基于卷积神经网络的轮对超声波探头倾斜检测

Wheelset ultrasonic probe tilt detection based on convolutional neural network

在使用球机检测车底轮对超声波探头时,摄像头角度会导致目标探头出现仿射形变;此外,目标探头的密集排列会导致检测结果不准确。针对以上两点问题,提出了一种基于卷积神经网络(CNN)的旋转探头检测方法。提出了一种结合多尺度特征融合和注意力机制的特征提取网络和一种针对形变物体的表示法。在验证过程中,使用实际条件下球机采集的轮对探头数据对改进后的深度卷积神经网络(DCNN)进行训练与测试。测试实验结果显示,提出的检测方法对密集的以及形变的探头目标均具有良好的检测效果,召回率可以达到92.15%,平均准确率可以达到86.39%。实验结果表明,所提的改进方法能够自动全面地提取探头目标特征,解决了仿射形变和密集探头目标的检测问题;而且检测精度和速度均能够满足实际需要,在不同的目标尺度以及模糊情况下,具有更强的适应性和更高的鲁棒性。

In order to solve the problem of affine deformation caused by camera angle and inaccuracy caused by dense array of probes when using ball head camera to detect the ultrasonic probe tilt of underbody wheelset,a detection method of rotating probe based on Convolutional Neural Network(CNN)was proposed. A feature extraction network was constructed by combining multi-scale feature fusion and attention mechanism,and a new six-parameter representation for deformable objects was proposed. The proposed network was trained and tested by wheelset probe data which were collected by ball head camera under actual conditions. In the experiment,the proposed detection method had a good detection effect on both dense and deformed probe targets,the recall reached 92. 15%,and the average accuracy reached 86. 39%. Experimental results show that the improved network can automatically and comprehensively extract the probe target features,solve the problem of affine deformation and dense probe target detection. The detection accuracy and speed can meet the practical needs,and it has strong adaptability and robustness under different target scales and fuzzy conditions.