Lithium battery health state assessment method and system based on deep network

Abstract

The invention provides a lithium battery health state assessment method and system based on a deep network in the technical field of lithium battery health state assessment, and the method comprises the steps: S1, building a health state assessment model based on a dual-time-sequence branch input module, a feature extraction module, a fusion and attention module, a time sequence modeling module and a prediction output module, setting a loss function; s2, acquiring a large amount of historical lithium battery monitoring data containing a cell-level time sequence signal and a packet-level time sequence signal to construct a data set; s3, training the health state evaluation model through the data set and the loss function, and deploying the trained health state evaluation model; and S4, collecting real-time lithium battery monitoring data, preprocessing the real-time lithium battery monitoring data, and inputting the preprocessed real-time lithium battery monitoring data into the deployed health state evaluation model to obtain a health state evaluation report. The method has the advantages that the accuracy, the real-time performance and the generalization ability of lithium battery health state evaluation are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery state of health evaluation, and particularly discloses a lithium battery state of health evaluation method and system based on a deep network. BACKGROUND

[0002] As the core infrastructure of the urban green travel system, the battery swap cabinet effectively alleviates the "range anxiety" problem commonly faced by electric bicycle users (especially instant delivery riders). At the same time, through the implementation of centralized charging management of lithium batteries, the facility can largely avoid fire safety hazards caused by private wire drawing, indoor illegal charging and other behaviors, significantly improving the safety of the charging process.

[0003] In the actual operation and maintenance process of the battery swap cabinet, accurate evaluation of the state of health (SOH) of the lithium battery is crucial in order to timely find and replace battery monomers or modules with deteriorating performance. At present, for the evaluation of the SOH of the lithium battery, the mainstream method is usually based on an equivalent circuit model or a preset empirical rule. Such a method mainly uses the charge and discharge voltage, current and temperature curves of the battery pack to extract single or limited feature parameters, and then estimates the SOH through Kalman filtering or linear regression algorithms. However, the existing technology has the following obvious defects:

[0004] 1. Lack of real-time performance: the equivalent circuit model or complex numerical integration and filtering algorithm has a large amount of calculation and takes a long time, which makes it difficult to meet the requirements of rapid screening of the battery at the second level in the battery swap cabinet scenario.

[0005] 2. Model dependency and poor generalization ability: the model parameters usually need to be carefully calibrated for a specific type of battery, and lack sufficient generalization ability for individual differences in battery aging paths and unknown aging modes.

[0006] 3. Insufficient use of high-dimensional time series information: the existing method mainly relies on the total voltage, current and temperature signals of the battery pack, and fails to fully utilize the subtle voltage / temperature difference information of the single battery cell, resulting in inaccurate SOH prediction due to the influence of the intra-group balancing effect.

[0007] 4. Difficulty in handling label missing problem: real and reliable battery SOH label data is extremely scarce in practical applications; although the service life, mileage or cycle number of the battery can be used as a reference, such data cannot be directly used as the input label of the SOH model, which restricts the effective training of the supervised learning model in the label missing scenario.

[0008] 5. Difficulty in fusing multi-source heterogeneous time series features: It is difficult to efficiently perform end-to-end feature learning and fusion in a unified algorithm framework for the multi-channel (such as single-cell voltage) and high-dimensional time series data generated during the operation of the battery, which limits the model's ability to capture complex health features.

[0009] Therefore, how to provide a lithium battery health state evaluation method and system based on a deep network to improve the accuracy, real-time performance and generalization ability of lithium battery health state evaluation has become a technical problem to be solved. SUMMARY

[0010] The technical problem to be solved by the present application is to provide a lithium battery health state evaluation method and system based on a deep network to improve the accuracy, real-time performance and generalization ability of lithium battery health state evaluation.

[0011] In a first aspect, the present application provides a lithium battery health state evaluation method based on a deep network, comprising the following steps:

[0012] Step S1, creating a health state evaluation model based on a double time series branch input module, a feature extraction module, a fusion and attention module, a time series modeling module and a prediction output module, and setting a loss function of the health state evaluation model;

[0013] Step S2, obtaining a large amount of historical lithium battery monitoring data containing cell-level time series signals and pack-level time series signals, and constructing a data set after preprocessing and labeling each of the historical lithium battery monitoring data;

[0014] Step S3, training the health state evaluation model through the data set and the loss function, and deploying the trained health state evaluation model;

[0015] Step S4, collecting real-time lithium battery monitoring data, inputting the preprocessed real-time lithium battery monitoring data into the deployed health state evaluation model, and obtaining a health state evaluation report.

