Safety early warning method for power battery based on transfer learning and thermodynamic coupling

Abstract

This invention relates to the field of battery management technology and discloses a power battery safety early warning method based on the coupling of transfer learning and thermodynamics. It collects multimodal data to obtain time-series data and divides it into source domain data and target domain data. A domain adaptive transfer network is constructed based on a domain adversarial neural network framework. A hybrid heterogeneous feature extraction mechanism and a phased adversarial training mechanism for source and target domain data are introduced to complete domain adaptive transfer learning and output a domain-independent safety feature representation. Based on the domain-independent safety feature representation, parameters are dynamically corrected and applied to the energy conservation equation to derive and calculate the core temperature of the new battery cell and the critical parameters for thermal runaway, thus determining the critical thermal runaway of the new battery. A bidirectional correction strategy is introduced to couple decision updates to the domain-independent safety feature representation and the core temperature of the cell, generating dynamic early warning indicators for dynamic early warning. This invention achieves rapid deployment of new batteries without massive labeling and ensures high-precision adaptive early warning throughout their entire life cycle.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, specifically to a power battery safety early warning method based on transfer learning and thermodynamic coupling. Background Technology

[0002] As the core energy carrier of electric vehicles and energy storage systems, the safe operation of power batteries is directly related to the safety of people and property and the large-scale application and promotion. In recent years, frequent fires and explosions caused by thermal runaway have made efficient, accurate, and adaptive safety early warning technologies a key bottleneck that battery management systems urgently need to overcome.

[0003] To address this need, existing technologies have developed various early warning methods, but they still suffer from the following significant drawbacks: First, limited multimodal data fusion capabilities: Current solutions typically involve simple channel splicing or feature stacking of sensor data such as voltage, current, and temperature, failing to optimize for the inherent correlations between heterogeneous data sources at the feature level. This results in low recognition accuracy of the fused early warning features for different thermal runaway states, making it difficult to capture early signs of anomalies. Second, poor cross-cell type adaptability: Different models and batches of batteries exhibit significant differences in electrochemical characteristics. Existing models often require separate collection and retraining of labeled data for each type of cell, lacking cross-model transfer capabilities. First, the system faces a serious "cold start" problem when new batteries are deployed, resulting in high deployment costs and long cycles. Second, the model-driven approach is singular: pure data-driven black-box models lack physical mechanism constraints, and their generalization ability decreases significantly when operating conditions change. Pure thermodynamic mechanism models rely on precise parameter calibration and cannot adapt to dynamic operating conditions, which can easily lead to delayed warnings or frequent false alarms. Third, the warning thresholds are statically fixed: existing systems use fixed warning thresholds and do not fully consider dynamic factors throughout the entire life cycle, such as cell aging, cycle count, and health status. This causes the accuracy of warnings to fluctuate significantly as the battery usage time increases, making it difficult to match the actual safety risk status of the battery. Summary of the Invention

[0004] This invention aims to provide a power battery safety early warning method based on the coupling of transfer learning and thermodynamics. Based on an optimized algorithm architecture of multimodal data perception, feature transfer learning, thermodynamic model coupling, and hierarchical early warning execution, it completely solves the problem of new battery deployment relying on a large amount of labeled data, enabling rapid deployment and ensuring continuous and high-precision adaptive early warning capabilities throughout the entire life cycle.

[0005] The basic solution provided by this invention is: a power battery safety early warning method based on the coupling of transfer learning and thermodynamics, comprising: S1. Collect and preprocess multimodal data characterizing the safety status of the power battery to obtain time-series data; divide the time-series data into source domain data characterized by historical battery life cycle data and target domain data characterized by new battery data. S2 constructs a domain adaptive transfer network based on a domain adversarial neural network framework, introduces a hybrid heterogeneous feature extraction mechanism and a phased adversarial training mechanism for source domain data and target domain data, completes domain adaptive transfer learning of time-series data, and finally outputs a domain-independent secure feature representation. S3. Construct a thermodynamic model, dynamically correct the preset parameters based on the domain-independent safety feature representation, the real-time surface temperature of the new battery cell and the fixed physical parameters of the new battery cell, and apply the corrected preset parameters to the energy conservation equation to derive and calculate the core temperature of the new battery cell and the critical parameters for thermal runaway, and make a critical determination of thermal runaway of the new battery. S4 introduces a bidirectional correction strategy for data features and physical models to perform coupled decision-making, and generates dynamic early warning indicators based on the updated domain-independent safety feature representation and the core temperature of the new battery cell; dynamic early warning is then performed using the dynamic early warning indicators.

