A battery thermal runaway failure prediction method based on abnormal heat generation and deep learning

Abstract

The application relates to a power battery thermal runaway fault prediction method for an electric vehicle, in particular to a battery thermal runaway early warning method based on abnormal heat generation and a deep learning algorithm. The method combines an improved long short-term memory neural network and a convolutional neural network to extract the space-time features in the temperature evolution process of the battery, and realizes high-precision multi-step temperature prediction; further based on a heat generation rate calculation model, the abnormal heat generation behavior of the battery is diagnosed, so that early prediction of thermal runaway is realized.

Description

Technical Field

[0001] This invention belongs to the field of electric vehicle battery safety management technology, specifically relating to a battery thermal runaway fault prediction method based on abnormal heat generation and deep learning. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles due to their high energy density and long lifespan. However, they are prone to thermal runaway under abuse conditions such as overcharging, overheating, and internal short circuits, leading to serious safety accidents. Traditional battery management systems (BMS) rely heavily on rule-based thresholds for fault detection, making it difficult to identify potential thermal runaway risks in their early stages. Existing temperature prediction methods are mostly based on laboratory environments, with short prediction step sizes and decreasing accuracy as the step size increases. Furthermore, they fail to effectively couple multiple influencing factors (such as vehicle status, driving behavior, and ambient temperature), limiting their applicability under complex real-world operating conditions. Summary of the Invention

[0003] To address the shortcomings of the existing technologies, this invention provides a method for early and accurate prediction of battery thermal runaway faults. By combining an improved deep learning algorithm with a thermal generation physics model, features are extracted from multidimensional data and online prediction is performed, significantly extending the warning time and ensuring vehicle safety.

[0004] A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning includes the following steps: Step 1, Data Acquisition and Preprocessing: Acquire historical operating data of electric vehicles, including battery parameters such as battery temperature, battery voltage, and current internal resistance; vehicle status parameters such as vehicle speed, motor speed, and SOC; driving behavior parameters such as accelerator pedal and brake pedal; and meteorological parameters such as local temperature, humidity, and air pressure. Preprocess the historical operating data, including missing value imputation, outlier correction, and noise smoothing.

[0005] Step 2, Feature dimensionality reduction: Principal component analysis is used to compress the preprocessed historical data and extract a few principal components with a cumulative contribution rate of over 95% as input to the CNN-LSTM neural network model. This can reduce computational complexity while retaining most of the information.

[0006] Preferably, the input to the CNN-LSTM neural network model is:

[0007] in, Let covariance matrix be the variance matrix. For historical operating data of electric vehicles related to battery temperature, including battery temperature and vehicle status driving behavior—local weather factors across all time steps—the expression is as follows:

[0008] Where n and m represent the factor and the time step, respectively; This represents the temperature of the battery cell q at time step t. This represents the j-th factor at time step t.

[0009] Step 3, Battery Temperature Prediction: An improved CNN-LSTM neural network model is constructed. This model extracts multi-parameter spatial features through CNN layers and extracts long-term dependencies of time series through improved LSTM layers. The model hyperparameters are automatically optimized using a stochastic neighborhood optimization method. The model uses a sliding window approach to input dimensionality-reduced historical data and outputs predicted battery temperatures for the next 48 time steps (corresponding to 8 minutes).

[0010] Preferably, in the improved CNN-LSTM model, the memory cells of the LSTM are modified to store the state of the memory cells from the previous time step. Introduced as additional input into the calculations of the forget gate, input gate, and output gate, it enhances the ability to remember key temporal information. The calculation process is as follows:

[0011]

[0012]

[0013]

[0014]

[0015]

[0016] in, and These represent the memory cell state and output from the previous time step, respectively. This is the current input; sigmoid is the activation function, and tanh is the hyperbolic tangent activation function; W f , W i , W c , W o These are the weight matrices for the corresponding gates (forget gate, input gate, candidate memory, output gate). b f , b i, b c , b o These are the bias terms for each gate; The output of the forget gate controls the memory from the previous moment. How much should be retained? The input gate output controls the current candidate memory. How much should be retained? The current candidate memory is the information to be stored, calculated based on the current input and the previous state. The current state of the memory cell is obtained by updating it together with the forget gate and the input gate; Output gate output, controls the current memory c t How many are in a hidden state; y t = h t This is the current output (hidden state), used for the next time step.

[0017] Preferably, a random neighborhood optimization method is used to automatically determine the hyperparameters of the improved CNN-LSTM model, including the number of convolutional kernels, the kernel size, and the sliding window size.

[0018] The stochastic neighborhood optimization method searches for the optimal combination of hyperparameters that minimizes the average relative error of temperature prediction by randomly searching for neighboring points in the hyperparameter grid space and evaluating their performance. This avoids the subjectivity and inefficiency of manual parameter tuning.

