Deep learning method and device for predicting cell failure in bms

Abstract

The application relates to the technical field of battery monitoring, in particular to a deep learning method and device for predicting cell failure in a BMS. The method comprises the following steps: collecting multi-modal operation data of the cell and dividing the data into time sequence samples; extracting time domain features, frequency domain features and differential features of the time sequence samples, selecting the features by using a Pearson correlation coefficient and mutual information, reducing the dimensions by principal component analysis, and constructing a data set; constructing a cell failure prediction model based on a gated recurrent unit model, optimizing the performance of the model by combining an attention mechanism; on the basis of predicting the cell failure, predicting the battery health state and failure time, introducing an adaptive failure threshold mechanism, and dynamically adjusting the failure judgment standard; constructing an incremental learning framework based on sample selection, and dynamically adjusting the prediction model by using new working condition samples. The method can significantly improve the accuracy, robustness and adaptability to complex working conditions of the cell failure prediction, and provides more reliable decision support for the battery management system.

Description

Deep learning methods and devices for cell failure prediction in BMS Technical Field

[0001] This application relates to the field of battery monitoring technology, specifically to a deep learning method and apparatus for predicting cell failure in a battery management system (BMS). Background Technology

[0002] The battery management system (BMS) is the core guarantee for the operation of the battery pack. As the basic unit of lithium battery, the operating status of the battery cell directly affects the performance, safety and life of the entire battery pack.

[0003] Traditional battery cell failure prediction methods are mostly based on empirical models or simple physical models. These methods typically rely on limited prior knowledge and struggle to fully capture the complex nonlinear dynamic changes during battery cell operation. In terms of data processing, existing technologies have low efficiency in utilizing multimodal battery cell operating data. Although modern battery cells can collect various operating parameters (such as voltage, current, and temperature) through sensors, existing methods generally lack the ability to effectively fuse and deeply analyze this high-dimensional data, making it difficult for models to fully utilize data information and affecting prediction performance.

[0004] Most existing cell failure prediction methods lack the ability to adapt to changes in operating conditions. The operating environment and conditions of cells may vary significantly in different application scenarios, and fixed models are unable to cope with such dynamic changes, resulting in unstable or inaccurate prediction results.

[0005] In summary, existing technologies have significant shortcomings in terms of accuracy of cell failure prediction, multimodal data processing capabilities, adaptability to operating conditions, and prediction range, and more advanced technologies are urgently needed to solve these problems.

[0006] In view of this, this application proposes a deep learning method and apparatus for predicting cell failure in a BMS. Summary of the Invention

[0007] To achieve the above objectives, this application provides a deep learning method and apparatus for predicting cell failure in a battery management system (BMS), the specific technical solution of which is as follows:

[0008] Deep learning methods for cell failure prediction in BMS include:

[0009] Multi-mode cell operation data is collected, and the collected multi-mode cell operation data is preprocessed and divided into time-series samples using a sliding window method.

[0010] The time-domain features, frequency-domain features, and difference features of the time-series samples are extracted. Feature selection is performed using Pearson correlation coefficient and mutual information, and dimensionality reduction is performed using principal component analysis to construct the dataset.

[0011] A cell failure prediction model based on the gated recurrent unit (GRU) model was constructed. The GRU model was trained using dataset data, and an attention mechanism was added to the GRU model to train the model and optimize the model parameters.

[0012] The cell failure prediction model is extended to predict battery health status and failure time while predicting cell failure, and an adaptive failure threshold mechanism is introduced.

[0013] To address the new operating conditions generated during battery cell operation, an incremental learning framework based on sample selection is constructed to adjust the battery cell failure prediction model using samples from these new operating conditions.

[0014] Preferably, multiple sensors are installed on the battery cell to collect multi-mode battery cell operation data during the battery cell operation process, including voltage, current, temperature and pressure;

[0015] The collected multi-mode cell operation data is preprocessed, including outlier detection and removal; the preprocessed data is time-aligned and then normalized; a sliding window method is used to divide the normalized multi-mode cell operation data into time-series samples.

