Search

CN-122017583-A - CNN-LSTM-SE battery thermal runaway early warning method and system based on transfer learning

CN122017583ACN 122017583 ACN122017583 ACN 122017583ACN-122017583-A

Abstract

The invention belongs to the technical field of power lithium battery application, and in particular relates to a CNN-LSTM-SE battery thermal runaway early warning method and system based on transfer learning, wherein the method comprises the steps of obtaining a source domain data set and a target domain data set, and carrying out standardized processing to obtain a source domain standardized data sequence and a target domain standardized data sequence; the method comprises the steps of constructing a CNN-LSTM-SE neural network model according to a source domain standardized data sequence, pre-training the CNN-LSTM-SE neural network model to obtain a source domain model, fine-adjusting weight parameters of the source domain model by utilizing a target domain standardized data sequence to obtain a target domain early-warning model, carrying out regression prediction on the target domain standardized data sequence according to the target domain early-warning model to obtain a predicted value, taking the difference value between the predicted value and a true value as a predicted error, comparing the predicted error with a preset multi-level threshold to generate an early-warning signal, and realizing accurate early warning when a battery is in thermal runaway failure.

Inventors

  • MA JIN
  • XU MING
  • DU WEI

Assignees

  • 国网山西省电力有限公司长治供电分公司

Dates

Publication Date
20260512
Application Date
20260129

Claims (10)

