Method and device for early life prediction of large capacity battery cell based on deep learning, and equipment

Abstract

The application provides a large-capacity battery early life prediction method and device based on deep learning, a device, equipment and a storage medium, by constructing a variety of temperature and charge-discharge conditions of large-capacity lithium ion battery full life cycle data set. Only the first N early cycle data of the battery is used to extract four types of features such as charging capacity, temperature, incremental capacity and capacity difference, and a three-dimensional feature tensor is constructed. The hybrid deep learning model containing convolutional neural network and recurrent neural network is used to extract space-time features, the physical information constraint of battery degradation dynamics is introduced into the model, the physical consistency loss function consistent with the real change rate is established, and the life prediction model is obtained by joint optimization. The method uses the degradation precursor information in the limited early cycle data and the physical prior knowledge, significantly improves the early prediction accuracy of the end point of the life of the large-capacity battery and the model generalization ability, and solves the reliability problem of the migration of the small-capacity battery model to the large-capacity scene.

Description

Technical Field

[0001] This application relates to the field of battery health management technology, specifically to a method for predicting the early lifespan of large-capacity battery cells based on deep learning. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have become core energy storage components in electric vehicles, large-scale energy storage systems, and other fields. During use, batteries undergo irreversible aging, manifested as capacity decay and increased internal resistance, ultimately reaching the end of their lifespan. Accurately predicting the remaining lifespan or end-of-life of a battery is crucial for optimizing battery management systems, developing preventative maintenance strategies, and assessing system economics and reliability. Early life prediction, in particular, which involves estimating the entire lifespan of a battery in the early stages of lifespan degradation (such as the first few dozen to hundreds of cycles), can significantly reduce testing costs, shorten development cycles, and provide early warnings of potential failure risks.

[0003] Currently, battery life prediction methods can be mainly divided into mechanistic model-based methods and data-driven methods. Mechanism-based methods (such as electrochemical models) attempt to describe battery aging based on physicochemical reaction principles. While they have the advantage of strong interpretability, the models are complex, parameter identification is difficult, and they struggle to accurately describe the nonlinear degradation behavior under complex real-world operating conditions. Data-driven methods, on the other hand, directly learn aging patterns from historical operating data, requiring no complex prior physical knowledge. In recent years, with the development of machine learning, especially deep learning technology, they have become a research hotspot.

[0004] However, existing data-driven battery life prediction technologies, especially early predictions, face the following prominent drawbacks:

[0005] First, there is a lack of high-quality training data and its limitations. Existing publicly available and widely cited battery aging datasets (such as the NASA and CALCE datasets) are mainly based on small cylindrical or pouch cells like the 18650, with capacities typically below 4Ah. The degradation behavior, temperature distribution uniformity, and manufacturing process sensitivity of these laboratory-grade small batteries differ significantly from those of commercially available large-capacity cells (such as 50Ah and above). The electrochemical reactions, thermal fields, and stress fields within large-capacity cells are more unevenly distributed, and their aging mechanisms and rates may differ. Directly applying models trained on small-capacity battery data to large-capacity scenarios often leads to decreased prediction accuracy and insufficient generalization ability. Currently, the lack of publicly available, large-capacity battery lifecycle datasets covering various operating conditions is a key bottleneck restricting the application of this technology in practical industrial settings.

[0006] Second, there are issues with insufficient information utilization and weak model generalization ability in early predictions. In the early cycling stages, battery capacity degradation is often not significant, resulting in sparse degradation feature information available for model learning. Some existing methods attempt to extract handcrafted features from early cycling data (such as voltage curve features and incremental capacity curve peaks), but this feature engineering relies on expert experience and may lose high-dimensional time-series information from the original data. Other methods use deep neural networks to automatically extract features, but with limited data (especially since large-capacity battery data is scarce), the model is prone to overfitting, the learned patterns lack clear physical meaning, and generalization ability for unseen operating conditions is poor.

[0007] Third, existing technical solutions have shortcomings in incorporating physical information. Purely data-driven deep learning models are like "black boxes," lacking physical interpretability in their predictions and potentially making predictions that violate fundamental physical laws under small sample sizes or extreme conditions. Although the concept of physical information neural networks has been proposed, how to effectively and lightweightly embed prior knowledge of battery degradation dynamics into deep learning frameworks to improve learning efficiency and prediction reliability under small data conditions, especially for early battery life predictions and particularly for large-capacity batteries, remains an unresolved challenge.

[0008] In summary, existing technologies lack a high-precision, high-generalization early lifetime prediction scheme specifically for large-capacity lithium-ion cells that can fully utilize limited early cycle data and integrate prior physical knowledge. Summary of the Invention

[0009] The main objective of this application is to provide a method for predicting the early lifespan of large-capacity battery cells based on deep learning, including the following steps:

[0010] Step S1: Construct a large-capacity lithium-ion battery full life cycle dataset, obtain complete charge and discharge cycle data of multiple lithium-ion batteries with a nominal capacity of not less than 40Ah under different ambient temperatures and different charge and discharge strategies, until the battery capacity decays to the preset failure threshold, and form a database containing battery identification, cycle number, and voltage, current, capacity, and temperature time series data for each cycle.

[0011] Step S2: For the battery to be predicted, extract the data of its first N complete charging cycles; for each cycle, extract and calculate four types of feature sequences within a preset constant voltage window: charging capacity Q, temperature T, incremental capacity dQ / dV, and the charging capacity difference ΔQ between the current cycle and the initial cycle; concatenate the four types of features in the order of the cycles into a three-dimensional feature tensor with dimensions [4, N, M], where M is the number of sampling points within the voltage window;

[0012] Step S3: Construct and train a deep learning prediction model constrained by physical information. The deep learning prediction model constrained by physical information includes a feature encoder, a temporal processor, and a prediction head. The feature encoder is used to extract spatial features from the three-dimensional feature tensor. The temporal processor is used to model the temporal dependency of the spatial features as they evolve with the cycle. The prediction head is used to output the predicted value of the battery's end-of-life point. In model training, physical information constraints based on the degradation dynamics of battery health state are introduced to construct a physical consistency loss term, which, together with the life prediction loss term, constitutes a joint loss function for model optimization.

[0013] Step S4: Input the three-dimensional feature tensor of the battery to be predicted into the trained deep learning prediction model constrained by the physical information to obtain the predicted value of the battery's lifespan termination point.

[0014] In one embodiment, the different ambient temperatures include, but are not limited to, at least two of 25°C, 45°C, and 60°C; the different charge-discharge strategies include, but are not limited to, at least one of constant current-constant voltage charge-discharge strategy and constant power charge-discharge strategy.

