Improved process monitoring and supervision based on a reliable multi-stage feature-based pattern recognition technique

Abstract

This paper investigates two new multisensor data fusion algorithms for object detection in monitoring of industrial processes. The goals were to reduce the rate of false detection and obtain reliable decisions on the presence of target objects. The monitoring system uses acceleration sensors and is used as a sensor-cluster. In principle the approach can include arbitrary data acquisition techniques. Two approaches were proposed. The first uses a short-time Fourier transform (STFT) as a prefilter to extract relevant features from the acceleration signals. The features extracted from different sensor channels are first classified using support vector machine (SVM)-based filters. A novel decision fusion process to combine individual decisions was developed. The second approach uses a continuous wavelet transform (CWT) as a prefilter to extract relevant features from the acceleration signals. The features extracted from different sensor signals are subjected to further prefiltering processes before SVM-based classification. The individual decision functions are then combined in a decision fusion module. The classification system was trained and validated using real industrial data. The two approaches were tested using the same data and their performance and modeling complexity are compared. The developed approaches show strong improvements in detection and false alarm rates.