Battery life prediction model and system and readable storage medium

Abstract

The invention discloses a battery life prediction model and system and a readable storage medium. The battery life prediction model comprises a data processing module, an interactive coding module, a multi-scale fusion module, a global prediction module, a trend modeling module and a hierarchical fusion module, and all the modules cooperate to achieve high-precision battery life prediction. The data processing module extracts core features and constructs an input sequence to provide high-quality support for prediction; and the interactive coding module adopts a double-axis attention mechanism to mine feature association and time sequence rules and output enhanced features. The multi-scale fusion module extracts and fuses long-term trend and local fluctuation features through a double-branch structure; and the global prediction module generates a global prediction result through feature convergence, adaptive weighted fusion and regression mapping. The trend modeling module separates a core performance index sequence and captures a battery aging rule through a dual-path network; and the hierarchical fusion module fuses the two types of prediction results by means of a gating network to avoid single-mode deviation. All the modules cooperatively improve the data utilization rate, and the prediction precision and reliability are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of battery analysis technology, and more specifically, to a battery life prediction model, system, and readable storage medium. Background Technology

[0002] Lithium-ion batteries, as the core of modern energy storage technology, have been widely used in key fields such as electric vehicles, energy storage systems, and consumer electronics. Accurate prediction of their remaining useful life (RUL) is of great engineering value and economic significance for ensuring the safe operation of the system, optimizing operation and maintenance costs, and improving user experience. In particular, achieving full life cycle performance prediction in the early stages of battery use is crucial for the intelligent decision-making and preventive maintenance of battery management systems.

[0003] Currently, battery remaining life prediction technologies mainly fall into three categories: physical model-based, data-driven, and hybrid methods. While some progress has been made, significant technical limitations remain. Physical modeling methods based on electrochemical mechanisms, though possessing good physical interpretability, face challenges such as complex parameter identification, excessive computational burden, and insufficient adaptability to individual differences, making them difficult to meet practical application needs. Data-driven methods have evolved from statistical learning to deep learning, but suffer from systemic defects: feature extraction lacks electrochemical mechanism support, often employing simple statistical features with ambiguous physical meanings and ignoring crucial electrochemical information such as relaxation processes; data utilization efficiency is low, with single-loop, single-feature vector patterns causing significant information loss; furthermore, the lack of effective time-series data augmentation strategies leads to insufficient training samples, hindering the full utilization of deep learning's advantages. Existing deep learning methods mostly employ generalized time-series architectures, lacking the ability to model interactions between features, making it difficult to integrate multi-scale time-series information such as long-term trends and short-term fluctuations in battery degradation; and insufficient domain knowledge integration results in model generalization ability and prediction accuracy that fail to meet practical engineering requirements.

[0004] Therefore, there is an urgent need for a new battery life prediction method to overcome the shortcomings of existing technologies and achieve high-precision and high-reliability prediction of remaining battery life. Summary of the Invention

[0005] This application provides a battery life prediction model, system, and readable storage medium. Through the collaborative work of six modules—data processing, interactive coding, multi-scale fusion, global prediction, trend modeling, and hierarchical fusion—it achieves high-precision and high-reliability prediction of remaining battery life.

[0006] A battery life prediction model includes: The data processing module is used to extract and preprocess features from historical battery cycle data and construct an input sequence, wherein the features include at least relaxation voltage features, temperature features and cycle-level overall features. The interactive encoding module employs a dual-axis attention mechanism to sequentially perform interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence, outputting an enhanced feature representation. The multi-scale fusion module includes a parallel temporal attention branch and a dynamic multi-scale convolution branch, which are used to extract and fuse long-term trend features and local fluctuation features of the enhanced feature representation to generate a comprehensive feature representation. The global prediction module is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, and then generate a global prediction result through regression mapping; The trend modeling module is used to separate the core performance index sequence from the input sequence, and perform attenuation trend prediction based on the core performance index sequence through a dual-path network structure to generate trend prediction results. The core performance index sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The hierarchical fusion module is used to dynamically weight and fuse the global prediction results and the trend prediction results through a trend fusion gating network to generate the final battery life prediction value.

[0007] Optionally, the interactive encoding module includes: The feature interaction attention unit is used to rearrange the dimensions of the input sequence, apply a multi-head self-attention mechanism on the feature dimensions, calculate the correlation weights between different features, and generate an intermediate feature matrix containing feature interaction information. The temporal dependency encoding unit is used to apply a Transformer encoder to the intermediate feature matrix in the time dimension to capture the long-term temporal dependencies between historical cyclic data and output the enhanced feature representation.

[0008] Optionally, the temporal attention branch includes: The temporal dependency encoder, which is a time-dimensional Transformer encoder, is used to encode the enhanced feature representation obtained by the feature interaction attention module to model the long-term temporal dependency relationship of the battery cycle sequence and output the temporal dependency enhanced representation.

[0009] Optionally, the dynamic multi-scale convolution branch includes: A lightweight controller network is used to dynamically generate weight coefficients corresponding to multiple predetermined convolutional kernel sizes based on the input feature representation; The multi-scale parallel convolutional layer, consisting of multiple one-dimensional convolutional layers with different kernel sizes, is used to extract the local temporal patterns of the enhanced features in parallel and output multiple convolutional features. The weighted fusion layer is used to perform weighted summation on the multiple convolutional features based on the weight coefficients generated by the lightweight controller network to generate dynamic convolutional features.

[0010] Optionally, the multi-scale fusion module further includes a branch fusion unit; The branch fusion unit adopts an adaptive attention mechanism to dynamically calculate the first fusion weight of the temporal dependency enhancement representation and the second fusion weight of the dynamic convolutional feature, and performs weighted fusion of the temporal dependency enhancement representation and the dynamic convolutional feature based on the first fusion weight and the second fusion weight to generate the comprehensive feature representation.

[0011] Optionally, the dual-path network structure includes: A transient state modeling path is used to extract the latest time step index value from the core performance index sequence and generate a first trend feature through a first multilayer perceptron. The temporal evolution modeling path is used to perform one-dimensional convolution and pooling operations on the complete sequence of the core performance indicators to extract temporal evolution features, and generate second trend features through a second multilayer perceptron. The trend internal fusion unit is used to dynamically fuse the first trend feature and the second trend feature through a trend fusion gating network to generate the trend prediction result.

[0012] Optionally, the calculation formula for generating the trend prediction result through the trend fusion gating network is as follows:

[0013] in, The fusion weights used to calculate the trend prediction results. The result of the trend fusion gating network processing time features. As the first trend feature, This is the second trend characteristic. This is the result of trend prediction.

[0014] Optionally, the data processing module constructs the input sequence using a sliding window mechanism; The sliding window mechanism is as follows: From the cyclic sequence of the entire battery life cycle, feature data from multiple consecutive cycles are extracted in sequence to form a training sample, and the battery performance index of the cycle after a specified interval in the future is used as the prediction target of the training sample. A training sample sequence with temporal correlation is generated by sliding a window of fixed length along the cyclic sequence, and used as the input sequence.

[0015] A battery life prediction system, wherein the system integrates the battery life prediction model described in any one of the above claims.

