Few-Shot Learning for Image-Based Nonintrusive Appliance Signal Recognition

Abstract

In this article, we present the recognition of nonintrusive disaggregated appliance signals through a reduced dataset computer vision deep learning approach. Deep learning data requirements are costly in terms of acquisition time, storage memory requirements, computation time, and dynamic memory usage. We develop our recognition strategy on Siamese and prototypical reduced data few-shot classification algorithms. Siamese networks address the 1-shot recognition well. Appliance activation periods vary considerably, and this can result in imbalance in the number of appliance-specific generated signal images. Prototypical networks address the problem of data imbalance in training. By first carrying out a similarity test on the entire dataset, we establish the quality of our data before input into the deep learning algorithms. The results give acceptable performance and show the promise of few-shot learning in recognizing appliances in the nonintrusive load-monitoring scheme for very limited data samples.