Implementation and Analysis of Interactive Myocontrol Algorithms Based on Incremental Ridge Regression with Random Fourier Features

Traditional control of myoelectric prostheses via surface electromyography (sEMG) signals utilizes standard regression models that rely on static batch learning, making them vulnerable to real-world sources of non-stationarities such as muscle fatigue, sweat, and electrode shift. This thesis applies and assesses the incremental Ridge Regression with Random Fourier Features (iRFFRR) algorithm as a real-time adaptive solution for continuous, multi-degree-of-freedom kinematic prediction to get rid of the necessity of frequent offline recalibration. To precisely replicate real-world deployment, the iRFFRR architecture was benchmarked against a linear incremental Ridge Regression (iRR) baseline and an exact nonlinear Kernel Ridge Regression (KRR) baseline using a sliding train/test set scheme with a prequential "predict-then-update" evaluation protocol. To quantify algorithmic resilience, the synthetic perturbations designed to reproduce the physical shocks - namely skin impedance, fatigue, channel mixing, and sensor failure - were simulated within the code and applied to highly redundant intact musculature (EMaGer dataset) and sparse amputee musculature (NinaPro DB8 dataset). The experimental results obtained for the 2 subjects demonstrate that the iRFFRR model successfully overcomes the exponential computational bottlenecks of exact kernel methods while achieving robust prediction performance in sEMG motion tracking. Importantly, the use of random Fourier features, specifically within the sparse amputee data of a single subject, helped prevent the "kernel collapse" observed in exact KRR. This made it possible for the iRFFRR algorithm to recover its tracking accuracy after multiple sensor failures by dynamically unlearning failed channels in sub-millisecond per-sample execution times. In the end, this work shows that iRFFRR offers a resilient, adaptive, and embedded-ready framework for long-term proportional myoelectric control.