly suppressing low-frequency sub-bands containing characteristic peaks. This approach suppresses random noise while maximizing the preservation of key electrochemical features such as peak values ​​and inflection points.

[0033] Using the inflection points of the voltage-capacity curves during multi-condition aging experiments as a benchmark, a dynamic time warping algorithm is used to align the feature data of multi-mode batteries under different aging cycles and operating parameters in both time and capacity dimensions. In the specific execution of Dynamic Time Warping (DTW), the characteristic inflection points (e.g., characteristic peak positions) on the voltage-capacity curves during the charging process are first used as a reference benchmark to initially align the sequences under different cycles or operating conditions on the capacity axis. Then, by non-linearly warping the time axis, the minimum distance path between the two sequences is calculated, ensuring that each sequence achieves feature point matching simultaneously in both the capacity and time dimensions. This eliminates time scaling and capacity accumulation errors caused by rate and temperature differences, ensuring the consistency of feature extraction.

[0034] Interval standardization was used to standardize voltage, current, and temperature data, while Z-score standardization was used to standardize battery impedance spectrum data. The data standardization employed a dual standardization strategy combining interval standardization and Z-score standardization. Voltage, current, and temperature data were standardized using interval standardization, mapping them to the [0,1] interval. Impedance data, due to its large differences in dimensions and numerical range, and the existence of distribution shift issues, underwent Z-score standardization to reduce the impact of dimensional differences and distribution shifts on the model.

[0035] To address missing values ​​and local outliers in multimodal battery feature data, multimodal attention interpolation is employed. This involves constructing a multi-scale attention weight matrix based on the feature peak positions, peak heights, and impedance change rates of the multimodal battery feature data. This matrix assigns dual weights based on the distance to adjacent valid features and impedance correlation for interpolation. Missing values ​​can be caused by factors such as loop interruptions and sensor malfunctions, while local outliers can be caused by factors such as tab corrosion and peak distortion due to separator damage. In multimodal attention interpolation, the feature peak positions, peak heights, and impedance change rates of the source domain multimodal data (after outlier removal, noise suppression, and interval standardization) are used as references to construct a multi-scale attention weight matrix. Time-series points in missing intervals are assigned dual weights based on both adjacent valid features and impedance correlation for interpolation, as shown in Formula 1. Formula 1; in, This represents the interpolation results for missing values ​​and local outliers to be filled, where k is the number of adjacent valid feature peaks involved in the interpolation. The multimodal attention weights for the i-th adjacent effective feature peaks are used to construct a multi-scale attention weight matrix. Let be the value at the i-th adjacent valid characteristic peak. , This is the weight adjustment coefficient (which is optimized to 0.8 or 0.3 through cross-validation). The distance between the missing value, local outlier, and the i-th adjacent valid feature peak to be filled. Let be the impedance value of the i-th adjacent effective characteristic peak. This is the reference impedance threshold.

[0036] Based on this, the repair process improves the integrity of the repaired data and reduces the characteristic peak distortion rate. The data integrity can reach over 98%, while the characteristic peak distortion rate can be reduced to below 3%.

[0037] S102: Based on multi-mode battery characteristic data, the real-time charging capacity is determined by the ampere-hour integration algorithm and impedance correction, and the capacity increment is determined by the adaptive sliding window algorithm to construct a dynamic capacity increment curve.

[0038] Specifically, based on multi-mode battery characteristic data, the real-time charging capacity is determined by using an ampere-hour integral algorithm that includes charge and discharge efficiency, and by including the impedance change at the current moment and the impedance correction coefficient, as shown in Formula 2: Formula 2; in, The real-time charge capacity is after impedance correction, and t is the duration of the charge / discharge process. For charging and discharging efficiency, express The charging current at any given moment, For the ampere-hour integral part, Here are the impedance data at time t. The initial impedance of the new battery (a battery in a fresh state) is... This is the impedance correction factor (its value is usually 0.002~0.005).

[0039] The sliding window size is adaptively and dynamically adjusted based on the voltage change rate (typically ranging from 0.01Ah to 0.05Ah). The corresponding sliding step size is determined based on the window size, and is half the window size. Based on the capacity change and voltage change corresponding to the sliding window, the corresponding capacity increment is determined, and a dynamic capacity increment curve is constructed. The formula for calculating the capacity increment (IC value) is shown in Formula 3. Formula 3; in, For capacity increment, This represents the change in capacity within the sliding window. This represents the voltage change within the sliding window. , These are the capacity values ​​for the k-th sampling point and the (k+1)-th sampling point, respectively. , These are the voltage values ​​at the k-th and (k+1)-th sampling points, respectively.

