Unsupervised Anomaly Detection of a Home Appliance by Monitoring EMI Data

We propose an anomaly detection method by approximating the system state using electromagnetic interference (EMI) data. Since the harmonics of switching frequency cause characteristic patterns in conducted emission (CE) currents, understanding CE patterns can be exploited to detect an anomaly state caused by physical or functional defects in the system. To capture the CE patterns that follow an intractable distribution, we introduce a method based on a variational generative model. The anomaly data in a real-world scenario is challenging to obtain, and as such, we determined an approximated distribution for the normal states of a system to detect an outlier. Further, we designed a manifold space with a multi-modal prior distribution, thus our method can be extended to consider the entire system. To evaluate our approach, we manually collected normal and anomaly EMI data from the outdoor unit of an air conditioner. Using the EMI data from a normal state, we approximated the manifold distribution that follows an tractable distribution and demonstrate the possibilities for outlier detection. While common-mode (CM) CE EMI noise are mainly used for configuring the state of a system, we also apply our approach to the differential-mode (DM) as well as for direct power line noise.