Distribution-Based Residual Monitoring Using Multi-Scale ICA and Kantorovich Distance for Anomaly Detection in Noisy Multivariate Processes

In the era of Industry 4.0, the rapid advancement of intelligent manufacturing has rendered data-driven anomaly detection (AD) a critical approach for ensuring reliable and efficient operations within industrial systems. This paper presents a non-Gaussian distribution residual monitoring strategy for anomaly detection in multivariate processes subjected to noisy conditions. Specifically, the framework of Independent Component Analysis (ICA) is integrated with multi-scale filtering utilizing wavelet techniques to manage measurement noise in process variables.Furthermore, the integration of the multi-scale (MS) ICA strategy with the distribution-based Kantorovich Distance (KD) indicator enhances the detection of anomalies, addressing the limitations of conventional anomaly indicators that struggle to identify abnormal changes in data amidst noisy environmental conditions. The effectiveness of the proposed framework is evaluated through its capacity to detect four distinct types of anomalies in a distillation column (DC) setup. Additional investigations are conducted to assess the influence of moving-window selection, kernel function choice for kernel density estimation (KDE), and the selection of wavelet families on the proposed strategy's performance. The results unequivocally demonstrate that the combination of MS ICA modeling with the KD distribution metric yields a robust framework for monitoring various anomalies in multivariate processes. Moreover, the proposed MSICA-KD strategy surpasses the performance of traditional anomaly indicators based on ICA and multi-scale principal component analysis (MSPCA) in anomaly detection strategies.