[0040] The dynamic capacity increment curve is smoothed by both Gaussian filtering and morphological filtering. The standard deviation of Gaussian filtering can be set to 0.01~0.03. Gaussian filtering is mainly used to suppress high-frequency noise and preserve the overall trend of the curve. Morphological filtering, on the other hand, performs erosion and dilation operations on the curve by setting reasonable structural elements (such as a circular structural element with a radius of 0.02Ah), further eliminating local spikes and baseline drift, making the characteristic peaks of the dynamic capacity increment curve more prominent and the outline clearer.

[0041] Baseline correction is achieved by using the capacity increment curve and impedance baseline under the preset battery health state. The IC curve and impedance baseline under the fresh battery state are used as dual references. Piecewise polynomial fitting is used to remove the baseline drift during the aging process, thereby achieving baseline correction and highlighting the characteristic peak signals related to aging.

[0042] Invalid curve segments are removed based on impedance data. Valid IC curve segments are selected based on impedance data, and invalid curve segments caused by impedance abrupt changes are removed. For example, when the rate of change of battery impedance in a short period of time exceeds a preset threshold (e.g., set to 5% / s), the capacity increment curve during that period is determined to be an invalid curve segment and is removed to avoid characteristic distortion caused by sudden faults such as micro-short circuits or poor contact inside the battery.

[0043] S103: Based on the dynamic capacity increment curve and the multi-mode battery feature data, extract dynamic features at multiple scales, and perform feature screening based on nonlinear correlation analysis to extract effective features.

[0044] Specifically, based on dynamic capacity increment curves and multimodal battery characteristic data, dynamic features at multiple scales are extracted. These dynamic features at multiple scales include: microscopic peak features describing the instantaneous morphology of a single characteristic peak, mesoscopic trend features describing the evolution trend of a single characteristic peak, and macroscopic correlation features describing the relationship between different characteristic peaks.

[0045] Among them, the microscopic peak characteristics can include the number of characteristic peaks and the voltage position corresponding to each characteristic peak (the voltage position of the first characteristic peak). Voltage position of the nth characteristic peak Peak height (peak height of the first characteristic peak) Peak height of the nth characteristic peak ), peak half width (half width of the peak of the first characteristic peak) The peak half-width of the nth characteristic peak Peak area (the peak area of ​​the first characteristic peak) The peak area of ​​the nth characteristic peak Peak rise / fall rate (dIC / dt) corresponds to the characteristic reaction of lithium insertion / deintercalation inside the battery.

[0046] Mesoscopic trend characteristics may include characteristic peak voltage drift, peak height attenuation rate, IC curve integral area, and impedance-capacitance correlation coefficient (i.e., the sliding window correlation between IC value and impedance). Among these, characteristic peak voltage drift and peak height attenuation rate are shown in Formulas 4 and 5, respectively: Formula 4; Formula 5; in, , These represent the voltage values ​​of each characteristic peak in the fresh state and under the i-th cycle of aging test. This represents the characteristic peak voltage shift. , These represent the peak heights of each characteristic peak in the fresh state and under the i-th cycle of aging test. This represents the peak height attenuation rate.

[0047] Macroscopic correlation characteristics can include the voltage difference between different characteristic peaks (the voltage difference between the m-th characteristic peak and the n-th characteristic peak). Peak height ratio (the ratio of the height of the m-th characteristic peak to the height of the n-th characteristic peak) ), the slope of impedance as a function of capacitance (dZ / dQ), and the stability index of characteristic peaks under multiple operating conditions (coefficient of variation CV).

[0048] Among the obtained dynamic features, a three-step method of "kernel correlation analysis - multi-level importance ranking - redundancy removal" is used to screen effective features.

[0049] For kernel correlation analysis, the degree of nonlinear correlation between each dynamic feature and the actual capacity label is determined by using a multinomial kernel correlation coefficient, and nonlinear characteristics that match the actual capacity label are selected. Specifically, the multinomial kernel correlation coefficient replaces the traditional Pearson correlation coefficient in calculating the degree of nonlinear correlation between each dynamic feature and the actual battery capacity, mapping the dynamic features and the actual battery capacity to a high-dimensional space, thereby capturing any potential nonlinear correlations between them. By calculating the multinomial kernel correlation coefficient between each dynamic feature and the capacity label, strongly correlated features (e.g., features with an absolute correlation coefficient greater than 0.85) are selected to fit the nonlinear characteristics of battery degradation.

