Lithium battery life prediction method based on combination of multilevel feature fusion and time sequence modeling

Abstract

The invention discloses a multi-level feature fusion and time sequence modeling combined lithium battery life prediction method, and relates to the technical field of lithium ion battery health management and life prediction. The method comprises the steps of receiving time sequence observation data in a battery operation process, inputting the time sequence observation data into a pre-constructed local feature extraction model, and introducing a one-dimensional convolutional neural network into the local feature extraction model to perform feature extraction on the time sequence observation data to obtain local feature representation. According to the method, three structures of local feature extraction, global context modeling and bidirectional time sequence modeling are fused, and the battery degradation modeling capability and prediction precision are effectively improved. The TFN adopts an end-to-end architecture design, has good feature perception capability and time-dependent modeling capability, and can adapt to various degradation modes and complex time sequence environments. The method is suitable for life evaluation and health state monitoring in an intelligent battery management system, and has relatively high practical value and popularization prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery health management and life prediction, and specifically to a lithium battery life prediction method combining multi-level feature fusion and time series modeling. Background Art

[0002] Lithium-ion batteries (LIBs) have been widely used in scenarios such as electric vehicles (EVs), portable electronic devices, and renewable energy systems due to their high energy density, long cycle life, and low self-discharge rate, becoming core energy storage components. However, as batteries undergo multiple charge and discharge cycles, irreversible physical and chemical reactions occur within them, such as loss of electrode materials, electrolyte decomposition, and increased internal resistance, leading to a gradual decline in performance, manifested in decreased capacity and increased safety risks. Therefore, accurately assessing the battery's state of health (SOH) and remaining useful life (RUL) is of great significance for ensuring system operation safety, optimizing battery scheduling, and extending the service life.

[0003] Most traditional battery life prediction methods are based on physical modeling, such as electrochemical models and equivalent circuit models (ECMs), which simulate the internal reaction mechanisms of batteries to predict degradation behavior. In recent years, data-driven methods have gained widespread attention because they do not require explicit modeling of the complex internal battery mechanisms. Deep learning (DL) methods, in particular, can automatically mine complex nonlinear patterns from historical monitoring data and have been widely used in battery RUL prediction tasks. For example, a multi-scale prediction method that combines variational mode decomposition (VMD) with integrated machine learning for lithium-ion battery RUL modeling can effectively reduce capacity prediction errors. Another method combines an improved particle filter (with an adaptive ant lion algorithm to optimize particle distribution) with Gaussian process regression to construct a model-data fusion mechanism. This achieves high-precision life prediction through error sequence decomposition and correction, overcoming the particle degradation and local error fluctuation issues of traditional filtering methods. Therefore, data-driven methods have been widely demonstrated to have significant advantages in characterizing the complex nonlinear evolution of battery degradation.

[0004] Although data-driven models have achieved many results in battery life prediction, there are still several key challenges: (1) Long-term dependencies are difficult to model: Battery degradation has a long-term cumulative effect. How to effectively model contextual dependencies in sequences with a long time span is still a major problem; (2) Data quality and robustness issues: Actual battery monitoring data often have noise interference, sample missing or distribution offset. How the model can maintain the stability and generalization ability of prediction on this basis needs further study. To this end, the present invention proposes a lithium battery life prediction method that combines multi-level feature fusion with time series modeling. Summary of the Invention

[0005] The purpose of the present invention is to provide a lithium battery life prediction method that combines multi-level feature fusion with time series modeling. By combining local perception, global attention and dynamic time series modeling structure, key feature extraction and fusion at different time scales are achieved, effectively improving the modeling accuracy and prediction robustness of complex life evolution patterns.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a lithium battery life prediction method combining multi-level feature fusion and time series modeling, comprising the following steps:

[0007] Receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data and obtain a local feature representation;

[0008] Construct a Transformer-based multi-head self-attention structure, input the local feature representation into the Transformer-based multi-head self-attention structure, and use the context-aware mechanism to model the semantic associations between different time steps in the sequence, thereby achieving multi-scale global fusion of local features and obtaining global fusion features.

[0009] In time series modeling, a bidirectional long short-term memory network structure is designed and the obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions.

[0010] A regression output module based on fully connected mapping is constructed, and the captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

[0011] Furthermore, the time series observation data during the battery operation process is received and input into a pre-built local feature extraction model. A one-dimensional convolutional neural network is introduced into the local feature extraction model to extract features from the time series observation data and obtain a local feature representation, as follows:

[0012] (21) Assume that the time series observation data sequence during battery operation is X = [x1, x2,…, x t ], each of which The battery capacity feature vector for the battery operation at the tth time step;

[0013] (22) The input sequence X = [x1, x2, ..., x t ] is input into a one-dimensional convolutional neural network for local feature extraction. The convolution layer uses a kkk-length sliding window to extract features from the local time region. The calculation formula is:

[0014] F local =ReLU(Conv1D(X×W conv +b conv )) (1)

[0015] in Represents local feature representation, Conv1D is a one-dimensional convolution operation, W conv and b conv are the convolution kernel and bias term respectively, and ReLU is the activation function;

[0016] (23) Downsampling is performed through the maximum pooling operation, and the maximum response value in the window is retained as the representative feature. The calculation formula is as follows:

[0017]

[0018] in Indicates representative local features.