[0006] The working principle and advantages of this invention are as follows: This invention, based on an optimized algorithm architecture of multimodal data perception, feature transfer learning, thermodynamic model coupling, and hierarchical early warning execution, solves key technical problems such as difficulty in adapting to different cell types, limitations of a single driving mode, and accuracy fluctuations throughout the entire lifecycle caused by fixed thresholds. It achieves rapid deployment of new batteries without massive annotation and ensures high-precision adaptive early warning throughout their entire lifecycle. Its core advantages are: First, a hybrid heterogeneous feature extraction and attention fusion mechanism is introduced to improve the early warning identification of thermal runaway. Existing technologies often use simple channel splicing for multi-source data such as voltage, current, temperature, impedance spectrum, and characteristic gas concentration, failing to fully explore multi-scale anomaly patterns in battery time-series data. This invention designs a hybrid heterogeneous feature extraction network, in which a ResNet18 convolutional neural network branch is used to extract local fluctuation features of battery data to capture short-term transient anomalies, and a Transformer branch is used to extract global evolution features to characterize long-term aging trends. An attention mechanism is used to adaptively weight and fuse the two types of features, enabling the model to automatically focus on key signs strongly correlated with thermal runaway, significantly improving the sensitivity and accuracy of early warning.

[0007] Second, a phased domain-adaptive transfer learning mechanism is constructed to achieve domain-independent safety feature generation across battery models. Existing technologies require separate collection and retraining of labeled data for different battery models, resulting in a significant "cold start" problem; standard domain adversarial neural networks employ end-to-end joint training, which can cause the feature extractor to prematurely sacrifice discriminative power. This invention utilizes only source domain full-lifecycle battery data and its labels in the first stage to train a hybrid heterogeneous feature extraction network and classifier, enabling the model to fully learn the safety status discrimination features during battery aging. In the second stage, new battery data from the target domain (without labels) is introduced. Through a gradient inversion layer, the feature extraction network and the domain classifier engage in an adversarial game, mapping the features to a domain-independent space while maintaining discriminative power, ultimately outputting a reusable domain-independent safety feature representation. This mechanism allows for rapid adaptation to new battery models without retraining, significantly reducing deployment costs and time.

[0008] Third, a unique design achieves deep bidirectional coupling between data features and the physical model, overcoming the limitations of a single-drive mode. Pure data-driven models lack physical constraints, while pure mechanistic models cannot adapt to dynamic operating conditions. This invention combines domain-independent safety feature representation with measured cell surface temperature and fixed physical parameters. By applying dynamic parameter correction to the energy conservation equation, it achieves adaptive derivation of core temperature and thermal runaway critical parameters. Furthermore, it establishes a bidirectional correction logic that adjusts the weights of data features constraining the physical model and corrects physical model parameters based on data features, thus ensuring both data adaptability and mechanistic reliability.

[0009] Fourth, it enables dynamic early warning throughout the entire battery lifecycle based on cell status. Existing systems use fixed early warning thresholds and do not consider dynamic factors such as cell aging, cycle count, and health status. This invention, based on the domain-independent safety feature representation updated by bidirectional correction and the core temperature of the cell, dynamically calculates the early warning index in real time by coupling the early warning index formula, ensuring that the accuracy of the early warning remains stable throughout the entire lifecycle, and achieving early, accurate, and adaptive early warning of battery thermal runaway risks. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the power battery safety early warning method based on the coupling of transfer learning and thermodynamics provided in an embodiment of the present invention. Detailed Implementation

[0011] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A power battery safety early warning method based on the coupling of transfer learning and thermodynamics includes: S1. Collect and preprocess multimodal data characterizing the safety state of the power battery to obtain time-series data; divide the time-series data into source domain data characterized by historical battery lifecycle data and target domain data characterized by new battery data; as detailed below: In S1, multimodal data acquisition and preprocessing are performed to eliminate sensor noise, dimensional differences and data disorder, and transform multi-source raw data into regular time-series data that can be directly input into the model, laying the foundation for subsequent feature extraction.

[0012] Raw multimodal data are collected using equipment such as integrated temperature and pressure sensors, impedance testers, and gas sensors.

