Similarity preserving hashing for appliance identification based on V-I trajectory

Abstract

Non-intrusive load monitoring (NILM) is a technique used to monitor energy consumption in buildings without requiring hardware installation on individual appliances. This approach offers a cost-effective and scalable solution to enhance energy efficiency and reduce energy usage. Recent advancements in NILM primarily employ deep-learning algorithms for appliance identification. However, the substantial number of parameters in deep learning models presents challenges in quickly and effectively identifying appliances. An effective technique for appliance identification is analyzing the appliances’ voltage-current (V-I) trajectory signature. This research introduces a novel hashing method that learns compact binary codes to achieve highly efficient appliance V-I trajectory identification. Specifically, this paper uses a profound structure to acquire V-I trajectory image features by acquiring multi-level non-linear transformations. Subsequently, we merge these intermediary traits with high-level visual data from the uppermost layer to carry out the V-I trajectory image retrieval process. These condensed codes are subjected to three distinct standards: minimal loss in quantization, uniformly distributed binary components, and autonomous bits that are not interdependent. As a result, the network easily encodes newly acquired query V-I images for appliance identification by propagating them through the network and quantizing the network’s outputs into binary code representations. Through extensive experiments conducted on the PLAID dataset, we demonstrate the promising performance of our approach compared to state-of-the-art methods.