Determination of the Optimal Neural Network Parameters for Human Movements Classification Using Data of the Wearable Personal Devices

The article is devoted to solving the problem of improving the quality of the classification of human movements according to accelerometric data of wearable monitoring devices. An algorithm for working with accelerometric data is described, which makes it possible to reduce the measured data arrays to a single dimension, as well as generate data for training a neural network for the selection and classification of movements. The choice of the optimal parameters of the neural network for the classification of movements was substantiated, which was based on a comparative assessment of the values of the percentage of errors in the classification of movements and the parameter of cross entropy. For training and testing of the neural network, the data of the results of experimental studies of the parameters of movements during the performance of squat and jump exercises were used by 20 subjects (10 women and 10 men at the age of 18 ±3.4 years). The average value of the probability of correct classification of movements based on the data for each of the subjects was 0.94, which allows concluding that there are prospects for using the developed algorithm in biotechnical systems and dynamic biometric systems.