Anomaly detection in rotating machinery via active learning of unlabeled vibration samples

In response to the challenges posed by the substantial volume of monitoring data from rotating machinery, the considerable effort required for manual interpretation, and the scarcity of labeled fault samples, this study proposes a vibration-based anomaly-detection method that applies active learning to unlabeled vibration signals. The key novelty is a redundancy-aware batch active learning scheme, in which predictive-entropy from a committee is combined with a long short-term memory fully convolutional network (LSTM–FCN) deep-clustering module. One most representative sample is selected from each cluster to increase diversity and reduce labeling cost. The method comprises two stages: first, predictive entropy is computed for all unlabeled samples to rank uncertainty and perform an initial screening; second, a deep-clustering procedure mitigates redundancy among high-uncertainty candidates, after which the highest-entropy instance in each cluster is selected for expert labeling. Evaluations on vibration datasets from a rolling-bearing accelerated-life rig and a centrifugal-compressor rig show consistent improvements in accuracy and recognition stability over conventional clustering-based anomaly-detection methods, alleviating the dependence on scarce labeled anomalies in real-world rotating machinery monitoring.