[0015] In one embodiment, the steps of extracting and calculating four types of feature sequences within a preset constant voltage window for each cycle specifically include:

[0016] Calculate the incremental capacity curve based on the voltage-capacity data of the charging curve;

[0017] Identify the main peaks on the incremental capacity curve and their corresponding voltage points;

[0018] Based on the main peak voltage point, the voltage window is dynamically divided into multiple sub-intervals to ensure that each main electrochemical reaction stage is independently covered;

[0019] Within each voltage sub-interval, the mean, variance, skewness, and kurtosis of the capacity values ​​are extracted as statistical features, and the first and second derivative approximations of the capacity with respect to voltage within that interval are extracted as differential features.

[0020] In one embodiment, the preset constant voltage window is 3.20V to 3.52V.

[0021] In one embodiment, extracting the charging capacity Q sequence includes:

[0022] For the charging capacity-voltage curve of each cycle, interpolation resampling and smoothing filtering are performed within the voltage window, and the processed capacity value is normalized.

[0023] In one embodiment, extracting the incremental capacity dQ / dV sequence includes:

[0024] Based on the smoothed charging capacity-voltage curve, the numerical derivative of the charging capacity with respect to voltage is calculated.

[0025] In one embodiment, the feature encoder is a convolutional neural network, and the timing processor is a long short-term memory network or a gated recurrent unit network.

[0026] In one embodiment, the convolutional neural network is composed of three alternating layers of two-dimensional convolutional layers and one layer of max pooling layer; the bidirectional long short-term memory network is composed of two stacked layers of Bi-LSTM.

[0027] In one embodiment, the introduction of physical information constraints based on battery health state degradation dynamics specifically includes:

[0028] Define the battery health state (SOH), and construct an SOH prediction sub-network within the deep learning prediction model to predict the SOH value based on the intermediate features extracted by the feature encoder and the number of iterations.

[0029] A physical function G is constructed to characterize the predicted rate of change of SOH over time, i.e. The estimate;

[0030] Define physical residuals And construct a physical consistency loss term L_phy, wherein L_phy includes at least a constraint that makes the physical residual f approach zero;

[0031] The joint loss function L_total is expressed as: L_total = λ1 * L_eol + λ2 * L_phy, where L_eol is the loss between the predicted value and the actual value at the end of the lifetime, and λ1 and λ2 are weighting coefficients.

[0032] In one embodiment, the physical consistency loss term L_phy further includes the derivative of the physical residual f with respect to time. Constraints approaching zero.

[0033] Therefore, this application has the following beneficial effects:

[0034] This application provides a deep learning-based method for predicting the early lifespan of large-capacity battery cells, comprising the following steps: constructing a full lifespan dataset of large-capacity lithium-ion batteries, acquiring complete charge-discharge cycle data of multiple lithium-ion batteries with a nominal capacity of not less than 40Ah under different ambient temperatures and different charge-discharge strategies, until the battery capacity decays to a preset failure threshold, forming a database containing battery identification, cycle number, and time-series data of voltage, current, capacity, and temperature for each cycle; for the battery to be predicted, extracting data from its first N complete charging cycles; for each cycle, extracting and calculating four types of feature sequences within a preset constant voltage window: charging capacity Q, temperature T, incremental capacity dQ / dV, and the difference in charging capacity ΔQ between the current cycle and the initial cycle; concatenating the four types of features in the order of the cycles into a three-dimensional array with dimensions [4, N, M]. A feature tensor, where M is the number of sampling points within the voltage window; a deep learning prediction model constrained by physical information is constructed and trained, the physical information-constrained deep learning prediction model including a feature encoder, a temporal processor, and a prediction head; the feature encoder is used to extract spatial features from the three-dimensional feature tensor; the temporal processor is used to model the temporal dependency of the spatial features as they evolve with the cycle; the prediction head is used to output the predicted value of the battery's end-of-life point; in model training, physical information constraints based on the degradation dynamics of battery health state are introduced, a physical consistency loss term is constructed, and together with the life prediction loss term, a joint loss function is formed for model optimization; the three-dimensional feature tensor of the battery to be predicted is input into the trained deep learning prediction model constrained by physical information to obtain the predicted value of the battery's end-of-life point.

[0035] To address the lack of full lifecycle data for large-capacity batteries in existing publicly available datasets, which leads to insufficient generalization ability of prediction models, this application first independently constructs a full lifecycle dataset for large-capacity lithium-ion batteries (capacity ≥ 40Ah) covering various temperatures and charge / discharge conditions. Based on this, using only the first N (e.g., 20) early cycle data of the battery, four types of features are extracted: charging capacity, temperature, incremental capacity, and capacity difference, constructing a three-dimensional feature tensor. Spatiotemporal features are extracted using a hybrid deep learning model incorporating convolutional neural networks and recurrent neural networks. Furthermore, the physical information constraints of battery degradation dynamics are innovatively introduced into the model, establishing a physical consistency loss function that matches the predicted capacity change rate with the actual change rate. Joint optimization yields the final lifecycle prediction model. This method fully utilizes the degradation precursor information and physical prior knowledge in the limited early cycle data, significantly improving the early prediction accuracy and model generalization ability for the lifespan termination point of large-capacity batteries, and solving the reliability problem of migrating small-capacity battery models to large-capacity scenarios. This application solves the data foundation bottleneck for model training by constructing a dedicated dataset for large-capacity batteries. The proposed hybrid deep learning model, combining three-dimensional spatiotemporal feature tensors with physical information constraints, can effectively extract degradation precursor information from a small amount of early cycle data and guide model learning using physical laws. This method significantly improves the early prediction accuracy of the lifespan termination point of large-capacity battery cells (e.g., reducing mean absolute error and root mean square error), while enhancing the model's generalization ability and robustness under different temperatures and charge / discharge strategies. It achieves high-reliability prediction with only a few dozen early cycles, greatly reducing testing costs and time. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a system flowchart of a deep learning-based method for predicting the early lifespan of large-capacity battery cells.

[0038] Figure 2 This is a charging and discharging flowchart of the CC-CV and CP strategies for predicting the early lifespan of large-capacity battery cells based on deep learning.

[0039] Figure 3 This is a curve showing the change in discharge capacity of a battery cell with the number of cycles, based on a deep learning-based method for predicting the early lifespan of large-capacity battery cells.

