A method and system for early warning of thermal runaway of a vehicle power battery

Abstract

This disclosure pertains to the field of batteries, specifically a method and system for early warning of thermal runaway in automotive power batteries. It involves collecting data on individual battery cell voltage, total voltage, current, and temperature to construct an original feature sequence. This sequence is then fused using multi-head one-dimensional convolution to integrate local temporal features at different time scales, generating a fused feature sequence. Temporal encoding is performed using a gated recurrent unit with an added anomaly response gate. Anomaly scores are calculated based on the degree to which the input features deviate from the historical mean, and the hidden state is updated accordingly. The output consists of a hidden state sequence and anomaly score sequence. A dual-channel attention module is used to calculate temporal and feature attention weights, which are then coupled through a Hadamard product to generate a coupled attention weight matrix. The hidden states are weighted and aggregated to obtain a context vector, which is decoded by a fully connected network to determine the thermal runaway risk probability. This probability, combined with the rate of change, generates an early warning threshold, enabling real-time and accurate thermal runaway early warning triggering.

Description

Technical Field

[0001] This disclosure pertains to the field of batteries, and particularly relates to a method and system for early warning of thermal runaway in automotive power batteries. Background Technology

[0002] Lithium-ion power batteries may experience thermal runaway under high temperatures or during their own aging and degradation, potentially leading to combustion or even explosion, posing a serious threat to the lives and property of passengers. Therefore, establishing a system capable of accurately and promptly predicting and warning of power battery thermal runaway is crucial for ensuring the safe operation of electric vehicles. Currently, early warning methods for power battery thermal runaway can be mainly categorized into threshold-based methods, model-based methods, and data-driven methods. Threshold-based methods monitor key parameters such as voltage and temperature, triggering an alarm when preset safety limits are exceeded. This method is simple and intuitive but easily affected by environmental factors, leading to false alarms and missed alarms. Model-based methods predict the state by establishing electrochemical or thermodynamic models of the battery. While the mechanisms are clear, the models are typically complex, computationally intensive, and poorly adaptable to individual battery differences and aging states, making large-scale, accurate applications difficult. Recurrent neural networks and their variants, such as Long Short-Term Memory networks and gated recurrent units, are widely used due to their advantages in processing time-series data. However, traditional network structures have limited ability to detect local correlation features of battery data at different time scales and may ignore multi-dimensional temporal patterns. Standard recurrent neural networks are not sensitive enough to early weak abnormal fluctuations that indicate thermal runaway when processing time-series data, and their internal state update mechanisms are difficult to reliably amplify key abnormal signals. When decoding encoded time-series information to assess risk, existing methods often fail to adequately distinguish the differences in the contribution of different time points and different monitoring parameters to risk determination. Moreover, fixed risk probability thresholds cannot adapt to the changing process of battery state, especially in the stage of rapid risk accumulation, which may lead to early warning delays. Summary of the Invention

[0003] To address the issue that existing methods fail to adequately differentiate the contributions of different time points and monitoring parameters to risk assessment, this disclosure proposes, in its first aspect, a method for early warning of thermal runaway in automotive power batteries, comprising the following steps: Data on the voltage, total voltage, current, and temperature of individual cells in an automotive power battery are acquired to form an original feature sequence with a preset time length. The original feature sequence is then subjected to multi-head one-dimensional convolution processing to extract and fuse local temporal features under different receptive fields at different times, generating a fused feature sequence. The fused feature sequence is input into an anomaly response gating loop unit for temporal encoding. This unit adds an anomaly response gate in addition to the standard update and reset gates. The anomaly response gate calculates an anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses this anomaly score to adjust the hidden state update process. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. The hidden state sequence and the anomaly score sequence are input into the dual-channel attention module to generate a coupled attention weight matrix; The hidden state sequence is weighted and aggregated based on the coupled attention weight matrix to generate a context vector; the context vector is input to a fully connected output network to decode the thermal runaway risk probability at the current moment; when the thermal runaway risk probability at the current moment exceeds the warning threshold at the current moment, a thermal runaway warning signal is triggered.