[0016] A readable storage medium storing parameters of a battery life prediction model, wherein the battery life prediction model is a model defined by any of the above.

[0017] As can be seen from the above technical solutions, the battery life prediction model, system, and readable storage medium provided in this application include a data processing module, an interactive encoding module, a multi-scale fusion module, a global prediction module, a trend modeling module, and a hierarchical fusion module. These modules work collaboratively to achieve high-precision prediction of battery life. The data processing module constructs an input sequence by extracting and preprocessing features from historical battery cycle data, focusing on extracting relaxation voltage features, temperature features, and cycle-level overall features. These features, directly related to the internal electrochemical processes of the battery, can fully capture key information throughout the charging and discharging process, effectively improving the effectiveness and relevance of the features. The interactive encoding module employs a dual-axis attention mechanism, sequentially performing interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence. This effectively uncovers the intrinsic correlations between different features and the temporal patterns of the data, outputting a more representative enhanced feature representation. The multi-scale fusion module, through parallel temporal attention branches and dynamic multi-scale convolution branches, extracts and fuses long-term trend features and local fluctuation features from the enhanced feature representation, generating a comprehensive feature representation. This achieves comprehensive capture of temporal information at different scales, improving the completeness of the features. The global prediction module aggregates features from the comprehensive feature representation, weights and fuses the feature representations of different branches using adaptive weights, and then generates a global prediction result through regression mapping. This fully leverages the value of each branch's features and improves feature utilization. The trend modeling module separates the core performance index sequence, including charging capacity, discharging capacity, charging energy, and discharging energy, from the input sequence. It then uses a dual-path network structure to predict the degradation trend, generating a trend prediction result that accurately captures the battery aging pattern. The hierarchical fusion module dynamically weights and fuses the global prediction result and the trend prediction result through a trend fusion gating network to generate the final battery life prediction value. This effectively avoids the bias of a single prediction mode and significantly improves prediction accuracy and reliability. The synergistic effect of these modules further improves data utilization efficiency, enabling efficient model training without a large number of training samples, and meeting the needs of early and online prediction in practical engineering. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a battery life prediction model disclosed in an embodiment of this application; Figure 2 This is a schematic diagram illustrating the training and verification results of a model disclosed in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application can be used in a wide variety of general-purpose or special-purpose computing device environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc.

[0022] The following section introduces the solution proposed in this application. The technical solution is as follows, and details are provided below.

[0023] Figure 1 This is a schematic diagram of a battery life prediction model disclosed in an embodiment of this application.

[0024] like Figure 1 As shown, the battery life prediction model may include: Data processing module 10 is used to extract and preprocess features from historical battery cycle data and construct an input sequence, wherein the features include at least relaxation voltage features, temperature features and cycle-level overall features; The interactive encoding module 20 employs a dual-axis attention mechanism to sequentially perform interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence, outputting an enhanced feature representation. The multi-scale fusion module 30 includes a parallel temporal attention branch and a dynamic multi-scale convolution branch, which are used to extract and fuse long-term trend features and local fluctuation features of the enhanced feature representation to generate a comprehensive feature representation. The global prediction module 40 is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, and then generate a global prediction result through regression mapping; The trend modeling module 50 is used to separate the core performance index sequence from the input sequence, and perform attenuation trend prediction based on the core performance index sequence through a dual-path network structure to generate trend prediction results. The core performance index sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The hierarchical fusion module 60 is used to dynamically weight and fuse the global prediction result and the trend prediction result through a trend fusion gating network to generate the final battery life prediction value.

[0025] Each module will be described in detail below.

[0026] Data processing module 10 is used to extract and preprocess features from historical battery cycle data to construct an input sequence. The features include at least relaxation voltage features, temperature features, and overall cycle-level features. The specific implementation process is as follows: (1) Data testing and collection: A standardized battery cycle test protocol is adopted to conduct full life cycle testing on the target battery. Each cycle test process includes a complete charging-resting-discharging-resting cycle. Key electrochemical parameters in each cycle are recorded: voltage, current, capacity, energy, temperature and other time-series data. The test continues until the battery capacity decays to 80% of the rated capacity, and complete battery aging trajectory data is obtained.

[0027] (2) Data preprocessing and quality control: Establish a multi-level data quality inspection mechanism. First, conduct data integrity inspection to identify and handle data missing and abnormal interruptions during the testing process. Second, implement data consistency verification to detect and correct problems such as sensor drift and measurement errors. Finally, conduct batch identification processing to automatically identify the problem of discontinuous cycle numbers caused by experimental interruption, and ensure the temporal continuity of data.

[0028] (3) Electrochemical feature extraction: Electrochemical features are extracted from multiple dimensions, including relaxation voltage, temperature, and overall cycle stage features, among which: Relaxation voltage feature extraction: For the two key stages of post-charge and post-discharge rest, a relaxation feature extraction method based on electrochemical mechanism is established. The last data point of the charge and discharge process is taken as the starting point of the relaxation voltage sequence to accurately capture ohmic voltage drop information. The extracted features include ohmic voltage drop features (the instantaneous voltage difference between the first and second points, which directly reflects the battery's ohmic internal resistance), ohmic voltage drop ratio (the proportion of ohmic voltage drop in the total relaxation amplitude, which characterizes the contribution of internal resistance to total polarization), pure relaxation slope (the slope of the relaxation curve after excluding the influence of ohmic voltage drop, which reflects the concentration polarization characteristics), as well as multiple basic statistical features such as mean, standard deviation, and quantiles. Temperature feature extraction: The temperature evolution features are systematically extracted throughout the entire charging, resting, discharging and resting cycle. For temperature time series data, multi-dimensional temperature features including maximum value, minimum value, mean value, standard deviation, range of change, and rate of change are extracted to comprehensively reflect the thermodynamic state information of the battery under different operating conditions. Cycle-level overall feature extraction: Extract the core performance indicators of each cycle, including key parameters that directly reflect the battery performance status, such as charging capacity, discharging capacity, charging energy, and discharging energy. These features will be used as inputs to a dedicated trend modeling branch to achieve in-depth modeling of the battery's core performance indicators.

[0029] Furthermore, the data processing module constructs the input sequence using a sliding window mechanism; The sliding window mechanism is as follows: From the cyclic sequence of the entire battery life cycle, feature data from multiple consecutive cycles are extracted in sequence to form a training sample, and the battery performance index of the cycle after a specified interval in the future is used as the prediction target of the training sample. A training sample sequence with temporal correlation is generated by sliding a window of fixed length along the cyclic sequence, and used as the input sequence.