[0019] Furthermore, we construct a Transformer-based multi-head self-attention structure, input the local feature representation into the Transformer-based multi-head self-attention structure, and use the context-aware mechanism to model the semantic associations between different time steps in the sequence, thus achieving multi-scale global fusion of local features and obtaining global fusion features, as follows:

[0020] (31) The Transformer-based multi-head self-attention structure maps local features into query, key, and value matrices:

[0021]

[0022] Where W Q ,W K ,WV is the learnable linear mapping weight;

[0023] (32) The dependency between different time steps within a time series is calculated by the scaled dot product self-attention mechanism, and the sequence representation is weighted by the key-value similarity between the time steps. The specific formula is:

[0024]

[0025] (33) The context dependencies of different time scales are captured through the multi-head attention mechanism, and the results of each head are finally concatenated and mapped into the global fusion feature F global :

[0026] F global =Concat(head1,head2,…,head h )×W O (7)

[0027] Among them, head i =Attention(Q i ,K i ,V i ) is the output of each head, h represents the number of attention heads, W O is the output mapping matrix.

[0028] Furthermore, a bidirectional long short-term memory network structure is designed in the time series modeling. The obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions. The details are as follows:

[0029] The global feature sequence F global =[f1,f2,…,f T ] is input to the bidirectional long short-term memory network, and the two LSTM sub-networks, forward and reverse, jointly model the bidirectional time dependency:

[0030] The forward process is as follows:

[0031] Forget Gate Determines how much of the current moment's information is retained from the historical information:

[0032]

[0033] Input Gate Determines whether new memory can be added to the cell state at the current moment:

[0034]

[0035] Then, the candidate cell state generates a temporary memory of the current moment

[0036]

[0037] The outputs of the forget gate and the input gate are combined to update the current cell state

[0038]

[0039] Finally, the output gate Output the feature representation of the current time step:

[0040]

[0041] Get the hidden state output of the forward process

[0042]

[0043] in, Represent the input weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. Represent the hidden state weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. are the bias terms of the forget gate, input gate, candidate memory unit, and output gate, respectively. σ(·) and tanh represent the activation function.

[0044] The reverse process is as follows:

[0045] In the same way, the features are processed backward from the end time of the sample t = T to 1 to obtain the reverse process hidden state Finally, the forward and reverse hidden state features are fused to obtain the fusion feature h t for:

[0046]

[0047] Furthermore, a regression output module based on fully connected mapping is constructed. The captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency. The details are as follows:

[0048] (51) The final time step of the bidirectional long short-term memory network is fused with the feature h T Input to the fully connected layer to generate the RUL prediction value, the calculation formula is:

[0049]

[0050] in is the predicted remaining useful life, is the weight matrix of the fully connected layer, is the bias term;

[0051] (52) With the true label y true and predicted value The difference between them is used to construct the regression loss function, and the mean square error loss is used for training optimization, which is defined as follows:

[0052]

[0053] (53) The training process uses the Adam adaptive gradient optimization algorithm, the learning rate is initialized to η, and all parameters in the network are dynamically updated to minimize as the goal.

[0054] According to a second aspect of the present invention, the present invention provides a lithium battery life prediction system combining multi-level feature fusion and time series modeling, which is used to implement the lithium battery life prediction method combining multi-level feature fusion and time series modeling, including:

[0055] The local feature extraction module is used to receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, and introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data to obtain a local feature representation;

[0056] The global feature fusion module is used to build a Transformer-based multi-head self-attention structure. The local feature representation is input into the Transformer-based multi-head self-attention structure. The semantic association between different time steps in the model sequence is modeled through a context-aware mechanism to achieve multi-scale global fusion of local features and obtain global fused features.

[0057] The time series modeling module is used to design a bidirectional long short-term memory network structure in time series modeling. The obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions;

[0058] The training output module is used to build a regression output module based on fully connected mapping. The captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

[0059] According to the third aspect of the present invention, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, the lithium battery life prediction method combining the above-mentioned multi-level feature fusion and time series modeling is adopted.

[0060] According to a fourth aspect of the present invention, the present invention provides a storage medium comprising computer executable instructions, which, when executed by a computer processor, are used to execute the lithium battery life prediction method combining multi-level feature fusion and time series modeling as described above.

[0061] The present invention has at least the following beneficial effects:

[0062] (1) The present invention integrates local modeling, global dependency modeling and time dynamic modeling capabilities through a multi-level structure, thereby achieving complete modeling of the complex degradation process of the battery, improving the response capability to key time segments and inflection points of change, and the combination of local convolution and attention mechanism enhances the robustness of feature representation. When faced with the noise, local fluctuations and degradation morphological diversity problems existing in actual sequences, the model exhibits stronger stability and learning ability. At the same time, the present invention introduces a bidirectional LSTM structure to enhance the model's memory ability for the long-term evolution pattern of the time series, effectively reducing the sensitivity to a single prediction starting point and enhancing the generalization ability of life prediction;

[0063] (2) The end-to-end deep learning framework proposed in this invention can achieve high-precision RUL prediction without physical modeling. It has the advantages of simple deployment and strong scalability. It is suitable for battery health status monitoring and intelligent prediction needs in various engineering scenarios such as electric vehicles and power storage systems.