[0019] Step 4, Calculation of heat generation rate: Based on the predicted battery temperature, the real-time heat generation rate of the battery is calculated using a model-based scheme. The model-based scheme calculates the battery heat absorption rate, the thermal conductivity with adjacent batteries and heat sinks, and the convective heat dissipation rate with the environment based on the heat balance equation, and then reverse-engineers the real-time battery heat generation rate.

[0020] Preferably, the model base scheme calculates the heat generation rate based on the following heat balance equation:

[0021] Among them, Q ge The total heat generation rate of the battery; Q ab The battery's heat absorption rate is calculated using the following formula:

[0022] Among them, C q For the specific heat capacity of the battery, mq For the mass of a single battery cell, Battery temperature T q The instantaneous rate of change with respect to time t, and These are differential symbols, representing minute changes in temperature and minute intervals of time, respectively, rather than the variables themselves; ΔT q Let Δt be the temperature change of the battery during the sampling time interval, and ΔT be the sampling time interval. q / Δt is the average rate of temperature change within the sampling time interval Δt. It is a discrete approximation of the differential form and is used for practical measurement and numerical calculation.

[0023] Q dis The heat dissipation rate between the battery and the external environment is calculated using the following formula:

[0024] Q cond The conduction heat dissipation rate is calculated based on Fourier's law, using Q as the coefficient of performance. conv The convective heat dissipation rate is calculated based on Newton's cooling formula between the battery surface and the ambient air.

[0025] Step 5, Diagnosis and Early Warning of Abnormal Heat Generation: The heat generation rate is compared with a pre-established database of heat generation rates under normal operating conditions, based on simulation. When the residual ΔQ ge When M (where M is a threshold determined by statistics, such as 0.935) is reached, it is determined to be abnormal heat generation, triggering a thermal runaway warning.

[0026] This invention tightly integrates an improved deep learning model with a battery thermophysics model, employing a data-driven approach to learn the spatiotemporal evolution of battery temperature from real vehicle operating data. It utilizes abnormal heat generation as a key intrinsic physical indicator for fault diagnosis, achieving early and accurate prediction of battery thermal runaway. Compared to traditional threshold alarms or laboratory models, this method can predict thermal runaway 27 minutes earlier in real-world data validation, significantly extending the warning time. Furthermore, the model exhibits good robustness and adaptability to complex real-world operating conditions, effectively improving the safety performance of electric vehicles.

[0027] The present invention has the following advantages: 1. Early warning time: Through a two-stage strategy of "temperature prediction + abnormal heat generation diagnosis", internal faults can be identified before the battery temperature becomes obviously abnormal, achieving an early warning of up to 27 minutes.

[0028] 2. High prediction accuracy: The improved CNN-LSTM model fully exploits the spatiotemporal features of the data, maintaining high accuracy for multi-step temperature prediction.

[0029] 3. Strong adaptability: The model adopts a stochastic neighborhood optimization method to automatically optimize hyperparameters, which reduces the model's dependence on manual parameter tuning and enhances its applicability to different vehicle models and operating conditions.

[0030] 4. Excellent engineering applicability: The model can be trained offline based on historical data and updated and fine-tuned online using real-time vehicle data, making it highly suitable for practical engineering applications. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of predicting battery thermal runaway faults according to the present invention. Figure 2 This is a schematic diagram of the improved LSTM memory cell structure in this invention; Figure 3 This is a schematic diagram of the configuration of the improved CNN-LSTM in this invention; Figure 4 This is a schematic diagram of the random neighborhood optimization method used in this invention. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0033] like Figure 1 As shown, the battery thermal runaway fault prediction method based on abnormal heat generation and deep learning includes five steps: data acquisition and preprocessing, feature dimensionality reduction, battery temperature prediction, heat generation rate calculation, and anomaly diagnosis. Step 1: Data Preprocessing Acquire historical operating data of electric vehicles, including battery parameters such as battery temperature, battery voltage, and current internal resistance; vehicle status parameters such as vehicle speed, motor speed, and SOC; driving behavior parameters such as accelerator pedal and brake pedal; and meteorological parameters such as local temperature, humidity, and air pressure. Preprocess the historical operating data, including missing value imputation, outlier correction, and noise smoothing.

[0034] Step 2: Feature dimensionality reduction: Principal component analysis was used to compress the preprocessed historical data, and a few principal components with a cumulative contribution rate of over 95% were extracted as input to the CNN-LSTM neural network model. The data was compressed to five main features to reduce computation and prevent overfitting.