[0050] Based on random forest and gradient boosting tree, a weighted fusion method is used to calculate the feature importance score of each nonlinear characteristic and filter candidate features. Random forest and gradient boosting tree (XGBoost) are used to calculate the importance score of each nonlinear characteristic separately, and then the scores are weighted and averaged to fused, thereby comprehensively ranking the nonlinear characteristics and selecting the top (e.g., top 30) nonlinear characteristics as candidate features. This combines the advantages of two ensemble learning models, avoids the bias of a single model, and thus more robustly selects candidate features more relevant to battery capacity decay.

[0051] By employing a dual analysis of mutual information entropy and distance correlation coefficient, redundant features are eliminated from candidate features, resulting in effective features strongly correlated with capacity decay. Mutual information entropy captures generalized dependencies, while distance correlation coefficient specifically detects nonlinear associations. Through this dual analysis, highly redundant feature pairs are identified, and candidate features with higher importance scores are retained (both metrics exceed a threshold; for example, in feature pairs with mutual information entropy > 0.8 and distance correlation coefficient > 0.75, the more important one is retained). This process eliminates redundancy, ensuring a concise and non-repetitive feature set. Ultimately, this yields multiple cross-scale dynamic feature sets strongly correlated with capacity decay, balancing model accuracy and computational efficiency.

[0052] S104: Construct a collaborative model to output the corresponding battery health status based on the dynamic capacity increment curve and the effective features; wherein, the collaborative model extracts the temporal features of the dynamic capacity increment curve through a self-attention enhanced one-dimensional convolutional neural network, and during the training process of the collaborative model, for the adaptive maximum mean difference domain adaptation module, the maximum mean difference value between the source domain and the target domain is determined by a hybrid kernel constructed by linear kernel and Gaussian kernel, and the weight of the maximum mean difference loss corresponding to the maximum mean difference value is dynamically adjusted.

[0053] In traditional approaches, SOH estimation techniques for lithium-ion batteries can utilize conventional one-dimensional convolutional neural networks (1D-CNNs) for feature extraction. However, traditional 1D-CNNs struggle to capture long-range temporal features in dynamic capacity increment curves (IC curves) (e.g., continuous drift of feature peaks with aging, and changes in coupling between peaks), leading to insufficient mining of microscopic electrochemical features. Furthermore, conventional maximum mean discrepancy (MMD) domain adaptation modules typically employ a single linear kernel and fixed loss weights, making it difficult to accurately quantify the nonlinear distribution differences between the source and target domains. This can easily result in over-alignment leading to decreased regression accuracy or under-alignment causing generalization failure. Additionally, the target domains of retired batteries often involve small sample sizes and low utilization of unlabeled data, further limiting model adaptability.

[0054] Based on this, this application addresses the above problems through collaborative model design: 1. It introduces a self-attention mechanism and integrates it with 1D-CNN to enhance the ability to capture micro-temporal dependent features and multi-scale correlation features of the IC curve, breaking through the limitation of traditional 1D-CNN that can only capture local features; 2. It designs an adaptive MMD domain adaptation module with hybrid kernels to dynamically adjust the domain difference loss weights and achieve a balance between regression accuracy and domain distribution alignment; 3. It embeds a semi-supervised pseudo-label adaptation mechanism to lay the model foundation for subsequent utilization of unlabeled data in the target domain and improve the data utilization rate and model generalization ability in small sample scenarios.

[0055] Specifically, such as Figure 2 As shown, the collaborative model architecture constructs a 5-layer 1D-CNN network, embedding 2 self-attention modules to form a three-level feature extraction architecture: convolution to extract local features, self-attention to enhance temporal correlation, and pooling to reduce dimensionality and suppress overfitting. This design is based on the feature scale difference between the local peak (short-term) of the IC curve and the global baseline to ensure the full mining of multi-scale electrochemical features. Its model architecture includes, in sequence, an input layer, a first 1D convolutional layer, a first self-attention module, a first 1D pooling layer, a second 1D convolutional layer, a third 1D convolutional layer, a second self-attention module, a fourth 1D convolutional layer, a global average pooling layer, and a fully connected layer, as detailed in Table 1 below. Table 1. Model Architecture Diagram ; ; The first and second self-attention modules assign corresponding weights to different features by calculating the correlation weights between temporal features. The self-attention modules assign high weights to features strongly correlated with battery degradation (e.g., feature peak drift, peak decay) and low weights to irrelevant noise features, thus reinforcing the temporal dependence of key degradation features. The specific weight calculation is achieved through scaled dot product attention, as shown in Equation 6. Formula Six; in, The output weights are Q, the query vector, K, and V, all derived from the feature maps of the previous convolutional layer. Using the key vector dimension, multi-head attention is used to concatenate and fuse multi-scale attention features, further improving the comprehensiveness of feature capture.