[0004] In some optional embodiments, the dual-channel attention module includes a temporal attention channel for fusing the hidden state sequence and the anomaly score sequence information to calculate the attention weights at each time step, and a feature attention channel for calculating the attention weights of each feature dimension in the hidden state sequence; the temporal attention weights and the feature attention weights are multiplied by a Hadamard product to generate a coupled attention weight matrix.

[0005] In some optional embodiments, the step of performing multi-head one-dimensional convolution processing on the original feature sequence to extract and fuse local temporal features under different receptive fields to generate a fused feature sequence includes:

[0006] The original feature sequence is processed by multiple parallel one-dimensional convolutional layers with different kernel sizes. After processing by a non-linear activation function, the outputs of the multiple one-dimensional convolutional layers are concatenated along the feature dimension to generate the fused feature sequence.

[0007] In some optional embodiments, the anomaly response gate calculates an anomaly score for the current time step based on the deviation of the input feature at the current time step from the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the update process of the hidden state, including: Let the input features at the current time step be... The average value within a historical sliding window of a preset length is Calculate the degree of deviation between the two. ; The degree of deviation Anomaly scores are generated by normalization using activation functions. ,in For the Sigmoid function, and These are learnable parameters; Based on the anomaly score Standard update gate and candidate hidden state Update hidden status The updated formula is: ; in This is the hidden state from the previous time step. It represents the Hadamardi (or Hadama) stack.

[0008] In some optional embodiments, in the temporal attention channel, the hidden state vector of each time step is concatenated with the corresponding anomaly score, and the attention weight of each time step is calculated through a fully connected network and normalization processing. In the feature attention channel, the hidden state sequence is pooled along the time step dimension to aggregate temporal information and generate a feature summary vector; the feature summary vector is input into a fully connected network and normalized to obtain the attention weight for each feature dimension.

[0009] In some optional embodiments, the step of weighting and aggregating the hidden state sequence based on the coupled attention weight matrix to generate a context vector includes: The coupled attention weight matrix is ​​multiplied element-wise with the hidden state sequence, and the result is summed along the time step dimension to generate the context vector.

[0010] In some optional embodiments, the step of inputting the context vector into a fully connected output network and decoding it to obtain the current thermal runaway risk probability includes: The fully connected output network includes at least one hidden layer with a nonlinear activation function and an output layer with an activation function that can output probability values, and decodes to obtain the thermal runaway risk probability.

[0011] In some optional embodiments, the method further includes: Based on the current probability of thermal runaway risk The probability of thermal runaway compared to the previous moment Calculate the rate of change of probability ; Based on the probability change rate, a preset basic threshold is applied. Adjustments will be made to lower the warning threshold when the probability of risk is on the rise. The calculation formula is as follows: ; in, The warning threshold at the current moment, where k is a preset adjustment coefficient and k>0. To correct the linear unit function; Adjusted warning thresholds Constrained within a preset range Inside.

[0012] In the second aspect, this disclosure proposes an early warning system for thermal runaway of automotive power batteries, comprising the following modules: The extraction module is used to acquire data on the voltage, total voltage, current, and temperature of individual cells of automotive power batteries, forming an original feature sequence with a preset time length; the original feature sequence is subjected to multi-head one-dimensional convolution processing to extract and fuse local time-series features under different time receptive fields to generate a fused feature sequence. The calculation module is used to input the fused feature sequence into the anomaly response gating loop unit for time-series encoding. The anomaly response gating loop unit adds an anomaly response gate in addition to the standard update gate and reset gate. The anomaly response gate calculates the anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the update process of the hidden state. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. The generation module is used to input the hidden state sequence and the anomaly score sequence into the dual-channel attention module to generate a coupled attention weight matrix; The early warning module is used to perform weighted aggregation on the hidden state sequence based on the coupled attention weight matrix to generate a context vector; input the context vector into a fully connected output network to decode the thermal runaway risk probability at the current moment; and trigger a thermal runaway early warning signal when the thermal runaway risk probability at the current moment exceeds the early warning threshold at the current moment.