[0040] Figure 4It is a feature curve representation of four types of features in a deep learning-based method for predicting the early lifespan of large-capacity battery cells;

[0041] Figure 5 It is a CNN-LSTM model framework based on deep learning for predicting the early lifespan of large-capacity battery cells;

[0042] Figure 6 It is a PI-CNN-LSTM model framework based on deep learning for predicting the early lifespan of large-capacity battery cells;

[0043] Figure 7 This is the lifetime prediction result of the PI-CNN-LSTM model, a method for predicting the early lifetime of large-capacity battery cells based on deep learning. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. 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.

[0045] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0046] To address the shortcomings of existing technologies, such as weak generalization ability, low accuracy with small samples, and poor adaptability to new operating conditions due to the scarcity of large-capacity battery samples in publicly available datasets, sparse early degradation feature information, and the lack of physical interpretability of purely data-driven models, this application provides a deep learning-based method for predicting the early lifespan of large-capacity battery cells. This method first constructs a dedicated dataset covering the entire lifespan of large-capacity (≥40Ah) lithium-ion batteries at multiple temperatures (e.g., 25℃, 45℃, 60℃) and multiple charge / discharge strategies (e.g., constant current-constant voltage, constant power), fundamentally solving the data foundation and generalization source problems for model training. Based on this, the method innovatively designs a multi-dimensional spatiotemporal feature engineering scheme for early cycles. It extracts charging data from the first N cycles of the battery and systematically extracts four types of feature sequences—charging capacity, temperature, incremental capacity, and capacity difference—within a selected constant electrochemical sensitive voltage window (e.g., 3.20V-3.52V). These sequences are then constructed as a three-dimensional tensor of [feature type, cycle number, voltage sampling point], thereby preserving the spatial and temporal degradation precursor information to the greatest extent possible in limited data.

[0047] Furthermore, this application constructs a hybrid deep learning model PI-CNN-LSTM that integrates physical information constraints. This model uses a Convolutional Neural Network (CNN) as the feature encoder to extract spatial local patterns from the aforementioned three-dimensional tensor; and a Long Short-Term Memory (LSTM) network or its variant as the temporal processor to capture the long-term dependencies of features as they evolve cyclically. Its core innovation lies in introducing physical constraints based on the degradation dynamics of battery state of health (SOH) during model training: by constructing an internal SOH prediction sub-network and a physical function G, and defining the physical residual. Construct a physical consistency loss term. This loss term not only requires f to approach zero, but also further requires... Approaching zero, the prior physical knowledge of battery degradation is embedded into the model optimization objective as a soft constraint. Finally, the model is trained by combining lifetime prediction loss and physical consistency loss, so that while the model fits the data, the degradation laws it learns internally conform to basic physical principles.

[0048] This application, through a systematic approach involving the construction of a dedicated dataset, the design of multidimensional features, and physical constraint models, achieves high-precision and robust prediction of the lifespan termination point of large-capacity battery cells using only dozens of early cycle data points. It effectively overcomes challenges such as overfitting with small samples and insufficient generalization ability across operating conditions, providing a reliable technical means for early health management of large-capacity battery packs in energy storage systems and electric vehicles.

[0049] This application provides a deep learning-based method for predicting the early lifespan of large-capacity battery cells, including steps S1-S4, as described above. Figure 1 , Figure 1 This is a system flowchart of a deep learning-based method for predicting the early lifespan of large-capacity battery cells.

[0050] Step S1: Construct a large-capacity lithium-ion battery full life cycle dataset, obtain complete charge and discharge cycle data of multiple lithium-ion batteries with a nominal capacity of not less than 40Ah under different ambient temperatures and different charge and discharge strategies, until the battery capacity decays to the preset failure threshold, and form a database containing battery identification, cycle number, and voltage, current, capacity, and temperature time series data for each cycle.

[0051] Step S2: For the battery to be predicted, extract the data of its first N complete charging cycles; for each cycle, extract and calculate four types of feature sequences within a preset constant voltage window: charging capacity Q, temperature T, incremental capacity dQ / dV, and the charging capacity difference ΔQ between the current cycle and the initial cycle; concatenate the four types of features in the order of the cycles into a three-dimensional feature tensor with dimensions [4, N, M], where M is the number of sampling points within the voltage window;

[0052] Step S3: Construct and train a deep learning prediction model constrained by physical information. The deep learning prediction model constrained by physical information includes a feature encoder, a temporal processor, and a prediction head. The feature encoder is used to extract spatial features from the three-dimensional feature tensor. The temporal processor is used to model the temporal dependency of the spatial features as they evolve with the cycle. The prediction head is used to output the predicted value of the battery's end-of-life point. In model training, physical information constraints based on the degradation dynamics of battery health state are introduced to construct a physical consistency loss term, which, together with the life prediction loss term, constitutes a joint loss function for model optimization.

[0053] Step S4: Input the three-dimensional feature tensor of the battery to be predicted into the trained deep learning prediction model constrained by the physical information to obtain the predicted value of the battery's lifespan termination point.

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0055] Specifically, in this embodiment, the present application provides a method for predicting the early lifespan of large-capacity battery cells based on deep learning, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a deep learning-based method for predicting the early lifespan of large-capacity battery cells. The method includes steps S1 to S4.

[0056] Step S1: Construct a large-capacity lithium-ion battery full life cycle dataset.

[0057] This step aims to provide foundational data support for model training and validation. Specifically, multiple commercial lithium-ion batteries with a nominal capacity of no less than 40Ah (e.g., square aluminum-cased or cylindrical cells based on lithium iron phosphate / graphite systems) are acquired. These batteries are subjected to cyclic aging tests under different environmental stress conditions, with ambient temperatures covering at least 25°C (room temperature) and 45°C (high temperature), and charge / discharge strategies covering at least two typical modes: constant current-constant voltage (CC-CV) and constant power (CP). Each battery is subjected to continuous charge-discharge cycles using a high-precision battery testing system, and real-time data such as voltage, current, surface temperature, and cumulative capacity are collected throughout each cycle. Cyclic testing continues until the battery's discharge capacity decays to a preset failure threshold (e.g., 70% or 80%) of its initial capacity, at which point the number of End of Life (EOL) cycles is recorded. Finally, the complete test data of all batteries is compiled into a structured database, with each data record containing a unique battery identifier, cycle number, and the complete sensor data sequence corresponding to that cycle.

[0058] The embodiments of this application utilize four self-constructed subset datasets, comprising a total of 39 high-capacity industrial-grade graphite / lithium iron phosphate cells. The compaction densities of the positive and negative electrodes are 2.35 g / cm³ and 1.50 g / cm³, respectively, with a negative-to-positive electrode capacity ratio designed to be 1.12:1.00. A cylindrical battery with a nominal capacity of 50 Ah is assembled. The battery's operating voltage window is 2.50-3.65V. The output power corresponding to a 1C rate is 160 W, calculated based on a current of 50A and a nominal voltage of 3.2V.

