4.Machine learning warning model: dual model collaborative warning system
Build a dual model architecture that combines a fusion model and an LSTM temporal model, and accurately identify the thermal runaway stage of lithium iron phosphate batteries based on multimodal data, adapting to different application scenarios:
4.1 Fusion ensemble model: high generalization ability to cope with complex working conditions
Adopting an integrated framework of Random Forest (RF), Gradient Boosting Tree (GBDT), Support Vector Machine (SVM), and Multi Layer Perceptron (MLP), the warning results are output through majority voting decision:
Input features: Integrate 8 key monitoring signals including temperature, voltage, CO concentration, deformation, sound, etc
Optimization strategy: Use 5-fold cross validation to optimize parameters, RF and GBDT to optimize tree structure parameters, SVM to configure RBF kernel function, MLP to design double-layer hidden layers
Performance indicators: In single rate testing, the accuracy rate of 0.5C condition is 89.18%, and the accuracy rate of 1.5C condition is 95.76%; In the multi magnification fusion scenario, the highest accuracy achieved under 0.75C working condition is 99.39%, which is suitable for complex working conditions in energy storage power plant scenarios
4.2 LSTM Time Series Analysis Model: Dynamic Feature Capture Expert
Based on a two-layer 64 unit stacked architecture, predict the trend of thermal runaway by focusing on the temporal variation of the signal
Data processing: Construct an 8-dimensional feature vector with a time dimension and extract dynamic features using a sliding time window
Training configuration: Using cross entropy as the loss function, Adam optimizer (initial learning rate 1 × 10 ⁻ ³) trains for 20 rounds
Application advantages: With a stable accuracy rate of over 95% in full magnification scenarios and 97.58% in low magnification conditions at 0.5C, it is particularly suitable for dynamic operating environments such as electric vehicles
Core warning indicators: Verified by feature importance analysis, temperature (weight ratio of 33% -43%) and CO concentration (21% -28%) are the most critical warning parameters, which are highly consistent with the observed “temperature CO concentration synergistic change” pattern in experiments.
5.Three dimensional emergency decision-making matrix: quantitative response plan
To address the severe thermal runaway scenario of lithium iron phosphate batteries, a three-dimensional emergency decision-making matrix based on toxicity, flammability, and visibility is constructed to achieve graded and precise response. This matrix follows NFPA standards and quantifies risk through the following indicators:
5.1 Calculation of Risk Index
Toxicity Index (TI): Calculated using a formula based on the concentrations of carbon monoxide (IDLH=1200ppm) and hydrogen fluoride (IDLH=30ppm), with a value range of 0-1.0. The higher the value, the stronger the toxicity
Flammability Index (FI): Take the maximum concentration/lower explosive limit of hydrogen (LEL=4%), methane (LEL=5%), and carbon monoxide (LEL=12.5%), with a corresponding increase in flammability risk from 0 to 1.0
Visibility (V): Converted through the extinction coefficient formula, it intuitively reflects the degree of smoke obscuration
5.2 Hierarchical Response Strategy
Risk dimension: Low risk interval, medium risk interval, high risk interval
Toxicity Index (TI) 0.0-0.3 0.3-0.6 0.6-1.0
Flammability index (FI) 0.0-0.4 0.4-0.7 0.7-1.0
Visibility (V) ≥ 3 meters 1.5-3 meters<1.5 meters
Disposal strategy personnel, internal disposal collaboration, professional strength, external remote suppression
Applicable scenarios: early thermal runaway, mid-term thermal runaway, severe thermal runaway
This matrix can be integrated into the early warning system to intelligently recommend disposal plans through real-time monitoring data. For example, when a severe thermal runaway state (TI=0.79, FI=0.8, V=0.8m) is detected at a charging rate of 1.5C, the system will automatically trigger an “external suppression” response command to maximize personnel safety.
6.Conclusion
This study conducted five sets of gradient charging rate experiments to deeply analyze the overcharging thermal runaway process of lithium iron phosphate batteries, and constructed a complete closed-loop system of “signal acquisition intelligent warning quantitative response”. The main innovative achievements are as follows:
Division of thermal runaway stages: A new three-stage analysis model is proposed to accurately define the three key stages of “before safety valve opening, before voltage rebound, and thermal stability”, providing theoretical support for the precise design of early warning systems.
Multimodal warning technology: Determine CO concentration, temperature changes, and battery deformation as core warning indicators, develop fusion models and LSTM time series models, achieve multi-dimensional data fusion analysis, and achieve warning accuracy of over 95%.
Quantitative emergency decision-making: Create a three-dimensional decision matrix to upgrade the traditional empirical judgment of thermal runaway emergency response to a data-driven model, significantly improving the scientific and effective nature of emergency response.
Subsequent research will focus on different environmental temperatures (25 ℃, 45 ℃), states of charge (50%, 80%), and thermal runaway propagation characteristics of battery packs, further optimizing the universality and engineering practicality of the proposed solutions.