Robust statistical industrial fault monitoring: A machine learning-based distributed CCA and low frequency control charts

Abstract

Over the past two decades, there has been a notable increase in the complexity and dynamism of industrial and manufacturing systems. Traditional fault detection strategies exhibit limited robustness in terms of detection accuracy, inefficiency, and poor detection outcomes. This paper presents an innovative, robust machine learning (ML) based distributed canonical correlation analysis (DCCA) and low-frequency generated univariate control charts CUSUM and EWMA using wavelet transforms (WT) to monitor industrial and manufacturing systems, specifically focusing on abnormalities and fault detection. The framework’s performance is tested using previous algorithms such as multiscale PCA and PLS, and the (CSTR) is utilized as the benchmark case application. The results illustrate that the ML-based multiscale low-frequency (DCCA) monitoring charts for fault cases exhibited an (MDR) of 0 %, (FAR) of 0.2 %, (FDR) of 100 %, precision of 99.8 %, and an F1-score of 99.9 %. This method enables robust detection and ensures complete monitoring of aberrant circumstances.