Intelligent control and automatic detection/prediction in sensor based systems

Sensors are increasingly used for collecting data from the field for monitoring and detecting anomalous behavior. In this thesis a network of sensors are used for data collection, analysis, and detecting abnormal situations in two domains of patient health monitoring and failures in rod pump systems in an oilfield. In these application domains, there are two challenging problems: intelligent control for wireless sensor operations to make decisions on sampling to optimize life time of a sensor network and accurate anomaly detection and prediction using the collected data.;For the health monitoring application domain, a new policy-based framework of Markov Decision Processes (MDP) is formulated for energy efficient optimization problem. The optimal global policy obtained from MDP formulation can be used by distributed sensors to achieve adaptive sampling for optimal and intelligent control of both energy consumption (system lifetime) and detection accuracy. The size of MDP policy may be large with increasing number of sensors having limited memory and discretization granularity of the problem. A decision tree-based learning algorithm is applied for a compact policy representation. Computational complexity is also exponential to the number of sensors and proportional to the discretization granularity of the problem, which causes the computational scalability problem and limits the application of MDP framework on large state space cases. In order to overcome computational scalability problem, three computationally efficient learning algorithms are developed based on approaches to learn local policies for each sensor: RLAA Learning Algorithm, AMRL Learning Algorithm and COL Learning Algorithm. We successfully applied our approaches to healthcare monitoring system, and compared the performance with other methods. The results show that all three learning algorithms are scalable to sensor networks with large state space.;For the oil field domain, learning-based automatic anomaly detection and prediction algorithms are developed for artificial lift rod pump systems which fail due to various reasons and fixing them can be costly and difficult because most parts are underground. Currently, failures in such systems are detected by field experts, which take time and incur labor costs. Our approach is supervised learning-based anomaly detection techniques from field data and we developed a novel combination of two supervised learning algorithms, AdaBNet and AdaDT for this problem. These techniques are successfully applied to detecting and predicting failures in rod pump systems with real data from oilfields. Our automated anomaly detection and prediction approach can allow automated surveillance of large number of wells in an oil field to reduce cost while monitoring wells remotely.