A Study on Rejecting Non-Target and Misclassified Motions for Robust Tactile-Sensor-Based Prosthetic Hand Control

Reliable motion classification is essential for practical prosthetic-hand control. Unintended activations caused by ambiguous motions, unknown motions, or non-target body movements can degrade controllability and compromise user safety. Mechanical-sensing approaches are attracting attention as alternatives or complements to surface electromyography, and tactile-sensor-based methods represent one such direction. However, despite extensive studies on prosthetic control, systematic investigations of computationally lightweight motion-rejection strategies remain limited. This study investigates rejection mechanisms to improve the robustness of polyvinylidene fluoride (PVDF) tactile-sensor-based prosthetic control. The proposed approach selectively withholds outputs for misclassified and non-target inputs. We compare three mechanisms: (1) one-class support vector machine (OCSVM) outlier detection, (2) entropy-based rejection using a multilayer perceptron (BPNN-Entropy), and (3) a parameter-free decision-consistency check for one-vs-rest support vector machines (SVMs) that withholds classification when the output sign pattern is inconsistent (one-vs-rest reject option (OvR-RO); proposed). Performance is evaluated for three sources of unintended activation: ambiguous target trials (retrospectively defined), unknown motions excluded from training, and non-target body movements. The results show that OvR-RO achieves a favorable balance between rejection rate and rejection precision for ambiguous motions, while maintaining responsiveness. Overall, explicitly rejecting misclassified and non-target motions is effective for enhancing robustness in tactile-sensor-based prosthetic control.