A practical muscle-signal–driven control mechanism for an affordable robotic prosthetic hand

This research presents a practical signal-processing framework designed to improve the recognition of surface electromyography signals for controlling an affordable robotic prosthetic arm. Focusing on essential hand movements such as grasping and opening, the system employs low-cost muscle sensors and a multi-stage processing pipeline to enhance signal quality and extract discriminative features. Noise reduction is achieved using a Chebyshev Type II filter with a 50–500 Hz passband, followed by a nine-level Discrete Wavelet Transform using the Daubechies 4 wavelet to capture meaningful temporal–frequency patterns. Statistical and temporal descriptors, including Root Mean Square, Mean Absolute Value, Integrated Electromyography, and Waveform Length, are extracted and classified using several machine-learning models: Support Vector Machine, Random Forest, Multi-Layer Perceptron, K-Nearest Neighbors, and a Convolutional Neural Network. Experimental results show that the Convolutional Neural Network provides the highest accuracy at 95.30 %, outperforming all other classifiers. The framework is further validated through integration with a robotic prosthetic arm prototype and tested in real-movement scenarios. Comparable performance across different electrode placements indicates robustness to variability in signal amplitude. While the system is demonstrated through the specific application of prosthetic arm control, the proposed framework also provides generic value for biosignal processing, rehabilitation devices, wearable interfaces, and broader human–machine interaction applications.