[0013] The raw multimodal data include temperature (sampling range -40℃ to 230℃, sampling frequency 100Hz), pressure (sampling range 0 to 2MPa, sampling frequency 100Hz), impedance spectrum (sampling range 1kHz to 1MHz, sampling frequency 1Hz), and characteristic gas concentrations (CO / HF / H2S, detection limit ≤1ppm, sampling frequency 10Hz).

[0014] Preprocessing includes: 1) Outlier Removal (3σ Criterion): Filter out extreme outliers caused by sensor malfunctions and electromagnetic interference. Calculate the mean for any data dimension x. and standard deviation Remove excess The values ​​are calculated, and outliers are replaced with the average of the five normal data points before and after them to avoid data gaps; the output is a multimodal data sequence without outliers.

[0015] 2) Data Standardization (Z-score): Eliminates dimensional differences and ensures fair weighting of data across modalities; for each data point x, the standardized value is... ( The mean of this dimension. (where is the standard deviation); after standardization, all data have a mean of 0 and a standard deviation of 1; output a dimensionless standardized multimodal data sequence.

[0016] 3) Time series data segmentation (sliding window): The continuous data stream is segmented into samples of fixed length while preserving short-term abrupt changes and temporal continuity; the window size is set to 5s (to cover short-term features such as sudden temperature rises), and the step size is 1s (to avoid missing key information); one window corresponds to one sample, containing 500 temperature and pressure data points (100Hz×5s), 5 impedance data points (1Hz×5s), and 50 gas data points (10Hz×5s); the output dimension is a regularized time series data matrix (denoted as X) with the dimensions of "number of samples × 3 (temperature / pressure / gas) × 500 (maximum time step)", where impedance and gas data are aligned to a time step of 500 through zero padding.

[0017] The output regularized time series data matrix X is divided into two categories: Historical battery lifecycle data was selected as the source domain data: 3000-cycle full lifecycle data of ternary lithium batteries (100,000 samples, including normal / aging / precursor thermal runaway). Select new battery data as the target domain data: 1000 sets of normal operating condition data for the new battery pack.

[0018] S2 constructs a domain-adaptive transfer network based on a domain adversarial neural network framework. It introduces a hybrid heterogeneous feature extraction mechanism and a phased adversarial training mechanism for source and target domain data to complete domain-adaptive transfer learning of time-series data, ultimately outputting a domain-independent secure feature representation (i.e., a feature vector). Details are as follows: Based on the feature extraction of the domain adversarial neural network architecture, gradient inversion layer (GRL) and domain classifier, through adversarial training and domain adaptive learning, the model can reuse mature battery cell knowledge and adapt with only a small amount of new battery cell data, outputting general safety features. By ignoring differences in battery cell type and focusing on safety features, the model solves the "cold start" problem of different battery models.

[0019] S21, a hybrid heterogeneous feature extraction network is used to extract at least the local fluctuation features and global evolution features from the time series data and fuse them to obtain an initial feature representation.

[0020] Specifically, a hybrid network structure of ResNet18 and Transformer can be used to process data with different characteristics. ResNet18 extracts local fluctuation features, while Transformer extracts global evolution features. Then, key features are fused through attention to obtain the initial feature representation.

[0021] ResNet18 extracts local features representing short-term abrupt changes: Temperature and pressure data (500 time steps) from the time-series data matrix X are input into ResNet18. The convolutional layers of ResNet18 capture short-term features such as sudden temperature increases and pressure fluctuations, outputting local feature representations. (256 dimensions).

[0022] The Transformer extracts long-term features representing long-term trends: Impedance and gas data (55 time steps) from the time-series data matrix X are input into the Transformer. Through the Transformer's multi-layer encoder (6 layers) and multi-head attention (8 heads), long-term dependencies such as slow impedance spectrum shifts and gas concentration accumulation are captured; the output is a long-term feature representation. (256 dimensions).

[0023] Based on the principle of assigning higher weight to key features of thermal runaway precursors (such as gas concentration), attention-weighted fusion is performed:

[0024] in, For local feature fusion weights, For long-term feature fusion weights, Greater than , usually take It is 0.6. It is 0.5.

[0025] The initial feature representation F (64-dimensional) of the fusion output is obtained by mapping the weighted fusion of 256-dimensional features to a 64-dimensional low-dimensional space through a projection layer, which reduces computational complexity while retaining key discriminative information.

