Transfer learning method of power battery thermal runaway early warning large model
Through the transfer learning method of the large-scale power battery thermal runaway warning model, the problem of warning accuracy under different battery models and complex environments has been solved, rapid adaptation and efficient warning on new models have been achieved, and battery safety and reliability have been improved.
Patent Information
- Application Number
- CN202510995682.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-18
- Publication Date
- 2025-10-28
AI Technical Summary
Existing thermal runaway warning technologies lack accuracy and adaptability when dealing with different battery models or complex environments, making it difficult to provide reliable warnings.
A transfer learning method is adopted for the large power battery thermal runaway warning model. The general thermal runaway characteristics are learned through the pre-trained model, and fine-tuning technology is used to quickly adapt to the data of new vehicle models. This includes data enhancement, hierarchical fine-tuning and adaptive attention mechanism, combined with cross-entropy loss function and Bayesian posterior analysis to dynamically adjust the warning threshold.
The model's adaptability to new vehicle models and warning accuracy are improved, data requirements are reduced, the real-time and accuracy of warnings are improved, and it is suitable for battery safety monitoring in complex environments.
Smart Images
Figure CN120851130A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery management technology, and more specifically to a transfer learning method for a large-scale model of early warning of thermal runaway in power batteries. Background Technology
[0002] Power batteries are the core components of electric vehicles and energy storage devices. However, their safety, especially thermal runaway, has become a significant factor affecting battery lifespan and safety. Thermal runaway refers to the uncontrolled chemical reaction within the battery caused by external environmental factors or improper use, ultimately leading to a rapid increase in battery temperature and potentially causing serious safety incidents such as explosions or fires. Thermal runaway not only severely damages the battery but also poses a significant threat to the lives and property of drivers, passengers, and those in the surrounding area.
[0003] Therefore, predicting the risk of thermal runaway in advance and taking corresponding preventive measures has become a core task of battery management systems (BMS). BMS aims to assess the battery's operating status in a timely manner by monitoring parameters such as voltage, current, temperature, SOC (state of charge), and SOH (state of health) in real time, predicting potential faults and risks, and ensuring the safe operation of the battery. Effective thermal runaway early warning technology can provide warnings before thermal runaway occurs, giving sufficient time to take effective safety measures, thereby greatly improving battery safety and reliability.
[0004] Currently, early warning technologies for battery thermal runaway mainly include physical models, data-driven methods, and sensor-based monitoring technologies. However, these methods often have the following problems when dealing with different battery models or new vehicle types: Physical models typically rely on complex internal electrochemical and thermodynamic models of batteries to predict thermal runaway by simulating the battery's thermal behavior. While physical models can provide a certain level of predictive accuracy when the battery structure is relatively simple and operating conditions are stable, the thermal runaway process is influenced by various factors, such as battery design, operating conditions, and environmental factors. Traditional physical models struggle to fully consider these complex factors, and they also suffer from high computational demands, poor real-time performance, and are difficult to apply to dynamic and complex environments.
[0005] Data-driven approaches train machine learning models using extensive historical battery data to identify battery thermal runaway risks. While capable of uncovering potential patterns from large datasets and exhibiting high accuracy, these models suffer from poor adaptability. Existing models are often trained on specific battery types (e.g., specific car models or brands), and their accuracy drops significantly when faced with data from new car models or different environmental conditions, making it difficult to provide reliable risk predictions.
[0006] Sensor-based thermal runaway early warning methods typically rely on real-time data from temperature and current sensors to determine changes in battery status. While sensor technology can provide real-time data support, limitations in sensor accuracy, data lag, or sensor malfunction mean that relying solely on sensor data cannot comprehensively reflect the battery's health status, and it is difficult to provide accurate early warnings in complex environments.
