A method and prediction system for constructing a thermal runaway prediction model for lithium-ion batteries

Abstract

The present invention discloses a method and a prediction system for constructing a thermal runaway prediction model for a lithium-ion battery. The method comprises the following steps: conducting a thermal runaway experiment on a lithium-ion battery, collecting battery sequence feature data including temperature, voltage and current, as well as sound signals and thermal imaging data; slicing the data to generate historical data; recording the battery state as y, and dividing the thermal runaway state y=1 and the non-thermal runaway state y=0 according to the temperature threshold; taking the historical data as input, taking the battery thermal runaway abnormal state as a label, and extracting data features; calculating the attention weights of all data features, and performing feature fusion; obtaining the thermal runaway state of the lithium-ion battery using a classifier; dividing the historical data into a training set and a test set and inputting them into the model for training and verification, and judging the model accuracy with the accuracy rate as the evaluation index, thereby constructing a thermal runaway prediction model. The multimodal model constructed by the present invention improves the accuracy of thermal runaway prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery safety prediction and identification, and specifically relates to a method and a prediction system for constructing a thermal runaway prediction model for a lithium-ion battery. Background Art

[0002] Lithium-ion batteries have been widely used as clean energy in electric vehicles and electronic devices due to their high energy density, long life and small size. However, in practical applications, lithium-ion batteries will encounter some abnormal conditions, including mechanical abuse, electrical abuse and thermal abuse. If a lithium-ion battery has thermal runaway, it will cause a serious accident. Therefore, battery thermal runaway warning is a safety issue that needs to be solved urgently.

[0003] At present, many researchers explore the internal reaction mechanism and external characteristics of lithium-ion batteries during thermal runaway based on experiments or simulations, and analyze the thermal runaway process. The equation-based method involves high computational complexity. The thermal runaway process of lithium-ion batteries is derived through mathematical formulas, and finally the state equation of thermal runaway of the battery is obtained. However, if the lithium-ion battery is in a relatively complex environment, it is difficult to derive the thermal runaway model. With the development of machine learning and deep learning, data-driven methods have overcome these problems. Only by focusing on the characteristic data such as voltage, current and temperature of lithium-ion batteries during the entire thermal runaway process, the abnormal state of the battery can be diagnosed. Voltage, current and temperature are all sequence data, which can only reflect the changes in these sequence data caused by abnormal reactions inside the battery when thermal runaway occurs, and cannot be predicted in advance before thermal runaway is reached. The patent document with publication number CN113344024A discloses a lithium-ion battery thermal runaway graded warning method and warning system. This method uses fault-free battery feature elements to train the thermal runaway prediction ability of the prediction model composed of long short-term memory network, time convolution network and GRU neural network; then the battery feature elements of the monitored battery in the time series are input into the prediction model to obtain the predicted value under normal state at a certain moment, and then compared with the actual value corresponding to the collected moment. The greater the difference between the two, the greater the risk of thermal runaway of the battery. However, since normal batteries are used for prediction, there is a large deviation from the actual thermal runaway state.

[0004] In addition to the characteristics of the battery itself, there may be sounds and abnormal heating in the battery before thermal runaway. These characteristics may appear earlier than abnormal measurement data such as voltage, current and temperature. Therefore, it is necessary to further comprehensively consider various abnormal signals and improve the prediction method of battery thermal runaway. Summary of the invention

[0005] In view of this, one of the objectives of the present invention is to provide a method and a prediction system for constructing a thermal runaway prediction model for a lithium-ion battery.

[0006] The technical solution is as follows:

[0007] A method for constructing a thermal runaway prediction model for a lithium-ion battery, the key of which is to collect thermal runaway experimental data and use the experimental data to construct a thermal runaway prediction model based on a neural network, the thermal runaway prediction model includes a feature extractor, a feature fusion device and a classifier, and the method specifically includes the following steps:

[0008] S1. Thermal runaway test data collection: Conduct thermal runaway tests on lithium-ion batteries and collect battery sequence feature data, as well as sound signals and thermal imaging data;

[0009] The battery sequence characteristic data includes temperature data, voltage data and current data;

