针对钢桥螺栓脱落病害人工巡检效率低和智能化检测样本数据集不足的问题,提出一种基于数据深度增强的钢桥螺栓脱落智能检测方法.该方法首先以采集的螺栓图像数据集为基础,利用深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks,DCGAN)对有限的螺栓图像进行增广;然后将生成的图像与原始图像合并构建增广后的数据集;再将数据集输入到单阶段目标检测网络YOLO(You Only Look Once,YOLO)中,结合迁移学习方法进行模型训练,并对训练后模型的性能进行验证;最后,进行螺栓脱落病害识别.为验证该方法的可行性,对螺栓脱落检测模型进行试验验证,并对不同采集环境下的某钢桥拼接板螺栓脱落病害进行检测.结果表明:DCGAN可有效生成逼真的螺栓图像,且与常规增广方式相比,DCGAN生成的图像质量更高、性能最优;检测模型受拍摄距离、角度及光照强弱影响且对角度最为敏感,控制拍摄距离在1.6 m内、拍摄角度在20°内、外界光线明亮可保证模型性能较优;与常规增广后训练的模型相比,利用生成图像增广后训练的模型检测性能更优且鲁棒性更好;该智能检测方法可以用于螺栓脱落病害检测,且检测精度明显提高.
An Intelligent Bolt Shedding Detection Method for Steel Bridges Based on Depth Data Enhancement
Bolt shedding is one of the typical deteriorations of steel bridges,which can be manually or intelligently detected,however the two ways of detection display intrinsic limitations,specifically the low efficiency with the former and insufficient datasets of detection samples with the latter.This paper presents an intelligent bolt shedding detection method based on depth data enhancement.In this method,the datasets of collected bolt images are augmented by using the Deep Convolutional Generative Adversarial Networks(DCGAN),afterwards the generated images are combined with the original images to construct augmented datasets,the latter are then input into YOLO(You Only Look Once),together with the transfer learning method to conduct model training and verify the performance of the trained models,and the final step is to identify the slack bolts.To verify the applicability of the presented method,specimens were prepared to verify the numerical models.The bolts in splices of an existing steel bridge were detected and images were collected in different environments.It is shown that DCGAN can effectively generate vivid bolt images,compared with the conventional augmentation method,and the images generated in DAGAN are of better quality.The sensitive factors to the detection model include the shooting distance(controlled ≤1.6 m),shooting angle(the most sensitive factor,controlled ≤20°)and lights(as bright as possible).Compared with the trained models with conventional augmentation,the trained model augmented with generated images exhibit better detection effect,in terms of performance,robustness and detection accuracy.
万方数据
维普
