Multi-device identification with V-I trajectory for non-intrusive load monitoring based on multi-scale feature fusion

The traditional methods for multi-device recognition are unable to detect devices that are not included in the training set. Furthermore, these previous methods assume that all data conforms to the same probability distribution, disregarding the variations in probability distribution among different categories of data. Additionally, when it comes to feature extraction, conventional approaches primarily focus on capturing semantic information from images and fail to extract spatial features necessary for observing global V-I trajectories. In our proposal, we use the feature extraction method of attention mechanism and channel attention convolution fusion to extract global spatial features and local semantic features. In addition, this paper considers that each type of V-I trajectory follows a different probability distribution, and uses the Variational Autoencoder (VAE) to generate the probability distribution. Then, according to the strategy of clustering, the probability distibution of the same class is gathered to assist the detection of unknown devices. The experimental results on the public data PLAID show that the average accuracy is improved by about 7%, and the average F1-score is improved by about 5.5%.