A new energy vehicle battery thermal runaway early warning method and system based on transfer learning

Abstract

The application discloses a new energy automobile battery thermal runaway early warning method and system based on transfer learning, belongs to the technical field of new energy automobile battery safety monitoring, and relates to the technical field of new energy automobile battery safety monitoring. The method comprises the following steps: collecting battery temperature, gas, smoke and electrochemical parameters through a plurality of source sensors and constructing a characteristic vector; pre-training a basic early warning model by using a laboratory and public data; migrating the model to a target vehicle through a domain self-adaptive technology to solve a small sample problem; learning a dynamic benchmark based on vehicle normal data to realize abnormal detection; executing graded early warning according to a risk level and cooperating with a teaching system; and finally, optimizing the model through a continuous learning mechanism. Corresponding systems comprise a data acquisition layer, an edge computing layer, a cloud platform layer, a transfer learning engine, a teaching adaptation module and a user interaction layer. The application realizes early and accurate early warning of battery thermal runaway, has a low false alarm rate, provides complete teaching adaptation functions, and is particularly suitable for practical teaching environments.

Description

Technical Field

[0001] This invention belongs to the field of battery safety monitoring technology, and in particular relates to a method and system for early warning of thermal runaway of new energy vehicle batteries based on transfer learning. Background Technology

[0002] With the rapid development of new energy vehicles, the safety of power batteries has become a key concern in the industry. Thermal runaway is the most serious safety accident involving power batteries, which can lead to serious consequences such as vehicle fires and explosions. Existing battery thermal runaway warning technologies have the following shortcomings: (1) Traditional methods are mostly based on fixed thresholds for a single temperature parameter, which usually only trigger alarms when thermal runaway has already occurred or during the thermal diffusion stage, thus missing the best time for intervention (early stage of thermal runaway). (2) The thermal runaway characteristics of batteries of different types (such as lithium iron phosphate and ternary lithium), different aging degrees, and different usage environments are significantly different, but the existing early warning models are mostly general designs and lack adaptability to specific batteries. (3) Obtaining sufficient thermal runaway sample data in actual vehicles is extremely difficult and dangerous, making it difficult to fully train machine learning-based early warning models; (4) Most existing commercial early warning systems are closed designs and do not provide the functions required for teaching, such as fault simulation and data access, making it difficult to integrate them into the new energy vehicle training system. Current technologies still rely on simple threshold judgments. In recent years, some studies have begun to try applying machine learning methods, but they still face the challenge of model training under small sample conditions. Especially in practical training environments, it is difficult to balance the accuracy of early warnings with the teaching needs of demonstrations and fault diagnosis. Summary of the Invention

[0003] To address the aforementioned technical issues, this invention proposes a method and system for early warning of thermal runaway in new energy vehicle batteries based on transfer learning. By using transfer learning to solve the model training problem under small sample conditions, and combining multi-source data fusion and adaptive threshold mechanism, it achieves early and accurate warning of battery thermal runaway, while also providing comprehensive practical training and teaching adaptation functions.

[0004] The technical solution adopted in this invention is as follows: A method for early warning of thermal runaway in new energy vehicle batteries based on transfer learning, comprising the following steps: Step S1: Multi-source data acquisition and feature construction, collecting temperature, gas, smoke and electrochemical parameters of the battery pack, and constructing a multi-dimensional feature vector; Step S2: Source domain model pre-training, using laboratory simulation data and large-scale public battery datasets to train a basic early warning model; Step S3: Target domain adaptive transfer, adapting the pre-trained model to the specific conditions of the target vehicle through domain adaptation techniques; Step S4: Dynamic benchmark learning and anomaly detection, learning a dynamic benchmark pattern based on the target vehicle's normal operating condition data; Step S5: Tiered early warning and teaching collaboration, triggering tiered responses based on risk levels and providing teaching analysis data; Step S6: Continuous learning and model updates, continuously optimizing the early warning model based on newly accumulated data.

