Battery Pack Fault Detection Using Fleet-Trained ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting battery faults in rechargeable batteries, particularly in electric vehicles, are inadequate as they rely on inaccurate modeling or lack real-time data from diverse battery populations, failing to address operational disruptions and safety hazards effectively.
Innovation Solution
A computer-implemented method using machine learning models trained on voltage and temperature sensor data from battery packs, transforming measurements into a time-series sequential window format, determining scores, and predicting fault conditions to enable timely preventative actions, with parameters updated via a cloud-based system for improved accuracy across large fleets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional fault detection methods are used, then device complexity is reduced, but measurement precision and reliability deteriorate due to inaccurate modeling and lack of real-time data
Solution Approach 1:
The patent introduces machine learning models as intermediary components between sensor data and fault detection decisions. These models process voltage and temperature measurements, transforming raw data into predictive insights without requiring complex manual analysis systems.
Solution Approach 2:
The patent replaces conventional mechanical/mathematical modeling approaches with data-driven machine learning systems. Instead of relying on inaccurate physical models, the system uses trained models that learn patterns from real-time sensor data, significantly improving measurement precision.
2Reliability
If real-time monitoring of all battery parameters is implemented, then reliability improves, but use of energy increases due to continuous data collection and processing
Solution Approach 1:
The patent extracts only the most critical features (voltage and temperature measurements) for fault detection, rather than processing all possible battery parameters. This selective approach maintains reliability while minimizing energy consumption for data collection and computation.
Solution Approach 2:
The system performs monitoring at selective intervals and focuses computation only when fault conditions are suspected, rather than continuously analyzing all parameters. This partial monitoring approach balances safety requirements with energy conservation.
3Speed
If machine learning models are trained locally in each vehicle, then response time improves, but device complexity and manufacturing difficulty increase
Solution Approach 1:
The patent performs machine learning model training in advance using historical data from the fleet, so that pre-trained models can be deployed to individual vehicles. This preliminary training action enables fast local inference without requiring complex real-time training capabilities in each vehicle.
Solution Approach 2:
The system creates copies of the trained machine learning model and deploys them to multiple vehicles in the fleet. This copying approach allows each vehicle to have its own local model for fast prediction, while avoiding the complexity of training unique models for each vehicle.
4Measurement precision
If fleet-wide data collection is implemented, then measurement precision improves through larger datasets, but loss of information increases due to data transmission and storage requirements
Solution Approach 1:
The patent extracts and transmits only essential features and aggregated statistics from fleet-wide sensor data, rather than transmitting complete raw datasets. This selective extraction maintains model training precision while minimizing data transmission overhead and information loss.
Data Source
AI summary
In one aspect, computer-implemented method may include receiving, at a cloud-based computing system, a set of measurements over a certain period of time. The set of measurements are received from a set of vehicles and pertain to voltages and temperatures of a set of battery packs associated with the set of vehicles. The method may include training, using the set of measurements, one or more machine learning models to predict a battery pack fault condition. The one or more machine learning models include a set of parameters that are modified during the training. The method may include transmitting the set of parameters to the set of vehicles to enable the set of vehicles to update, based on the set of parameters, one or more respective in-vehicle machine learning models configured to predict the battery pack fault condition.