[0007] In summary, while existing thermal runaway warning technologies have addressed battery safety issues to some extent, they still have significant limitations when dealing with new vehicle models and complex application environments, and cannot provide accurate and reliable thermal runaway warnings. Summary of the Invention
[0008] The present invention aims to provide a transfer learning method for a large-scale model for early warning of thermal runaway in power batteries. By learning common thermal runaway characteristics through pre-trained models and using fine-tuning techniques to quickly adapt to data from new vehicle models, the cross-domain adaptability and early warning accuracy of the model are improved.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: Transfer learning methods for large-scale models of thermal runaway early warning for power batteries include: S1. Use the real vehicle operation data of the power battery obtained from the cloud data platform to train a large model for thermal runaway early warning and obtain upstream pre-training parameters; S2 collects real-vehicle operating data of the power battery of the new model and preprocesses the data; S3 enhances the original real-vehicle operation data through a diffusion model; S4 combines the real-vehicle operating data of the new model's power battery with the enhanced data to obtain a new dataset; S5: Load the parameters of the pre-trained thermal runaway early warning model, fix the parameters of the bottom network, fine-tune the parameters of the high-level network using a hierarchical fine-tuning strategy, and add an adaptive attention mechanism to the high-level network. S6, cross-entropy loss function and Adam optimizer are selected to optimize the large thermal runaway early warning model, and Warmup strategy and cosine annealing strategy are combined with dynamic scheduling of learning rate; S7, use the test set to evaluate the fine-tuned thermal runaway early warning large model; S8 inputs real-time data from the new model into the finely tuned thermal runaway warning big model to obtain the probability of thermal runaway risk of the power battery; S9 combines real-time data with historical warning records and uses Bayesian posterior analysis to dynamically adjust the model and warning threshold. When the risk probability exceeds the set warning threshold, a warning is triggered.
[0010] The principle and advantages of this solution are as follows: In practical applications, a large-scale thermal runaway early warning model is trained using a large amount of real-vehicle operating data of power batteries obtained from a cloud data platform. This learns the general patterns of battery thermal runaway risk under different vehicle models, obtaining upstream pre-trained parameters. These parameters capture the common characteristics of battery thermal runaway, laying the foundation for subsequent model transfer. Real-vehicle operating data of power batteries for new vehicle models are collected, and the original data is enhanced using a diffusion model to generate samples similar to the actual data distribution, expanding the diversity of the dataset. This step helps alleviate the problem of data scarcity for new vehicle models and enhances the robustness of the model. Pre-trained model parameters are loaded, fixing the underlying network parameters to retain general features, while fine-tuning the high-level network parameters. An adaptive attention mechanism is introduced into the high-level network to enhance the model's ability to capture key features. This hierarchical fine-tuning strategy can quickly adapt to the characteristics of new vehicle models while retaining pre-trained knowledge. The model is optimized using a cross-entropy loss function and the Adam optimizer, and the learning rate is dynamically scheduled using a Warmup strategy and a cosine annealing strategy to improve the model's convergence speed. The fine-tuned model is evaluated using a test set to ensure its accuracy and generalization ability. Real-time data from the new vehicle model is input into the fine-tuned model, and the probability of thermal runaway risk is output. By combining real-time data with historical warning records, Bayesian posterior analysis is used to dynamically adjust the warning threshold to ensure the real-time performance and accuracy of the warning system.
[0011] Technical effects: (1) High efficiency in cross-domain adaptability: Through transfer learning fine-tuning technology, the existing thermal runaway model can be quickly transferred to the new vehicle model, which greatly improves the model's adaptability in unknown environments. The model can quickly adapt to the target task (new vehicle model data) by leveraging the knowledge learned from the source task (existing vehicle model data), effectively improving the prediction accuracy.
[0012] (2) Significantly reduced data requirements: Through transfer learning fine-tuning, only a small amount of target task data is needed to complete the fine-tuning, which significantly reduces the dependence on a large amount of new model data and can more flexibly deal with the data scarcity problem of different models. It is especially suitable for the situation where new models are launched in the early stage and data is not fully collected.
[0013] (3) Improved model training efficiency: By fine-tuning the high-level network and freezing the low-level feature extraction layer, this solution can reduce training time and computing resources, while improving training efficiency and model accuracy, especially when facing new car models or new environments.