[0016] Preferably, for each time series sample, feature extraction and calculation are performed, including: extracting time domain features, extracting frequency domain features, and calculating difference features;

[0017] Feature selection is performed on the extracted and computed time-domain features, frequency-domain features, and difference features using Pearson correlation coefficient and mutual information;

[0018] Principal component analysis is performed on the selected features to reduce their dimensionality, resulting in a feature matrix. The reduced feature matrix is ​​then combined with the corresponding label vectors to form the dataset.

[0019] Preferably, a cell failure prediction model based on a gated recurrent unit (GRU) model is constructed. The input layer of the GRU model receives the feature vector of time-series samples, and the output layer outputs the cell failure probability. Multiple GRU hidden layers are set between the input and output layers of the GRU model, and multiple GRU units are set in each hidden layer.

[0020] An attention mechanism is introduced into the constructed GRU model to make the GRU model focus on temporal sample information; the hyperparameters of the GRU model are tuned using cross-validation.

[0021] Preferably, while predicting cell failure, two auxiliary tasks are introduced: battery health status prediction and failure time prediction. Battery health status is defined as the ratio of the current capacity to the initial capacity, and failure time is defined as the remaining number of cycles from the current time to the failure time. Battery health status prediction and failure time prediction are modeled as regression problems.

[0022] Preferably, based on the constructed GRU model, output branches for battery health state prediction and failure time prediction are introduced. The feature representation of time-series samples is extracted through the GRU encoder in the GRU model, and then mapped to the output of different tasks through a fully connected layer.

[0023] An adaptive failure threshold mechanism is introduced to dynamically adjust the cell failure probability threshold based on the predicted battery health status and failure time.

[0024] Preferably, an incremental learning framework based on sample selection is constructed, and the parameters of the pre-trained cell failure prediction model are adjusted by selecting new operating condition samples; the selection strategy for new operating condition samples includes random selection and diversity selection.

[0025] An incremental learning objective function is constructed. By balancing the fitting of new operating condition samples and the parameter update magnitude, the updated model parameters are obtained. The incremental learning objective function is optimized using a mini-batch gradient descent algorithm to obtain a cell failure prediction model adapted to the new operating conditions.

[0026] The deep learning device for predicting cell failure in BMS, which is used to implement the deep learning method for predicting cell failure in BMS, includes: a data acquisition module, a feature extraction module, a GRU model module, a function extension module, and an incremental learning module.

[0027] The data acquisition module is used to collect multi-mode cell operation data, preprocess the collected multi-mode cell operation data, and divide it into time-series samples using a sliding window method.

[0028] The feature extraction module is used to extract the time-domain features, frequency-domain features, and difference features of the time-series samples, perform feature selection using Pearson correlation coefficient and mutual information, and use principal component analysis to perform dimensionality reduction to construct a dataset.

[0029] The GRU model module is used to construct a cell failure prediction model based on the gated cyclic unit (GRU) model. The GRU model is trained using dataset data, an attention mechanism is added to the GRU model, the GRU model is trained, and the model parameters are optimized.

[0030] The functional extension module is used to extend the cell failure prediction model, predicting the battery health status and failure time while predicting cell failure, and introducing an adaptive failure threshold mechanism.

[0031] The incremental learning module constructs an incremental learning framework based on sample selection to address new operating conditions generated during battery cell operation, and adjusts the battery cell failure prediction model using new operating condition samples.

[0032] An electronic device includes a processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the deep learning method for cell failure prediction in the BMS by calling the computer program stored in the memory.

[0033] A computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the deep learning method for predicting cell failure in the BMS.

[0034] The beneficial effects of this application are: This application comprehensively captures the dynamic changes in the operating status of battery cells through multimodal data acquisition and sliding window partitioning, thereby improving the temporal integrity of the samples and the accuracy of prediction.

[0035] This application extracts multiple features and performs selection and dimensionality reduction to reduce redundant information, improve data quality, reduce model complexity, and enhance the model's sensitivity to key features.

