Toward intelligent cyber-physical systems: Algorithms, architectures, and applications

Cyber-physical systems (CPS) are the new generation of engineered systems integrated with computation and physical processes. The integration of computation, communication and control adds new capabilities to the systems being able to interact with physical world. The uncertainty in physical environment makes future CPS to be more reliant on machine learning algorithms which can learn and accumulate knowledge from historical data to support intelligent decision making. Such CPS with the incorporation of intelligence or smartness are termed as intelligent CPS which are safer, more reliable and more efficient. This thesis studies fundamental machine learning algorithms in supervised and unsupervised manners and examines new computing architecture for the development of next generation CPS. Two important applications of CPS, including smart pipeline and smart grid, are also studied in this thesis. Particularly, regarding supervised machine learning algorithms, several generative learning and discriminative learning methods are proposed to improve learning performance. For the generative learning, we build novel classification methods based on exponentially embedded families (EEF), a new probability density function (PDF) estimation method, when some of the sufficient statistics are known. For the discriminative learning, we develop an extended nearest neighbor (ENN) method to predict patterns according to the maximum gain of intra-class coherence. The new method makes a prediction in a “two-way communication” style: it considers not only who are the nearest neighbors of the test sample, but also who consider the test sample as their nearest neighbors. By exploiting the generalized class-wise statistics from all training data, the proposed ENN is able to learn from the global distribution, therefore improving pattern recognition performance and providing a powerful technique for a wide range of data analysis applications. Based on the concept of ENN, an anomaly detection method is also developed in an unsupervised manner. CPS usually have high-dimensional data, such as text, video, and other multimodal sensor data. It is necessary to reduce feature dimensions to facilitate the learning. We propose an optimal feature selection framework which aims to select feature subsets with maximum discrimination capacity. To further address the information loss issue in feature reduction, we develop a novel learning method, termed generalized PDF projection theorem (GPPT), to reconstruct the distribution in high-dimensional raw data space from the low-dimensional feature subspace. To support the distributed computations throughout the CPS, it needs a novel computing architecture to offer high-performance computing over multiple spatial and temporal scales and to support Internet of Things for machine-to-machine communications. We develop a hierarchical distributed Fog computing architecture for the next generation CPS. A prototype of such architecture for smart pipeline monitoring is implemented to verify its feasibility in real world applications. Regarding the applications, we examine false data injection detection in smart grid. False data injection is a type of malicious attack which can threaten the security of energy systems. We examine the observability of false data injection and develop statistical models to estimate underlying system states and detect false data injection attacks under different scenarios to enhance the security of power systems.