One Class Fault Detection in Rotating Machinery Using Distributional Features from Triggered Acoustic Emission Data

Acoustic Emission (AE) monitoring offers high sensitivity in detecting faults and defects in rotating machinery. Hit-based AE acquisition systems produce only short, intermittent waveform signals rather than continuous recordings. This work addresses the challenge of fault detection from such fragmented sensor data. It is demonstrated that individual AE hits are insufficient for reliable fault detection, as waveforms from different fault conditions overlap substantially in feature spaces. To overcome this, a distributional feature aggregation approach is proposed. AE hit features are extracted from each waveform, grouped into non-overlapping sequential bags, and summarised through order statistics and distribution moments. Four one-class classifiers namely: (i) Isolation Forest, (ii) PCA one class classifier, (iii) K-nearest neighbour one class classifier, and (iv) Local Outlier Factor were evaluated on a Drivetrain Dimulator (DTS) test rig that simulates faulty gear and bearing conditions. AE signals coming from the healthy condition were used to train the classifiers. Signal deviations due to the presence of gear and bearing faults were subsequently identified. Fault detection results show that bag-level distributional features substantially outperform per-hit features: four of five faults achieve 100% single-bag detection under five-fold cross-validation, while per-hit detection rates remain below 50% for four of five faults. The PCA one class classifier achieves 100% detection for all faults individually but is excluded from the final ensemble due to a 99% false alarm rate caused by the limited healthy training sample in a high-dimensional feature space. A CUSUM sequential detector applied to a three-method ensemble (Isolation Forest, KNN, and LOF) is evaluated on a chronological 80/20 split of the healthy data, triggering alarms for all five fault conditions with a false alarm rate of 0% at the recommended setting. Four faults are detected with sustained alarm patterns (54–100% alarm rates), while bearing outer race, the most challenging fault, is detected intermittently (10.5% alarm rate), demonstrating that sequential evidence accumulation can identify faults that are invisible to single-bag thresholding.