[0016] Further, in step S1, the double time series branch input module is constructed based on a pack signal input channel and a cell signal input channel; the pack signal input channel is used to standardize the pack-level time series signal in the input lithium battery monitoring data and output to the feature extraction module; the cell signal input channel is used to standardize the cell-level time series signal in the input lithium battery monitoring data and output to the feature extraction module;

[0017] The feature extraction module is constructed based on a pack-level feature extraction unit and a single-body feature extraction unit; the pack-level feature extraction unit is configured to extract pack-level features from the pack-level time series signal through a first one-dimensional convolutional neural network and a channel attention subunit; and the single-body feature extraction unit is configured to extract single-body features from the cell-level time series signal through a second one-dimensional convolutional neural network;

[0018] The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit; the feature fusion unit is configured to splice the pack-level features and the single-body features, and then perform nonlinear transformation and dimension reduction on the spliced features through a fully connected layer to output a fusion feature vector; and the timestamp attention unit is configured to apply a self-attention mechanism to the fusion feature vector, calculate the importance weights of different time points, and output an enhanced feature after weighting;

[0019] The time series modeling module is configured to infer the enhanced feature through a bidirectional gated recurrent unit network to obtain a time series feature representation;

[0020] The prediction output module is configured to map the time series feature representation through a fully connected regressor to output a health state evaluation report carrying a SOH prediction value and auxiliary task prediction values (service life, driving mileage, and cycle number);

[0021] The formula of the loss function is as follows:

[0022]

[0023] wherein, L represents a loss value of the loss function; k represents a task index; λ k represents a task weight of the kth task; L k represents an average loss of each task; mask k (i) represents a binary mask variable, mask k (i) = 1 indicates that the task k label of the sample i exists, mask k (i) = 0 indicates that the task k label of the sample i does not exist; represents a sub-loss of the task k, and an error function is adopted; represents a true value of the task k; represents a prediction value of the task k.

[0024] Further, the step S2 is specifically as follows:

[0025] A large amount of historical lithium battery monitoring data containing cell-level time series signals and pack-level time series signals is acquired, wherein the cell-level time series signals at least include single-body voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position; and the pack-level time series signals at least include total voltage, total current, SOC, and average temperature;

[0026] The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

[0027] Furthermore, step S3 specifically includes:

[0028] The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

[0029] Furthermore, step S4 specifically includes:

[0030] Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

[0031] Secondly, the present invention provides a lithium battery health status assessment system based on deep networks, comprising the following modules:

[0032] A health status assessment model creation module is used to create a health status assessment model based on a dual-temporal-branch input module, a feature extraction module, a fusion and attention module, a temporal modeling module, and a prediction output module, and to set the loss function of the health status assessment model.

[0033] The dataset construction module is used to acquire a large amount of historical lithium battery monitoring data containing cell-level and package-level timing signals, and to construct the dataset after preprocessing and labeling the historical lithium battery monitoring data.

[0034] The health status assessment model training module is used to train the health status assessment model using the dataset and loss function, and to deploy the trained health status assessment model.

[0035] The health status assessment module is used to collect real-time lithium battery monitoring data, preprocess the real-time lithium battery monitoring data, and input it into the deployed health status assessment model to obtain a health status assessment report.

[0036] Furthermore, in the health status assessment model creation module, the dual-time-branch input module is constructed based on the package-level signal input channel and the cell signal input channel; the package-level signal input channel is used to standardize the package-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module; the cell signal input channel is used to standardize the cell-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module.

[0037] The feature extraction module is constructed based on a packet-level feature extraction unit and a single-unit feature extraction unit; the packet-level feature extraction unit is used to extract packet-level features from the packet-level time-series signal through a first one-dimensional convolutional neural network and a channel attention subunit; the single-unit feature extraction unit is used to extract single-unit features from the cell-level time-series signal through a second one-dimensional convolutional neural network.

[0038] The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit. The feature fusion unit is used to concatenate package-level features and individual features, and then perform nonlinear transformation and dimensionality reduction through a fully connected layer to output a fused feature vector. The timestamp attention unit is used to apply a self-attention mechanism to the fused feature vector, calculate the importance weights at different time points, and output enhanced features after weighting.

[0039] The temporal modeling module is used to infer the enhanced features through a bidirectional gated recurrent unit network to obtain a temporal feature representation;

[0040] The prediction output module is used to map the time-series feature representation through a fully connected regressor and output a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values ​​(years of use, mileage, number of cycles).

[0041] The formula for the loss function is:

[0042]

[0043] Where L represents the loss value of the loss function; k represents the task index; λ k L represents the task weight of the k-th task; k Mask represents the average loss for each task; k (i) represents the binary mask variable, mask k (i) = 1 indicates that the task k label for sample i exists, mask k (i) = 0 indicates that the task k label for sample i does not exist; The sub-loss for task k is represented by the mean squared error function; This represents the actual value of task k; This represents the predicted value for task k.

[0044] Furthermore, the dataset construction module is specifically used for:

[0045] Acquire a large amount of historical lithium battery monitoring data, including cell-level timing signals and pack-level timing signals. The cell-level timing signals include at least single cell voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position. The pack-level timing signals include at least total voltage, total current, SOC, and average temperature.