[0056] During the training process of the collaborative model, such as Figure 2 As shown, the adaptive maximum mean difference domain adaptation module dynamically adjusts the weight of the maximum mean difference loss corresponding to the maximum mean difference value based on the distribution similarity between the source and target domains. The adaptive MMD domain adaptation module uses a hybrid kernel function to calculate the distribution difference of features between the source and target domains. Its kernel functions include linear kernels and radial basis function (RBF) kernels. The radial basis function kernel is also called a Gaussian kernel, and its definition is shown in Formula 7. Formula 7; in, For linear kernels, For Gaussian kernel, , For their respective kernel weights, the values ​​that can be taken are... , , Let x be the Gaussian kernel bandwidth, and let x and y be the feature vectors of samples taken from the source and target domains, respectively.

[0057] Meanwhile, the distributional difference of MMD is calculated as shown in Formula 8: Formula 8; Where X and Y are the source domain feature set and the target domain feature set, respectively. The difference in distribution between the source domain feature set and the target domain feature set. , Let be the feature vectors of the i-th sample in the source domain and the j-th sample in the target domain, respectively, and let n and m be the number of samples in the source domain feature set and the target domain feature set, respectively. The mapping function maps the original feature vectors to the reproducing kernel Hilbert space. This is the Reproducing Kernel Hilbert Space (RKHS).

[0058] Simultaneously, an adaptive weight adjustment mechanism is introduced to dynamically adjust the MMD loss weights based on the distribution similarity between the source and target domains. As shown in Formula Nine: Formula Nine; in, As the baseline weight, it can be set to 0.5. This represents the maximum value of MMD during training. To achieve the desired loss weights, an adaptive adjustment is made so that the greater the domain difference, the greater the loss weight of the MMD, thus avoiding the imbalance between regression accuracy and domain alignment caused by fixed weights.

[0059] And, as Figure 2 As shown, multiple loss functions are set up for collaborative optimization. The total loss function includes the maximum mean difference loss, the mean squared error regression loss, and the L2 regularization loss. The total loss function is obtained by weighted summation of multiple loss functions, as shown in Formula 10: Formula 10; in, For the total loss function, The mean squared error regression loss, For the maximum mean difference loss, For L2 regularization loss, , For the corresponding weights, It can be set to 0.001 to suppress model overfitting.

[0060] 1. Improved estimation accuracy: By combining ampere-hour integration with impedance correction and constructing a dynamic capacity increment curve, the large error of the traditional voltage-current integration method is overcome. The self-attention-enhanced one-dimensional convolutional neural network effectively captures temporal features, improving the accuracy of capacity estimation. In experiments, with the target domain data volume being only 8% (128 groups) of the source domain, the average relative error (MRE) of NCM811 battery SOH estimation was 3.2%, which is 61.0%, 52.9%, and 41.8% higher than that of traditional 1D-CNN (without transfer learning) (8.2%), naive transfer learning (without self-attention and adaptive MMD) (6.8%), and traditional scheme (five-layer perceptron + fixed MMD) (5.5%).

[0061] 2. Strong adaptability across operating conditions and states: Employing a hybrid kernel adaptive maximum mean difference module, the model dynamically adjusts the domain adaptation loss weights, enabling it to exhibit good transferability across different aging paths, operating conditions, and batches of retired batteries, reducing estimation bias caused by distribution differences. In experiments, in scenarios involving cross-model and retired battery cascade utilization, the SOH estimation MRE was <4.0%, a reduction of over 42.9% compared to traditional cross-model errors (generally >7.0%).

[0062] 3. High correlation between features and degradation mechanism: Multi-scale dynamic features are extracted and screened based on nonlinear correlation analysis, retaining effective features that are essentially related to the internal electrochemical changes of the battery, thus enhancing the model's interpretability and its ability to characterize the capacity degradation process. In experiments, through self-attention enhanced 1D-CNN and cross-scale feature extraction, the feature dimension was increased to 520 dimensions, and the average correlation (kernel correlation coefficient) with the battery degradation mechanism increased from 0.72 in the traditional scheme to 0.92, significantly enhancing the model's interpretability.

[0063] 4. High detection efficiency, suitable for rapid grading: It does not rely on complete charge-discharge tests; rapid estimation can be achieved solely based on dynamic capacity increment curves and multimodal features, meeting the production needs of real-time monitoring of electric bicycle BMS and rapid sorting of large-scale retired batteries. In experiments, the multimodal attention interpolation method achieved ≥98% data restoration integrity of retired batteries. Combined with semi-supervised pseudo-label training, the utilization rate of unlabeled data increased by 60%, and the data dependence was reduced by 50% in small sample scenarios.

