Appliance Identification With Ambiguously Labeled Data via Contrastive Partial Label Learning

Non-intrusive load monitoring (NILM) technology enables the estimation of both the power consumption and real-time operational states of individual (or target) appliances by analyzing voltage and aggregate current data at the power inlet. The notable performance of modern deep neural networks in energy disaggregation and appliance identification largely stems from the availability of large-scale, appliance-level consumption data with fully accurate annotations. For the widely adopted event-based appliance identification methods, as the construction of customized label data relying on intrusive sub-metering is cost-prohibitive, the aggregate power profiles are often used to acquire this dataset, which tends to result in inaccurate labels. In this work, we propose a contrastive partial label learning method for appliance identification, where each training instance is associated with a coarse candidate label set, which consists of several possible labels including the ground truth. Firstly, a momentum contrastive learning module is employed, which enriches the feature learning space by maintaining a dynamic memory queue of positive and negative samples. This mechanism enables the model to learn more discriminative and consistent representations for the same appliance class, thereby enhancing intra-class compactness and inter-class separability in the embedding space. Secondly, a specific label disambiguation cross-entropy loss function is designed. By dynamically restricting the predicted results only to the candidate label set, it enables the model to progressively increase the predicted probability of the true label during training, thereby effectively facilitating label disambiguation. By integrating the momentum contrastive learning with the label disambiguation cross-entropy loss, the proposed method effectively addresses ambiguously labeled data and achieves comparable appliance identification accuracy in real-world NILM scenarios. Extensive experiments on publicly available NILM datasets validate the effectiveness of the proposed method under the coarse label data.