sEMG-Based Upper Limb Elbow Force Estimation Using CNN, CNN-LSTM, and CNN-GRU Models

Abstract

Muscle fibers act as a source of force in the human body. The joint actions occur due to contraction of these muscle fibers controlled by central nervous system (CNS). Surface electromyography (EMG) signals produced during any muscular activity can play a vital role in measuring the muscular force under non-fatigued and fatigued conditions. Due to nonlinear, irregular, and continuous characteristics, sEMG signals need to be preprocessed to estimate muscle force correctly. The upper limb Biceps and Triceps muscles controls the movement of the elbow joints. Any disorder in this muscle can lead to difficulty in bending the elbow, turning palms in and out, and movement of arms. This research paper proposes different deep learning-based sEMG-to-force estimation models. The proposed models estimate the force from the sEMG signals acquired from the Biceps and Triceps muscles during maximal and submaximal contractions. The proposed deep learning models are based on hybrid 1D Convolutional Neural Network-Long Short-term Memory (CNN-LSTM) and 1D Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) approaches along with self-attention mechanism. The dataset contains raw sEMG signals acquired from the human arm’s Biceps and Triceps muscles. These signals are first preprocessed to remove DC offset and high frequency noises. The models are then trained with the pre-processed signals. Results of the proposed deep learning models are compared with the conventional machine learning algorithms including Support Vector regression (SVR) and Gaussian process regression (GPR) algorithms. Different performance metrics including Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Coefficient of Determination (R2) are utilized to measure the performance of the models based on the predicted and actual force. Results shows that the proposed CNN-LSTM and CNN-GRU models achieved better performance (MSE: 0.001, RMSE: 0.023, MAE: 0.017, and R2: 0.987) and (MSE: 0.001, RMSE: 0.025, MAE: 0.019, and R2: 0.985), respectively, as compared to GPR (MSE: 0.080, RMSE: 0.282, MAE: 0.162, and R2: 0.856) and SVR (MSE: 0.008, RMSE: 0.082, MAE: 0.067, and R2: 0.482). The proposed research study is helpful in developing a rehabilitation control system beneficial for people suffering with muscular disorder.

Keywords

• Muscles force
• electromyography
• signal pre-processing
• machine learning
• deep learning
• CNN-LSTM