[0064] 5. Improved Data Utilization Efficiency: By enhancing preprocessing and the adaptive sliding window algorithm, data information is fully mined under limited sample conditions, reducing reliance on complete aging experiment data and improving generalization ability in small sample scenarios. In experiments, single-sample estimation takes less than 3 seconds, supporting rapid grading of large-scale retired batteries (processing ≥1200 batteries per hour), adapting to the needs of electric bicycle BMS real-time monitoring and retired battery cascade utilization production lines.

[0065] In one embodiment, data domain definition and adaptation are performed.

[0066] For the source domain, it is a benchmark battery multimodal dataset with sufficient data, comprehensive operating condition coverage, and complete aging information. It includes 10 NCM523 (20Ah) batteries and 10 LFP (15Ah) batteries. Each battery has undergone full life cycle aging tests from SOH>95% to SOH=70%. Data is collected once every 50 cycles. The data volume of a single model is ≥800 sets, and the total data volume of the source domain is ≥1600 sets. The operating conditions cover 0.5C~2C charge and discharge rates, -20℃~60℃ ambient temperature, and 20%~100% depth of discharge. The input is IC curve (also known as IC timing sequence) + effective features, and the labels are actual capacity and SOH value (SOH=actual capacity / rated capacity×100%).

[0067] Specifically, its core function is to enable the model to learn the general mapping relationship between IC curve time-series features and battery capacity decay through pre-training, accumulating electrochemical characteristic knowledge under different operating conditions and aging stages, laying the foundation for cross-domain transfer. Specific specifications include: Battery type: Select mainstream and widely used lithium-ion battery models, covering different cathode material systems to ensure the representativeness of source domain data, such as ternary lithium system (NCM523, NCM811) and lithium iron phosphate system (LFP). Battery quantity and specifications: 10 NCM523 batteries (rated capacity 20Ah, rated voltage 3.7V), 10 NCM811 batteries (rated capacity 18Ah, rated voltage 3.7V), and 10 LFP batteries (rated capacity 15Ah, rated voltage 3.2V). Data scale: Each battery completes a full life cycle aging test (from initial SOH>95% to SOH=70%), and complete multimodal data and baseline capacity tags are collected every 50 cycles. The data volume of a single battery model is ≥800 sets, and the total data volume of the source domain is ≥1600 sets. Operating conditions coverage: Includes multi-dimensional aging operating condition combinations, comprehensively covering the relevant operating conditions of electric bicycle batteries in service and retired, including: charge and discharge rate (0.5C, 1C, 2C), ambient temperature (-20℃, 0℃, 25℃, 45℃, 60℃), and depth of discharge (DOD) (20%, 50%, 80%, 100%). Data content: The input features are a 500-dimensional IC curve + 20-dimensional cross-scale effective features, and the label data are the actual capacity and SOH value obtained from the baseline capacity test (0.2C constant current charge and discharge) (SOH = actual capacity / rated capacity × 100%).

[0068] For the target domain, it is a small-sample, special-condition battery dataset to be evaluated, including electric bicycle batteries, retired energy storage batteries, and ternary soft-pack batteries (18Ah), with 5 batteries of each model, each battery having 10-40 aging cycles, for a total of 50-200 sets of data (including a small number of baseline SOH labels); the data format is uniformly 500-dimensional IC curve + 20-dimensional effective features, and the sequence length difference is filled by linear interpolation.

[0069] Specifically, its core is the scenario related to retired batteries, which is the ultimate application object of transfer learning, and its specific specifications include: Battery types: retired electric bicycle batteries, retired energy storage batteries, and retired ternary soft-pack batteries (rated capacity 18Ah, with significant batch variations), tailored to actual application scenarios; Battery quantity and data scale: 5 batteries per model, 10-40 aging cycles per battery, a total of 50-200 sets of data, including only a small number (≥10 sets) of baseline SOH labels, to simulate the actual small sample and insufficient label retired battery scenario; Data format: After domain adaptation preprocessing, it is unified into a 520-dimensional input vector consisting of a 500-dimensional IC time series sequence and 20-dimensional cross-scale effective features, which is consistent with the source domain data format and adapts to the model input requirements.

[0070] To ensure the compatibility of the collaborative model with subsequent training data, domain adaptation preprocessing and clear data domain division are carried out simultaneously after the model is built, laying the foundation for subsequent semi-supervised transfer training.

[0071] Domain adaptation preprocessing is performed to address the differences in format, units, and abnormal time series between the source and target domain data. Specific adaptation preprocessing steps are designed to ensure the consistency and validity of the input model data. These steps may include time series normalization, abnormal time series removal, and sequence length standardization.