[0026] S22 performs forward propagation through a gradient inversion layer (GRL) without altering the features; then it performs backward propagation to invert the gradient of the domain classifier, forcing feature extraction to generate domain-independent features.

[0027] S23 introduces a loss function that balances two objectives: classification of safe states and elimination of domain differences.

[0028] in, Total loss; Apply risk level classification loss (cross-entropy loss) to ensure the model can identify normal / warning states; λ is the domain discrimination loss, which measures the model's ability to distinguish between different types of battery cells; λ is the balance coefficient, which avoids excessive adversarial behavior that could lead to classification failure (value 0.6). use ,in Let c be the true label of the c-th sample. Let be the predicted probability of the c-th sample, and n be the total number of samples. c is the number of modes in the multimodal data, which must include at least three types of samples corresponding to temperature, pressure, and gas concentration, and can be expanded according to the actual sensor configuration.

[0029] S24, Phased training of the transfer network: The first stage is source domain pre-training: using source domain data (100,000 sets) to train the hybrid feature extraction network and domain classifier to ensure convergence (validation set accuracy ≥ 98%), and the source domain model is obtained after training.

[0030] Specifically, source domain data with category labels is input into a hybrid feature extraction network for local fluctuation hybrid feature extraction; the extracted fused feature vector is input into a task classifier to calculate the source domain classification loss; the adversarial strength coefficient λ=0 of the gradient inversion layer is set so that the domain discriminator branch does not participate in parameter updates; by minimizing the source domain classification loss, the parameters of the hybrid feature extraction network and the task classifier are updated to obtain the source domain model.

[0031] The second stage is target domain adaptation: 1000 sets of target domain data are introduced for adversarial training. After the initial feature representation, a gradient inversion layer and a domain classifier are connected. The gradient inversion layer reverses the gradient, causing the hybrid feature extraction network and the domain classifier to form an adversarial game. The source domain model is fine-tuned so that the domain classification accuracy is ≤55% (close to randomness, domain differences are eliminated).

[0032] Specifically, unlabeled target domain data is input into the source domain model, where some network parameters in the pre-trained source domain model are selectively frozen; source domain samples and target domain samples are respectively input into a hybrid feature extraction network to extract and fuse features to obtain a fused feature vector; the fused feature vector is simultaneously input into a task classifier and a gradient reversal layer, where the gradient reversal layer is set with an adversarial strength coefficient λ>0; the features output by the gradient reversal layer are input into a domain discriminator to calculate the domain discrimination loss; the total loss is calculated, where source domain data is calculated with both source domain classification loss and domain discrimination loss, while target domain data is calculated with only domain discrimination loss; the unfrozen network parameters are updated through backpropagation, so that the fused feature vector gradually learns domain-independent class discrimination features.

[0033] Selective freezing should employ at least the following strategies: Strategy A: Freeze all parameters of the local fluctuation feature extraction branch, and only update the parameters of the global evolution feature extraction branch and the feature fusion structure; Strategy B: Freeze all parameters of the global evolution feature extraction branch, and only update the parameters of the local fluctuation feature extraction branch and the feature fusion structure; Strategy C: Freeze the low-level parameters of the local fluctuation feature extraction branch and the global evolution feature extraction branch, and only update the high-level parameters and the parameters of the feature fusion structure; Strategy D: Based on the contribution of each feature extraction branch to the classification task in the first stage of source domain pre-training, dynamically select branches with a contribution value higher than the threshold for freezing.

[0034] Output: Domain-independent security feature representation (64 dimensions).

[0035] S3. Construct a thermodynamic model. Based on the domain-independent safety feature representation, the real-time surface temperature of the new battery cell, and the fixed physical parameters of the new battery cell, dynamically correct the preset parameters. Apply the corrected preset parameters to the energy conservation equation to derive and calculate the core temperature and critical parameters for thermal runaway of the new battery cell, and determine the critical thermal runaway of the new battery. Details are as follows: In S3, the thermodynamic model calculates the core temperature and critical state of thermal runaway of the battery cell based on physical mechanisms, providing "physical constraints" for subsequent coupled decisions and avoiding "black box misjudgments" of pure data models.

