UNI-DIRECTIONAL AND BI-DIRECTIONAL LSTM COMPARISON ON SENSOR BASED SWIMMING DATA

This paper aim is to present the deep learning model comparison for swimming style recognition using publicly available sensor data and provide a comparison of Uni-directional LSTM(Long-Short Term Memory) and Bi-directional LSTM. Both neural networks were constructed using MATLAB neural network toolbox. Data for the neural networks was prepared by segmenting data into fixed size windows with overlap. To reduce the computational cost five features from time domain signal were extracted: Signal Magnitude Area (SMA), median absolute deviation (MAD), interquartile range (IQR), mean and standart deviation. And five features from frequency domain signal: entropy, energy, kurtosis, skewness and index of frequency domain signal. These features were extracted from every window. The Uni-directional LSTM was able to perform with F1-score of 87.66 % and Bi-directional LSTM with F1-score of 90.35 %.