Machine learning based data analytics for IoT devices

• Main Author: Xu, Hang
• Other Authors: Yu Hao
• Format: Theses and Dissertations
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/72342
• Institution: Nanyang Technological University

The Internet of Things (IoT) is a new emerging Eco-system of networked devices with the potential of delivering significant benefits to individuals and societies. By incorporating of IoT system into human lives, physical world information can be actively collected and analyzed to facilitate the decision making process towards ultimate benefits. To perform data analytics and smart decisions, machine learning based algorithm becomes popular for IoT system. In this thesis, machine learning based algorithm has been applied to IoT application in smart home/building field and biomedical field. From machine learning algorithm perspective, the thesis first has studied single hidden layer neural network (extreme learning machine) for data regression. Then such neural network is further extended to multi-layer neural network (convolutional neural network) for more complex data analytics. From IoT application perspective, smart building energy management system and lensless blood-cell counting system have been designed and built with machine learning based data analytics. •For modern residential buildings, the energy can be provided by main electricity power-grid as well as additional power-grid of solar energy. As it is common for main electricity power-grid to experience extremely huge load demand during the peak period, renewable energy, such as solar energy, should be optimally allocated for alleviating the effect of peak load demand on the main power-grid. Therefore, the development of smart building energy management system (BEMS) becomes strongly needed with ambient intelligence (AmI). Here, the AmI of BEMS refers to data analytics within an indoor environment in response to occupants' presence. For current indoor AmI, the main challenge is how to response dynamic ambient changes in real time because of the latency of processing backend in cloud. Another challenge is how to deal robustness and scalability in large space by using a distributed solution. Therefore, a real-time data analytics platform with distributed architecture is needed for smart building with AmI. Moreover, occupant behavior with random nature, which has significant effect on energy consumption in residential buildings, is also a challenge for load forecasting. It is necessary to proposed an efficient and feasible approach to detect and analyze the occupant behavior in real time. To tackle aforementioned challenges, this thesis has proposed an on-line sequential machine learning based data analytics platform with distributed architecture. It can build and update forecasting model continuously without increasing computation complex due to incremental least-square solver. Then, occupant behavior is monitored by real-time indoor positioning system through Wi-Fi signal detection and analytics. Based on extracted occupant behavior profile and energy consumption profile, solar energy allocation problem can be solved which aims at decreasing the peak load demand in main electricity power-grid. The flow of the machine learning based data analytics for BEMS can be summarized: sensing real-time data, predicting occupant behavior profile and energy consumption profile, and allocating solar energy. Instead of using cloud, all these operations are running on the smart-gateway network with limited computational resource.In the experiment, the proposed approach can achieve better in load forecasting accuracy by 14:83% comparing with support vector regression (SVR). Based on the extracted energy consumption profile and occupant behavior profile, solar energy can be allocated, achieving reduction of peak load by 15:20% and saving of energy cost by 51:94% with comparison of static prediction based allocation strategy. •With the rapid development of biomedical IoT system, lensless blood-cell counting system (LBCS) has become a potential solution for blood cell counting in point-of-care testing (POCT) that can offer critical information for rapid on-site disease diagnosis and monitoring. Composed by the complementary metal oxide semiconductor (CMOS) and microfluidic channel, the LBCS has competitive advantage over conventional lens based imaging system due to mass production of inexpensive CMOS and rapid development of lab-on-a-chip technology. However, such a system can only capture low-resolution images with loss of detailed information in cell morphology because there is no optical lens. It is imperative to improve the image resolution through super resolution (SR) processing in system-level for cell counting and recognition, For current SR technology, the main challenge is how to improve captured image resolution with low cost of processing resource and also without degrading throughput. In this part of thesis, convolutional neural network based super-resolution (CNNSR), which is lightweight, feed-forward and potential in implementation of on-chip hardware, has been employed to tackle challenges mentioned above. Moreover, a lensless blood cell counting prototype is demonstrated, which not only utilizes custom designed backside-illuminated CMOS image sensors but commercial CMOS image sensors. Based on such a prototype, when the captured low- resolution cell image is input, a corresponding high-resolution image is reconstructed through CNNSR. In experiment, the image resolution can be improved by 4. Comparing with another machine learning technique, extreme learning machine (ELM), CNNSR can perform better in resolution improvement quality by 9:5%. Such experiment results can demonstrate that CNNSR approach possesses the potential of high-efficiency and high-quality resolution improvement in the LBCS towards POCT. In summary, the main contribution of this thesis can be summarized as follows. Firstly, to manage residential building energy consumption and optimize renewable energy source allocation yet without sacrificing occupants' comfort, an on-line sequential machine learning based data analytics platform with distribution structure has been proposed. It can build model with continuously update capability for both occupant behavior profile and energy consumption profile extraction. Considering the result of short-term load forecasting, solar energy allocation decision is made to reduce peak load and save energy cost. Using such a platform, data can be analyzed and predicted in real time with low computation complex. Secondly, to solve the low-resolution limitation of lensless microfluidic imaging for point-of-care testing blood cell counting, CNNSR is employed to recover high-resolution cell images from corresponding captured low-resolution images. Using such an approach, the cost of processing resources is low and the system throughput is not degraded, which is matched well with the developed lensless blood cell counting system.