[0030] Its sliding window data preprocessing workflow consists of two parts. The first part is the design of a sliding window mechanism to increase the number of training samples. Specifically, this involves establishing an innovative sliding window data augmentation strategy to transform the entire lifecycle data of a single battery into multiple progressive prediction tasks. Let the complete lifecycle of a battery contain N cycles, denoted as the sequence {C1, C2, ..., C...}. N}, where C i Let W represent the feature vector of the i-th cycle. The sliding window construction formula is defined as follows: For the window length W and the prediction interval G, the k-th training sample is defined as:

[0031] In the k-th training sample, X kThe input feature window contains feature data for W consecutive cycles, y k To predict the target, namely the discharge capacity of the (k+W+G-1)th cycle, the value range of k conforms to the sequential time logic. The data augmentation effect can be quantified as follows: for a single battery, traditional methods can only generate 1 training sample, while the sliding window mechanism can generate S=NW-G+1 training samples. For example, when N=2000, W=50, and G=1000, S=2000-50-1000+1=951 samples can be generated. Secondly, for multi-battery data integration, for a dataset containing M batteries, the total number of training samples is the sum of the number of training samples contributed by all individual batteries. The total number of training samples is:

[0032] Where, N i S represents the lifespan length of the i-th battery. i The number of training samples contributed to the i-th battery.

[0033] Interactive encoding module 20, including: The feature interaction attention unit is used to rearrange the dimensions of the input sequence, apply a multi-head self-attention mechanism on the feature dimensions, calculate the correlation weights between different features, and generate an intermediate feature matrix containing feature interaction information. The temporal dependency encoding unit is used to apply a Transformer encoder to the intermediate feature matrix in the time dimension to capture the long-term temporal dependencies between historical cyclic data and output the enhanced feature representation.

[0034] Specifically, the interactive coding module is the core functional module for realizing the collaborative modeling of electrochemical feature interaction and temporal dependency. Its core design purpose is to solve the technical pain points of existing general time series models in the process of battery life prediction, which ignore the inherent physical coupling relationship between different electrochemical features, resulting in low feature utilization and insufficient temporal modeling accuracy. Specifically, through the dual-axis attention design logic of feature first and time later, it integrates the feature interaction attention unit and the temporal dependency coding unit to complete the feature enhancement and temporal coding of the input sequence, and provides a high-quality enhanced feature representation for the feature fusion of subsequent hybrid convolutional attention.

[0035] The interactive coding module specifically includes two functionally coordinated sub-units: a feature interaction attention unit and a temporal dependency coding unit. The structure, function, and technical implementation details of each unit are as follows: The Feature Interaction Attention Unit (FIA) is designed to reconstruct the dimensions and mine the features of the time-series feature matrix input to the interaction encoding module. It is specifically used to learn the interdependencies and importance weights of different electrochemical features within the same time step (i.e., a single charge-discharge cycle of the battery), thereby generating an intermediate feature matrix containing complete feature interaction information, laying the foundation for subsequent time-series dependency modeling. The technical implementation of this unit strictly follows a feature dimension-first modeling logic, given the input time-series feature matrix X∈R. B×W×F (Where B represents the batch size during training, W represents the sliding window length, i.e., the number of historical charge-discharge cycles used for prediction, and F represents the number of electrochemical features extracted in a single cycle, covering multiple dimensions such as relaxation voltage features, temperature features, and cycle-level overall features.) The three key steps are executed sequentially: dimension rearrangement, multi-head self-attention calculation, and dimension recovery. First, dimension rearrangement is performed, transforming the original dimensions (B, W, F) of the input feature matrix into (B, F, W), making feature dimension F the new sequence dimension, thus achieving feature dimension-priority modeling and breaking the inherent limitation of time dimension priority in traditional time-series models. Then, the standard multi-head self-attention mechanism is applied to the rearranged feature matrix.

[0036] The queries (Q), keys (K), and values ​​(V) required for attention computation all originate from the feature dimension, ensuring that the model can accurately calculate the correlation strength between any two electrochemical features (such as ohmic voltage drop and temperature, pure relaxation slope and discharge capacity, etc.), autonomously identify and strengthen feature combinations that play a key role in battery aging prediction, and weaken the interference of irrelevant features. Finally, after multi-head self-attention computation, the feature matrix is ​​restored to its original dimensions (B, W, F), resulting in the output intermediate feature matrix X′∈R. B×W×F Each row of this matrix corresponds to a time step (a single charge-discharge cycle), and each row of features contains the interactive coding information between all electrochemical features within that time step. Compared with the original input feature matrix, its feature expression is more targeted and informative, and can effectively characterize the inherent physical coupling relationship between electrochemical features.

[0037] The temporal dependency encoding unit, as the core of the temporal modeling in the interaction encoding module, encodes the intermediate feature matrix output by the feature interaction attention unit in the temporal dimension. It is specifically designed to capture long-term temporal dependencies between different historical cyclic data, ultimately outputting an enhanced feature representation for subsequent model computation. This unit uses a standard Transformer encoder as its core structure. Its input is the intermediate feature matrix X′ generated by the feature interaction attention unit. Since X′ already contains complete feature interaction information, the temporal dependency encoding unit can fully utilize the fused feature interaction relationships when encoding it in the temporal dimension. This avoids the problem of feature interaction information being smoothed or weakened due to prior temporal encoding in traditional temporal models, thereby improving the accuracy of capturing long-term temporal dependencies. Specifically, the temporal-dependent encoding unit uses the multi-layer self-attention mechanism and feedforward network of the Transformer encoder to explore the correlation between different time steps layer by layer, accurately capturing the long-term evolution trend of battery capacity decay, including the overall capacity decline pattern, the gradual increase trend of internal resistance, and the feature change patterns of different aging stages. At the same time, it effectively suppresses the interference of data noise on temporal modeling, and finally outputs an enhanced feature representation that has undergone dual encoding of feature interaction and temporal dependence. This enhanced feature representation retains the feature interaction information within a single loop and integrates the temporal correlation information between different loops.

[0038] The interactive coding module employs a feature-first, time-later dual-axis attention design, which offers four significant advantages over existing general time-series model coding methods: First, it enhances feature representation capabilities. By prioritizing the learning of interactions between features at each time step, it enriches and targets feature representations, accurately depicting the physical coupling relationships between electrochemical features and aligning with the inherent laws of battery aging. Second, it improves the accuracy of time-series modeling. Since time-series dependency coding is based on an intermediate feature matrix that already contains feature interaction information, it can fully utilize the correlations between features, effectively improving the accuracy of capturing long-term time-series dependencies and avoiding time-series dependency issues. The design addresses several key issues: first, it avoids disconnect between modeling and feature interaction; second, it prevents information loss by preserving feature interaction information before temporal encoding, unlike the previous approach of encoding time first and then features later. This prevents feature interaction information from being smoothed or weakened in the early temporal encoding process, ensuring the maximum utilization of feature value; and third, it enhances model robustness by accurately modeling feature interactions and effectively capturing temporal dependencies. This allows the interaction encoding module to stably output high-quality enhanced feature representations under various operating conditions (such as different ambient temperatures and charge / discharge rates) and data noise interference, thus ensuring the robustness of the early prediction model for battery capacity degradation.

[0039] The multi-scale fusion module 30 includes parallel temporal attention branches and dynamic multi-scale convolution branches.

[0040] Temporal attention branches include: The temporal dependency encoder, which is a time-dimensional Transformer encoder, is used to encode the enhanced feature representation obtained by the feature interaction attention module to model the long-term temporal dependency relationship of the battery cycle sequence and output the temporal dependency enhanced representation.