[0046] The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

[0047] Furthermore, the health status assessment model training module is specifically used for:

[0048] The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

[0049] Furthermore, the health status assessment module is specifically used for:

[0050] Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

[0051] The advantages of this invention are:

[0052] 1. A health status assessment model is created based on a dual-time-branch input module, a feature extraction module, a fusion and attention module, a time-series modeling module, and a prediction output module, and a loss function is set for the health status assessment model. Then, a large amount of historical lithium battery monitoring data containing cell-level and package-level time-series signals is acquired. After preprocessing and labeling each historical lithium battery monitoring data, a dataset is constructed. The health status assessment model is trained using the dataset and the loss function, and then deployed. Next, real-time lithium battery monitoring data is collected, preprocessed, and input into the deployed health status assessment model to obtain a health status assessment report. In short, through an end-to-end deep network architecture, combined with a dual-time-branch input module, package-level and cell-level time-series signals are processed simultaneously. For heterogeneous signals at the core level (such as cell voltage and temperature distribution location), multi-scale features are automatically extracted using a one-dimensional convolutional neural network. Then, key time point information is dynamically weighted through feature fusion and self-attention mechanisms, significantly improving the accuracy of SOH prediction. Lightweight bidirectional gated recurrent units are used for time series modeling, and a streaming computing engine is used to preprocess real-time data, achieving a second-level response and solving the real-time deficiency of traditional methods. At the same time, based on the multi-task mask loss function, SOH and auxiliary tasks (such as the number of iterations) are jointly learned, effectively utilizing incomplete labeled data. Combined with large-scale historical data training and sliding window verification mechanisms, the model's generalization ability to different aging paths and battery models is greatly enhanced, ultimately greatly improving the accuracy, real-time performance, and generalization ability of lithium battery health status assessment.

[0053] 2. Packet-level timing signals (such as total voltage and total current) and cell-level timing signals (such as individual cell voltage and detection point temperature) are processed separately through the package-level signal input channel and the cell-level signal input channel. This parallel design allows the model to simultaneously capture the macroscopic and microscopic state information of the lithium battery, solving the problem of incomplete evaluation caused by a single signal source in traditional methods, improving the accuracy of SOH prediction (e.g., it can more accurately identify inconsistencies between cells), and ensuring the consistency of input data through standardization processing (such as signal standardization), thereby enhancing the model's generalization ability.

[0054] 3. In the feature fusion unit, the packet-level features and individual features are concatenated and then subjected to nonlinear transformation and dimensionality reduction through a fully connected layer. Combined with the timestamp attention unit, a self-attention mechanism is applied to calculate the importance weights of different time points. This mechanism can automatically focus on key time points (such as battery charging and discharging peaks), suppress noise interference, and improve feature quality. This solves the problem of redundant information in time series signals, enabling the model to capture dynamic changes more efficiently and improve prediction accuracy (e.g., reducing SOH prediction error). Furthermore, the introduction of the attention mechanism enhances the interpretability of the model.

[0055] 4. The loss function design supports multi-task prediction, including the main SOH task and auxiliary tasks (such as years of use, mileage, and number of cycles). Missing labels are handled by task weights and binary mask variables, that is, by utilizing the correlation between tasks (such as the correlation between mileage and SOH), sharing feature representations, and reducing the risk of overfitting. The masking mechanism allows the model to flexibly handle incomplete data (for example, some auxiliary labels are missing in historical data), which improves its applicability and robustness in real-world scenarios. At the same time, it outputs a comprehensive health status report, which facilitates users' maintenance decisions.

[0056] 5. Using a time-series modeling module (such as a bidirectional gated recurrent unit network) to infer enhanced features addresses the long-term dependency problem of sequence data. The bidirectional structure can simultaneously consider information from previous and subsequent time points, better modeling the degradation process of lithium battery state over time (such as capacity decay). This solves the gradient vanishing problem of traditional RNN models in long sequences, improves the expressive power of time-series features, and thus supports more reliable SOH prediction.

[0057] 6. Consistent preprocessing steps are implemented in both historical and real-time data, including missing value handling, outlier handling, time alignment and synchronization, noise reduction and smoothing, which reduces data noise and bias and improves the reliability of model training and prediction; combined with a streaming computing engine, real-time processing is achieved, supporting efficient data stream processing, ensuring the timeliness and practicality of evaluation, while reducing the impact of data quality issues on model performance.

[0058] 7. The training process uses a sliding window method to divide the dataset (training set, validation set, and test set), and sets loss thresholds and early stopping conditions to optimize the model; after deployment, real-time data is processed through a streaming engine; the sliding window method enhances the utilization efficiency of time series data, and the early stopping mechanism prevents overfitting and improves the model convergence speed; the entire process (training-validation-testing-deployment) is compactly designed, achieving seamless integration into actual systems, reducing deployment costs, and supporting large-scale industrial applications.

[0059] 8. The prediction output module not only outputs the SOH prediction value, but also includes the prediction value of auxiliary tasks (such as service life and mileage), forming a multi-dimensional health status assessment report, providing more comprehensive battery status information (for example, the predicted mileage can guide users to optimize their usage habits), which enhances the practical value of the solution; the auxiliary tasks, as regularization terms, indirectly improve the accuracy of SOH prediction, meeting the needs of multi-indicator monitoring in commercial applications.

