Lightweight Load Identification Based on Fusion of Electrical Signal Image Features

Load identification is a fundamental aspect of non-intrusive load monitoring (NILM), which seeks to discern the operational condition of individual electrical devices from aggregated power signals. In recent years, deep learning-based image classification techniques have demonstrated considerable promise in enhancing the precision and efficacy of load identification tasks. This research presents a unique strategy that combines electrical-signal image encoding with lightweight network optimization to improve image-driven recognition performance. Voltage, total current, and reactive current signals are initially normalized to create $U-I$ greyscale maps, while the periodic active power is represented using the Gramian Angular Field (GAF) approach. The features are subsequently integrated into a multimodal RGB image, offering abundant and complementary information for categorization. MobileNetV3 serves as the backbone network because of its equilibrium between performance and efficiency, while the Efficient Channel Attention (ECA) mechanism is integrated into bottleneck layers to improve feature discrimination. The classification head is reconfigured by removing superfluous upsampling layers and implementing pruning techniques, which substantially decrease model parameters and computing demands while maintaining recognition accuracy. Experimental results confirm the method’s advantages in both performance and efficiency.