[0064] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 Schematic diagram of the process of the prediction method of the present invention;

[0066] Figure 2 This is a schematic diagram of the structure of the prediction method of the present invention;

[0067] Figure 3 Schematic diagram of the degradation curves of the CALCE dataset and the NASA dataset in the present invention, where (a) represents the degradation curve of the CALCE dataset and (b) represents the degradation curve of the NASA dataset;

[0068] Figure 4 Schematic diagram of the prediction results of the Maryland dataset of the present invention;

[0069] Figure 5 Schematic diagram of the evaluation index radar of the Maryland dataset of the present invention, where (a) represents the MAE evaluation index, (b) represents the RMSE evaluation index, and (c) represents the R 2 Evaluation indicators;

[0070] Figure 6 Schematic diagram of the prediction results of the NASA data set of the present invention;

[0071] Figure 7 Schematic diagram of the NASA dataset evaluation index radar of the present invention, where (a) represents the MAE evaluation index, (b) represents the RMSE evaluation index, and (c) represents the R 2 Evaluation indicators. DETAILED DESCRIPTION

[0072] The following will be combined with the accompanying drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the embodiments described are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0073] See also Figures 1 to 2 The present invention provides a technical solution: a lithium battery life prediction method combining multi-level feature fusion and time series modeling, comprising the following steps:

[0074] S1. Receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data and obtain a local feature representation;

[0075] (S11) Let the input sequence be X = [x1, x2, ..., x t ], each of which The battery capacity feature vector for the battery operation at the tth time step;

[0076] (S12) The input sequence is input into a one-dimensional convolutional neural network for local feature extraction. The convolution layer uses a kkk-length sliding window to extract features from the local time region. The calculation formula is:

[0077] F local =ReLU(Conv1D(X×W conv +b conv )) (1)

[0078] in is the local feature representation, W conv and bconv They are convolution kernel and bias term respectively, ReLU is the activation function, which enhances the nonlinear expression ability;

[0079] (S13) To enhance robustness and reduce the risk of overfitting, a maximum pooling operation is introduced for downsampling, retaining the maximum response value within the window as the representative feature. The calculation formula is as follows:

[0080]

[0081] This process completes the efficient extraction and compression of local degradation features in time series signals.

[0082] Provides an input basis for subsequent global context modeling and improves the ability to represent subtle changes and local patterns;

[0083] S2. Construct a Transformer-based multi-head self-attention structure, input the local feature representation into the Transformer-based multi-head self-attention structure, and use the context-aware mechanism to model the semantic associations between different time steps in the sequence, achieving multi-scale global fusion of local features to obtain global fused features.

[0084] (S21) Let the local feature be expressed as The Transformer module first maps it into query, key, and value matrices:

[0085]

[0086] Where W Q ,W K ,W V is the learnable linear mapping weight;

[0087] (S22) The dependencies between different time steps within the time series are then calculated using the scaled dot product self-attention mechanism:

[0088] Calculate the scaled dot product attention and fuse the sequence representation weighted by the key-value similarity between time steps. The formula is:

[0089]

[0090] (S23) In order to enhance the model’s ability to model multi-layer semantic features, a multi-head attention mechanism is further introduced, and the outputs of each head are spliced ​​and projected into a unified space to obtain the global fusion feature F global :

[0091] F global=Concat(head1,head2,…,head h )×W O (7)

[0092] Among them, head i =Attention(Q i ,K i ,V i ) is the output of each head, h represents the number of attention heads, W O is the output mapping matrix;

[0093] This module can model long-term dependencies and global interactions in time series, enhance the global perception of complex degradation trends, and enable the model to extract deep semantic representations from multiple hierarchical structures, providing dynamic structured features for subsequent time series modeling.

[0094] S3. Design a bidirectional long short-term memory (LSTM) network structure for time series modeling and input the obtained global fusion features into the bidirectional LSTM network structure to capture the dynamic evolution characteristics of the battery operating state in the forward and reverse time dimensions.

[0095] The global feature sequence F global =[f1,f2,…,f T ] is input to the bidirectional long short-term memory network, and the two LSTM sub-networks, forward and reverse, jointly model the bidirectional time dependency:

[0096] The forward process is as follows:

[0097] Forget Gate Determines how much of the current moment's information is retained from the historical information:

[0098]

[0099] Input Gate Determines whether new memory can be added to the cell state at the current moment:

[0100]

[0101] Then, the candidate cell state generates a temporary memory of the current moment

[0102]

[0103] The outputs of the forget gate and the input gate are combined to update the current cell state

[0104]

[0105] Finally, the output gate Output the feature representation of the current time step:

[0106]

[0107] Get the hidden state output of the forward process

[0108]

[0109] in, Represent the input weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. Represent the hidden state weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. are the bias terms of the forget gate, input gate, candidate memory unit, and output gate, respectively. σ(·) and tanh represent the activation function.

[0110] The reverse process is as follows:

[0111] In the same way, the features are processed backward from the end time of the sample t = T to 1 to obtain the reverse process hidden state Finally, the forward and reverse hidden state features are fused to obtain the fusion feature h t for:

[0112]

[0113] This stage is through the F global Perform bidirectional modeling to capture the sequential evolution patterns of lithium batteries at different stages of their life cycle, and achieve a deep fusion of three types of information: local features, global context, and temporal modeling.

