With the rapid development of robotics technology, intelligent robots are widely used in substation inspections. Aiming at the problem that the current target detection algorithms have too many parameters and the performance of embedded devices is limited. It is difficult to achieve real-time detection on the embedded platform. A n improved YOLOv4 embedded substation instrument detection algorithm is proposed. The algorithm is based on YOLOv4 and uses MobileNetV3 as the backbone feature extraction network. It reduces the amount of calculation and increases the detection speed while ensuring that the model can effectively extract features. At the same time, the convolution operation in the path aggregation network (PANet) is replaced with a depthwise separable convolution after feature extraction; the training strategy of transfer learning is used to overcome the difficult problem of model training. Finally, the improved model is optimized by TensorRT to achieve fast and efficient deployment reasoning. The improved algorithm is tested on the embedded NVIDIA Jetson Nano, and the experimental results show that the detection speed is increased by 2 times to 15 FPS at the expense of less accuracy. This provides the possibility for real-time instrument detection in edge computing scenarios.