A lithium-ion battery remaining useful life prediction method of capacity regeneration-aware degradation semantic modal enhancement and mamba selective state space modeling

Abstract

The application relates to a lithium ion battery residual service life prediction method based on capacity regeneration-aware degradation semantic modal enhancement and Mamba selective state space modeling, which comprises the following steps: collecting experimental data of a lithium ion battery in a cycle aging process, and constructing a capacity regeneration-aware lithium ion battery degradation feature sequence; performing constraint variation narrowband modal decomposition on the lithium ion battery degradation feature sequence, and constructing a degradation semantic modal enhancement representation; using a trend-fluctuation semantic grouping channel re-labeling mechanism to group and re-label the degradation semantic modal enhancement representation, and obtaining a channel re-labeled degradation representation; constructing a lithium ion battery residual service life prediction model based on Mamba selective state space modeling; inputting the channel re-labeled degradation representation into the lithium ion battery residual service life prediction model, and obtaining a lithium ion battery residual service life prediction result.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery remaining life prediction technology, and in particular to a lithium-ion battery remaining life prediction method based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles, energy storage power stations, portable electronic devices, and new energy grid-connected systems due to their advantages such as high energy density, long cycle life, and low self-discharge rate. With the increase in the number of charge-discharge cycles, lithium-ion batteries gradually experience aging phenomena such as active lithium loss, interface film thickening, active material degradation, and intensified polarization, leading to a continuous decline in their usable capacity. When the battery capacity decays to below a certain percentage of its rated capacity, the battery is generally considered to have reached the end of its lifespan. The remaining lifespan of a lithium-ion battery refers to the number of remaining cycles from the current cycle or the predicted start point until the battery first reaches the end of its lifespan. Therefore, accurately predicting the remaining lifespan of lithium-ion batteries is a crucial foundation for battery management systems to achieve safety warnings, maintenance decisions, and operational optimization.

[0003] Existing methods for predicting the remaining lifespan of lithium-ion batteries mainly include model-based methods and data-driven methods. Model-based methods typically rely on electrochemical models, equivalent circuit models, or empirical degradation models, which can reflect the internal degradation mechanism of the battery to a certain extent. However, these methods usually require complex parameter identification processes and strong prior knowledge, and are easily affected by temperature, rate capability, aging degree, and individual differences under complex operating conditions and across battery application scenarios. Data-driven methods, on the other hand, utilize data collected during battery operation, such as capacity, voltage, current, temperature, internal resistance, and charge / discharge time, to construct health indicators or degradation sequences, and then use machine learning or deep learning models to predict battery lifespan. These methods do not rely on precise electrochemical mechanism models and have strong data adaptability, thus they are widely used in battery lifespan prediction.

[0004] However, in actual battery aging processes, the capacity degradation trajectory is not always strictly monotonic. Influenced by factors such as reversible electrochemical reactions, resting conditions, temperature changes, and measurement errors, local capacity rebounds or short-term fluctuations often occur in the capacity sequence, causing the degradation sequence to exhibit non-stationary, multi-scale, and non-monotonic characteristics. If the original capacity sequence or health indicator sequence is directly input into the prediction model, the model is easily affected by local fluctuations and high-frequency noise, making it difficult to accurately distinguish between long-term degradation trends, local capacity regeneration fluctuations, and random disturbances, thus affecting the accuracy and stability of the remaining service life prediction.

[0005] To improve the modeling capabilities of complex degradation sequences, existing methods have attempted to use mode decomposition, attention mechanisms, and deep temporal networks for battery lifetime prediction. Mode decomposition can break down the original degradation signal into modal components of multiple scales, helping to characterize long-term decay trends and local fluctuations. Attention mechanisms can assign different weights to different feature channels or time positions, thereby enhancing the representation of key degradation features. Deep temporal networks can learn the temporal dependencies in the degradation sequence. However, existing methods still have several shortcomings. First, the multiple modal components obtained from decomposition contribute differently to the battery degradation state. If all modal components are directly concatenated into the prediction model, redundant or noisy modes are easily introduced into the model, affecting the stability of the lifetime prediction results. Second, existing channel attention mechanisms usually treat different input channels as homogeneous features for unified modeling, making it difficult to reflect the degradation semantic differences between the original degradation signal channel, the dominant modal channel, and the residual channel, and also making it difficult to differentiate and recalibrate long-term trend information and local fluctuation information. Third, some time series models based on self-attention mechanisms have high computational complexity when modeling long sequences, which is not conducive to efficient modeling and prediction of battery degradation data throughout the entire life cycle. Fourth, although some state space or deep time series models can perform sequence recursion, if there is a lack of an input recalibration mechanism that matches the degradation semantic modal representation, it is still difficult to stably identify the dominant degradation trend in capacity regeneration and local non-monotonic fluctuation scenarios.

[0006] Therefore, it is necessary to propose a method for predicting the remaining lifespan of lithium-ion batteries that can simultaneously characterize long-term degradation trends, local capacity regeneration fluctuations, and multi-scale perturbation information, and can differentiate the processing of different degradation semantic channels, so as to improve the accuracy, stability, and robustness of lifespan prediction under complex non-monotonic degradation scenarios. Summary of the Invention

[0007] To address the shortcomings of existing lithium-ion battery remaining life prediction methods under complex degradation scenarios, particularly the difficulties in accurately identifying dominant degradation trends, insufficient utilization of modal information, easy interference of prediction results by redundant or noise-sensitive channels, and inefficient modeling of long-range degradation dependencies under the combined effects of measurement noise, capacity regeneration, local non-monotonic fluctuations, and multi-scale degradation features, this invention proposes a lithium-ion battery remaining life prediction method based on capacity regeneration-aware degradation semantic modal enhancement and Mamba selective state-space modeling. This method constructs a lithium-ion battery degradation feature sequence, performs modal decomposition and dominant modal selection on key degradation signals, forming a degradation semantic modal enhancement representation containing the original degradation signal, dominant modal components, and residual information. Then, trend-related channels and modality-related channels are grouped and recalibrated according to channel degradation semantics. Finally, the recalibrated degradation representation is input into the Mamba model for time-series modeling to obtain the predicted remaining life of the lithium-ion battery.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration-aware degradation semantic modality enhancement and Mamba selective state-space modeling includes: Experimental data of lithium-ion batteries during cycle aging were collected to construct a lithium-ion battery degradation characteristic sequence with capacity regeneration sensing. Constrained variational narrowband mode decomposition is performed on the degradation feature sequence of the lithium-ion battery to construct a degradation semantic mode enhanced characterization. The degenerate semantic modality enhancement representation is recalibrated by using a trend-fluctuation semantic grouping channel recalibration mechanism to obtain the channel recalibrated degenerate representation. Construct a prediction model for the remaining service life of lithium-ion batteries based on Mamba selective state space modeling; The channel is recalibrated to reflect degradation characteristics, and then input into the lithium-ion battery remaining lifespan prediction model to obtain the lithium-ion battery remaining lifespan prediction results.