[0041] Dynamic multi-scale convolution branches include: A lightweight controller network is used to dynamically generate weight coefficients corresponding to multiple predetermined convolutional kernel sizes based on the input feature representation; The multi-scale parallel convolutional layer, consisting of multiple one-dimensional convolutional layers with different kernel sizes, is used to extract the local temporal patterns of the enhanced features in parallel and output multiple convolutional features. The weighted fusion layer is used to perform weighted summation on the multiple convolutional features based on the weight coefficients generated by the lightweight controller network to generate dynamic convolutional features.

[0042] In addition, the multi-scale fusion module also includes a branch fusion unit; The branch fusion unit adopts an adaptive attention mechanism to dynamically calculate the first fusion weight of the temporal dependency enhancement representation and the second fusion weight of the dynamic convolutional feature, and performs weighted fusion of the temporal dependency enhancement representation and the dynamic convolutional feature based on the first fusion weight and the second fusion weight to generate the comprehensive feature representation.

[0043] The global prediction module 40 is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, thereby generating a global prediction result through regression mapping.

[0044] Specifically, the multi-scale fusion module aims to address the technical challenges of existing battery life prediction models. These models struggle to simultaneously and accurately capture both local temporal patterns (such as short-term capacity fluctuations and instantaneous electrochemical feature mutations) and global long-term temporal trends (such as continuous capacity decay) during battery aging. Furthermore, the fixed feature fusion methods fail to adapt to the dynamic feature requirements of different battery aging stages. This module employs parallel temporal attention branches and dynamic multi-scale convolution branches to mine global temporal dependencies and multi-scale local temporal features, respectively. Then, an adaptive weighted fusion of these two types of features is achieved through a branch fusion unit, ultimately generating a comprehensive feature representation that combines global trends with local details. This provides comprehensive and accurate feature support for the capacity decay trend prediction of the subsequent dedicated trend modeling branch. The module works synergistically with the interactive coding module and the dedicated trend modeling branch to further improve the accuracy and robustness of battery life prediction.

[0045] The multi-scale fusion module specifically includes a parallel temporal attention branch, a dynamic multi-scale convolution branch, and a branch fusion unit for integrating the output features of the two branches. The structure, function, and technical implementation details of each branch and unit are as follows. All parts work together to ensure effective mining and adaptive fusion of multi-scale features: The temporal attention branch, as the core branch for capturing global long-term temporal dependencies, has the core function of further temporal encoding the enhanced feature representation obtained by the feature interaction attention unit in the interactive encoding module, deeply mining the long-term temporal dependencies in the battery cycle sequence (i.e., charge and discharge cycle time series data), and finally outputting the temporal dependency enhanced representation, providing global trend information for the comprehensive feature representation. The core execution component of this branch is a time-dependent encoder, which employs a time-dimensional Transformer encoder. Its input is the enhanced feature representation output by the interactive encoding module (this feature already contains interactive information of different electrochemical features within a single time step). Through the multi-layer self-attention mechanism and feedforward network built into the Transformer encoder, the correlation between different time steps (i.e., different charge-discharge cycles) is mined layer by layer, accurately capturing the long-term evolution trend of battery capacity decay, including the overall capacity decline pattern, the gradual increase trend of internal resistance, and the feature change patterns of different aging stages. At the same time, it effectively suppresses the interference of data noise on time-series modeling, and finally outputs a time-dependent enhanced representation that can fully characterize the global long-term time-series dependency. This representation can fully reflect the macroscopic trend of battery aging and provide global feature support for subsequent feature fusion.

[0046] The dynamic multi-scale convolution branch, as the core branch for capturing multi-scale local temporal patterns, has the core function of adaptively mining local temporal features (including short-term high-frequency fluctuations and long-term low-frequency trends) at different time scales during battery aging. By dynamically adjusting the convolutional kernel weights, it adapts to the feature changes required at different aging stages of the battery, ultimately generating dynamic convolutional features that provide local detail information for the comprehensive feature representation. This branch specifically includes three functionally coherent sub-components: a lightweight controller network, multi-scale parallel convolutional layers, and a weighted fusion layer. The technical implementation and collaborative working process of each sub-component are as follows: The lightweight controller network's core function is to dynamically generate weight coefficients corresponding to multiple predetermined convolutional kernel sizes based on the enhanced feature representation input to the dynamic multi-scale convolutional branches. This enables adaptive selection of convolutional kernels, adapting to the feature extraction needs of different battery aging stages. The technical implementation of this controller network is lightweight and efficient. It performs feature analysis on the input time-series feature matrix, focusing on capturing the feature changes of the current battery aging stage, and then predicts the weight coefficient β corresponding to each predetermined convolutional kernel size. The predetermined convolutional kernel size is selected as k∈{3,7,11,15}, where small-sized convolutional kernels (k=3) are used to capture short-term high-frequency fluctuations (such as instantaneous changes in voltage and temperature within a single charge-discharge cycle), and large-sized convolutional kernels (k=11,15) are used to capture long-term low-frequency trends (such as the slow capacity decay characteristics over multiple charge-discharge cycles). The dynamic generation of the weight coefficient β allows the model to adaptively emphasize the feature weights of short-term fluctuations or long-term trends at different battery aging stages, overcoming the limitation of traditional fixed convolutional kernels being unable to adapt to dynamic aging features.

[0047] The multi-scale parallel convolutional layer consists of multiple one-dimensional convolutional layers, each with a different predetermined kernel size. These layers operate in parallel, their core function being to extract local temporal features from the input enhanced feature representation in parallel, outputting multiple convolutional features corresponding to different time scales. Each one-dimensional convolutional layer performs convolution operations on the temporal dimension of the enhanced feature representation, mining local temporal patterns at its corresponding scale. Simultaneously, a ReLU activation function is applied to the output of each convolutional layer to achieve a non-linear transformation of the features, suppressing irrelevant noise interference and enhancing the expressive power of local features. Ultimately, each convolutional layer outputs a convolutional feature at a corresponding scale, forming a multi-scale convolutional feature set, providing the foundation for subsequent weighted fusion. The results from each convolutional kernel are weighted by the weight coefficients predicted by the controller network before fusion.

[0048] The weighted fusion layer performs a weighted summation operation on multiple convolutional features output from the multi-scale parallel convolutional layers based on the weight coefficients β generated by the lightweight controller network, ultimately generating dynamic convolutional features. This weighted fusion process strictly follows the adaptive principle; the weight coefficients β directly determine the contribution of the convolutional features at the corresponding scale, enabling the dynamic convolutional features to adaptively integrate local temporal information at different time scales, achieving an optimal balance between multi-scale information, noise suppression, and computational efficiency. Through this weighted fusion method, the dynamic convolutional features can capture both short-term high-frequency fluctuations during battery aging and retain long-term low-frequency trends, effectively compensating for the limitations of single-scale convolutional features and providing high-quality local feature support for subsequent fusion with temporally dependent augmented representations.