[0014] (4) Enhanced accuracy and real-time performance of early warnings: Existing technologies, especially those based on traditional physical models or simple data-driven models, often fail to provide highly accurate thermal runaway warnings in complex and variable environments. This solution utilizes transfer learning fine-tuning technology to enable the model to flexibly adapt to different battery characteristics and operating conditions, thereby improving the accuracy and real-time performance of the warning. Through fine-tuning, the model can not only accurately identify the thermal runaway risk of new vehicle models but also respond quickly in real-time monitoring, avoiding the delays and errors inherent in traditional methods.
[0015] Preferably, as an improvement, the actual vehicle operating data includes voltage, current, temperature, SOC, SOH, charge / discharge status, mileage, speed, internal resistance, maximum cell voltage, and minimum cell voltage.
[0016] Technical benefits: Facilitates the comprehensive acquisition of real-vehicle operation data, providing a data foundation for subsequent early warning.
[0017] Preferably, as an improvement, the preprocessing includes data cleaning and normalization. Data cleaning includes removing missing values and outliers, and the normalization formula is:
[0018] in, Represents the original data. This represents the mean of the data. The standard deviation of the data This represents the normalized data.
[0019] Technical effect: It facilitates the elimination of dimensional differences between different features and ensures that each feature contributes equally to the training of the model.
[0020] Preferably, as an improvement, S3 includes: S31, Construct a diffusion process, using the preprocessed data as initial data, and apply a forward diffusion process to generate a series of noisy data; S32, the reverse training process, trains a neural network model to learn the mapping from noisy data to the original data, and the training objective is to maximize the variational lower bound. S33, generate new data, sample noise from the prior noise distribution, and calculate using the reverse process in sequence.
[0021] Technical effect: It facilitates the gradual addition of noise through a forward diffusion process, and the reverse process uses a neural network to learn the mapping from the noisy data to the original data, thereby achieving data augmentation.
[0022] Preferably, as an improvement, the adaptive attention mechanism calculates attention weights and performs feature weighting.
[0023] in, For attention weights, Calculated by the adaptive layer, This represents the input features.
[0024] Technical benefits: It facilitates the model's ability to capture key features.
[0025] Preferably, as an improvement, the method of combining the Warmup strategy and the cosine annealing strategy with dynamic scheduling of the learning rate includes: In the early stages of training, a warmup strategy is used to gradually increase the learning rate to the target learning rate.
[0026] In the later stages of model training, a cosine annealing strategy is used to adjust the learning rate:
[0027] in, To maximize the learning rate, To minimize the learning rate, For Warmup time, Total training time This is the current training time.
[0028] Technical effect: Using the Warmup strategy in the early stage of training avoids the model's unstable convergence due to an excessively large learning rate. The learning rate is gradually reduced in the later stage of training, which helps the model converge better.
[0029] Preferably, as an improvement, in S8, the output layer uses a sigmoid activation function to convert the risk probability output of thermal runaway into a probability value ranging from 0 to 1. The conversion model is as follows:
[0030] in, These are model parameters. It is a weight matrix. It is a bias term. It is the Sigmoid activation function. This indicates the probability of thermal runaway occurring.
[0031] Technical benefit: Improves the visibility of output results.
[0032] Preferably, as an improvement, S9 further includes: S91, set an initial risk threshold and update it using Bayesian methods:
[0033] in, This is the initial risk threshold; S92, at predetermined time windows, the error between the model's predicted probability and the actual event within the window is statistically analyzed, and the threshold is adjusted based on the error rate; S93: When the predicted probability exceeds the dynamically updated warning threshold, an early warning is triggered.
[0034] Technical benefits: Facilitates improved accuracy of early warning systems.
[0035] Preferably, as an improvement, a feedback mechanism is also included, which combines the feedback information of each warning with the current warning threshold to update the threshold and update the model as follows:
[0036] in, Indicates the error of the current window. It is the target error level. To adjust the step size.