[0060] 9. A high-precision assessment of lithium battery health status is achieved through an innovative deep learning architecture (dual-time-branch input module, fusion and attention mechanism, and time-series modeling module). Its core advantages are: by utilizing deep fusion and adaptive feature enhancement of multi-scale signals (packet level and cell level), combined with a multi-task learning framework (SOH main task and auxiliary tasks such as driving range), the comprehensiveness and robustness of the prediction are significantly improved; at the same time, the standardized data preprocessing process, the masking mechanism to handle missing labels, and the real-time deployment of the streaming computing engine effectively solve the limitations of traditional methods in terms of noise sensitivity, incomplete data, and real-time performance, providing an efficient, reliable, and industrially applicable assessment solution for battery management systems. Attached Figure Description

[0061] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0062] Fig. 1 This is a flowchart of a lithium battery health status assessment method based on deep networks according to the present invention.

[0063] Fig. 2 This is a schematic diagram of the structure of a lithium battery health status assessment system based on deep networks according to the present invention. Detailed Implementation

[0064] The overall approach of the technical solution in this application is as follows: An end-to-end deep network architecture is used, combined with a dual-time-branch input module to synchronously process heterogeneous signals at the package and cell levels (such as individual cell voltage and temperature distribution). A one-dimensional convolutional neural network is used to automatically extract multi-scale features, and then key time-point information is dynamically weighted through feature fusion and self-attention mechanisms, significantly improving the accuracy of SOH prediction. Lightweight bidirectional gated recurrent units are used for time-series modeling, and a streaming computing engine is used for real-time data preprocessing, achieving second-level response and overcoming the shortcomings of traditional methods in terms of real-time performance. Simultaneously, based on a multi-task mask loss function, SOH and auxiliary tasks (such as loop counts) are jointly learned, effectively utilizing incomplete labeled data. Combined with large-scale historical data training and a sliding window verification mechanism, the model's generalization ability to different aging paths and battery models is greatly enhanced, thereby improving the accuracy, real-time performance, and generalization ability of lithium battery health status assessment.

[0065] Please refer to Figs. 1-2 As shown, a preferred embodiment of the lithium battery health status assessment method based on deep networks of the present invention includes the following steps:

[0066] Step S1: Create a health status assessment model based on the dual temporal branch input module, feature extraction module, fusion and attention module, temporal modeling module and prediction output module, and set the loss function of the health status assessment model;

[0067] Step S2: Obtain a large amount of historical lithium battery monitoring data containing cell-level timing signals and package-level timing signals; preprocess and label the historical lithium battery monitoring data to construct a dataset.

[0068] Step S3: Train the health status assessment model using the dataset and loss function, and deploy the trained health status assessment model.

[0069] Step S4: Collect real-time lithium battery monitoring data, preprocess the real-time lithium battery monitoring data and input it into the deployed health status assessment model to obtain a health status assessment report.

[0070] In step S1, the dual-time-branch input module is constructed based on a packet-level signal input channel and a cell-level signal input channel. The packet-level signal input channel is used to standardize (e.g., normalize) the packet-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The cell-level signal input channel is used to standardize the cell-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The dual-branch design allows for parallel processing of signals of different granularities, avoiding information mixing. Data standardization improves the model's convergence speed and robustness.

[0071] The feature extraction module is constructed based on a packet-level feature extraction unit and a single-unit feature extraction unit. The packet-level feature extraction unit is used to extract packet-level features from the packet-level time-series signal using a first one-dimensional convolutional neural network (1D-CNN) and a channel attention subunit. The single-unit feature extraction unit is used to extract single-unit features from the cell-level time-series signal using a second one-dimensional convolutional neural network. The feature extraction module is used to extract local features and short-time dynamic features of the signal. The packet-level feature extraction unit extracts features from the packet-level time-series signal and applies a channel attention mechanism to dynamically weight different physical quantities. The single-unit feature extraction unit extracts features from the cell-level time-series signal. 1D-CNN has weight sharing and translation invariance, making it suitable for processing high-dimensional time-series signals. The channel attention mechanism strengthens discriminative features (such as the contribution of total voltage to SOH) and improves feature representation ability. Residual connections (implemented in CNN) ensure the stability of deep feature extraction.

[0072] The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit. The feature fusion unit concatenates (Concat) packet-level features and individual features, then performs nonlinear transformation and dimensionality reduction through a fully connected layer (Dense) to output a fused feature vector. The timestamp attention unit applies a self-attention mechanism (e.g., through query, key, and value calculation) to the fused feature vector, calculates the importance weights of different time points, and outputs enhanced features after weighting. The fusion and attention module fuses packet-level and individual features and captures global dependencies in the time dimension through a self-attention mechanism, strengthening features at key time points. The feature fusion unit (Concat+Dense) combines features of different granularities to enrich the feature dimensions. The timestamp attention unit models long-term dependencies, improving the ability to capture battery state evolution trends. The overall design takes into account both local mutations and long-term trends.

[0073] The temporal modeling module is used to infer the enhanced features through a bidirectional gated recurrent unit network (Bi-GRU) to obtain temporal feature representations (hidden states); the temporal modeling module is used to model long-term temporal dependencies and capture the evolution trend of battery state (such as capacity decay); Bi-GRU combines forward and backward information to effectively process long sequences; the gating mechanism avoids gradient vanishing and improves the stability and accuracy of temporal modeling.

