Data-Driven Identification of Friction Dynamics in Mechatronic Systems using Machine Learning

Abstract

High-performance control and precise functioning of mechatronic systems depend on accurate modelling of friction dynamics; nevertheless, non-linear, time-varying, and system- specific friction effects are frequently missed by traditional friction models. In this paper, a data-driven model for machine learning-based friction dynamics identification in mechatronic systems is presented. The suggested method learns friction behaviour from the measured system data, allowing for an accurate depiction of complicated non-linear dynamics without the need for explicit theoretical friction formulations. A variety of neural network-based models, such as feedforward neural networks (FNNs), convolutional neural networks (CNNs), long short-term memory (LSTM) networks, as well as transformers, as well as cutting-edge machine learning methods like physics-informed neural networks (PINNs) and sparse identification of nonlinear dynamics (SINDy), are included in the framework. The integration of these methods to real-world systems is the main focus. The efficacy of the FNN, CNN, LSTM, transformers, SINDy, & PINN methods for data-driven friction modelling and system identification is assessed using a geared DC motor as a case study. The findings show that for this traditional nonlinear dynamical system, all machine learning techniques under consideration provide excellent predictive performance. Furthermore, compared to solely data-driven black-box models, the SINDy and PINN models provide improved interpretability. The comparison analysis reveals each approach's advantages and disadvantages with regard to computing complexity, interpretability, and forecast accuracy. Potential uses and future research paths are explored, and the suggested models offer a versatile basis for friction-aware modelling as well as control of mechatronic systems.