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

Abstract:

This study introduces a novel approach for axle-box bearing time-series anomaly detection (TSAD) that operates under a one-class learning setting, requiring only normal samples for training. The proposed framework, termed Two-stage Pseudo Anomaly-guided Anomaly Detection (TPA-AD, is designed to enhance detection accuracy by leveraging pseudo-anomalies.

The methodology proceeds in two distinct phases. Initially, a reconstruction model combined with per-feature target-error control is employed to generate pseudo-anomalous data windows situated near the boundary of normal data distributions. Subsequently, the model refines anomaly-sensitive feature representations by applying contrastive learning to differentiate between genuine normal windows and the generated pseudo-anomalous ones. Final anomaly scoring is achieved at both the window and point levels via k-nearest neighbors (KNN).

Unlike conventional techniques that depend on predefined fault categories, existing real anomaly priors, or arbitrary anomaly injection, TPA-AD enhances the separability of the normal boundary by strategically constructing pseudo-anomalies in its vicinity. Furthermore, the method is capable of simultaneously processing mixed-variable scenarios containing both continuous and discrete features.

Empirical evaluations were performed on bearing fault detection and degradation-process datasets, supplemented by an exploratory analysis across 13 public TSAD benchmarks. The findings indicate that TPA-AD delivers consistent anomaly detection responses, exhibits high sensitivity to degradation trends, and shows promising generalizability when applied to public TSAD standards as well as real-world bearing data from high-speed trains.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC