Real-Time Implementation of a Multidomain Feature Fusion Model Using Inherently Available Large Sensor Data

This paper presents the real-time implementation of a feature fusion based learning using multidomain discriminant correlation analysis (MDCA) for accurate diagnosis of nonlinear processes. The algorithm is also implemented in a 8-b PIC-microcontroller (PIC18F45k22) from the perspective of online applications. In the MDCA, a set of multiple features is evaluated in direct and wavelet domain for each process employing large-volume available signals. Features are subjected for correlation analysis to demonstrate the variation of features and to extract low-order statistics that represent underlying phenomena. The locally evaluated statistics are fused using MDCA and obtained discriminant features are subjected to linear transformation. Finally, a set of efficient key statistics are derived for accurate characterization of various processes. The algorithm is validated statistically and integrated with decision models-k-nearest neighbor, discriminant analysis, and neural network for real-time process diagnosis. The method achieves accuracy in the range of 98.75-100%. Results and comparison analysis show the effectiveness and reliability of the proposed model.