The real-time detection of defects in nuclear power pipeline thermal insulation glass fiber by deep-learning

Glass fiber, prized for its high-temperature thermal insulation and radiation resistance, serves as a crucial material for insulating nuclear power pipelines. However, the harsh operational conditions often lead to material defects, underscoring the importance of defect detection for energy efficiency and personnel safety, and manually segmenting and classifying defects can be time-consuming and increase risks. Hence, there is a pressing need for a real-time and accurate detection method. In this work, infrared images of nuclear power pipeline thermal insulation glass fiber defects were collected to establish the dataset, and the damage mechanisms were analyzed. Besides, various prevalent object detection models were tested and found that YOLOv8n exhibited significant potential for improvement with exceptional speed performance and detection accuracy. Through integrated EMA attention blocks, incorporating the FasterNet blocks into the backbone, retrofitting the neck layers with the slim-neck structure, and implementing DyHead in the YOLOv8n's head, our improved model achieves the highest values of mean Average Precision (mAP) scores with 0.5:0.95 intersection over union (IoU) of 57.6 %, and 0.5 IoU of 86.8 %, while maintaining the original high detection speed and low number of parameters, ensures suitability for real-time detection deployment on edge devices of nuclear power plants.