[0059] The embodiments of this application employ a CT-4064T (5V / 12A) battery testing system to conduct charge-discharge cycle tests under different operating conditions. Specifically, 15 cells were cycled at 25℃ / 1C using a constant current-constant voltage (CC-CV) charge-discharge strategy, 12 cells were cycled at 45℃ / 1C (CC-CV), 5 cells were cycled at 60℃ / 1C (CC-CV), and 8 cells were cycled at 45℃ / 1C using a constant power (CP) charge-discharge strategy, labeled as 25-1CCCV, 45-1CCCV, 60-1CCCV, and 45-1CP, respectively. Detailed charge-discharge procedures for CC-CV and CP are as follows... Figure 2As shown. To ensure the reliability of the model, both the training set and the test set contain the above-mentioned subsets of data. The training set contains 29 cells, and the test set contains 10 cells. The training set contains 11 25-1CCCV cells, 9 45-1CCCV cells, 4 60-1CCCV cells, and 5 45-1CP cells, while the test set contains 4 -1CCCV cells, 2 45-1CCCV cells, 1 60-1CCCV cell, and 3 45-1CP cells.

[0060] The curve showing the change in discharge capacity of the battery cell with the number of cycles is as follows: Figure 3 As shown in (a), the end-of-life (EOL) capacity threshold for each cell is set at 37.5 Ah. Figure 3 (b) shows the distribution of EOL (Effective Online Period) as a function of initial capacity: the initial capacity of the cells ranged from 46.8 to 53.4 Ah, while the EOL ranged from 1601 to 6510 cycles. Figure 3 As shown, there is no obvious linear correlation between EOL and initial capacity, and the effects of different cycling strategies and temperature conditions on EOL do not exhibit a consistent trend. Therefore, to achieve high-precision end-of-life prediction for large-capacity battery cells, it is urgent to construct a feasible and reliable EOL prediction model through reasonable feature engineering design and algorithm selection.

[0061] Step S2: Extract early recurrent multidimensional features.

[0062] For any battery whose lifespan is to be predicted (including training batteries and newly added unknown batteries), this step extracts features containing early signs of degradation from its early life stages. First, the raw data of the first N complete charging cycles (e.g., N=20, 50, the specific value can be adjusted according to the prediction timeliness requirements) of the battery are extracted. For each cycle i (i=1,2,…,N), a constant voltage analysis window that sensitively reflects electrode phase transitions and kinetic changes is selected in the charging curve, for example, 3.20V to 3.52V. Within this window, the following feature processing and calculations are performed:

[0063] The charging capacity characteristic sequence Q_i: Obtain the capacity-voltage curve Q(V) for this cycle. Perform linear interpolation resampling on this curve within a selected voltage window to obtain capacity values ​​at M equal voltage intervals (e.g., M=100 or 128). Then, perform smoothing filtering (e.g., Savitzky-Golay filter) to suppress noise, and normalize all capacity values ​​(e.g., divide by the nominal capacity) to obtain a sequence Q_i of length M.

[0064] Temperature feature sequence T_i: The temperature-voltage curve T(V) of this cycle is acquired synchronously. Interpolation resampling is performed at the same voltage point as the capacity feature, and smoothing is applied to obtain a sequence T_i of the same length M.

[0065] Incremental capacity characteristic sequence dQ / dV_i: Based on the smoothed capacity sequence Q_i, its first-order numerical derivative with respect to voltage is calculated to obtain the incremental capacity curve dQ / dV_i, which also has a length of M.

[0066] Capacity difference feature sequence ΔQ_i: Calculate ΔQ_i = Q_i - Q_1, which is the point-by-point difference between the current cycle capacity sequence and the first cycle (initial state) capacity sequence to obtain the sequence ΔQ_i.

[0067] Finally, the above N loops are organized as follows: the four types of features in each loop are treated as "slices" and stacked in loop order to form a three-dimensional feature tensor with dimensions [4, N, M]. This tensor integrates information from different types of features (channel dimension 4), time evolution (loop dimension N), and voltage spatial distribution (sampling point dimension M). This feature calculates the capacity difference point by point based on the smoothed charging capacity curve, thus generating 107 ΔQ feature points in each loop. By comparing with the initial state, ΔQ can directly reflect the distribution characteristics of capacity decay in the voltage dimension.

[0068] After extracting the above four types of features, the input features of each battery cell can be uniformly organized into a three-dimensional tensor: [4, n_cycle, 107]; where 4 represents the number of feature types (Q, T, dQ / dV, ΔQ), n_cycle=20 represents the number of input cycles, and 107 is the number of sampling points for each feature in the voltage dimension. Therefore, each battery cell contains a total of 8560 feature dimensions. Figure 4 As shown, the distribution of characteristic curves reveals that the main changes in the electrochemical reaction are concentrated in the voltage range of 3.36-3.44V. The changes in different characteristics are most significant within this range, reflecting the dominant role of electrode reaction kinetics and material structure evolution in lifetime degradation. This provides high-information-density input features for subsequent model learning of key degradation modes.

[0069] Step S3: Construct and train a deep learning prediction model constrained by physical information.

[0070] The core of this application lies in constructing a hybrid model that integrates data-driven learning with physical law constraints. The model structure includes:

[0071] Feature encoders typically employ convolutional neural networks, whose convolutional kernels perform one-dimensional convolution operations on the voltage sampling point dimension (M), aiming to extract local spatial correlations and abstract features from the four features of each cycle.

[0072] Temporal processors typically employ recurrent neural networks, such as long short-term memory networks or gated recurrent units (GRUs). They take as input the cyclically ordered feature sequences extracted by CNNs and learn the temporal evolution patterns of battery degradation features.

[0073] The prediction head typically consists of a fully connected layer that maps the final implicit state output by the timing processor to a single battery end-of-life prediction.

[0074] The model's innovation lies in introducing physical information constraints during training. Specifically, an auxiliary battery health state (SOH) prediction sub-pathway is constructed within the model. This sub-pathway also takes intermediate features extracted by the CNN as input and outputs a predicted SOH value. Simultaneously, a learnable physical function G is defined to approximate the rate of change of SOH over time (cycles). This is achieved by constructing physical residuals. A physical consistency loss term, L_phy, is designed, for example, requiring the mean squared value of f to be as small as possible. L_phy is combined with the lifetime prediction loss L_eol (e.g., mean squared error) of the main task to form a joint loss function L_total = λ1 * L_eol + λ2 * L_phy for training. By optimizing this joint loss, while learning the statistical regularities of the data, the model's internal representation and prediction process are guided towards a direction consistent with the basic dynamics of battery degradation, thereby improving learning efficiency and generalization ability under small sample sizes. The model parameters are trained end-to-end using the dataset constructed in step S1.

