Research And Implementation Of Deep Learning Method For Defect Detection Of Castings Based On X-Ray

Precision castings have a high utilization rate in industrial production,and their quality problems directly affect the safety performance of industrial equipment.Defective castings can even result in catastrophic consequences.Therefore,the quality inspection of castings is a priority in the production of castings.At present,most of the defects detection of precision castings still rely on manual operation.The recognition accuracy is influenced by personal experience,and the visual fatigue produced by long time work still has a great influence on the detection results.In addition there is a problem of low detection efficiency.In recent years,deep learning has become a research hotspot in various circles.Among the tasks of image recognition,the convolutional neural network has performed most prominently.In this paper,according to the testing requirements of the casting factory,this paper proposes an intelligent detection method for casting defects based on deep learning.For this method,the main work has been done in the following areas:Firstly,this paper uses the existing X-ray inspection platform to collect two types of images.They are images of castings with defective and normal castings.The images are treat as classification targets.In order to expand the number of two types of samples through a series of operations such as rotation,cutting,and contrast enhancement.Then establish a sample database for casting inspections.Secondly,for the method of casting detection based on deep learning,complete the configuration of its software environment.Construct a Caffe deep learning framework on a GPU-equipped server and complete the construction of a convolutional neural network for castings inspection.Thirdly,for the ray image database of castings in this paper,the convolutional neural network is trained,and better experimental results are obtained through parameter adjustment.Fourthly,for the time cost problem in industrial applications,thispaper studies the influence of the size of the convolution kernel and the complexity of the network structure on the model performance and time cost.Propose an improved form of caffenet and carry out experiments.Experiments show that the improved network improves the recognition accuracy and reduces training time.Finally,configure the deep learning software environment of the deep learning embedded development board-Jetson TX1.Transplant the trained model to the Jetson TX1,and perform the test experiments on the Jetson TX1.The test results are ideal and the feasibility of the development board for defect identification is verified.After completing the transplantation of model,test the Jetson TX1 development board.The accuracy rate of the castings recognition is 97%.Among them,the identification of the defective castings is up to 100%accuracy.Therefore,there is no missing phenomenon.The accuracy of casting without defect identification is up to 95.5%.