Unknown Load Detection for Nonintrusive Load Monitoring Based on Feature Embedding and Nearest Neighbor Learning

Nonintrusive load monitoring (NILM) can effectively evaluate and track fine-grained energy consumption at the appliance level by using instruments or devices such as smart meters, providing a reasonable and safe energy measurement solution. Existing NILM algorithms, however, rely on closed-set recognition, which cannot detect the continuous addition of new home appliances in the real world. Conventional NILM methods, moreover, face challenges in feature selection, such as information loss and ambiguous distinguishing features, which further reduce the recognition rate on the measuring device. In order to address unknown load detection, this article proposes a feature embedding and nearest neighbor learning method using dual-channel LUNAR. The dual-channel network first leverages feature embeddings extracted from the original signal to reduce the intraclass distance for known devices, thereby distinguishing them from untrained unknown devices. Learnable unified neighborhood anomaly ranking (LUNAR) then detects unknown devices by employing nearest neighbor distance learning in the embedded feature space. In order to improve the distinction between appliances with similar running modes in feature selection, a new kind of three-channel load signature is, furthermore, generated by constructing a relative position matrix (RPM), V-I trajectory, and Mel-spectrogram. Last, for feature learning, the dual-channel network consists of a residual convolutional network embedded with coordinate attention (CA) for learning multilevel semantic features and a transformer network along with rotary position embedding (RoPE) for better adaptation to current sequence tasks. Comprehensive experiments conducted on two public datasets highlight the advantages of our proposed approach.