Method and system for handling overall shift of lithium battery thermal runaway risk

Abstract

The application discloses a kind of lithium battery thermal runaway risk overall deviation processing method and system, method includes target domain risk label generation, multi-source migration model training, risk deviation detection, normalization deviation adjustment and deviation secondary correction and tertiary correction steps.By calculating thermal runaway risk label value to determine normal operating condition risk boundary, the model of feature engineering module, shared feature extractor and fully connected layer regressor is constructed, the optimal model is obtained by training the prediction error loss and distribution consistency loss fusion.This application effectively solves the problem of multi-source data migration adaptation and dynamic deviation correction, improves the accuracy and robustness of processing thermal runaway risk overall deviation.

Description

Technical Field

[0001] This invention relates to the field of battery safety monitoring, specifically to a method and system for handling the overall shift in the risk of thermal runaway in lithium batteries. Background Technology

[0002] With the widespread application of battery technology, battery thermal runaway has received increasing attention. Thermal runaway refers to a rapid increase in the internal temperature of a battery, leading to an uncontrollable exothermic reaction that can potentially cause fires or explosions. Therefore, accurate prediction of battery thermal runaway risks is of great importance.

[0003] However, existing technologies suffer from the following problems when predicting thermal runaway risk: First, when normalizing thermal runaway cases, the large differences in the maximum and minimum values ​​of temperature and voltage among each case lead to data distribution bias in the predicted thermal runaway risk values ​​obtained using models trained with normalized data. Specifically, when predicting the target domain, the inconsistency between the normalized range of the training set and the target domain causes deviations in the predicted risk values. Second, in practical applications, the highest temperature at the time of thermal runaway cannot be predicted in advance, making it impossible to determine whether there is a data range shift between the target and source domains, thus affecting the reliability of the model's predictions. Furthermore, existing technologies lack effective mechanisms to dynamically adjust normalization parameters to adapt to changes in data distribution under different operating conditions, and do not consider the extraction of common features and distribution consistency among data from multiple source domains, resulting in insufficient accuracy in cross-domain predictions. Summary of the Invention

[0004] To address the data distribution shift issue that occurs when normalizing thermal runaway cases and to ensure the accuracy of thermal runaway risk prediction, this invention provides a method for processing the overall shift in lithium battery thermal runaway risk.

[0005] The technical solution adopted by this invention to solve its technical problem is: a method for handling the overall shift in the risk of thermal runaway in lithium batteries, including:

[0006] The steps for generating risk labels for the target domain are as follows: collect and preprocess normal operating condition data of the target domain, calculate the thermal runaway risk label value, and determine the maximum value of the risk label as the risk boundary of normal operating conditions.

[0007] The training steps for the multi-source transfer model are as follows: construct the multi-source transfer model, input multi-source domain data into the model for feature engineering, common feature extraction and risk value prediction, and use the fusion loss function to train and obtain the optimal model;

[0008] The risk offset detection step involves using a trained model to predict the real-time risk value of the target domain, calculating the degree of deviation between the predicted risk value and the maximum value of the risk label, and determining whether it exceeds a first preset threshold.

[0009] In the normalization offset adjustment step, if the deviation exceeds the first preset threshold, the maximum temperature value in the normalization process is adjusted according to the offset, the data is re-normalized and the model is trained again.

[0010] In the second offset correction step, if the degree of deviation does not exceed the first preset threshold, calculate the offset after the first correction and determine whether it exceeds the second preset threshold.

[0011] The process involves three offset correction steps. If the offset does not exceed the second preset threshold, the predicted risk value is compared with the risk boundary under normal operating conditions. If the offset exceeds a preset multiple, the predicted risk value is reduced proportionally, and the final result is output. This solves the problems of multi-source data migration and adaptation, as well as dynamic offset correction.

[0012] Furthermore, in the target domain risk label generation step, determining the maximum value of the risk label includes:

[0013] Select the highest risk value from historical data as the initial value;

[0014] If the risk value of multiple consecutive time steps is less than the current initial value, update the risk value of the current step to the new initial value;

[0015] The maximum value of the risk label is obtained by averaging all initial values. This solves the problem of determining the dynamic risk boundary.

