基于深度学习的正交异性钢桥面板疲劳裂纹智能识别方法研究

Intelligent Fatigue Crack Identification Methods for Orthotropic Steel Bridge Deck Based on Deep Learning

为提高正交异性钢桥面板疲劳裂纹的识别效果,基于深度学习方法,进行正交异性钢桥面板疲劳裂纹识别方法研究。制作疲劳裂纹试件,通过室内检测采集检测结果图像建立数据集,结合图像预处理改进数据集中图像;基于单一回归目标检测(You Only Look Once,YOLO)系列中的YOLOv5s图像识别算法,引入轻量级卷积注意力机制(Convolutional Block Attention Module,CBAM),建立疲劳裂纹超声相控阵检测图像的智能识别方法;将该方法应用于超声波相控阵检测结果判读。结果表明:结合限制性对比度直方图均衡(Contrast Limited Adaptive Histogram Equalization,CLAHE)算法与自适应中值滤波的图片预处理方法较为适合超声相控阵检测图像的图片预处理;通过引入CBAM改进图像识别算法,图像识别的精确度相较改进前提升1.28%,达到98.74%;建立的YOLOv5s-CBAM图像识别算法能够准确地框选疲劳裂纹区域,对四类常见疲劳裂纹类型的识别置信度达到0.91,对其它裂纹类型的识别置信度达到0.85,在室内环境下能够满足疲劳裂纹的智能检测需求。

This study presents a method to identify fatigue cracks in the orthotropic steel deck panels.The method is developed on the basis of deep learning and can improve the fatigue crack recognition efficiency.Specimens were prepared for indoor fatigue crack detection and image collection to create data sets,and the data sets were in turn processed by the image pre-processing technique.Combined with YOLOv5s(an image detection algorithm)and CBAM(Convolutional Block Attention Module)attention mechanism,a method to intelligently recognize ultrasonic phased array detection images is developed to complete the interpretation of the test results.It is shown that the image pre-processing method combining the CLAHE(Contrast Limited Adaptive Histogram Equalization)and adaptive median filtering suits well the pre-processing of ultrasonic phased arrays.Through introducing the CBAM attention mechanism,the image recognition accuracy is improved by 1.28%,reaching 98.74%.The presented YOLOv5s-CBAM image recognition algorithm can accurately frame and select the fatigue cracking zones,and the recognition confidence of the four types of common fatigue cracks reaches up to 0.91,and 0.85 for other types of fatigue cracks,proving that it can meet the intelligent detection requirements of fatigue cracks in indoor environments.