[0036] This application constructs and trains a GRU model with an attention mechanism to effectively capture long-term dependencies in time-series data, thereby improving the accuracy of cell failure prediction and the model's generalization ability.

[0037] This application extends the functionality of the constructed GRU model, simultaneously predicting SOH and failure time, and introduces an adaptive failure threshold mechanism to improve the model's applicability and prediction reliability in multi-task scenarios.

[0038] This application uses an incremental learning framework to adapt to new operating condition samples, improving the model's adaptability and dynamic adjustment capabilities, and enabling real-time optimization of the cell failure prediction model.

[0039] The method in this application constructs a highly accurate and adaptable cell failure prediction system through multimodal data acquisition, feature extraction, attention mechanism optimization, task expansion, and incremental learning framework, providing a more intelligent and comprehensive monitoring and prediction capability for the battery management system (BMS). Attached Figure Description

[0040] Figure 1 is a flowchart of the deep learning method for cell failure prediction in BMS provided in this application;

[0041] Figure 2 is a flowchart of the deep learning method for cell failure prediction in BMS provided in this application, which is used to construct the cell failure prediction model.

[0042] Figure 3 is a flowchart of the incremental learning framework for constructing the deep learning method for cell failure prediction in BMS provided in this application.

[0043] Figure 4 is a structural diagram of the deep learning device for cell failure prediction in the BMS provided in this application. Detailed Implementation

[0044] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of this application. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that excludes other embodiments.

[0047] Example 1

[0048] Referring to Figures 1 to 3, the first embodiment of this application provides a deep learning method for predicting cell failure in a BMS.

[0049] Step 1: Collect multi-mode cell operation data, preprocess the collected multi-mode cell operation data, and divide it into time series samples using a sliding window method.

[0050] For example, by installing multiple sensors on the battery cell, multi-mode battery cell operation data during the battery cell operation process is collected, including voltage, current, temperature and pressure; wherein, voltage and current data are collected by sensors with an accuracy of 1mV and 1mA, and the sampling frequency is 100Hz; temperature data is collected by thermocouples with an accuracy of 1℃, and the sampling frequency is 1Hz; pressure data is collected by pressure sensors with an accuracy of 1kPa, and the sampling frequency is 1Hz.

[0051] Outlier detection and removal are performed on the collected multi-mode cell operation data; an outlier detection method based on interquartile range (IQR) is adopted for the data. Calculate its relationship with the first quartile. The difference Sum and third quartile The difference :

[0052] ;

[0053] like or Then Mark as an outlier and remove; where .

[0054] Time alignment is necessary because the data sampling frequencies of different modes are different. Using voltage and current data as a reference, temperature and pressure data are linearly interpolated to align them with voltage and current data in the time dimension.

[0055] Let the timestamps for voltage and current data be... The timestamp for the temperature data is For any time Find the two closest temperature data timestamps. and Calculated by linear interpolation Temperature value at time :

[0056] ;

[0057] in, and They are time points and The time alignment method for pressure data is the same as that for temperature data.

[0058] Numerical normalization is performed on the time-aligned multi-mode cell operation data, mapping the data of each mode to... Within the interval; the maximum-minimum normalization method is used for the data of each modality. Calculate its maximum value and minimum value Normalize it to: ;in, These are the normalized data values.

[0059] A sliding window approach is used to divide the normalized multi-mode cell operation data into time-series samples; let the window size be... The sliding step size is Then the first The time series samples are:

[0060] ;

[0061] in, For the first A multimodal data vector of time series samples, i.e., time series samples; the label of a time series sample is the cell failure state after the last time step of that sample. tags Used to mark the failure status of the battery cell after the sample is completed. If the battery cell fails in the first... If it fails after a certain time step, then ,otherwise .

[0062] Step 2: Extract the time-domain features, frequency-domain features, and difference features of the time-series samples, perform feature selection using Pearson correlation coefficient and mutual information, and use principal component analysis to perform dimensionality reduction to construct the dataset.

