Multi-View Complementary Chimeric Image Encoding and a Lightweight Vision Model for Appliance-Level Load Identification

Abstract

With the development of smart grids and demand-side energy-efficiency management, load identification in non-intrusive load monitoring (NILM) provides important support for intelligent electricity management. However, existing identification methods still suffer from insufficient accuracy when dealing with resistive loads with highly similar electrical characteristics and nonlinear loads operating under complex conditions. To address this issue, this paper proposes a multi-view complementary chimeric imaging method for appliance-level load identification from segmented individual-load signals. First, by exploiting the structural symmetry of the transformation matrices, the upper triangular region of the Gramian Angular Field (GAF) and the lower triangular region of the Markov Transition Field (MTF) are spatially combined, thereby preserving the global temporal geometric features while capturing the macroscopic statistical transition patterns of signal amplitudes. Second, to compensate for the loss of high-frequency details caused by the discretized binning operation in MTF, a high-resolution time-frequency matrix generated by the Synchrosqueezing Transform (SST) is introduced to provide microtexture compensation for locally redundant regions formed by repeated states, thereby constructing a two-dimensional load fingerprint representation that integrates time-domain, statistical-domain, and frequency-domain features. Finally, the constructed chimeric image-based load fingerprint is fed into a lightweight improved YOLOv8 network for classification and identification. Experimental results on the public PLAID and COOLL datasets demonstrate that the proposed method can effectively alleviate the limitations of single-representation feature expressiveness, while maintaining high identification performance in scenarios involving similar waveforms and signal disturbances, and exhibiting good applicability and identification stability under different data conditions.