Lithium battery performance prediction method and device based on deep learning, equipment and medium

Abstract

The application is suitable for the technical field of lithium batteries, and provides a lithium battery performance prediction method based on deep learning, which comprises the following steps: fine-tuning training of a trained lithium battery performance prediction model by using early cycle data of a target battery; online prediction of the performance of the target battery by using the lithium battery performance prediction model after fine-tuning training, to obtain the remaining service life and the health state of the target battery; wherein, before fine-tuning training of the lithium battery performance prediction model, a sample screening strategy is adopted to screen a preset battery sample data set, to obtain a training sample set; and the lithium battery performance prediction model constructed in advance is trained offline by using the training sample set, to obtain the trained lithium battery performance prediction model, so that the prediction accuracy and the prediction effect of the lithium battery performance prediction are improved, and the accurate prediction of the lithium battery capacity degradation trajectory is realized.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery technology, and particularly relates to a method, apparatus, device and medium for predicting the performance of lithium batteries based on deep learning. Background Technology

[0002] Lithium-ion batteries are widely used in various 3C products and are one of the most widely used mobile energy storage systems. Lithium-ion batteries release and store energy through the insertion and extraction of lithium ions between the positive and negative electrodes. However, this chemical process is not completely reversible. As the number of battery cycles increases, irreversible changes occur in the microstructure of the positive and negative electrodes, causing permanent degradation of the battery's capacity until the battery reaches its failure threshold. Accidental battery failure can not only cause permanent damage to electrical appliances but also potentially lead to safety issues. Therefore, it is essential to predict the remaining useful life (RUL) and state of health (SOH) of lithium-ion batteries in advance.

[0003] Currently, the main prediction methods for lithium batteries are divided into physicochemical modeling methods and data-driven methods. For physicochemical modeling methods, most work is based on three traditional approaches: electrochemical models, equivalent circuit models, and empirical models. These methods start from the physicochemical properties of the battery, capture model parameters based on the battery's cycling performance, and thus make predictions, exhibiting good interpretability. However, these models themselves are very complex, and determining experimental parameters often requires a large number of experiments, resulting in lower model accuracy. Data-driven methods, on the other hand, often focus only on actual battery cycling data without requiring additional experiments to determine specific parameters, saving time and efficiency, and offering higher prediction accuracy, thus gradually becoming a hot topic. Common data-driven methods mainly consist of three steps: feature creation, feature sampling, and time-series modeling. Severson et al. extracted data from the 10th and 100th cycles of the battery to create features and demonstrated a strong correlation between these features and the remaining battery life using Pearson correlation coefficients. Wang et al. used a multivariate long short-term memory (LSTM) network to capture battery degradation information, achieving better prediction results than networks such as LSTM. However, these works heavily rely on knowledge of the battery field, and the quality of feature creation directly impacts the prediction results. Furthermore, the machine learning methods used by Severson et al. often only provide end-to-end point estimates of remaining lifespan, failing to predict battery capacity degradation trajectories and thus unable to accurately grasp battery degradation information. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, device, and storage medium for predicting lithium battery performance based on deep learning, aiming to solve the problem that the existing technology cannot predict the capacity degradation trajectory of lithium batteries, resulting in poor performance prediction of lithium batteries.

[0005] On the one hand, the present invention provides a method for predicting the performance of lithium batteries based on deep learning, the method comprising the following steps:

[0006] The trained lithium battery performance prediction model was fine-tuned using early cycle data of the target battery.

[0007] The lithium battery performance prediction model, which has been fine-tuned and trained, is used to predict the performance of the target battery online, thereby obtaining the remaining lifespan and health status of the target battery.

[0008] The lithium battery performance prediction model is trained in the following way:

[0009] A sample selection strategy is used to select from the pre-set battery sample dataset to obtain the training sample set;

[0010] The pre-built lithium battery performance prediction model is trained offline using the training sample set to obtain a trained lithium battery performance prediction model.

[0011] Preferably, the lithium battery performance prediction model includes a dual-channel feature extraction module, a feature aggregation module, and a temporal processing and prediction module. The dual-channel feature extraction module includes a first channel module and a second channel module, and both the first channel module and the second channel module embed channel attention and spatial attention modules. The feature aggregation module includes a flattening layer and a splicing layer. The temporal processing and prediction module includes a long short-term memory network, two fully connected layers, and a temporal attention module.

[0012] Preferably, both the first channel module and the second channel module contain two convolutional layers, and each convolutional layer is followed by a max pooling layer. At the same time, a channel attention and spatial attention module is embedded between each convolutional layer and the max pooling layer.

[0013] Preferably, the step of offline training of the pre-built lithium battery performance prediction model using the training sample set includes:

[0014] Based on the training sample set, construct the quasi-Gram angle field matrix and degradation matrix of the battery samples in the training sample set;

[0015] The lithium battery performance prediction model is trained offline using the quasi-Gram angle field matrix and the degradation matrix.