[0063] For each time series sample Extract time-domain features, including: mean Standard deviation Root mean square Peak Peak factor skewness and kurtosis .

[0064] Extract for each time series sample The signal is then transformed to the frequency domain using a Fast Fourier Transform (FFT) to obtain the frequency domain signal. ;from Extracting frequency domain features, including frequency domain energy. Frequency domain entropy Frequency domain peak and frequency domain centroid .

[0065] For each time series sample Calculate its first-order difference sequence : ;from Temporal statistical features are extracted as difference features. Indicates the first The time series sample at the time series is A multimodal data vector at each time step.

[0066] Pearson correlation coefficient and mutual information were used to select features from the extracted time-domain features, frequency-domain features, and difference features.

[0067] For each feature Calculate its relationship with the label Pearson correlation coefficient :

[0068] ;

[0069] in, Indicates the first The time series sample of the th time series The value of each feature, and Representing features respectively and tags The mean.

[0070] For each feature Calculate its relationship with the label mutual information :

[0071] ;

[0072] in, Representation of features and tags The joint probability distribution, and Representing features respectively and tags The marginal probability distribution.

[0073] Based on the Pearson correlation coefficient and mutual information, the top results with the highest relevance to the labels are selected. One feature;

[0074] Principal component analysis (PCA) is used to reduce the dimensionality of the selected features; let the original feature matrix be... , in This represents the number of time-series samples. It is a sparse matrix; for After centralization and standardization, a matrix is ​​obtained. ,right Perform eigenvalue decomposition: ;in, The eigenvector matrix, It is the transpose of the eigenvector matrix. It is a diagonal matrix, and the diagonal elements are the eigenvalues.

[0075] Sort the eigenvalues ​​and select the first eigenvalue with the largest value. eigenvectors form a matrix , the original feature matrix Mapping to a lower-dimensional space: ;in, This is the feature matrix after dimensionality reduction.

[0076] The reduced feature matrix With the corresponding label vector Combined, they form a complete dataset. , in For the first Feature vectors of time series samples For the first Labels for each time series sample.

[0077] Step 3: Construct a cell failure prediction model based on the gated recurrent unit (GRU) model. Train the GRU model using the dataset, add an attention mechanism to the GRU model, train the GRU model, optimize the model parameters, and perform hyperparameter tuning on the GRU model.

[0078] The input layer of the gated recurrent unit (GRU) model receives the feature vector of the time-series sample, and the output layer outputs the cell failure probability. Between the input layer and the output layer, multiple GRU hidden layers are set, and each hidden layer contains multiple GRU units.

[0079] For the Input vector at each time step The update formula for the GRU cell is as follows:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] in, To reset the door, To update the door, In the candidate hidden state, Currently in a hidden state. This is the weight matrix. For bias vectors, It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This is element-wise multiplication.

[0085] Based on the GRU model, an attention mechanism is introduced, enabling the GRU model to adaptively focus on key information in time-series samples. The Bahdanau attention mechanism is employed for the first... Hidden state at each time step Calculate its attention weights with respect to the hidden states at all time steps:

[0086] ;

[0087] ;

[0088] ;

[0089] in, Indicates the first The time step for the first Attention score at each time step These are learnable parameters for the attention mechanism. Indicates the first The hidden state vector at each time step, For bias vectors, For the first The time step for the first Attention weights at each time step For the first The context vector at each time step. The total number of time steps. This represents an exponential function with the natural constant as its base.

[0090] context vector With the current hidden state Concatenation, as the final feature representation: ;in, This indicates a vector concatenation operation.

[0091] Representing the features of the last time step The data is passed into a fully connected layer, and the cell failure probability is obtained using the Sigmoid activation function. ;in, The weight matrix of the output layer. This is the bias vector.

[0092] The mean squared error (MSE) is used as the loss function to measure the difference between the predicted and the actual values.

