Battery thermal runaway time prediction model training method, prediction device and prediction method thereof

Abstract

The invention belongs to the technical field of battery safety detection, and discloses a battery thermal runaway time prediction model training method, a prediction device and a prediction method thereof. The invention discloses a battery thermal runaway time prediction model training method. The method comprises the following steps: collecting battery related multi-source time sequence monitoring data, and aligning and integrating the data into a structured input sequence according to a unified time reference; a staged fusion neural network model is established, a TCN module extracts multi-scale local time sequence features by adopting a stacked residual block and expansion causal convolution, a Transform encoder captures global long-term dependency of a sequence through a multi-head self-attention mechanism, and the TCN module and the Transform encoder are deeply fused through a progressive structure; and after the model is trained and optimized by using the training set, the trained model is deployed to a battery management system, and real-time online accurate prediction of the thermal runaway remaining time is realized. According to the method, the accuracy and timeliness of thermal runaway early warning are remarkably improved through advantage complementation of local feature extraction and global correlation modeling.

Description

Technical Field

[0001] This invention belongs to the field of battery safety testing technology, specifically relating to a battery thermal runaway time prediction model training method, prediction device and prediction method. Background Technology

[0002] Lithium iron phosphate (LFP) batteries have become key energy devices in electric vehicles, large-scale energy storage, and other fields due to their high safety and long cycle life. However, under extreme abuse conditions, these LFP batteries can still experience thermal runaway, leading to serious safety accidents.

[0003] Currently, thermal runaway early warning mainly relies on threshold monitoring of single parameters such as voltage or temperature. While simple, this method has limitations. Thermal runaway is the result of complex coupling of multiple physical fields, and single parameters are easily affected by interference, leading to low accuracy and insufficient reliability in early warning. Furthermore, fixed threshold methods struggle to characterize the temporal evolution of parameters, failing to provide early warning. Although some electrochemical-thermal coupling models have been proposed, their computational complexity and reliance on precise internal parameters make them unsuitable for practical online applications.

[0004] In recent years, deep learning models, such as Recurrent Neural Networks (RNNs) and Long Short-Term Memory Networks (LSTMs), have been attempted for battery health state prediction due to their powerful ability to capture temporal features. However, these sequence models generally suffer from vanishing or exploding gradients when dealing with long sequence predictions, and their recursive structures lead to low computational efficiency. Accumulated errors spread with increasing prediction step size, affecting the stability of long-term predictions. Furthermore, thermal runaway processes are often accompanied by drastic, nonlinear data mutations, requiring prediction models to not only possess excellent long-term memory capabilities but also high sensitivity to key mutation point features and strong parallel computing efficiency—qualities that traditional time-series models struggle to achieve simultaneously.

[0005] Therefore, it is necessary to construct a predictive model that can deeply integrate temporal local patterns and global multi-parameter correlations to provide more accurate and earlier warnings of thermal runaway in lithium iron phosphate batteries and improve the reliability of safety management. Summary of the Invention

[0006] The purpose of this invention is to provide a training method for a battery thermal runaway time prediction model. The constructed prediction model is a hybrid framework model that integrates Temporal Convolutional Network (TCN) and Transformer to deeply mine the temporal dynamic features and global dependencies of thermal runaway precursors. It aims to solve the problems of low accuracy and poor timeliness of existing early warning methods and achieve earlier and more reliable prediction of thermal runaway.

[0007] The present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for training a battery thermal runaway time prediction model. The battery thermal runaway time prediction model is a battery thermal runaway time prediction model based on a staged fusion neural network. Specifically, the model training method includes the following steps:

[0009] Step S1: Acquire multi-source time-series monitoring data of the battery under operating or thermal abuse conditions. For example, in a closed combustion chamber experimental platform, heat the test battery by applying a constant power of 600W to trigger thermal runaway. The multi-source time-series monitoring data includes at least one or more of temperature, temperature rise rate, voltage, gas release parameters, mass change, or expansion force. A multi-dimensional time-series input sequence is constructed according to a unified time scale. The original dataset consisting of the multi-dimensional time-series input sequence and labels is divided into a training set, a validation set, and a test set. Normalization is used to remove the influence of data dimensions. It should be noted that the multi-dimensional time-series input sequence consists of multiple monitoring parameters aligned by timestamps. The label is the remaining time from each timestamp to the point of thermal runaway, and its value is determined by the difference between the thermal runaway initiation determination timestamp and the current timestamp.

