LARGE SCALE SENSOR DATA PROCESSING BASED ON DEEP STACKED AUTOENCODER NETWORK

摘要

Recently, Internet of Things (IoT) extremely populated by massive amounts of connected embedded devices, which are gathering large volumes of real-time heterogeneous data. Hence, IoT becomes an archetypal instance of Big Data. The collected IoT Big Data may not be profitable unless we evaluate and accurately exploit them. Providing mining for large scales of raw sensor data is an open challenge. To cope with this challenge, we proposed a system that operates in two modes, which are preparation and processing. The preparation mode converges on reducing the factors that hinder making efficient processing by focusing on three stages. First, handling missing data by applying interval-valued fuzzy-rough feature selection methodology. It highlights the most important features that contain missing data and gets rid of the others. Then, Maximum Likelihood (ML) approach is used for estimating the missing values. Second, anomalies are detected by initially utilizing K-nearest neighbors (KNN) algorithm then removing the detected ones from the data. Third, the dimensionality of nonlinearly separable data is reduced by exploiting Self-Organizing Map (SOM) network. In the processing mode, we passed the prepared data to a straightforward classifier based on a Deep Learning (DL) approach. We used autoencoder networks in constructing a deep network, which is the Deep Stacked Autoencoder (DSAE). The extracted features by the DSAE are non-handcrafted and task dependent, which gives it the most discriminative power to work as an efficient classifier. We apply the proposed model to PAMAP2 Physical Activity Monitoring data set. The results show that DSAE achieves high accuracy (99.8%) compared to the state-of-the-art classifiers.