[0114] S4. Build a regression output module based on fully connected mapping. Input the captured dynamic evolution features into the fully connected regression layer of the regression output module to output the remaining battery life. Use mean squared error as the loss function and combine it with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

[0115] (S41) The hidden state h of the last time step of BiLSTM T Input to the fully connected layer to generate the remaining useful life (RUL) prediction value of the lithium battery, which is defined as:

[0116]

[0117] This layer undertakes the key conversion task from high-dimensional time series features to target value regression, ensuring that the model output is physically interpretable, that is, there is a direct mapping relationship between the predicted value and the remaining battery life.

[0118] (S42) with the true label ytrue and predicted value The regression loss function is constructed based on the difference between them, and the mean square error (MSE) loss is used for training optimization, which is defined as follows:

[0119]

[0120] (S43) The training process adopts the Adam adaptive gradient optimization algorithm, the learning rate is initialized to η, and all parameters in the network are dynamically updated to minimize For the goal:

[0121] (S44) This module implements the mapping conversion from features to predicted values, ensuring that the life prediction output has physical interpretability and numerical stability, and improving the generalization performance and convergence efficiency of the model in multiple rounds of iterative training.

[0122] It should be further explained that the optimizer selection and training mechanism configuration

[0123] To improve training efficiency and convergence stability, this embodiment uses the Adam optimizer as the parameter update strategy. Its adaptive learning rate mechanism can quickly converge and obtain a stable solution when facing complex battery degradation data. In this embodiment, the initial learning rate is set to 0.001, and the Dropout technique is introduced during the training process to suppress overfitting and enhance the generalization ability of the model.

[0124] The entire training process continuously optimizes the learnable parameters in the network including W through the back propagation algorithm. fc and b fc And the weights of each layer, ultimately achieving accurate modeling and prediction of lithium battery RUL;

[0125] Furthermore, in the lithium battery life prediction task, the performance evaluation of the model depends on the degree of difference between the predicted value and the true value. To comprehensively measure the prediction ability of the model at different starting points (SP), this embodiment adopts three classic regression evaluation indicators:

[0126] Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Coefficient of Determination (R 2 )The specific calculation formula is as follows:

[0127]

[0128] Among them, y i and Represent the true remaining life and predicted value of the i-th sample, is the average of all true values, and n is the total number of samples;

[0129] From the perspective of evaluation, the smaller the MAE and RMSE, the lower the prediction error and the stronger the model robustness; 2 The closer it is to 1, the better the model can explain the real trend of data changes and the better the prediction fitting effect. In order to more comprehensively evaluate the generalization ability of the model at different life cycle stages, this embodiment performs predictions at multiple prediction starting points (SP, that is, the number of battery cycles at the beginning of each prediction) and compares the corresponding MAE, RMSE and R 2 The average is taken as the overall performance indicator of the final model.

[0130] Next, the technical solution of the present invention is further described with reference to specific embodiments:

[0131] 1. Battery Capacity Degradation Dataset Description

[0132] In order to verify the effectiveness and adaptability of the prediction method (Time Fusion Net, TFN) of this embodiment in the actual battery degradation modeling task, this embodiment selected two public battery degradation datasets that are widely used in remaining useful life (RUL) prediction research, namely the University of Maryland Center for Advanced Life Cycle Engineering (CALCE) dataset and the National Aeronautics and Space Administration (NASA) battery dataset.

[0133] Step (11) CALCE dataset sample construction

[0134] 11.1) Data source selection and experimental subject setting;

[0135] In this example, four prismatic LiCoO2 batteries numbered CS2_35, CS2_36, CS2_37, and CS2_38 were selected from the CALCE dataset as experimental subjects. The nominal capacity of each battery was 1.1 Ah. The experiment was conducted at a constant temperature of 1°C, which meets the standard test conditions for the battery capacity degradation process.

[0136] 11.2) Charge and discharge strategy and end-of-life determination;

[0137] During the experiment, the battery was charged at a constant current of 0.55A to 4.2V, then switched to constant voltage charging until the charging current decayed to 50mA. The discharge phase was carried out at a constant current of 1.1A, and the discharge cut-off voltage was set at 2.7V. According to the CALCE dataset standard, when the battery capacity drops to 70% of the nominal capacity (i.e., 0.77Ah), it is considered to have reached the end of life (EOL) state.

[0138] 11.3) Data characteristics and application significance;

[0139] This dataset fully records the voltage, current, temperature, and capacity changes during the battery's operating cycle. It features high resolution, long-term coverage, and good stability, providing a reliable foundation for the training and validation of lithium battery life prediction models.

[0140] Step (12) NASA dataset sample construction

[0141] 12.1) Data source selection and experimental subject setting;

[0142] In this example, four 18650 cylindrical LiCoO2 batteries numbered B0005, B0006, B0007, and B0018 were selected from the NASA battery dataset as experimental objects. The nominal capacity of each battery was 2Ah, and the experiment was conducted at room temperature of 24°C.