[0072] For time series normalization, voltage range normalization (mapped to 2.5V~4.3V) and capacity range normalization (mapped to 0~rated capacity) are performed on the IC time series sequences of the source and target domains. At the same time, the 20-dimensional effective features are range normalized (mapped to the [0,1] range) to eliminate the interference of dimensional differences on domain distribution calculation and model training.

[0073] To address the removal of anomalous timing sequences, the Dynamic Time Warping (DTW) algorithm is used to calculate the similarity between the target domain IC timing sequence and the source domain sequence. Anomalous sequences with a similarity of <0.6 are removed, such as incomplete IC curves caused by charging / discharging interruptions or sensor failures, and distorted sequences caused by impedance abrupt changes, ensuring the effectiveness of domain adaptation and the stability of model training.

[0074] To standardize the sequence length, all IC time series sequences in the source and target domains are completed or truncated through linear interpolation and standardized to 500 dimensions. These are then concatenated with 20-dimensional cross-scale effective features to form a 520-dimensional input vector, which meets the dimensional requirements of the model's input layer.

[0075] Define and adapt the data domains, clarify the dataset specifications and coverage of the source and target domains, and ensure the generalization ability of the source domain data and the scenario adaptability of the target domain data. The specific definitions of the source and target domains have been described above and will not be repeated here.

[0076] To ensure the stability, convergence speed, and generalization ability of the model training, basic training parameters are set based on the characteristics of the dataset and the model structure, providing a parameter benchmark for subsequent semi-supervised transfer training. Optimizer: Adam optimizer is selected, learning rate = 0.001 (optimized by grid search to balance convergence speed and training accuracy), weight decay coefficient = 0.0001; Training batch and iteration count: Batch size = 32, maximum number of iterations = 500, set early stopping strategy (stop training if the mean relative error (MRE) on the validation set does not decrease for 10 consecutive rounds) to avoid overtraining and overfitting; Regularization and dropout settings: In addition to L2 regularization, the dropout rate of the self-attention module is 0.15, and no additional dropout is set for the convolutional layers to balance feature capture capability and overfit suppression effect; Training environment: Based on Intel Core i7-12700H processor and NVIDIA RTX 3060 graphics card, the training framework uses PyTorch 1.12.0 to ensure model training efficiency and real-time performance.

[0077] Furthermore, the collaborative model uses semi-supervised transfer training, including source domain pre-training, semi-supervised domain adaptation fine-tuning, and target domain battery health state estimation.

[0078] Source domain pre-training: A self-attention-enhanced one-dimensional convolutional neural network is trained using 1600 sets of labeled data from the source domain; optimization objectives include mean squared error regression loss and L2 regularization loss. ); retain the fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, and the fully connected layer, and freeze the other layers (including the first one-dimensional convolutional layer, the first self-attention module, the first one-dimensional pooling layer, the second one-dimensional convolutional layer, the third one-dimensional convolutional layer, and the second self-attention module) for training to obtain the preliminary feature extractor.

[0079] Semi-supervised domain adaptation fine-tuning involves using a pre-trained model obtained from source domain pre-training for prediction. Training samples with prediction confidence higher than a preset confidence level (e.g., 0.9) are selected as high-confidence pseudo-label samples and merged with the target domain real label samples to form a mixed label set for mixed training. The optimization objective is the total loss function. Fine-tuning is performed on the fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, the first self-attention module, the second self-attention module, and the fully connected layer to achieve collaborative optimization of "source domain feature transfer - target domain distribution alignment - unlabeled data utilization".

[0080] The target domain battery health status estimation uses a semi-supervised domain adaptive fine-tuned collaborative model applied to sample data in the target domain. The inputs are dynamic capacity increment curves and effective features, and the output is an estimated value of battery health status. The single sample estimation time is <3s (based on an Intel Core i7-12700H processor), which meets the requirements for real-time monitoring and rapid grading.

[0081] In one embodiment, the contribution of each effective feature to capacity decay is determined using the Shapley Additive exPlanations (SHAP) method. Based on the contribution and electrochemical prior knowledge, each effective feature is mapped to each decay mechanism, quantifying the first contribution percentage of each effective feature to each decay mechanism. In one example, the graphite lithium intercalation peak voltage shift can be obtained. The dynamic correlation coefficient between lithium loss (LL) and SHOP is ≥0.95, and the contribution of SHOP is ≥35%; the decay rate of cathode delithiation peak area (LL) The dynamic correlation coefficient between impedance-capacity correlation coefficient and the rate of increase of battery internal resistance is ≥0.90, and the contribution of SHAP is ≥30%; the dynamic correlation coefficient between impedance-capacity correlation coefficient and the rate of increase of battery internal resistance is ≥0.88, and the contribution of SHAP is ≥20%.