[0037] Technical benefits: It facilitates further integration of early warning thresholds and early warning information, thereby improving the accuracy of early warnings. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the transfer learning method for a large-scale model of thermal runaway early warning for power batteries. Detailed Implementation
[0039] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: Transfer learning methods for large-scale models of thermal runaway early warning for power batteries include: S1 utilizes real-vehicle operating data of power batteries obtained from the cloud data platform to train a large-scale thermal runaway early warning model and obtain upstream pre-training parameters; the real-vehicle operating data of power batteries includes multi-dimensional information such as battery voltage, current, temperature, SOC (state of charge), SOH (state of health), charge and discharge status, mileage, speed, internal resistance, maximum cell voltage, and minimum cell voltage.
[0040] S2 collects real-vehicle operating data of the power battery of the new model and preprocesses the data.
[0041] The preprocessing includes data cleaning and normalization. Data cleaning includes removing missing and outlier values, and the normalization formula is as follows:
[0042] in, Represents the original data. This represents the mean of the data. The standard deviation of the data This represents the normalized data. Normalization helps eliminate dimensional differences between different features, ensuring a balanced contribution of each feature to the model's training.
[0043] S3, enhances the original real-vehicle operation data through a diffusion model; specifically, S3 includes: S31, Construct the diffusion process, using the preprocessed data As initial data, the forward diffusion process is applied. Generate a series of noisy data ;
[0044] in, It represents the data state at time step t, and can be a vector or tensor representing a data sample during the diffusion process; This indicates the data state at time step t−1, which is the data from the previous time step; This represents the noise figure at time step t, a scalar value between 0 and 1, used to control the intensity of the noise and determine the amount of noise added at the current time step. This indicates the proportion of data retained, used to scale the data from the previous time step to ensure that the variance of the data remains stable during diffusion. This represents a scaling factor for the noise, used to adjust the magnitude of the noise to match the variance of the data. This represents the random noise added at time step t, which is a variable... Random vectors of the same dimension are often assumed to be standard Gaussian noise (i.e., a normal distribution with a mean of 0 and a variance of 1).
[0045] This indicates that at time step t, the data retains a portion of the data from the previous time step, calculated using a scaling factor. The variance of the data remains stable during the diffusion process, preventing the data from rapidly degenerating into noise.
[0046] This part indicates that at time step t, a certain proportion of random noise was added. The intensity of the noise is determined by... Control the noise to ensure that its variance matches the data variance. (Added noise) It is typically standard Gaussian noise, used to simulate random variations in data.
[0047] In this process, noise is gradually added to each data sample, causing its distribution to gradually transition from the original data to a Gaussian noise distribution. This process can be viewed as a Markov chain, where the state at each step depends only on the state of the previous step and the added noise.
[0048] S32, Training the reverse process, through training the neural network model. This allows the model to learn the mapping from noisy data to the original data. The training objective is to maximize the variational lower bound. Maximizing the variational lower bound aims to minimize the difference between the forward and backward diffusion processes while maintaining a good fit to the original data. This enables the model to effectively recover the original data from noise, thereby achieving data augmentation and generation.
[0049] in, This represents the variational lower bound. By maximizing the variational lower bound, the model can better recover the original data from noise. Indicates the original data Expectations Sampled from the real data distribution, the desired operation is to calculate the average performance of the model across the entire dataset. The model represents the original data. The log-likelihood. It is the probability distribution of the original data predicted by the model. Maximizing the log-likelihood means making the model as close as possible to the true data distribution. The KL divergence is a measure of the difference between two probability distributions. The smaller the KL divergence value, the closer the two distributions are. It constrains the proximity of the data distribution generated by the reverse process to the prior distribution. The conditional probability distribution representing the forward diffusion process describes the distribution from the original data. Initially, through a diffusion process of a series of time steps t=1,2,…,T, the noise data is finally obtained. The probability distribution. This represents the conditional probability distribution of the reverse process.
[0050] This indicates how well the model fits the original data. By maximizing this term, the model can better predict the original data. This measures the difference between the forward and reverse diffusion processes. By maximizing this term, the model can make the reverse process as close as possible to the inverse of the forward diffusion process, thus better recovering the original data from noise.
