Defect Recognition For DR Images Of Castings Based On Machine Learning

With the rapid development of China’s economy,the ability of rail transport has been greatly improved.The defect detection and quality evaluation of the railway freight trains of castings have become an important part in the production of foundry industry.Large-scale industrial digital radiography(DR)technology not only can detect the surface and internal defect information of the parts without damaging the casting performance,also achieve large-scale production of bulk inspection,which provides a powerful technical support for the quality control of castings.At present,most of the foundry manufacturers at home and abroad mainly uses the manual eye-measurement method based on the traditional X-ray films to identify the casting image defect information,which not only requires large labor intensity and holds low efficiency,but also subjects to the subjective impact of inspectors’ ability and experience.It also leads to misjudgment and omission and can’t meet the large-scale product quality real-time examination.China’s railway industry began to install large-scale linear accelerator digital ray nondestructive testing equipment for railway bolts,side frames,fork and other key components of the detection since 2007,but the defect is still determined with the artificial method.Under the support of the National Natural Science Fund Project and Chongqing Science and Technology Key Project,taking the digital X-ray inspection and assembly of cast steel parts of typical large-scale freight trains as the research object,and the results are in accordance with ASTM standards.From the point of view of machine learning,this paper realizes the automatic recognition of casting defects by constructing defective clustering classifier.The main contents and innovation of this paper mainly include the following aspects:First,the production of the standard defect maps.In order to conduct the study of the defects of railway castings easily,this paper uses the casting defective X-ray film as the object provided in the ASTM E446,and references to the prior knowledge of the classification of railway castings in the standards.It is mainly aimed at the industrial digital radiography images of bolts and side frames castings of the large-scale railway freight trains,which are divided into a large number of 64×64 pixels’ and 128×128 pixels’ size of standard defect images according to the principle of scale normalization.At last,we use the image preprocessing algorithms such as median filter and CLAHE algorithm to remove the noise and enhance the contrast of the defect areas in the standard defect images,which constitutes the final standard defect map.Second,the feature extraction and selection.In order to detect the defects of castings better,this paper takes the characteristics of blowholes,cracks and slags defects as the prior knowledge and analyzes the importance of the features with the different targets,and then takes the feature extraction technology to acquire many related features information from the defect texture,grayscale,geometry,spatial orientation and so on,and finally gets four kinds of 22-dimensional defect recognition features.Finally in order to improve the processing speed of the data and the accuracy of the model,it takes the performance of the defect clustering identifier as the evaluation index,the ReliefF algorithm and the PCA algorithm are respectively used to select and analyze the features,which realizes reducing of the redundant data.Finally,the construction of a defect clustering identifier based on machine learning clustering algorithm to identify the defects of castings.It takes the performance of the defect clustering identifier based on the improved K-Means++ algorithm as the evaluation index,the performance of the defective clustering recognizer is used as the evaluation index.The four kinds of features are respectively used to analyze the real recognition effect of the casting defects,and then carrying out the training experiment of the defect recognition.It divides the 180 standard defect samples selected from the standard defect map into the training set and the verification set with method of the "hold-out" and the stratified sampling.The PCA-ReliefF algorithm is used to train the identifier in 30 training subsets to get the optimal feature subset.At last,a test set consisting of 200 standard image sub-blocks is obtained from the 9C-type side frames’ and the K6-type bolsters’ industrial DR image and then takes the test of the defect identification of the castings in the identifier to see the generalization ability of the recognizer.The accuracy of the final defect clustering identifier reached 86.5%,which making a good recognition result.