  1. 1. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning is characterized by comprising the following steps of: s1, acquiring a source domain data set and a target domain data set, and slicing and standardizing the source domain data set and the target domain data set to respectively obtain a source domain standardized data sequence and a target domain standardized data sequence; S2, constructing a CNN-LSTM-SE neural network model based on the source domain standardized data sequence, wherein the constructed CNN-LSTM-SE neural network model is provided with a full-connection layer, a parallel feature extraction layer, a feature fusion layer and a double-target prediction output layer, the parallel feature extraction layer comprises a CNN layer and a bidirectional LSTM layer which are arranged in parallel, and a mixed attention module is fused behind the bidirectional LSTM layer and consists of a channel attention module and a time sequence attention module, wherein the feature fusion layer comprises a MTLA attention module; S3, dividing a source domain standardized data sequence into source domain training data and source domain verification data according to a ratio of 7:3, pre-training the CNN-LSTM-SE neural network model through the source domain training data, verifying the CNN-LSTM-SE neural network model through the source domain verification data immediately after each iteration round is finished, monitoring the change trend of source domain training loss and source domain verification loss, automatically stopping a pre-training program of the CNN-LSTM-SE neural network model after a plurality of iteration rounds, and judging that the pre-training of the CNN-LSTM-SE neural network model is finished, so that the CNN-LSTM-SE neural network model learns and extracts the voltage, the temperature, the change rate of the voltage and the time-space characteristics of the change rate of the temperature of a battery to obtain a source domain model; S4, freezing a full connection layer, a parallel feature extraction layer and a feature fusion layer in the source domain model based on transfer learning, taking data before 1500S in a target domain standardized data sequence as target domain training data, taking data after 1500S in the target domain standardized data sequence as target domain verification data, and utilizing the target domain training data to finely tune the weights of a time sequence attention module, a MTLA attention module and a double-target prediction output layer in the source domain model, wherein after each fine tuning is completed for one iteration round, the source domain model is verified by the target domain verification data immediately, meanwhile, the change trend of target domain training loss and target domain verification loss is monitored, and after a plurality of iteration rounds, the change amplitude of the target domain training loss and the target domain verification loss is smaller than a preset threshold value 1e-5, and the time sequence attention module, the MTLA attention module and the double-target prediction output layer which are judged to be the source domain model are completed, so that a target domain early warning model is obtained; S5, carrying out regression prediction on the temperature and the voltage of the battery in the target domain standardized data sequence by using the target domain early warning model to obtain predicted values of the future time step temperature and the voltage; And S6, comparing the temperature prediction error and the voltage prediction error with preset multilevel temperature and voltage error thresholds, judging the battery state and generating corresponding early warning signals.
  2. 2. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 1 is characterized in that in the step S1, the source domain data set is voltage data, voltage change rate data, temperature data and temperature change rate data acquired during normal operation of a vehicle, the source domain data set is sliced and standardized to obtain a source domain standardized data sequence, the source domain standardized data sequence is divided according to the ratio of source domain training data to source domain verification data 7:3, the target domain data set is obtained through a single battery thermal runaway experiment and comprises temperature data and voltage data of a battery before and after thermal runaway, the acquired temperature data and voltage data are filled and calculated, and parameters of voltage, voltage change rate, temperature and temperature change rate are obtained.
  3. 3. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 2 is characterized in that when the source domain data set and the target domain data set are sliced, the length of a time window is set to N time steps, continuous subsequences are extracted in a sliding window mode, each subsequence comprises four characteristics of voltage, voltage change rate, temperature and temperature change rate, a Z-score method is adopted for standardization processing of the source domain data set and the target domain data set, and a calculation formula is as follows: where x is the raw data collected, μ and σ are the mean and standard deviation of the data features, respectively, and z is the normalized data.
  4. 4. The CNN-LSTM-SE battery thermal runaway warning method based on transfer learning according to claim 3, wherein in step S2, the CNN-LSTM-SE neural network model includes a full connection layer, a parallel feature extraction layer, a feature fusion layer and a dual-target prediction output layer; inputting the source domain standardized data sequence obtained by pretreatment into the full-connection layer, wherein the number of neurons arranged on the full-connection layer is 20, introducing nonlinear transformation by using the ReLU activation function, and converting the data into space-time data consisting of a batch size B, a time step T and a characteristic dimension F; the parallel feature extraction layer is formed by connecting a CNN layer and a bidirectional LSTM layer in parallel, wherein the bidirectional LSTM layer is used for extracting the time features of the space-time data, the bidirectional LSTM layer is fused with a mixed attention module, the mixed attention module is used for enabling the CNN-LSTM-SE neural network model to dynamically focus on the key information of the space-time data in time, and the CNN layer is used for extracting the space features of the space-time data; The bidirectional LSTM layer expands the space-time data obtained by processing the full-connection layer according to time dimension and is used for acquiring voltage, temperature, voltage change rate and history and future time sequence information of the temperature change rate; the bidirectional LSTM layer is provided with four LSTM units to form a bidirectional and stacked combined structure, the four LSTM units comprise a forward 1 st layer LSTM unit, a forward 2 nd layer LSTM unit, a backward 1 st layer LSTM unit and a backward 2 nd layer LSTM unit, wherein the forward 1 st layer LSTM unit and the forward 2 nd layer LSTM unit are used for extracting forward time sequence characteristics of battery operation parameters from the past to the current time step, and the backward 1 st layer LSTM unit and the backward 2 nd layer LSTM unit are used for extracting reverse time sequence characteristics from the future to the current time step so as to capture historical dependence and future trend of battery states at the same time; the number of hidden units of each LSTM unit is 2, each LSTM unit generates a hidden layer state at each time step, the hidden layer state of each time step is output as a two-dimensional vector, the two-dimensional vector is used for representing the electric characteristic and the thermal characteristic in the battery operation process, the first unit in the hidden units is used for learning the time dependence of the electric characteristic and comprises a dynamic change rule of voltage and voltage change rate, the second unit in the hidden units is used for learning the time dependence of the thermal characteristic and comprises a time evolution characteristic of temperature and temperature change rate, and the two-way LSTM layer acquires two hidden layer states with opposite time sequences through the forward 1 st layer LSTM unit, the forward 2 nd layer LSTM unit, the backward 1 st layer LSTM unit and the backward 2 nd layer LSTM unit, and the output dimension is a time feature vector comprising a batch size B, a time step T and a time feature dimension F1; the mixed attention module comprises a channel attention module and a time sequence attention module which are arranged in parallel, wherein the time sequence attention module is provided with a self-adaptive attention network, the mixed attention module is added behind the bidirectional LSTM layer, adopts a mixed attention mechanism, is connected with the self-adaptive time sequence attention module for extracting time sequence characteristics in parallel on the basis of the channel attention module, and enables the bidirectional LSTM layer to pay attention to the importance of the feature dimension and the dynamic change of the time dimension, and enables the bidirectional LSTM layer to be self-adaptively endowed with higher weight of key time sequence information; The CNN layer is used for extracting local mode characteristics of voltage, temperature and voltage change rate and temperature change rate in space-time data obtained by processing the full-connection layer in the space dimension; the CNN layer comprises