[0013] This disclosure utilizes multi-head one-dimensional convolution to deeply explore the local temporal correlations of battery data from multiple scale perspectives, extracting more state features. Combined with a gated recurrent unit that adds an anomaly response gate, it enhances the model's ability to detect weak anomaly signals in the early stages of thermal runaway, amplifying key deviation signs while suppressing interference from normal fluctuation data. By employing a dual-channel attention mechanism of temporal and feature data, the model can focus on key time nodes and core feature dimensions that contribute most to early warning decisions, achieving the filtering and aggregation of large amounts of temporal information and improving the accuracy of model decisions. Furthermore, by combining the absolute value and rate of change of risk probability to jointly determine the early warning threshold, it can sensitively respond to the rapid accumulation of risk while avoiding frequent false alarms under stable high-risk conditions, ensuring the timeliness and reliability of early warning results. Attached Figure Description

[0014] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram illustrating the principle of multi-head one-dimensional convolution feature extraction. Figure 3 This is a schematic diagram illustrating the process from data collection to triggering an early warning. Figure 4 This is a schematic diagram of the dual-channel attention mechanism and threshold decision. Detailed Implementation

[0015] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this disclosure.

[0016] In the first embodiment, this disclosure proposes a method for early warning of thermal runaway in automotive power batteries, such as... Figure 1 This includes the following steps: S1. Acquire data on the voltage, total voltage, current, and temperature of individual cells of the automotive power battery to form an original feature sequence with a preset time length; perform multi-head one-dimensional convolution processing on the original feature sequence to extract and fuse local temporal features under different receptive fields to generate a fused feature sequence. Specifically, the vehicle battery management system (BMS) collects the voltage and temperature of all individual cells, as well as the total voltage and current of the battery pack, and sends the data to the vehicle data acquisition terminal via the CAN bus. The data acquisition terminal combines all the data collected at each moment into a feature vector, and continuously extracts the feature vectors of the past 60 time steps to obtain an original feature sequence matrix multiplied by the number of features.

[0017] Three parallel one-dimensional convolutional layers with different kernel sizes of 3, 5, and 7 are used to detect features at different time scales. The original feature sequence matrix is ​​simultaneously input into the three convolutional layers, and each convolutional layer performs sliding computation along the time step dimension to extract local temporal feature maps for short, medium, and long time spans, respectively. The feature maps output by the three convolutional layers are concatenated along the feature dimension to obtain a fused feature sequence containing multi-scale temporal information. (See [link to documentation]). Figure 2 .

[0018] More specifically, in order to extract features at different time scales from the original time-series data, in an optional embodiment, the original feature sequence is subjected to multi-head one-dimensional convolution processing to extract and fuse local temporal features under different temporal receptive fields to generate a fused feature sequence, including: The original feature sequence is processed by multiple parallel one-dimensional convolutional layers with different kernel sizes. After processing by a non-linear activation function, the outputs of the multiple one-dimensional convolutional layers are concatenated along the feature dimension to generate the fused feature sequence.

[0019] Three parallel one-dimensional convolutional layers were configured, preferably using kernels of sizes 3, 5, and 7, as described above. This multi-branch structure enables the simultaneous detection of short-, medium-, and long-term dependencies in battery data. For example, small kernels can identify instantaneous spikes in voltage or current, while large kernels can detect slow temperature increases. Each convolutional layer contains multiple kernels.

[0020] The input consists of a raw feature sequence with 100 time steps and eight monitored variables, such as voltage, current, and temperature, with a data dimension of 100×8. This sequence is simultaneously fed into three parallel convolutional layers. The first layer uses 32 convolutional kernels of size 3, the second layer uses 32 convolutional kernels of size 5, and the third layer uses 32 convolutional kernels of size 7. After convolution operations and ReLU nonlinear activation function processing, each convolutional layer outputs a feature map with a dimension of 100×32. The three feature maps are then concatenated along their respective feature dimensions to generate a fused feature sequence with a dimension of 100×96.

