In response to the problems of extensive number of parameters as well as the intricate model complexity prevalent in traditional deep learning detection algorithms, and taking the faults such as broken dials, blurred dials, and broken housings of substation instrumentation equipment as the research object, this paper proposes a lightweight substation instrumentation equipment defect detection method based on YOLOv4-Mobilenetv2-S. Firstly, the Mobilenetv2 backbone network is used to replace the CSPDarknet53 in the YOLOv4 backbone feature extraction network to decrease the quantity of parameters of the model; then, the SE attention mechanism module is added to the three preliminary effective feature layers of the backbone extraction network and the PANet to scale up the information extraction capability of the feature layer; lastly, all the $3 \times 3$ standard convolutions in the PANet and the $5 \times 5$ convolutions in PANet are adjusted to depth-separable convolutions to streamline the model and computation. The practical outcomes demonstrate that compared with the original YOLOv4 network, the algorithm reduces the amount of parameters of the network by 83.1% and compresses the model by 83.7% under the premise of guaranteeing the accuracy, and the average recognition time of the image reaches 13.14ms with an accuracy of 85.99%, which can address the requirements for detecting defects in substation instrumentation equipment.