Abstract
This paper proposes a novel probabilistic neural network and a prosthetic arm control system designed to prevent unintended forearm motions caused by anomalies. The network incorporates a Gaussian mixture model and a one-versus-the-rest classifier model, and can be used to estimate the posterior probability of predefined and undefined classes through training with given data. The control system incorporates the proposed network, thereby preventing unintended prosthetic arm motion in which an unexpected action is performed. In experiments conducted with a forearm amputee and two healthy subjects, electromyogram (EMG) classification ability of the proposed network was demonstrated, including for cases with unlearned motions. The results of the experiments also showed that the system enables smooth prosthetic arm control using EMG signals while preventing malfunctions caused by anomalies.
Introduction: The Problem of the Closed-Set Assumption
Most modern electromyogram (EMG) controlled prosthetics rely on pattern recognition to translate muscle signals into hand movements. However, traditional classifiers—such as standard Neural Networks or SVMs—typically operate under a "closed-set" assumption: they assume every input must belong to one of the pre-defined training classes.
In real-world applications, this is dangerous. If a user performs an untrained movement, sneezes, or experiences sensor noise, a standard system will force that signal into a known category (e.g., "Close Hand"), leading to unintended and potentially hazardous prosthetic motions.
As shown in the scatter plot above, signals from an unlearned motion (Flexion, $M_3$) can easily be misclassified as a learned motion (Grasping, $M_2$) because they fall on the wrong side of a simple decision boundary in the feature space.
Proposed Methodology
The authors propose a hierarchical architecture designed to explicitly detect anomalies and reject unlearned motions, solving the closed-set problem.
Network Architecture
The proposed system consists of three main components:
- Master Classifier: A Log-Linearized Gaussian Mixture Network (LLGMN) that calculates the posterior probabilities for all pre-defined (learned) classes.
- Servant Classifiers: A set of "one-versus-the-rest" models. For every Gaussian component in the master model, a servant classifier estimates the probability that a given signal does not belong to that specific component.
- Probability Revision: This layer combines the outputs of the Master and Servant classifiers to calculate $P(\tilde{c}=0|x)$—the probability that the input belongs to an "unlearned" class (Class 0).
Training via Bootstrap Data Generation
To define the boundaries for unknown regions, the authors used a bootstrap-based data generation method. Since it is impossible to provide training data for every potential anomaly, they mathematically generated "pseudo-unlearned" data points that sit just outside the distribution of the learned classes. This creates a "safety buffer" around known motions, allowing the network to learn where the "known" world ends.
Entropy-Based Rejection
Beyond class probability, the system calculates the entropy ($H$) of the output. If the network is too "confused" (high entropy), it holds the current state rather than making a risky guess, further preventing unintended movements.
Experimental Evaluation
The system was evaluated using EMG data from two healthy subjects (6 electrode pairs) and one forearm amputee (4 electrode pairs). Learned tasks included motions like "Open," "Grasp," and "Extension," while motions like "Pronation" and "Supination" were withheld during training to serve as anomalies.
Key Results
- High Accuracy: For learned motions, the system achieved an average classification rate of 96.7% ± 3.8% across all subjects.
- Robust Anomaly Detection: The system correctly identified unlearned motions as "unknown" with 90.0% ± 12.3% accuracy.
- Superiority over Baselines: The method significantly outperformed standard LLGMN and Mahalanobis distance-based classifiers, which frequently misclassified unlearned motions as learned ones.
- Real-time Stability: In live tests, the prosthetic remained stationary when unlearned motions were performed, effectively preventing the malfunctions seen in traditional systems.
Conclusion
This research presents a significant advancement in prosthetic control by addressing the safety risks of unlearned muscle signals. By combining master and servant classifiers with synthetic anomaly training, the system provides a reliable mechanism for rejecting unknown inputs. Future work will focus on automating parameter optimization and testing the system's limits with a wider variety of complex, real-world anomalies and muscle fatigue conditions.
Summary updated