two one-dimensional convolution layers, wherein the two one-dimensional convolution layers respectively extract the spatial characteristics of voltage and the spatial characteristics of temperature, each one-dimensional convolution layer in the CNN layer is provided with 7 convolution kernels, the coverage area of the convolution kernels is 1 multiplied by 7, 7 convolution kernel weight parameters are formed, the step size is 1, the filling mode is SAME, the spatial dimension of output data of each one-dimensional convolution layer is consistent with the spatial dimension of space-time data processed by a fully connected layer, each one-dimensional convolution layer is connected with a BN module, the two one-dimensional convolution layers in the CNN layer adopt one-dimensional convolution structures with a local perception mechanism and a convolution kernel weight sharing mechanism, the local perception mechanism only extracts characteristics in the local range of input space-time data by setting the coverage area of the convolution kernels, captures the mutation modes of voltage, temperature and change rate of the voltage and the temperature in a short time sequence, the SAME convolution kernel weight is repeatedly used at different positions of the whole time sequence to form a convolution kernel weight sharing mechanism; The feature fusion layer performs weighted combination on time information extracted by the bidirectional LSTM layer and space information extracted by the CNN layer, performs dimensional splicing on time feature vectors output by the bidirectional LSTM layer and space feature vectors output by the CNN layer, outputs space-time feature vectors with the size of batch B, time step T and the sum of the time feature dimension and the space feature dimension F1+F2, introduces MTLA attention module into the feature fusion layer, distributes weight to the spliced space-time feature vectors according to the time step and the feature dimension, wherein the scaling factor of the MTLA attention module is that Embedding a corrected full-connection layer at the tail end of the characteristic fusion layer, wherein the corrected full-connection layer has an activation function of LeakyReLU and a slope of 0.2, carrying out weighted average on the space-time fusion characteristics output by the MTLA attention module through the corrected full-connection layer, mapping the weighted space-time fusion characteristics to target dimensions to form voltage high-dimensional characteristics and temperature high-dimensional characteristics; the dual-target prediction output layer is used for outputting the voltage and temperature predicted values of the future time steps, and the voltage high-dimensional features and the temperature high-dimensional features output by the feature fusion layer are mapped into the dual-target predicted values for predicting the voltage and the temperature of the future time steps through the dual-target prediction output layer.
  5. 5. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 4 is characterized in that in a mixed attention module, the channel attention module comprises a one-dimensional average pooling branch and a one-dimensional maximum pooling branch, the channel attention module extracts global and local information of each characteristic channel, the global and local information is processed through two layers of fully-connected networks and then fused to obtain channel attention weights, the time sequence attention module is provided with an adaptive time weight computing network, the time sequence attention weights are obtained by distributing weights to input features of different time steps through a Softmax function, and the time sequence attention weights and the channel attention weights are fused to carry out weighting processing on the input features.
  6. 6. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 5 is characterized in that in step S3, a CNN-LSTM-SE neural network model is pre-trained through a source domain standardized data sequence, the pre-training process comprises the steps of taking voltage, voltage change rate, temperature and temperature change rate of a sliced battery as input, predicting temperature and voltage of a future time step, adopting a mean square error as a loss function, adopting an Adam optimizer to update network parameters until a mean square error loss value is not reduced, wherein parameters in the CNN-LSTM-SE neural network model are converged and stabilized, and mapping and storing the parameters in a weight matrix and bias parameters of the CNN-LSTM-SE neural network model through the Adam optimizer, so that the obtained source domain model has the capability of extracting and expressing general space-time characteristics of the battery.
  7. 7. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 6, wherein the specific steps of step S4 are as follows: s41, loading the weight parameters of the source domain model obtained by training in the step S3, freezing the parameters of the full-connection layer, the bidirectional LSTM layer, the CNN layer and the feature fusion layer in the source domain model, and simultaneously setting the weight parameters of the time sequence attention module and the MTLA attention module in the feature fusion layer in the dual-target prediction output layer and the parallel feature extraction layer as trainable states; S42, taking a target domain standardized data sequence as input, setting super parameters, wherein the number of samples is 128, the initial learning rate is 0.0001, iteration is 20 to 30, mean square error is adopted as a loss function, and an Adam optimizer is utilized to iteratively update the weight parameters of a double-target prediction output layer, a time sequence attention module and a MTLA attention module until convergence is achieved under the condition that parameters of a full-connection layer, a bidirectional LSTM layer, a CNN layer and a feature fusion layer in a parallel feature extraction layer are kept unchanged, so that a target domain early warning model capable of representing the running state features of a target domain battery is obtained.
  8. 8. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 7 is characterized in that in step S5, regression prediction is carried out on the temperature and the voltage of target domain data by using the target domain early warning model obtained in step S4, the process comprises the steps of taking a target domain data slice as input of the target domain early warning model, directly obtaining temperature and voltage predicted values of future time steps at a double-target predicted output layer of the target domain early warning model after processing of a fully-connected layer, a parallel feature extraction layer and a feature fusion layer of the target domain early warning model, measuring differences between the predicted values and actual acquired temperature and voltage actual values by means of a mean square error loss function, and calculating differences between the predicted values and actual values as predicted errors.
  9. 9. The CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning according to claim 8, wherein the multi-stage temperature and voltage error thresholds include a primary temperature error threshold and a primary voltage error threshold, a secondary temperature error threshold and a secondary voltage error threshold, and a tertiary temperature error threshold and a tertiary voltage error threshold, and according to the error between the predicted value and the actual collected value, the error is firstly compared with the temperature threshold, and then compared with the voltage threshold, and according to the comprehensive comparison result of the error and the threshold, the battery state is judged to be normal, abnormal, about to fail or a failure state, and a monitoring instruction or an alarm signal is correspondingly sent.
  10. 10. CNN-LSTM-SE battery thermal runaway early warning system based on migration study, characterized by comprising: The data acquisition and preprocessing module is used for acquiring voltage and temperature data of the source domain and the target domain, slicing and standardized preprocessing the source domain data and the target domain data to obtain a standardized data sequence of the source domain and a standardized data sequence of the target domain; The model construction module is used for constructing a CNN-LSTM-SE neural network model based on a source domain standardized data sequence, and the CNN-LSTM-SE neural network model sequentially comprises a full-connection layer, a parallel characteristic extraction layer, a characteristic fusion layer and a double-target prediction output layer; The model pre-training module is used for pre-training the CNN-LSTM-SE neural network model by utilizing a source domain standardized data sequence to obtain a source domain model; the migration learning module is used for carrying out fine tuning training on the source domain model by utilizing a target domain standardized data sequence based on migration learning to obtain a target domain early warning model; The prediction and error calculation module is used for predicting the temperature and the voltage of the target domain standardized data sequence based on the target domain early warning model and calculating a prediction error; And the early warning judgment module is used for comparing the prediction error with a preset multi-stage temperature and voltage error threshold value and outputting a corresponding early warning signal according to a comparison result.