[0009] Optionally, constructing a capacity regeneration-sensing lithium-ion battery degradation feature sequence includes: Based on experimental data of lithium-ion batteries during cycle aging, a multi-source health index sequence is constructed. The multi-source health index is used to characterize the degradation state of the battery at different cycle stages. The multi-source health index sequence includes capacity, charging time, equivalent internal resistance, average temperature, incremental capacity curve characteristics, and degradation-related parameters obtained by fitting the charge-discharge curve. For multi-source health indicator scenarios, key health indicator sequences with a stronger deterioration correlation than a preset threshold are selected based on the correlation between candidate health indicators and the prediction target. When only the capacity sequence can be stably obtained from the battery data, the discharge capacity sequence is used as the core degradation feature sequence. The key health indicator sequence or capacity degradation sequence is subjected to outlier detection, missing value interpolation, smoothing and normalization to obtain the lithium-ion battery degradation feature sequence.

[0010] Optionally, performing constrained variational narrowband mode decomposition on the lithium-ion battery degradation feature sequence includes: For multi-source health indicator scenarios, the key health indicator sequences after correlation screening are used as decomposition objects; for single capacity sequence scenarios, the capacity degradation sequence is used as the decomposition object. The constrained variational narrowband mode decomposition is achieved by variational mode decomposition. Through this decomposition process, the key degradation signal is decomposed into several mode components with different center frequencies and finite bandwidths. Dominant mode components are selected from the decomposed modal components, and residual channels are constructed based on the differences between the original degenerate signal and the selected dominant mode components, thereby forming a degenerate semantic mode enhancement representation that includes the original degenerate signal channel, the dominant mode channel, and the residual channel.

[0011] Optionally, the grouping and recalibration of the degenerate semantic modality enhancement representation using the trend-fluctuation semantic grouping channel recalibration mechanism includes: Based on the degenerate semantics of the channels, the degenerate semantic modal enhancement representation is divided into trend-related channels and modality-related channels; For the aforementioned trend-related channels, global information compression is performed along the time dimension to obtain the trend channel description vector; For modal-related channels, global information compression is performed along the time dimension to obtain modal channel description vectors; The trend channel description vector and the modal channel description vector are nonlinearly mapped to obtain the corresponding channel weights. Based on the obtained channel weights, the trend-related channels and modality-related channels are recalibrated respectively; The recalibrated trend-related channels and modal-related channels are spliced ​​together to obtain the degenerate characterization after channel recalibration.

[0012] Optionally, the trend correlation channel includes: an original degradation signal channel and a residual channel, wherein the trend correlation channel is used to characterize the overall trend of battery capacity degradation, low-frequency changes, and remaining degradation information; The modal correlation channel includes the dominant modal components obtained by constrained variational narrowband mode decomposition. The modal correlation channel is used to characterize local capacity regeneration, multi-scale fluctuations, and short-term disturbance information.

[0013] Optionally, the lithium-ion battery remaining life prediction model includes: an input embedding layer, a sequence position information supplementation layer, several Mamba residual modeling modules, a normalization layer, and an output prediction layer; The input embedding layer is used to map the recalibrated multi-channel degradation features to a high-dimensional latent space; The sequence position information supplementation layer is used to supplement sequence position information; The Mamba residual modeling module is used to capture long-range dependencies and local dynamic changes in the battery degradation process along the time dimension, and adjust the state update intensity, input mapping method and readout method through selective parameters adaptively generated from the current input. The normalization layer is used to stabilize the feature distribution; The output prediction layer is used to generate lithium-ion battery capacity prediction results or remaining lifespan prediction results.

[0014] Optionally, the Mamba residual modeling module is constructed based on a selective state-space model; the Mamba residual modeling module adaptively generates selective parameters from the current input; In each of the Mamba residual modeling modules, the input sequence is first normalized, and then enters the state space modeling branch and the gated branch respectively; The state space modeling branch extracts local neighborhood information through short convolutions and captures long-range dependencies through selective state space recursion or selective scanning. The gated branch is used to control the information flow at different time steps. The module output is added to the residual path to stabilize model training and enhance the deep temporal representation capability.

[0015] Optionally, after constructing the lithium-ion battery remaining life prediction model based on Mamba selective state-space modeling, the model further includes: The remaining lifespan prediction model for the lithium-ion battery was trained as follows: The lithium-ion battery degradation feature sequence is divided into a training set, a validation set, and a test set. The training set is used to update the prediction model parameters, the validation set is used to monitor the model convergence state and determine the model configuration, and the test set is used to evaluate the model's ability to predict unknown battery degradation sequences after training. During training, the model parameters are updated based on the training set error, and the model is judged whether to continue training based on the validation set error. When the validation error no longer improves within a set number of rounds, training is stopped and the model parameters with the best validation performance are retained. During training, a hyperparameter optimization method based on tree-structured Parzen estimation is used for optimization.

[0016] The beneficial effects of this invention are as follows: This invention utilizes degraded semantic mode enhancement to construct a structured multi-channel degradation representation from the original degradation signal, dominant mode components, and residual information. This allows long-term capacity decay trends, local capacity regeneration fluctuations, and residual perturbation information to be expressed in different channels. Compared to directly inputting the original capacity sequence or directly splicing all mode components, this representation method reduces the coupling effect between trend information, fluctuation information, and noise information, and minimizes the interference of redundant or noisy modes on subsequent lifetime prediction.

[0017] This invention constructs a trend-fluctuation semantic grouping channel recalibration mechanism based on the degradation semantics of different input channels, performing channel description and weight generation for trend-related channels and modality-related channels respectively. This mechanism can enhance trend information and effective modality information closely related to capacity decay and lifetime evolution, suppress the influence of weakly correlated channels and noise-sensitive channels, thereby improving the relevance and stability of feature representation in complex non-monotonic degradation scenarios.