[0016] Preferably, the step of filtering the preset battery sample dataset using a sample filtering strategy includes:

[0017] The battery sample dataset is cleaned according to a preset capacity drop threshold to obtain the remaining sample dataset;

[0018] Based on the remaining sample dataset, extract the capacity degradation matrix for all battery samples in the remaining sample dataset;

[0019] Principal component analysis was used to reduce the dimensionality of the capacity degradation matrix to obtain a capacity dimensionality-reduced matrix.

[0020] The capacity dimensionality reduction matrix is ​​clustered using an unsupervised learning algorithm to obtain the clustering results.

[0021] Based on the preset prior features and the clustering results, the remaining sample dataset is filtered to obtain the training sample set.

[0022] On the other hand, the present invention provides a lithium battery performance prediction device based on deep learning, the device comprising:

[0023] The model fine-tuning unit is used to fine-tune the trained lithium battery performance prediction model using early cycling data of the target battery.

[0024] The performance prediction unit is used to predict the performance of the target battery online using the lithium battery performance prediction model that has been fine-tuned and trained, so as to obtain the remaining service life and health status of the target battery.

[0025] The lithium battery performance prediction model is trained in the following way:

[0026] A sample selection strategy is used to select from the pre-set battery sample dataset to obtain the training sample set;

[0027] The pre-built lithium battery performance prediction model is trained offline using the training sample set to obtain a trained lithium battery performance prediction model.

[0028] Preferably, the lithium battery performance prediction model includes a dual-channel feature extraction module, a feature aggregation module, and a temporal processing and prediction module. The dual-channel feature extraction module includes a first channel module and a second channel module, and both the first channel module and the second channel module embed channel attention and spatial attention modules. The feature aggregation module includes a flattening layer and a splicing layer. The temporal processing and prediction module includes a long short-term memory network, two fully connected layers, and a temporal attention module.

[0029] Preferably, the sample screening unit includes:

[0030] The data cleaning unit is used to clean the battery sample dataset according to a preset capacity drop threshold to obtain the remaining sample dataset.

[0031] The matrix extraction unit is used to extract the capacity degradation matrix of all battery samples in the remaining sample dataset based on the remaining sample dataset.

[0032] The matrix dimensionality reduction unit is used to perform dimensionality reduction on the capacity degradation matrix using principal component analysis to obtain a capacity dimensionality reduction matrix.

[0033] The matrix clustering unit is used to perform clustering processing on the capacity dimensionality reduction matrix using an unsupervised learning algorithm to obtain the clustering results.

[0034] The sample filtering subunit is used to filter the remaining sample dataset based on preset prior features and the clustering results to obtain the training sample set.

[0035] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the steps described in the above-described deep learning-based lithium battery performance prediction method.

[0036] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the above-described deep learning-based lithium battery performance prediction method.

[0037] This invention utilizes early cycle data of the target battery to fine-tune a pre-trained lithium battery performance prediction model. The finely-tuned model is then used to predict the target battery's performance online, yielding its remaining lifespan and health status. Before fine-tuning the lithium battery performance prediction model, a sample selection strategy is employed to filter a pre-defined battery sample dataset, resulting in a training sample set. This training sample set is then used to train the pre-built lithium battery performance prediction model offline, thus improving the accuracy and effectiveness of lithium battery performance prediction and enabling accurate prediction of lithium battery capacity degradation trajectories. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the implementation of the deep learning-based lithium battery performance prediction method provided in Embodiment 1 of the present invention.

[0039] Figure 2This is a flowchart illustrating the implementation of the deep learning-based lithium battery performance prediction method provided in Embodiment 2 of the present invention.

[0040] Figure 3 This is a schematic diagram of the structure of the lithium battery performance prediction model provided in Embodiment 2 of the present invention;

[0041] Figure 4 This is a schematic diagram of the dual-channel feature extraction module in the lithium battery performance prediction model provided in Embodiment 2 of the present invention;

[0042] Figure 5 This is a schematic diagram of the structure of the lithium battery performance prediction device based on deep learning provided in Embodiment 3 of the present invention;

[0043] Figure 6 This is a schematic diagram of the structure of the electronic device provided in Embodiment 4 of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0045] The specific implementation of the present invention will be described in detail below with reference to specific embodiments:

[0046] Example 1:

[0047] Figure 1 The implementation flow of the lithium battery performance prediction method based on deep learning provided in Embodiment 1 of the present invention is illustrated. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0048] In step S101, the trained lithium battery performance prediction model is fine-tuned using early cycle data of the target battery.

[0049] This invention is applicable to electronic devices, such as personal computers and servers. In this embodiment, early cycle data of the target battery (e.g., data from the 1% capacity decay stage or data from the first 50 cycles) is acquired. This data may include charge / discharge curves, capacity decay curves, coulombic efficiency, voltage change data, etc. The early cycle data is processed according to the model structure of the lithium battery performance prediction model and the characteristics of the input data. The processed target data is then input into a pre-trained lithium battery performance prediction model to achieve model fine-tuning. Furthermore, the specific implementation details of the model structure of the lithium battery performance prediction model and the characteristics of the input data are described in Embodiment 2 below and will not be repeated here.

