Microcontroller ML State Determination With External Model Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training machine learning models on microcontrollers for state determination in closed-loop control is inefficient and error-prone due to resource constraints and the need for multiple iterations, which can impact control performance and introduce errors.
Innovation Solution
Implement a training agent on an external data processing system to update the machine learning model on the microcontroller, using a trace system to capture operational and state information without interfering with the microcontroller's operation, allowing for iterative updates without additional computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training is performed externally with multiple iterations, then model accuracy can be improved, but efficiency deteriorates and errors increase
Solution Approach 1:
The training process is segmented into two distinct parts: (1) initial comprehensive training performed externally on a data processing system to achieve high accuracy, and (2) lightweight incremental updates performed on the microcontroller using captured operational data. This segmentation allows each part to operate in its optimal environment, resolving the contradiction between achieving high accuracy through external training and maintaining efficiency.
Solution Approach 2:
The machine learning model is pre-trained externally before being deployed on the microcontroller. This preliminary action ensures the model achieves high accuracy in a controlled environment with sufficient computational resources, avoiding the need for multiple error-prone iterations on the resource-constrained microcontroller.
2Measurement precision
If training is performed externally with multiple iterations, then model accuracy can be improved, but the process becomes more error-prone
Solution Approach 1:
The error-prone iterative training process is extracted from the microcontroller environment and performed externally on a data processing system. Only the finalized, reliable model and subsequent incremental updates are transferred back to the microcontroller, eliminating the accumulation of errors from multiple deployment-iteration cycles.
3Speed
If the microcontroller performs training, then model updates can be done in real-time, but computational resources are insufficient
Solution Approach 1:
Instead of performing complete retraining on the microcontroller, the system performs partial incremental updates using newly captured operational data. This partial action approach allows the model to adapt to changing conditions in real-time without requiring the full computational resources needed for complete retraining.
4Measurement precision
If external training is performed, then model accuracy improves, but the number of training steps increases
Solution Approach 1:
The system maintains continuous learning by capturing operational data in the background and performing incremental model updates without interrupting the control loop operation. This continuous useful action allows the model to improve accuracy over time while minimizing training time and maintaining real-time control performance.
Data Source
AI summary
According to various embodiments, a training arrangement for training a machine learning model for state determination in closed-loop control is described, comprising a microcontroller and a data processing system external to the microcontroller. The microcontroller comprises a trace system configured to capture operational information state information determined by a machine learning model of the microcontroller and to transmit them to the data processing system which is configured to compare the determined state information with ground truth sensor information, to determine whether an update for the machine learning model is required depending on the comparison and, if it has determined that an update for the machine learning model is required, to determine an update for the machine learning model and transmit the update to the microcontroller. The microcontroller is configured to update the machine learning model in response to the reception of an update from the data processing system.