[0049] The branch fusion unit, acting as an interactive multi-scale fusion module, adaptively weights and fuses the temporal dependency enhancement representation output by the temporal attention branch and the dynamic convolutional features output by the dynamic multi-scale convolution branch. This generates a comprehensive feature representation that combines global long-term temporal trends with multi-scale local temporal patterns, providing core input for capacity decay prediction in subsequent dedicated trend modeling branches. This unit employs an adaptive attention pooling mechanism, using a small feedforward network to dynamically calculate the fusion weights corresponding to the two types of features based on their characteristic properties: the first fusion weight for the temporal dependency enhancement representation and the second fusion weight for the dynamic convolutional features.

[0050] The specific fusion process is as follows: based on the dynamically calculated first and second fusion weights, a weighted summation operation is performed on the temporal dependency enhancement representation and the dynamic convolutional features to obtain the final comprehensive feature representation. The core advantage of this adaptive fusion method lies in its ability to automatically balance the roles of global long-term trends (captured by the temporal attention branch) and local temporal patterns (captured by the dynamic multi-scale convolution branch) according to the current aging state of the battery. In the early stages of battery aging, local temporal patterns (such as short-term feature fluctuations) contribute more to capacity prediction, and the model automatically increases the second fusion weight. In the middle and later stages of battery aging, global long-term trends (such as the continuous capacity decay pattern) contribute more, and the model automatically increases the first fusion weight. This achieves dynamic adaptation of feature fusion weights for different aging stages, further enhancing the relevance and effectiveness of the comprehensive feature representation, and providing a core guarantee for improving the accuracy of subsequent battery life prediction.

[0051] The trend modeling module 50 is used to separate the core performance indicator sequence from the input sequence, and to predict the decay trend based on the core performance indicator sequence through a dual-path network structure, generating a trend prediction result. The core performance indicator sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The dual-path network structure includes: A transient state modeling path is used to extract the latest time step index value from the core performance index sequence and generate a first trend feature through a first multilayer perceptron. The temporal evolution modeling path is used to perform one-dimensional convolution and pooling operations on the complete sequence of the core performance indicators to extract temporal evolution features, and generate second trend features through a second multilayer perceptron. The trend internal fusion unit is used to dynamically fuse the first trend feature and the second trend feature through a trend fusion gating network to generate the trend prediction result.

[0052] The trend modeling module extracts the core performance index sequence from the original time-series feature matrix input to the model, and predicts the degradation trend based on this sequence through a dual-path network structure, ultimately generating the trend prediction result. The core performance index sequence is the key basis for battery life prediction, and includes at least four core electrochemical performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. These four indicators directly reflect the battery's health status and aging degree, and are the core data source for capacity degradation trend prediction. This effectively avoids interference from irrelevant features in trend modeling, ensuring the purity of the trend prediction.

[0053] The trend modeling module employs a dual-path network structure, which models the battery performance degradation trend from different dimensions through two parallel paths, taking into account both short-term instantaneous state and long-term time-series evolution. Furthermore, the trend modeling module uses an internal fusion unit to achieve dynamic adaptive fusion of the two types of features. The specific structure, functions, and technical implementation details are as follows: Firstly, the instantaneous state modeling path aims to capture the degradation dynamics of the battery in its latest health state, extracting the latest time step index values ​​from the core performance index sequence and focusing on the impact of short-term state changes on trend prediction. This path is technically simple and efficient. Firstly, it accurately extracts the values ​​of four core indicators from the separated core performance index sequence at the latest moment within the time window (i.e., the most recent charge-discharge cycle). These values ​​directly reflect the battery's current instantaneous health state. Then, the extracted latest moment index values ​​are input into a first multilayer perceptron (MLP). Through nonlinear transformation and feature learning of the multilayer neural network, the degradation trend information contained in the latest instantaneous state is mined, ultimately generating a first trend feature (i.e., instantaneous trend feature). This feature accurately characterizes the battery's short-term degradation dynamics, providing immediate support for trend prediction.

[0054] Secondly, the temporal evolution modeling path aims to capture the long-term degradation patterns of core battery performance indicators. It extracts global temporal features from the complete core performance indicator sequence, focusing on the impact of long-term evolution trends on prediction results. This path employs a one-dimensional convolutional neural network (Conv1D) combined with pooling operations. The specific implementation process is as follows: First, the separated core performance indicator sequence X_trend∈R^(B×W×4) (where B represents the batch size, W represents the window length, and 4 represents the four core performance indicators) is transposed to obtain the transposed sequence X_trendT∈R^(B×4×W), adapting to the input requirements of one-dimensional convolution. Then, the transposed sequence is input into the Conv1D_Extractor, which contains multiple layers of one-dimensional convolution and adaptive pooling. The module extracts time-series evolution features at different scales through multi-layer one-dimensional convolution in parallel, capturing the decay patterns of core indicators with charge-discharge cycles (such as the gradual decrease in capacity and the slow decay of energy). Then, the extracted time-series features are reduced in dimensionality and filtered through an adaptive pooling module, retaining key long-term evolution information. Finally, the time-series evolution features processed by convolution and pooling are input into the second multi-layer perceptron to further enhance the feature representation capability and generate a second trend feature (i.e., time-series evolution trend feature). This feature can completely depict the long-term decay trend of the battery's core indicators, providing global support for trend prediction.

[0055] Thirdly, the trend fusion unit's core function is to dynamically fuse the first and second trend features, avoiding the isolated use of these two types of features while preventing coupling with the main network features, thus ensuring the accuracy and purity of the trend prediction results. This unit employs a specially designed trend fusion gating network, which dynamically calculates the fusion weights corresponding to the first and second trend features. The weight values ​​are adaptively adjusted by the gating network based on the characteristics of the two types of trend features, intelligently balancing the contribution of short-term instantaneous states and long-term evolutionary trends to the prediction results. The final trend prediction result is obtained by weighted summation using these dynamic weights. This fusion method fully integrates the advantages of the two types of trend features, preserving the accuracy of short-term instantaneous states while taking into account the globality of long-term evolutionary trends, further improving the reliability of the trend prediction results.

[0056] The core design advantage of the interactive trend modeling module lies in its independence and focus. As an independent lightweight neural network branch, it can focus on modeling core performance indicators, avoiding interference from global features of the main model (such as temperature and voltage curves) on the trend modeling of core indicators. This effectively solves the technical problem of information confusion in existing models and ensures the purity and accuracy of trend prediction results. At the same time, the design of dual-path parallel modeling and dynamic gating fusion takes into account both short-term and long-term trends, further improving the accuracy of trend prediction and model robustness, and providing high-quality trend prediction input for subsequent hierarchical fusion.

[0057] The hierarchical fusion module 60 is used to dynamically weight and fuse the global prediction result and the trend prediction result through a trend fusion gating network to generate the final battery life prediction value. The calculation formula corresponding to the trend prediction result generated by the gating network fusion is as follows:

[0058] in, The fusion weights used to calculate the trend prediction results. The result of the trend fusion gating network processing time features. As the first trend feature, This is the second trend characteristic. This is the result of trend prediction.

