Motif Difference Field: An Effective Image-based Time Series Classification and Applications in Machine Malfunction Detection

Abstract

Time series motifs are widely used in time series analysis. The time series motifs are used for the discovery of order structures and patterns in time series. The motif difference field (MDF) is proposed based on time series motifs. Compared to the other image representations of time series such as Recurrence Plots, MDF images are simple and effective to be construct. Taking the Fully Convolution Network (FCN) as the classifier, MDF demonstrates the state-of-the-art performance on the UCR time series datasets compared with other time series classification methods. The triadic MDF-FCN classifier gives the best result in the test. Furthermore, the MDF-FCN classifier is used for the malfunctioning industrial machine investigation and inspection sound dataset (MIMII). The MDF-FCN classification AUC demonstrates that it can have outstanding performance in realistic scenarios. Given the recent development of the Cloud Tensor Processing Units (TPU), the MDF-FCN classifier has great application potential in edge computing.