[0035] The input to the CNN-LSTM neural network model is:

[0036] Among them, P (n+1) ×K Let H be the covariance matrix. m×(n+1)For historical operating data of electric vehicles related to battery temperature, including battery temperature and vehicle status driving behavior—local weather factors across all time steps—the expression is as follows:

[0037] Where n and m represent the factor and time step, respectively; T t, q (t=1, 2, …, m) represents the temperature of battery cell q at time step t, F t，j (t=1, 2, …,m;j=1, 2 ,…, n) represents the j-th factor at time step t.

[0038] Step 3: Battery Temperature Prediction An improved CNN-LSTM neural network model is constructed. This model extracts multi-parameter spatial features through CNN layers and extracts long-term dependencies of time series through improved LSTM layers. The model hyperparameters are automatically optimized using a stochastic neighborhood optimization method. The model uses a sliding window approach to input dimensionality-reduced historical data and outputs predicted battery temperatures for the next 48 time steps (corresponding to 8 minutes).

[0039] The dimensionality-reduced data is then fed into the improved CNN-LSTM model for training and prediction. For example... Figure 3 As shown, the model structure includes an input layer, two convolutional layers, two pooling layers, three improved LSTM layers, two fully connected layers, and an output layer. The model's hyperparameters (such as the number and size of convolutional kernels, and the sliding window size) are automatically optimized using a stochastic neighborhood optimization method. The trained model can be used to predict battery temperature over the next 48 time steps (i.e., 8 minutes), with an average relative error of 0.28%.

[0040] In the improved CNN-LSTM model, the memory units of the LSTM (such as...) Figure 2 The state of the memory cell c from the previous moment (as shown) has been modified. t-1 Introduced as additional input into the calculations of the forget gate, input gate, and output gate, it enhances the ability to remember key temporal information. The calculation process is as follows:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] Among them, c t-1 and h t-1 These represent the memory cell state and output from the previous time step, x. t This is the current input; σ is the sigmoid activation function, and tanh is the hyperbolic tangent activation function; W f W i W c W o These are the weight matrices for the corresponding gates (forget gate, input gate, candidate memory, output gate), b f , b i , b c , b o These are the bias terms for each gate; f t The output of the forget gate controls the memory c from the previous time step. t-1 How much should be retained? i t The input gate output controls how much of the current candidate memory g2 is retained; g2 is the current candidate memory, which is the information to be stored calculated based on the current input and the previous state. c t The current state of the memory cell is obtained by updating it together with the forget gate and the input gate; o t Output gate output, controls the current memory c t How many are in a hidden state; y t = h t This is the current output (hidden state), used for the next time step.

[0047] like Figure 4 As shown, the stochastic neighborhood optimization method is used to automatically determine the hyperparameters of the improved CNN-LSTM model, including the number of convolutional kernels, the kernel size, and the sliding window size.

[0048] Step 4: Calculation of heat generation rate: Based on the predicted battery temperature, the real-time heat generation rate of the battery is calculated using a model-based scheme. The model-based scheme calculates the battery heat absorption rate, the thermal conductivity with adjacent batteries and heat sinks, and the convective heat dissipation rate with the environment based on the heat balance equation, and then reverse-engineers the real-time battery heat generation rate.

[0049] The model-based scheme calculates the heat generation rate based on the following heat balance equation:

[0050] Among them, Q ge The total heat generation rate of the battery; Q ab The battery's heat absorption rate is calculated using the following formula:

[0051] Among them, C q For the specific heat capacity of the battery, m q For the mass of a single battery cell, Battery temperature T q The instantaneous rate of change with respect to time t, and These are differential symbols, representing minute changes in temperature and minute intervals of time, respectively, rather than the variables themselves; ΔT q Let Δt be the temperature change of the battery during the sampling time interval, and ΔT be the sampling time interval. q / Δt is the average rate of temperature change within the sampling time interval Δt. It is a discrete approximation of the differential form and is used for practical measurement and numerical calculation.

[0052] Q dis The heat dissipation rate between the battery and the external environment is calculated using the following formula:

[0053] Q cond The conduction heat dissipation rate is calculated based on Fourier's law, using Q as the coefficient of performance. conv The convective heat dissipation rate is calculated based on Newton's cooling formula between the battery surface and the ambient air.

[0054] Step 5, Diagnosis and Early Warning of Abnormal Heat Generation: Q ge,real The heat generation rate Q under normal operating conditions, pre-calculated using the Simulink electrothermal model. ge,normal Compare and calculate the residual ΔQ ge If ΔQ ge If M (where M is a threshold obtained from historical normal data statistics, such as 0.935) is reached, then abnormal heat generation is determined to have occurred, and the system triggers a thermal runaway warning.