[0021] To further improve the accuracy of early warnings and eliminate false alarms that may be caused by drastic fluctuations in battery parameters during conditions such as rapid acceleration, rapid deceleration, high-power charging, or forced energy recovery, when acquiring data on the voltage of individual battery cells, total voltage, current, and temperature of individual battery cells, data under conditions such as rapid acceleration, rapid deceleration, high-power charging, or forced energy recovery are removed, and only data under stable vehicle conditions are retained.

[0022] S2, The fused feature sequence is input into the anomaly response gating loop unit for time-series encoding. The anomaly response gating loop unit adds an anomaly response gate in addition to the standard update gate and reset gate. The anomaly response gate calculates the anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the update process of the hidden state. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. Specifically, for each time step in the fused feature sequence, following the calculation method of the standard gated recurrent unit, the outputs of the update gate and reset gate are calculated based on the hidden state of the previous time step and the input features of the current time step. To calculate the anomaly score, the input features of the 10 time steps prior to the current time step are selected to form a historical sliding window, and the element-wise mean of all feature vectors within this window is calculated. The Euclidean distance between the current input feature vector and this mean vector is calculated as the degree of deviation. This deviation value is input into a fully connected layer and the subsequent Sigmoid activation function to obtain an anomaly score between 0 and 1. When updating the hidden state, the anomaly score is used to adjust the weight of the candidate hidden states of the standard gated recurrent unit. The larger the anomaly score, the higher the weight of the current input feature in the update. This unit outputs the adjusted hidden state and the calculated anomaly score at each time step. After traversing the entire sequence, a complete hidden state sequence and anomaly score sequence are obtained. In the anomaly response gating loop unit, the update gate and reset gate adopt the gating structure of the standard GRU, respectively controlling the retention ratio of historical hidden state information and the generation method of candidate hidden states. The anomaly response gate is a third type of gating unit introduced in parallel on this basis. It does not participate in the state calculation of the reset gate, but generates anomaly scores based on the deviation between the current time step input features and the historical sliding window statistical features, which are used to adjust the hidden state update process under the control of the update gate. The anomaly response gate, update gate, and reset gate are at the same gating level in the same time step, but functionally it is specifically responsible for characterizing and amplifying anomaly dynamics. In each time step, the anomaly response gating loop unit calculates the current hidden state based on the current input features and the hidden state of the previous time step through the joint calculation of the update gate, reset gate, and anomaly response gate. At the same time, the anomaly response gate independently calculates the corresponding anomaly score based on the deviation between the input features and the historical statistical features. By sequentially expanding the loop calculation along the time step, the hidden states of all time steps constitute the encoded hidden state sequence, and the anomaly scores of all time steps are synchronously collected to form the anomaly score sequence.

[0023] In some embodiments, the anomaly response gate calculates an anomaly score for the current time step based on the deviation of the input feature at the current time step from the mean of the feature within a historical sliding window, and uses the anomaly score to adjust the update process of the hidden state, including: Let the input features at the current time step be... The average value within a historical sliding window of a preset length is Calculate the degree of deviation between the two. ; The degree of deviation Anomaly scores are generated by normalization using activation functions. ,in For the Sigmoid function, and These are learnable parameters; Based on the anomaly score Standard update gate and candidate hidden state Update hidden status The updated formula is: ; in This is the hidden state from the previous time step. It represents the Hadamardi (or Hadama) stack.

[0024] This embodiment enhances the recurrent neural network's ability to detect data mutations by utilizing an exceptionally sensitive update mechanism. At each time step, the current input feature vector is calculated. The sliding window mean of the feature over the past, for example, 20 time steps The Euclidean distance between them gives the degree of deviation. This represents the degree of anomaly of the current data point compared to recent historical data. A sudden drop in voltage or a surge in temperature can cause a significant [anomaly / abnormality]. value.

