Optimized unsupervised deep learning assisted reconstructed coder in the on-nodule wearable sensor for human activity recognition

摘要

Presently Analysing clinical data of patients using machine learning techniques enhances both outcomes for patients and operations in hospitals. Moreover, the implementation of machine-learning-based patient data processing is influenced by heterogeneous patient data and inefficient in analysing feature-learning methods. Recently, Smart healthcare applications are being fitted with wearable sensors, which are mainly used to monitor and strengthen the Human activity recognition (HAR) using supervised and unsupervised learning methods which fail to attain minimized computation time to on-nodule wearable sensors and during the processing of data in the network, it fails to reduce the reconstruction error rate with optimized accuracy during classification. Therefore, this paper suggested, an innovative, unsupervised Deep learning assisted reconstructed coder (UDR-RC) which optimize the data during pre-processing at on-nodule wearable sensors to get minimized computation time of 11.25 ns for test set size and improves recognition performance in the feature selection and extraction inside the neural network for HAR activity mechanism. In this work, Coder architecture has been fused with a Z-layer scheme to model the deep learning framework to improve accuracy and to reduce reconstruction error, Further, data analytics technique has been introduced during pre-processing to minimize the computation time. Evidence of the proposed research is performed on a Wireless Sensor Data Mining (WISDM) laboratory dataset which is open to the public. Furthermore, the findings indicate that the classification accuracy of 97.5% and Mean Squared Error rate of 0.52% has been numerically validated on-nodule wearable sensor at lab scale analysis. (C) 2020 Elsevier Ltd. All rights reserved.