[0059] The hierarchical fusion module uses a trend fusion gating network to dynamically weight and fuse the global prediction results and the trend prediction results output by the interactive trend modeling module, ultimately generating accurate and reliable battery life prediction values. This achieves intelligent arbitration between the global complex pattern and the pure trend of the core indicator, solving the technical problems of the limitations and insufficient adaptability of single prediction results.

[0060] The technical implementation of the layered fusion module and the trend fusion unit within the interactive trend modeling module adopt a homogeneous trend fusion gating network to ensure the consistency and uniformity of the technical solution. The specific implementation process is as follows: First, the global prediction result output by the main model and the trend prediction result output by the interactive trend modeling module are input into the final fusion gating network. Then, based on the characteristics of the two prediction results (such as the completeness of the global prediction result and the purity of the trend prediction result), the gating network dynamically generates corresponding fusion weights γ. The value of weight γ is adaptively adjusted to balance the contribution of the two prediction results. Finally, based on the dynamically calculated fusion weight γ, the global prediction result and the trend prediction result are weighted and summed to obtain the final battery capacity prediction result. The specific calculation formula is as follows: final_output=γ main_output+(1-γ) trend_prediction.

[0061] The core design advantage of the hierarchical fusion module lies in its dynamic adaptability, which automatically adjusts the fusion weights based on the actual characteristics of the input samples. When the global features of the input samples (such as temperature, voltage curve anomalies, etc.) contain key predictive information and contribute more to battery life prediction, the model automatically increases the fusion weight γ, placing greater trust in the global prediction results of the main model. Conversely, when the degradation trend of core performance indicators is very clear and the trend prediction results are more reliable, the model automatically decreases the fusion weight γ, increasing the trust in the trend prediction results. This dynamic fusion strategy fully leverages the advantages of both global and trend predictions, avoids the limitations of a single prediction result, further improves the accuracy, robustness, and adaptability of battery life prediction, and ensures that the model can output stable and accurate prediction results under different operating conditions and aging stages, ultimately achieving accurate battery life prediction.

[0062] As can be seen from the above technical solutions, the battery life prediction model, system, and readable storage medium provided in this application include a data processing module, an interactive encoding module, a multi-scale fusion module, a global prediction module, a trend modeling module, and a hierarchical fusion module. These modules work collaboratively to achieve high-precision prediction of battery life. The data processing module constructs an input sequence by extracting and preprocessing features from historical battery cycle data, focusing on extracting relaxation voltage features, temperature features, and cycle-level overall features. These features, directly related to the internal electrochemical processes of the battery, can fully capture key information throughout the charging and discharging process, effectively improving the effectiveness and specificity of the features. The interactive encoding module employs a dual-axis attention mechanism, sequentially performing interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence. This effectively uncovers the intrinsic correlations between different features and the temporal patterns of the data, outputting a more representative enhanced feature representation. The multi-scale fusion module, through parallel temporal attention branches and dynamic multi-scale convolution branches, extracts and fuses long-term trend features and local fluctuation features from the enhanced feature representation, generating a comprehensive feature representation. This achieves comprehensive capture of temporal information at different scales, improving the comprehensiveness of the features. The global prediction module aggregates features from the comprehensive feature representation, weights and fuses the feature representations of different branches using adaptive weights, and then generates a global prediction result through regression mapping. This fully leverages the value of each branch's features and improves feature utilization. The trend modeling module separates the core performance indicator sequence, including charging capacity, discharging capacity, charging energy, and discharging energy, from the input sequence. It then uses a dual-path network structure to predict the degradation trend, generating a trend prediction result that accurately captures the battery aging pattern. The hierarchical fusion module dynamically weights and fuses the global prediction result and the trend prediction result through a trend fusion gating network to generate the final battery life prediction value. This effectively avoids the bias of a single prediction mode and significantly improves prediction accuracy and reliability. The synergistic effect of these modules further improves data utilization efficiency, enabling efficient model training without a large number of training samples, thus meeting the needs of early and online prediction in practical engineering.

[0063] The following example further illustrates this application. This example details the specific implementation process of the battery remaining useful life (RUL) prediction method based on enhanced electrochemical feature extraction and hybrid deep learning networks. Through a complete data collection, feature extraction, data preprocessing, model training, and application process, the effectiveness and high accuracy of the method of this invention are verified. The experimental samples, parameter settings, and experimental results used in this example are all obtained from actual verification and can be directly used as a reference for the specific implementation of the method of this invention.

[0064] 1. Data Collection

[0065] The experimental samples used for model training, validation, and testing in this example are all derived from the ORT (Ongoing Reliability Test) test data of a 104Ah square aluminum-cased battery throughout its entire lifecycle. A total of 100 complete samples were collected, with each sample corresponding to the entire lifecycle cycle test process of one battery, ensuring the representativeness and completeness of the samples. Each battery's individual charge-discharge cycle strictly followed the standard operating condition of "constant current charging – rest (15 min) – constant current discharging – rest (15 min)," synchronously collecting real-time data such as battery voltage, current, temperature, and time throughout the process. This provides comprehensive and accurate raw data support for subsequent feature extraction, ensuring that the extracted features accurately reflect the internal electrochemical state and aging patterns of the battery.

[0066] 2. Feature Extraction

[0067] This example extracts four main categories of features for each "constant current charging - resting - constant current discharging - resting" cycle obtained from the data collection, constructing a comprehensive battery state feature set. The specific extraction process is as follows: (1) Physical characteristics based on enhanced resting period analysis: This type of characteristic is the core innovation of this feature extraction. The focus is on optimizing the definition of the resting period and mining the voltage relaxation law during the resting period to extract characteristics that directly reflect the internal electrochemical mechanism of the battery. First, the starting point of the resting period is strictly optimized and defined as the precise moment when constant current charging (or discharging) ends and the current is completely cut off. This improvement ensures that the collected relaxation voltage sequence can completely include the two key physical processes of instantaneous ohmic voltage drop and subsequent polarization recovery, avoiding the loss of feature information due to the ambiguity of the resting period definition. Based on this enhanced relaxation voltage sequence, two types of sub-features are extracted simultaneously: basic statistical features and key physical mechanism features. Basic statistical characteristics: A comprehensive statistical description of the complete relaxation voltage curve is provided to capture the overall shape and numerical distribution characteristics of the voltage relaxation curve. A total of 13 statistical indicators are extracted, including mean, standard deviation, median, skewness, kurtosis, and the overall linear regression slope of the sequence, which can comprehensively characterize the macroscopic variation law of relaxation voltage.

[0068] Key physical mechanism characteristics: Directly quantifies key parameters of the internal electrochemical reaction process of the battery, focusing on reflecting core aging-related indicators such as battery internal resistance and polarization characteristics. These mainly include ohmic drop and pure relaxation slope: Ohmic drop is defined as the difference between the voltage value at the moment the charge / discharge ends and the voltage value immediately following it, i.e., the difference between the voltage value at the moment the charge / discharge ends and the voltage value at the first resting time point. This characteristic directly reflects the current ohmic internal resistance of the battery and is a core indicator for assessing the battery's health status. Pure relaxation slope is defined as the slope obtained by linear regression of the remaining voltage recovery curve (i.e., the pure polarization recovery stage) after the ohmic drop occurs. This characteristic reflects the recovery rate of internal electrochemical polarization and concentration polarization, indirectly reflecting the activity level of the internal electrochemical reaction of the battery.