[0025] Degree of deviation The anomalous score is transformed into a value between 0 and 1 by a fully connected layer and a sigmoid activation function. .when A value close to 1 indicates that the data at the current moment is highly abnormal. This occurs when updating the hidden state. At that time, a score composed of abnormal values The regulated hybrid term will replace the standard recurrent neural network update term. .if The update will rely more heavily on the original input features. Instead of candidate states generated internally by the network. Optionally, the input features are transformed into the same dimension as the hidden state through a linear mapping before participating in the hidden state update, for example, through a learnable parameter matrix. right After mapping, input the update formula into... This allows the model to force potential anomalous signals into its memory state, preventing important information from being smoothed out or forgotten during cyclical transmission, thereby improving the ability to detect weak features in the early stages of thermal runaway. Furthermore, when the anomalous score is high, the model's immediate response to abrupt changes is enhanced by incorporating current input features.

[0026] S3, The hidden state sequence and the anomaly score sequence are input into the dual-channel attention module. The module includes a temporal attention channel for fusing the information of the hidden state sequence and the anomaly score sequence to calculate the attention weights at each time step, and a feature attention channel for calculating the attention weights of each feature dimension in the hidden state sequence. The temporal attention weights and the feature attention weights are multiplied by a Hadamard product to generate a coupled attention weight matrix. Specifically, in the temporal attention channel, the hidden state sequence and the anomaly score sequence are first concatenated along the feature dimension. The concatenated sequence is then input into a multilayer perceptron network. After the network output is normalized by the Softmax function, a temporal attention weight vector representing the importance of each time step is generated. In the feature attention channel, the hidden state sequence is globally averaged along the time dimension to obtain a feature vector that can represent the information of the entire sequence. This vector is then input into another multilayer perceptron network. After the output is processed by the Sigmoid function, a feature attention weight vector representing the importance of each feature dimension is generated. The temporal attention weight vector and the feature attention weight vector are expanded into matrices of the same size as the hidden state sequence through a broadcast mechanism. Then, the two matrices are multiplied element-wise, i.e., the Hadamard product, to obtain the coupled attention weight matrix.

[0027] In an optional embodiment, the hidden state sequence and the anomaly score sequence are input into a dual-channel attention module. The module includes a temporal attention channel for fusing the hidden state sequence and the anomaly score sequence information to calculate the attention weights at each time step, and a feature attention channel for calculating the attention weights for each feature dimension in the hidden state sequence. In the temporal attention channel, the hidden state vector of each time step is concatenated with the corresponding anomaly score, and the attention weight of each time step is calculated through a fully connected network and normalization processing. In the feature attention channel, the hidden state sequence is pooled along the time step dimension to aggregate temporal information and generate a feature summary vector; The feature summary vector is input into a fully connected network and normalized to obtain the attention weights for each feature dimension.

[0028] This implementation evaluates the importance of information in parallel from both temporal and feature dimensions. In the temporal attention channel, for a sequence of length 100 and a hidden state dimension of 64, the hidden state vector at each time step is... It will be related to the abnormal score at that moment. The sequences are concatenated into a 65-dimensional vector. The entire sequence forms a 100×65 matrix, which is then input into a multilayer perceptron. The matrix is ​​normalized in the time dimension using the Softmax function, generating a 100-dimensional weight vector. This weight vector highlights the most critical time points for predicting thermal runaway, such as the moment when the temperature begins to rise sharply.

[0029] In the feature attention channel, the 100×64 hidden state sequence undergoes global average pooling along the time dimension to obtain a 64-dimensional feature summary vector, which represents the overall feature representation within the entire time window. This summary vector is input into another independent multilayer perceptron and processed by a sigmoid or softmax function to generate a 64-dimensional weight vector. Each element corresponds to an importance score for a feature dimension, enabling the model to identify which monitored variables are important, such as the temperature of a specific battery cell or the total current of the battery pack, which is crucial for risk assessment.

[0030] S4, the hidden state sequence is weighted and aggregated based on the coupled attention weight matrix to generate a context vector; the context vector is input to the fully connected output network to decode the thermal runaway risk probability at the current moment; when the thermal runaway risk probability at the current moment exceeds the warning threshold at the current moment, a thermal runaway warning signal is triggered.

