Perturbation-driven data augmentation for time series anomaly detection improvement in predictive maintenance

In predictive maintenance (PdM), time-series anomaly detection using IoT sensors has become a common approach for monitoring the condition of industrial machines in real-time. However, it is difficult to determine the condition of the machine accurately due to complex system infrastructure and the insufficient number of anomalies obtained for the training data. In particular, class imbalance derived from a very small set of anomalies leads to the degradation of the anomaly detection model’s performance. To resolve class imbalance, we propose data augmentation based on additive perturbation to improve anomaly detection model performance. Adding generated abnormal data to the training set enhances the model’s ability to capture anomalies based on similarity with noise perturbation. Case studies of industrial air compressors and rectifiers in the rotary machine were conducted to evaluate the proposed perturbation-driven augmentation method. The results showed that the proposed augmentation method could achieve anomaly detection accuracy over 30% higher than recent data augmentation techniques. Thus, we believe that the proposed method can enhance the performance of time-series anomaly detection in predictive maintenance.