[0050] In step S102, the performance of the target battery is predicted online using the fine-tuned lithium battery performance prediction model to obtain the remaining lifespan and health status of the target battery.

[0051] In this embodiment of the invention, a finely tuned lithium battery performance prediction model is used to predict the subsequent capacity changes of the target battery online, obtaining the remaining useful life (RUL) and state of health (SOH) of the target battery. RUL refers to the time or number of cycles the lithium battery can continue to operate normally in its current state of health, while SOH represents the current state of the lithium battery's health, providing specific power and energy. Specifically, the calculation formula for RUL is used... α =CC EOL -CC α and RUL and SOH were calculated separately, where C actual C represents the current capacity of the target battery predicted by the lithium battery performance prediction model. nom The initial rated capacity of the target battery is given by α, where α is the current number of revolutions, and RUL is the current rated capacity. α CC represents the remaining service life of the current revolution number a. EOL To reach the total number of cycles required to reach end-of-life (EOL), CC α To reach the number of cycles completed for the current lap.

[0052] In this embodiment of the invention, the trained lithium battery performance prediction model is fine-tuned using early cycle data of the target battery. The fine-tuned lithium battery performance prediction model is then used to predict the performance of the target battery online, thereby obtaining the remaining service life and health status of the target battery. This improves the prediction accuracy and effect of lithium battery performance prediction and enables accurate prediction of the lithium battery capacity degradation trajectory.

[0053] Example 2:

[0054] Figure 2 The implementation flow of the lithium battery performance prediction method based on deep learning provided in Embodiment 2 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, and are described in detail below:

[0055] In step S201, a sample screening strategy is used to screen the preset battery sample dataset to obtain a training sample set.

[0056] This invention is applicable to electronic devices, such as personal computers and servers. In this invention, the battery sample dataset is a raw battery cycling dataset containing detailed records of commercial lithium-ion batteries cycling to failure under fast-charging conditions. Because the cycle life varies significantly between different batteries, directly using all samples to model battery degradation may incur unnecessary burden. Therefore, a suitable method is needed to select appropriate battery samples. Based on this, a sample selection strategy is employed to filter the battery sample dataset, i.e., removing or selectively retaining certain samples, in order to more accurately reflect the overall characteristics of the dataset or optimize for specific problems (such as lithium battery capacity degradation). This results in a training sample set that improves the training efficiency and generalization ability of the deep learning model.

[0057] In one feasible embodiment, an open-source dataset was used as the battery sample dataset. This dataset contained 124 commercial lithium-ion batteries that had been cycled to failure under fast-charging conditions. These batteries were manufactured by A123 Systems (specifically model APR18650M1A), using lithium phosphate (LFP) / graphite as the positive and negative electrode materials, and each battery had a nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V. These batteries were cycle-tested in a horizontal cylindrical fixture of a 48-channel Arbin LBT potentiostat at a test temperature set at 30°C. Furthermore, all batteries employed either a one-step or two-step fast-charging strategy, with the charging strategy randomly assigned. For a one-step fast charge, the battery was first charged to 80% SOC (State of Charge) at a constant current, and then charged to the cutoff voltage at 1C constant current-constant voltage (CC-CV) mode. For a two-step fast charge, the battery was first charged to an intermediate capacity Q1 at a constant current, then charged to 80% SOC at a constant current, and finally charged to the cutoff voltage at 1C. Charged to the cutoff voltage in CC-CV mode, all batteries were discharged at a constant current of 4C, with an upper limit voltage of 3.6V and a lower limit voltage of 2.0V. It is worth noting that the 124 lithium batteries in this dataset were tested in three batches, with a testing interval of up to six months. Furthermore, noise was introduced during the testing process due to factors such as power outages and equipment restarts.

[0058] In a feasible embodiment, the battery sample dataset is filtered to obtain the training sample set through the following steps:

[0059] (1) Clean the battery sample dataset according to the preset capacity drop threshold to obtain the remaining sample dataset;

[0060] In this embodiment of the invention, the original battery sample dataset contains many abnormal capacity drops. These capacity drops are caused by factors such as power outages and restarts of the cycling equipment during cycle testing. For short-term equipment anomalies, the capacity drops are small; for long-term equipment anomalies, the capacity drops are large. Since long-term equipment anomalies may cause serious deviations in the cycle performance of the battery and affect the overall prediction accuracy, the battery sample dataset is cleaned according to a preset capacity drop threshold to delete severely abnormal battery samples and obtain the remaining sample dataset.

[0061] In one feasible embodiment, the capacity drop threshold is set to 0.035Ah. Battery samples with a capacity drop less than this threshold are considered to be tolerable device noise, while battery samples with a capacity drop greater than this threshold are deleted.

[0062] (2) Based on the remaining sample dataset, extract the capacity degradation matrix of all battery samples in the remaining sample dataset;

[0063] In this embodiment of the invention, the battery capacity degradation mode refers to analyzing the battery's charge and discharge signals to understand the various characteristics and data exhibited by its performance gradually declining during long-term use, thereby enabling data screening and feature establishment. Lithium battery capacity degradation is influenced by a combination of factors, such as the ambient temperature during battery cycling, the nonlinear degradation characteristics of the battery, external factors related to battery cycling, manufacturing errors, charge and discharge strategies, and differences in the battery's components. Therefore, directly using the lithium battery capacity degradation curve is insufficient to accurately capture the degradation situation under different circumstances. Based on this, to analyze the lithium battery degradation mode, for each battery sample in the remaining sample dataset, a capacity degradation matrix is ​​constructed based on its capacity change data over time. This capacity degradation matrix reflects the trend of capacity change during battery use. Specifically, the capacity degradation matrix S is represented as... Where i represents the battery sample number, j represents the cycle number of the battery with the longest lifespan, and Qi j This represents the capacity of the i-th battery sample in the j-th cycle. Additionally, for battery samples with a cycle number less than j, the matrix is ​​filled with 0s.

[0064] (3) Principal component analysis was used to reduce the dimensionality of the capacity degradation matrix to obtain the capacity dimensionality reduction matrix;

[0065] In this embodiment of the invention, principal component analysis (PCA) is performed on the capacity degradation matrix. By selecting to retain 95% of the degradation information, the capacity degradation matrix S is dimensionality reduced to obtain a capacity dimensionality-reduced matrix. The capacity dimensionality-reduced matrix retains the main variation features in the original data while reducing the dimensionality of the data.

[0066] In a feasible embodiment, when performing dimensionality reduction on the capacity degradation matrix using principal component analysis (PCA) by selecting to retain 95% of the degradation information, specifically, firstly, the covariance matrix of the capacity degradation matrix S is calculated. This covariance matrix reflects the correlation between features and is crucial for PCA analysis. Next, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvalues ​​represent the importance (i.e., the magnitude of variance) of the corresponding eigenvectors. In PCA, larger eigenvalues ​​mean that the corresponding principal components (i.e., eigenvectors) have a larger variance contribution to the dataset, thus containing more information. These eigenvalues ​​are then arranged in descending order, i.e., eigenvalues ​​λ1>=λ2>=…>=λ m Then, starting with the largest eigenvalue, the contribution rate of each eigenvalue is accumulated sequentially until the cumulative contribution rate reaches or exceeds a specified threshold (e.g., 0.95). Where m is the dimension of the original data (i.e., the total number of eigenvalues), and k is the number of eigenvalues ​​considered during the accumulation process. When the cumulative contribution rate reaches or exceeds a specified threshold, the corresponding number of eigenvalues ​​k is the eigendimensionality that is retained. Finally, the eigenvectors corresponding to the first k eigenvalues ​​are used as the basis vectors of the new space, and the original data is projected onto this new low-dimensional space to obtain the dimensionality reduction matrix R, i.e. Where i = {1, 2, ..., the maximum number of battery samples}, j = {1, 2, ..., k}, and the maximum number of battery samples is the total number of battery samples in the remaining sample dataset.

[0067] (4) Use an unsupervised learning algorithm to cluster the capacity dimensionality reduction matrix to obtain the clustering results;

[0068] In this embodiment of the invention, an unsupervised graph clustering algorithm is used to cluster the capacity dimensionality reduction matrix to divide the battery samples in the remaining sample dataset into multiple clusters, thereby obtaining the clustering results.

[0069] In one feasible embodiment, Spectral Clustering (SC) is used to automatically identify degradation patterns in the capacity dimensionality reduction matrix, resulting in three cluster groups. These groups correspond to different degradation patterns in the long, medium, and short cycle life, respectively, and each cluster group contains battery samples with similar capacity degradation characteristics.

[0070] (5) Based on the preset prior features and clustering results, the remaining sample dataset is filtered to obtain the training sample set.

[0071] In this embodiment of the invention, although unsupervised learning spectral clustering can classify batteries with different RULs with zero errors, batteries with similar RULs do not necessarily have similar degradation patterns. To further accurately identify the degradation patterns of different batteries, prior features based on domain knowledge are introduced to evaluate the accuracy of the clustering results and assist in classification. These prior features refer to features related to battery degradation signals extracted manually using prior knowledge to identify battery degradation patterns. Here, three prior features as shown in Table 1 are used, and the Pearson correlation coefficient is used to evaluate the correlation between these three features and the battery RUL. The closer the Pearson correlation coefficient is to 1, the more positively correlated the feature is with RUL; the closer the Pearson correlation coefficient is to -1, the more negatively correlated the feature is with RUL; and the closer the Pearson correlation coefficient is to 0, the less correlated the feature is with RUL. Here, ΔQ... 100-10 (V) represents the sequence obtained by subtracting the linear interpolated voltage from the voltage-capacity curve of the battery in the 100th discharge cycle from the linear interpolated voltage from the voltage-capacity curve of the 10th discharge cycle, ΔV. 100-10 (Q = 0.85Ah) represents the difference between the voltage closest to Q = 0.85Ah in the voltage-capacity curve of the battery's 100th discharge cycle and the voltage closest to Q = 0.85Ah in the battery's 10th discharge cycle. As shown in Table 1, the three extracted prior features all have a correlation of over 90% with capacity degradation, indicating that degradation pattern recognition based on these manually extracted features has high reliability. It is worth noting that these features were obtained from the first 100 battery cycles, demonstrating significant differences in batteries under different charging strategies in the early stages of cycling, providing strong evidence for the rationality of early battery prediction.