[0005] Further, in step S1, the multi-source data acquisition includes: Temperature monitoring: A distributed temperature sensor network is used to deploy multiple temperature measurement points at key locations such as the battery pack surface, cell gaps, and connection terminals to construct temperature field distribution characteristics; Gas monitoring: Equipped with a metal oxide semiconductor gas sensor to monitor early characteristic gases of thermal runaway, such as hydrogen and carbon monoxide; Smoke monitoring: A photoelectric smoke sensor is used to detect particulate matter concentration; Electrochemical parameters: The voltage, current, internal resistance, and SOC status parameters of the battery management system are obtained through the CAN bus interface; Environmental parameters: Collect auxiliary information such as ambient temperature and humidity.

[0006] Further, in step S2, the source domain model pre-training specifically includes: Multi-source data construction: Integrating laboratory thermal runaway simulation data, publicly available battery test data, and simulation-generated data to form a source domain dataset; Feature engineering processing: preprocessing the raw sensor data such as filtering and noise reduction, feature extraction, and time alignment; Basic model training: Deep neural networks are used to learn early feature patterns of thermal runaway. The network structure includes convolutional layers to extract spatial features, LSTM layers to extract temporal features, and attention mechanism layers for feature selection. Knowledge distillation and compression: Transferring knowledge from complex models to lightweight models to adapt to embedded device deployments.

[0007] Further, in step S3, the adaptive migration of the target domain specifically includes: Domain difference measure: Calculates the degree of difference between the data distribution of the source domain and the target domain, using the maximum mean difference or Wasserstein distance as the metric; Feature space alignment: Through adversarial training or correlation alignment methods, the feature distributions of the source domain and the target domain are aligned at the feature extraction layer; Small sample fine-tuning: Fine-tuning the aligned model using a small amount of labeled data of the target vehicle. A conservative learning rate strategy is used for fine-tuning to prevent overfitting. Uncertainty estimation: Provides confidence assessment for model prediction results, and triggers a manual review mechanism when the confidence level is low.

[0008] Furthermore, in step S4, the dynamic benchmark learning and anomaly detection specifically include: Normal pattern learning: Based on the normal operating condition data of the target vehicle over a long period of time, an autoencoder or Gaussian mixture model is used to learn the normal data distribution; Adaptive threshold setting: Dynamically adjusts the abnormal thresholds of each monitoring parameter based on battery health status, environmental conditions, and usage history; Multi-indicator fusion judgment: Combine multiple indicators such as the rate of temperature rise, the trend of gas concentration change, and voltage consistency for joint judgment; False alarm suppression mechanism: The false alarm rate is reduced by using time window filtering, multi-sensor consistency verification and other methods.

[0009] Furthermore, in step S5, the hierarchical early warning and teaching collaboration specifically includes: The three-level early warning system consists of: Level 1 warning (potential anomaly, mobile APP notification), Level 2 warning (confirmed anomaly, audible and visual alarm + SMS notification), and Level 3 warning (emergency danger, automatic power cut-off + fire extinguishing activation). Teaching data recording: Completely record sensor data, model reasoning process, response measures, etc. before and after the warning is triggered, forming a teaching case library; Fault simulation function: Provides a software configuration interface to simulate various training scenarios such as sensor failure, communication interruption, and battery abnormality; API Open Interface: Provides standardized data and control interfaces to support students in secondary development and algorithm verification.

[0010] Furthermore, in step S6, the continuous learning and model update specifically include: Incremental learning mechanism: After new data has accumulated to a certain scale, the model parameters are updated using incremental learning to avoid catastrophic forgetting; Model performance monitoring: Continuously monitor key indicators such as early warning accuracy, false alarm rate, and response time; A / B testing verification: Before deploying the new model, A / B testing was conducted in shadow mode to verify the performance improvement effect; Version management and traceability: Establish a model version management system to support model rollback and effect traceability.