[0010] Step S2: The multidimensional temporal input sequence preprocessed in Step S1 is input into the Temporal Convolutional Network (TCN) module. The TCN module contains multiple residual blocks, through which feature extraction is performed. Each residual block includes at least two layers of dilated causal convolution, an activation function, and a 1×1 convolution. The dilation rate of the dilated causal convolution gradually increases in different residual blocks according to a predetermined rule of geometric progression, to form a causal receptive field covering multiple time scales. In the TCN module, the input sequence is first subjected to two layers of dilated causal convolution and then activated by the activation function through the first residual block. The input and output are processed by a 1×1 convolutional layer for channel transformation, and then the input and output are added through residual connections to achieve feature fusion. The input and output are then processed by the same hierarchical structure through subsequent residual blocks, with the expansion rate of each subsequent residual block increasing according to the predetermined rule. Finally, the local dynamics modeling of the multidimensional temporal input sequence is completed, and the risk candidate feature sequence representing the thermal runaway evolution trend is output. The 1×1 convolutional layer is set to introduce 1×1 convolution in the residual branch to achieve feature channel alignment or channel reconstruction, thereby enhancing the consistency of risk features at different time scales before entering the Transformer encoder.

[0011] Step S3: The risk candidate feature sequence obtained in step S2 is used as the input to the Transformer encoder, instead of directly inputting the original multi-source time-series data into the Transformer. The Transformer encoder consists of multiple encoder blocks stacked sequentially. Each encoder block includes a multi-head attention module, a first Dropout and normalization module, a feedforward network module, and a second Dropout and normalization module. Specifically, positional encoding is introduced into the Transformer encoder. The risk candidate feature sequence with added positional encoding is input into the multi-head attention module, and the query (Q), key (K), and value (V) are calculated. The attention weights between the layers are used to capture global temporal dependencies. Then, the output of the multi-head attention module is added to the input of the module through a residual connection, and then processed by Dropout and layer normalization. Next, the processed features are input to the feedforward network module for nonlinear transformation and feature enhancement. Finally, the output of the feedforward network module is added to the output of the previous layer through a residual connection, and then processed by Dropout and layer normalization to output the result of the current encoder block. This result will be used as the input of the next encoder block and processed by all encoder blocks in sequence, thereby completing the global temporal dependency modeling of the risk candidate feature sequence to obtain a global risk representation of the thermal runaway evolution process.

[0012] Step S4: Perform global average pooling processing on the global risk representation obtained in step S3 in the time dimension to eliminate the influence of instantaneous fluctuations on the prediction results and obtain the overall risk state representation; input the overall risk state representation into the multilayer perceptron network, and output the predicted value of the remaining time of the lithium-ion battery from thermal runaway at the current moment through nonlinear mapping.

[0013] Step S5: Train the staged fusion neural network model from steps S2 to S4 using the training set, and validate the trained model using the validation set data to check whether the model has been trained to the optimal state.

[0014] Step S6: Input the test set into the trained staged fusion neural network model to obtain the prediction results, and use the mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination R² to evaluate and analyze the prediction accuracy.

[0015] As a specific embodiment of this application, in step S1, taking a lithium iron phosphate battery as an example, the thermal runaway data, i.e., the multi-source time-series monitoring data, includes the battery's front surface temperature, front surface temperature rise rate, rear surface temperature, rear surface temperature rise rate, left surface temperature, left surface temperature rise rate, right surface temperature, right surface temperature rise rate, terminal voltage, mass loss, hydrogen concentration, carbon monoxide concentration, methane concentration, ethylene concentration, and expansion force. Specifically, the above-mentioned multi-source time-series monitoring data can be based on, for example... Figure 1 The physical simulation experimental platform shown was used to trigger thermal runaway processes under different SOC states, simultaneously acquiring multiphysics time-series data. Four surface temperatures were directly measured using K-type thermocouples placed on the battery surface, and the temperature rise rate was calculated to characterize the spatiotemporal evolution of local thermal behavior. High-temperature resistant voltage acquisition lines were directly connected to the positive and negative terminals of the battery and integrated into a high-precision data acquisition system (JK360) to continuously monitor terminal voltages to identify internal degradation and short-circuit initiation. A high-precision electronic balance was fixed to the experimental platform and maintained in a sealed combustion chamber with a dedicated exhaust channel to maintain environmental stability, allowing for real-time tracking of mass loss and effective elimination of interference from gas release, accurately reflecting the degree of exhaust and ejection. A multi-component gas analyzer was used to monitor gas concentration online to assess the decomposition reaction and exothermic intensity. A high-temperature resistant load sensor and its matching digital controller were used to measure expansion force in real time, capturing the thermomechanical deformation of the battery casing. These parameters collectively constitute key characteristic variables characterizing the thermal runaway evolution process, serving as inputs for subsequent prediction models.