[0072] Table 1

[0073] Prior features symbol Pearson correlation coefficient <![CDATA[Standard deviation ΔQ 100-10 (V)]]> std(ΔQ) -0.92473 <![CDATA[Average value|ΔQ 100-10 (V)|]]> mean(ΔQ) -0.90906 <![CDATA[|ΔV 100-10 (Q=0.85Ah)|]]> abs(ΔV) -0.90962

[0074] In this embodiment of the invention, when filtering the remaining sample dataset based on preset prior features and clustering results, the three prior features in Table 1 are used to further filter the clustering results of spectral clustering. For the filtered clustering results, the K-Nearest Neighbor (KNN) algorithm is used to identify batteries with similar cycle life as the training sample set for the target battery. Specifically, for each cluster group of the clustering results, battery samples in the cluster group that do not meet the thresholds corresponding to these prior features are deleted, that is, the battery samples are removed from the remaining sample dataset. For the remaining sample dataset after deletion, the KNN algorithm is used to calculate the Euclidean distance between the target battery and each battery sample in the remaining sample dataset in the feature space. Based on the distance, a specified number of nearest neighbor samples are found as training samples. These samples have higher similarity to the target battery in terms of capacity degradation, and these training samples constitute the training sample set.

[0075] By using the above steps (1)-(5), the battery sample dataset is filtered, thereby achieving efficient and accurate feature extraction by combining domain knowledge and unsupervised learning. This effectively reduces redundant data and noise signals in the training sample set, thereby improving the training effect and training speed of subsequent models.

[0076] In step S202, the pre-built lithium battery performance prediction model is trained offline using the training sample set to obtain the trained lithium battery performance prediction model.

[0077] In this embodiment of the invention, battery samples from the training sample set are used as image features input into a pre-constructed lithium battery performance prediction model. The lithium battery performance prediction model is then trained offline to learn the battery degradation mode, thereby obtaining a trained lithium battery performance prediction model.

[0078] In one feasible embodiment, such as Figure 3 As shown, the lithium battery performance prediction model (CNN-Attention-LSTM) includes a dual-channel feature extraction module, a feature aggregation module, and a temporal processing and prediction module. The dual-channel feature extraction module includes a first-channel module and a second-channel module, both of which embed channel attention and spatial attention modules. The feature aggregation module includes a flattening layer and a splicing layer. The temporal processing and prediction module includes a long short-term memory network, two fully connected layers, and a temporal attention module, thereby improving the prediction accuracy of the model.

[0079] In this embodiment of the invention, a dual-channel feature extraction module is used to extract features from the input matrix, and the extracted features are processed by a channel attention and spatial attention module (Convolutional Block Attention Module, CBAM) to obtain image features. A feature aggregation module is used to flatten and stitch together the image features output by the dual-channel feature extraction module to obtain a set of one-dimensional feature vectors. This one-dimensional feature vector represents the fusion of dual-channel features. The flattening layer is used to flatten the two-dimensional image features along the axis, and the stitching layer is used to stitch together the flattened one-dimensional vectors of the two channels to meet the input conditions of the temporal network. A temporal processing and prediction module is used to perform temporal modeling representation on the aggregated one-dimensional feature vectors. A temporal attention module is also used to improve the modeling effect and predict the final result. A Long Short-Term Memory (LSTM) network is used to perform temporal modeling on the feature vectors obtained by the feature aggregation module to obtain the temporal pattern of each original image feature. The temporal attention module is used to improve the temporal modeling effect, and a fully connected layer module is used to integrate the module for temporal output.

[0080] In yet another feasible embodiment, such as Figure 4 As shown, both the first and second channel modules contain two convolutional layers (Conv) for feature extraction, and each convolutional layer is followed by a max pooling layer (MaxPool) to prevent overfitting. Additionally, a channel attention and spatial attention module (CBAM) is embedded between each convolutional layer and the max pooling layer to accelerate training by filtering features through CBAM. The convolutional kernel size of the two Conv layers is 3*3, and the pooling window size of the two MaxPool layers is 2*2.

