Residual-based unsupervised bearing fault detection using ICA-enhanced LSTM autoencoder

Abstract

Early fault detection in rolling element bearings remains a challenging problem, particularly under unsupervised conditions where labeled fault data are unavailable. Incipient defects often generate weak impulsive vibration signatures that are easily masked by operational noise. This paper proposes a residual-based unsupervised fault detection method that integrates Independent Component Analysis (ICA) with a Long Short-Term Memory (LSTM) autoencoder. ICA is employed to decompose vibration signals into statistically independent components, enhancing fault-related impulsive features while suppressing redundant background vibration. An LSTM autoencoder is trained exclusively on healthy-condition data to learn normal temporal dynamics. Bearing anomalies are identified through reconstruction residuals, which quantify deviations from learned healthy behavior. Instead of fixed or heuristic thresholds, the decision boundary is determined via F1-score optimization, framing fault detection as a data-driven residual decision problem. The proposed approach is validated using the Case Western Reserve University (CWRU) bearing dataset under a 2 hp load condition, focusing on inner race faults. Experimental results demonstrate perfect fault recall and an ROC-AUC of 1.000, confirming the effectiveness of ICA-enhanced residual learning for early fault detection. The method is computationally efficient, interpretable, and suitable for practical predictive maintenance applications.