Detection of Cracks in the Industrial System Using Adaptive Principal Component Analysis and Wavelet Denoising

In evaluating industrial cyber-physical systems, there is an emerging inclination toward employing data-driven exploration techniques for the purpose of fault detecting and diagnosing typical procedural behaviors. The primary objective is to identify occurrences of typical behavior of the fundamental factors contributing to the aberrations. Multivariate statistical analysis, namely Principal Component Analysis (PCA), has con-siderable importance within the realm of research methodologies. This method is frequently employed in the context of outlier detection metrics, including the Hotelling
$T^{2}$
statistic and the Squared Prediction Error (SPE), which are commonly utilized for fault identification purposes. The present work presents a novel and adaptable thresholding technique that utilizes a modified variant of the PCA with wavelet denoising. This method en-ables enhanced sensitivity and robustness to abnormalities while minimizing the occurrence of false alarms and missed detection rate. In this paper, the novel methodology employs a dynamic threshold that undergoes adjustments in accordance with the available data of the Hanoi University of Science and Technology ball-bearing experimental setup. The findings indicated that the novel strategy exhibited significantly more efficacy in identifying anomalies compared to the conventional technique. Moreover, it exhibited a reduced probability of producing erroneous alerts raising sensitivity and robustness for fault detection.