To address the existing deep neural network models for insulator defect detection are large in size, high in computational resource consumption, low in detection accuracy and difficult to deploy at the edge end. In this paper, we propose a lightweight insulator defect detection model ACAM-YOLOv5s with asymmetric convolution and attention mechanism based on channel pruning and YOLOv5s method. The ACAM-YOLOv5s model uses the asymmetric convolution module ACBlock to replace the standard convolution in the residual structure of the YOLOv5s backbone network, combined with the attention CBAM of channel and spatial blending for feature fusion to enhance the expressiveness, feature extraction and robustness of the backbone network. PIoU, which is highly sensitive to the size and position of the bounding box, was introduced as a localisation regression loss to address the problem of low defect detection localisation accuracy due to high insulator aspect ratios. The experimental results show that the pruned ACAM-YOLOv5s model has relative advantages over the original YOLOv5s in terms of detection accuracy, computational volume and model size, which can meet the needs of edge device deployment and has potential value in the field of UAV aerial insulator defect detection.