[0018] This invention inputs the recalibrated degradation characterization into a Mamba model for time-series modeling. It dynamically adjusts the state update and output readout processes using relevant selective parameters to capture long-range dependencies and local dynamic changes during battery degradation. Simultaneously, the Mamba model reduces computational and storage overhead in long-sequence modeling through selective scanning, making it more suitable for efficient modeling of full-lifecycle battery degradation sequences compared to strictly sequential recursive models and time-series models based on self-attention mechanisms. Therefore, this invention improves the accuracy, stability, and modeling efficiency of predicting the remaining lifespan of lithium-ion batteries. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the lithium-ion battery remaining life prediction model architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process for predicting the remaining lifespan of a lithium-ion battery according to an embodiment of the present invention. Figure 3 This is a comparison chart of the capacity predictions for the B0005 battery in this embodiment of the invention, with a prediction starting point of 50. Figure 4 This is a comparison chart of the capacity predictions for the B0005 battery in this embodiment of the invention, with a prediction starting point of 90. Figure 5 This is a comparison chart of the capacity predictions for the B0007 battery in this embodiment of the invention, with a prediction starting point of 50. Figure 6 This is a comparison chart of the capacity predictions for the B0007 battery in this embodiment of the invention, with a prediction starting point of 90. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 2 As shown, this embodiment proposes a method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling, including: Step S1: Collect experimental data of lithium-ion batteries during the cycle aging process and construct a capacity regeneration sensing lithium-ion battery degradation characteristic sequence; Step S2: Perform constrained variational narrowband mode decomposition on the lithium-ion battery degradation feature sequence to construct a degradation semantic mode enhanced representation; Step S3: Using the trend-fluctuation semantic grouping channel recalibration mechanism, the degenerate semantic modality enhancement representation is grouped and recalibrated to obtain the channel recalibrated degenerate representation; Step S4: Construct a prediction model for the remaining lifespan of lithium-ion batteries based on Mamba selective state-space modeling; Step S5: Recalibrate the degradation characterization of the channel, input it into the lithium-ion battery remaining life prediction model, and obtain the lithium-ion battery remaining life prediction result.

[0024] Furthermore, the construction of a capacity regeneration-sensing lithium-ion battery degradation characteristic sequence includes: Based on experimental data of lithium-ion batteries during cycle aging, a multi-source health index sequence is constructed. The multi-source health index is used to characterize the degradation state of the battery at different cycle stages. The multi-source health index sequence includes capacity, charging time, equivalent internal resistance, average temperature, incremental capacity curve characteristics, and degradation-related parameters obtained by fitting the charge-discharge curve. For multi-source health indicator scenarios, key health indicator sequences with a stronger deterioration correlation than a preset threshold are selected based on the correlation between candidate health indicators and the prediction target. When only the capacity sequence can be stably obtained from the battery data, the discharge capacity sequence is used as the core degradation feature sequence. The key health indicator sequence or capacity degradation sequence is subjected to outlier detection, missing value interpolation, smoothing and normalization to obtain the lithium-ion battery degradation feature sequence.

[0025] Specifically, in step S1, experimental data of lithium-ion batteries during cycle aging is acquired, and a health indicator sequence is constructed based on multi-source measurement information such as voltage, current, temperature, capacity, internal resistance, and charging time. The health indicators include capacity, charging time, equivalent internal resistance, average temperature, incremental capacity curve characteristics, and degradation-related parameters obtained by fitting charge-discharge curves. For multi-source health indicator scenarios, health indicators with strong degradation correlation are selected based on the correlation between candidate health indicators and the prediction target to reduce the interference of weakly correlated indicators and noise-sensitive indicators on subsequent mode enhancement and lifetime prediction processes. The correlation selection is preferably based only on training batteries or training data to avoid information leakage during the verification and testing phases. When only the capacity sequence can be stably obtained from the battery data, the discharge capacity sequence is used as the core degradation feature sequence. Outlier detection, missing value interpolation, smoothing, and normalization are performed on the multi-source health indicator sequence or capacity degradation sequence to obtain a lithium-ion battery degradation feature sequence suitable for model input.

[0026] Furthermore, performing constrained variational narrowband mode decomposition on the lithium-ion battery degradation characteristic sequence includes: For multi-source health indicator scenarios, the key health indicator sequences after correlation screening are used as decomposition objects; for single capacity sequence scenarios, the capacity degradation sequence is used as the decomposition object. The constrained variational narrowband mode decomposition is achieved by variational mode decomposition. Through this decomposition process, the key degradation signal is decomposed into several mode components with different center frequencies and finite bandwidths. Dominant mode components are selected from the decomposed modal components, and residual channels are constructed based on the differences between the original degenerate signal and the selected dominant mode components, thereby forming a degenerate semantic mode enhancement representation that includes the original degenerate signal channel, the dominant mode channel, and the residual channel.

[0027] Specifically, in step S2, key degradation signals that represent the battery degradation process are selected as the objects of constrained variational narrowband mode decomposition. For multi-source health indicator scenarios, the key health indicators selected in step S1 are used as the decomposition objects; for single capacity sequence scenarios, the capacity degradation sequence is directly used as the decomposition object. The constrained variational narrowband mode decomposition can be implemented using variational mode decomposition, or it can be implemented using variational mode decomposition with adaptive mode number estimation, parameter optimization, or recursive window update. Through this decomposition process, the key degradation signal is decomposed into several modal components with different center frequencies and finite bandwidths, thereby separating the long-term degradation trend, local capacity regeneration fluctuations, and random disturbance information in the mode space. Subsequently, the dominant modal component is selected from the decomposed modal components, and a residual channel is constructed based on the difference between the original degradation signal and the selected dominant modal component. This forms a degradation semantic mode enhancement representation that includes the original degradation signal channel, the dominant modal channel, and the residual channel, so that the physical meaning of the original degradation trajectory, multi-scale fluctuation information, and residual information not fully represented by the dominant mode can be preserved simultaneously.

[0028] First, a candidate multi-source health indicator sequence is constructed based on measurement data such as voltage, current, temperature, capacity, internal resistance, and charging time. Then, using only training data, the correlation between each candidate health indicator and the prediction target is calculated. From this, key health indicators with strong degradation correlations are selected as inputs for subsequent mode decomposition and lifetime prediction. If only the capacity sequence can be stably obtained, the capacity sequence is directly used as the key degradation signal.