[0036] S31, Domain-independent security feature representation obtained from input S2 (Includes information related to cell aging, used to correct for internal resistance), the actual cell surface temperature measured by the sensor. (Data after S1 preprocessing), and fixed physical parameters of the battery cells (battery pack factory information data: For cell density, Where is the specific heat capacity of the battery cell, V is the volume of the battery cell, A is the heat dissipation area, and h is the heat transfer coefficient.

[0037] S32, because the internal resistance of the battery cell changes with the degree of aging, feature correction is required (to avoid calculation deviations caused by fixed parameters); a domain-independent security feature representation is then performed. The dynamic parameter correction (internal resistance R update) process includes: Will The fully connected layers (2 layers) of the input network are mapped to an internal resistance correction coefficient k (range 0.8~1.5, the coefficient increases with more severe aging); dynamic internal resistance , The initial internal resistance of the new battery cell is given; the final output is the dynamically updated internal resistance of the battery cell.

[0038] S33 calculates the core temperature of the battery cell based on the energy conservation equation. It derives the core temperature, which cannot be directly measured, through energy conservation, reflecting the true heating state of the battery cell. The energy conservation equation is as follows:

[0039] in, I represents the core temperature of the battery cell, and I represents the measured current. The internal resistance of the battery cell is a preset parameter. The heat generated by the decomposition of the SEI membrane The value is the surface temperature of the battery cell; the left side of the equation is the rate of heat accumulation inside the battery cell, which determines the core temperature change; the three terms on the right side of the equation are Joule heat plus heat generated by the decomposition of the SEI film (QSEI) minus heat loss.

[0040] Input I at each preset time step (e.g., 1 second). , The equations were solved using the Euler method for numerical integration, ultimately outputting the core temperature of the battery cell. and thermal runaway critical parameters (rate of temperature rise) ).

[0041] S34, thermal runaway criticality determination, determines whether the cell has entered the initial stage of thermal runaway based on the thermal runaway critical parameter (temperature rise rate); Judgment criteria: When the temperature rise rate exceeds the set threshold (e.g., 1℃ / s), it is judged as "thermal runaway initiation critical state"; the final output is the physical critical judgment result (whether it is close to thermal runaway), core temperature, and critical temperature (200℃ for ternary lithium, 250℃ for lithium iron phosphate).

[0042] S4 introduces a bidirectional correction strategy combining data features and the physical model for coupled decision-making, and generates dynamic early warning indicators based on the updated domain-independent safety feature representation and the core temperature of the new battery cell; dynamic early warning is then implemented using these indicators. Details are as follows: In S4, a coupled decision-making process is carried out based on a two-way correction strategy of data characteristics and physical models. The physical parameters are corrected by the dynamics of data, and the data prediction is constrained by physical laws. Finally, dynamic early warning indicators that are adapted to the entire life cycle are generated.

[0043] S41, Input: Domain-independent security feature representation, dynamic internal resistance, core cell temperature, critical temperature.

[0044] S42, bidirectional correction (data → physics, physics → data) includes: The physical model constrains the data features. When the core temperature of the battery cell approaches the critical temperature, the weights of key temperature-related features in the domain-independent safety feature representation are adjusted (e.g., the weights of features related to temperature and heat are increased by 1.2 times to strengthen key precursor features); the weighted domain-independent safety feature representation is output.

[0045] The physical model is corrected using data features. The weighted feature representation is then used to re-correct dynamic parameters to update the energy conservation equation (thermodynamic physical model) and recalculate the updated core temperature of the new battery cell. This correction ensures that the physical calculations accurately reflect the real-time state of the battery cell.

[0046] The above bidirectional correction forms an iterative closed loop. The iteration termination condition is as follows: if the temperature difference of the new battery cell core obtained from two adjacent iterations is less than the preset temperature threshold (e.g., 0.5°C), convergence is determined (the temperature difference of the new battery cell core obtained from the last iteration is used as the final output), or if the number of iterations reaches the preset maximum number of iterations (e.g., 5~10 times), it is forcibly terminated. If convergence is not achieved after reaching the maximum number of iterations, the average of the core temperatures of the battery cell obtained from the two most recent iterations is taken as the final output, and the confidence level of the warning indicator at that time step is marked as reduced.