[0081] In another feasible embodiment, based on the training sample set, a quasi-Gram angle field matrix and a degradation matrix of the battery samples in the training sample set are constructed; the quasi-Gram angle field matrix and the degradation matrix are used to train the lithium battery performance prediction model offline, thereby achieving accurate prediction of the capacity degradation trajectory.

[0082] In this embodiment of the invention, based on the training sample set, a quasi-Gramian Angular Field (GAF) matrix and a degenerate matrix (DM) for the battery samples in the training sample set are first constructed. The GAF is used to capture the fluctuating characteristics of the battery, and the DM is used to capture the long-term degradation trend characteristics. Then, the GAF is input into the first channel module of the lithium battery performance prediction model, and the DM is input into the second channel module for offline training of the lithium battery performance prediction model. Specifically, when constructing the capacity GAF matrix, firstly, the degradation matrix C = [C1, C2, ..., C...] of each battery sample in the training sample set is... n Normalize to Then, using the polarization equation right After polar coordinate conversion, according to the equation The GAF matrix is ​​obtained, where each point of the GAF matrix represents the nonlinear relationship between the i-th cycle capacity and the j-th cycle capacity of the battery sample within an n-sliding window, t. i is the timestamp, and n is the width of the sliding window. When constructing the degradation matrix DM, the discharge capacity voltage curve of each battery sample in the training sample set is first filtered using linear interpolation so that the discharge signal is regularized into the same shape. Then, the sliding window sampling is used to obtain the signal, thereby effectively capturing the capacity degradation of the battery.

[0083] In step S203, the trained lithium battery performance prediction model is fine-tuned using early cycle data of the target battery.

[0084] In this embodiment of the invention, early cycle data of the target battery is obtained. This data may include charge-discharge curves, capacity decay curves, coulombic efficiency, voltage change data, etc. Based on the early cycle data, the corresponding GAF matrix and DM matrix are constructed. The GAF matrix and DM matrix of the target battery are input into the offline trained lithium battery performance prediction model, and the learning rate of fine-tuning training is adjusted to 0.1 times that of the offline training process to achieve fine-tuning of the model.

[0085] In step S204, the performance of the target battery is predicted online using the fine-tuned lithium battery performance prediction model to obtain the remaining lifespan and health status of the target battery.

[0086] In this embodiment of the invention, the specific implementation of step S204 can be referred to the description of step S102 in Embodiment 1, and will not be repeated here.

[0087] In this embodiment of the invention, a sample screening strategy is used to screen a preset battery sample dataset to obtain a training sample set. The pre-built lithium battery performance prediction model is then trained offline based on the training sample set to learn the battery degradation pattern. Based on the early cycle data of the target battery, the offline-trained lithium battery performance prediction model is fine-tuned. The fine-tuned lithium battery performance prediction model is then used to predict the performance of the target battery online, obtaining the remaining lifespan and health status of the target battery. Thus, by using a two-step training strategy of offline training and fine-tuning, the goal of reducing the amount of input data and balancing prediction accuracy is achieved. This overcomes the limitations of existing technologies that can only estimate the remaining number of cycles end-to-end, enabling accurate prediction of the lithium battery capacity degradation trajectory and improving the prediction accuracy and effect of lithium battery performance prediction.

[0088] Example 3:

[0089] Figure 5 The structure of the deep learning-based lithium battery performance prediction device provided in Embodiment 3 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown, including:

[0090] Model fine-tuning unit 51 is used to fine-tune the trained lithium battery performance prediction model using early cycle data of the target battery.

[0091] The performance prediction unit 52 is used to predict the performance of the target battery online using a fine-tuned lithium battery performance prediction model, and to obtain the remaining lifespan and health status of the target battery.

[0092] Preferably, for training the lithium battery performance prediction model, the deep learning-based lithium battery performance prediction device of this embodiment further includes:

[0093] The sample filtering unit is used to filter the preset battery sample dataset using a sample filtering strategy to obtain a training sample set.

[0094] The model training unit is used to train a pre-built lithium battery performance prediction model offline using a training sample set to obtain a trained lithium battery performance prediction model.

[0095] Preferably, the lithium battery performance prediction model includes a dual-channel feature extraction module, a feature aggregation module, and a time-series processing and prediction module. The dual-channel feature extraction module includes a first-channel module and a second-channel module, and both the first-channel module and the second-channel module embed channel attention and spatial attention modules. The feature aggregation module includes a flattening layer and a splicing layer. The time-series processing and prediction module includes a long short-term memory network, two fully connected layers, and a time-series attention module.

[0096] Preferably, both the first channel module and the second channel module contain two convolutional layers, and each convolutional layer is followed by a max pooling layer. At the same time, a channel attention and spatial attention module are embedded between each convolutional layer and the max pooling layer.

[0097] Preferably, the model training unit includes:

[0098] The matrix construction unit constructs the quasi-Gram angle field matrix and the degradation matrix of the battery samples in the training sample set based on the training sample set.

[0099] The model training subunit uses the quasi-Gram angle field matrix and degradation matrix to train the lithium battery performance prediction model offline.