[0029] Furthermore, the grouping and recalibration of the degenerate semantic modality enhancement representation using the trend-fluctuation semantic grouping channel recalibration mechanism includes: Based on the degenerate semantics of the channels, the degenerate semantic modal enhancement representation is divided into trend-related channels and modality-related channels; For the aforementioned trend-related channels, global information compression is performed along the time dimension to obtain the trend channel description vector; For modal-related channels, global information compression is performed along the time dimension to obtain modal channel description vectors; The trend channel description vector and the modal channel description vector are nonlinearly mapped to obtain the corresponding channel weights. Based on the obtained channel weights, the trend-related channels and modality-related channels are recalibrated respectively; The recalibrated trend-related channels and modal-related channels are spliced ​​together to obtain the degenerate characterization after channel recalibration.

[0030] Specifically, in step S3, to address the issue of inconsistent contributions from different channels to battery health status in the degradation semantic modal enhancement representation, a trend-fluctuation semantic grouping channel recalibration mechanism is constructed. Based on the channel's degradation semantics, the modal enhancement degradation representation is divided into trend-related channels and modal-related channels. Trend-related channels include the original degradation signal channel and the residual channel, used to characterize the overall trend of battery capacity degradation, low-frequency changes, and remaining degradation information; modal-related channels include the dominant modal components obtained through constrained variational narrowband modal decomposition, used to characterize local capacity regeneration, multi-scale fluctuations, and short-term disturbance information. Global information compression is performed along the time dimension on both trend-related and modal-related channels to obtain corresponding channel description vectors. Then, a lightweight weight generation structure is used to obtain the importance weights of different channels within each group, and these weights are applied to the corresponding input channels to achieve channel recalibration. After this step, trend channels and modal channels closely related to lifetime degradation are enhanced, while redundant modal channels and noise-sensitive channels are suppressed, thus obtaining the recalibrated degradation representation.

[0031] Furthermore, the lithium-ion battery remaining life prediction model includes: an input embedding layer, a sequence position information supplementation layer, several Mamba residual modeling modules, a normalization layer, and an output prediction layer; The input embedding layer is used to map the recalibrated multi-channel degradation features to a high-dimensional latent space; The sequence position information supplementation layer is used to supplement sequence position information; The Mamba residual modeling module is used to capture long-range dependencies and local dynamic changes in the battery degradation process along the time dimension, and adjust the state update intensity, input mapping method and readout method through selective parameters adaptively generated from the current input. The normalization layer is used to stabilize the feature distribution; The output prediction layer is used to generate lithium-ion battery capacity prediction results or remaining lifespan prediction results.

[0032] Specifically, in step S4, the recalibrated degradation characterization obtained in step S3 is input into a lithium-ion battery remaining life prediction model based on Mamba selective state space modeling. The prediction model includes an input embedding layer, a sequence position information supplementation layer, several Mamba residual modeling modules, a normalization layer, and an output prediction layer. The input embedding layer maps multi-channel degradation features to a high-dimensional latent space; the sequence position information supplementation layer supplements sequence position information; the Mamba model captures long-range dependencies and local dynamic changes in the battery degradation process along the time dimension, and adjusts the state update intensity, input mapping method, and readout method through selective parameters adaptively generated from the current input; the normalization layer stabilizes the feature distribution; and the output prediction layer generates lithium-ion battery capacity prediction results or remaining life prediction results. This step improves the model's ability to characterize local non-monotonic fluctuations and capacity regeneration phenomena while maintaining long-range time-series modeling capabilities.

[0033] Furthermore, the model for predicting the remaining lifespan of lithium-ion batteries based on Mamba selective state-space modeling also includes: The remaining lifespan prediction model for the lithium-ion battery was trained as follows: The lithium-ion battery degradation feature sequence is divided into a training set, a validation set, and a test set. The training set is used to update the prediction model parameters, the validation set is used to monitor the model convergence state and determine the model configuration, and the test set is used to evaluate the model's ability to predict unknown battery degradation sequences after training. During training, the model parameters are updated based on the training set error, and the model is judged whether to continue training based on the validation set error. When the validation error no longer improves within a set number of rounds, training is stopped and the model parameters with the best validation performance are retained. During training, a hyperparameter optimization method based on tree-structured Parzen estimation is used for optimization.

[0034] Specifically, in step S5, the lithium-ion battery degradation feature sequence is divided into a training set, a validation set, and a test set. The training set is used to learn the parameters of the prediction model, and the validation set is used to monitor changes in model error, resulting in a well-trained prediction model. During training, a hyperparameter optimization method based on tree-structured Parzen estimation is used to optimize the constrained variational narrowband mode decomposition parameters, channel recalibration parameters, and model training parameters to improve the stability of the model configuration. After training, the degradation feature sequence of the test battery is processed in the same manner as during the training phase and input into the well-trained prediction model to obtain the predicted remaining lifespan of the lithium-ion battery. The prediction model can output a normalized remaining lifespan label, a health status label, or a capacity prediction target; when the output is a capacity prediction target, the remaining lifespan of the lithium-ion battery is further calculated based on the cycle position where the capacity reaches the lifespan termination threshold.

[0035] This invention discloses a method for predicting the remaining lifespan of lithium-ion batteries by combining capacity regeneration-aware degradation semantic mode enhancement with a Mamba model. The method mainly includes the following processes: constructing a degradation feature sequence of the lithium-ion battery, constrained variational narrowband mode enhancement, trend-fluctuation semantic grouping channel recalibration, Mamba model time-series modeling, and model training and lifespan prediction output. First, the method constructs a degradation feature sequence based on battery cycle aging data. Then, it uses constrained variational narrowband mode decomposition to convert key degradation signals into a degradation semantic mode enhancement representation containing the original degradation signal, dominant mode components, and residual information. Next, it recalibrates different degradation semantic channels through a trend-fluctuation semantic grouping channel recalibration mechanism. Finally, it inputs the recalibrated degradation representation into the Mamba model to obtain the predicted remaining lifespan of the lithium-ion battery.

[0036] Step S1: Construct a capacity regeneration sensing sequence of lithium-ion battery degradation characteristics; First, experimental data on lithium-ion batteries during cycle aging were acquired. For data including multi-source measurement information such as voltage, current, temperature, capacity, internal resistance, and charging time, a multi-source health index sequence was constructed. These multi-source health indices characterize the battery's degradation state at different cycle stages, including capacity, charging time, equivalent internal resistance, average temperature, incremental capacity curve characteristics, and degradation-related parameters obtained by fitting charge-discharge curves.