[0055] This implementation method has been verified on real electric vehicle operating data. It can predict thermal runaway 27 minutes in advance, and has high accuracy, strong robustness and real-time performance, which significantly improves the safety of electric vehicles.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: Acquire historical operating data of electric vehicles and preprocess the historical operating data; Step 2, Feature dimensionality reduction: Principal component analysis was used to compress the preprocessed historical data and extract a few principal components with a cumulative contribution rate of over 95% as input to the CNN-LSTM neural network model. Step 3, Battery Temperature Prediction: An improved CNN-LSTM neural network model is constructed. This model extracts multi-parameter spatial features through CNN layers and extracts long-term dependencies of time series through improved LSTM layers. The model hyperparameters are automatically optimized using a stochastic neighborhood optimization method. Historical data after dimensionality reduction is input using a sliding window method, and the predicted battery temperature values ​​for the next 48 time steps are output. Step 4, Calculation of heat generation rate: Based on the predicted battery temperature, the real-time heat generation rate of the battery is calculated using the model-based scheme. Step 5, Diagnosis and Early Warning of Abnormal Heat Generation: The heat generation rate is compared with a pre-established database of heat generation rates under normal operating conditions through simulation to diagnose and warn of abnormal heat generation.

2. The battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithm according to claim 1, characterized in that: Historical operating data includes battery parameters such as battery temperature, battery voltage, and internal resistance; vehicle status parameters such as vehicle speed, motor speed, and SOC; driving behavior parameters such as accelerator pedal and brake pedal; and meteorological parameters such as local temperature, humidity, and air pressure.

3. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms according to claim 2, characterized in that: Preprocessing includes missing value imputation, outlier correction, and noise smoothing.

4. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms according to claim 1, characterized in that: The input to the CNN-LSTM neural network model in step 2 is: in, Let covariance matrix be the variance matrix. For historical operating data of electric vehicles related to battery temperature, the expression for local weather factors at all time steps of battery temperature and vehicle state driving behavior is as follows: Where n and m represent the factor and the time step, respectively; This represents the temperature of the battery cell q at time step t. This represents the j-th factor at time step t.

5. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms according to claim 1, characterized in that: In the improved CNN-LSTM model described in step 3, the memory cells of the LSTM are modified by storing the state of the memory cells from the previous time step. Introduced as additional input into the calculations of the forget gate, input gate, and output gate, it enhances the ability to remember key temporal information. The calculation process is as follows: in, and These represent the memory cell state and output from the previous time step, respectively. This is the current input; sigmoid is the activation function, and tanh is the hyperbolic tangent activation function; , , , These are the weight matrices corresponding to the forget gate, input gate, candidate memory, and output gate, respectively. , , , These are the bias terms corresponding to the forget gate, input gate, candidate memory, and output gate, respectively. The output of the forget gate controls the memory from the previous moment. How much should be retained? The input gate output controls the current candidate memory. How much should be retained? The current candidate memory is the information to be stored, calculated based on the current input and the previous state. The current state of the memory cell is obtained by updating it together with the forget gate and the input gate; Output gate output, controls the current memory. How many are in a hidden state; This is the current output, to be used in the next time step.

6. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms according to claim 5, characterized in that: The stochastic neighborhood optimization method is used to automatically determine the hyperparameters of the improved CNN-LSTM model, including the number of convolutional kernels, the kernel size, and the sliding window size.

7. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithms according to claim 6, characterized in that: The model base scheme calculates the battery heat absorption rate, the heat conduction rate with adjacent batteries and heat sinks, and the convective heat dissipation rate with the environment based on the heat balance equation, and then calculates the real-time battery heat generation rate.

8. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithm according to claim 7, characterized in that: The model-based scheme calculates the heat generation rate based on the following heat balance equation: in, The total heat generation rate of the battery; The battery's heat absorption rate is calculated using the following formula: in, For the specific heat capacity of the battery, For the mass of a single battery cell, Battery temperature The instantaneous rate of change with respect to time t, and These are differential symbols, representing minute changes in temperature and minute intervals of time, respectively, rather than the variables themselves; This represents the temperature change of the battery during the sampling time interval. The sampling time interval, It is the sampling time interval. The average rate of temperature change within the range is a discrete approximation of the differential form, used for practical measurement and numerical calculation. The heat dissipation rate between the battery and the external environment is calculated using the following formula: The conduction heat dissipation rate is calculated based on Fourier's law between the battery and adjacent batteries and the heat sink. The convective heat dissipation rate is calculated based on Newton's cooling formula between the battery surface and the ambient air.

9. A battery thermal runaway fault prediction method based on abnormal heat generation and deep learning algorithm according to claim 1, characterized in that: In step 5, when the residual When M is a threshold determined through statistics, it is judged as abnormal heat generation, triggering a thermal runaway warning.

Images