[0047] The following considerations are taken into account when determining iterative convergence: The preset temperature threshold is determined based on the accuracy of the temperature sensor and the accuracy of solving the energy conservation equation. It should be greater than the sensor noise level to avoid false convergence, while being less than the accuracy required for the temperature percentage calculation in S43. A typical value is 0.3°C to 1.0°C. The preset maximum number of iterations is determined based on the time budget of the BMS control cycle and the time consumed per iteration. It also takes into account the statistical results of the number of iterations required for more than 95% of the samples in offline simulation, plus a safety margin. A typical value is 5 to 10 iterations.

[0048] S43, Calculation of Coupled Early Warning Indicators (Core Formula), Comprehensive Data Trends ( Rate of change) and physical state ( (Relative critical value percentage), generating dynamic early warning indicator S:

[0049] in, This represents the rate of change of the updated domain-independent security feature representation (at the data level, reflecting risk trends, such as...). (A rapid increase in gas characteristics increases the value); represents the proportion of the updated core temperature of the battery cell relative to the critical value (physical level, reflecting the degree of risk); is the data feature weight, and β is the physical parameter weight; α=0.6 and β=0.4 can be selected.

[0050] When the S42 iteration fails to converge, the dynamic early warning index S calculated by S43 is marked with a reduced confidence level. This is used to increase the safety margin when judging the early warning level in the subsequent S44. That is, the corresponding early warning threshold is appropriately reduced (e.g., the first-level threshold is reduced from 0.3 to 0.25, and the second-level threshold is reduced from 0.8 to 0.7) to provide a more conservative safety warning when the calculation uncertainty increases.

[0051] S44, based on S, at least one level is defined: Normal: S < Level 1 threshold (0.3); Level 1 warning: Level 1 threshold ≤ S < Level 2 threshold (0.8); Level 2 warning: S ≥ Level 2 threshold (0.8, precursor to thermal runaway); Furthermore, when bidirectional correction fails to converge, the input dynamic warning index S is accompanied by a reduced confidence level marker, and the thresholds at each level are reduced by a preset offset (e.g., 0.05~0.1) to provide a more conservative warning judgment when bidirectional correction fails to converge, ensuring safety as a priority. In other embodiments, the thresholds at each level are finely adjusted according to the aging degree of the battery cell.

[0052] The final output warning level result (normal / level 1 / level 2) is used to determine the subsequent safety enforcement strategy.

[0053] S45, tiered early warning execution based on security level: Level 1 warning (parameter control): When S is between 0.3 and 0.8 (0.3 ≤ S < 0.8), the BMS controls the liquid cooling system flow rate to increase by 50%, the charging current to drop to 0.3C, and pushes a "Battery risk warning, please avoid fast charging and high-speed driving" to the vehicle terminal.

[0054] Level 2 Warning (Emergency Response): When S≥0.8, immediately cut off the charging and discharging circuit, activate the battery pack directional pressure relief valve to release the fire extinguishing agent, and send an "emergency stop request" to the vehicle control system, and synchronize the warning information to the cloud management platform.

[0055] The power battery safety early warning method based on the coupling of transfer learning and thermodynamics provided in this embodiment solves key technical problems such as the difficulty of cold start adaptation across cell models, the large limitations of a single driving mode, and the fluctuation of accuracy throughout the entire life cycle caused by fixed thresholds. It is based on an optimized algorithm architecture of multimodal data perception, feature transfer learning, thermodynamic model coupling, and hierarchical early warning execution. It achieves the technical effect of rapid deployment of new battery models without the need for a large amount of labeled data and maintaining high-precision adaptive early warning throughout the entire life cycle.

[0056] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A power battery safety early warning method based on the coupling of transfer learning and thermodynamics, characterized in that, include: S1, Collect and preprocess multimodal data characterizing the safety status of the power battery to obtain time-series data; The time-series data is divided into source domain data, characterized by historical battery lifecycle data, and target domain data, characterized by new battery data. S2 constructs a domain adaptive transfer network based on a domain adversarial neural network framework, introduces a hybrid heterogeneous feature extraction mechanism and a phased adversarial training mechanism for source domain data and target domain data, completes domain adaptive transfer learning of time-series data, and finally outputs a domain-independent secure feature representation. S3. Construct a thermodynamic model, dynamically correct the preset parameters based on the domain-independent safety feature representation, the real-time surface temperature of the new battery cell and the fixed physical parameters of the new battery cell, and apply the corrected preset parameters to the energy conservation equation to derive and calculate the core temperature of the new battery cell and the critical parameters for thermal runaway, and make a critical determination of thermal runaway of the new battery. S4 introduces a bidirectional correction strategy for data features and physical models to perform coupled decision-making, and generates dynamic early warning indicators based on the updated domain-independent safety feature representation and the core temperature of the new battery cell; dynamic early warning is then performed using the dynamic early warning indicators.

2. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, In S2, a hybrid heterogeneous feature extraction network is used to extract at least the local fluctuation features and global evolution features from the time series data and fuse them to obtain an initial feature representation. In the first stage, the hybrid feature extraction network and the domain classifier are trained using source domain data to obtain the source domain model. In the second stage, target domain data is introduced, and after the initial feature representation, a gradient inversion layer and the domain classifier are connected. By inverting the gradient through the gradient inversion layer, the hybrid feature extraction network and the domain classifier form an adversarial game, which fine-tunes the source domain model and finally outputs a domain-independent secure feature representation.

3. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, In S2, the hybrid heterogeneous feature extraction adopts a hybrid network structure of ResNet18 and Transformer to process data with different characteristics. ResNet18 extracts local fluctuation features, while Transformer extracts global evolution features. Then, key features are fused through attention to obtain the initial feature representation.

4. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, In S2, a loss function is introduced that balances two objectives: classification of safe states and elimination of domain differences. in, Total loss; Losses are categorized by risk level to ensure the model can identify normal / warning states; The domain discrimination loss measures the model's ability to distinguish between different cell types; λ is the balance coefficient. use ,in Let c be the true label of the c-th sample. Let be the predicted probability of the c-th sample, and n be the total number of samples.

5. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, Preset parameters include cell internal resistance; The correction process involves representing the domain-independent security features as input to the fully connected layer of the adaptive migration network, mapping them to an internal resistance correction coefficient k, and then calculating the output dynamic cell internal resistance using the following formula. : in, This represents the initial internal resistance of the new battery cell.

6. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, The core temperature of the new battery cell is calculated based on the following energy conservation equation: in, For cell density, V is the specific heat capacity of the battery cell, and V is the volume of the battery cell. I represents the core temperature of the new battery cell, and I represents the measured current. The internal resistance of the battery cell is a preset parameter. The heat generated is due to the decomposition of the SEI film, where h is the heat transfer coefficient and A is the heat dissipation area. This refers to the surface temperature of the battery cell. Input I for each preset time step. , The equations were solved using the Euler method of numerical integration, and the core temperature of the new battery cell was finally output. and the rate of temperature rise as a critical parameter for thermal runaway .

7. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, Bidirectional correction includes: The physical model constrains the data features. When the core temperature of the new battery cell approaches the critical temperature, the weights of the temperature-related key characteristics in the domain-independent safety feature representation are adjusted, and the weighted updated domain-independent safety feature representation is output. The physical model is corrected by data features. The preset parameters are dynamically corrected again by using the weighted updated domain-independent security feature representation to update the energy conservation equation and recalculate the updated core temperature of the new battery cell.

8. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, The bidirectional correction forms an iterative closed loop. The convergence condition for the iteration is: the temperature difference between the core of the new battery cell obtained from two adjacent iterations is less than the preset temperature threshold, or the number of iterations reaches the preset maximum number of iterations. If convergence is not achieved after reaching the maximum number of iterations, the average of the core temperatures of the new battery cells obtained from the two most recent iterations will be taken as the final output.

9. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, The dynamic early warning index S is calculated using the following formula: in, This represents the updated domain-independent security feature representation. The rate of change reflects risk trends at the data level; Indicates the updated core temperature of the battery cell. Relative critical temperature The proportion reflects the degree of risk at the physical level; β represents the weight of the data features, and β represents the weight of the physical parameters.

10. The power battery safety early warning method based on transfer learning and thermodynamic coupling according to claim 1, characterized in that, Based on the dynamic early warning indicator S, at least the following levels should be defined: Normal: S < Level 1 threshold; Level 1 warning: Level 1 threshold ≤ S < Level 2 threshold; Level 2 warning: S ≥ Level 2 threshold; The final output level is used to determine the subsequent security enforcement strategy; When the bidirectional correction fails to converge, the input dynamic early warning index S is marked with a reduced confidence level, and the thresholds at each level are reduced by a preset offset from their original values.

Images