[0074] The prediction output module maps the temporal feature representation through a fully connected regressor, outputting a health status assessment report carrying SOH prediction values ​​(main task) and auxiliary task prediction values ​​(years of use, mileage, number of cycles). The prediction output module maps temporal features to prediction results, including the main task (SOH) and auxiliary tasks (years of use, mileage, number of cycles). The fully connected layer implements a non-linear mapping from features to output. The multi-task output design only requires SOH prediction during the inference phase, simplifying deployment.

[0075] The formula for the loss function is:

[0076]

[0077] Where L represents the loss value of the loss function; k represents the task index (k=1: SOH main task; k=2: usage years auxiliary task; k=3: mileage auxiliary task; k=4: number of iterations auxiliary task); λ k L represents the task weight of the k-th task; k Mask represents the average loss for each task; k (i) represents the binary mask variable, mask k (i) = 1 indicates that the task k label for sample i exists, mask k(i) = 0 indicates that the task k label for sample i does not exist; The sub-loss of task k is represented by the mean squared error (MSE) function; This represents the actual value of task k; This represents the predicted value for task k.

[0078] By selectively guiding labeled samples to participate in training through the Mask mechanism, the interference of unlabeled samples on the main task is avoided. At the same time, auxiliary tasks such as service life, mileage, and number of cycles are introduced to collaboratively learn and share temporal feature representations, thereby effectively improving the accuracy and generalization ability of SOH prediction even when labels are incomplete.

[0079] Step S2 specifically involves:

[0080] Acquire a large amount of historical lithium battery monitoring data, including cell-level timing signals and pack-level timing signals. The cell-level timing signals include at least single cell voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position. The pack-level timing signals include at least total voltage, total current, SOC, and average temperature.

[0081] The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

[0082] Step S3 specifically involves:

[0083] The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

[0084] Step S4 specifically involves:

[0085] Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

[0086] A preferred embodiment of the lithium battery health status assessment system based on deep networks of the present invention includes the following modules:

[0087] A health status assessment model creation module is used to create a health status assessment model based on a dual-temporal-branch input module, a feature extraction module, a fusion and attention module, a temporal modeling module, and a prediction output module, and to set the loss function of the health status assessment model.

[0088] The dataset construction module is used to acquire a large amount of historical lithium battery monitoring data containing cell-level and package-level timing signals, and to construct the dataset after preprocessing and labeling the historical lithium battery monitoring data.

[0089] The health status assessment model training module is used to train the health status assessment model using the dataset and loss function, and to deploy the trained health status assessment model.

[0090] The health status assessment module is used to collect real-time lithium battery monitoring data, preprocess the real-time lithium battery monitoring data, and input it into the deployed health status assessment model to obtain a health status assessment report.

[0091] In the health status assessment model creation module, the dual-time-branch input module is constructed based on a packet-level signal input channel and a cell-level signal input channel. The packet-level signal input channel is used to standardize (e.g., normalize) the packet-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The cell-level signal input channel is used to standardize the cell-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The dual-branch design allows for parallel processing of signals of different granularities, avoiding information mixing. Data standardization improves the model's convergence speed and robustness.

[0092] The feature extraction module is constructed based on a packet-level feature extraction unit and a single-unit feature extraction unit. The packet-level feature extraction unit is used to extract packet-level features from the packet-level time-series signal using a first one-dimensional convolutional neural network (1D-CNN) and a channel attention subunit. The single-unit feature extraction unit is used to extract single-unit features from the cell-level time-series signal using a second one-dimensional convolutional neural network. The feature extraction module is used to extract local features and short-time dynamic features of the signal. The packet-level feature extraction unit extracts features from the packet-level time-series signal and applies a channel attention mechanism to dynamically weight different physical quantities. The single-unit feature extraction unit extracts features from the cell-level time-series signal. 1D-CNN has weight sharing and translation invariance, making it suitable for processing high-dimensional time-series signals. The channel attention mechanism strengthens discriminative features (such as the contribution of total voltage to SOH) and improves feature representation ability. Residual connections (implemented in CNN) ensure the stability of deep feature extraction.

[0093] The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit. The feature fusion unit concatenates (Concat) packet-level features and individual features, then performs nonlinear transformation and dimensionality reduction through a fully connected layer (Dense) to output a fused feature vector. The timestamp attention unit applies a self-attention mechanism (e.g., through query, key, and value calculation) to the fused feature vector, calculates the importance weights of different time points, and outputs enhanced features after weighting. The fusion and attention module fuses packet-level and individual features and captures global dependencies in the time dimension through a self-attention mechanism, strengthening features at key time points. The feature fusion unit (Concat+Dense) combines features of different granularities to enrich the feature dimensions. The timestamp attention unit models long-term dependencies, improving the ability to capture battery state evolution trends. The overall design takes into account both local mutations and long-term trends.