[0016] Furthermore, the multi-source transfer model training steps include:

[0017] The feature engineering module is constructed to normalize temperature data to [0,1], voltage data to [-1,1], and time data to [0,1].

[0018] Common features from multiple source domains are extracted using a shared feature extractor;

[0019] The common characteristics are input into the fully connected layer regressor, which outputs predicted values ​​for thermal runaway risk, temperature, and voltage.

[0020] A linear combination of prediction error loss and distribution consistency loss is used as the total loss, and the model is trained using the Adam optimizer. This solves the problems of multi-source data standardization and joint training.

[0021] Furthermore, the construction of the multi-source migration model includes:

[0022] The shared feature extractor adopts a VGG16 or ResNet50 network architecture;

[0023] The fully connected layer regressor contains at least two fully connected layers;

[0024] The feature engineering module includes a data normalization layer and a feature construction layer. This, in turn, solves the problem of model architecture concretization.

[0025] Furthermore, the extracted common features include:

[0026] Input the multi-source domain feature tensor output by the feature engineering module into the Siamese network;

[0027] By generating common features across multiple source domains through comparative learning, the problem of cross-domain common feature extraction is thus solved.

[0028] Furthermore, the adoption of the fusion loss function includes:

[0029] The prediction error is defined as the Euclidean distance or mean square error between the true value and the predicted value.

[0030] Distribution consistency is defined as the KL divergence between the predicted distribution and the true distribution;

[0031] The two loss functions are linearly combined and then trained using the Adam optimizer. This solves the problem of designing the loss function for the transfer model.

[0032] Furthermore, the degree of deviation in the calculation includes:

[0033] The difference between the predicted risk value and the maximum risk label value is calculated.

[0034] The degree of deviation is quantified using either absolute or relative values. This solves the problem of quantifying risk offset.

[0035] Furthermore, the normalization offset adjustment step includes:

[0036] If the deviation is greater than the third preset threshold, the maximum temperature value will be shifted to the right by the first predetermined step.

[0037] If the deviation is less than a negative third preset threshold, the maximum temperature value is shifted to the left by a first predetermined step. This solves the problem of dynamic adjustment of the normalized parameters.

[0038] Furthermore, the method also includes:

[0039] Based on the offset prediction and pre-adjustment steps of digital twins, when the offset exceeds the second preset threshold:

[0040] A digital twin model is constructed based on the battery's state of health (SOH), ambient temperature, and the number of charge-discharge cycles.

[0041] Predict the risk distribution trend for a future scheduled time period;

[0042] The maximum temperature value in the normalized parameters is adjusted based on the trend. This solves the problem of pre-adaptation to long-term risk trends.

[0043] Furthermore, the predicted risk distribution trend includes:

[0044] Set the time steps for SOH to decrease at a predetermined rate;

[0045] Accumulate the risk values ​​at each time step to generate a monthly risk trend;

[0046] The maximum temperature value is shifted to the left by a predetermined percentage based on the trend magnitude. This resolves the issue of adjusting digital twin parameters.

[0047] Accordingly, embodiments of the present invention also provide a system for processing the overall shift in the risk of thermal runaway in lithium batteries, employing the method described above, including:

[0048] The data preprocessing unit is used to collect and preprocess normal operating condition data of the target domain, and calculate thermal runaway risk label values ​​to determine the risk boundary of normal operating conditions.

[0049] A multi-source transfer model training unit is used to construct and train a multi-source transfer model. The model includes a feature engineering module, a shared feature extractor, and a fully connected layer regressor. The optimal model is obtained by training a loss function that integrates prediction error loss and distribution consistency loss.

[0050] The risk prediction and detection unit is used to predict the risk value of real-time data in the target domain using a trained model, and to calculate the degree of deviation between the predicted risk value and the risk boundary.

[0051] The offset adjustment unit is used to adjust the normalization parameters and retrain the model when the deviation exceeds the first preset threshold, and to perform secondary and tertiary offset corrections when the deviation does not exceed the threshold.