[0011] This invention also provides a new energy vehicle battery thermal runaway early warning system based on transfer learning, comprising: Data acquisition layer: Composed of a distributed sensor network and a CAN bus interface, responsible for raw data acquisition; Edge computing layer: Deployed in in-vehicle embedded devices to complete data preprocessing, lightweight model inference, and local alarms; Cloud platform layer: Provides cloud services such as data storage, model training, rule engine, and alarm push; Transfer learning engine: responsible for source domain model training, target domain adaptation, and continuous model optimization; Teaching adaptation module: Provides teaching functions such as fault simulation, data access, and experiment configuration; User interaction layer: includes mobile APP, web management backend, and training platform display interface.

[0012] The beneficial effects of the present invention after adopting the above structure are as follows: (1) Solved the problem of model training under small sample conditions: By using transfer learning technology, we can make full use of the knowledge of laboratory simulation data and public datasets. Only a small amount of target vehicle data is needed to build a high-precision early warning model. Under the condition of only 50 sets of target abnormal samples, the model accuracy reached 96.2%, which is 38% higher than training from scratch, breaking through the technical bottleneck of scarce thermal runaway samples.

[0013] (2) Early and accurate warning is achieved: Through multi-source data fusion and dynamic benchmark learning, the system can identify subtle early characteristics of thermal runaway. Tests show that the average warning time is advanced to 8 minutes and 23 seconds before thermal runaway occurs, providing a sufficient time window for emergency response. Multi-indicator fusion judgment keeps the false alarm rate below 1.9%.

[0014] (3) It provides complete teaching adaptation functions: The system is specially designed with teaching functions such as fault simulation, data openness, and experimental configuration, which supports the multi-level needs of new energy vehicle training courses. The open API interface allows students to carry out secondary development, which cultivates students' system integration and algorithm optimization capabilities.

[0015] (4) It has good practicality and scalability: The system adopts a modular design, and the hardware cost is controlled within RMB 1,000, which is suitable for large-scale deployment in the training room. Through the continuous learning mechanism, the model can be continuously optimized as data accumulates, and the system architecture supports expansion to more application scenarios such as battery health management and charging safety monitoring. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0017] Figure 1This is a flowchart of the core early warning method of the new energy vehicle battery thermal runaway early warning method and system based on transfer learning proposed in this invention. Figure 2 This is a flowchart illustrating the transfer learning implementation of the new energy vehicle battery thermal runaway early warning method and system based on transfer learning proposed in this invention. Figure 3 This invention provides the teaching and system deployment process for the new energy vehicle battery thermal runaway early warning method and system based on transfer learning. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Example 1, see Figures 1-3 The method for early warning of thermal runaway in new energy vehicle batteries based on transfer learning provided by this invention, when implemented in a training environment, includes the following steps: Step S1: Multi-source data acquisition and feature construction. Twelve temperature monitoring points are set up inside the battery pack of the training vehicle, including six surface temperature points and six cell gap points, using DS18B20 digital temperature sensors; two MQ-9 gas sensors are configured to monitor hydrogen and carbon monoxide; one photoelectric smoke sensor is installed; battery parameters are obtained by connecting to the CAN bus through the OBD-II interface, with the sampling frequency set to 10Hz. After the data is filtered by moving average, a 120-dimensional feature vector containing time-series features is constructed.

[0020] Step S2: Source domain model pre-training. The source domain dataset consists of three parts: laboratory simulation data (obtained by heating to trigger battery thermal runaway, containing complete thermal runaway process data), public datasets (NASA battery dataset, CALCE battery dataset), and simulation-generated data (generated based on electrochemical-thermal coupling model simulation). Pre-training adopts a hybrid network structure of ResNet-18 combined with BiLSTM. The input is a 120-dimensional feature vector, and the output is four states: normal, potential risk, early thermal runaway, and thermal runaway occurred. The training set, validation set, and test set are divided in a 7:2:1 ratio.

[0021] Step S3: Target domain adaptive transfer. For specific training vehicles, collect 2 weeks of normal operation data (100,000 records) and 50 simulated abnormal scenarios. Use domain adversarial neural network for transfer learning. Add gradient reversal layer and domain classifier after feature extractor. The loss function is designed as: L=Ltask+γLdomain. Where Ltask is the state classification loss, Ldomain is the domain classification loss, and γ is a tradeoff parameter (initial value 0.1, increasing by 0.02 per training round, up to a maximum of 1.0). After transfer learning, the model achieves an accuracy of 96.2% on the target vehicle, which is 38% higher than direct training.

