Multi-Task Load Identification and Signal Denoising via Hierarchical Knowledge Distillation

Complex neural networks with deep structures are beneficial for solving problems such as load classification in Non-intrusive load monitoring (NILM) due to their powerful feature extraction capabilities. Unfortunately, corresponding complex models designed based on deep learning algorithms require high computational and memory resources. Additionally, the external noise interference during practical load identification poses a challenge. To solve these difficulties with practical industrial significance, this paper proposes a multi-task-knowledge distillation (MTL-KD) framework for NILM. The main contributions within this framework include a new feature extraction method that combines variational mode extraction (VME) and mutual information (MI) to extract unique features and filter out noise interference, an attention-based MTL model to simultaneously perform the load identification and signal de-noising tasks, and new KD modules to transfer knowledge from a complex teacher model to a small student model. Experimental evaluations conducted on public datasets such as the plug-load appliance identification dataset (PLAID) and the worldwide household and industry transient energy dataset (WHITED), as well as a private load dataset collected in the lab, demonstrate that the proposed MTL-KD framework surpasses state-of-the-art approaches.