[0143] 12.2) Charge and discharge strategy and end-of-life determination;

[0144] During the charging phase, the battery is charged at a constant current of 1.5A to 4.2V, then enters the constant voltage phase and continues until the current drops to 20mA. During the discharging phase, the battery is discharged at a constant current of 2A. The cut-off voltages set for different battery packs are 2.7V, 2.5V, and 2.2V, respectively. To unify the evaluation criteria, this embodiment uses the battery capacity dropping to 1.4Ah (i.e., 70% of the nominal capacity) as the basis for end-of-life (EOL) determination.

[0145] 12.3) Data characteristics and application significance;

[0146] The NASA dataset records in detail the voltage, current, and capacity changes during each charge and discharge cycle. With comprehensive data collection and rigorous experimental design, it serves as a benchmark data source for validating the generalization and robustness of battery degradation prediction models.

[0147] Step (13) Summary of data set parameters and experimental conditions

[0148] 13.1) Establish unified comparison standards;

[0149] To facilitate consistent processing of data from different sources, this example applies a unified end-of-life judgment standard (capacity drops to 70% of the nominal capacity) and a consistent data cleaning and preprocessing process to battery samples selected from the CALCE and NASA datasets.

[0150] 13.2) Experimental conditions comparison table;

[0151] Table 1 summarizes the charge and discharge parameters and experimental environment conditions of each battery sample in the CALCE and NASA datasets used in the experiment of this embodiment to ensure the comparability of the experimental design and the scientific nature of the results.

[0152] Table 1 Battery charge and discharge parameters and experimental conditions of NASA and CALCE datasets

[0153]

[0154] 2. Data Preprocessing and Experimental Setup

[0155] To ensure the accuracy and repeatability of the lithium battery life prediction method (Time Fusion Net, TFN) proposed in this example, which integrates multi-level feature extraction and time series modeling, in the experiment, necessary data preprocessing was performed on the original dataset before formal training and testing, and a systematic experimental setup and comparative verification process were designed.

[0156] Step (21) Data preprocessing process

[0157] 21.1) CALCE dataset preprocessing;

[0158] In the CALCE dataset, each charge-discharge cycle contains multiple steps, each corresponding to a different Step_Index number. To extract valid discharge capacity information, this example selects the discharge step with Step_Index equal to 7 in each cycle and calculates the discharge capacity of each cycle based on the Coulomb counting method. After processing, the resulting capacity change curve can truly reflect the degradation trajectory of the battery during use, with good trend and representativeness.

[0159] 21.2) NASA dataset preprocessing;

[0160] The NASA dataset has a clear structure. Each cycle consists of three stages: charging, discharging, and impedance testing. This example directly extracts the capacity value from the discharge record of each cycle and uses it as the main tracking indicator of the lithium battery health status to construct a degradation sequence.

[0161] 21.3) Visualization of capacity degradation curve;

[0162] In order to intuitively show the degradation trend of the processed data, Figure 3 The capacity degradation curves of representative battery samples in the CALCE dataset and NASA dataset are plotted. Figure 3 (a) shows the degradation curve of the CALCE dataset. Figure 3(b) shows the degradation curve of the NASA data set. As the number of cycles increases, different batteries all show a continuous capacity decline trend, which is consistent with the natural aging law of batteries;

[0163] Step (22) Experimental design and training setup

[0164] 22.1) Repeated experiments and robustness control of results;

[0165] To reduce the impact of accidental bias, this example conducted 10 independent retrainings on each model under the same hyperparameter settings, and took the average of the experimental results as the final evaluation indicator to ensure the robustness and credibility of the experimental conclusions.

[0166] 22.2) Starting point setting and cross-life cycle prediction;

[0167] In the CALCE dataset, the prediction starting point (SP) is set to the 300th cycle; in the NASA dataset, the prediction starting point is set to the 30th cycle. The starting point is selected in the middle and late degradation stage of the battery to ensure the prediction performance of the evaluation model in the decay process in the second half of the life;

[0168] 22.3) Unified parameter configuration;

[0169] To verify the comprehensive performance of the TFN model proposed in this example, all experiments were trained with a uniform number of training rounds (100 epochs) and consistent hyperparameter settings. To adapt to the characteristics of different datasets, the time step length was set to 5 in the CALCE dataset and 3 in the NASA dataset.

[0170] Step (23) Training and testing division and experimental arrangement

[0171] 23.1) Independent battery cross-sample test design;

[0172] To further verify the generalization ability of the method in this embodiment under cross-sample and cross-experimental conditions, four independent training-test split experiments were designed. The specific arrangements are shown in Table 2:

[0173] Table 2 Lithium battery degradation experiment arrangement

[0174]

[0175] 23.2) Test strategy and performance evaluation objectives;

[0176] The four sets of experiments above cover independent test scenarios of different batteries under similar operating conditions, aiming to comprehensively evaluate the adaptability, transferability, and predictive stability of the TFN model proposed in this example for unknown battery samples in practical applications.

[0177] 3. Experimental Verification on the CALCE Dataset

[0178] To verify the generalized prediction capabilities of the lithium battery life prediction method (Time Fusion Net, TFN) proposed in this example, which integrates multi-level feature extraction and time series modeling, under different degradation modes, an experimental study was conducted using two lithium-ion batteries, CS2-35 and CS2-36, from the CALCE dataset. The experimental results are presented and analyzed using capacity prediction curves, quantitative error indicators, and radar charts to comprehensively evaluate the prediction accuracy and stability of the proposed method.

