Battery Thermal Runaway Prediction from Multi-Sensor ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current thermal runaway prediction strategies for batteries rely heavily on empirical threshold settings, are subjective, and inefficiently utilize massive data, requiring frequent model modifications to account for sensor failures and lack integration with cloud technologies for big data updates.
Innovation Solution
A machine learning-based method using linear regression predictors trained on thermal runaway data from real-world sources, including module temperature, pressure, and voltage, with iterative updates and mini-batch gradient descent algorithms to predict thermal runaway events accurately.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If empirical threshold settings are used for thermal runaway prediction, then the implementation is simple, but the prediction accuracy is low and the method is subjective
Solution Approach 1:
The patent replaces the mechanical/empirical threshold-setting system with a machine learning-based predictive system. Instead of manually setting fixed thresholds based on experience, the system uses trained models (such as random forest, support vector machine, or neural network models) that automatically learn optimal prediction boundaries from historical battery data, thereby eliminating subjectivity and improving prediction accuracy while maintaining implementation feasibility through automated deployment
Solution Approach 2:
The patent transforms the static empirical threshold parameters into dynamic predictive parameters generated by machine learning models. The system continuously processes multiple battery parameters (temperature, voltage, current, pressure) and their derivatives through trained models to generate dynamic risk assessment results, allowing the prediction criteria to adapt to different battery states and conditions rather than relying on fixed empirical values
2Quantity of substance
If massive battery data is collected, then the information availability increases, but the data utilization efficiency decreases due to lack of effective processing methods
Solution Approach 1:
The patent segments the massive battery data into structured feature dimensions including basic parameters (temperature, voltage, current, pressure), derivative parameters (rates of change, second derivatives), and contextual parameters (charging state, ambient conditions). This segmentation allows the machine learning models to process different types of data efficiently and extract meaningful patterns from each dimension, transforming raw data volume into actionable predictive insights
Solution Approach 2:
The patent creates a virtual copy of the battery system through machine learning models that replicate battery behavior patterns. By training models on historical data, the system creates digital twins that simulate battery thermal runaway characteristics, allowing efficient analysis of massive datasets without requiring direct processing of all raw data in real-time, thus improving data utilization efficiency
3Adaptability or versatility
If manual model modification is performed to account for sensor failures, then the system adaptability improves, but the R&D time and effort increase significantly
Solution Approach 1:
The patent implements self-service through automated machine learning pipelines that handle model adaptation without manual intervention. The system automatically detects sensor failures through data quality monitoring, selects appropriate imputation strategies, retrains models with adjusted parameters, and deploys updated prediction algorithms. This automation maintains high adaptability to sensor failures while eliminating the need for manual R&D time and effort
Solution Approach 2:
The patent incorporates feedback mechanisms where the system continuously monitors prediction performance and sensor data quality. When sensor failures or anomalies are detected, the feedback loop automatically triggers model retraining with adjusted parameters or alternative data sources. This closed-loop feedback system enables the model to adapt to sensor failures dynamically without requiring manual intervention, maintaining both adaptability and efficiency
4Measurement precision
If cloud technologies are integrated for big data updates, then the prediction capability improves, but the system complexity increases
Solution Approach 1:
The patent introduces a cloud-based model training and deployment platform as an intermediary between data collection and local prediction systems. The cloud platform handles heavy computational tasks including model training on aggregated data from multiple sources, hyperparameter optimization, and model version management. Local edge devices receive pre-trained models and perform real-time predictions with minimal computational resources, thereby improving prediction capability while managing system complexity through distributed architecture
Data Source
Figure 1
AI summary
The disclosure provides a method and system for predicting thermal runaway of a battery based on a machine learning. The method for predicting thermal runaway of the battery based on the machine learning includes: S01: obtaining characteristic parameters related to thermal runaway data, wherein, thermal runaway data are stored in a pre-established battery real database; S02: dividing a training set and a testing set, wherein, both the training set and the testing set come from the real battery database; S03: constructing a linear regression predictor for evolutionary operation; S04: training the linear regression predictor based on the training set; S05: predicting real parameters of the thermal runaway based on the testing set, which means inputting real data of the battery in S01 to the linear regression predictor, and comparing a predicted result of the linear regression predictor with the real data of the battery in S01. The disclosure utilizes massive data to enable a model to iterate quickly, improve an eigenvalue of a thermal runaway state, which effectively improves a determining accuracy of the thermal runaway and predicts a thermal event.