[0069] The physical feature extraction method based on enhanced resting period analysis was applied to two stages: post-charging resting and post-discharging resting. It ultimately obtained two sets of approximately 36 high-information-density relaxation features (including multiple basic statistical features and multiple physical mechanism features), providing core support for the accurate identification of battery aging status.

[0070] (2) Core capacity / energy characteristics: These characteristics are the most intuitive and core indicators for evaluating battery performance, directly reflecting the battery's energy storage capacity and health status. A total of four characteristics are extracted, namely discharge capacity (Ah), charging capacity (Ah), discharge energy (Wh), and charging energy (Wh). These four characteristics serve as the core inputs for the subsequent trend modeling module, used to directly capture the core trend of battery capacity decay.

[0071] (3) Process monitoring features: These features are used to reflect the operating conditions and process status of the battery in each charge and discharge cycle, and help improve the prediction accuracy and robustness of the model. They mainly include temperature rise features (such as the maximum temperature rise in the charging and discharging stages), time features (such as the constant current charging duration) and cycle index. The cycle index is used to mark the degradation stage of the battery and provide a reference for the model to distinguish the feature patterns of different aging stages.

[0072] Through the above feature extraction steps, each charge-discharge cycle of each battery is transformed into an information-rich multidimensional feature vector, which includes not only deep features reflecting the internal electrochemical mechanism of the battery, but also intuitive performance indicators and operating condition features, providing a solid and comprehensive data foundation for the subsequent construction of a high-precision hybrid deep learning prediction model.

[0073] 3. Data Preprocessing

[0074] To transform the extracted discrete multidimensional feature vectors into an input format suitable for time-series deep learning models, while maximizing data utilization and eliminating data interference, this example employs the following two data preprocessing steps to ensure the data quality and suitability of the input model: (1) Feature Normalization: Due to the large differences in physical units and numerical ranges of different types of features, it is easy to interfere with model training and affect the convergence speed of the model. Therefore, this example uses the Min-Max Scaling method to process all features obtained in the feature extraction stage (except for category features such as the cycle number), linearly scaling the value of each feature to a uniform interval of [0,1]. This eliminates the dimensional differences between features and preserves the original variation law of the features, effectively accelerating the convergence speed of model training and improving the training efficiency of the model.

[0075] (2) Construction of time series dataset based on sliding window: This step is the key to achieving efficient data utilization and capturing the dynamic law of battery degradation in this example. The discrete cyclic feature sequence of a single battery is transformed into input sample-target label pairs suitable for time series deep learning models through the sliding window strategy. The specific implementation process is as follows: First, a fixed window length W and sliding step size St are set. In this example, the specific parameters are set as follows: window length W = 10 cycles, sliding step size St = 1 cycle. Second, input samples are constructed: for the entire life cycle feature sequence of a single battery, the sliding window starts from the beginning of the sequence and slides backward step by step according to the set sliding step size St. At each sliding position, the feature vectors of the W consecutive cycles covered by the window together constitute a time-series input sample X. The dimension of each input sample X is W × the number of features, which contains the state evolution history of the battery over the past W cycles and can fully reflect the short-term aging dynamics of the battery. Third, target labels are constructed: the prediction target (i.e., label) Y corresponding to each input sample X is defined as the last... The Real Remaining Lifetime (RUL) at each cycle point is calculated as: RUL = Number of cycles at the end of the battery's design life - Current cycle number, ensuring that the label accurately reflects the battery's current remaining lifetime. Finally, data augmentation is achieved through this sliding window strategy: a battery with only N cycles of test data can be transformed into a training set containing approximately N-W+1 (input-output) sample pairs (X,Y). For the 100 battery samples in this example, this method can generate a training dataset of tens of thousands of samples, which not only greatly expands the amount of data and solves the problem of scarce samples throughout the battery's lifespan, but also enables the model to effectively learn the complex mapping relationship of inferring future lifespan from a small segment of local degradation history, thereby improving the model's generalization ability.

[0076] 4. Model Training and Application

[0077] This example uses a dataset of tens of thousands of time-series sample pairs built in the preceding steps to train, validate, and test the model. The experimental process is strictly controlled to ensure the authenticity and reliability of the model performance evaluation. The specific implementation process is as follows: First, a battery-level data partitioning strategy is adopted: 80 of the 100 battery samples (and all time-series samples generated through a sliding window) are used as the model training set for training and optimizing model parameters; the remaining 20 battery samples (and all corresponding time-series samples) are used as the test set to evaluate the model's final predictive performance. This partitioning method ensures that the batteries in the test set are completely unknown during the training process, effectively avoiding data leakage issues, and can truly reflect the model's predictive performance on novel battery individuals, thus improving the reliability of model evaluation.

[0078] Secondly, model training parameter settings: During model training, mean squared error (MSE) is used as the loss function to measure the deviation between the model's predicted values ​​and the true values, guiding the optimization of model parameters; the Adam optimizer is used to update model parameters, and by adaptively adjusting the learning rate, it accelerates model convergence and avoids the model getting trapped in local optima. The loss function change curve during model training is shown below. Figure 2 As shown in (a) (the blue curve represents the training set loss and the orange curve represents the validation set loss), the trend of the curves shows that both curves decrease rapidly with the increase of training epochs and eventually converge stably. The validation set loss does not show a significant rebound phenomenon, indicating that the hybrid deep learning network structure proposed in this invention can effectively learn the battery degradation law in the data and does not produce overfitting. The model training effect is good.

[0079] Finally, model performance testing and application: After model training, the final performance evaluation was conducted on 20 completely independent test sets of cells. Evaluation metrics included the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), while comparing the model's predicted trajectory with the actual decay trajectory. Figure 2 Figure (b) shows the regression scatter plot of all predicted points (predicted capacity) and true values ​​(actual capacity) on the test set. As can be seen from the figure, all data points are closely clustered around the diagonal (y=x), showing a high degree of linear correlation, indicating that the model's predicted values ​​deviate little from the true values. The calculated coefficient of determination (R²) of the model on the entire test set reaches 0.9669, and the root mean square error (RMSE) is only 0.2935 Ah, proving that the prediction method proposed in this invention has high overall prediction accuracy and robustness.

[0080] To further demonstrate the model's effectiveness in predicting the entire lifecycle of a single battery, Figure 2(c) and Figure 2 Figure (d) shows a comparison between the actual capacity degradation trajectory and the model-predicted trajectory of two randomly selected batteries from the test set. The comparison clearly shows that the model's predicted value (red curve) can very accurately fit the downward trend of the actual value, and even successfully capture some nonlinear fluctuations and local drops in capacity degradation, demonstrating the model's precise ability to capture details of battery aging. In these two representative samples, the model's mean absolute error (MAE) is less than 0.25 Ah, indicating that the prediction method proposed in this invention can not only provide accurate single-point RUL prediction, but also reliably track and depict the degradation trajectory throughout the battery's entire life cycle. This provides strong technical support for condition-based battery maintenance and lifespan early warning, verifying the practicality and superiority of the method.