[0093] The Adam optimization algorithm is used to train the model, adaptively adjusting the learning rate to accelerate convergence; the update formula of the Adam optimization algorithm is as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] in, For the first First-moment estimate of the gradient at each time step For the first First-moment estimate of the gradient at each time step For the first Second-moment estimation of the gradient at each time step. For the first Second-moment estimation of the gradient at each time step. It is the correction value for the first moment estimate of the gradient. It is the correction value for the second moment estimate of the gradient. The attenuation rate, For the first The gradient at each time step For learning rate, For smoothing terms, For the first Model parameters at each time step, For the first Model parameters at each time step.

[0100] The hyperparameters of the model were tuned using cross-validation, including: the number of hidden layers in the GRU model, the number of units per layer, the dimension of the attention mechanism, the number of fully connected layers, the number of units per layer, the learning rate, the batch size, and the regularization coefficient.

[0101] By tuning hyperparameters, the optimal model parameter configuration is obtained for subsequent prediction tasks.

[0102] Step 4: Extend the cell failure prediction model to a multi-task learning framework. While predicting cell failure, predict the state of health (SOH) and failure time, and introduce an adaptive failure threshold mechanism.

[0103] While predicting cell failure, the system also introduces battery state of health (SOH) prediction and failure time prediction.

[0104] Battery State of Health (SOH) Prediction: SOH is defined as the ratio of the battery's current capacity to its initial capacity, ranging from [0, 1]. SOH prediction is modeled as a regression problem, with the loss function being the mean squared error (MSE).

[0105] ;

[0106] in, It is the mean square error (MSE) value. For the first Predicted State of Health (SOH) values ​​for each time series sample. For the first The true state of health (SOH) values ​​of the battery for each time series sample.

[0107] Failure time prediction: Failure time is defined as the remaining number of cycles the battery has from the current time to the failure time, with a value ranging from [0, +∞). Failure time prediction is modeled as a regression problem, with the loss function being the mean absolute error (MAE).

[0108] ;

[0109] in, Mean Absolute Error (MAE) value For the first Predicted failure time for each time series sample. For the first The actual failure time value of each time series sample.

[0110] Based on the existing GRU model, two auxiliary task output branches are introduced into the GRU model; the feature representations of the temporal samples are extracted through the GRU encoder in the GRU model, and then mapped to the outputs of different tasks through task-specific fully connected layers.

[0111] The battery health state (SOH) prediction branch in the GRU model is represented as follows: The failure time prediction branch in the GRU model is represented as follows: ;in, This is the hidden state of the last time step of the GRU encoder. These are the weight vectors for the State of Health (SOH) and Time to Failure (TFF) prediction branches, respectively. For bias terms, These are the predicted values ​​for battery health status and failure time, respectively.

[0112] Joint loss function for: ;in, Let the cross-entropy loss function be used for the cell failure prediction task. The weighted hyperparameters are used to balance the losses of the three tasks.

[0113] An adaptive failure threshold mechanism is introduced to dynamically adjust the cell failure probability threshold based on the predicted battery health status and failure time.

[0114] Define cell failure probability threshold for: ;in, The weight vector calculated for the threshold. For bias terms, The Sigmoid activation function maps the threshold to the range (0, 1).

[0115] To encourage the GRU model to learn a suitable failure probability threshold, a threshold penalty term is introduced into the loss function: ;in, For the preset target threshold, The weighted hyperparameters are used to balance the threshold penalty term with the losses from other tasks.

[0116] Add a threshold penalty term to the joint loss function:

[0117] ;

[0118] In the prediction phase, the cell failure probability is compared with a dynamic threshold to obtain the final failure prediction result.

[0119] ;

[0120] in, This is the predicted probability of cell failure. This represents the probability of cell failure.

[0121] By employing a multi-task learning framework and an adaptive failure threshold mechanism, the accuracy and interpretability of cell failure prediction can be improved by fully utilizing SOH and failure time information.

[0122] Step 5: To address the new operating conditions generated during cell operation, construct an incremental learning framework based on sample selection, and adjust the cell failure prediction model using samples from the new operating conditions.