[0094] The temporal modeling module is used to infer the enhanced features through a bidirectional gated recurrent unit network (Bi-GRU) to obtain temporal feature representations (hidden states); the temporal modeling module is used to model long-term temporal dependencies and capture the evolution trend of battery state (such as capacity decay); Bi-GRU combines forward and backward information to effectively process long sequences; the gating mechanism avoids gradient vanishing and improves the stability and accuracy of temporal modeling.

[0095] The prediction output module maps the temporal feature representation through a fully connected regressor, outputting a health status assessment report carrying SOH prediction values ​​(main task) and auxiliary task prediction values ​​(years of use, mileage, number of cycles). The prediction output module maps temporal features to prediction results, including the main task (SOH) and auxiliary tasks (years of use, mileage, number of cycles). The fully connected layer implements a non-linear mapping from features to output. The multi-task output design only requires SOH prediction during the inference phase, simplifying deployment.

[0096] The formula for the loss function is:

[0097]

[0098] Where L represents the loss value of the loss function; k represents the task index (k=1: SOH main task; k=2: usage years auxiliary task; k=3: mileage auxiliary task; k=4: number of iterations auxiliary task); λ k L represents the task weight of the k-th task; k This represents the average loss for each task; mask k (i) represents the binary mask variable, mask k (i) = 1 indicates that the task k label for sample i exists, maskk (i) = 0 indicates that the task k label for sample i does not exist; The sub-loss of task k is represented by the mean squared error (MSE) function; This represents the actual value of task k; This represents the predicted value for task k.

[0099] By selectively guiding labeled samples to participate in training through the Mask mechanism, the interference of unlabeled samples on the main task is avoided. At the same time, auxiliary tasks such as service life, mileage, and number of cycles are introduced to collaboratively learn and share temporal feature representations, thereby effectively improving the accuracy and generalization ability of SOH prediction even when labels are incomplete.

[0100] The dataset construction module is specifically used for:

[0101] Acquire a large amount of historical lithium battery monitoring data, including cell-level timing signals and pack-level timing signals. The cell-level timing signals include at least single cell voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position. The pack-level timing signals include at least total voltage, total current, SOC, and average temperature.

[0102] The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

[0103] The health status assessment model training module is specifically used for:

[0104] The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

[0105] The health status assessment module is specifically used for:

[0106] Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

[0107] In summary, the advantages of this invention are as follows:

[0108] 1. A health status assessment model is created based on a dual-time-branch input module, a feature extraction module, a fusion and attention module, a time-series modeling module, and a prediction output module, and a loss function is set for the health status assessment model. Then, a large amount of historical lithium battery monitoring data containing cell-level and package-level time-series signals is acquired. After preprocessing and labeling each historical lithium battery monitoring data, a dataset is constructed. The health status assessment model is trained using the dataset and the loss function, and then deployed. Next, real-time lithium battery monitoring data is collected, preprocessed, and input into the deployed health status assessment model to obtain a health status assessment report. In short, through an end-to-end deep network architecture, combined with a dual-time-branch input module, package-level and cell-level time-series signals are processed simultaneously. For heterogeneous signals at the core level (such as cell voltage and temperature distribution location), multi-scale features are automatically extracted using a one-dimensional convolutional neural network. Then, key time point information is dynamically weighted through feature fusion and self-attention mechanisms, significantly improving the accuracy of SOH prediction. Lightweight bidirectional gated recurrent units are used for time series modeling, and a streaming computing engine is used to preprocess real-time data, achieving a second-level response and solving the real-time deficiency of traditional methods. At the same time, based on the multi-task mask loss function, SOH and auxiliary tasks (such as the number of iterations) are jointly learned, effectively utilizing incomplete labeled data. Combined with large-scale historical data training and sliding window verification mechanisms, the model's generalization ability to different aging paths and battery models is greatly enhanced, ultimately greatly improving the accuracy, real-time performance, and generalization ability of lithium battery health status assessment.

[0109] 2. Packet-level timing signals (such as total voltage and total current) and cell-level timing signals (such as individual cell voltage and detection point temperature) are processed separately through the package-level signal input channel and the cell-level signal input channel. This parallel design allows the model to simultaneously capture the macroscopic and microscopic state information of the lithium battery, solving the problem of incomplete evaluation caused by a single signal source in traditional methods, improving the accuracy of SOH prediction (e.g., it can more accurately identify inconsistencies between cells), and ensuring the consistency of input data through standardization processing (such as signal standardization), thereby enhancing the model's generalization ability.

[0110] 3. In the feature fusion unit, the packet-level features and individual features are concatenated and then subjected to nonlinear transformation and dimensionality reduction through a fully connected layer. Combined with the timestamp attention unit, a self-attention mechanism is applied to calculate the importance weights of different time points. This mechanism can automatically focus on key time points (such as battery charging and discharging peaks), suppress noise interference, and improve feature quality. This solves the problem of redundant information in time series signals, enabling the model to capture dynamic changes more efficiently and improve prediction accuracy (e.g., reducing SOH prediction error). Furthermore, the introduction of the attention mechanism enhances the interpretability of the model.