[0179] Step (31) CS2-35 battery degradation prediction experiment

[0180] 31.1) Comparison of capacity prediction curves;

[0181] Figure 4 The capacity prediction results of the TFN model on CS2-35 and CS2-36 battery data are shown. Figure 4 It can be observed that TFN maintains a good fitting effect throughout the entire discharge cycle of the CS2-35 battery, especially in the rapid degradation period near EOL (approximately 630 to 680 cycles), accurately tracking the capacity decline trend and demonstrating excellent dynamic response capabilities. Compared with CS2-35, the CS2-36 has more obvious fluctuations during the degradation process, which places higher fitting requirements on the model. TFN also shows strong tracking capabilities on this dataset. In multiple small fluctuation sections, the predicted curve is highly consistent with the true curve, indicating that it has good dynamic perception and local detail restoration capabilities.

[0182] 31.2) Quantitative error index analysis;

[0183] Table 3 summarizes the main error indicators of the TFN model on CS2-35 and CS2-36 batteries, including the mean absolute error (MAE), root mean square error (RMSE) and determination coefficient (R 2 ), for easy distinction, the corresponding validation dataset is marked in the table;

[0184] Table 3 Evaluation of battery capacity prediction error of CALCE dataset

[0185]

[0186] As can be seen from Table 3, TFN achieved excellent prediction results in both experiments. For the CS2-35 data, the RMSE is 0.0086 and the R 2 The RMSE is 0.9982, indicating that the model can stably and accurately fit the capacity degradation curve. On the CS2-36 data, TFN further improves the performance, reducing the RMSE to 0.0069 and R2 It reaches 0.9992, which fully verifies the strong generalization ability and high prediction accuracy of the model under different degradation modes.

[0187] Step (32) Comprehensive performance radar chart analysis

[0188] 32.1) Performance visualization and indicator comparison;

[0189] Figure 5 The comprehensive radar chart shows the prediction results of TFN for different batteries (CS2-35 and CS2-36) on the CALCE dataset. Figure 5 (a)MAE, Figure 5 (b)RMSE, Figure 5 (c)R 2 The model performance was visually evaluated in three dimensions. The radar chart results showed that the TFN model reached the optimal or near-optimal boundary position in each indicator dimension, with excellent overall performance.

[0190] 32.2) Performance summary;

[0191] A comprehensive analysis of the capacity prediction curves and error indicator radar charts shows that the TFN model demonstrates stable and accurate fitting capabilities for both CS2-35 and CS2-36 battery degradation prediction tasks. By introducing multi-level feature extraction and dynamic time series modeling, this embodiment's method effectively balances local change perception with long-term degradation trend fitting, demonstrating leading advantages across multiple evaluation metrics and promising potential for engineering applications.

[0192] 4. NASA Dataset Experimental Verification

[0193] To further validate the generalization ability of the lithium battery life prediction method (Time Fusion Net, TFN) proposed in this example, which integrates multi-level feature extraction and time series modeling, across diverse datasets, independent tests were conducted on two battery groups, B0005 and B0006, from the NASA battery dataset. The experimental results are presented and analyzed using capacity prediction curves, quantitative error metrics, and radar plots, comprehensively evaluating the proposed method's prediction performance on real, complex degradation sequences.

[0194] Step (41) Analysis of battery degradation prediction results

[0195] 41.1) Comparison of capacity prediction curves;

[0196] Figure 6 The capacity prediction results of the TFN model on the B0005 and B0006 battery data are shown respectively. Figure 6As can be seen in the figure, the TFN model maintains a good fit throughout the degradation cycle of the B0005 battery. In particular, the predicted curve closely matches the true curve during the 110-130 cycle, a period of rapid capacity decline, demonstrating excellent trend capture and local response capabilities. Despite the frequent fluctuations and multiple mutation points in the B0006 degradation process, the TFN model is still able to accurately fit the multi-stage degradation process, maintaining good prediction stability and noise immunity at local discharge mutation points (such as around the 120th cycle).

[0197] 41.2) Quantitative error index analysis. Table 4 summarizes the main error evaluation indicators of the TFN model on B0005 and B0006 batteries, including mean absolute error (MAE), root mean square error (RMSE) and determination coefficient (R 2 ) and mark the corresponding validation dataset in the table.

[0198] Table 4 Evaluation of battery capacity prediction error of NASA dataset

[0199]

[0200] As can be seen from Table 4, TFN achieved excellent prediction performance in both battery samples. For B0005 battery, TFN performed well in all three indicators, with an RMSE of 0.0133 and R 2 The TFN still maintains an extremely low error level (RMSE is 0.0170) for the B0006 battery with stronger degradation volatility, and the R 2 It reaches 0.9911, which verifies the stability and generalization ability of the method in this embodiment when facing complex nonlinear degradation data;

[0201] Step (42) Comprehensive performance radar chart analysis

[0202] 42.1) Performance visualization and indicator comparison;

[0203] Figure 7 (a), (b), and (c) show the results based on MAE, RMSE, and R 2 The comprehensive performance radar charts drawn using the three evaluation indicators correspond to the battery data of B0005 and B0006. It can be intuitively observed from the radar charts that the TFN model has the largest envelope area for each indicator, fully demonstrating its comprehensive advantages in accuracy, robustness, and overall trend fitting ability.