[0010] S2, data preprocessing: slicing the data collected in step S1 to generate historical data;

[0011] The battery state is recorded as y, and a temperature threshold is set. If the temperature exceeds the threshold, thermal runaway occurs, y=1; otherwise, thermal runaway does not occur, y=0;

[0012] S3, extracting feature data: taking the data preprocessed in step S2 as input, taking the abnormal state of thermal runaway of the lithium-ion battery as a label, and using the feature extractor to extract data features;

[0013] S4, feature fusion: using the feature fuser to calculate the attention weights of all the data features and perform feature fusion;

[0014] S5, thermal runaway state judgment: using the classifier to obtain the thermal runaway state of the lithium-ion battery;

[0015] S6, model training: divide the historical data in step S2 into a training set and a test set, input the training set into the thermal runaway prediction model for training, input the test set for verification, and judge the accuracy of the model with the accuracy rate as the evaluation index, so as to update the thermal runaway prediction model.

[0016] Preferably, the specific process of slicing in the above step S2 is: slice the data collected in step S1 using a sliding window of size T to generate historical voltage data X V , current data X I , temperature data X T , sound dataX S and thermal imaging dataX P .

[0017] Preferably, the above feature extractor includes a long short-term memory neural network (LSTM), a one-dimensional convolutional neural network (1D CNN), and a ResNet50;

[0018] In step S3, the long short-term memory neural network (LSTM) is used to extract the feature vectors of the voltage, current, and temperature sequence data, denoted as h V ∈R N×m , h I ∈R N×m , and h T ∈R N×m ;

[0019] The one-dimensional convolutional neural network (1D CNN) is used to extract the feature vector of the sound signal, denoted as h S ∈R N×m ;

[0020] The ResNet50 is used to extract the feature vector of the thermal imaging, denoted as h P ∈R N×m .

[0021] Preferably, in step S3 above, the features of the voltage, current, and temperature sequence data are calculated and extracted according to equations (1) to (6),

[0022] g t =σ(W f ·[h t-1 , x t ) + b g (1)

[0023] i t =σ(W i ·[h t-1 , x t ) + b i (2)

[0024]

[0025]

[0026] o t =σ(W o ·[h t-1 , x t ) + b o (5)

[0027] h t = o t *tanh(S t ) (6)

[0028] where g t is the forgetting gate, it is the input gate, o t is the output gate, σ is the activation function, W is the weight matrix, initialized with normal distribution, and b is the bias, initialized to 0;

[0029] h t is the characteristic of voltage or current or temperature series data;

[0030] When extracting voltage series data features, input x t ∈X V , when extracting the features of current series data, input x t ∈X I , input x when extracting temperature series data features t ∈X T ;

[0031] and S t It is the intermediate value during the operation.

[0032] Preferably, the feature fusion device is used for concatenation, weight calculation and fusion of feature vectors;

[0033] The specific process of step S4 is to first transform the feature vector h V 、h I 、h T 、h S and h P Concatenate into feature matrix h∈R N ×5×m ;

[0034] The feature matrix h∈R N×5×m Input to the fully connected layer and multiply it by the trainable weight matrix W Q , W K and W T , calculate the query matrix Q, key matrix K and value matrix V;

[0035] Then according to the attention operation function

[0036]

[0037] Calculate the attention score matrix α = [α 1 ,α 2 ,α 3 ,α 4 ,α 5 ]∈R N×5×5 , where α 1 , α 2 , α 3 , α 4 and α 5They are the attention scores of voltage feature, current feature, temperature feature, sound signal feature, and thermal imaging feature, d is the hidden dimension of the fully connected layer, and K T is the transposed matrix of the key matrix K;

[0038] The attention mechanism is introduced to calculate the attention weight feature matrix H = α·V;

[0039] Finally, the feature fusion is completed by weighting and summing each feature using formula (8).

[0040] H=α 1 ·h 1 +α 2 ·h 2 +α 3 ·h 3 +α 4 ·h 4 +α 5 ·h 5 (8)

[0041] where h is the eigenvector corresponding to α.