[0037] In one implementation, incremental capacity characteristics are constructed based on capacity and voltage sampling data during the constant current charging phase. Let the... The capacity of each sampling point is The voltage is The degradation-related fitting parameters are extracted through the relationship of capacity changes, specifically: in, and To fit the obtained degradation-related parameters, This is for measuring the noise term. Parameters Parameters are used to characterize the nonlinear degradation intensity during capacity change processes. This is used to characterize the baseline offset in the fitted relationship. Further, it is calculated based on the capacity-voltage curve. And extract features such as the maximum value, minimum value, and variance of the incremental capacity curve. Let the first... The average temperature corresponding to each cycle is The equivalent internal resistance is The constant current charging time is Then the multi-source health indicator vector is represented as: in, , and They represent the first The maximum, minimum, and variance of the incremental capacity curve under each cycle. Indicates the first Each cycle corresponds to a multi-source health indicator vector.

[0038] To reduce the interference of weakly correlated and noise-sensitive indicators on the lifetime prediction process, in multi-source health indicator scenarios, only the training battery or training data is used to calculate the Pearson and Spearman correlation coefficients between candidate health indicators and the normalized prediction target. The Pearson correlation coefficient reflects the linear association between the candidate health indicator and the prediction target, while the Spearman correlation coefficient reflects their monotonic consistency. Weakly correlated or unstable health indicators are screened out according to the bicorrelation criterion, while health indicators with strong degradation correlations are retained for subsequent modality enhancement.

[0039] When only the capacity sequence is stably available from the battery data, the discharge capacity sequence is used as the core degradation feature sequence. After outlier handling, missing value interpolation, and smoothing, it is directly input into the subsequent constrained variational narrowband mode decomposition module. Subsequently, the multi-source health index sequence or capacity degradation sequence is normalized to obtain a lithium-ion battery degradation feature sequence suitable for model input. The statistical parameters in the normalization process are determined based on the training data and remain consistent during the validation and testing phases to ensure the independence of the training and testing processes.

[0040] Step S2: Construct a degenerate semantic mode enhancement representation based on constrained variational narrowband mode decomposition; Key degradation signals representative of the battery degradation process are selected as the objects of constrained variational narrowband mode decomposition. For multi-source health indicator scenarios, key health indicators after correlation screening are used as the decomposition objects; for single capacity sequence scenarios, the capacity degradation sequence is used as the decomposition object. Let the key degradation signal be... One implementation of constrained variational narrowband mode decomposition is variational mode decomposition, which will... Decomposed into There are 1 modal component, each denoted as . The corresponding center angular frequency is Then we have: The decomposition method solves for the modal components using constrained variational optimization, ensuring that each modal component has a finite bandwidth and is distributed around its corresponding center frequency. Its optimization objective is expressed as: in, Represents the number of decomposed modes. Indicates the first One modal component, Indicates the corresponding center angular frequency. This represents the Dirac function. Represents the imaginary unit. This represents the convolution operation. Represents the time derivative operator. This represents the L2 norm.

[0041] To solve the above constrained optimization problem, we introduce a quadratic penalty term and Lagrange multipliers to construct the augmented Lagrange function: in, The augmented Lagrangian function represents the decomposition problem. As a penalty factor, For Lagrange multipliers, The inner product is represented by the above modal decomposition. Through this modal decomposition, the long-term degradation trend, local capacity regeneration fluctuations, and random perturbation information in the key degradation signal can be separated in the modal space, thus providing a structured multi-channel degradation characterization for subsequent models.

[0042] After obtaining the modal component set Then, the modal components are filtered. Let the length be... Within the time window, the first The modal component in the th ... The discrete values ​​at each position are ,in The energy of this modal component is expressed as: Select from all modal components based on modal energy or degradation correlation. There are 1 dominant mode components. Let the set of indices of the selected dominant mode components be . To avoid the complete discarding of unselected modal information, a residual channel is constructed: Therefore, within the time window, the first Construct modality-enhanced input vectors at each of the following locations: The length is The input vectors at all positions within the time window are stacked to obtain the degenerate semantic modality enhancement representation: in, This represents the total number of channels after modal enhancement, including the retained original degraded signal channels or health indicator channels, the dominant modal channels selected through modal decomposition, and the residual channels when residual modeling is enabled. This degraded semantic modal enhancement representation simultaneously includes the original degraded signal channels, the dominant modal channels, and the residual channels, preserving the physical meaning of the original degraded signal while introducing multi-scale modal information and residual degradation information.

[0043] Step S3: Construct a feature enhancement method based on trend-fluctuation semantic grouping channel recalibration; The channels in the degradation semantic modality enhancement representation are not homogeneous features, but rather correspond to the original degradation signal, the dominant modal component, and residual information, respectively. Different channels contribute differently to the prediction of remaining battery life. Directly inputting all channels with the same weight into the time series model can easily lead to redundant modes and noise-sensitive channels interfering with subsequent predictions. Therefore, this invention constructs a trend-fluctuation semantic grouping channel recalibration mechanism to recalibrate different degradation semantic channels in groups.

[0044] Specifically, modal enhancement degradation characterization Based on channel degradation semantics, channels are divided into trend-related channels and modality-related channels: in, The trend-related channels mainly include the original degradation signal channel and the residual channel, which are used to characterize the overall trend, low-frequency changes and remaining degradation information of battery capacity degradation; The modal correlation channel mainly includes the dominant modal component, which is used to characterize local capacity regeneration, multi-scale fluctuations and short-term disturbance information. and These represent the number of trend-related channels and the number of modal-related channels, respectively.

[0045] For trend-related channels, global information compression is performed along the time dimension to obtain the trend channel description vector: For modality-related channels, global information compression is performed along the time dimension to obtain the modality channel description vector: in, and These represent the description vectors for the trend-related channel and the modality-related channel, respectively.

[0046] Subsequently, nonlinear mappings are performed on the trend channel description vector and the modal channel description vector respectively to obtain the corresponding channel weights: in, and These represent the weight vectors for the trend-related channel and the modality-related channel, respectively. , , and This is the learnable parameter matrix within the corresponding group; Represents a non-linear activation function. This represents the Sigmoid function.