[0051] S33, Generate new data from the prior noise distribution. Medium sampling noise Calculate using the reverse process sequentially :
[0052] The final generated data The new data matches the multidimensional distribution of the original power battery data.
[0053] S4 combines the real-vehicle operation data of the new model's power battery with the enhanced data to obtain a new dataset. This dataset is then divided into training, validation, and testing sets to ensure the model can be trained and evaluated on different datasets. The dataset division ratio is adjusted according to requirements; in this embodiment, 70% is used for training, 15% for validation, and 15% for testing.
[0054] S5 loads the parameters of the pre-trained thermal runaway early warning model, fixes the parameters of the lower-level networks, and uses a hierarchical fine-tuning strategy to fine-tune the parameters of the higher-level networks, adding an adaptive attention mechanism to the higher-level networks. The core objective of transfer learning is to utilize the pre-trained thermal runaway early warning model. The underlying feature extraction capability is used for fine-tuning on new vehicle model data. Specifically, this includes: S51, loaded with the large thermal runaway early warning model trained in S1. and fix its underlying network parameters. It retains the general features learned by the pre-trained model.
[0055] S52, for high-level network parameters After fine-tuning to adapt it to the characteristics of the new car model, the model's loss function becomes:
[0056] S53 introduces an adaptive attention mechanism into the high-level network to enhance the model's ability to capture key features. This adaptive attention mechanism calculates attention weights and performs feature weighting.
[0057] in, For attention weights, Calculated by the adaptive layer, This represents the input features.
[0058] S6. The cross-entropy loss function and Adam optimizer are selected to optimize the large thermal runaway early warning model. The Warmup strategy and cosine annealing strategy are combined with dynamic scheduling of the learning rate.
[0059] The cross-entropy loss function is calculated as follows:
[0060] To accelerate the training process and avoid local optima, the Adam optimizer is used. The update rule of the Adam optimizer is:
[0061] in, and These are the momentum and variance of the gradient, respectively. It's the learning rate. It is a small constant that avoids division by zero.
[0062] The learning rate is dynamically scheduled using a combination of Warmup and cosine annealing strategies. In the early stages of training, a warmup strategy is used to gradually increase the learning rate to the target learning rate, avoiding excessively large learning rates that could lead to unstable model convergence.
[0063] In the later stages of model training, a cosine annealing strategy is used to adjust the learning rate, causing it to gradually decrease and helping the model converge better.
[0064] in, To maximize the learning rate, To minimize the learning rate, For Warmup time, Total training time This is the current training time.
[0065] S7 uses a test set to evaluate the finely tuned thermal runaway early warning model. The evaluation includes calculating metrics such as the model's precision, recall, and F1 score.
[0066] S8 inputs real-time data from the new vehicle model into a finely tuned thermal runaway warning model to obtain the probability of thermal runaway risk from the power battery. The output layer uses a Sigmoid activation function to convert the thermal runaway risk probability output into a probability value ranging from 0 to 1. The converted model is as follows:
[0067] in, These are model parameters. It is a weight matrix. It is a bias term. It is the Sigmoid activation function. This indicates the probability of thermal runaway occurring.
[0068] S9, combining real-time data and historical warning records, dynamically adjusts the model and warning threshold using Bayesian posterior analysis. When the risk probability exceeds the set warning threshold, an warning is triggered. S9 also includes: S91, set an initial risk threshold and update it using Bayesian methods:
[0069] in, This is the initial risk threshold.
[0070] S92, every predetermined time window, such as one day or one week, the error between the model's predicted probability and the actual event within the window is statistically analyzed, and the threshold is adjusted according to the error rate; if the false alarm rate is high, the threshold is increased; if the false alarm rate is high, the threshold is decreased.
[0071] S93: When the predicted probability exceeds the dynamically updated warning threshold, an early warning is triggered.
[0072] It also includes a feedback mechanism that combines the feedback information from each warning with the current warning threshold. The feedback information, such as operator feedback and actual event records, updates the threshold and the model as follows:
[0073] in, Indicates the error of the current window. It is the target error level. To adjust the step size.
