An Attention-based Deep CNN-BiLSTM Model for Forecasting of Fatigue-induced Surface Electromyography Signals During Isotonic Contractions

 An accurate estimation of muscle fatigue is critical  for adaptive control of existing assistive devices, such as an  exoskeleton, prosthesis, and functional electrical stimulation  (FES)-based neuroprostheses. However, the estimation of muscle  fatigue using surface electromyography (sEMG) for a long  duration of time becomes challenging due to loosening of sEMG  sensors, sweating, and other accidental failures. These problems  can be potentially solved by forecasting future sEMG signals using  initially recorded high-quality data points. For the first time, we  attempt to forecast the fatigue-induced electromyography signal  using the initial sEMG recorded for a shorter interval of time,  during biceps curl with weights of 1 kg, 2 kg, 3 kg, and 4 kg. An  attention-based deep CNN-BiLSTM neural network model that  captures input sEMG dynamics to forecast future sEMG signals  corresponding to fatigue state was trained and tested. An average mean absolute percentage error (MAPE) of 26.7% between  forecasted and recorded sEMG was observed across eight  subjects, five muscles, and four weights. In addition, the time  domain features like integrated EMG (IEMG), root-mean-square  (RMS) value, and variance of EMG (VEMG) were compared  between forecasted and recorded sEMG (fatigue state), which  yielded an average MAPE of 8%, 19.2%, and 31.7%, across eight  subjects, five muscles, and four weights, for (IEMG and MAV),  RMS, and (VEMG and SSI) respectively. The results encourage  combining the proposed approach with wearable technology for  forecasting fatigue-induced sEMG to drive stimulation devices like  FES and robotic devices.