[0047] Based on the obtained channel weights, the trend-related channels and modality-related channels are recalibrated respectively: Finally, the recalibrated trend correlation channels and modality correlation channels are concatenated to obtain the degenerate representation of the channels after recalibration: in, , , This indicates splicing along the channel dimension. This represents the degradation representation after channel recalibration via trend-fluctuation semantic grouping. After this step, the trend and modal channels, which are closely related to lifetime degradation, are enhanced, while redundant modal channels and noise-sensitive channels are suppressed, thus obtaining a degradation representation suitable for input into the Mamba model.

[0048] Step S4: Construct a prediction model for the remaining lifespan of lithium-ion batteries based on Mamba selective state-space modeling, such as... Figure 1 As shown.

[0049] The channel recalibration degradation characterization obtained in step S3 is input into the lithium-ion battery remaining life prediction model based on Mamba selective state space modeling. The prediction model includes an input embedding layer, a sequence position information supplementation layer, several Mamba residual modeling modules, a normalization layer, and an output prediction layer.

[0050] The input embedding layer is used to map the recalibrated multi-channel degradation features to a high-dimensional latent space. Let the channel recalibrated degradation representation of the batch input be... ,in, Indicates batch size. Indicates the length of the time window. This indicates the number of input channels. For a single sample, after passing through the input embedding layer and the sequence position information supplementation layer, the sequence representation is obtained: in, Indicates the embedding dimension. Indicates the first Embedding vectors for each discrete time step.

[0051] The Mamba residual modeling module is built upon a selective state-space model. Its core lies in the adaptive generation of selective parameters from the current input, enabling state updates and output readouts to dynamically change with different degradation stages and local fluctuations. A typical continuous-time state-space model is represented as follows: in, Indicates input, Indicates a hidden state. Indicates the output. , , and This represents the system parameter matrix. After zero-order hold-bar discretization, the continuous system can be mapped to a discrete state-space form, whose discretization parameters are determined by the time step. Confirmed, the expression is: in, This represents the identity matrix. The discretization parameters of traditional structured state-space models are typically relatively fixed during sequence processing, making it difficult to dynamically adjust them for different degradation stages and local fluctuations. Therefore, the Mamba model described in this invention further introduces input-related parameters, enabling state updates and output mappings to dynamically change with the current input. Its discrete state update process (i.e., the expression corresponding to the selective state-space model) is expressed as: in, , and For dynamic parameters related to the current input, and For shared learnable parameters, and They represent the first Discretized state transition parameters and input mapping parameters corresponding to each discrete time step. Because , and It can be adaptively generated from the current discrete-time step input, so the state update intensity and readout method at different cycle positions can be dynamically adjusted according to the input content. When there is capacity regeneration or random perturbation in the degradation sequence, this structure can reduce the impact of irrelevant perturbations on state updates; when the degradation trend is obvious, this structure can enhance the effective information related to lifetime evolution.

[0052] The selective state-space model corresponds to the discrete state update equation and the generation relationship of input-related parameters. Among them, , and The input-dependent mechanism, which adaptively generates the discretization parameters, input mapping, and output readout based on the current input, allows the discretization parameters, input mapping, and output readout to change dynamically with the input at different time steps. This mechanism reflects the selectivity of the Mamba state-space model.

[0053] In each Mamba residual modeling module, the input sequence is first normalized before entering the state-space modeling branch and the gating branch, respectively. The state-space modeling branch extracts local neighborhood information through short convolutions and captures long-range dependencies through selective state-space recursion or selective scanning. The gating branch controls the information flow at different time steps, and the module output is added to the residual path to stabilize model training and enhance deep temporal representation capabilities. Furthermore, the selective scanning process can transform stepwise state recursion into a scanning process suitable for parallel computing, thereby reducing the computational and storage overhead in long sequence modeling. Compared to recurrent neural networks or long short-term memory networks, it can alleviate the efficiency limitations caused by strict serial recursion. Compared to time-series models based on self-attention mechanisms, it can avoid the high complexity problem caused by the increase in self-attention computation as the sequence length grows. Therefore, this module can capture long-range dependencies in the battery degradation process and improve the computational efficiency of full-lifecycle degradation sequence modeling. After several Mamba residual modeling modules, a high-level degradation time series representation is obtained, and the remaining lifespan prediction results of the lithium-ion battery are generated through a normalization layer and an output prediction layer.

[0054] Step S5: Train, optimize, verify, and output the prediction results of the remaining lifespan of the lithium-ion battery; The lithium-ion battery degradation feature sequences were divided into training, validation, and test sets. The training set was used to update the prediction model parameters, the validation set was used to monitor the model convergence status and determine the model configuration, and the test set was used to evaluate the model's ability to predict unknown battery degradation sequences after training.

[0055] Let the prediction model parameters be... The hyperparameter vector is The training set is The validation set is The model predicts the value. The actual predicted target label is The parameter learning process is represented as follows: in, Represents the regression loss function. This represents the weight decay coefficient. This indicates the number of training samples. The predicted target label... It can be a normalized remaining useful life label, a health status label, or a capacity prediction target. During training, the model parameters are updated based on the training set error, and the model continues training based on the validation set error. When the validation error no longer improves within a set number of rounds, training stops and the model parameters with the best validation performance are retained.

[0056] To obtain stable modal enhancement, channel recalibration, and prediction model configuration, a hyperparameter optimization method based on tree-structured Parzen estimation is employed. Let the validation evaluation function be... Then the hyperparameter optimization objective is expressed as: in, Represents the hyperparameter search space. This represents the validation error obtained from a complete training and validation process. The tree-structured Parzen estimation method builds a probabilistic model based on historical experimental results and generates new candidate hyperparameter combinations, thereby improving search efficiency with limited computational resources. In actual training, the parameters related to the constrained variational narrowband mode decomposition are first optimized to determine a suitable degenerate semantic mode enhancement representation; then, the model training parameters are optimized to determine a stable temporal prediction model configuration.

[0057] After training, the degradation feature sequence of the test battery is processed in the same way as in the training phase, sequentially through degradation feature sequence construction, constrained variational narrowband mode enhancement, trend-fluctuation semantic grouping channel recalibration, and Mamba model, outputting a capacity prediction sequence. The predicted lifespan termination cycle is determined based on the cycle position where the capacity reaches the lifespan termination threshold, and the remaining lifespan prediction result is calculated based on the cycle position where the capacity prediction sequence reaches the lifespan termination threshold.