[0082] For each operating condition, the second contribution percentage of each degradation mechanism is determined. In one example, we can obtain the following: High temperature condition (45℃~60℃): Lithium loss (LL) dominates capacity degradation, contributing ≥70%, and the characteristic peak voltage shift is... ≥0.05V; High-rate operation (1.5C~2C): Active material failure (LAM) dominates capacity decay, contributing ≥65%, with characteristic peak area decay rate ≥30%; Retired battery scenario: Non-uniform degradation (local polarization, tab corrosion), lithium loss and material failure synergistic effect, significant characteristic peak fusion phenomenon, impedance-capacity correlation coefficient ≤0.6.

[0083] Based on the first contribution proportion, the second contribution proportion, and the feature change threshold, the full life cycle of the battery in multiple stages is constructed. In an example, it can be finely divided into: Fresh stage (SOH>85%): Solid Electrolyte Interphase (SEI) growth causes slight lithium loss, <0.02V, <5%, impedance increase ≤10%; Mid-term aging (75%<SOH≤85%): Lithium loss and active material dissolution act synergistically, =0.02V~0.04V, =5%~20%, impedance increase 10%~30%; Late aging (70%<SOH≤75%): Active material shedding intensifies, =0.04V~0.06V, =20%~35%, impedance increase 30%-50%; Deep aging (SOH≤70%): Tab corrosion and separator damage dominate, the half-width of the characteristic peak increases by ≥40%, some characteristic peaks disappear, and the impedance increases by ≥50%.

[0084] As shown in Figure 3 , on the other hand, the embodiments of the present application further propose a retired lithium-ion battery capacity rapid estimation device, comprising: At least one processor; and, A memory communicatively connected with the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can perform the retired lithium-ion battery capacity rapid estimation method according to any one of the above embodiments.

[0085] On the other hand, the embodiments of the present application further propose a non-volatile computer storage medium, storing computer-executable instructions, which are configured to implement the retired lithium-ion battery capacity rapid estimation method according to any one of the above embodiments.

[0086] The various embodiments in the present application are described in a progressive manner, the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively brief, and the relevant parts can be referred to the description of the method embodiments.

[0087] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for rapid estimation of the capacity of retired lithium-ion batteries, characterized in that, include: For lithium-ion batteries in different initial states, corresponding actual capacity tags and multi-mode battery characteristic data are collected through multi-condition aging experiments, and the multi-mode battery characteristic data is enhanced and preprocessed. Based on multi-mode battery characteristic data, the real-time charging capacity is determined through an ampere-hour integration algorithm and impedance correction, and the capacity increment is determined through an adaptive sliding window algorithm to construct a dynamic capacity increment curve. Specifically, this includes: determining the real-time charging capacity based on multi-mode battery characteristic data using an ampere-hour integration algorithm that includes charging and discharging efficiency, and using impedance changes and impedance correction coefficients at the current moment; adaptively and dynamically adjusting the window size of the sliding window according to the voltage change rate, and determining the corresponding sliding step size based on the window size, so as to determine the corresponding capacity increment based on the capacity change and voltage change corresponding to the sliding window, and constructing a dynamic capacity increment curve; performing double smoothing processing on the dynamic capacity increment curve through Gaussian filtering and morphological filtering, and performing baseline correction by using the capacity increment curve and impedance baseline under a preset battery health state, and removing invalid curve segments based on impedance data; Based on the dynamic capacity increment curve and the multimodal battery characteristic data, dynamic features at multiple scales are extracted, and feature selection is performed based on nonlinear correlation analysis to obtain effective features. Specifically, this includes: extracting dynamic features at multiple scales based on the dynamic capacity increment curve and the multimodal battery characteristic data; the dynamic features at multiple scales include: microscopic peak features describing the instantaneous morphology of a single characteristic peak, mesoscopic trend features describing the evolution trend of a single characteristic peak, and macroscopic correlation features describing the correlation between different characteristic peaks; determining the degree of nonlinear correlation between each dynamic feature and the actual capacity label through polynomial kernel correlation coefficient, and selecting nonlinear characteristics that match the actual capacity label; calculating the feature importance score of each nonlinear characteristic through weighted fusion based on random forest and gradient boosting tree, and selecting candidate features; and filtering redundant features from the candidate features through dual analysis of mutual information entropy and distance correlation coefficient to obtain effective features strongly correlated with capacity decay. A collaborative model is constructed to output the corresponding battery health status based on the dynamic capacity increment curve and the effective features. The collaborative model extracts the temporal features of the dynamic capacity increment curve through a self-attention enhanced one-dimensional convolutional neural network. During the training process of the collaborative model, for the adaptive maximum mean difference domain adaptation module, the maximum mean difference value between the source domain and the target domain is determined by a hybrid kernel constructed by linear kernel and Gaussian kernel, and the weight of the maximum mean difference loss corresponding to the maximum mean difference value is dynamically adjusted.