[0022] Step S4: Dynamic benchmark learning and anomaly detection. A variational autoencoder is used to learn the data distribution under normal operating conditions. The encoder compresses 120-dimensional features into a 32-dimensional latent space, and the decoder reconstructs the input. The anomaly score is defined as the weighted sum of the reconstruction error and the deviation from the latent space: AnomalyScore=α∥xx∥2+β∥z-μ∥2. Where x is the input feature, x is the reconstructed feature, z is the latent variable, μ is the latent space center of normal data, α=0.7 and β=0.3 are weight coefficients, and an early warning analysis is triggered when the abnormal score exceeds the dynamic threshold (95th percentile of normal data).

[0023] Step S5: Tiered early warning and teaching collaboration, establishing a three-tiered response mechanism guided by teaching principles: Level 1 Warning (Score 0.7-0.85): The training platform display screen flashes a yellow warning, a notification is pushed to the mobile APP, and it is recorded as a teaching case without interrupting the training; Level 2 Warning (Score 0.85-0.95): Triggers audible and visual alarms, sends a text message to the instructor, and recommends suspending operation; Level 3 warning (score > 0.95): Automatically disconnects the main battery relay (requires coordination with BMS), activates emergency ventilation, and sends an emergency notification.

[0024] All early warning events generate detailed reports, including trigger time, sensor readings, model reasoning process, response measures, etc., which are stored in the teaching case library for course use.

[0025] Step S6: Continuous learning and model update; the system is set to automatically learn the model, triggering a fine-tuning of the model every 5,000 new data points. Elastic weight consolidation is employed to prevent catastrophic forgetting. The importance weights are calculated using the Fisher information matrix: Fi = 1 / N * (∂logp(yn∣xn, θ)∂θi)²; where Fi represents the importance weight of the i-th parameter θi in the neural network, with a larger value indicating greater importance to the learned task; N represents the number of training samples in the old task (source domain); yn represents the true label of the n-th sample; xn represents the input features of the n-th sample; θ represents the set of all parameters of the model, and (∂logp(yn∣xn, θ)∂θi) reflects the sensitivity of the parameter to the prediction result; the loss function during update is: Lnew = Lcurrent + γewciFi(θi - θi∗)²; where Lcurrent is the loss function for the new task; θi∗ is the optimal value of the parameter on the old task; and γewc is the penalty coefficient for elastic weight consolidation, with a value of 0.1.

[0026] Example 2: Hardware configuration of the system of the present invention: Data acquisition layer hardware: Main controller: ESP32-S3 development board, dual-core 240MHz, built-in WiFi / Bluetooth, supports neural network acceleration; Temperature sensors: 12 DS18B20 sensors connected via 1-Wire bus, measuring range -55℃~125℃, accuracy ±0.5℃; Gas sensors: 2 MQ-9, detection range 10~1000ppmH2, 100~10000ppmCO; Smoke sensor: 1 GP2Y1014AU0F, detecting 0.1~0.9 mg / m³ 3 Particulate matter; CAN bus interface: MCP2515+TJA1050 module, supporting SAE J1939 protocol; Local display: 7-inch IPS touchscreen, 1024×600 resolution; Alarm device: 85dB buzzer + RGB LED warning light.

[0027] Installation and layout plan: Temperature sensors 1-6 are arranged on the surface of the battery pack (3 on the top surface, 2 on the side, and 1 on the bottom), and sensors 7-12 are inserted into the gaps between the battery cells through special probes; gas sensors are installed near the pressure relief valve of the battery pack; smoke sensors are located at the top of the battery compartment; all wiring harnesses use automotive-grade high-temperature resistant wires, and signal lines have additional shielding layers.