[0058] To evaluate the prediction performance of the method of this invention, four indicators are used: root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²). Let the number of test samples be... , No. The true value of each test sample is The predicted value is The true mean is The four evaluation indicators are defined as follows: RMSE and MAE are used to measure the absolute error between the predicted and actual values, while MAPE is used to measure the relative error of the prediction. Used to measure how well a model fits real-world degradation changes. In the NASA dataset example, Indicates the first The actual capacity value corresponding to each test sample The values ​​represent the model's predicted capacity. RMSE and MAE are in Ah, while MAPE is presented as a decimal in the table.

[0059] In this embodiment, to verify the effectiveness of the method of the present invention, the publicly available NASA lithium-ion battery degradation dataset is used for experimental verification. The NASA dataset contains test data on capacity, voltage, current, and temperature of 18650 type lithium-ion batteries such as B0005, B0006, B0007, and B0018 during cyclic charge-discharge processes. The rated capacity of the batteries in this dataset is 2.0 Ah. In this embodiment, capacity decay to 70% of the rated capacity is considered the end-of-life state, i.e., the end-of-life threshold is set to 1.40 Ah. When the predicted capacity sequence first falls below the end-of-life threshold, the corresponding cycle position is determined as the predicted end-of-life cycle, and the remaining service life of the lithium-ion battery is calculated accordingly.

[0060] Based on step S1, in the NASA dataset example, since the auxiliary health indicators that can be stably used for cross-cell modeling are relatively limited, the cycle-by-cycle discharge capacity sequence is selected as the core degradation signal. Outlier processing, missing value interpolation, smoothing, and normalization are then performed on the capacity sequence to obtain a capacity degradation sequence suitable for model input. To verify the prediction stability at different prediction stages, this example sets two prediction starting points: the 50th cycle and the 90th cycle. The prediction starting point represents the cycle position where capacity prediction begins and the remaining lifetime is further calculated.

[0061] Based on step S2, the processed capacity degradation sequence is used as the object of constrained variational narrowband mode decomposition to obtain several modal components; then, the dominant modal component is selected according to the modal energy or degradation correlation, and residual channels are constructed to obtain the degradation semantic modality enhancement representation composed of the original capacity channel, the dominant modal channel and the residual channel.

[0062] Based on step S3, the trend-related channels and mode-related channels in the degraded semantic modality enhancement representation are recalibrated to obtain the degraded representation after trend-fluctuation semantic grouping channel recalibration. This recalibration process can enhance the trend channels and effective modal channels that are closely related to capacity decay and lifetime evolution, and reduce the impact of redundant modes and noise-sensitive channels on subsequent predictions.

[0063] Based on step S4, the degradation characterization after channel recalibration is input into the Mamba model. The state update intensity, input mapping method and readout method are adjusted by selective parameters adaptively generated from the current input. The long-range dependency and local non-monotonic fluctuations in the battery capacity degradation process are modeled. The computational and storage overhead in long sequence modeling is reduced by selective scanning. The capacity prediction sequence or remaining service life prediction result is output.

[0064] Based on step S5, during training, the training set is used to update model parameters, the validation set is used to monitor the model's convergence state and retain the model parameters with the best validation performance, and the test set is used to evaluate the model's predictive ability for unknown battery degradation sequences. To obtain a stable model configuration, this embodiment employs a hyperparameter optimization method based on tree-structured Parzen estimation to optimize the constrained variational narrowband mode decomposition parameters, model training parameters, and preprocessing parameters. The main parameter settings used in this embodiment are shown in Tables 1 and 2.

[0065] Table 1 Model training parameter settings in embodiments of the present invention Table 2. Optimization, modal enhancement, and preprocessing parameter settings of the embodiments of the present invention. To further illustrate the effectiveness of the method of this invention, this embodiment compares the method of this invention with comparative methods such as MambaLithium, BiLSTM, TimesNet, and DTNN under the same test battery and the same prediction starting point. Specifically, battery B0005 is used to verify the prediction ability under a relatively smooth degradation trajectory, and battery B0007 is used to verify the prediction ability under a degradation trajectory with capacity regeneration and local non-monotonic fluctuations. The prediction performance comparison of different methods on NASA dataset batteries B0005 and B0007 is shown in Tables 3 and 4, respectively, and the corresponding capacity prediction curve comparison results are as follows. Figures 3 to 6 As shown.

[0066] Table 3. Comparison of prediction performance of the method of this invention and the comparative method on the NASA dataset B0005 battery. Table 4. Comparison of prediction performance of the method of this invention and the comparative method on the NASA dataset B0007 battery. From Table 3 and Figure 3 , Figure 4 It can be seen that, for the B0005 battery, with prediction starting points of 50 and 90, the method of this invention achieves low RMSE, MAE, and MAPE, and high R². Specifically, when the prediction starting point is 50, the RMSE of the method of this invention is 0.0095 Ah, the MAE is 0.0072 Ah, and the R² is 0.9950; when SP=90, the RMSE of the method of this invention is 0.0096 Ah, the MAE is 0.0071 Ah, and the R² is 0.9835. Combined with the corresponding capacity prediction curves, it can be seen that the method of this invention can track the actual capacity curve well under a relatively smooth degradation trajectory and maintain a relatively stable prediction trend in the later degradation stage near the failure threshold.

[0067] From Table 4 and Figure 5 , Figure 6 As can be seen, for the B0007 battery, where capacity regeneration and local non-monotonic fluctuations are more pronounced, the method of this invention still achieves superior results at both prediction starting points. Specifically, when the prediction starting point = 50, the RMSE of the method of this invention is 0.0092 Ah, MAE is 0.0070 Ah, MAPE is 0.0045, and R² is 0.9924; when the prediction starting point = 90, the RMSE of the method of this invention is 0.0092 Ah, MAE is 0.0067 Ah, MAPE is 0.0044, and R² is 0.9767. Combined with the corresponding capacity prediction curves, it can be seen that the method of this invention can still stably track actual capacity changes even under degradation trajectories with capacity regeneration and local non-monotonic fluctuations. Compared with the comparative methods, the method of this invention has a lower overall error index and maintains a higher level in the goodness-of-fit index, indicating that it can improve the accuracy and stability of lithium-ion battery capacity prediction and remaining service life prediction under complex degradation scenarios.

