Weixin Yuan, Kouqiang Zhang, and Haokai Wang
Inspection robots, electricity theft detection, deep learning,lightweight squeeze-and-excitation residual network (SE-ResNet)∗ School of Mechanical and Electrical Engineering, ChinaUniversity of Mining & Technology-Beijing, Beijing 100083,China; e-mail: [email protected]∗∗ School of Logistics and Management Engineering, YunnanUniversity of Finance and Economics, Kunming 650221, China;e-mail: [email protected]∗∗∗ School of Information Engineering, Zhejiang University
To improve the ability of detection robots to identify electricity theft behaviours and abnormal equipment images in complex power envi- ronments, this paper proposes a lightweight squeeze-and-excitation residual network (SE-ResNet) module and applies it to the design of a deep learning algorithm for a robot vision system. The proposed module retains the residual structure and the channel-attention mechanism while reducing computational overhead and model size through channel compression and hierarchical feature reconstruction, thereby improving adaptability to diverse inspection scenarios. A dataset covering electricity theft behaviours and abnormal equipment images was constructed. Model training and field validation were conducted in representative robot inspection scenarios. Experimental results show that the model size is reduced to 15.3 MB, corresponding to a 41.5% reduction, and the recognition speed reaches 2.86 images/s while maintaining recognition accuracy. The desktop GPU was mainly used for offline training and baseline comparison, while deployment feasibility was further evaluated through constrained inference settings with batch size fixed at 1, limited memory occupation, and single-image latency measurement. This study provides an efficient and lightweight model design scheme for visual recognition tasks in resource-constrained power inspection robots.
Important Links:
Go Back