| Exploring wear status detection methods is of great significance to the intelligent operation and health management of mechanical equipment.At present,image processing technology is commonly used to achieve wear localization.However,limited by factors such as light and exposure,it is difficult to achieve batch wear positioning of wear images.On the other hand,high-resolution imaging devices such as optical microscopes and electron microscopes are used to determine the wear mechanism,but their discrimination relies on the subjective experience of the observer and the detection efficiency is low.In recent years,artificial intelligence and deep learning technologies have developed rapidly,and deep learning-based target detection methods with excellent feature extraction capabilities have provided new ideas for wear detection research.In this paper,an image wear detection method based on an improved Faster R-CNN(Region-based Convolutional Neural Network)network is proposed to achieve macroscopic wear localization and microscopic wear mechanism identification of wear images.To address the problems of difficult extraction of deep wear features and the difficulty of accurate localization of small-sized wear marks,the feature extraction structure based on Res Net101 with fusion structure of deep and shallow features is used to improve the classical Faster R-CNN network.Deeper and more abstract wear semantic features are mined by enhancing the network depth.And the detection capability of the network for different classes and scales of wear is improved by fusing deep features with shallow features.The target detection performance of the improved Faster R-CNN network is tested using public datasets.The results show that compared with the classical Faster R-CNN network,the improved Faster R-CNN network has better detection ability for different classes and scales,especially for small-scale targets.To verify the wear localization performance of the improved Faster R-CNN network,a reciprocating friction wear test device with a wear image acquisition system was built independently,and continuous wear image acquisition was realized.The improved Faster RCNN network is trained and tested using this dataset.The results show that the network can accurately locate abrasions at different scales with higher accuracy compared to image processing methods.The results show that the network can accurately locate the wear marks at different scales with higher accuracy than the image processing method.The error between the network and the automatically measured wear mark width is within 2% as measured by optical microscope.On this basis,the wear localization performance of the network was further verified by using the network to continuously measure the wear width to obtain the wear width variation curve and compare it with the friction coefficient variation curve.To verify the wear mechanism recognition capability of the improved Faster R-CNN network,a wear image dataset based on the classification of different wear mechanisms was produced.The improved Faster R-CNN network is trained and tested using it.The results show that the network can achieve an accuracy of over 98% for recognizing different wear mechanisms of wear images,which has a very strong ability to detect wear mechanisms.At the same time,the network still shows excellent wear mechanism detection ability for the coexistence of multiple wear mechanisms.Further,using the activation-like thermal feature visualization technique,it is demonstrated that the improved Faster R-CNN network achieves wear mechanism identification by focusing on the wear feature analysis inside the wear region.This both explains the rationality of this wear mechanism identification method and validates the superior performance of the wear mechanism identification method based on the improved Faster R-CNN network.Based on the idea of small-sample migration learning,the target detection network trained based on laboratory linear wear marks is migrated to the actual wear detection in engineering,and the Mask branch is added on the basis of the improved Faster R-CNN network so that the network can segment the wear marks of different sizes and shapes by contour and improve the accuracy of localization.It is also verified that the improved Mask R-CNN network based on migration learning still has a strong ability to locate wear and identify the wear mechanism using industrial typical wear images. |
