Ferrography Image Classification Based On Deep Learning

With the development of science and technology,modern industry is developing rapidly in the direction of intelligence and automation.Fault diagnosis technology is particularly important in modern manufacturing industry.Ferrography analysis technology can determine the site of wear and failure in the equipment.The morphology and distribution of wear particle in ferrography image can reflect the wear status of equipment.The analysis of ferrography image based on computer technology is an important way to improve the automation level of oil analysis and popularize ferrographic analysis techniques.Ferrography image analysis methods based on artificially designed wear particle parameters and machine learning algorithms for wear particle identification have the disadvantages of complicated processes and poor universality.Deep learning in computer vision has been widely used,the advantages of this method include high accuracy,wide range of applications,the process is simple and so on.Therefore,in this paper,deep learning technology is applied to the analysis of ferrography image.The convolution neural network model is used to classify the ferrography image to determine the wear status of the equipment.The main contents of this paper are as follows:Aiming at the problem of traditional wear particle recognition method,this paper proposes a ferrographic image classification method based on deep learning.Firstly,the ferrography image are divided into seven categories according to the distribution of the wear particle and the morphology of the wear particle in the ferrography image.Spatial geometric transformation of the image and get the appropriate data by crop the subgraph from image.Adopting spatial filtering,image warping and color transformation to augment the data properly.Label the ferrography image data to obtain the training dataset,verification dataset and test dataset required for deep learning.Research convolution neural network and its optimization methods.Convolution neural network was constructed based on Caffe.When training network model,the loss value and classification accuracy were recorded,and the model training process and model performance are analyzed accordingly.In this paper,the structure of the network and the parameters and training parameters in the network are experimentally determined and described in detail in the article.Compare the accuracy of the model in the validation data sets with different convolution kernel size and number of convolution layers.The performance comparison between the model and the classic convolutional neural network model AlexNet is carried out.The network model is tested through examples.The results show that the model results of the classification of ferrography images are consistent with the results of artificial classification.The output feature map of convolution layer is extracted and visualized.The process of convolutional neural network acquiring feature parameters with distinguishing ability through layer-bylayer feature extraction and dimension reduction processing of the original image through the network layer is analyzed.Through the dimensionality reduction and visualization of the output data of the full connection layer,the distribution of the data in the high dimensional space is also analyzed.The example test and network model analysis verify the accuracy and validity of the proposed classification model of ferrography images.