Information Fusion for Pattern Classification in Complex Interconnected Systems

Typical complex interconnected systems consist of several interconnected components with several heterogeneous sensors. Classification problems (e.g., fault diagnosis) in these systems are challenging because sensor data might be of high dimension and/or convey misleading or incomplete information. So, this thesis performs optimal sensor selection and fusion to lower computational complexity and improve classification accuracy. First, the thesis presents the novel Unsupervised Embedded algorithm for optimal sensor selection. The algorithm uses the minimum Redundancy Maximum Relevance (mRMR) criterion to select the candidate list of sensors, then uses K-means clustering algorithm and entropy criterion to select the optimal sensors. The Unsupervised Embedded algorithm is applied to heat exchanger fouling severity level diagnosis in an aircraft. Second, the thesis presents a fusion algorithm for classification improvement called the Better-than- the-Best Fusion (BB-Fuse) algorithm, which is analytically proven to outperform the best sensor Correct Classification Rate (CCR). Using confusion matrices of individual sensors, the BB-Fuse selects the optimal sensor-class pairs and organizes them in a tree structure where one class is isolated at each node. The BB-Fuse algorithm was tested on two human activity recognition data sets and showed CCR improvement as expected. Using the confusion matrices is a limitation of the BB-Fuse algorithm because of high computation complexity and pertinence to specific classifiers. So, the thesis next presents the Decomposed Sensor-Class Pair Tree with maximum Admissible-Relevance for Fusion (D-Fuse) algorithm, which is a novel sensor-class pair selection and fusion algorithm. Instead of using the confusion matrices, the D-Fuse algorithm uses a novel classifier-independent information-theoretic criterion (i.e., Admissible-Relevance (AR) criterion) to obtain the sensor-class pair. As a result, any classifiers can be used with the D-Fuse Tree. In comparison to other sensor selection algorithms in literature, the novel AR criterion has two major advantages. First, the AR criterion ensures selection of non-redundant sensors by selecting optimal sensor-class pairs. This pair selection ensures that the selected sensors carry signatures about different classes; i.e., they convey non-redundant classification information. Second, the AR criterion outputs a general fusion tree that suggests an order in which the sensors should be used.