[0016] In a preferred embodiment of this application, step S1, the step of dividing the dataset into training set, validation set and test set, is as follows: the dataset is divided into training set, validation set and test set in a ratio of 7:1:2.

[0017] In a preferred embodiment of this application, the specific calculation formula for the normalization of the normal distribution in step S1 is as follows:

[0018]

[0019] In the formula: X * This represents the normalized data; This represents the sample data; μ represents the mean. It represents the standard deviation.

[0020] In a preferred embodiment of this application, in step S2, the expansion rate of the expanded causal convolution in the multiple residual blocks of the temporal convolutional network module is set in an increasing manner, and the expansion rate includes at least 1, 2 and 4, so as to form a causal receptive field covering different time scales, thereby characterizing the multi-scale local dynamic features in the thermal runaway evolution process of lithium-ion batteries.

[0021] In a preferred embodiment of this application, in step S3, the self-attention mechanism module uses the query vector Q, key vector K, and value vector V, along with a correspondingly changing learnable weight matrix, to predict the encoded dataset; the output generated by this module is:

[0022]

[0023] In the formula: softmax(i) V represents the weighted sum; d k K represents the dimension of the query vector Q and the key vector K; T This represents the transpose of the K vector.

[0024] In a preferred embodiment of this application, in step S5, the staged fusion neural network model structure framework is as follows: Figure 1 As shown, training a staged fusion neural network model includes the following steps: optimizing the staged fusion neural network model using the Adam optimizer; performing multiple iterations on the model using the training set and updating the parameters with a specified batch size; and evaluating the staged fusion neural network model using a validation set after each iteration.

[0025] Secondly, based on the battery thermal runaway time prediction model training method described in the first aspect, the present invention also provides a battery thermal runaway time prediction device, the prediction device comprising: a data acquisition module, a data preprocessing module, a model training module, and a prediction module;

[0026] The data acquisition module is used to synchronously acquire multi-source time-series monitoring data, such as multi-parameter time-series measurement data recorded at 10Hz.

[0027] The data preprocessing module is used to perform normalization processing on the multi-source time-series monitoring data and construct features and labels, normalize the data to eliminate numerical differences between them, and divide the dataset into training set, validation set and test set.

[0028] The model training module is used to train the staged fusion neural network model by using the preprocessed training set as input to obtain the thermal runaway prediction model.

[0029] The prediction module is used to input test set data into the trained staged fusion neural network model to obtain prediction results.

[0030] Thirdly, the present invention provides a method for predicting battery thermal runaway time, which involves training the optimal staged fusion neural network model using the battery thermal runaway time prediction model training method described in the first aspect and then deploying and applying it. Specifically, the model is embedded into the monitoring unit of the target battery system, receiving preprocessed real-time sensor data, continuously calculating and outputting the remaining thermal runaway time, and generating thermal runaway early warning information based on the predicted remaining time value to guide the safety management or emergency response of the battery system.

[0031] Technical effects:

[0032] 1) Effective decoupling modeling of local dynamics and global dependencies. This invention integrates temporal convolutional networks and Transformer models in stages. First, multi-scale dilated causal convolution is used to model the causal constraints of local temporal features in the thermal runaway evolution process of lithium-ion batteries. Then, Transformer is used to model the global dependencies of the selected risk candidate features. This effectively avoids the interference of long-term temporal noise on the attention mechanism and improves the ability to identify the thermal runaway evolution trend.

[0033] 2) Stability and reliability of early stage prediction of thermal runaway. By introducing residual structure, channel reconstruction and global average pooling mechanism in the model fusion process, this invention can weaken the impact of instantaneous abnormal fluctuations on the prediction results, making the model output more reflective of the overall risk state of the battery, thereby improving the stability of thermal runaway early warning in the early stage.