[0031] The coupling attention weight matrix and the hidden state sequence matrix are multiplied element-wise to obtain a weighted hidden state sequence. The value of each feature dimension at each time step in this sequence is scaled according to its importance. The vectors of all time steps in the weighted hidden state sequence are summed to compress and aggregate the temporal information into a fixed-dimensional context vector.

[0032] The context vector is taken as input and passed to a multilayer perceptron network consisting of two fully connected layers. The first fully connected layer uses the ReLU activation function for nonlinear transformation, and its output is passed to the second fully connected layer. The second fully connected layer, the output layer, contains only one neuron and uses the Sigmoid activation function to compress the output value to the range of 0 to 1. The output value represents the probability of thermal runaway at the current moment. See also Figure 3 After data collection, multi-scale feature extraction, abnormal response encoding, dual-channel attention coupling, risk probability calculation, and dynamic threshold adjustment, it is determined whether an alert has been triggered. If it has, an alert is issued, and the alert method includes, but is not limited to, voice.

[0033] Optionally, store the risk probability value calculated at the previous moment. And obtain the risk probability value at the current moment. The difference between the two yields the rate of change of probability. Set a fixed basic early warning threshold. The value is 0.8, and an adjustment coefficient k is 0.5, obtained through formulas such as... To calculate the warning threshold at the current moment If the current risk probability Greater than the warning threshold If a thermal runaway warning command is generated immediately, the vehicle controller will be notified via the CAN bus to illuminate the warning lights and send a warning message to the driver; the current risk probability will be... Save as , used for calculations at the next time step.

[0034] To fuse the results of temporal attention and feature attention, the hidden state sequence is weighted. In an optional embodiment, the weighted aggregation of the hidden state sequence based on the coupled attention weight matrix to generate a context vector includes: The coupled attention weight matrix is ​​multiplied element-wise with the hidden state sequence, and the result is summed along the time step dimension to generate the context vector.

[0035] A coupled attention weight matrix is ​​generated by calculating the outer product of the temporal attention weight vector and the feature attention weight vector. For example, the outer product of a 100-dimensional temporal weight vector and a 64-dimensional feature weight vector yields a 100×64 coupled weight matrix. Each element in this matrix reflects the importance of both the corresponding time step and the feature dimension. The generated 100×64 coupled attention weight matrix is ​​then multiplied element-wise with the hidden state sequence, which also has a dimension of 100×64. This operation readjusts the salience of each value in the hidden state sequence according to the combined weights, enhancing key features at critical moments while suppressing secondary information. The weighted 100×64 matrix is ​​then summed along the time step dimension, compressing the sequence information into a unique 64-dimensional context vector. This context vector contains the information most relevant to the risk of thermal runaway in the entire input sequence.

[0036] In an optional embodiment, the step of inputting the context vector into a fully connected output network and decoding it to obtain the current thermal runaway risk probability includes: The fully connected output network includes at least one hidden layer with a nonlinear activation function and an output layer with an activation function that can output probability values, and decodes to obtain the thermal runaway risk probability.

[0037] This network acts as a decoder, mapping the generated context vector to specific risk probability values. Preferably, the network is a multilayer perceptron. For example, the input is a 64-dimensional context vector generated in the previous stage. This vector enters a fully connected hidden layer containing 32 neurons and is processed using the ReLU nonlinear activation function. The output of the hidden layer, a 32-dimensional vector, is fed into an output layer containing only a single neuron. This output layer uses the Sigmoid activation function. The Sigmoid function maps input values ​​to between 0 and 1, and the output value can be interpreted as a probability. Therefore, the output of this neuron is a scalar value representing the probability of thermal runaway occurring within the current evaluation time window. For example, an output of 0.85 indicates that the model predicts an 85% probability of thermal runaway.

[0038] In an optional embodiment, it further includes: Based on the current probability of thermal runaway risk The probability of thermal runaway compared to the previous moment Calculate the rate of change of probability ; Based on the probability change rate, a preset basic threshold is applied. Adjustments will be made to lower the warning threshold when the probability of risk is on the rise. The calculation formula is as follows: ; in, Let k be the warning threshold at the current moment, and k be a preset adjustment coefficient, where k > 0. To correct the linear unit function; Adjusted warning thresholds Constrained within a preset range Inside.

