基于Faster R-CNN的齿轮箱铁谱磨粒识别

在利用铁谱技术对齿轮箱磨损状态进行评估时,存在磨粒特征提取困难且磨粒识别的数量有限的问题,基于铁谱磨粒图像特性,提出基于Faster R-CNN算法的铁谱磨粒识别。该算法采用ResNet-34网络完成铁谱磨粒特征自动提取,并采用RPN网络实现对图像中多个磨粒的识别。通过实验对Batch_Size和学习率超参数进行优化,使用超参数优化后的Faster R-CNN算法进行实验。结果表明:该方法克服了磨粒交叉引起的识别难点,能识别一副图像中的多个磨粒,能统计各类磨粒数量,且准确率较高;在磨粒背景颜色不同、存在噪声干扰等情况下,该方法仍能够准确判断磨粒类型,具有较好的稳定性。

Ferrography Wear Particle Recognition of Gearbox Based on Faster R-CNN

When using ferrography to evaluate the wear state of gear box,it is difficult to extract the characteristics of wear particles in gearbox oil tank,and the number of wear particles identification is limited.In view of the above problems,based on the image characteristics of ferrography,the identification of abrasive particles with ferrography based on the Faster R-CNN algorithm was proposed.The algorithm adopts ResNet-34 network to complete the automatic extraction of ferrography abrasive grain features,and adopts RPN network to realize the recognition of multiple abrasive grains in the image.The Batch_Size and learning rate hyperparameters were optimized through experiments,and the experiments were carried out by using the Faster R-CNN algorithm after hyperparameter optimization.The results show that this method overcomes the difficulties of identification caused by abrasive grain intersection,which can identify multiple grinding particles in the image,and count the number of all kinds of grinding grain with high accuracy.Even in the case of different background color of the abrasive grain and noise interference,this method still can accurately judge the type of abrasive grains and has good stability.