Anomaly Detection Using Autoencoders for Movement Prediction

The smaller the time window, the faster the response of a prosthesis to the user's movement. However, very small windows have very little information, making it difficult to classify the surface electromyography signal (sEMG). This article presents the use of autoencoders for the detection of motion in real-time processing. For this purpose, a time window of 0.01 s window (i.e., ten samples per window). The difference between the number of peaks and the distance between them in the resulting latent space makes it possible to classify the moment when the patient starts to move. Through an autoencoder as an anomaly detector, it was possible to classify the beginning of the user's movement, thus managing to improve the classification in real-time.