[0068] The experimental results above show that the present invention obtains the original capacity information, dominant mode information and residual information by enhancing the degradation semantic mode, enhances the effective channels related to lifetime evolution by recalibrating the trend-fluctuation semantic grouping channel, and dynamically models the long-range dependencies and local dynamic changes in the battery capacity degradation process by using the Mamba model. While maintaining a low long-sequence modeling overhead, it improves the accuracy, stability and robustness of the prediction of the remaining life of lithium-ion batteries.

[0069] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the remaining lifespan of lithium-ion batteries using capacity regeneration sensing-based degradation semantic modality enhancement and Mamba selective state-space modeling, characterized in that, include: Experimental data of lithium-ion batteries during cycle aging were collected to construct a lithium-ion battery degradation characteristic sequence with capacity regeneration sensing. Constrained variational narrowband mode decomposition is performed on the degradation feature sequence of the lithium-ion battery to construct a degradation semantic mode enhanced characterization. The degenerate semantic modality enhancement representation is recalibrated by using a trend-fluctuation semantic grouping channel recalibration mechanism to obtain the channel recalibrated degenerate representation. Construct a prediction model for the remaining service life of lithium-ion batteries based on Mamba selective state space modeling; The channel is recalibrated to reflect degradation characteristics, and then input into the lithium-ion battery remaining lifespan prediction model to obtain the lithium-ion battery remaining lifespan prediction results.

2. The method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 1, characterized in that, The construction of a capacity regeneration-sensing lithium-ion battery degradation feature sequence includes: Based on experimental data of lithium-ion batteries during cycle aging, a multi-source health index sequence is constructed. The multi-source health index is used to characterize the degradation state of the battery at different cycle stages. The multi-source health index sequence includes capacity, charging time, equivalent internal resistance, average temperature, incremental capacity curve characteristics, and degradation-related parameters obtained by fitting the charge-discharge curve. For multi-source health indicator scenarios, key health indicator sequences with a stronger deterioration correlation than a preset threshold are selected based on the correlation between candidate health indicators and the prediction target. When only the capacity sequence can be stably obtained from the battery data, the discharge capacity sequence is used as the core degradation feature sequence. The key health indicator sequence or capacity degradation sequence is subjected to outlier detection, missing value interpolation, smoothing and normalization to obtain the lithium-ion battery degradation feature sequence.

3. The method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 2, characterized in that, The constrained variational narrowband mode decomposition of the lithium-ion battery degradation feature sequence includes: For multi-source health indicator scenarios, the key health indicator sequences after correlation screening are used as decomposition objects; for single capacity sequence scenarios, the capacity degradation sequence is used as the decomposition object. The constrained variational narrowband mode decomposition is achieved by variational mode decomposition. Through this decomposition process, the key degradation signal is decomposed into several mode components with different center frequencies and finite bandwidths. Dominant mode components are selected from the decomposed modal components, and residual channels are constructed based on the differences between the original degenerate signal and the selected dominant mode components, thereby forming a degenerate semantic mode enhancement representation that includes the original degenerate signal channel, the dominant mode channel, and the residual channel.

4. The method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 1, characterized in that, The grouping and recalibration of the degenerate semantic modality enhancement representation using the trend-fluctuation semantic grouping channel recalibration mechanism includes: Based on the degenerate semantics of the channels, the degenerate semantic modal enhancement representation is divided into trend-related channels and modality-related channels; For the aforementioned trend-related channels, global information compression is performed along the time dimension to obtain the trend channel description vector; For modal-related channels, global information compression is performed along the time dimension to obtain modal channel description vectors; The trend channel description vector and the modal channel description vector are nonlinearly mapped to obtain the corresponding channel weights. Based on the obtained channel weights, the trend-related channels and modality-related channels are recalibrated respectively; The recalibrated trend-related channels and modal-related channels are spliced ​​together to obtain the degenerate characterization after channel recalibration.

5. The method for predicting the remaining lifespan of lithium-ion batteries based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 4, characterized in that, The trend correlation channel includes: an original degradation signal channel and a residual channel. The trend correlation channel is used to characterize the overall trend, low-frequency changes and remaining degradation information of battery capacity degradation. The modal correlation channel includes the dominant modal components obtained by constrained variational narrowband mode decomposition. The modal correlation channel is used to characterize local capacity regeneration, multi-scale fluctuations, and short-term disturbance information.

6. The method for predicting the remaining lifespan of a lithium-ion battery based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 1, characterized in that, The lithium-ion battery remaining life prediction model includes: an input embedding layer, a sequence position information supplementation layer, several Mamba residual modeling modules, a normalization layer, and an output prediction layer. The input embedding layer is used to map the recalibrated multi-channel degradation features to a high-dimensional latent space; The sequence position information supplementation layer is used to supplement sequence position information; The Mamba residual modeling module is used to capture long-range dependencies and local dynamic changes in the battery degradation process along the time dimension, and adjusts the state update intensity, input mapping method and readout method through selective parameters adaptively generated from the current input. The normalization layer is used to stabilize the feature distribution; The output prediction layer is used to generate lithium-ion battery capacity prediction results or remaining lifespan prediction results.

7. The method for predicting the remaining lifespan of a lithium-ion battery based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 6, characterized in that, The Mamba residual modeling module is constructed based on a selective state-space model; the Mamba residual modeling module adaptively generates selective parameters from the current input; In each of the Mamba residual modeling modules, the input sequence is first normalized, and then enters the state space modeling branch and the gated branch respectively; The state space modeling branch extracts local neighborhood information through short convolutions and captures long-range dependencies through selective state space recursion or selective scanning. The gated branch is used to control the information flow at different time steps. The module output is added to the residual path to stabilize model training and enhance the deep temporal representation capability.

8. The method for predicting the remaining lifespan of a lithium-ion battery based on capacity regeneration sensing degradation semantic modality enhancement and Mamba selective state space modeling according to claim 1, characterized in that, After constructing the lithium-ion battery remaining life prediction model based on Mamba selective state-space modeling, the following steps are also included: The remaining lifespan prediction model for the lithium-ion battery was trained as follows: The degradation feature sequences of lithium-ion batteries are divided into training set, validation set and test set. The training set is used to update the prediction model parameters, the validation set is used to monitor the model convergence state and determine the model configuration, and the test set is used to evaluate the model's ability to predict unknown battery degradation sequences after training. During training, the model parameters are updated based on the training set error, and the model is judged whether to continue training based on the validation set error. When the validation error no longer improves within a set number of rounds, training is stopped and the model parameters with the best validation performance are retained. During training, a hyperparameter optimization method based on tree-structured Parzen estimation is used for optimization.

Images