Theses Doctoral

Signal Processing and Machine Learning Methods for Internet of Things: Smart Energy Generation and Robust Indoor Localization

Chen, Leian

The application of Internet of Things (IoT) where sensors and actuators embedded in physical objects are linked through wired and wireless networks has shown a rapid growth over the past years in various domains with the benefits of improving efficiency and productivity, reducing cost, providing mobility and agility, etc. This dissertation focuses on developing signal processing and machine learning based techniques in IoT with applications to 1) smart energy generation and 2) robust indoor localization in smart city.

Smart grids, in contrast to legacy grids, facilitate more efficient electricity generation and consumption by allowing two-way information exchange among various components in the grid and the users based on the measurements from numerous sensors located at different places. Due to the introduction of information communications, a smart grid is faced with the risk of external attacks which is aimed to take control of the grid. In particular, electricity generation from photovoltaic (PV) systems is a mature power generation technology utilizing renewable resources, owning to its advantages in clean production, reduced cost and high flexibility. However, the performance of a PV system can be susceptible and unstable due to various physical failures and dynamic environments (internal circuit faults, partial shading, etc.).

To safeguard the system security, fault or attack detection technologies are of great importance for PV systems and smart grids. Existing approaches on fault or attack detection either rely on the prediction by a predetermined system model which acts as reference data for comparison or can be applied only within a certain set of component (e.g., several PV strings) based on local statistical properties without the capability of generalization. Furthermore, the output performance of a PV system is dynamic under different environmental conditions (irradiance level, temperature, etc.), which can be optimized by the technique of maximum power point tracking (MPPT). However, previous studies on MPPT usually require prior knowledge of the system model or high computational complexity for iterative optimization.

Smart city, as another important application of IoT, relies on analysis of the measurement data from sensors located at users and environments to provider intelligent solutions in our daily life. One of the fundamental tasks for advanced location-based services is to accurately localize the user in a certain environment, e.g., on a certain floor inside a building. Indoor localization is faced with challenges of moving users, limited availability of sensors and noisy measurements due to hardware constraints and external interferences.

This dissertation first describes our advanced fault/attack detection and localization methods for PV systems and smart grids, then develops our enhanced MPPT techniques for PV systems, and finally presents our robust indoor localization methods for smartphone users, based on statistical signal processing and machine learning approaches.

In Chapter 2 and Chapter 3, we proposes fault/attack detection method in PV systems and smart grids respectively in the framework of abrupt change detection utilizing sequential output measurements without assuming any prior knowledge of the system characteristics or particular faulty/attack patterns, such that an alarm will triggered regardless of the magnitude or the type of faulty/attack signals. Starting from the proposed fault detection method in Chapter 2, we present our fault localization method for PV systems in Chapter 4 where the central controller is able to identify the faulty PV strings without full knowledge of each local measurements.

Chapter 5 studies the MPPT method under dynamic shading conditions where we adopt neural networks to assist the identification of the global maximum power point by depicting the relationship between the system output power and the operating voltage. In Chapter 6, to tackle the challenge of accurate and robust indoor localization for smart city when sensors provides noisy measurement data, we propose a cooperative localization method which exploits the readings of the received strengths of Wi-Fi signals at the smartphone users and the relative distances among neighboring users to combat the deterioration due to aggregated measurement errors.

Throughout the dissertation, our proposed methods are followed by simulations (of a PV system or a grid under various operating conditions) or experiments (of localizing moving users with smartphones to record sensors' measurements). The results demonstrate that our proposed fault/attack detection and localization methods and MPPT schemes can achieve higher adaptivity and efficiency with robustness against various external conditions an lower computational complexity, and our cooperative localization methods have high localization accuracy even given large measurement errors and limited measurement data.


  • thumnail for Chen_columbia_0054D_17109.pdf Chen_columbia_0054D_17109.pdf application/pdf 2.71 MB Download File

More About This Work

Academic Units
Electrical Engineering
Thesis Advisors
Wang, Xiaodong
Wright, John
Ph.D., Columbia University
Published Here
April 13, 2022