[0204] 42.2) Performance summary;

[0205] By comparing the prediction results of the two sets of degradation data, B0005 and B0006, it can be found that the TFN model not only maintains high-precision predictions on regular degradation sequences, but also maintains good stability and trend tracking capabilities when facing complex battery samples with obvious volatility and irregularity. The method in this embodiment improves the adaptability of the model in different degradation scenarios through multi-level feature fusion and dynamic time series modeling mechanism, demonstrating excellent cross-sample migration capabilities and practical application potential.

[0206] In summary, the present invention addresses the problems of nonlinear degradation, complex timing dependence, and insufficient life prediction accuracy of lithium batteries during long-term operation, and constructs a hybrid deep learning framework that integrates local feature extraction, global context perception, and bidirectional time series modeling. By introducing a convolutional neural network (CNN) to extract short-term evolution features, combining a Transformer-based multi-head attention mechanism to model global associations, and utilizing a bidirectional long short-term memory network (BiLSTM) to capture the dynamic evolution of the battery aging process, the accuracy and stability of the remaining life estimation are significantly improved. The method of the present invention does not need to rely on complex physical modeling parameters or a priori battery aging models, and has end-to-end prediction capabilities. It is suitable for intelligent health management and life assessment of lithium batteries in electric vehicles, power storage systems, and renewable energy equipment.

[0207] Example 2:

[0208] This embodiment provides a lithium battery life prediction system that combines multi-level feature fusion with time series modeling, which is used to implement the lithium battery life prediction method that combines multi-level feature fusion with time series modeling, including:

[0209] The local feature extraction module is used to receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, and introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data to obtain a local feature representation;

[0210] The global feature fusion module is used to build a Transformer-based multi-head self-attention structure. The local feature representation is input into the Transformer-based multi-head self-attention structure. The semantic association between different time steps in the model sequence is modeled through a context-aware mechanism to achieve multi-scale global fusion of local features and obtain global fused features.

[0211] The timing modeling module is used to design a bidirectional long short-term memory network structure in the timing modeling layer. The obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions.

[0212] The training output module is used to build a regression output module based on fully connected mapping. The captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

[0213] Specifically, the above-mentioned local feature extraction module, global feature fusion module, timing modeling module and training output module can be embedded in a computer processing system. The computer calls the above-mentioned modules to complete the task of battery life prediction based on the lithium battery life prediction method combining multi-level feature fusion and timing modeling provided above; the above-mentioned local feature extraction module, global feature fusion module, timing modeling module and training output module can perform operations according to the specific steps given in the lithium battery life prediction method combining multi-level feature fusion and timing modeling.

[0214] It should be noted that it should be understood that the division of the various modules of the above system is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. Moreover, these modules can all be implemented in the form of software called by processing elements; they can also all be implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the program processing module can be a separately established processing element, or it can be integrated into a certain chip of the above-mentioned device. In addition, it can also be stored in the memory of the above-mentioned device in the form of program code, and called and executed by a certain processing element of the above-mentioned device to perform the functions of the above-mentioned signal processing module. The implementation of other modules is similar. In addition, these modules can all or partly be integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed by the hardware integrated logic circuit in the processor element or the instructions in the form of software.

[0215] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application-specific integrated circuits (ASICs), one or more microprocessors (DSPs), or one or more field programmable gate arrays (FPGAs). For another example, when a module is implemented by scheduling program code through a processing element, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0216] Example 3:

[0217] The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores a computer program capable of running on the processor. When the processor loads and executes the computer program, it adopts the lithium battery life prediction method combining the above-mentioned multi-level feature fusion and time series modeling.

[0218] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer or a cloud server, and the terminal device includes but is not limited to a processor and a memory. For example, the terminal device can also include input and output devices, network access devices and buses, etc.

[0219] Furthermore, the processor may adopt a central processing unit (CPU). Of course, depending on the actual usage, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. may also be adopted. The general-purpose processor may adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0220] Example 4:

[0221] This embodiment provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the lithium battery life prediction method combining multi-level feature fusion and time series modeling as described above.

[0222] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0223] For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. When an element is referred to as being "assembled on", "installed on", "fixed on" or "set on" another element, it can be directly on the other element or there can be a central element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there can be a central element at the same time. The terms "vertical", "horizontal", "up", "down", "left", "right" and similar expressions used herein are for illustrative purposes only and are not intended to be the only embodiment.

[0224] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0225] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Claims

1. A lithium battery life prediction method combining multi-level feature fusion and time series modeling, characterized by: The following steps are involved: Receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data and obtain a local feature representation; Construct a Transformer-based multi-head self-attention structure, input the local feature representation into the Transformer-based multi-head self-attention structure, and use the context-aware mechanism to model the semantic associations between different time steps in the sequence, thereby achieving multi-scale global fusion of local features and obtaining global fusion features. In time series modeling, a bidirectional long short-term memory network structure is designed and the obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions. A regression output module based on fully connected mapping is constructed, and the captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

2. The lithium battery life prediction method combining multi-level feature fusion and time series modeling according to claim 1 is characterized by: Receive the time series observation data during battery operation, input the time series observation data into the pre-built local feature extraction model, introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data, and obtain local feature representation, as follows: (21) Assume that the time series observation data sequence during battery operation is X = [x1, x2,…, x t ], each of which The battery capacity feature vector for the battery operation at the tth time step; (22) The input sequence X = [x1, x2, ..., x t ] is input into a one-dimensional convolutional neural network for local feature extraction. The convolution layer uses a kkk-length sliding window to extract features from the local time region. The calculation formula is: F local =ReLU(Conv1D(X×W conv +b conv )) (1) in Represents local feature representation, Conv1D is a one-dimensional convolution operation, W conv and b conv are the convolution kernel and bias term respectively, and ReLU is the activation function; (23) Downsampling is performed through the maximum pooling operation, and the maximum response value in the window is retained as the representative feature. The calculation formula is as follows: in Indicates representative local features.