[0042] Preferably, in the above step S3, the 1D CNN is composed of a convolution layer, a pooling layer and a fully connected layer, and convolution and pooling operations are performed on the input data through multiple convolution kernels to extract the potential features of the data. Specifically, the sound signal features are calculated and extracted according to formulas (9) to (11).

[0043] y i =f(u i *k+b i ) (9)

[0044] z i (j) = maxy i (k),k∈D j (10)

[0045] h i =W i z i +b i (11)

[0046] Among them, y i represents the feature representation obtained by the i-th filter, k represents the convolution kernel, u represents the input, u∈X S , * represents the convolution operation, f(·) represents the tanh activation function, z i (j) represents the feature of the i-th filter after pooling, D j represents the jth pooling area, y i (k) represents the feature representation of the i-th filter in the pooling kernel, W is the weight matrix, b is the bias, and hi It is the sound signal characteristic.

[0047] Preferably, in the above step S5, a multi-layer fully connected network is used as a classifier to obtain the thermal runaway state of the lithium-ion battery. During the model training process, a corresponding relationship is established between the fused features and the corresponding battery state y in the classifier, and the model is continuously trained using the Adam optimizer until convergence.

[0048] Preferably, the above step S6 specifically includes taking 80% of the data in step S2 as a training set and the remaining 20% ​​as a test set;

[0049] The training set is input into the multimodal model for training, the model training epoch is 3000, the learning rate is 8e-5, and the Adam optimizer is used to obtain a trained thermal runaway prediction model;

[0050] Then input the test set for verification to get the prediction result, and use the accuracy rate ACC as the evaluation index to judge the accuracy of the model.

[0051]

[0052] Among them, TP represents the number of thermal runaway states predicted as thermal runaway, TN represents the number of normal states predicted as normal states, FP represents the number of normal states predicted as thermal runaway, and FN represents the number of thermal runaway states predicted as normal states.

[0053] A second object of the present invention is to provide a lithium-ion battery thermal runaway prediction system.

[0054] A lithium-ion battery thermal runaway prediction system, the key of which is that it includes a data acquisition module, a data transmission module and a thermal runaway prediction module;

[0055] The data acquisition module is used to collect voltage, current, temperature, sound and thermal imaging data of the battery;

[0056] The data transmission module is used to transmit the data collected by the data acquisition module to the thermal runaway prediction module;

[0057] The thermal runaway prediction module includes a memory, a processor, and a data processing mechanism stored in the memory and operable on the processor, wherein the data processing mechanism is used to:

[0058] Inputting the data transmitted by the data transmission module into the thermal runaway prediction model, wherein the thermal runaway prediction model outputs a predicted thermal runaway state of the battery according to the input data;

[0059] The thermal runaway prediction model is constructed and trained and updated according to any one of the above methods.

[0060] Compared with the prior art, the present invention has the following beneficial effects: by collecting various types of data such as battery sequence characteristics, sound signals and thermal imaging data during the thermal runaway experiment, the battery thermal runaway process is considered from different feature angles, different deep learning models are used to mine the potential features of the data, and the attention mechanism is used to calculate the weights between features. The importance of different features to the classification results is considered, a multimodal model is constructed, and the accuracy of thermal runaway prediction is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flowchart of a method for constructing a prediction model according to Embodiment 1;

[0062] Figure 2 It is a structural diagram of embodiment 2. DETAILED DESCRIPTION

[0063] The present invention is further described below in conjunction with embodiments and drawings.

[0064] Embodiment 1

[0065] like Figure 1 A method for constructing a thermal runaway prediction model for a lithium-ion battery, comprising collecting thermal runaway experimental data and using the experimental data to construct a thermal runaway prediction model based on a neural network, wherein the thermal runaway prediction model comprises a feature extractor, a feature fusion device and a classifier, and the method specifically comprises the following steps:

[0066] S1. Thermal runaway test data collection: Conduct thermal runaway tests on lithium-ion batteries and collect battery sequence feature data, as well as sound signals and thermal imaging data;

[0067] The battery sequence characteristic data includes voltage, current and temperature data;

[0068] S2, data preprocessing: slicing the data collected in step S1 to generate historical data;

[0069] The specific process of slicing is: slice the data collected in step S1 using a sliding window of size T to generate historical voltage data X V , current data X I , temperature data X T , sound dataX S and thermal imaging dataX P ;

[0070] The prediction of thermal runaway of lithium-ion batteries is defined as a binary classification task. The battery state is recorded as y. A temperature threshold is set. If the temperature exceeds the threshold, thermal runaway occurs, y=1. Otherwise, thermal runaway does not occur, y=0.

