Data Augmentation for Industrial Multivariate Time Series via a Spatial and Frequency Domain Knowledge GAN

Abstract

The data imbalance is an intractable issue industrial fault diagnosis. This imbalance between a small amount of fault data and a large amount of normal data usually leads to the unsatisfactory performance of the fault diagnosis. In order to solve this issue, many approaches have been presented including the cost-sensitive learning and the data augmentation approaches. Due to the fact that the common type of the industrial data is time series, however many current methods do not take advantage of their characteristics. Thus, we propose a Spatial and Frequency domain Knowledge Generative Adversarial Networks (SFKGAN), extracting information from the spatial and the frequency domain, to take account of the spatial and temporal characteristics in industrial time series. When treating the temporal characteristics, distinctive from many current GANs that almost solely focus on the time or frequency domain, our method takes into account these two domains simultaneously. Our rationale is that the frequency information is better at capturing the inner information and especially the impacts that these faults have on the system, and meanwhile consider the temporal characteristics. As to the spatial characteristic, a special auto-encoder (AE) is adopted to process the correlation among the variables. We use the Tennessee Eastman Process (TEP) dataset, a benchmark dataset in the chemical engineering field, to demonstrate the advantage of our approach over many other classical generative methods. Index Terms– Sliding Window, Wavelet Decomposition, Autoencoder (AE), Generative Adversarial Networks (GANs), Maximum Mean Discrepancy(MMD), Fault Diagnosis