Anomaly Detection of Cross-Operational State Streaming Data in an Edge Computing Platform

Electrical load detection provides important insights into energy consumption and operating modes of equipment. This paper proposes an integrated method combining state anomaly detection with edge computing, applied in a distributed energy station project for real-time anomaly detection of electrical loads. On the hardware side, considering the actual deployment environment, a new energy detection and sensing terminal, the Edge Industrial Unit Detector, is designed based on an edge single-board computing platform. An anomaly detection method that can run directly and completely on this hardware platform is also developed. The method first uses current spikes during state transitions as the starting point, employing extreme value theory and performing maximum likelihood estimation on the distribution of spike data to quickly respond to spikes in the data stream without making prior assumptions about the data distribution. After filtering out normal current spikes, an unsupervised clustering method is applied. By extracting features from real-time flow data through differential processing, window smoothing, and lag processing, clustering models are used to identify and differentiate normal and abnormal operating patterns of electrical loads. Unlike traditional anomaly detection techniques, which require centralized training and subsequent deployment on edge devices, the proposed method can be directly deployed on edge devices with limited computational resources. This effectively reduces the pressure on aggregation platforms and communication networks, improving the efficiency and response speed of real-time detection. After testing, the method is capable of marking abnormal regions with 100% accuracy, and the processing speed can reach one millisecond per data point.