Abstract: In the smart grid, precise monitoring of the operational status of critical equipment for transmission, distribution and power supply is essential for effective online maintenance. Faced with the inefficiencies of manual recording and inspection, as well as the challenges associated with the complex installation, high cost and lengthy periods required for digital upgrades of monitoring devices, a novel approach that integrates image capture devices with image processing technology has been developed. This approach, leveraging the allocation of computational resources for task distribution, introduces a Light-Resnet-based numerical recognition algorithm, which enhances network training through the optimization of the D-Add loss function, enabling remote reading of electrical equipment monitoring data. Experiments have demonstrated that Light-Resnet achieves a rigorous accuracy rate of 98.8% on the MNIST dataset with only 6090 parameters. When combined with edge computing collaboration mechanisms, it resulted in a 20.73% reduction in power consumption on the terminal side. The proposed algorithm not only proves its adaptability and efficiency in resource-constrained environments but also significantly improves the network's accuracy with design of the D-Add loss function.