Primawati, Ferra Yanuar, Dodi Devianto, Remon Lapisa, Dwiprima Elvanny Myori, Fazrol Rozi
This study proposes a hybrid anomaly detection framework based on Independent Component Analysis (ICA) and Bayesian Long Short-Term Memory (LSTM) networks for early fault diagnosis in rolling element bearings. The model is trained exclusively on normal condition data from the Case Western Reserve University (CWRU) Bearing Dataset, enabling unsupervised deployment suitable for real-world industrial environments. Vibration signals are preprocessed using ICA to extract statistically independent features, followed by a Bayesian LSTM that captures temporal dynamics and provides uncertainty-quantified predictions. A dynamic thresholding mechanism based on Interquartile Range (IQR) autonomously distinguishes between normal and anomalous behavior without manual calibration. Experimental results demonstrate exceptional performance with 99.1% accuracy, perfect recall (100%) on fault conditions, and zero false negatives, ensuring no faults are missed. The composite anomaly score effectively tracks degradation progression, thereby rendering the system highly reliable and practical for predictive maintenance applications. This approach combines statistical signal processing with probabilistic deep learning to deliver a robust, explainable, and adaptive solution for bearing health monitoring. Copyright © 2026 by the authors.
Department of Mechanical Engineering, Faculty of Engineering, Universitas Negeri Padang, Indonesia; Department of Mathematics and Data Science, Universitas Andalas, Indonesia; Department of Electrical Engineering, Faculty of Engineering, Universitas Negeri Padang, Indonesia; Department of Information Technology, Politeknik Negeri Padang, Indonesia