[0111] 4. The loss function design supports multi-task prediction, including the main SOH task and auxiliary tasks (such as years of use, mileage, and number of cycles). Missing labels are handled by task weights and binary mask variables, that is, by utilizing the correlation between tasks (such as the correlation between mileage and SOH), sharing feature representations, and reducing the risk of overfitting. The masking mechanism allows the model to flexibly handle incomplete data (for example, some auxiliary labels are missing in historical data), which improves its applicability and robustness in real-world scenarios. At the same time, it outputs a comprehensive health status report, which facilitates users' maintenance decisions.

[0112] 5. Using a time-series modeling module (such as a bidirectional gated recurrent unit network) to infer enhanced features addresses the long-term dependency problem of sequence data. The bidirectional structure can simultaneously consider information from previous and subsequent time points, better modeling the degradation process of lithium battery state over time (such as capacity decay). This solves the gradient vanishing problem of traditional RNN models in long sequences, improves the expressive power of time-series features, and thus supports more reliable SOH prediction.

[0113] 6. Consistent preprocessing steps are implemented in both historical and real-time data, including missing value handling, outlier handling, time alignment and synchronization, noise reduction and smoothing, which reduces data noise and bias and improves the reliability of model training and prediction; combined with a streaming computing engine, real-time processing is achieved, supporting efficient data stream processing, ensuring the timeliness and practicality of evaluation, while reducing the impact of data quality issues on model performance.

[0114] 7. The training process uses a sliding window method to divide the dataset (training set, validation set, and test set), and sets loss thresholds and early stopping conditions to optimize the model; after deployment, real-time data is processed through a streaming engine; the sliding window method enhances the utilization efficiency of time series data, and the early stopping mechanism prevents overfitting and improves the model convergence speed; the entire process (training-validation-testing-deployment) is compactly designed, achieving seamless integration into actual systems, reducing deployment costs, and supporting large-scale industrial applications.

[0115] 8. The prediction output module not only outputs the SOH prediction value, but also includes the prediction value of auxiliary tasks (such as service life and mileage), forming a multi-dimensional health status assessment report, providing more comprehensive battery status information (for example, the predicted mileage can guide users to optimize their usage habits), which enhances the practical value of the solution; the auxiliary tasks, as regularization terms, indirectly improve the accuracy of SOH prediction, meeting the needs of multi-indicator monitoring in commercial applications.

[0116] 9. A high-precision assessment of lithium battery health status is achieved through an innovative deep learning architecture (dual-time-branch input module, fusion and attention mechanism, and time-series modeling module). Its core advantages are: by utilizing deep fusion and adaptive feature enhancement of multi-scale signals (packet level and cell level), combined with a multi-task learning framework (SOH main task and auxiliary tasks such as driving range), the comprehensiveness and robustness of the prediction are significantly improved; at the same time, the standardized data preprocessing process, the masking mechanism to handle missing labels, and the real-time deployment of the streaming computing engine effectively solve the limitations of traditional methods in terms of noise sensitivity, incomplete data, and real-time performance, providing an efficient, reliable, and industrially applicable assessment solution for battery management systems.

[0117] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for assessing the health status of lithium batteries based on deep networks, characterized in that: Includes the following steps: Step S1: Create a health status assessment model based on the dual temporal branch input module, feature extraction module, fusion and attention module, temporal modeling module and prediction output module, and set the loss function of the health status assessment model; Step S2: Obtain a large amount of historical lithium battery monitoring data containing cell-level timing signals and package-level timing signals; preprocess and label the historical lithium battery monitoring data to construct a dataset. Step S3: Train the health status assessment model using the dataset and loss function, and deploy the trained health status assessment model. Step S4: Collect real-time lithium battery monitoring data, preprocess the real-time lithium battery monitoring data and input it into the deployed health status assessment model to obtain a health status assessment report.

2. The lithium battery health status assessment method based on deep networks as described in claim 1, characterized in that: In step S1, the dual-time-branch input module is constructed based on the packet-level signal input channel and the cell signal input channel; the packet-level signal input channel is used to standardize the packet-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module; the cell signal input channel is used to standardize the cell-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The feature extraction module is constructed based on a packet-level feature extraction unit and a single-unit feature extraction unit; the packet-level feature extraction unit is used to extract packet-level features from the packet-level time-series signal through a first one-dimensional convolutional neural network and a channel attention subunit; the single-unit feature extraction unit is used to extract single-unit features from the cell-level time-series signal through a second one-dimensional convolutional neural network. The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit. The feature fusion unit is used to concatenate package-level features and individual features, and then perform nonlinear transformation and dimensionality reduction through a fully connected layer to output a fused feature vector. The timestamp attention unit is used to apply a self-attention mechanism to the fused feature vector, calculate the importance weights at different time points, and output enhanced features after weighting. The temporal modeling module is used to infer the enhanced features through a bidirectional gated recurrent unit network to obtain a temporal feature representation; The prediction output module is used to map the time-series feature representation through a fully connected regressor and output a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values ​​(years of use, mileage, number of cycles). The formula for the loss function is: Where L represents the loss value of the loss function; k represents the task index; λ k L represents the task weight of the k-th task; k This represents the average loss for each task; mask k (i) represents the binary mask variable, mask k (i) = 1 indicates that the task k label for sample i exists, mask k (i) = 0 indicates that the task k label for sample i does not exist; The sub-loss for task k is represented by the mean squared error function; This represents the actual value of task k; This represents the predicted value for task k.

