Incipient Fault Detection in Industrial Processes Using Sliding Window K-Means Shared Dictionary Learning

Incipient faults characterized by low magnitude and slow evolution are often undetectable by conventional detection methods in their early stages. If left unaddressed, these subtle anomalies could escalate into serious safety incidents. Therefore, achieving early and sensitive detection of incipient faults is critical for ensuring safe industrial operations. This paper proposes an incipient fault detection method that integrates sliding window–based feature extraction with a shared dictionary constructed via K-means clustering. The core steps of the approach are as follows: First, multivariate time-series data are collected; Next, multiple statistical features are computed within overlapping sliding windows; Then, a shared dictionary is learned from all normal historical data using K-means clustering; Subsequently, the Euclidean distance between each new window and its nearest atom in the dictionary is used as the reconstruction error to measure the severity of anomalies; Finally, a control limit is derived to make fault detection decisions. The proposed approach is evaluated on the Tennessee Eastman (TE) process. Experimental results show that, under comparable experimental settings, the method achieves higher fault detection rates and lower false alarm rates than conventional approaches such as k-Nearest Neighbors (KNN) and Principal Component Analysis (PCA). Moreover, the method is computationally lightweight, easy to deploy online, and demonstrates strong robustness across varying operating conditions.

Keywords: Incipient fault detection, K-means clustering, Shared dictionary, Reconstruction error, Principal component analysis.