Building occupant sensing : occupancy prediction and behavior recognition

• Main Author: Zhu, Qingchang
• Other Authors: Soh Yeng Chai
• Format: Theses and Dissertations
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/88236; http://hdl.handle.net/10220/45814
• Institution: Nanyang Technological University

Minimizing energy consumption and maintaining occupants’ comfort levels are of topical and growing interest as people spend much time inside buildings. To achieve these goals in smart buildings, it is necessary to study the problem of occupant sensing by leveraging machine learning methods to understand occupants based on sensor signals. In this dissertation, we study occupant sensing by answering this question in two aspects: what are the occupants’ levels and behaviors? Existing methods based on feature engineering for occupant sensing are still un- satisfactory. In this thesis, we propose four frameworks based on feature learning to perform occupancy level prediction and occupant behavior recognition. For occupancy level prediction, we propose a framework of neural networks with local receptive fields to process the environmental data, especially those of CO2. Real experiments are conducted in a laboratory to evaluate the effectiveness of the proposed framework. Besides, we demonstrate that when the time-domain data is combined with frequency-domain ones, it could provide much valuable information during the feature extraction process to improve the performance of the classification of occupancy levels. Occupancy levels are the coarse information, but occupants’ behaviors are the fine- grained ones. The occupant behavior recognition problem is further divided into simple motion-related activity recognition with inertial measurements and energy- related sound event recognition. For simple motion-related activity recognition based on Inertial Measurement Unit, i.e. accelerometer, gyroscope and magnetometer, we propose a semi-supervised learning framework using Deep Long Short- Term Memory method. As the unlabeled data is pervasively available, the pro- posed method could enhance performance of the classification of simple motions. Moreover, for lightweight computing purpose, a feature learning framework based on dictionary learning is also proposed to recognize simple motion-related activities based on built-in inertial sensors of smartphones. Real experiments based on our collected inertial sensing data and some public data demonstrate that the performance of the proposed simple occupant behavior recognition is significantly improved. As simple behavior alone is not enough to enrich the understanding of occupants and indoor environments, energy-related sound event recognition is also studied for the interactive activities between occupants and appliances based on built-in microphones of smartphones. We propose a dictionary learning framework for the computationally lightweight sound event recognition application on smartphones. Furthermore, to improve the performance based on distinctive representations of recorded sound signals, the feature learning framework based on neural networks with local receptive fields is also proposed to recognize energy-related sound events. Experimental results based on our collected audio dataset and some public dataset demonstrate the effectiveness of the proposed frameworks.