Deep Learning Approach for Smart Home Appliances Monitoring and Classification

Abstract

With the global rise in the adoption of smart grids and smart homes, it is imperative to find approaches to efficiently monitor and manage load profiles across households. We consider aggregated load profiles during mutual operation, where the rationale is to provide a relatively robust and adaptable deep learning method that can perform appliance classification without constraints on the consumer behavior. We propose an enhanced real-time single-sensor home appliance classification and monitoring system leveraging convolutional neural networks and transfer learning. The real-time information obtained from smart meters is input into a pre-trained learning model which classifies multiple concurrently active home appliances. The convolutional neural network architectures of VGG16, ResNet50, and the InceptionV3 are trained individually by the transfer-learning paradigm with the image features of V-I trajectories, spectrograms, continuous wavelet transforms, and Fryze decomposed active components respectively. This approach effectively realizes end-to-end learning, and mitigates the need to disaggregate load before the identification process. Experimental results suggest that the utilization of transfer learning improves the multi-label classification performance of aggregate load. This model is made accessible to consumers through a mobile application, which is used to interface with smart meter data and provide subsequent appliance usage insights. This is one of of the first works to re-purpose pre-trained deep learning networks used for image processing high frequency concurrent load classification in the context of an Advanced Metering Infrastructure (AMI).