2. The method according to claim 1, characterized in that, For lithium-ion batteries in different initial states, corresponding actual capacity labels and multi-mode battery characteristic data were collected through multi-condition aging experiments, specifically including: Multiple lithium-ion batteries of different models and in different initial states were selected; the initial state was determined based on the degree of aging. Based on different operating parameters, corresponding multi-condition aging experiments are determined to conduct cycle aging tests on the multiple lithium-ion batteries. For each aging cycle node, the corresponding actual capacity label is obtained through benchmark capacity testing, and the corresponding multi-mode battery characteristic data is collected; the multi-mode battery characteristic data includes voltage data, current data, temperature data, and battery impedance spectrum data.

3. The method according to claim 2, characterized in that, Enhanced preprocessing of the multimodal battery feature data specifically includes: For the aforementioned multimodal battery feature data, outlier removal is performed using an improved Raida criterion; Noise suppression is achieved through wavelet packet transform. Based on the inflection point of the voltage-capacity curve during the multi-condition aging test, the characteristic data of multi-mode batteries under different cycle aging tests and different operating conditions are aligned in time and capacity dimensions through a dynamic time warping algorithm. The voltage data, current data, and temperature data are standardized by interval standardization, and the battery impedance spectrum data is standardized by Z-score standardization. For missing values ​​and local outliers in the multimodal battery feature data, multimodal attention interpolation is used to construct a multi-scale attention weight matrix based on the feature peak position, peak height, and impedance change rate of the multimodal battery feature data. This matrix assigns dual weights based on the distance between adjacent valid features and the impedance correlation for interpolation processing.

4. The method according to claim 1, characterized in that, The collaborative model architecture sequentially includes an input layer, a first one-dimensional convolutional layer, a first self-attention module, a first one-dimensional pooling layer, a second one-dimensional convolutional layer, a third one-dimensional convolutional layer, a second self-attention module, a fourth one-dimensional convolutional layer, a fifth one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer. The first self-attention module and the second self-attention module assign corresponding weights to different features by calculating the correlation weights between temporal features; During the training process of the collaborative model, the adaptive maximum mean difference domain adaptation module dynamically adjusts the weight of the maximum mean difference loss corresponding to the maximum mean difference value based on the distribution similarity between the source domain and the target domain, and the total loss function includes the maximum mean difference loss, mean squared error regression loss, and L2 regularization loss.

5. The method according to claim 4, characterized in that, The collaborative model uses semi-supervised transfer training, including source domain pre-training, semi-supervised domain adaptation fine-tuning, and target domain battery health state estimation. The source domain pre-training involves training the self-attention-enhanced one-dimensional convolutional neural network using labeled data from the source domain; the optimization objectives include the mean squared error regression loss and the L2 regularization loss. The fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, and the fully connected layer are retained, while the other layers are frozen for training. The semi-supervised domain adaptation fine-tuning involves using a pre-trained model obtained from the source domain pre-training for prediction, selecting training samples with prediction confidence higher than a preset confidence level as high-confidence pseudo-label samples, merging them with the target domain's real label samples to form a mixed label set, and performing mixed training; the optimization objective is the total loss function; fine-tuning is performed on the fourth one-dimensional convolutional layer, the fifth one-dimensional convolutional layer, the first self-attention module, the second self-attention module, and the fully connected layer; The target domain battery health state estimation is achieved by applying a semi-supervised, adaptively fine-tuned collaborative model to the sample data of the target domain. The inputs are the dynamic capacity increment curve and the effective features, and the output is an estimated value of the battery health state.

6. The method according to claim 1, characterized in that, The method further includes: The contribution of each effective feature to capacity decay is determined by the Shapley addition method. Based on the contribution and electrochemical prior knowledge, each effective feature is mapped to each decay mechanism, and the first contribution percentage of each effective feature to each decay mechanism is quantified. For each operating condition, determine the proportion of the second contribution of each attenuation mechanism; Based on the first contribution ratio, the second contribution ratio, and the feature change threshold, a multi-stage battery life cycle is constructed.

7. A device for rapid capacity estimation of retired lithium-ion batteries, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the rapid capacity estimation method for retired lithium-ion batteries as described in any one of claims 1 to 6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to implement the method for rapid capacity estimation of retired lithium-ion batteries as described in any one of claims 1 to 6.

Images