3. The lithium battery health status assessment method based on deep networks as described in claim 1, characterized in that: Step S2 specifically involves: Acquire a large amount of historical lithium battery monitoring data, including cell-level timing signals and pack-level timing signals. The cell-level timing signals include at least single cell voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position. The pack-level timing signals include at least total voltage, total current, SOC, and average temperature. The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

4. The lithium battery health status assessment method based on deep networks as described in claim 1, characterized in that: Step S3 specifically involves: The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

5. The lithium battery health status assessment method based on deep networks as described in claim 1, characterized in that: Step S4 specifically involves: Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

6. A lithium battery health status assessment system based on deep networks, characterized in that: Includes the following modules: A health status assessment model creation module is used to create a health status assessment model based on a dual-temporal-branch input module, a feature extraction module, a fusion and attention module, a temporal modeling module, and a prediction output module, and to set the loss function of the health status assessment model. The dataset construction module is used to acquire a large amount of historical lithium battery monitoring data containing cell-level and package-level timing signals, and to construct the dataset after preprocessing and labeling the historical lithium battery monitoring data. The health status assessment model training module is used to train the health status assessment model using the dataset and loss function, and to deploy the trained health status assessment model. The health status assessment module is used to collect real-time lithium battery monitoring data, preprocess the real-time lithium battery monitoring data, and input it into the deployed health status assessment model to obtain a health status assessment report.

7. The lithium battery health status assessment system based on deep networks as described in claim 6, characterized in that: In the health status assessment model creation module, the dual-time-branch input module is constructed based on the package-level signal input channel and the cell signal input channel; the package-level signal input channel is used to standardize the package-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module; the cell signal input channel is used to standardize the cell-level time-series signal in the input lithium battery monitoring data and then output it to the feature extraction module. The feature extraction module is constructed based on a packet-level feature extraction unit and a single-unit feature extraction unit; the packet-level feature extraction unit is used to extract packet-level features from the packet-level time-series signal through a first one-dimensional convolutional neural network and a channel attention subunit; the single-unit feature extraction unit is used to extract single-unit features from the cell-level time-series signal through a second one-dimensional convolutional neural network. The fusion and attention module is constructed based on a feature fusion unit and a timestamp attention unit. The feature fusion unit is used to concatenate package-level features and individual features, and then perform nonlinear transformation and dimensionality reduction through a fully connected layer to output a fused feature vector. The timestamp attention unit is used to apply a self-attention mechanism to the fused feature vector, calculate the importance weights at different time points, and output enhanced features after weighting. The temporal modeling module is used to infer the enhanced features through a bidirectional gated recurrent unit network to obtain a temporal feature representation; The prediction output module is used to map the time-series feature representation through a fully connected regressor and output a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values ​​(years of use, mileage, number of cycles). The formula for the loss function is: Where L represents the loss value of the loss function; k represents the task index; λ k L represents the task weight of the k-th task; k This represents the average loss for each task; mask k (i) represents the binary mask variable, mask k (i) = 1 indicates that the task k label for sample i exists, mask k (i) = 0 indicates that the task k label for sample i does not exist; The sub-loss for task k is represented by the mean squared error function; This represents the actual value of task k; This represents the predicted value for task k.

8. The lithium battery health status assessment system based on deep networks as described in claim 6, characterized in that: The dataset construction module is specifically used for: Acquire a large amount of historical lithium battery monitoring data, including cell-level timing signals and pack-level timing signals. The cell-level timing signals include at least single cell voltage, detection point temperature, highest temperature position, lowest temperature position, highest voltage position, and lowest voltage position. The pack-level timing signals include at least total voltage, total current, SOC, and average temperature. The historical lithium battery monitoring data are preprocessed, including missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed historical lithium battery monitoring data are then labeled with the actual SOH value and the actual value of the auxiliary task, and a dataset is constructed.

9. The lithium battery health status assessment system based on deep networks as described in claim 6, characterized in that: The health status assessment model training module is specifically used for: The dataset is divided into a training set, a validation set, and a test set using the sliding window method. The health status assessment model is trained using the training set until the loss value of the loss function is less than a preset loss threshold or a preset early stop condition is met. The trained health status assessment model is then validated and tested using the validation set and the test set. The health status assessment model that passes the test is then deployed.

10. The lithium battery health status assessment system based on deep networks as described in claim 6, characterized in that: The health status assessment module is specifically used for: Real-time lithium battery monitoring data, including cell-level and package-level timing signals, is collected. The real-time lithium battery monitoring data is preprocessed using a streaming computing engine, including at least missing value processing, outlier processing, time alignment and synchronization, noise reduction and smoothing. The preprocessed real-time lithium battery monitoring data is then input into a deployed health status assessment model to obtain a health status assessment report carrying SOH prediction values ​​and auxiliary task prediction values.

Images