Description

CNN-LSTM-SE battery thermal runaway early warning method and system based on transfer learning Technical Field The invention belongs to the technical field of application of power lithium batteries, and particularly relates to a CNN-LSTM-SE battery thermal runaway early warning method and system based on transfer learning. Background The power lithium battery is widely applied to electric automobiles and energy storage power stations due to the advantages of high energy density, long cycle life and the like, however, under extreme working conditions such as electric abuse, mechanical needling, thermal abuse and the like, the internal temperature of the battery can be rapidly increased and trigger thermal runaway to cause serious accidents such as explosion, combustion and the like, and in recent years, the thermal runaway event of an energy storage system frequently occurs to cause casualties and property loss, so that the method for monitoring the running state of the battery in real time and early warning in advance is an important technical direction for guaranteeing the safety of the battery. The existing battery thermal runaway prediction method mainly depends on a data driving model, a machine learning or deep learning algorithm is used for extracting rules from operation data, however, in practical application, the problems of insufficient data acquisition, data missing and early warning response lag exist, in order to save energy consumption, a plurality of battery management systems start data recording only when parameters such as voltage, temperature and the like exceed a threshold value, continuous complete data is lacked under stable working conditions, the storage space of an embedded sensor and the data transmission flow are limited, normal operation data are covered or ignored, parameter fluctuation is small during normal charging and discharging, sampling frequency is low, so that training data is scarce, meanwhile, the traditional method based on threshold value judgment only alarms when the parameters exceed a set value, the time sequence characteristics of battery operation cannot be fully utilized to discover abnormality in advance, and the method cannot adapt to different types of batteries and changeable operation conditions, so that false alarm or missing alarm is caused. Disclosure of Invention The invention aims to provide a CNN-LSTM-SE battery thermal runaway early warning method and system based on transfer learning, which effectively solve the problems of untimely prediction, false alarm, delayed alarm and the like of an early warning system in the prior art. The technical scheme of the invention is that the CNN-LSTM-SE battery thermal runaway early warning method based on transfer learning comprises the following steps: s1, acquiring a source domain data set and a target domain data set, and slicing and standardizing the source domain data set and the target domain data set to respectively obtain a source domain standardized data sequence and a target domain standardized data sequence; S2, constructing a CNN-LSTM-SE neural network model based on the source domain standardized data sequence, wherein the constructed CNN-LSTM-SE neural network model is provided with a full-connection layer, a parallel feature extraction layer, a feature fusion layer and a double-target prediction output layer, the parallel feature extraction layer comprises a CNN layer and a bidirectional LSTM layer which are arranged in parallel, and a mixed attention module is fused behind the bidirectional LSTM layer and consists of a channel attention module and a time sequence attention module, wherein the feature fusion layer comprises a MTLA attention module; S3, dividing a source domain standardized data sequence into source domain training data and source domain verification data according to a ratio of 7:3, pre-training the CNN-LSTM-SE neural network model through the source domain training data, verifying the CNN-LSTM-SE neural network model through the source domain verification data immediately after each iteration round is finished, monitoring the change trend of source domain training loss and source domain verification loss, automatically stopping a pre-training program of the CNN-LSTM-SE neural network model after a plurality of iteration rounds, and judging that the pre-training of the CNN-LSTM-SE neural network model is finished, so that the CNN-LSTM-SE neural network model learns and extracts the voltage, the temperature, the change rate of the voltage and the time-space characteristics of the change rate of the temperature of a battery to obtain a source domain model; S4, freezing a full connection layer, a parallel feature extraction layer and a feature fusion layer in the source domain model based on transfer learning, taking data before 1500S in a target domain standardized data sequence as target domain training data, taking data after 1500S in the target domain standardized data sequence