Pattern Recognition of sEMG Signals by CNN-BiLSTM Algorithm for Bio-Robotics Applications

Abstract

This paper presents a hybrid CNN-BiLSTM model for surface Electromyography (sEMG) signal classification towards bio-robotic applications. The EMG signals were segmented into four-channels, 50 -sample windows and then passed through a deep convolutional network for spatial feature extraction followed by a bidirectional LSTM layer to address temporal dependencies. Dynamic learning rate adaptation and class weight balancing were employed to train the model in terms of addressing dataset variability. Evaluation on the test set demonstrated an overall testing accuracy of (87 %) with Average False Negative Rate (FNR) 13.5%, and Average False Positive Rate (FPR) 5.0%. The gestures in the data predicted as rest (96 %), extension (95 %), flexion (95 %), ulnar (93 %) and radial deviation of the wrist (82 %), grip (87 %), supination (77 %), and pronation (67 %). The results indicate the promise of the model for reliable real-time control in human-machine interface (HMI) and assistive robotic systems. This is an application that extends beyond signal classification which emphasizes the application of Knowledge Representation and Reasoning (KRR) within AI as communication principles. Human-Robot Interaction (HRI) is improved while broad scope prospects are opened for use in assistive and prosthetic devices. Surface electromyography (sEMG) signals from human muscles contain complex spatiotemporal patterns that are challenging to classify accurately for real-time robotic control applications. The inherent variability, noise, and non-stationary nature of sEMG signals pose significant challenges for reliable gesture recognition in human-machine interfaces.