Self-Supervised Event Detection and Labeling Method for Nonintrusive Load Monitoring

Traditional event detection methods in nonintrusive load monitoring (NILM) often rely on manual labeling and predefined statistical assumptions, limiting their adaptability to dynamic industrial environments with complex and variable loads. To address these limitations, we propose a novel self-supervised event detection and labeling method based on contrastive learning (CL). This approach automatically tags events from collected data, without requiring manual intervention or assumptions about the data’s statistical properties, making it adaptable to new environments. The method works by learning feature representations through CL, using cosine similarity between adjacent window embeddings to detect events, where significant changes are identified when the similarity drops below an adaptive threshold learned during training. In addition, the method is implemented on low-cost devices suitable for industrial deployment (Raspberry Pi 5) to demonstrate experimental results. The method outperforms state-of-the-art algorithms on the public building-level fully labeled dataset for electricity disaggregation (BLUED) dataset, achieving superior $F1$ and accuracy scores, and achieves a 96.67% event detection accuracy on a private dataset.