Design of Cost-Effective Auto-Encoder for Electric Motor Anomaly Detection in Resource Constrained Edge Device

Abstract

The electric motor failure triggers the system paralyzation for various fields such as industry or transportation. Thus, continuous management is necessary. Recently, the automated anomaly detection system is adopted for reducing human exhaustion. Moreover, for improving the cost-effectiveness and monitoring stability, edge device computing is considered on system construction. For enabling edge computing, we need to achieve high performance with a low-complex anomaly detection model, considering the constrained resource. In this paper, we empirically evaluate various anomaly detection architectures from two perspectives for designing a cost-effective model. One of the perspectives is the feature aggregation method and the other one is whether to adopt the bottleneck structure or not for constructing autoencoder. The effectiveness and efficiency are improved by adopting linear feature aggregation and non-bottleneck structured auto-encoder. By combining the above two methods, the computational cost is reduced by 2 in 10k, while losing only 1.972% of the averaged anomaly detection performance.