[0052] Furthermore, the multi-source transfer model training unit includes: a feature engineering submodule, used to normalize the input multi-source domain data, mapping temperature data to [0,1], voltage data to [-1,1], and time data to [0,1];

[0053] The shared feature extraction submodule adopts the ResNet50 or VGG16 network architecture and extracts common features from multiple source domains through Siamese network contrastive learning.

[0054] The regression prediction submodule contains at least two fully connected layers and uses the ReLU activation function to perform nonlinear transformation on common features to output thermal runaway risk and temperature prediction values.

[0055] The loss optimization submodule defines the prediction error loss as the mean square error between the true and predicted values, and the distribution consistency loss as the KL divergence between the predicted and true distributions. The two losses are linearly combined and then used to train the model through the Adam optimizer.

[0056] This invention proposes a method for detecting data distribution shifts. This method is simple to implement and can accurately detect the degree of shift in thermal runaway risk. By calculating the deviation between the predicted risk value and the maximum value of the risk label, the reliability of the model's prediction results can be quantitatively evaluated, providing a basis for subsequent shift correction.

[0057] A solution to the risk bias problem is proposed, namely, a bias adjustment method under normalization processing. This method adjusts the maximum value by modifying the boundaries of normal operating conditions, thereby effectively correcting the risk bias. This approach can dynamically adapt to the data distribution characteristics of different thermal runaway cases, improving the model's prediction accuracy in the target domain.

[0058] A method for secondary and tertiary offset correction after model training is proposed. This multi-level offset detection and correction mechanism ensures the accuracy of thermal runaway risk prediction. If offset still exists after the initial correction, the system will further adjust the scaling factor or perform predictive adjustments based on digital twins to ensure the reliability of the risk prediction results.

[0059] This invention effectively solves the problem of data distribution differences among different thermal runaway cases through a multi-source transfer learning framework, realizes cross-domain knowledge transfer, and improves the model's generalization ability and prediction accuracy for novel thermal runaway cases. Attached Figure Description

[0060] Figure 1 is a schematic diagram of the overall process of an embodiment of the present invention.

[0061] Figure 2 is a temperature change diagram of a thermal runaway case according to an embodiment of the present invention.

[0062] Figure 3 is a temperature change diagram of the target domain according to an embodiment of the present invention.

[0063] Figure 4 is a flowchart of the offset secondary correction process according to an embodiment of the present invention.

[0064] Figure 5 is a flowchart of multi-source transfer learning according to an embodiment of the present invention. Detailed Implementation

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] In this invention, the use of terms such as "first" and "second" is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0067] Example 1

[0068] Please refer to Figures 1 to 5. The method for processing the overall offset of lithium battery thermal runaway risk according to the present invention includes the following steps: target domain risk label generation step, multi-source migration model training step, risk offset detection step, normalization offset adjustment step, offset secondary correction step, and offset tertiary correction step.

[0069] Step 1: Target Domain Risk Label Generation Steps

[0070] First, data on battery temperature, voltage, and current under normal operating conditions in the target domain are collected. The collected data undergoes preprocessing, including outlier removal, missing value imputation, and smoothing. Then, based on the preprocessed data, a thermal runaway risk label value is calculated. Specifically, when calculating the maximum risk label value, the highest risk value from historical data is selected as the initial value. If the risk value for 30 consecutive basic time units (e.g., seconds) is lower than the current initial value, the risk value for the current step is updated to the new initial value. Finally, the average of all initial values ​​is taken to obtain the maximum risk label value, which serves as the risk boundary under normal operating conditions.

[0071] Furthermore, by preprocessing the normal operating condition data and dynamically updating the initial risk value and taking the average, the risk boundary under normal operating conditions can be accurately characterized, effectively eliminating interference factors such as outliers and short-term fluctuations, and providing a reliable reference threshold for the accurate identification of subsequent thermal runaway risks.

[0072] Step 2: Training steps for the multi-source transfer model

[0073] A multi-source migration model is constructed, comprising a feature engineering module, a shared feature extractor, and a fully connected layer regressor. The feature engineering module normalizes temperature data to the range [0,1], voltage data to the range [-1,1], and time data to the range [0,1]. The shared feature extractor uses a VGG16 network architecture to extract common features from multiple source domains. The fully connected layer regressor contains three fully connected layers to output predicted values ​​for thermal runaway risk, temperature, and voltage.