[0100] Preferably, the sample screening unit includes:

[0101] The data cleaning unit is used to clean the battery sample dataset according to a preset capacity drop threshold to obtain the remaining sample dataset.

[0102] The matrix extraction unit is used to extract the capacity degradation matrix of all battery samples in the remaining sample dataset based on the remaining sample dataset.

[0103] The matrix dimensionality reduction unit is used to perform dimensionality reduction on the capacity degradation matrix using principal component analysis to obtain the capacity dimensionality reduction matrix.

[0104] The matrix clustering unit is used to perform clustering on a capacity-reduced dimensionality matrix using an unsupervised learning algorithm to obtain the clustering results.

[0105] The sample filtering subunit is used to filter the remaining sample dataset based on preset prior features and clustering results to obtain the training sample set.

[0106] In this embodiment of the invention, each unit of the deep learning-based lithium battery performance prediction device can be implemented by corresponding hardware or software units. Each unit can be an independent hardware or software unit, or it can be integrated into a single hardware or software unit, which is not intended to limit the invention. Specifically, the implementation methods of each unit can be referred to the description of the foregoing method embodiments, and will not be repeated here.

[0107] Example 4:

[0108] Figure 6 The structure of the electronic device provided in Embodiment 4 of the present invention is shown. For ease of explanation, only the parts related to the embodiments of the present invention are shown.

[0109] The electronic device 6 of this embodiment includes a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60. When the processor 60 executes the computer program 62, it implements the steps described in the embodiment of the deep learning-based lithium battery performance prediction method, for example... Figure 1 The steps S101 to S102 are shown. Alternatively, when the processor 60 executes the computer program 62, it implements the functions of each unit in the above-described device embodiments, for example... Figure 5 The functions of units 51 to 52 shown.

[0110] In this embodiment of the invention, the trained lithium battery performance prediction model is fine-tuned using early cycle data of the target battery. The fine-tuned lithium battery performance prediction model is then used to predict the performance of the target battery online, obtaining the remaining lifespan and health status of the target battery. Before fine-tuning the lithium battery performance prediction model, a sample screening strategy is used to screen a preset battery sample dataset to obtain a training sample set. The pre-built lithium battery performance prediction model is then trained offline using the training sample set to obtain the trained lithium battery performance prediction model.

[0111] The electronic device in this embodiment of the invention can be a personal computer or a server. The steps implemented by the processor 60 in the electronic device 6 when executing the computer program 62 to implement the deep learning-based lithium battery performance prediction method can be referred to the description of the foregoing method embodiments, and will not be repeated here.

[0112] Example 5:

[0113] In this embodiment of the invention, a computer-readable storage medium is provided, which stores a computer program. When executed by a processor, the computer program implements the steps described in the embodiment of the deep learning-based lithium battery performance prediction method. For example... Figure 1 The steps S101 to S102 are shown. Alternatively, when the computer program is executed by the processor, it implements the functions of each unit in the above-described device embodiments, for example... Figure 5 The functions of units 51 to 52 shown.

[0114] In this embodiment of the invention, the trained lithium battery performance prediction model is fine-tuned using early cycle data of the target battery. The fine-tuned lithium battery performance prediction model is then used to predict the performance of the target battery online, obtaining the remaining lifespan and health status of the target battery. Before fine-tuning the lithium battery performance prediction model, a sample screening strategy is used to screen a preset battery sample dataset to obtain a training sample set. The pre-built lithium battery performance prediction model is then trained offline using the training sample set to obtain the trained lithium battery performance prediction model.

[0115] The computer-readable storage medium in embodiments of the present invention may include any entity or device capable of carrying computer program code, a recording medium, such as ROM / RAM, disk, optical disk, flash memory, etc.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting performance of a lithium battery based on deep learning, characterized in that, The method comprises the following steps: Fine-tuning training is performed on the trained lithium battery performance prediction model using early cycle data of the target battery, the early cycle data including charge-discharge curves, capacity attenuation curves, coulomb efficiency and voltage change data, the lithium battery performance prediction model including a dual-channel feature extraction module, a feature aggregation module and a time series processing and prediction module, wherein the dual-channel feature extraction module includes a first channel module and a second channel module, and the first channel module and the second channel module both embed a channel attention and spatial attention module, the feature aggregation module includes a flattening layer and a concatenation layer, and the time series processing and prediction module includes a long short-term memory network, two fully connected layers and a time series attention module, the first channel module and the second channel module both include two convolution layers, and each convolution layer is connected with a maximum pooling layer, and a channel attention and spatial attention module is embedded between each convolution layer and the maximum pooling layer; The performance of the target battery is predicted online using the lithium battery performance prediction model after fine-tuning training, and the remaining useful life and the health state of the target battery are obtained; The lithium battery performance prediction model is trained in the following manner: The preset battery sample data set is filtered using a sample filtering strategy to obtain a training sample set, and when the preset battery sample data set is filtered using the sample filtering strategy, the following steps are included: data cleaning is performed on the battery sample data set according to a preset capacity drop threshold to obtain a remaining sample data set; a capacity degradation matrix of all battery samples in the remaining sample data set is extracted according to the remaining sample data set; the capacity degradation matrix is dimensionally reduced using principal component analysis to obtain a capacity reduced matrix; the capacity reduced matrix is clustered using an unsupervised learning algorithm to obtain a clustering result; the remaining sample data set is filtered based on a preset prior feature and the clustering result to obtain the training sample set; The training sample set is used to perform offline training on a pre-constructed lithium battery performance prediction model to obtain a trained lithium battery performance prediction model; The step of performing offline training on the pre-constructed lithium battery performance prediction model using the training sample set comprises: The quasi-Gram angular field matrix and the degradation matrix of the battery samples in the training sample set are constructed according to the training sample set; The quasi-Gram angular field matrix and the degradation matrix are used to perform offline training on the lithium battery performance prediction model; The quasi-gramian angle field matrix is constructed by polarizing each battery sample in the training sample set normalizing to obtain a normalized degradation matrix using a polarization equation polarizing polarizing according to the equation obtaining the quasi-gramian angle field matrix, wherein GAF represents the quasi-gramian angle field matrix, represents a timestamp, represents the width of the sliding window, represents the battery capacity at the nth cycle.