[0074] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A transfer learning method for a large-scale model of thermal runaway early warning for power batteries, characterized in that, include: S1. Use the real vehicle operation data of the power battery obtained from the cloud data platform to train a large model for thermal runaway early warning and obtain upstream pre-training parameters; S2 collects real-vehicle operating data of the power battery of the new model and preprocesses the data; S3 enhances the original real-vehicle operation data through a diffusion model; S4 combines the real-vehicle operating data of the new model's power battery with the enhanced data to obtain a new dataset; S5: Load the parameters of the pre-trained thermal runaway early warning model, fix the parameters of the bottom network, fine-tune the parameters of the high-level network using a hierarchical fine-tuning strategy, and add an adaptive attention mechanism to the high-level network. S6, cross-entropy loss function and Adam optimizer are selected to optimize the large thermal runaway early warning model, and Warmup strategy and cosine annealing strategy are combined with dynamic scheduling of learning rate; S7, use the test set to evaluate the fine-tuned thermal runaway early warning large model; S8 inputs real-time data from the new model into the finely tuned thermal runaway warning big model to obtain the probability of thermal runaway risk of the power battery; S9 combines real-time data with historical warning records and uses Bayesian posterior analysis to dynamically adjust the model and warning threshold. When the risk probability exceeds the set warning threshold, a warning is triggered.
2. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 1, characterized in that: The actual vehicle operating data includes voltage, current, temperature, SOC, SOH, charge / discharge status, mileage, speed, internal resistance, maximum cell voltage, and minimum cell voltage.
3. The transfer learning method for the large-scale model of thermal runaway early warning of power batteries according to claim 1, characterized in that, The preprocessing includes data cleaning and normalization. Data cleaning includes removing missing and outlier values, and the normalization formula is as follows: in, Represents the original data. This represents the mean of the data. The standard deviation of the data This represents the normalized data.
4. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 1, characterized in that, S3 includes: S31, Construct a diffusion process, using the preprocessed data as initial data, and apply a forward diffusion process to generate a series of noisy data; S32, the reverse training process, trains a neural network model to learn the mapping from noisy data to the original data, and the training objective is to maximize the variational lower bound. S33, generate new data, sample noise from the prior noise distribution, and calculate using the reverse process in sequence.
5. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 1, characterized in that: The adaptive attention mechanism calculates attention weights and performs feature weighting. in, For attention weights, Calculated by the adaptive layer, This represents the input features.
6. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 1, characterized in that, The method of using Warmup and cosine annealing strategies combined with dynamic scheduling of the learning rate includes: In the early stages of training, a warmup strategy is used to gradually increase the learning rate to the target learning rate. In the later stages of model training, a cosine annealing strategy is used to adjust the learning rate: in, To maximize the learning rate, To minimize the learning rate, For Warmup time, Total training time This is the current training time.
7. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 1, characterized in that: In S8, the output layer uses the Sigmoid activation function to convert the probability output of thermal runaway risk into a probability value ranging from 0 to 1. The transformed model is as follows: in, These are model parameters. It is a weight matrix. It is a bias term. It is the Sigmoid activation function. This indicates the probability of thermal runaway occurring.
8. The transfer learning method for the large-scale model of thermal runaway early warning of power batteries according to claim 1, characterized in that, S9 further includes: S91, set an initial risk threshold and update it using Bayesian methods: in, This is the initial risk threshold; S92, at predetermined time windows, the error between the model's predicted probability and the actual event within the window is statistically analyzed, and the threshold is adjusted based on the error rate; S93: When the predicted probability exceeds the dynamically updated warning threshold, an early warning is triggered.
9. The transfer learning method for the large-scale model of early warning of thermal runaway in power batteries according to claim 8, characterized in that: It also includes a feedback mechanism that combines the feedback information from each warning with the current warning threshold to update the threshold and the model, as follows: in, Indicates the error of the current window. It is the target error level. To adjust the step size.