3. The lithium battery life prediction method combining multi-level feature fusion and time series modeling according to claim 1 is characterized by: A Transformer-based multi-head self-attention structure is constructed. The local feature representation is input into the Transformer-based multi-head self-attention structure. The semantic association between different time steps in the sequence is modeled through the context-aware mechanism to achieve multi-scale global fusion of local features and obtain global fusion features. The details are as follows: (31) The Transformer-based multi-head self-attention structure maps local features into query, key, and value matrices: Where W Q ,W K ,W V is a learnable linear mapping weight; (32) The dependency between different time steps within a time series is calculated by the scaled dot product self-attention mechanism, and the sequence representation is weighted by the key-value similarity between the time steps. The specific formula is: (33) The context dependencies of different time scales are captured through the multi-head attention mechanism, and the results of each head are finally concatenated and mapped into the global fusion feature F global : F global =Concat(head1,head2,…,head h )×W O (7) Among them, head i =Attention(Q i ,K i ,V i ) is the output of each head, h represents the number of attention heads, W O is the output mapping matrix.

4. The lithium battery life prediction method combining multi-level feature fusion and time series modeling according to claim 1 is characterized by: In time series modeling, a bidirectional long short-term memory network structure is designed. The obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions. The details are as follows: The global feature sequence F global =[f1,f2,…,f T ] is input to the bidirectional long short-term memory network, and the two LSTM sub-networks, forward and reverse, jointly model the bidirectional time dependency: The forward process is as follows: Forget Gate Determines how much of the current moment's information is retained from the historical information: Input Gate Determines whether new memory can be added to the cell state at the current moment: Then, the candidate cell state generates a temporary memory of the current moment The outputs of the forget gate and the input gate are combined to update the current cell state Finally, the output gate Output the feature representation of the current time step: Get the hidden state output of the forward process in, Represent the input weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. Represent the hidden state weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. are the bias terms of the forget gate, input gate, candidate memory unit, and output gate, respectively. σ(·) and tanh represent the activation function. The reverse process is as follows: In the same way, the features are processed backward from the end time of the sample t = T to 1 to obtain the reverse process hidden state Finally, the forward and reverse hidden state features are fused to obtain the fusion feature h t for:

5. The lithium battery life prediction method combining multi-level feature fusion and time series modeling according to claim 1 is characterized by: A regression output module based on fully connected mapping is constructed. The captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency. The details are as follows: (51) The final time step of the bidirectional long short-term memory network is fused with the feature h T Input to the fully connected layer to generate the RUL prediction value, the calculation formula is: in is the predicted remaining useful life, is the weight matrix of the fully connected layer, is the bias term; (52) With the true label y true The difference between the predicted value y and the predicted value y is used to construct the regression loss function, and the mean square error loss is used for training optimization, which is defined as follows: (53) The training process uses the Adam adaptive gradient optimization algorithm, the learning rate is initialized to η, and all parameters in the network are dynamically updated to minimize as the goal.

6. A lithium battery life prediction system combining multi-level feature fusion and time series modeling, used to implement the lithium battery life prediction method combining multi-level feature fusion and time series modeling according to any one of claims 1 to 5, characterized in that: include: The local feature extraction module is used to receive time series observation data during battery operation, input the time series observation data into a pre-built local feature extraction model, and introduce a one-dimensional convolutional neural network into the local feature extraction model to extract features from the time series observation data to obtain a local feature representation; The global feature fusion module is used to build a Transformer-based multi-head self-attention structure. The local feature representation is input into the Transformer-based multi-head self-attention structure. The semantic association between different time steps in the model sequence is modeled through a context-aware mechanism to achieve multi-scale global fusion of local features and obtain global fused features. The timing modeling module is used to design a bidirectional long short-term memory network structure in the timing modeling layer. The obtained global fusion features are input into the bidirectional long short-term memory network structure to capture the dynamic evolution characteristics of the battery operating status in the forward and reverse time dimensions. The training output module is used to build a regression output module based on fully connected mapping. The captured dynamic evolution features are input into the fully connected regression layer of the regression output module to output the remaining service life of the battery. The mean square error is used as the loss function combined with the Adam optimizer for training to ensure prediction stability and convergence efficiency.

7. A terminal device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the lithium battery life prediction method combining multi-level feature fusion and time series modeling as described in any one of claims 1 to 5 is adopted.

8. A storage medium containing computer-executable instructions, characterized in that: When executed by a computer processor, the computer executable instructions are used to execute the lithium battery life prediction method combining multi-level feature fusion and time series modeling as described in any one of claims 1 to 5.

Images