[0123] When new operating conditions arise during the operation of the battery cell, representative samples are selected from the data collected under these new conditions to form a new operating condition sample set. ,in For the first The feature vector of a new working condition sample For the corresponding tags.

[0124] Randomly selected from the new operating condition data A new working condition sample set is formed from a set of samples; diversity selection: samples that are significantly different from the existing training samples are selected. For each sample, metrics such as Euclidean distance or cosine similarity can be used.

[0125] Let the parameters of the pre-trained model be... The goal of incremental learning is to fine-tune the model parameters on new working condition samples to obtain updated model parameters. The loss function for incremental learning is:

[0126] ;

[0127] in, The loss function for incremental learning, This is the sample number for the new operating condition. , For the model at input The predicted output on, For incremental learning tasks, the loss function is (e.g., cross-entropy loss or mean squared error loss). This is the regularization coefficient, used to balance the magnitude of new sample fitting and parameter updates.

[0128] The optimization objective of the incremental learning objective function is: , Indicates taking The value at its smallest value.

[0129] The incremental learning objective function is optimized using the mini-batch gradient descent algorithm. Input: pre-trained model parameters. New working condition sample set Learning rate Regularization coefficient Small batch size Number of iterations Output: Updated model parameters .

[0130] initialization: , ,from Random selection A small batch of samples consists of 10 samples. Calculate the loss function on mini-batch. :

[0131] ;

[0132] Calculate the gradient of the loss function: ; This indicates that the loss function is applied to the parameters. The gradient;

[0133] Update model parameters: , , This represents the parameter value updated in the current iteration step. The parameter values ​​of the previous iteration step in the current iteration step. Indicates the final iteration step The parameter value after that.

[0134] The incremental learning framework can effectively adapt the pre-trained cell failure prediction model to new operating conditions, thereby improving the model's generalization performance and practical application value.

[0135] Example 2

[0136] Referring to Figure 4, the second embodiment of this application provides a deep learning device for predicting cell failure in a BMS, including: a data acquisition module, a feature extraction module, a GRU model module, a function extension module, and an incremental learning module.

[0137] The data acquisition module is used to collect multi-mode cell operation data, preprocess the collected multi-mode cell operation data, and divide it into time-series samples using a sliding window method.

[0138] The feature extraction module is used to extract the time-domain features, frequency-domain features, and difference features of the time-series samples, perform feature selection using Pearson correlation coefficient and mutual information, and use principal component analysis to perform dimensionality reduction to construct a dataset.

[0139] The GRU model module is used to construct a cell failure prediction model based on the gated cyclic unit (GRU) model. The GRU model is trained using dataset data, an attention mechanism is added to the GRU model, the GRU model is trained, and the model parameters are optimized.

[0140] The functional extension module is used to extend the cell failure prediction model, predicting the state of health (SOH) and failure time of the battery while predicting cell failure, and introducing an adaptive failure threshold mechanism.

[0141] The incremental learning module constructs an incremental learning framework based on sample selection to address new operating conditions generated during battery cell operation, and adjusts the battery cell failure prediction model using new operating condition samples.

[0142] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of this application without departing from the spirit and scope of protection of the claims. All of these variations are within the protection scope of this application.