[0039] Assuming a base threshold For example, 0.8, and an adjustment coefficient k, for example, 5.0. At time t, the model outputs the risk probability. .like It is 0.75. If the value is 0.70, then the rate of change of probability is... A positive value of 0.05 indicates that the risk is increasing. The result of correcting the linear unit function ReLU(0.05) is still 0.05.

[0040] According to the formula, the new warning threshold It is 0.75. Because... Equal to the adjusted threshold This will trigger an alert. In contrast, if a fixed threshold of 0.8 is used, no alert will be triggered. If the risk probability decreases or remains unchanged, Less than or equal to zero If the threshold is zero, the threshold will remain the base threshold. To ensure system stability, the calculated threshold will be... This will be limited to a preset, reasonable range, such as [0.6, 0.9], to avoid the threshold becoming too high or too low due to drastic fluctuations in probability. Figure 4 .

[0041] In a second embodiment, this disclosure also provides an automotive power battery thermal runaway early warning system, characterized in that it includes the following modules: The extraction module is used to acquire data on the voltage, total voltage, current, and temperature of individual cells of automotive power batteries, forming an original feature sequence with a preset time length; the original feature sequence is subjected to multi-head one-dimensional convolution processing to extract and fuse local time-series features under different time receptive fields to generate a fused feature sequence. The calculation module is used to input the fused feature sequence into the anomaly response gating loop unit for time-series encoding. The anomaly response gating loop unit adds an anomaly response gate in addition to the standard update gate and reset gate. The anomaly response gate calculates the anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the update process of the hidden state. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. The generation module is used to input the hidden state sequence and the anomaly score sequence into the dual-channel attention module to generate a coupled attention weight matrix; The early warning module is used to perform weighted aggregation on the hidden state sequence based on the coupled attention weight matrix to generate a context vector; input the context vector into a fully connected output network to decode the thermal runaway risk probability at the current moment; and trigger a thermal runaway early warning signal when the thermal runaway risk probability at the current moment exceeds the early warning threshold at the current moment.

[0042] In a third embodiment, this disclosure also provides a thermal runaway warning program for automotive power batteries, which, when executed by a processor, implements the method described in the first embodiment.

[0043] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0044] The above description of the disclosed embodiments enables those skilled in the art to make or use this disclosure. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for thermal runaway early warning of an automotive traction battery, characterized in that, Includes the following steps: Data on the voltage, total voltage, current, and temperature of individual cells in an automotive power battery are acquired to form an original feature sequence with a preset time length. The original feature sequence is then subjected to multi-head one-dimensional convolution processing to extract and fuse local temporal features under different receptive fields at different times, generating a fused feature sequence. The fused feature sequence is input into an anomaly response gating loop unit for temporal encoding. This unit adds an anomaly response gate in addition to the standard update and reset gates. The anomaly response gate calculates an anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses this anomaly score to adjust the hidden state update process. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. The hidden state sequence and the anomaly score sequence are input into the dual-channel attention module to generate a coupled attention weight matrix; The hidden state sequence is weighted and aggregated based on the coupled attention weight matrix to generate a context vector; The context vector is input into a fully connected output network and decoded to obtain the probability of thermal runaway at the current moment. When the probability of thermal runaway at the current moment exceeds the warning threshold at the current moment, a thermal runaway warning signal is triggered.

2. The method of claim 1, wherein, The dual-channel attention module includes a temporal attention channel for fusing the hidden state sequence and the anomaly score sequence information to calculate the attention weights at each time step, and a feature attention channel for calculating the attention weights of each feature dimension in the hidden state sequence. Perform a Hadamard product between the temporal attention weights and the feature attention weights to generate a coupled attention weight matrix.

3. The method of claim 1, wherein, The process of performing multi-head one-dimensional convolution on the original feature sequence to extract and fuse local temporal features under different receptive fields to generate a fused feature sequence includes: The original feature sequence is processed by multiple parallel one-dimensional convolutional layers with different kernel sizes. After processing by a non-linear activation function, the outputs of the multiple one-dimensional convolutional layers are concatenated along the feature dimension to generate the fused feature sequence.

