Deep-Learning LSTM Mechanism and Wearable Devices based Virtual Fitness-Coach Information System for Barbell Bench Press

This study aims to design and develop a virtual fitness-coach information system for barbell bench press based on deep-learning Long Short Term Memory (LSTM) mechanism and wearable devices. We utilizes a set of three-axis accelerometers, gyroscopes and Electromyography (EMG) sensing modules to design our proposed wearable devices. Through computer and smartphone, the analysis and real-time assessment of the weight training in barbell free bench press can be performed to avoid injury in weight training and improve the quality of training performance.In this study, 21 subjects are recruited to use our proposed wearable devices for weight training in barbell free bench press. In the training, the subject's physiological signals and videos are captured, and the subject's signals are extracted according to the 11 most common kinds of errors marked by the fitness instructor, including 7 posture errors and 4 kinds of muscle force errors. After the extracted signal is normalized, the data is fed for the Recurrent Neural Network (RNN) training through the Long Short Term Memory (LSTM) to classify the weight training errors. The experimental results show that the classification threshold used in the classification has the best classification result when set at 0.5, and the overall average accuracy, accuracy, recall rate, F1 Score, FPR and FNR are 91.84%, 89.25%, 88.17%, 88.18%, 6.50% and 11.83%, respectively. We found that in some categories, because the sensors are not powerful enough to capture the characteristics of the errors, the accuracy is low. While the overall accuracy of the other categories is higher than 85%.In order to accelerate the training speed of LSTM, we also try to use the common factor extraction analysis to reduce the data of accelerometers and gyroscopes from 24 to 18, 12 and 6 dimensions for training. When the total dimension including EMG is 30 dimensions, there is not much difference in the accuracy when the dimension is reduced to 24 or 18. However when it is reduced to 12 dimensions, the evaluation metrics are reduced to below 70%, and the False Negative Rate (FNR) has risen sharply to 30.21%. We therefore choose to reduce the training data from 30 dimensions to 18 dimensions to maintain recognition accuracy and to accelerate LSTM training.To verify the feasibility of our Virtual Fitness-Coach Information System, we have further recruited 5 subjects for user satisfaction survey of the instant voice feedback and our wearable devices. The users show relatively high satisfiaction about our instant feedback system in the following aspects: helpfulness, clearance, reliability, correctness, and performance. The users also feel relatively comfortable for our wearable devices and suggest further simplification of our wearable devices for ease of wearing.