[0074] When extracting common features, the multi-source domain feature tensor output by the feature engineering module is input into the Siamese network, and multi-source domain common features are generated through comparative learning. A fusion loss function is used to train the model, which is a linear combination of prediction error loss and distribution consistency loss. In a preferred embodiment, to prioritize the accuracy of predicted values, a higher weight (e.g., 0.9) is given to the prediction error loss (e.g., Euclidean distance), while a relatively lower weight (e.g., 0.1) is given to the distribution consistency loss (KL divergence). The combined loss is trained using the Adam optimizer with a learning rate of 0.001 for 100 training epochs to obtain the optimal model.

[0075] Furthermore, by normalizing multi-source data through feature engineering and extracting common features through comparative learning with the Siamese network, the distribution differences of multi-source data are effectively reduced, improving the model's adaptability to different source domains. The fusion loss function promotes the synergistic optimization of prediction error and distribution consistency, reducing the Euclidean distance between predicted and true values, while also narrowing the KL divergence between the predicted and true distributions, thus enhancing the accuracy and reliability of predictions. The optimal model trained by the Adam optimizer can fully learn the patterns of multi-source data, achieving accurate predictions of thermal runaway risk, temperature, and voltage.

[0076] Step 3: Risk Offset Detection Steps

[0077] The trained model is used to predict real-time risk values ​​in the target domain, and the deviation between the predicted risk value and the maximum value of the risk label is calculated. To calculate the deviation, the difference between the predicted risk value and the maximum value of the risk label is taken, and the absolute value difference is used to quantify the degree of deviation. It is then determined whether the deviation exceeds a first preset threshold of 0.15.

[0078] Furthermore, it can monitor in real time the significant deviation between the predicted risk value and the maximum value of the risk label in the target domain.

[0079] Step 4: Normalization Offset Adjustment Steps

[0080] If the deviation exceeds the first preset threshold of 0.15, the maximum temperature value in the normalization process is adjusted according to the deviation. Specifically, if the deviation is greater than the third preset threshold of 0.2, the maximum temperature value is shifted to the right by a first predetermined step of 5°C; if the deviation is less than the negative third preset threshold of -0.2, the maximum temperature value is shifted to the left by a first predetermined step of 5°C. After adjustment, the data is re-normalized and the model is trained again until the deviation is less than the first preset threshold. The settings of parameters such as the first preset threshold, the third preset threshold, and the adjustment step are determined based on the historical data distribution and model stability requirements, aiming to ensure that when the deviation is lower than the first threshold, the data distribution deviation does not affect the effective learning of the model.

[0081] Furthermore, by dynamically adjusting the maximum temperature value in the normalization process and re-normalizing the training, the deviation is controlled within the first preset threshold, thereby improving the stability and accuracy of model training.

[0082] Step 5: Secondary Offset Correction Step

[0083] If the deviation does not exceed the first preset threshold of 0.15, calculate the offset after the initial correction and determine whether it exceeds the second preset threshold of 0.1. If the offset exceeds the second preset threshold, perform the offset prediction and pre-adjustment steps based on digital twins: construct a digital twin model based on the battery health state (SOH), ambient temperature, and charge / discharge cycle count; predict the risk distribution trend over a predetermined period; and pre-adjust the maximum temperature value in the normalized parameters according to the trend.

[0084] When predicting the risk distribution trend, the time steps for SOH to decrease at a predetermined rate of 0.5% / month are set, the risk values ​​of each time step are accumulated to generate a monthly risk trend, and the maximum temperature value is shifted to the left by a predetermined percentage of 3% based on the trend magnitude.

[0085] Furthermore, by predicting the risk distribution trend and pre-adjusting the maximum temperature, potential deviations can be accurately corrected in advance, effectively controlling the degree of deviation within a preset threshold.

[0086] Step Six: Three Offset Correction Steps

[0087] If the offset does not exceed the second preset threshold of 0.1, the predicted risk value is compared with the risk boundary under normal operating conditions. If it exceeds the preset multiple of 1.2, the predicted risk value is reduced proportionally and the final result is output. The reduction ratio is 80% of the excess portion.