[0034] 3) It enables continuous prediction of the remaining time of thermal runaway, rather than state judgment. The output of this invention is the remaining time of a lithium-ion battery in its current state before thermal runaway occurs. Compared to simply judging whether thermal runaway has occurred or predicting a single physical parameter, this is more conducive to early intervention and safety decisions by the battery management system.

[0035] 4) Balancing prediction accuracy and computational efficiency, suitable for engineering applications. This invention combines dilated causal convolution with staged attention modeling, maintaining model prediction accuracy while avoiding direct attention calculation on extremely long original sequences, thus reducing computational complexity and making it suitable for online monitoring and embedded deployment.

[0036] 5) Adapted to the characteristics of thermal runaway, a critical instability process in lithium-ion batteries. Compared to general time-series prediction models, this invention is designed around the criticality, abruptness, and irreversibility of thermal runaway in both model structure and output format, which better meets the actual needs of lithium-ion battery safety early warning. Attached Figure Description

[0037] Figure 1 Diagram of the experimental setup;

[0038] Figure 2 This is a structural framework diagram of the staged fusion neural network in this invention;

[0039] Figure 3 This is a flowchart of the present invention;

[0040] Figure 4 This is a comparison chart of the prediction results of the present invention; Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1: Staged Fusion Neural Network Model Training Method

[0043] This invention proposes a training method for a battery thermal runaway time prediction model based on a staged fusion neural network. The prediction process is as follows: Figure 3 As shown, the specific steps include:

[0044] 1. Data Collection

[0045] In a closed combustion chamber experimental platform, thermal runaway was triggered by applying a constant power of 600W to a 314Ah lithium iron phosphate battery. The experimental platform consisted of the test battery, a thermal abuse triggering unit, a multi-parameter synchronous data acquisition unit, and a safety protection unit. The thermal abuse triggering unit was a constant power heating plate attached to one side of the battery. The multi-parameter synchronous data acquisition unit included: multiple thermocouples connected to the front, back, left, and right surfaces of the battery to collect temperature data; high-temperature resistant wires connected to the positive and negative terminals of the battery to collect terminal voltage data; both thermocouples and voltage wires were connected to a multi-channel data acquisition instrument; an overload sensor for measuring the expansion force of the battery casing, which was parallel to the other side of the battery via rigid pads and insulated with mica sheets; the electrical signal output by the sensor was conditioned by an amplifier before being connected to the data acquisition instrument; a high-precision electronic balance supporting the entire experimental setup, connected to a computer via a data cable to transmit mass data; and a multi-component gas analyzer that extracted gas from the combustion chamber through a high-temperature sampling tube and analyzed the gas products online. The safety protection unit consists of a sealed explosion-proof combustion chamber body, a pressure relief ventilation system, and an external fireproof blanket.

[0046] Through the aforementioned platform, multi-source time-series monitoring data of lithium-ion batteries under thermal abuse conditions are acquired. This multi-source time-series monitoring data includes temperature, temperature rise rate, voltage, gas release parameters, mass change, and expansion force. The experimental data acquisition method for the lithium iron phosphate battery in Example 1 is as follows: Temperature data is collected by thermocouples directly contacting the battery surface, and the thermoelectric potential signal is directly converted into a temperature value and recorded by a data acquisition instrument; the temperature rise rate is calculated from the recorded temperature-time series using numerical differentiation; voltage data is directly collected from the battery's terminal voltage signal via wires; expansion force data is obtained from the analog voltage signal output by an overload sensor, which is then converted into a force value after amplifier conditioning; mass data is obtained through an electronic balance, which is zeroed before the experiment, and its output real-time digital signal represents the mass loss of the battery system. Gas concentration data is directly measured and output by an online gas analyzer, which outputs the volume concentration signals of hydrogen, carbon monoxide, methane, and ethylene. Each acquisition device independently records data, which is then integrated into a standardized synchronous time-series dataset after alignment and interpolation processing using a unified time reference.

[0047] Characteristic variables are determined based on the acquired and preprocessed synchronous data. These characteristic variables include the battery's front surface temperature, front surface temperature rise rate, rear surface temperature, rear surface temperature rise rate, left surface temperature, left surface temperature rise rate, right surface temperature, right surface temperature rise rate, terminal voltage, mass loss, hydrogen concentration, carbon monoxide concentration, methane concentration, ethylene concentration, and expansion force.