[0028] Example 3: Specific parameter configuration for transfer learning: Source domain training data scale: 800 sets of laboratory data (each set containing a complete thermal runaway process), 1200 sets of publicly available data, and 2000 sets of simulation data; the basic model training cycle is 50 rounds, batch size is 32, initial learning rate is 0.001, and decays by 0.5 times every 10 rounds. In the knowledge distillation process, the teacher model is an 8-layer ResNet, the student model is a 4-layer MobileNet, the distillation temperature T=3, and the distillation loss weight is 0.7.

[0029] When adapting to the target domain, a progressive transfer strategy is adopted: the first stage freezes the feature extraction layer and trains only the classification layer (learning rate 0.01); the second stage unfreezes the last two convolutional layers (learning rate 0.001); the third stage fine-tunes the entire network (learning rate 0.0001). Each stage is trained for 10 rounds, and the next stage begins when the accuracy on the validation set no longer improves.

[0030] Example 4: Details of the implementation of the teaching adaptation function: The fault simulation module provides 6 types of configurable faults: Sensor malfunction: Simulates abnormal readings of one or more sensors (offset, increased noise, complete failure); Communication failure: Simulated CAN bus communication interruption, WiFi connection disconnection; Battery abnormalities: simulated voltage imbalance, sudden increase in internal resistance, and abnormal temperature distribution; Model failure: Intentionally using a low-precision model or incorrect parameters; Environmental interference: Simulating the effects of high temperature and high humidity environments on sensors; Combined failure: Multiple failures occur simultaneously.

[0031] The teaching data interface provides: Real-time data stream: WebSocket interface pushes real-time sensor data; Historical data query: The RESTful API supports querying by time and event type; Model transparency: Provides feature importance analysis and visualization of the decision-making process; Experimental setup: Allows students to modify warning thresholds and adjust model parameters.

[0032] Security protection mechanism: In the practical training mode, the automatic power-off function of the three-level early warning needs to be manually confirmed before it is executed; All configuration changes are logged, and rollback to previous versions is supported. Provides an emergency stop button, allowing you to restore all default settings with a single click.

[0033] Example 5: Cloud Platform Implementation Architecture It adopts a microservice architecture and includes the following service modules: Data access service: MQTT agent receives vehicle data, supporting up to 100,000 concurrent devices; Time-series database: InfluxDB stores historical sensor data and retains it for 6 months; Model Services: TensorFlowServing provides a model inference API; Rule engine: Implements complex alarm rules based on Drools; Push notifications: Integrates multiple notification methods including WeChat template messages, SMS, and voice calls; Teaching Management: Provides functions for class management, student experiments, and grade evaluation.

[0034] Example 6: Cost and Implementation Results In a training lab environment, the hardware cost of a single system is approximately 850 yuan (main controller 180 yuan + sensor 350 yuan + display screen 200 yuan + other 120 yuan). The cloud platform can be deployed at zero cost using an open-source solution, or a commercial IoT platform can be used with an annual fee of approximately 500 yuan.

[0035] Actual test results: Early warning timeliness: The average advance warning time is 8 minutes and 23 seconds (compared to about 2-3 minutes using traditional methods); Accuracy: In a 6-month test, the correct alert rate was 98.7%, and the false alarm rate was 1.9%. Pedagogical value: The system supports 12 experimental projects across 3 courses (Battery Technology, Internet of Things Applications, and Fault Diagnosis); Scalability: Through modular design, it can be expanded to applications such as battery health status assessment and charging safety monitoring.

Claims

1. A method for early warning of thermal runaway in new energy vehicle batteries based on transfer learning, characterized in that, Includes the following steps: Step S1: Multi-source data acquisition and feature construction; collect temperature, gas, smoke and electrochemical parameters of the battery pack, and construct a multi-dimensional feature vector; Step S2: Source domain model pre-training; train the basic early warning model using laboratory simulation data and large-scale public battery datasets; Step S3: Adaptive migration of the target domain; Domain adaptation techniques are used to fit pre-trained models to specific conditions of the target vehicle. Step S4: Dynamic benchmark learning and anomaly detection; Learn dynamic benchmark patterns based on normal operating condition data of the target vehicle; Step S5: Tiered early warning and teaching collaboration; triggering tiered responses based on risk levels and providing teaching analysis data; Step S6: Continuous learning and model updates; continuously optimize the early warning model based on newly accumulated data.

