Recognition of Unmeasured Motions via Synthetic EMG Generation

A major barrier to the practical use of EMG-based classification of hand and wrist motions is the cost of collecting training data. Most real upper-limb movements are "combined motions" in which, for example, wrist flexion and hand grasping occur simultaneously. Because the number of possible combinations of basic motions is enormous, measuring all of them in advance is not realistic. In this research, we synthesize EMG patterns of unmeasured combined motions from measured basic motion data and use them for training, making it possible to classify motions that have never been measured. The most straightforward idea is to superimpose basic motion patterns directly in the signal domain. However, this assumes that combined motions can be represented as linear combinations of basic motions; when that assumption breaks down due to nonlinear neuromuscular phenomena such as muscle co-contraction, the synthesized patterns diverge from signals that are actually measured. To overcome this limitation, we encode EMG signals into a low-dimensional latent space using generative models such as variational autoencoders (VAEs), and structure the latent space itself so that combined motions lie between their constituent basic motions. Generating synthetic patterns inside such a well-structured latent space yields patterns closer to real muscle activity than simple superposition of signals, which in turn improves classification performance for unmeasured motions. Technical approaches: - Synthesis of combined motions by superimposing basic motion patterns - Structuring of the latent space with consideration of mixing consistency - Synthetic data generation in the structured latent space and classifier training