A CNN-LSTM neural network model for upper limb joint angle estimation based on mechanomyograph

Many stroke patients suffer from the sequelae of hemiplegia, and the emergence of rehabilitation exoskeleton robots has brought new hope for the recovery of these patients. Mirror rehabilitation training is one of the rehabilitation model, which drives the exoskeleton to perform the same movements by recognizing the biomedical signals on the healthy side, has a better effect. The traditional mirror rehabilitation training is mainly controlled by the pattern recognition results of the healthy side movement, but it limits the variety of rehabilitation training movements, and the incorrect recognition results can easily cause injury to the patients. To solve the continuous motion control problem of rehabilitation robots, we propose a lightweight model that combines CNN and LSTM to estimate the angles of upper limb joint movements using deep learning techniques. As for the selection of biomedical signals, many studies have used electromyographic signals (sEMG) to obtain the intention of movement, but due to the limitations of sEMG, in this paper we choose another biomedical signal – MMG as the object of study. We extract 8 time-domain features, 3 midfrequency domain features, and 12 wavelet packet energy features as inputs to the model to estimate the angles of the shoulder, elbow, and wrist joints of the upper limb. The experimental results show that the CNN-LSTM model has good prediction performance with higher R2 values and smaller root mean square (RMSE) values than the Back Propagation Neural Network (BPNN) model, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network. Therefore, it is feasible to use the CNN-LSTM model proposed in this paper to predict the joint angles based on the measured MMG of the upper limb and apply it to the continuous motion control of upper limb exoskeleton robots.