Wear Particle Analysis Based On Convolutional Neural Network And Image Saliency

Ferrography can reflect the wear state of equipment and diagnose faults by analyzing the type,quantity,concentration,and other characteristics of wear particles.The process of wear particle analysis based on computer image processing technology is a one-way linear processing: wear particle segmentation,feature extraction,classification and recognition.In order to improve the automation level of ferrograph image classification and wear particle detection,the convolution neural network(CNN)technology is introduced into ferrograph image analysis to overcome the disadvantages of its wear particle segmentation difficulty,manual design feature method involving many links,cumbersome and inefficient,error accumulation and transmission,and algorithm difficulty in overall optimization.Based on the image saliency of different types of wear particles and the flow from overall judgment to detail recognition,the CNN model of wear particle image classification and the CNN model of wear particle detection are designed,constructed and optimized.The intelligent analysis of wear particle image is realized by using the characteristic of CNN model such as self-learning,deep abstraction and end-to-end processing.Wear-Net,a CNN model for wear particle image classification,is designed and constructed.Firstly,based on actual wear particle images,the data set for image classification is created by image augmentation techniques such as color transformation and image jitter.Secondly,AlexNet and VGG16 are reproduced and compared on the data set of wear particles.Wear-Net is constructed by optimizing the structure and parameters.The model consists of four convolution layers and the number of convolution kernels is 64,128,256 and 128,respectively.Five kinds of wear particle images are classified.Finally,the visual analysis of the network convolution layer and the fully-connected layer is carried out,and the Wear-Net classification results are compared with the linear flow method based on image processing.It can be seen from the experimental results that the average accuracy of the Wear-Net model on the test set is 98.75%,which can be used to classify the actual wear particle images.Wear-SSD,a CNN model for wear particle detection,is designed and constructed.Firstly,the data set for wear particle detection is created.Then,based on the SSD model,the Wear-SSD model is constructed,which can detect three kinds of abnormal wear particles,by optimizing and improving the size of input image,design of default box,transfer learning method and etc.This model combines the advantages of two kinds of target detection CNN models based on region proposals and bounding box regression.It can detect wear particles in small size and complex background.The accuracy and recall rate of the test set are all above 85%.In addition,the model can also estimate concentration values of abnormal wear particles,extending the model's function.The experimental results show that the Wear-Net model and Wear-SSD model constructed and trained on the Caffe are characterized by simple,end-to-end processing,generalization ability and high accuracy.They have certain reference value for the practical application of intelligent ferrograph analysis.