[0071] S3, extracting feature data: taking the data preprocessed in step S2 as input, taking the abnormal state of thermal runaway of the lithium-ion battery as a label, and using the feature extractor to extract data features;

[0072] S4, feature fusion: using the feature fusion device to calculate the attention weights of all the data features and perform feature fusion;

[0073] S5. Thermal runaway state judgment: using a classifier to obtain the thermal runaway state of the lithium-ion battery;

[0074] S6, model training: the historical data in step S2 is divided into a training set and a test set, the training set is input into the thermal runaway prediction model for training, and the test set is input into the thermal runaway prediction model for verification, and the accuracy rate is used as the evaluation index to judge the accuracy of the model, thereby updating the thermal runaway prediction model.

[0075] Since the voltage, current and temperature data are sequence data, which are in different formats from the sound signal and thermal imaging data, the feature extractor includes a long short-term memory neural network (LSTM), a one-dimensional convolutional neural network (1D CNN) and a residual network ResNet50. Therefore:

[0076] In step S3, a long short-term memory neural network (LSTM) is used to extract the characteristic vectors of the voltage, current and temperature sequence data, which are respectively denoted as h V ∈R N×m 、h I ∈R N×m and h T ∈R N×m ;

[0077] LSTM is a special recurrent neural network (RNN) consisting of an input gate, an output gate, a forget gate, and a storage unit. It can solve the gradient vanishing and gradient exploding problems in the long sequence training process and has better performance in long sequence time prediction.

[0078] According to equations (1) to (6), the characteristics of voltage, current and temperature series data are extracted.

[0079] g t =σ(W f ·[h t-1 ,x t ])+b g (1)

[0080] i t =σ(W i ·[h t-1 ,x t ])+b i (2)

[0081]

[0082]

[0083] o t =σ(W o ·[h t-1 ,x t ])+b o (5)

[0084] h t =o t *tanh(S t ) (6)

[0085] Among them, g t It is the forget gate, t is the input gate, o t is the output gate, σ is the activation function;

[0086] W is the weight matrix, which is initialized with normal distribution and continuously updated during the calculation process;

[0087] b is the bias, initialized to 0;

[0088] h t is the characteristic of voltage or current or temperature series data;

[0089] When extracting voltage series data features, input x t ∈X V , when extracting the features of current series data, input x t ∈X I , input x when extracting temperature series data features t ∈X T ;

[0090] and S t It is the intermediate value during the operation.

[0091] In step S3, a one-dimensional convolutional neural network (1D CNN) is used to extract the feature vector of the sound signal, denoted as h S ∈R N×m 1D CNN consists of convolutional layers, pooling layers, and fully connected layers. It uses multiple convolutional kernels to perform convolution and pooling operations on the input data to extract the potential features of the data. Specifically, the sound signal features are extracted according to equations (9) to (11).

[0092] y i =f(u i *k+b i ) (9)

[0093] z i (j) = maxyi (k),k∈D j (10)

[0094] h i =W i z i +b i (11)

[0095] Among them, y i represents the feature representation obtained by the i-th filter, k represents the convolution kernel, u represents the input, u∈X S , * represents the convolution operation, f(·) represents the tanh activation function, z i (j) represents the feature of the i-th filter after pooling, D j represents the jth pooling area, y i (k) represents the feature representation of the i-th filter in the pooling kernel, W is the weight matrix, initialized with a normal distribution, b is the bias, initialized to 0, and h i It is the sound signal characteristic.

[0096] In step S3, the residual network ResNet50 is used to extract the feature vector of thermal imaging, which is denoted as h P ∈R N×m The emergence of the ResNet network proves that the network can develop in a deeper direction. Its core idea is to repeatedly stack convolutional blocks and identify blocks.