[0075] Step S4: Battery life prediction.

[0076] For a new, high-capacity battery with only early cycle data, a corresponding three-dimensional feature tensor is first constructed based on its first N cycles, strictly following the method described in step S2. This tensor is then input into a convergent, stable physical information-constrained deep learning prediction model trained in step S3. After forward propagation, the prediction head outputs a scalar value, representing the predicted cycle number for the battery's end-of-life, thus completing the early-life prediction. Through the above embodiments, this application utilizes only limited early-cycle data of the battery, and through carefully designed feature engineering and a deep learning model incorporating physical priors, achieves high-precision and robust early prediction of the end-of-life of high-capacity lithium-ion cells, effectively solving the problems of data scarcity and model generalization inherent in traditional methods.

[0077] In one embodiment, the different ambient temperatures include, but are not limited to, at least two of 25°C, 45°C, and 60°C; the different charge-discharge strategies include, but are not limited to, at least one of constant current-constant voltage charge-discharge strategy and constant power charge-discharge strategy.

[0078] Specifically, in this embodiment, in order to construct a training dataset that can comprehensively characterize the aging behavior of large-capacity cells under different stress conditions, the different ambient temperatures and different charging and discharging strategies were specifically designed and combined.

[0079] Three representative temperature gradients were selected: 25℃ as the standard room temperature baseline, simulating normal operating conditions; 45℃ as a typical high-temperature stress condition, used to examine the aging mechanism and rate change under accelerated temperature; and 60℃ as a high-temperature accelerated aging condition, aiming to enhance heat-related side reactions to obtain more significant degradation data within a limited time. Two widely adopted charging and discharging strategies were selected: the constant current-constant voltage (CC-CV) strategy, as a benchmark testing method for evaluating battery performance, with a clear charging process and easy extraction of stable characteristics; and the constant power (CP) strategy, which more closely reflects the transient characteristics of battery power reception or release in actual application scenarios such as electric vehicles and energy storage systems. In the specific implementation of this embodiment, the above temperatures and strategies are cross-combined, for example, setting multiple sets of parallel test conditions such as "25℃-CC-CV", "45℃-CC-CV", "45℃-CP", and "60℃-CC-CV", thereby systematically acquiring full life-cycle data of the battery under different thermo-electric coupling stresses, ensuring that the subsequently trained prediction model can learn and generalize to diverse actual working environments.

[0080] In one embodiment, the steps of extracting and calculating four types of feature sequences within a preset constant voltage window for each cycle specifically include:

[0081] Calculate the incremental capacity curve based on the voltage-capacity data of the charging curve;

[0082] Identify the main peaks on the incremental capacity curve and their corresponding voltage points;

[0083] Based on the main peak voltage point, the voltage window is dynamically divided into multiple sub-intervals to ensure that each main electrochemical reaction stage is independently covered;

[0084] Within each voltage sub-interval, the mean, variance, skewness, and kurtosis of the capacity values ​​are extracted as statistical features, and the first and second derivative approximations of the capacity with respect to voltage within that interval are extracted as differential features.

[0085] Specifically, in this embodiment, the feature extraction step has been significantly optimized compared to the traditional fixed wide voltage window method, aiming to capture key electrochemical features during battery aging more precisely and with greater physical targeting. The specific implementation process is as follows:

[0086] First, for each charging cycle, the incremental capacity (IC) curve is calculated using numerical differentiation based on the raw voltage-capacity data. Then, a peak detection algorithm is applied to identify one or more major peaks on the IC curve and record their corresponding peak voltages (V_peak). These peak voltages typically correspond to specific phase transition reactions in the electrode material (such as the FePO4 / LiFePO4 two-phase reaction platform in lithium iron phosphate), and are the most sensitive regions reflecting changes in electrochemical state. Next, centered on each identified V_peak, multiple voltage sub-intervals are dynamically divided according to a preset voltage range (e.g., [V_peak - 0.05V, V_peak + 0.05V] for a given major peak). This dynamic division based on reaction peaks ensures that each critical electrochemical reaction stage is covered by an independent voltage window closely surrounding its characteristic points, avoiding the problem of a fixed window potentially including irrelevant voltage regions or missing critical reaction areas.

[0087] Finally, within each dynamically divided voltage sub-interval, multi-dimensional feature extraction is performed on the original capacity-voltage data: on the one hand, the statistical characteristics of the capacity values ​​within the interval are calculated, including the mean (reflecting the average energy storage level), variance (reflecting data volatility), skewness (reflecting distribution asymmetry), and kurtosis (reflecting the steepness of the distribution); on the other hand, based on the trend of capacity change within the interval, its first derivative with respect to voltage (approximately the characteristics of the IC value within the interval) and second derivative approximation (reflecting the curvature change of the IC curve) are calculated. In this way, this embodiment extracts a set of composite feature vectors containing both statistical distribution information and differential change information from each key electrochemical stage, thereby providing richer and more focused input information on the aging mechanism for subsequent deep learning models.

[0088] In one embodiment, the preset constant voltage window is 3.20V to 3.52V.

[0089] Specifically, in this embodiment, the preset constant voltage window is defined as 3.20V to 3.52V. This range is selected based on the optimized design of the unique electrochemical behavior of the target high-capacity lithium iron phosphate graphite system cell during charging. The charging voltage plateau of lithium iron phosphate batteries is mainly concentrated between 3.2V and 3.5V, which corresponds to the core electrochemical reaction of the coexistence of LiFePO4 and FePO4 phases. The starting point of 3.20V marks the beginning of the battery entering this main phase transition reaction region, while the ending point of 3.52V ensures that the end of the plateau is fully covered and the transition region to higher voltages is reached. This window accurately defines the key voltage segment that best reflects the evolution of electrode material structure, lithium-ion insertion / extraction kinetics, and related side reactions. Compared to wider or narrower voltage ranges, this window offers significant advantages: it effectively excludes the nonlinear capacity recovery region during the initial charging phase (below 3.20V) and the voltage range dominated by severe polarization and side reactions during the final charging phase (above 3.52V), thus avoiding interference from irrelevant noise signals. This allows the extracted features to be highly concentrated on the electrochemical information most relevant to the intrinsic aging of the battery. By performing uniform feature sampling and calculation within this optimized window, a high degree of comparability and consistency of feature data across different batteries and cycles is ensured, laying a reliable data foundation for subsequent model learning of stable and universal aging patterns.

