Transfer Function Identification with Few-Shots Learning and a Genetic Algorithm Using Proposed Signal-Signature

Abstract

Mathematical models help simulate system dynamics and identifying parameters impacts prediction and control. Various techniques exist for parameter estimation. The DC motor is commonly used for position and speed control due to its ease of use and precision, especially in low-power applications. However, accurately parameterizing with low error can be challenging. Few-shot learning, which involves identifying relevant parameters from a small amount of data, has gained popularity and is particularly useful when working with limited datasets. This work presents a new technique for identifying system transfer functions using few-shot learning. It assigns a unique signature value to each system and has been successfully tested on 1500 randomly generated systems. The approach reduced the search space significantly, enabling successful identification of all systems using a genetic algorithm. R-square values ranged from 0.99 to 1.0, with only 5% of samples falling out of range.