[0088] This method improves prediction accuracy by processing the overall shift of thermal runaway risk and establishing a dynamic risk boundary determination mechanism, while the three-level shift correction strategy effectively reduces prediction bias.

[0089] Furthermore, by proportionally reducing the predicted risk value that exceeds the risk boundary of normal operating conditions, the prediction deviation is effectively reduced, and the accuracy of risk prediction is improved.

[0090] Example 2

[0091] This embodiment provides a processing system for the overall offset of lithium battery thermal runaway risk. The system adopts a processing method based on the overall offset of lithium battery thermal runaway risk, including a data preprocessing unit, a multi-source migration model training unit, a risk prediction and detection unit, and an offset adjustment unit.

[0092] The data preprocessing unit collects and preprocesses normal operating condition data from the target domain, calculates thermal runaway risk label values, and determines the maximum risk label value as the risk boundary under normal operating conditions. When determining the maximum risk label value, this unit first selects the maximum risk value from historical data as the initial value; when the risk value at multiple consecutive time steps is less than the current initial value, the risk value at the current step is updated to the new initial value; finally, the average of all initial values ​​is taken to obtain the maximum risk label value. This method can dynamically determine the risk boundary.

[0093] The multi-source transfer model training unit is used to build and train a multi-source transfer model. This model includes a feature engineering module, a shared feature extractor, and a fully connected layer regressor. The optimal model is obtained through training using a loss function that fuses prediction error loss and distribution consistency loss. This unit contains four sub-modules:

[0094] The feature engineering submodule normalizes the input multi-source domain data, mapping temperature data to the [0,1] interval, voltage data to the [-1,1] interval, and time data to the [0,1] interval, thus achieving standardized processing of multi-source data.

[0095] The shared feature extraction submodule employs a ResNet50 or VGG16 network architecture and extracts common features from multiple source domains through a Siamese network contrastive learning approach. In a preferred embodiment, the multi-source domain feature tensor output by the feature engineering module is input into the Siamese network, and common features from multiple source domains are generated through contrastive learning, effectively solving the problem of cross-domain common feature extraction.

[0096] The regression prediction submodule contains at least two fully connected layers and uses the ReLU activation function to perform nonlinear transformation on common features, ultimately outputting the predicted values ​​of thermal runaway risk and temperature.

[0097] The loss optimization submodule defines prediction error loss as the mean squared error between the true and predicted values, and distribution consistency loss as the KL divergence between the predicted and true distributions. These two losses are linearly combined and then used to train the model through the Adam optimizer. During optimization, prediction error can be defined as commonly used regression metrics such as the Euclidean distance between the true and predicted values, mean squared error, or mean absolute error (MAE); distribution consistency is defined as the KL divergence between the predicted and true distributions. These two losses are linearly combined and then used to train the model through the Adam optimizer. These metrics are used to comprehensively evaluate the model's accuracy in predicting target values ​​such as the risk of thermal runaway.

[0098] The risk prediction and detection unit uses a trained model to predict risk values ​​from real-time data in the target domain and calculates the degree of deviation between the predicted risk value and the risk boundary. When calculating the deviation, the difference between the predicted risk value and the maximum value of the risk label is taken, and the degree of deviation is quantified using either the absolute difference or the relative value, thus achieving precise quantification of risk shift.

[0099] The offset adjustment unit is responsible for handling the deviation between the predicted risk value and the risk boundary. When the deviation exceeds a first preset threshold, the system adjusts the normalization parameters and retrains the model; when the deviation does not exceed the first preset threshold, the system performs secondary and tertiary offset corrections.

[0100] During the normalization offset adjustment process, if the deviation is greater than the third preset threshold, the system will shift the maximum temperature value to the right by a first predetermined step; if the deviation is less than the negative third preset threshold, the system will shift the maximum temperature value to the left by a first predetermined step, thereby realizing the dynamic adjustment of the normalization parameters.

[0101] In the second offset correction step, the system calculates the offset after the initial correction and determines whether it exceeds a second preset threshold. If the offset does not exceed the second preset threshold, the system proceeds to the third offset correction step.

