Noise-robust anomaly detection and classification for rotating machinery with deep filter fusion

To address the challenges of noise and data scarcity in rotating machinery diagnostics, this study proposes a deep filter fusion framework for robust anomaly detection and unsupervised clustering. The framework trains anomaly transformer models on normal data refined by complementary noise filters, then applies K-means clustering to the resulting multidimensional anomaly scores, separating samples into clusters that distinguish among different fault types without labels. The method was validated on the bearing and aluminum disk datasets, covering various bearing and structural faults. Across noise levels, the fusion approach consistently achieves higher Macro-F1 than single-filter baselines. Under severe noise at a signal to noise ratio of 4 dB, the method remains effective and the strongest cases exceed 90 percent Macro-F1, while typical cases still show clear Macro-F1 gains over baselines. This framework can therefore support industrial diagnostics by helping engineers classify fault types, leading to more informed maintenance decisions.