Federated ECU Model Calibration for Real-Time Vehicle Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic control systems in vehicles face challenges in maintaining safe and effective operation due to increasing complexity, as incremental software updates fail to adapt in real-time to changing vehicle behavior reflected in data exchanged between vehicle functional units.
Innovation Solution
The implementation of distributed and/or federated machine intelligence systems that construct trainable models using pre-trained templates, dynamically vary control parameters of vehicle functional units, and modify these models based on observational data, enabling adaptive behavior modification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If incremental software updates are used to maintain vehicle operation, then system reliability is improved, but the system cannot adapt in real-time to changing vehicle behavior
Solution Approach 1:
The system pre-trains machine learning models offline using historical vehicle data to create template models. These pre-trained models are then deployed to electronic control units, enabling them to perform real-time inference and adaptation without requiring complex real-time training computations. This preliminary action resolves the contradiction by preparing adaptive capabilities in advance while maintaining simple real-time operation.
Solution Approach 2:
The patent introduces template models as an intermediary between the offline training process and real-time control. The template models serve as a bridge that captures learned patterns from historical data and translates them into actionable control strategies. This intermediary enables real-time adaptation by providing pre-processed knowledge that can be quickly applied without requiring full training computations during operation.
2Adaptability or versatility
If machine learning models are trained in real-time using vehicle data, then adaptability is improved, but computational complexity and processing time increase
Solution Approach 1:
The system performs the computationally intensive model training offline using historical vehicle data, creating pre-trained template models before deployment. This preliminary action moves the time-consuming training process away from real-time operation, allowing the system to achieve both high adaptability and fast response times by separating training from inference.
Solution Approach 2:
The patent segments the machine learning process into two distinct phases: offline training phase where template models are created using historical data, and online inference phase where pre-trained models are applied in real-time. This segmentation resolves the time-complexity contradiction by performing heavy computations during offline preparation and simple lookups during real-time operation.
3Adaptability or versatility
If complex software is distributed among multiple vehicle functional units, then system versatility is improved, but system complexity and difficulty of maintenance increase
Solution Approach 1:
The patent creates universal template models that can be deployed across multiple vehicle functional units. These template models serve multiple purposes: they can be trained offline on aggregated data from multiple sources, then deployed to various control units to provide consistent adaptive behavior. This universality reduces complexity by using a common model architecture and training approach across different functional units.
Solution Approach 2:
The system creates copies of pre-trained template models and distributes them to multiple electronic control units. Rather than each unit independently training complex models, they receive standardized template model copies that have already been trained on comprehensive data. This copying approach maintains versatility across units while significantly reducing the computational and maintenance complexity at each individual unit.
Data Source
AI summary
Systems, devices, computer-implemented methods, and/or computer program products that facilitate modifying electronic control system behavior using distributed and/or federated machine intelligence. In one example, a system can comprise a process that executes computer executable components stored in memory. The computer executable components can comprise a model manager, a control component, and a learning component. The model manager can construct a trainable model using a pre-trained template model that is received via an in-vehicle network from a domain chief. The control component can dynamically vary a control parameter of a vehicle functional unit using the pre-trained template model to calibrate an output of the vehicle functional unit. The learning component can modify the trainable model based on observational data of the vehicle functional unit.