Anomaly Detection Utilizing a Riemann Metric for Robust Myoelectric Pattern Recognition

TL;DR

This paper introduces a novel anomaly detection method for myoelectric pattern recognition using a simplified log-Euclidean distance on symmetric positive definite manifolds, significantly improving robustness against unseen motions.

Contribution

It proposes a new anomaly detection approach based on Riemannian metrics that enhances discrimination and boundary shaping for target and novel motion separation in MPR systems.

Findings

Achieved 89.7% and 93.9% accuracy in novel motion detection with different feature extractors.

Maintained 90% target motion classification accuracy.

Outperformed existing methods with statistical significance.

Abstract

Traditional myoelectric pattern recognition (MPR) systems excel within controlled laboratory environments but they are interfered when confronted with anomaly or novel motions not encountered during the training phase. Utilizing metric ways to distinguish the target and novel motions based on extractors compared to training set is a prevalent idea to alleviate such interference. An innovative method for anomaly motion detection was proposed based on simplified log-Euclidean distance (SLED) of symmetric positive definite manifolds. The SLED enhances the discrimination between target and novel motions. Moreover, it generates a more flexible shaping of motion boundaries to segregate target and novel motions, therefore effectively detecting the novel ones. The proposed method was evaluated using surface-electromyographic (sEMG) armband data recorded while performing 6 target and 8 novel hand motions. Based on linear discriminate analysis (LDA) and convolution prototype network (CPN) feature extractors, the proposed method achieved accuracies of 89.7% and 93.9% in novel motion detection respectively, while maintaining a target motion classification accuracy of 90%, outperforming the existing ones with statistical significance (p<0.05). This study provided a valuable solution for improving the robustness of MPR systems against anomaly motion interference.