2. The method of claim 1, wherein, The step of performing offline training on the lithium battery performance prediction model using the quasi-Gram angular field matrix and the degradation matrix comprises: The quasi-Gram angular field matrix is input into the first channel module of the lithium battery performance prediction model, and the degradation matrix is input into the second channel module of the lithium battery performance prediction model, and offline training is performed on the lithium battery performance prediction model.

3. A lithium battery performance prediction device based on deep learning, characterized by, The device comprises: The model fine-tuning unit is configured to fine-tune a trained lithium battery performance prediction model by using early cycle data of the target battery, the early cycle data including charge-discharge curves, capacity attenuation curves, coulomb efficiency, and voltage change data, the lithium battery performance prediction model including a dual-channel feature extraction module, a feature aggregation module, and a time series processing and prediction module, wherein the dual-channel feature extraction module includes a first channel module and a second channel module, and the first channel module and the second channel module are embedded with a channel attention and spatial attention module, the feature aggregation module includes a flattening layer and a concatenation layer, and the time series processing and prediction module includes a long short-term memory network, two fully connected layers, and a time series attention module, the first channel module and the second channel module each include two convolutional layers, and each convolutional layer is connected with a maximum pooling layer, and a channel attention and spatial attention module is embedded between each convolutional layer and the maximum pooling layer. The performance prediction unit is configured to perform online prediction on the performance of the target battery by using the fine-tuned lithium battery performance prediction model to obtain the remaining useful life and the state of health of the target battery. The lithium battery performance prediction model is trained in the following manner: The sample screening strategy is used to screen the preset battery sample data set to obtain a training sample set, and when the sample screening strategy is used to screen the preset battery sample data set, the following steps are included: data cleaning is performed on the battery sample data set according to a preset capacity drop threshold to obtain a remaining sample data set; a capacity degradation matrix of all battery samples in the remaining sample data set is extracted according to the remaining sample data set; a capacity dimensionality reduction matrix is obtained by performing dimensionality reduction processing on the capacity degradation matrix by using principal component analysis; a clustering result is obtained by performing clustering processing on the capacity dimensionality reduction matrix by using an unsupervised learning algorithm; and the remaining sample data set is screened based on a preset prior feature and the clustering result to obtain the training sample set; The training sample set is used to perform offline training on a pre-constructed lithium battery performance prediction model to obtain a trained lithium battery performance prediction model, and when the training sample set is used to perform offline training on the pre-constructed lithium battery performance prediction model, the following steps are included: a quasi-gram angular field matrix and a degradation matrix of battery samples in the training sample set are constructed according to the training sample set; and the lithium battery performance prediction model is trained offline by using the quasi-gram angular field matrix and the degradation matrix; The quasi-gramian angle field matrix is constructed by polarizing the degradation matrix of each battery sample in the training sample set normalizing to obtain a normalized degradation matrix using a polarization equation polarizing polarizing according to the equation obtaining the quasi-gramian angle field matrix, wherein GAF represents the quasi-gramian angle field matrix, represents a timestamp, represents the width of the sliding window, represents the battery capacity at the nth cycle.

4. The apparatus of claim 3, wherein, The quasi-gram angular field matrix and the degradation matrix are used to perform offline training on the lithium battery performance prediction model, including: The quasi-gram angular field matrix is input into a first channel module of the lithium battery performance prediction model, and the degradation matrix is input into a second channel module of the lithium battery performance prediction model, and the lithium battery performance prediction model is trained offline.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 2.

6. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 5. The computer program, which is executed by a processor, implements the steps of the method as claimed in any of claims 1 to 2.

Images