[0102] In the offset correction steps, the system compares the predicted risk value with the risk boundary of normal operating conditions. If it exceeds a preset multiple, the predicted risk value is reduced proportionally and the final result is output.

[0103] Through the coordinated operation of the aforementioned units and steps, this system effectively solves the problems of multi-source data migration and adaptation as well as dynamic offset correction, providing reliable technical support for thermal runaway risk prediction. The system's advantage lies in its ability to dynamically adjust model parameters to adapt to risk prediction needs under different operating conditions, while ensuring the accuracy and reliability of prediction results through a multi-level correction mechanism. (Glossary of technical terms follows.)

[0104] Source domain: refers to the set of historical thermal runaway data containing different battery types (such as lithium iron phosphate and ternary lithium batteries), different experimental platforms (such as cell abuse certification platforms and laboratory testing platforms), and different operating conditions (such as overcharging, heating, and thermal propagation). It is used to train the multi-source migration model and provide common knowledge support for risk prediction in the target domain.

[0105] Target domain: refers to the real-time operating data set of the energy storage project to be evaluated (including large-scale energy storage, residential or commercial energy storage), which includes parameters such as battery temperature, voltage and current under normal operating conditions. The data format is the same as the source domain, but the operating conditions are different. It is the object of thermal runaway risk prediction.

[0106] Multi-source transfer learning refers to a machine learning framework that standardizes multi-source domain data through feature engineering modules, extracts common features from each source domain using shared feature extractors (such as CNN+Siamese networks), and trains the model using a fusion loss function (prediction error loss + distribution consistency loss) to adapt the common knowledge of the multi-source domains to the target domain, thus solving the prediction bias problem caused by the data distribution offset between the source and target domains.

[0107] Digital twin model: refers to a virtual mapping model built based on the actual operating parameters of lithium battery (such as state of health, ambient temperature, and number of charge-discharge cycles). It can simulate the battery's operating state, predict the distribution trend of thermal runaway risk over a predetermined period of time (such as 1 month), and provide a basis for pre-adjusting normalized parameters (such as maximum temperature) to achieve pre-adaptation to long-term risk trends.

[0108] Risk boundary: refers to the maximum value of the thermal runaway risk label under normal operating conditions in the target domain (which can be denoted as a). It is determined by dynamically updating the initial value (historical maximum risk value) and taking the average value after preprocessing the normal operating data and calculating the thermal runaway risk label value. It serves as the benchmark threshold for judging whether the predicted risk value has deviated and whether it needs to be corrected.

[0109] Furthermore, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. (This description is only necessary for software and communication methods.)

[0110] Furthermore, a portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. The program instructions invoking the methods of the invention may be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal-carrying medium, and / or stored in the working memory of a computer device operating according to the program instructions. Here, an embodiment of the invention includes an apparatus comprising a memory for storing computer program instructions and a processor for executing the program instructions, wherein, when the computer program instructions are executed by the processor, the apparatus is triggered to operate the methods and / or technical solutions based on the foregoing embodiments of the invention.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for handling the overall shift in the risk of thermal runaway in lithium batteries, characterized in that, include: The steps for generating risk labels for the target domain are as follows: collect and preprocess normal operating condition data of the target domain, calculate the thermal runaway risk label value, and determine the maximum value of the risk label as the risk boundary of normal operating conditions. The training steps for the multi-source transfer model are as follows: construct the multi-source transfer model, input multi-source domain data into the model for feature engineering, common feature extraction and risk value prediction, and use the fusion loss function to train and obtain the optimal model; The risk offset detection step involves using a trained model to predict the real-time risk value of the target domain, calculating the degree of deviation between the predicted risk value and the maximum value of the risk label, and determining whether it exceeds a first preset threshold. In the normalization offset adjustment step, if the deviation exceeds the first preset threshold, the maximum temperature value in the normalization process is adjusted according to the offset, the data is re-normalized and the model is trained again. In the second offset correction step, if the degree of deviation does not exceed the first preset threshold, calculate the offset after the first correction and determine whether it exceeds the second preset threshold. If the offset exceeds the second preset threshold, perform the offset prediction and pre-adjustment steps based on digital twins: construct a digital twin model based on battery health status (SOH), ambient temperature, and charge / discharge cycle count; predict the risk distribution trend over a predetermined time period; and pre-adjust the maximum temperature value in the normalized parameters according to the trend. The offset is corrected in three steps. If the offset does not exceed the second preset threshold, the predicted risk value is compared with the risk boundary of normal working conditions. If it exceeds the preset multiple, the predicted risk value is reduced proportionally and the final result is output.