[0097] The specific process of step S4 is to first transform the feature vector h V 、h I 、h T 、h S and h P Concatenate into feature matrix h∈R N ×5×m ;

[0098] The feature matrix h∈R N×5×m Input to the fully connected layer and multiply it by the trainable weight matrix W Q , W K and W T , the query matrix Q, key matrix K and value matrix V are calculated; in the initial stage, the weight matrix W is artificially adjusted according to experience Q , W K and W T Assignment, during the model training process, the weight matrix W Q , W K and W T Constantly updated;

[0099] Then according to the attention operation function

[0100]

[0101] Calculate the attention score matrix α = [α 1 ,α 2 ,α 3 ,α 4 ,α 5 ]∈R N×5×5 , where α 1 , α 2 , α 3 , α 4 and α 5 They are the attention scores of voltage feature, current feature, temperature feature, sound signal feature, and thermal imaging feature, d is the hidden dimension of the fully connected layer, and K T is the transposed matrix of the key matrix K;

[0102] The attention mechanism is introduced to calculate the attention weight feature matrix H = α·V;

[0103] Finally, the feature fusion is completed by weighting and summing each feature using formula (8).

[0104] H=α 1 ·h 1 +α 2 ·h 2 +α 3 ·h 3 +α 4 ·h 4 +α 5 ·h 5 (8)

[0105] where h is the eigenvector corresponding to α.

[0106] In step S5, a multi-layer fully connected network is used as a classifier to obtain the thermal runaway state of the lithium-ion battery. During the model training process, a corresponding relationship is established between the fused features and the corresponding battery state y in the classifier, and the model is continuously trained using the Adam optimizer until convergence.

[0107] Specifically, step S6 includes taking 80% of the data in step S2 as a training set and the remaining 20% ​​as a test set;

[0108] The training set is input into the thermal runaway prediction model for training, the model training epoch is 3000, the learning rate is 8e-5, and the Adam optimizer is used to obtain a trained thermal runaway prediction model, and all parameters are automatically learned during the model training process;

[0109] Then input the test set for verification to get the prediction results. Comparing the predicted state with the actual battery state, we can see that there are four prediction results:

[0110] The actual state is thermal runaway, and the predicted state is thermal runaway;

[0111] The actual state is thermal runaway, and the predicted state is normal;

[0112] The actual state is normal, but the predicted state is thermal runaway;

[0113] The actual state is normal, and the predicted state is normal;

[0114] Count the number of various prediction results, and use the accuracy rate ACC as the evaluation index to judge the accuracy of the model.

[0115]

[0116] Among them, TP represents the number of thermal runaway states predicted as thermal runaway, TN represents the number of normal states predicted as normal states, FP represents the number of normal states predicted as thermal runaway, and FN represents the number of thermal runaway states predicted as normal states;

[0117] The obtained accuracy ACC should be greater than 80% to consider that the model training has met the requirements and then conduct subsequent tests.

[0118] During thermal runaway, lithium-ion batteries will emit some abnormal sounds and generate localized heat inside the battery. These features may appear earlier than abnormalities in measurement data such as voltage, current, and temperature. Therefore, the present invention proposes for the first time that sound and thermal imaging should also be important features that need to be considered in data-driven methods. Sequence data, sound, and thermal imaging are in different data formats. Therefore, a single model cannot extract the features of these data to achieve thermal runaway prediction of lithium-ion batteries.

[0119] Based on three different types of features, namely sequence data, sound and thermal imaging, the present invention proposes a method for constructing a multi-modal and multi-feature lithium-ion battery thermal runaway prediction model. The battery thermal runaway process is considered from different feature perspectives, different deep learning models are used to mine the potential features of the data, and the attention mechanism is used to calculate the weights between features. The importance of different features to the classification results is taken into account, thereby improving the accuracy of thermal runaway prediction.

