Unsupervised Anomaly Detection in Industrial Engine Under Varying Load Conditions

Abstract

Reliable identification of abnormal behavior in industrial machines is important for ensuring operational reliability of the systems. Traditional condition monitoring methods rely on manually derived features and fixed thresholds, which are not suitable in varying operating conditions. To address these limitations, an anomaly detection architecture has been proposed for acoustic signals in an industrial engine using unsupervised learning. Acoustic data has been collected, where raw audio recordings have been aligned with the engine data using timestamps. A structured feature extraction pipeline has been developed, where time-domain and spectral characteristics have been extracted from each audio segment. Four unsupervised anomaly detection methods have been implemented and evaluated: K-Means clustering, Local Outlier Factor, Isolation Forest, and an Autoencoder. A load-based modeling method has been used, where each anomaly detection model has been trained separately for each engine load condition. Samples with anomaly scores above the 95th percentile of the score distribution are treated as anomalous. In addition, the original decision boundaries of Isolation Forest and Local Outlier Factor are also evaluated to examine how different thresholding strategies influence the detection results. Findings of the thesis demonstrate that anomaly detection in acoustic signals can be carried out effectively through unsupervised learning methods, without relying on labeled fault data. The developed framework opens several directions for future investigation, including validation on real fault data, applicability to different kinds of industrial machinery, and integration into real-time condition monitoring systems operating under continuously changing conditions.