OVERSAMPLING TECHNIQUE-BASED DATA AUGMENTATION AND 1D-CNN AND BIDIRECTIONAL GRU ENSEMBLE MODEL FOR HUMAN ACTIVITY RECOGNITION

Abstract

The success of the human activity recognition (HAR) algorithm of a wearable inertial sensor depends on the extraction of good temporal dynamics from the sequential motor movement of humans. In the past, a heuristic method was used for the feature extraction process. Recently, deep neural network (DNN)-based studies, in which this process is automatically performed in the model, have shown good results. The performance of the DNN models is good and almost all processes are automated; however, they have some limitations—data-imbalance problems in the dataset and the small data decrease the performance. In this study, the HAR public dataset was improved by introducing an oversampling technique-based data-augmentation algorithm, and a stacking ensemble model with a one-dimensional convolutional neural network and a bidirectional gated recurrent unit as the head was used for classification. The HAR public dataset used Wireless sensor data mining (WISDM) and the University of California, Irvine human activity recognition using smartphones data set (UCI-HAR). Consequently, the proposed model showed better performance than that of previous studies for both datasets. The accuracy of the proposed model was 99.49% and 99.48%; the F1-score of the model was 99.84% and 99.78%, respectively.