Image Formation, Deep Learning, and Physical Implication of Multiple Time-Series One-Dimensional Signals: Method and Application

Abstract

Time-series 1-D signals are ubiquitous in industrial applications for monitoring and control. However, it is lacking of efficient tools to deal with simultaneously multiple time-series 1-D signals. To this end, in this article, a novel theory of image formation is proposed that converts multiple 1-D signals to 2-D images and takes advantages of convolutional neural network for feature extraction and classification of a sequence of images. A case study is carried out for the classification of working conditions in photovoltaic power systems. In total, 23 1-D signals are mapped to a sequence of 2-D images to derive six different models through image formation-based deep learning. They are tested through the outdoor experiments under time varying working conditions. We discover that physical implication in 2-D images affects significantly the classification performance such that 2-D images with the clustered currents or voltages tend to create better results while randomly arranged image patterns are prone to generate worse results. Excellent performance with an accuracy 96.09% is guaranteed when physical advantages are incorporated in the proposed tools. Driven by the deep learning approaches, the proposed tools are promising for complicated industrial applications.