[0120] Embodiment 2

[0121] like Figure 2 , a lithium-ion battery thermal runaway prediction system, comprising a data acquisition module 100, a data transmission module 200 and a thermal runaway prediction module 300;

[0122] The data acquisition module 100 is used to collect voltage, current, temperature, sound and thermal imaging data of the battery: for example, a current / voltage signal collector 101 may be used to acquire voltage and current data, a temperature sensor 102 may be used to acquire temperature data, a sound sensor 103 may be used to acquire sound signal data, and a thermal imager 104 may be used to acquire thermal imaging data;

[0123] The data transmission module 200 is used to transmit the data collected by the data collection module 100 to the thermal runaway prediction module 300;

[0124] The thermal runaway prediction module 300 includes a memory 301, a processor 302, and a data processing mechanism stored in the memory 301 and operable on the processor 302. The data processing mechanism may be a computer program. The data processing mechanism is used to:

[0125] Inputting the data transmitted by the data transmission module 200 into the thermal runaway prediction model, the thermal runaway prediction model outputting a predicted thermal runaway state of the battery according to the input data;

[0126] The thermal runaway prediction model is constructed and trained according to the method described in Example 1.

[0127] Finally, it should be noted that the above description is only a preferred embodiment of the present invention. Under the guidance of the present invention, ordinary technicians in this field can make various similar expressions without violating the purpose and claims of the present invention, and such changes all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a thermal runaway prediction model for lithium-ion batteries. Features The method includes collecting thermal runaway experimental data and using the experimental data to construct a thermal runaway prediction model based on a neural network, wherein the thermal runaway prediction model includes a feature extractor, a feature fusion device and a classifier. The method specifically includes the following steps: S1. Thermal runaway test data collection: Conduct thermal runaway tests on lithium-ion batteries and collect battery sequence feature data, as well as sound signals and thermal imaging data; The battery sequence characteristic data includes temperature data, voltage data and current data; S2, data preprocessing: slicing the data collected in step S1 to generate historical data; The battery state is recorded as y, and a temperature threshold is set. If the temperature exceeds the threshold, thermal runaway occurs, y=1; otherwise, thermal runaway does not occur, y=0; S3, extracting feature data: taking the data preprocessed in step S2 as input, taking the abnormal state of thermal runaway of the lithium-ion battery as a label, and using the feature extractor to extract data features; S4, feature fusion: using the feature fuser to calculate the attention weights of all the data features and perform feature fusion; S5, thermal runaway state judgment: using the classifier to obtain the thermal runaway state of the lithium-ion battery; S6, model training: dividing the historical data in step S2 into a training set and a test set, inputting the training set into the thermal runaway prediction model for training, and inputting the test set into the thermal runaway prediction model for verification, and judging the model accuracy by using the accuracy rate as the evaluation index, thereby updating the thermal runaway prediction model; The specific process of slicing in step S2 is: slice the data collected in step S1 using a sliding window of size T to generate historical voltage data X V , current data X I , temperature data X T , sound dataX S and thermal imaging dataX P ; The feature extractor includes a long short-term memory neural network (LSTM), a one-dimensional convolutional neural network (1D CNN) and a residual network ResNet50; In step S3, a long short-term memory neural network (LSTM) is used to extract the characteristic vectors of the voltage, current and temperature sequence data, which are respectively denoted as h V ∈R N×m 、h I ∈R N×m and h T ∈R N×m ; A one-dimensional convolutional neural network (1D CNN) is used to extract the feature vector of the sound signal, denoted as h S ∈R N×m ; The residual network ResNet50 is used to extract the feature vector of thermal imaging, denoted as h P ∈R N×m .

2. A method for constructing a thermal runaway prediction model for a lithium-ion battery according to claim 1, Features: In step S3, the characteristics of the voltage, current and temperature sequence data are calculated and extracted according to equations (1) to (6). g t =σ(W f ·[h t-1 ,x t ])+b g (1) i t =σ(W i ·[h t-1 ,x t ])+b i (2) the t =σ(W o ·[h t-1 ,x t ])+b o (5) h t =o t *fishy(S) t ) (6) Among them, g t It is the forget gate, t is the input gate, o t is the output gate, σ is the activation function; W is the weight matrix, initialized with normal distribution; b is the bias, initialized to 0; h t is the characteristic of voltage or current or temperature series data; When extracting voltage series data features, input x t ∈X V , when extracting the features of current series data, input x t ∈X I , input x when extracting temperature series data features t ∈X T ; and S t It is the intermediate value during the operation.