[0048] A multidimensional time-series input sequence was constructed according to a unified time scale, and the remaining thermal runaway time corresponding to the occurrence of thermal runaway in the battery was used as the label data. The datasets collected under all five SOC conditions were concatenated into a unified dataset, which was then split into training, validation, and test sets, and normalized using a normal distribution to remove the influence of data dimensions.

[0049] 2. Data Preprocessing

[0050] The specific processing steps for constructing a multidimensional time series input sequence according to a unified time scale are as follows: integrate multivariate time series data into a model input sequence with a unified time base, wherein the time window length is set to 60 time steps, and each input sequence contains feature data of 60 time steps.

[0051] The input sequence of the model with a unified time base constructed in step 1 is normalized to remove the influence of data units and reduce the impact on the prediction results. The specific formula for normalization is as follows:

[0052]

[0053] In the formula: X * This represents the normalized data; This represents the original sample data in the input sequence; μ represents the mean of all sample data corresponding to the feature in the training set; This represents the standard deviation of all sample data corresponding to the feature in the training set.

[0054] 3. Model Building

[0055] TCN module: such as Figure 2 As shown, a TCN module is constructed using three stacked residual blocks. The kernel size of the dilated causal convolution in each residual block is 3, with dilation rates set to 1, 2, and 4 respectively to capture feature information at different time scales. After the dilated causal convolution, the ReLU activation function is used to increase the model's nonlinear expressive power, and then a 1x1 convolution is used to adjust the number of channels to ensure the feasibility of residual connections. The data processing procedure of the TCN module is as follows: the preprocessed multi-dimensional temporal input sequence is input into the temporal convolutional network TCN module, and feature extraction is performed through multiple residual blocks. First, the input sequence is subjected to two layers of dilated causal convolution and activated by the first residual block, followed by channel transformation through a 1×1 convolution layer. Then, the input and output are added through residual connections to achieve feature fusion. This process is repeated in subsequent residual blocks, performing the same hierarchical processing. Finally, the local dynamics modeling of the multi-dimensional temporal input sequence is completed, and a risk candidate feature sequence representing the thermal runaway evolution trend is output.

[0056] Transformer module: such as Figure 2As shown, it consists of two Transformer encoder blocks. The multi-head attention mechanism in each encoder block has four heads, enabling the model to simultaneously focus on features from different locations. The feedforward network contains two fully connected layers, with a ReLU activation function in between to further process and transform the features output by the attention mechanism. The data processing procedure of the Transformer module is as follows: Risk candidate feature sequences are used as input to the Transformer encoder, instead of directly inputting the original multi-source time-series data. Positional encoding is introduced into the Transformer encoder, and the risk candidate feature sequences with added positional encoding are input to the multi-head attention module. Global temporal dependencies are captured by calculating the attention weights between queries (Q), keys (K), and values ​​(V). The output of the multi-head attention module is then added to the module's input via a residual connection, and subsequently processed by Dropout and layer normalization. The processed features are then input to the feedforward network module for nonlinear transformation and feature enhancement. Finally, the output of the feedforward network module is again added to the output of the previous layer via a residual connection, and subsequently processed by Dropout and layer normalization to output the result of the current encoder block. This result serves as the input to the next encoder block, and is processed by all encoder blocks sequentially, thereby completing the global temporal dependency modeling of the risk candidate feature sequences to obtain a global risk representation of the thermal runaway evolution process.

[0057] Multilayer perceptron (MLP): such as Figure 2 As shown, the outputs of the TCN and Transformer modules are subjected to global average pooling, and the pooled features are input into a multilayer perceptron. The multilayer perceptron contains two hidden layers with 64 and 32 neurons respectively, using the ReLU activation function, and finally outputs the predicted time remaining in thermal runaway (TTR). The obtained global risk representation is subjected to global average pooling in the time dimension to eliminate the influence of instantaneous fluctuations on the prediction results, obtaining an overall risk state representation; the overall risk state representation is input into the multilayer perceptron network, and the predicted time remaining before thermal runaway of the lithium-ion battery at the current moment is output through nonlinear mapping.

