A novel method of non-stationary sEMG signal analysis and decomposition using a latent process model

To solve the problems of conventional signal analysis methods about non-stationary and frequency characteristics of surface electromyogrphy (sEMG) is of great significance to rehabilitation robot control with EMG-based human-computer interfaces (HCI). In this paper, the latent process models of sEMG signals were developed based on the combination of time-varying auto-regression (TVAR) model and dynamic linear model (DLM), which decomposed the signals into several components, and each component represents different time-frequency behavior of sEMG signals. On the basis of the latent process model, time-varying parameters, modulus and wavelength features were extracted. The fusing features of sEMG signals in two elbow movement conditions (elbow flexion and elbow extension) were adopted for clustering analysis and classification of data was visualized by using self-organizing map (SOM). An experiment with 9 healthy participants was carried out to verify the validity of this algorithm. The result implied that latent process model is a meaningful and valuable non-stationary sEMG signal analysis method which may be promising in rehabilitation robot control.