4. The method of claim 1, wherein, The anomaly response gate calculates an anomaly score for the current time step based on the deviation of the input feature at the current time step from the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the hidden state update process, including: Let the input feature of the current time step be , the mean value in the preset length of the historical sliding window be , and the deviation degree of the two be ; The degree of deviation Anomaly scores are generated by normalization using activation functions. ,in For the Sigmoid function, and These are learnable parameters; Based on the anomaly score Standard update gate and candidate hidden state Update hidden status The updated formula is: in This is the hidden state from the previous time step. It represents the Hadamardi (or Hadama) stack.

5. The method according to claim 2, characterized in that, In the temporal attention channel, the hidden state vector of each time step is concatenated with the corresponding anomaly score, and the attention weight of each time step is calculated through a fully connected network and normalization processing. In the feature attention channel, the hidden state sequence is pooled along the time step dimension to aggregate temporal information and generate a feature summary vector; the feature summary vector is input into a fully connected network and normalized to obtain the attention weight for each feature dimension.

6. The method according to claim 2 or 5, characterized in that, The step of weighting and aggregating the hidden state sequence based on the coupled attention weight matrix to generate a context vector includes: The coupled attention weight matrix is ​​multiplied element-wise with the hidden state sequence, and the result is summed along the time step dimension to generate the context vector.

7. The method according to claim 1, characterized in that, The step of inputting the context vector into a fully connected output network and decoding it to obtain the current thermal runaway risk probability includes: The fully connected output network includes at least one hidden layer with a nonlinear activation function and an output layer with an activation function that can output probability values, and decodes to obtain the thermal runaway risk probability.

8. The method according to claim 1, characterized in that, Also includes: Based on the current probability of thermal runaway risk The probability of thermal runaway compared to the previous moment Calculate the rate of change of probability ; Based on the probability change rate, a preset basic threshold is applied. Adjustments will be made to lower the warning threshold when the probability of risk is on the rise. The calculation formula is as follows: ; in, The warning threshold at the current moment, where k is a preset adjustment coefficient and k>

0. To correct the linear unit function; Adjusted warning thresholds Constrained within a preset range Inside.

9. A thermal runaway early warning system for automotive power batteries, characterized in that, Includes the following modules: The extraction module is used to acquire data on the voltage, total voltage, current, and temperature of individual cells of automotive power batteries, forming an original feature sequence with a preset time length; the original feature sequence is subjected to multi-head one-dimensional convolution processing to extract and fuse local time-series features under different time receptive fields to generate a fused feature sequence. The calculation module is used to input the fused feature sequence into the anomaly response gating loop unit for time-series encoding. The anomaly response gating loop unit adds an anomaly response gate in addition to the standard update gate and reset gate. The anomaly response gate calculates the anomaly score for the current time step based on the deviation between the input feature at the current time step and the mean of the feature within the historical sliding window, and uses the anomaly score to adjust the update process of the hidden state. The anomaly response gating loop unit outputs the encoded hidden state sequence and the corresponding anomaly score sequence. The generation module is used to input the hidden state sequence and the anomaly score sequence into the dual-channel attention module to generate a coupled attention weight matrix; The early warning module is used to perform weighted aggregation on the hidden state sequence based on the coupled attention weight matrix to generate a context vector; The context vector is input into a fully connected output network and decoded to obtain the probability of thermal runaway at the current moment. When the probability of thermal runaway at the current moment exceeds the warning threshold at the current moment, a thermal runaway warning signal is triggered.

10. The system according to claim 9, characterized in that, The dual-channel attention module includes a temporal attention channel for fusing the hidden state sequence and the anomaly score sequence information to calculate the attention weights at each time step, and a feature attention channel for calculating the attention weights of each feature dimension in the hidden state sequence. Perform a Hadamard product between the temporal attention weights and the feature attention weights to generate a coupled attention weight matrix.

Images