[0058] Overall model architecture: such as Figure 2 As shown, the constructed TCN module, Transformer module, and multilayer perceptron are connected sequentially to form a complete TCN-Transformer model.

[0059] 4. Model Training

[0060] The staged fusion neural network model is optimized using the Adam optimizer. Multiple iterations are performed on the model using the training set, with parameters updated using a specified batch size. After each iteration, the staged fusion neural network model is evaluated using a validation set. The trained staged fusion neural network model is validated using validation set data to verify whether it has reached its optimal state. The constructed staged fusion neural network model is then trained using the training dataset. Mean squared error (MSE) is chosen as the loss function, and the parameters are optimized using the Adam optimizer with a learning rate of 0.001. During training, after each epoch, the model is validated using the validation dataset, and the loss value on the validation set is calculated. Training stops when the loss value on the validation set does not decrease for 50 consecutive epochs, and the model parameters with the smallest loss value on the validation set are selected as the optimal parameters.

[0061] The test set data was input into the staged fusion neural network model trained in Example 1 to obtain the final thermal runaway prediction result. The prediction effect of the proposed model was verified using an experimentally obtained multi-physics thermal runaway time-series dataset of lithium iron phosphate batteries. This dataset covers various typical abuse conditions and has a sampling frequency of 10 Hz. All experimental data were randomly divided into a training set, a validation set, and an independent test set in a 7:1:2 ratio to ensure the generalization and fairness of the model evaluation. Error analysis was performed on the prediction results using four evaluation metrics: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²). This comprehensively quantifies the model's prediction accuracy and goodness of fit for the remaining thermal runaway time. The calculation formulas for each metric are as follows:

[0062]

[0063]

[0064]

[0065]

[0066] In the formula: N represents the number of samples; and These represent the actual and predicted values ​​of the remaining time of thermal runaway, respectively. This represents the average remaining time of thermal runaway.

[0067] To fully demonstrate the effectiveness of the battery thermal runaway time prediction method based on a staged fusion neural network proposed in this invention, the proposed staged fusion neural network model was compared with six other models: GRU, LSTM, Bi-LSTM, TCN-LSTM, TCN-Attention, and Bi-LSTM-Attention. Thermal runaway predictions were then performed on each model, and the prediction results of the proposed model are shown below. Figure 4 As shown in Table 1, the evaluation index results are as follows.

[0068] from Figure 4 As can be seen, the prediction results of the seven prediction models all fluctuate around the actual thermal runaway curve. However, the model proposed in this invention performs the best, demonstrating the superiority and effectiveness of this invention.

[0069] As shown in Table 1, compared to other models, the proposed model achieves the lowest MAE (0.4937), MSE (0.8350), and RMSE (0.9138), while obtaining the highest coefficient of determination (R² = 0.999999). Compared to traditional recurrent models, the improvement is significant, with RMSE reductions ranging from 52.9% (relative to GRU) to 57.5% (relative to LSTM). Compared to advanced hybrid baseline models such as TCN-LSTM and TCN-Attention, the proposed model still achieves consistent improvements, reducing RMSE by 72.1% and 62.6%, respectively. Table 1: Error analysis and evaluation metrics for different models.

[0070]

[0071] While the Bi-LSTM-Attention model also demonstrated competitiveness, the staged fusion neural network model surpassed it with an additional 7.8% reduction in RMSE, while maintaining comparable computational efficiency. Taken together, these results demonstrate that combining temporal convolutional networks with a Transformer-based attention mechanism enables the proposed architecture to extract multi-scale temporal features more effectively, thereby improving accuracy and robustness.