2. The method according to claim 1, characterized in that, In the target domain risk label generation step, determining the maximum value of the risk label includes: Select the highest risk value from historical data as the initial value; If the risk value of multiple consecutive time steps is less than the current initial value, update the risk value of the current step to the new initial value; The maximum value of the risk label is obtained by averaging all the initial values.

3. The method according to claim 1, characterized in that, The training steps for the multi-source transfer model include: The feature engineering module is constructed to normalize temperature data to [0,1], voltage data to [-1,1], and time data to [0,1]. Common features from multiple source domains are extracted using a shared feature extractor; The common characteristics are input into the fully connected layer regressor, which outputs predicted values ​​for thermal runaway risk, temperature, and voltage. The model is trained using a linear combination of prediction error loss and distribution consistency loss as the total loss, and the loss is calculated using the Adam optimizer.

4. The method according to claim 3, characterized in that, The construction of the multi-source migration model includes: The shared feature extractor adopts a VGG16 or ResNet50 network architecture; The fully connected layer regressor contains at least two fully connected layers; The feature engineering module includes a data normalization layer and a feature construction layer.

5. The method according to claim 3, characterized in that, The extracted common features include: Input the multi-source domain feature tensor output by the feature engineering module into the Siamese network; Common features from multiple source domains are generated through comparative learning.

6. The method according to claim 3, characterized in that, The fusion loss function adopted includes: The prediction error is defined as the Euclidean distance or mean square error between the true value and the predicted value. Distribution consistency is defined as the KL divergence between the predicted distribution and the true distribution; The two losses are linearly combined and then trained using the Adam optimizer.

7. The method according to claim 1, characterized in that, The degree of deviation in the calculation includes: The difference between the predicted risk value and the maximum risk label value is calculated. The degree of deviation is quantified by the difference in absolute or relative values.

8. The method according to claim 1, characterized in that, The normalization offset adjustment step includes: If the deviation is greater than the third preset threshold, the maximum temperature value will be shifted to the right by the first predetermined step. If the deviation is less than the negative third preset threshold, the maximum temperature value is shifted to the left by the first predetermined step.

9. A system for handling the overall shift in thermal runaway risk in lithium batteries, characterized in that, The method according to any one of claims 1 to 8 comprises: The data preprocessing unit is used to collect and preprocess normal operating condition data of the target domain, and calculate thermal runaway risk label values ​​to determine the risk boundary of normal operating conditions. A multi-source transfer model training unit is used to construct and train a multi-source transfer model. The model includes a feature engineering module, a shared feature extractor, and a fully connected layer regressor. The optimal model is obtained by training a loss function that integrates prediction error loss and distribution consistency loss. The risk prediction and detection unit is used to predict the risk value of real-time data in the target domain using a trained model, and to calculate the degree of deviation between the predicted risk value and the risk boundary. The offset adjustment unit is used to adjust the normalization parameters and retrain the model when the deviation exceeds the first preset threshold, and to perform secondary and tertiary offset corrections when the deviation does not exceed the threshold.

10. The system according to claim 9, characterized in that, The multi-source transfer model training unit includes: a feature engineering submodule, used to normalize the input multi-source domain data, mapping temperature data to [0,1], voltage data to [-1,1], and time data to [0,1]; The shared feature extraction submodule adopts the ResNet50 or VGG16 network architecture and extracts common features from multiple source domains through Siamese network contrastive learning. The regression prediction submodule contains at least two fully connected layers and uses the ReLU activation function to perform nonlinear transformation on common features to output thermal runaway risk and temperature prediction values. The loss optimization submodule defines the prediction error loss as the mean square error between the true and predicted values, and the distribution consistency loss as the KL divergence between the predicted and true distributions. The two losses are linearly combined and then used to train the model through the Adam optimizer.

Images