[0090] In one embodiment, extracting the charging capacity Q sequence includes:

[0091] For the charging capacity-voltage curve of each cycle, interpolation resampling and smoothing filtering are performed within the voltage window, and the processed capacity value is normalized.

[0092] Specifically, in this embodiment, a systematic and rigorous data preprocessing was performed on the extraction of the charging capacity Q sequence to ensure the regularity and effectiveness of the input features. First, for the raw charging capacity-voltage data of each cycle, linear interpolation resampling was performed within a preset constant voltage window (e.g., 3.20V to 3.52V).

[0093] Specifically, M voltage points are uniformly generated within the voltage range at fixed voltage intervals (e.g., 0.003V), and the capacity value corresponding to each voltage point is calculated using a linear interpolation algorithm. This transforms the original non-uniformly sampled curve into a discrete sequence with a fixed length M and strictly aligned voltage points, eliminating deviations caused by inconsistent sampling frequencies or data point distributions. Next, the resampled capacity sequence undergoes smoothing filtering. Preferably, a Savitzky-Golay filter is used for convolutional smoothing. This filter effectively suppresses random noise while preserving the original shape and key features (such as inflection points) of the capacity curve, avoiding the loss of electrochemical information due to over-smoothing. The filter window size and polynomial order can be optimized based on the data characteristics.

[0094] Finally, the smoothed capacity sequence is normalized. Each capacity value in the sequence is divided by a baseline value, preferably the battery's nominal capacity (e.g., 50 Ah) or the average capacity of the first cycle within the voltage window. This operation aims to eliminate the scaling effect caused by differences in the initial absolute capacity values ​​between different batteries, scaling the features to a similar numerical range, thereby accelerating the training and convergence process of the subsequent deep learning model and improving the model's generalization ability to absolute capacity values. After the above three steps, a standardized, denoised, and scale-uniform charging capacity feature sequence is finally obtained.

[0095] In one embodiment, extracting the incremental capacity dQ / dV sequence includes:

[0096] Based on the smoothed charging capacity-voltage curve, the numerical derivative of the charging capacity with respect to voltage is calculated.

[0097] Specifically, in this embodiment, the extraction of the incremental capacity (dQ / dV) sequence is a key feature engineering step, aiming to capture sensitive signals reflecting electrode phase transitions and reaction kinetics during battery charging. This process is strictly based on the aforementioned smoothed charging capacity-voltage curve. Specifically, numerical differentiation is used to estimate the derivative of capacity (Q) with respect to voltage (V). For M equally spaced voltage points within a voltage window, the incremental capacity value (dQ / dV) at the j-th point is calculated using a formula. This calculation method directly reflects the rate of capacity change caused by a unit voltage change, and the peaks and inflection points on the curve are closely related to specific electrochemical reactions within the battery. By differentiating the smoothed curve rather than the original noisy data, the noise amplification effect is effectively suppressed, resulting in a smooth, physically meaningful dQ / dV sequence. This sequence, as an input feature, can explicitly provide subtle, early electrochemical characteristic changes during battery aging to subsequent deep learning models.

[0098] In one embodiment, the feature encoder is a convolutional neural network, and the timing processor is a long short-term memory network or a gated recurrent unit network.

[0099] Specifically, in this embodiment, the feature encoder and the timing processor adopt the classic "CNN+RNN" hybrid architecture in deep learning to collaboratively process the spatiotemporal two-dimensional information contained in the input three-dimensional feature tensor. The feature encoder uses a convolutional neural network (CNN), specifically configured as a one-dimensional convolutional layer (Conv1D). Its core design lies in treating the "voltage sampling point dimension (M)" of the input tensor as a sequence of "space" to be processed. The CNN uses multiple convolutional layers, with convolutional kernels of a specific size (e.g., length 3), to slide along this voltage dimension, performing local feature perception and fusion on the data of the four feature types (Q, T, dQ / dV, ΔQ) at each loop (time point) in parallel. This operation can automatically extract key local patterns reflecting the electrochemical state at different voltage points (such as reaction plateau shape, peak region features) and cross-feature type correlations, effectively compressing and abstracting the original, high-dimensional voltage spatial information into a series of more discriminative high-level spatial feature vectors.

[0100] The timing processor employs a Long Short-Term Memory (LSTM) network or its variant, a Gated Recurrent Unit (GRU) network. Its function is to receive the sequence of feature vectors, arranged in cyclic order, from the CNN's encoded output. Leveraging its sophisticated internal gating mechanism (input gate, forget gate, output gate), the LSTM / GRU can selectively memorize, forget, and transmit information, thereby specifically modeling the long-term temporal dependencies of these spatial features as the number of cycles increases. It can capture key temporal patterns such as the onset of early degradation trends and subtle decay accelerations. Finally, the hidden state of the last time step of the timing processor, converging the spatiotemporal degradation information of the entire early cyclic sequence, is passed to the subsequent prediction head for lifetime value regression. This combination of CNN and LSTM / GRU fully leverages the respective advantages of CNN in local spatial feature extraction and LSTM / GRU in long-range temporal modeling, achieving end-to-end joint learning of the complex spatiotemporal evolution patterns during early battery aging.

[0101] In one embodiment, the convolutional neural network is composed of three alternating layers of two-dimensional convolutional layers and one layer of max pooling layer; the bidirectional long short-term memory network is composed of two stacked layers of Bi-LSTM.

[0102] Specifically, in this embodiment, the internal structure of the feature encoder and the timing processor has been further clarified and optimized, forming a deeper feature extraction network.

[0103] The specific architecture of the feature encoder consists of three alternating layers of two-dimensional convolutional layers (Conv2D) and one layer of max pooling layers (MaxPooling2D). The two-dimensional convolution is the core design: the input three-dimensional feature tensor (dimensions [4, N, M]) can be reshaped or directly viewed as a two-dimensional image (height 4, width M) with "feature type - voltage sampling points," while the cyclic dimension N serves as the batch or sequence length. The first Conv2D layer performs convolution on the (feature, voltage) plane using small-sized kernels (e.g., 3x3) to extract low-level spatial correlation features; the subsequent MaxPooling2D layer downsamples, compressing data volume, improving computational efficiency, preserving the most significant feature responses, and introducing a certain degree of translation invariance. Next, a second Conv2D layer and MaxPooling2D layer are stacked to extract higher-level, more abstract spatial patterns. After processing by this encoder, the input is transformed into a series of more compact and semantically rich feature map sequences.

