A GAN-Based Anomaly Detection Approach for Imbalanced Industrial Time Series

Imbalanced time series are universally found in industrial applications, where the number of normal samples is far larger than that of abnormal cases. Traditional machine learning algorithms, such as support vector machine and convolutional neural networks, are struggling to attain high classification accuracies for class-imbalanced problems, because they tend to ensure the accuracy of the majority class. Hereby, this paper proposes a novel anomaly detection approach based on generative adversarial networks (GAN) to overcome this problem. In particular, an encoder-decoder-encoder three-sub-network generator is trained involving the elaborately extracted features from normal samples alone. Anomaly scores for anomaly detection are made up of apparent loss and latent loss. Without having any knowledge of the abnormal samples, our approach can diagnose faults by generating much higher anomaly scores when a fault sample is fed into the trained model. Experimental studies are conducted to verify the validity and feasibility of our approach, including a benchmark rolling bearing dataset acquired by Case Western Reserve University and another rolling bearing dataset which is acquired by our laboratory. Our approach can distinguish abnormal samples from normal samples with 100% accuracies on both datasets.