Automatic Detection Of Casting Defects Based On Convolutional Neural Network

Social and economic development has entered a stage of rapid development,and casting products are widely used in many industrial fields.The quality of the cast product is directly related to the immediate interests of the user.Therefore,the cast product must be inspected strictly in accordance with national standards before leaving the factory,and regular quality inspections are required during use.At present,the identification of casting defects in images formed by X-ray inspection still depends on manual operations.The detection system is inefficient and the detection accuracy is poor.In recent years,the tide of artificial intelligence has swept the world,and Deep Learning has become a major trend in the research direction of universities and research institutes.In the process of image recognition research,convolutional neural networks are the most prominent,making it possible to automatic detection of casting defects.In this thesis,the X-ray image of a casting is taken as the research object,and the automatic detection of casting defects is studied and realized.The thesis mainly studies in the following aspects:1.A research on Deep Learning methods based on X-ray casting inspection is proposed.Capture model images of castings required for model training,including defective and non-defective categories.Through a series of operations,the number of two types of samples is expanded to establish a sample database for casting inspection.2.The convolutional neural network Alexnet model was selected and the convolutional neural network was built on TensorFlow.Complete the training of the data set and verify the correctness of the network.3.The influence of the size of the convolution kernel and the complexity of the network structure on the performance and time cost of the model is studied.The improved form of Alexnet is proposed and tested.Experiments show that the improved network improves recognition accuracy and reduces training time.Tests have shown that the detection system has a recognition rate of 98%.