Anomaly detection of multivariate image time series based on Gramian angular field using convolutional autoencoder

Abstract

This paper covers the implementation of predictive maintenance (PdM) predicting the wear of technical components of an industrial thermal system. Object of investigation of the technical components here is the motor of a fan. Basic requirement for the PdM system is the availability of extensive data from operating parameters of the motor, which must be recorded in advance through sensor technology. The sensor used here is a MPU 92.65 motion tracking multi sensor with three-axis acceleration, three-axis yaw rate, three-axis magnetic field and temperature measurement. Goal of this work is to develop an unsupervised Deep Learning (DL) model for anomaly detection of multivariate time series. By using the MPU 92.65 the operating parameters of the fan motor are recorded at specific time intervals ensuring sufficient data quality for the development of the DL model. In selecting a suitable DL algorithm for the anomaly detector a specific Convolutional Neural Network (CNN) called Convolutional Autoencoder (ConvAE) with 2-D convolutions has been selected. CNNs with 2-D convolutions are the typical type of neural networks normally used for images (object detection, segmentation, medical imaging etc.). However, the recorded sensor time series here are in 1-D and must therefore be transformed appropriately. The investigated approach of image transformation is the Gramian Angular Field (GAF), which encodes temporal correlation structures spatially as images. A total of four datasets is used with one representing the baseline for the deployment of the detector. The remaining datasets have been set for testing purposes with their results presented and discussed here.