3. A method for constructing a thermal runaway prediction model for a lithium-ion battery according to claim 1 or 2, Features: The feature fusion device is used for concatenation, weight calculation and fusion of feature vectors; The specific process of step S4 is to first transform the feature vector h V 、h I 、h T 、h S and h P Concatenate into feature matrix h∈R N×5×m ; The feature matrix h∈R N×5×m Input to the fully connected layer and multiply it by the trainable weight matrix W Q , W K and W T , calculate the query matrix Q, key matrix K and value matrix V; Then according to the attention operation function Calculate the attention score matrix α = [ɑ 1 ,α 2 ,α 3 ,α 4 ,α 5 ]∈R N×5×5 , where α 1 , α 2 、ɑ 3 、ɑ 4 and 5 They are the attention scores of voltage feature, current feature, temperature feature, sound signal feature, and thermal imaging feature, d is the hidden dimension of the fully connected layer, and K T is the transposed matrix of the key matrix K; The attention mechanism is introduced to calculate the attention weight feature matrix H = ɑ·V; Finally, the feature fusion is completed by weighting each feature using formula (8). H=a 1 ·h 1 +a 2 ·h 2 +a 3 ·h 3 +a 4 ·h 4 +a 5 ·h 5 (8) where h is the eigenvector corresponding to α.

4. A method for constructing a thermal runaway prediction model for a lithium-ion battery according to claim 1, Features: In step S3, the 1D CNN is composed of a convolution layer, a pooling layer and a fully connected layer. Multiple convolution kernels are used to perform convolution and pooling operations on the input data to extract the potential features of the data. Specifically, the sound signal features are calculated and extracted according to equations (9) to (11). y i =f(u i *k+b i ) (9) from i (j)=maxy i (k) ,k∈D j (10) h i = W i z i + b i (11) Among them, y i represents the feature representation obtained by the i-th filter, k represents the convolution kernel, u represents the input, u∈X S , * represents the convolution operation, f(·) represents the tanh activation function, z i (j) represents the feature of the i-th filter after pooling, D j represents the jth pooling area, y i (k) represents the feature representation of the i-th filter in the pooling kernel, W is the weight matrix, b is the bias, and h i It is the sound signal characteristic.

5. A method for constructing a thermal runaway prediction model for a lithium-ion battery according to claim 3, Features: In step S5, a multi-layer fully connected network is used as a classifier to obtain the thermal runaway state of the lithium-ion battery.

6. A method for constructing a thermal runaway prediction model for a lithium-ion battery according to claim 3, Features: Specifically, step S6 includes taking 80% of the data in step S2 as a training set and the remaining 20% ​​as a test set; The training set is input into the multimodal model for training, the model training epoch is 3000, the learning rate is 8e-5, and the Adam optimizer is used to obtain a trained thermal runaway prediction model; Then input the test set for verification to get the prediction result, and use the accuracy rate ACC as the evaluation index to judge the accuracy of the model. Among them, TP represents the number of thermal runaway states predicted as thermal runaway, TN represents the number of normal states predicted as normal states, FP represents the number of normal states predicted as thermal runaway, and FN represents the number of thermal runaway states predicted as normal states.

7. A lithium-ion battery thermal runaway prediction system, Features: It comprises a data acquisition module (100), a data transmission module (200) and a thermal runaway prediction module (300); The data acquisition module (100) is used to collect voltage, current, temperature, sound and thermal imaging data of the battery; The data transmission module (200) is used to transmit the data collected by the data collection module (100) to the thermal runaway prediction module (300); The thermal runaway prediction module (300) comprises a memory (301), a processor (302), and a data processing mechanism stored in the memory (301) and operable on the processor (302), wherein the data processing mechanism is used to: Inputting data transmitted by the data transmission module (200) into a thermal runaway prediction model, wherein the thermal runaway prediction model outputs a predicted thermal runaway state of the battery according to the input data; The thermal runaway prediction model is constructed and trained and updated according to the method described in any one of claims 1 to 6 above.

Images