[0104] The temporal processor's architecture employs a deep structure consisting of two stacked layers of bidirectional long short-term memory (Bi-LSTM) networks. Bi-LSTM can process the feature sequences output by the encoder simultaneously from both forward and backward directions, thereby capturing complete temporal dependencies at each time step (recurrence point) that depend on both past information and future context (within the sequence). The first Bi-LSTM layer learns preliminary temporal patterns, and its output hidden state sequence is passed as input to the second Bi-LSTM layer, which learns more complex and longer-term degradation trend evolution based on this. This two-layer stacked Bi-LSTM structure significantly enhances the model's ability to model complex nonlinear temporal relationships in the early degradation dynamics of batteries. Finally, the final hidden state of the last Bi-LSTM layer at the end of the sequence (or after aggregation) is passed to the prediction head to complete the regression prediction of lifetime values. This combined structure, through deep convolutions for spatial refinement and deep temporal modeling through deep bidirectional recurrent networks, together constitutes a powerful spatiotemporal feature learning engine. In this embodiment, the CNN-LSTM model adopts a structure consisting of three convolutional layers. Figure 5 As shown, the convolutional kernel size is 3×1, and the number of output channels is 8, 16, and 64, respectively. The extracted features are then input into a two-layer LSTM network with a hidden layer dimension of 64. Finally, the features are mapped through two fully connected layers (64 neurons per layer) to obtain the battery RUL prediction result.

[0105] In one embodiment, the introduction of physical information constraints based on battery health state degradation dynamics specifically includes:

[0106] Define the battery health state (SOH), and construct an SOH prediction sub-network within the deep learning prediction model to predict the SOH value based on the intermediate features extracted by the feature encoder and the number of iterations.

[0107] A physical function G is constructed to characterize the predicted rate of change of SOH over time, i.e. The estimate;

[0108] Define physical residuals And construct a physical consistency loss term L_phy, wherein L_phy includes at least a constraint that makes the physical residual f approach zero;

[0109] The joint loss function L_total is expressed as: L_total = λ1 * L_eol + λ2 * L_phy, where L_eol is the loss between the predicted value and the actual value at the end of the lifetime, and λ1 and λ2 are weighting coefficients.

[0110] Specifically, in this embodiment, the introduction of physical information constraints is a key mechanism for the model to achieve high accuracy and high generalization ability, and its specific implementation process is as follows:

[0111] First, a State of Health (SOH) prediction subnetwork is defined and constructed within the model. This subnetwork is a lightweight feedforward neural network whose input consists of two concatenated parts: one part is the high-level intermediate features representing the current cycle's electrochemical and thermal states extracted by the feature encoder (such as a CNN); the other part is the current cycle number (or normalized timestamp). The output of this subnetwork is a scalar, namely the predicted state of health (SOH) value of the current cycle, typically defined as the percentage of the current normalized charge capacity relative to the initial value. This subnetwork shares the feature encoder with the main network, enabling the reuse and collaborative learning of feature representations.

[0112] Secondly, a physical function G is constructed. This function aims to derive the pattern of SOH variation from the learned features. The physical function G is also a differentiable parameterized function (e.g., a small neural network). It takes the output (predicted SOH value) of the SOH prediction subnetwork and its internal features, as well as the cycle number, as input. Its output is the rate of change of SOH over time (cycles), i.e. The estimated value of the function G is obtained. The learning objective of the function G is to simulate the intrinsic dynamics of battery degradation as accurately as possible.

[0113] Next, the physical residuals are defined and the physical consistency loss is constructed. Based on the predicted SOH value, its automatic derivative with respect to the number of iterations can be calculated. (Predicted SOH) / Define the physical residual as: Based on the physical principles of battery degradation, an ideal, kinetic-compliant model should have a highly consistent predicted SOH rate of change (characterized by G) with the rate of change calculated from the predicted SOH sequence itself; that is, the residual f should approach zero. Therefore, a physical consistency loss term L_phy is constructed.

[0114] Finally, the joint loss function is constructed and optimized. The overall training objective of the model is driven by both the lifetime prediction task and the physical consistency task. During training, all parameters of the main network, the SOH prediction sub-network, and the physical function G are simultaneously optimized using the backpropagation algorithm. This ensures that while the model accurately predicts the end of the lifetime, the degradation dynamics learned internally also satisfy basic physical consistency. This constraint effectively utilizes domain prior knowledge, guiding the model to learn more generalizable and interpretable degradation patterns from limited early data.

[0115] In one embodiment, the physical consistency loss term L_phy further includes the derivative of the physical residual f with respect to time. Constraints approaching zero.

[0116] Specifically, in this embodiment, to further strengthen the strictness of physical constraints and improve the smoothness and continuity of the degradation dynamics learned by the model, the physical consistency loss term L_phy has been significantly enhanced: In addition to requiring the physical residual f itself to approach zero (first-order consistency), a term has been added that makes the derivative of the physical residual f with respect to time (number of iterations). It also approaches zero strong constraints.

[0117] This enhancement has profound physical and engineering value. Residual The consistency between the predicted rate of change (G) within the model and the actual rate of change calculated based on the predicted SOH was measured. If only f is required to approach zero, instantaneous matching of the first derivative at discrete cycle points can be guaranteed. However, introducing [a certain condition]... The constraint implies that the rate of change of the residual f over time must also be as small as possible. This essentially forces the predicted rate of change function G of SOH within the model to be not only numerically similar to... They should be equal, and their changing trends (i.e., the acceleration or curvature of the dynamic process) should also be consistent with the second derivative values ​​derived from the predicted SOH sequence, that is, satisfying higher-order dynamic self-consistency.

[0118] In this embodiment, the updated and optimized joint loss function is defined as:

[0119] ,

[0120] , ,

[0121] , .