[0081] The following describes a battery life prediction system provided by an embodiment of this application. The battery life prediction system described below integrates the battery life prediction model described in any one of the above claims. The battery life prediction model includes: The data processing module is used to extract and preprocess features from historical battery cycle data and construct an input sequence, wherein the features include at least relaxation voltage features, temperature features and cycle-level overall features. The interactive encoding module employs a dual-axis attention mechanism to sequentially perform interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence, outputting an enhanced feature representation. The multi-scale fusion module includes a parallel temporal attention branch and a dynamic multi-scale convolution branch, which are used to extract and fuse long-term trend features and local fluctuation features of the enhanced feature representation to generate a comprehensive feature representation. The global prediction module is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, and then generate a global prediction result through regression mapping; The trend modeling module is used to separate the core performance index sequence from the input sequence, and perform attenuation trend prediction based on the core performance index sequence through a dual-path network structure to generate trend prediction results. The core performance index sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The hierarchical fusion module is used to dynamically weight and fuse the global prediction results and the trend prediction results through a trend fusion gating network to generate the final battery life prediction value.

[0082] Optionally, the refinement and extension functions of the model can be referred to the above description.

[0083] This application embodiment also provides a readable storage medium storing parameters of a battery life prediction model, wherein the battery life prediction model is the model defined in any of the above claims, and the battery life prediction model includes: The data processing module is used to extract and preprocess features from historical battery cycle data and construct an input sequence, wherein the features include at least relaxation voltage features, temperature features and cycle-level overall features. The interactive encoding module employs a dual-axis attention mechanism to sequentially perform interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence, outputting an enhanced feature representation. The multi-scale fusion module includes a parallel temporal attention branch and a dynamic multi-scale convolution branch, which are used to extract and fuse long-term trend features and local fluctuation features of the enhanced feature representation to generate a comprehensive feature representation. The global prediction module is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, and then generate a global prediction result through regression mapping; The trend modeling module is used to separate the core performance index sequence from the input sequence, and perform attenuation trend prediction based on the core performance index sequence through a dual-path network structure to generate trend prediction results. The core performance index sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The hierarchical fusion module is used to dynamically weight and fuse the global prediction results and the trend prediction results through a trend fusion gating network to generate the final battery life prediction value.

[0084] Optionally, the refinement and extension functions of the model can be referred to the above description.

[0085] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A battery life prediction model, characterized in that, include: The data processing module is used to extract and preprocess features from historical battery cycle data and construct an input sequence, wherein the features include at least relaxation voltage features, temperature features and cycle-level overall features. The interactive encoding module employs a dual-axis attention mechanism to sequentially perform interactive attention calculations in the feature dimension and temporal dependency encoding in the time dimension on the input sequence, outputting an enhanced feature representation. The multi-scale fusion module includes a parallel temporal attention branch and a dynamic multi-scale convolution branch, which are used to extract and fuse long-term trend features and local fluctuation features of the enhanced feature representation to generate a comprehensive feature representation. The global prediction module is used to perform feature aggregation on the comprehensive feature representation, and to perform weighted fusion of feature representations from different branches based on adaptive weights, and then generate a global prediction result through regression mapping; The trend modeling module is used to separate the core performance index sequence from the input sequence, and perform attenuation trend prediction based on the core performance index sequence through a dual-path network structure to generate trend prediction results. The core performance index sequence includes at least four core performance indicators: charging capacity, discharging capacity, charging energy, and discharging energy. The hierarchical fusion module is used to dynamically weight and fuse the global prediction results and the trend prediction results through a trend fusion gating network to generate the final battery life prediction value.

2. The model according to claim 1, characterized in that, The interactive encoding module includes: The feature interaction attention unit is used to rearrange the dimensions of the input sequence, apply a multi-head self-attention mechanism on the feature dimensions, calculate the correlation weights between different features, and generate an intermediate feature matrix containing feature interaction information. The temporal dependency encoding unit is used to apply a Transformer encoder to the intermediate feature matrix in the time dimension to capture the long-term temporal dependencies between historical cyclic data and output the enhanced feature representation.

3. The model according to claim 1, characterized in that, The temporal attention branch includes: The temporal dependency encoder, which is a time-dimensional Transformer encoder, is used to encode the enhanced feature representation obtained by the feature interaction attention module to model the long-term temporal dependency relationship of the battery cycle sequence and output the temporal dependency enhanced representation.

4. The model according to claim 3, characterized in that, The dynamic multi-scale convolution branch includes: A lightweight controller network is used to dynamically generate weight coefficients corresponding to multiple predetermined convolutional kernel sizes based on the input feature representation; The multi-scale parallel convolutional layer, consisting of multiple one-dimensional convolutional layers with different kernel sizes, is used to extract the local temporal patterns of the enhanced features in parallel and output multiple convolutional features. The weighted fusion layer is used to perform weighted summation on the multiple convolutional features based on the weight coefficients generated by the lightweight controller network to generate dynamic convolutional features.

5. The model according to claim 4, characterized in that, The multi-scale fusion module also includes a branch fusion unit; The branch fusion unit adopts an adaptive attention mechanism to dynamically calculate the first fusion weight of the temporal dependency enhancement representation and the second fusion weight of the dynamic convolutional feature, and performs weighted fusion of the temporal dependency enhancement representation and the dynamic convolutional feature based on the first fusion weight and the second fusion weight to generate the comprehensive feature representation.

6. The model according to claim 1, characterized in that, The dual-path network structure includes: A transient state modeling path is used to extract the latest time step index value from the core performance index sequence and generate a first trend feature through a first multilayer perceptron. The temporal evolution modeling path is used to perform one-dimensional convolution and pooling operations on the complete sequence of the core performance indicators to extract temporal evolution features, and generate second trend features through a second multilayer perceptron. The trend internal fusion unit is used to dynamically fuse the first trend feature and the second trend feature through a trend fusion gating network to generate the trend prediction result.

7. The model according to claim 6, characterized in that, The calculation formula corresponding to the trend prediction result generated by the trend fusion gating network is as follows: in, The fusion weights used to calculate the trend prediction results. The result of the trend fusion gating network processing time features. As the first trend feature, This is the second trend characteristic. This is the result of trend prediction.

8. The model according to claim 1, characterized in that, The data processing module constructs the input sequence using a sliding window mechanism; The sliding window mechanism is as follows: From the cyclic sequence of the entire battery life cycle, feature data from multiple consecutive cycles are extracted in sequence to form a training sample, and the battery performance index of the cycle after a specified interval in the future is used as the prediction target of the training sample. A training sample sequence with temporal correlation is generated by sliding a window of fixed length along the cyclic sequence, and used as the input sequence.

9. A battery life prediction system, characterized in that, The system integrates a battery life prediction model as described in any one of claims 1 to 8.

10. A readable storage medium storing parameters of a battery life prediction model, characterized in that, The battery life prediction model is the model defined in any one of claims 1 to 8.

Images