TPA-AD: A Two-Stage Pseudo Anomaly-Guided Method for Bearing Time-Series Anomaly Detection

Researchers introduce TPA-AD, a two-stage machine learning method for detecting anomalies in bearing time-series data using only normal training samples. The approach generates synthetic anomalies near normal boundaries and uses contrastive learning to identify degradation patterns, demonstrating improved performance on bearing fault detection and broader applicability across 13 public anomaly detection datasets.

TPA-AD addresses a significant challenge in industrial predictive maintenance: detecting bearing faults when only healthy operational data exists for model training. This constraint reflects real-world conditions where anomalies are rare and expensive to collect, making traditional supervised approaches impractical. The method's innovation lies in its pseudo-anomaly generation strategy, which creates synthetic anomalous samples near decision boundaries rather than using random injection or pre-defined fault categories. This targeted approach forces the model to learn more nuanced separations between normal and degraded states.

The research builds on established anomaly detection techniques but combines them strategically—reconstruction models for generating candidates, per-feature error control for boundary proximity, and contrastive learning for representation learning. The use of k-nearest neighbors for final scoring provides interpretability compared to black-box deep learning alternatives. Testing on bearing datasets and degradation processes validates the method's core application, while extension to 13 public datasets suggests broader applicability beyond specialized industrial use cases.

For industrial IoT and predictive maintenance sectors, improved bearing fault detection directly reduces unplanned downtime and maintenance costs. The method's ability to handle mixed continuous-discrete features addresses practical deployment scenarios where sensor data varies in type. The sensitivity to degradation evolution enables early intervention before catastrophic failure, extending asset lifetime and improving safety in high-speed rail and other critical infrastructure.

Future validation should focus on real-world deployment with actual bearing degradation trajectories and comparison against current industrial maintenance schedules to quantify economic impact.

• →TPA-AD generates pseudo-anomalies near normal data boundaries to improve fault detection without requiring labeled anomalies
• →The method combines reconstruction models, contrastive learning, and KNN to achieve window-level and point-level anomaly scoring
• →Approach handles mixed continuous and discrete features, addressing practical industrial sensor scenarios
• →Demonstrates competitive performance on bearing fault detection and 13 public time-series anomaly detection benchmarks
• →Particularly valuable for predictive maintenance in critical infrastructure where anomalies are rare and expensive to obtain