Enhancing Feature Extraction in Sensor Fault Detection

Abstract

The large amount of data provided by sensors and sensor networks deployed in industrial plants presents an opportunity for advancing process monitoring and informed asset maintenance management. Data-driven anomaly detection systems can be developed for both process anomalies and monitoring systems but the approach must be tailored to effectively capture and categorize relevant features, establishing a robust approach. The use of multi-source signals facilitates the retention of maximal information without reducing the number of available features, as sensor signals may show high degrees of correlation even if the sensors measure different variables across different equipment. The scope of the paper is to execute a preliminary analysis of industrial data, obtaining a logical clustering from a physical one using Canonical Correlation Analysis. This data mining methodology is employed to gather highly correlated data for training an Artificial Neural Network aimed at identifying sensor anomalies. A benchmark is provided by comparing the performance of the ANN trained using logical clusters rather than physical clusters. This comparison serves to evaluate the effectiveness and advantages of utilizing logical clusters for anomaly detection and classification.