2. The method according to claim 1, characterized in that, The multi-source data acquisition in step S1 specifically includes: Temperature monitoring employs a distributed temperature sensor network, with multiple temperature measurement points arranged on the battery pack surface, cell gaps, and connection terminals. Gas monitoring: Equipped with a metal oxide semiconductor gas sensor to monitor hydrogen and carbon monoxide concentrations; Smoke monitoring uses photoelectric smoke sensors to detect particulate matter concentration; Electrochemical parameters are acquired by obtaining the voltage, current, internal resistance, and SOC status parameters of the battery management system through the CAN bus interface.

3. The method according to claim 2, characterized in that, The source domain model pre-training in step S2 specifically includes: The source domain dataset is constructed by integrating laboratory thermal runaway simulation data, publicly available battery test data, and simulation-generated data. Preprocessing of raw sensor data includes filtering and noise reduction, feature extraction, and time alignment. A deep neural network containing convolutional layers, LSTM layers, and attention mechanism layers is used for basic model training; Knowledge distillation is used to migrate complex models to lightweight models that are suitable for embedded device deployments.

4. The method according to claim 3, characterized in that, The adaptive migration of the target domain in step S3 specifically includes: Calculate the domain difference measure between the data distribution of the source domain and the target domain; Feature space alignment can be achieved through adversarial training or relevance alignment methods. The aligned model was fine-tuned using a small number of labeled data from the target vehicle; Provide confidence assessment for model prediction results, and trigger a manual review mechanism when the confidence level is low.

5. The method according to claim 4, characterized in that, The dynamic benchmark learning and anomaly detection in step S4 specifically include: Based on the normal operating condition data of the target vehicle over a long period of time, the normal data distribution is learned by using an autoencoder or Gaussian mixture model. The abnormal thresholds of each monitoring parameter are dynamically adjusted based on the battery health status, environmental conditions, and usage history. The assessment is based on a combination of multiple indicators, including the rate of temperature rise, the trend of gas concentration change, and voltage consistency. Time window filtering and multi-sensor consistency verification methods are used to reduce the false alarm rate.

6. The method according to claim 5, characterized in that, The graded early warning and teaching collaboration in step S5 specifically includes: Establish a three-tiered early warning system, including Level 1, Level 2, and Level 3 early warnings, each triggering different levels of response measures; Completely record sensor data, model reasoning process, and response measures before and after the early warning is triggered, forming a teaching case library; It provides configurable fault simulation functions to simulate various training scenarios such as sensor failure, communication interruption, and battery abnormality; It provides standardized data interfaces and open control APIs to support secondary development and algorithm verification.

7. The method according to claim 6, characterized in that, The continuous learning and model update in step S6 specifically include: Once new data has accumulated to a set scale, the model parameters are updated using incremental learning. Continuously monitor key indicators such as early warning accuracy, false alarm rate, and response time; Before deploying the new model, A / B testing was conducted in shadow mode to verify the performance improvement effect. Establish a model version management system to support model rollback and effect tracking.

8. A new energy vehicle battery thermal runaway early warning system based on transfer learning, used to implement the method according to any one of claims 1 to 7, characterized in that, include: The data acquisition layer, consisting of a distributed sensor network and a CAN bus interface, is used for raw data acquisition. Edge computing layer, deployed in in-vehicle embedded devices, is used for data preprocessing, lightweight model inference, and local alarms; The cloud platform layer provides cloud services for data storage, model training, rule engine, and alarm push. The transfer learning engine is responsible for training the source domain model, adapting it to the target domain, and continuously optimizing the model. The teaching adaptation module provides teaching functions such as fault simulation, data access, and experimental configuration. The user interaction layer includes a mobile app, a web management backend, and a training platform display interface.

Images