Claims

1. A deep learning method for cell failure prediction in BMS, characterized in that, include: Multi-mode cell operation data is collected, and the collected multi-mode cell operation data is preprocessed and divided into time-series samples using a sliding window method. The time-domain, frequency-domain, and difference features of the time-series samples are extracted. Feature selection is performed using Pearson correlation coefficient and mutual information, and dimensionality reduction is achieved through principal component analysis to construct a dataset. A cell failure prediction model based on a gated recurrent unit (GRU) model is constructed. The GRU model is trained using the dataset, and an attention mechanism is added to the GRU model. The model parameters are then optimized. The input layer of the GRU model receives the feature vectors of the time-series samples, and the output layer outputs the cell failure probability. Multiple GRU hidden layers are set between the input and output layers of the GRU model, with multiple GRU units in each hidden layer. An attention mechanism is introduced into the constructed GRU model to make it focus on the information of the time-series samples. The hyperparameters of the GRU model are tuned using cross-validation. The cell failure prediction model is extended to predict battery health status and failure time simultaneously with cell failure, and an adaptive failure threshold mechanism is introduced. Battery health status is defined as the ratio of the battery's current capacity to its initial capacity, and failure time is defined as the number of remaining cycles from the current time to the time of failure. Battery health status prediction and failure time prediction are modeled as regression problems. Based on the existing GRU model, output branches for battery health status prediction and failure time prediction are introduced. Feature representations of time-series samples are extracted through the GRU encoder in the GRU model and then mapped to the outputs of different tasks through a fully connected layer. An adaptive failure threshold mechanism is introduced to dynamically adjust the cell failure probability threshold according to the battery health status prediction value and failure time prediction value. For new operating conditions generated during cell operation, an incremental learning framework based on sample selection is constructed to adjust the cell failure prediction model through new operating condition samples.

2. The deep learning method for cell failure prediction in BMS according to claim 1, characterized in that, By installing multiple sensors on the battery cell, multi-mode battery cell operation data, including voltage, current, temperature, and pressure, is collected during the battery cell's operation. The collected multi-mode battery cell operation data is preprocessed, including outlier detection and removal. The preprocessed data is then time-aligned and normalized. Finally, a sliding window method is used to divide the normalized multi-mode battery cell operation data into time-series samples.

3. The deep learning method for cell failure prediction in BMS according to claim 2, characterized in that, For each time series sample, feature extraction and computation are performed, including: extracting time-domain features, extracting frequency-domain features, and calculating difference features; using Pearson correlation coefficient and mutual information to select features from the extracted and computed time-domain features, frequency-domain features, and difference features; performing principal component analysis to reduce the dimensionality of the selected features to obtain the feature matrix; and combining the dimensionality-reduced feature matrix with the corresponding label vector to form the dataset.

4. The deep learning method for cell failure prediction in BMS according to claim 3, characterized in that, An incremental learning framework based on sample selection is constructed. By selecting new operating condition samples, the parameters of the pre-trained cell failure prediction model are adjusted. The selection strategies for new operating condition samples include random selection and diversity selection. An incremental learning objective function is constructed. By balancing the fitting of new operating condition samples and the parameter update magnitude, the updated model parameters are obtained. The incremental learning objective function is optimized using a mini-batch gradient descent algorithm to obtain a cell failure prediction model adapted to the new operating conditions.

5. A deep learning device for predicting cell failure in a BMS, used to implement the deep learning method for predicting cell failure in a BMS as described in any one of claims 1 to 4, characterized in that, include: The module includes a data acquisition module, a feature extraction module, a GRU model module, a function extension module, and an incremental learning module. The data acquisition module is used to collect multimodal cell operation data, preprocess the collected multimodal cell operation data, and divide it into time-series samples using a sliding window method. The feature extraction module is used to extract the time-domain features, frequency-domain features, and difference features of the time-series samples, perform feature selection using Pearson correlation coefficient and mutual information, and perform dimensionality reduction using principal component analysis to construct a dataset. The GRU model module is used to construct a cell failure prediction model based on a gated recurrent unit (GRU) model, train the GRU model using the dataset data, add an attention mechanism to the GRU model, train the GRU model, and optimize the model parameters. The function extension module is used to extend the cell failure prediction model, predicting battery health status and failure time while predicting cell failure, and introducing an adaptive failure threshold mechanism. The incremental learning module constructs an incremental learning framework based on sample selection for new operating conditions generated during cell operation, and adjusts the cell failure prediction model using new operating condition samples.

6. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the deep learning method for cell failure prediction in a BMS according to any one of claims 1 to 4 by calling the computer program stored in the memory.

7. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform a deep learning method for cell failure prediction in a BMS as described in any one of claims 1 to 4.

Images