Claims

1. A method for training a battery thermal runaway time prediction model, characterized in that, The model training method includes the following steps: Step S1: Obtain multi-source time-series monitoring data of the battery under operating or thermal abuse conditions; the multi-source time-series monitoring data includes at least one or more of temperature, temperature rise rate, voltage, gas release parameters, mass change or expansion force, construct a multi-dimensional time-series input sequence according to a unified time scale, divide the original dataset consisting of the multi-dimensional time-series input sequence and labels into training set, validation set and test set, and use normal distribution normalization to remove the influence of data dimensions; Step S2: The multidimensional temporal input sequence preprocessed in Step S1 is input into the temporal convolutional network (TCN) module to model the local dynamics of the multidimensional temporal input sequence and output a risk candidate feature sequence characterizing the thermal runaway evolution trend. The TCN module contains multiple residual blocks in sequence, and feature extraction is performed through multiple residual blocks. Each residual block includes at least two layers of dilated causal convolution, an activation function, and a 1×1 convolution. The dilation rate of the dilated causal convolution gradually increases in different residual blocks according to a predetermined rule of geometric progression. Step S3: The risk candidate feature sequence obtained in step S2 is used as the input of the Transformer encoder. The Transformer encoder consists of multiple encoder blocks stacked in sequence. Each encoder block includes a multi-head attention module, a first Dropout and normalization module, a feedforward network module, and a second Dropout and normalization module. After being processed by all encoder blocks in sequence, the global time dependency model of the risk candidate feature sequence is completed, and the global risk characterization of the thermal runaway evolution process is obtained. Step S4: Perform global average pooling processing on the global risk representation obtained in step S3 in the time dimension to eliminate the influence of instantaneous fluctuations on the prediction results and obtain the overall risk state representation; input the overall risk state representation into the multilayer perceptron network, and output the predicted value of the remaining time of the lithium-ion battery from thermal runaway at the current moment through nonlinear mapping. Step S5: Train the staged fusion neural network model from steps S2 to S4 using the training set, and validate the trained model using the validation set data to check whether the model has been trained to the optimal state. Step S6: Input the test set into the trained staged fusion neural network model to obtain the prediction results, and use the mean absolute error, mean square error, root mean square error and coefficient of determination R² to evaluate and analyze the prediction accuracy.

2. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, The battery is a lithium iron phosphate battery. In this case, the thermal runaway data in step S1, i.e., the multi-source time-series monitoring data, includes the battery's front surface temperature, front surface temperature rise rate, rear surface temperature, rear surface temperature rise rate, left surface temperature, left surface temperature rise rate, right surface temperature, right surface temperature rise rate, terminal voltage, mass loss, hydrogen concentration, carbon monoxide concentration, methane concentration, ethylene concentration, and expansion force.

3. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, In step S1, the ratio of the division into training set, validation set and test set is 7:1:

2.

4. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, In step S1, the specific calculation formula for the normalization of the normal distribution is as follows: ； In the formula: X * This represents the normalized data; This represents the sample data; μ represents the mean. It represents the standard deviation.

5. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, In step S2, the dilation rates of the dilated causal convolutions in the multiple residual blocks of the temporal convolutional network module are 1, 2 and 4, respectively.

6. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, In step S3, the self-attention mechanism module uses the query vector Q, key vector K, and value vector V, along with a corresponding changing learnable weight matrix, to predict the encoded dataset; the output generated by this module is: ； In the formula: softmax(i) V represents the weighted sum; d k K represents the dimension of the query vector Q and the key vector K; T This represents the transpose of the K vector.

7. The battery thermal runaway time prediction model training method as described in claim 1, characterized in that, In step S5, training the staged fusion neural network model includes the following steps: optimizing the staged fusion neural network model using the Adam optimizer; performing multiple iterations on the model using the training set and updating the parameters with a specified batch size; and evaluating the staged fusion neural network model using a validation set after each iteration.

8. A battery thermal runaway time prediction device, characterized in that, The prediction device is used to execute the battery thermal runaway time prediction model training method according to claim 1. The prediction device includes: a data acquisition module, a data preprocessing module, a model training module, and a prediction module. The data acquisition module is used to synchronously acquire multi-source time-series monitoring data. The data preprocessing module is used to perform normalization processing on the multi-source time-series monitoring data and construct features and labels, normalize the data to eliminate numerical differences between them, and divide the dataset into a training set, a validation set, and a test set. The model training module is used to train a staged fusion neural network model by using the preprocessed training set as input to obtain a thermal runaway prediction model. The prediction module is used to input the test set data into the trained staged fusion neural network model to obtain the prediction result.

9. A method for predicting battery thermal runaway time, characterized in that, The prediction method involves deploying and applying the optimal staged fusion neural network model trained in the battery thermal runaway time prediction model training method of claim 1. Specifically, it is embedded into the monitoring unit of the target battery system, receives preprocessed real-time sensor data, continuously calculates and outputs the remaining thermal runaway time, and generates thermal runaway early warning information based on the predicted remaining time value to guide the safety management or emergency response of the battery system.

Images