[0122] in, , , , and These are the weighting coefficients; This is a penalty term used to constrain the weighted noise in the output. and To control the parameters of the two weights. In terms of engineering effects, this enhancement effectively avoids the model learning a degradation trajectory that, while barely matching on the first derivative, oscillates violently or is unstable and physically unreasonable in the time dimension. It guides the model to learn a smoother, more continuous, and physically more reliable SOH degradation curve, thereby improving the model's ability to reliably extrapolate between early, limited data points and enhancing the stability and robustness of the final EOL prediction. This constraint on the consistency of second-order dynamics is a key manifestation of the deep embedding of domain knowledge into the data-driven model in this application. Figure 6 As shown in the figure, this illustrates a PI-CNN-LSTM model framework for battery health management that integrates data-driven approaches and physical knowledge. Taking an electric vehicle battery as an example, the CNN analyzes the voltage curve shape of a single charge-discharge cycle, while the LSTM learns the temporal trend of long-term capacity decay, providing a preliminary prediction of the current state of health (SOH) and remaining cycle life (RUL). The core innovation lies in the physical constraints introduced on the right: a chemical kinetic model of battery capacity decay (such as the temperature-dependent Arrhenius equation) is used as a surrogate model for the neural network to validate the preliminary predictions. During training, the total loss function simultaneously optimizes prediction accuracy and physical consistency, ensuring that the SOH decline trajectory conforms to scientific principles. This allows the model to provide reliable and interpretable lifetime predictions even with limited data, effectively avoiding absurd results that might occur with pure data models. This application uses four evaluation metrics to assess model performance: MAE, MAPE, RMSE, and RUL. 2 MAE measures the average magnitude of absolute error, reflecting overall prediction bias; MAPE expresses prediction error as a percentage and is a dimensionless accuracy indicator; RMSE imposes a stronger penalty on larger errors, thus being more sensitive to outliers; R 2 This value is used to quantify the extent to which the model explains the variance of the actual data; the closer it is to 1, the better the model's predictive performance. The lifetime prediction results of the final physically constrained deep learning prediction model PI-CNN-LSTM in this application are as follows: Figure 7As shown, the predicted MAE, MAPE, RMSE, and R2 were 163 cycles, 5.33%, 234 cycles, and 0.94, respectively. This indicates that, under conditions of scarce data, compared to the experimental results using the CNN-LSTM model (MAE, MAPE, RMSE, and R2 were 225 cycles, 7.30%, 296 cycles, and 0.78, respectively), the information embedding with physical constraints enables the model to achieve higher prediction accuracy while maintaining strong generalization ability.

[0123] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system 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 system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0124] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0125] It should be particularly noted that, through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, or of course, by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the early lifespan of large-capacity battery cells based on deep learning, characterized in that, Includes the following steps: Step S1: Construct a large-capacity lithium-ion battery full life cycle dataset, obtain complete charge and discharge cycle data of multiple lithium-ion batteries with a nominal capacity of not less than 40Ah under different ambient temperatures and different charge and discharge strategies, until the battery capacity decays to the preset failure threshold, and form a database containing battery identification, cycle number, and voltage, current, capacity, and temperature time series data for each cycle. Step S2: For the battery to be predicted, extract the data of its first N complete charging cycles; for each cycle, extract and calculate four types of feature sequences within a preset constant voltage window: charging capacity Q, temperature T, incremental capacity dQ / dV, and the charging capacity difference ΔQ between the current cycle and the initial cycle; concatenate the four types of features in the order of the cycles into a three-dimensional feature tensor with dimensions [4, N, M], where M is the number of sampling points within the voltage window; Step S3: Construct and train a deep learning prediction model constrained by physical information. The deep learning prediction model constrained by physical information includes a feature encoder, a temporal processor, and a prediction head. The feature encoder is used to extract spatial features from the three-dimensional feature tensor. The timing processor is used to model the temporal dependency of the spatial features as they evolve with the cycle; the prediction head is used to output the predicted value of the battery's end-of-life point; in the model training, physical information constraints based on the dynamics of battery health state degradation are introduced to construct a physical consistency loss term, which together with the life prediction loss term constitutes a joint loss function for model optimization. Step S4: Input the three-dimensional feature tensor of the battery to be predicted into the trained deep learning prediction model constrained by the physical information to obtain the predicted value of the battery's lifespan termination point.

2. The method according to claim 1, characterized in that, In step S1, the different ambient temperatures include, but are not limited to, at least two of 25°C, 45°C, and 60°C; the different charging and discharging strategies include, but are not limited to, at least one of constant current-constant voltage charging and discharging strategy and constant power charging and discharging strategy.

3. The method according to claim 1, characterized in that, In step S2, for each cycle, within a preset constant voltage window, the steps of extracting and calculating four types of feature sequences specifically include: Calculate the incremental capacity curve based on the voltage-capacity data of the charging curve; Identify the main peaks on the incremental capacity curve and their corresponding voltage points; Based on the voltage point corresponding to the main peak, the voltage window is dynamically divided into multiple sub-intervals to ensure that each main electrochemical reaction stage is independently covered; Within each voltage sub-interval, the mean, variance, skewness, and kurtosis of the capacity values ​​are extracted as statistical features, and the first and second derivative approximations of the capacity with respect to voltage within that interval are extracted as differential features.

4. The method according to claim 1, characterized in that, In step S2, the preset constant voltage window is 3.20V to 3.52V.

5. The method according to claim 1, characterized in that, In step S2, extracting the charging capacity Q sequence includes: For the charging capacity-voltage curve of each cycle, interpolation resampling and smoothing filtering are performed within the voltage window, and the processed capacity value is normalized.

6. The method according to claim 1, characterized in that, In step S2, extracting the incremental capacity dQ / dV sequence includes: Based on the smoothed charging capacity-voltage curve, the numerical derivative of the charging capacity with respect to voltage is calculated.

7. The method according to claim 1, characterized in that, In step S3, the feature encoder is a convolutional neural network, and the timing processor is a long short-term memory network or a gated recurrent unit network.

8. The method according to claim 7, characterized in that, The convolutional neural network consists of three alternating layers of two-dimensional convolutional layers and one layer of max pooling; the long short-term memory network consists of two stacked layers of Bi-LSTM.

9. The method according to claim 1, characterized in that, In step S3, the introduction of physical information constraints based on the dynamics of battery health state degradation specifically includes: Define the battery health state (SOH); construct an SOH prediction sub-network within the deep learning prediction model to predict the SOH value based on the intermediate features extracted by the feature encoder and the number of iterations. Construct a physical function G to characterize the predicted rate of change of SOH over time, i.e., the estimate of ∂SOH / ∂t; Define the physical residual f = G-∂SOH / ∂t, and construct the physical consistency loss term L_phy, wherein L_phy includes at least the constraints that make the physical residual f approach zero; The joint loss function L_total is expressed as: L_total = λ1 * L_eol + λ2 * L_phy, where L_eol is the loss between the predicted value and the actual value at the end of the lifetime, and λ1 and λ2 are weighting coefficients.

10. The method according to claim 9, characterized in that, The physical consistency loss term L_phy also includes a constraint that makes the derivative of the physical residual f with respect to time, ∂f / ∂t, approach zero.

Images