Research On Intelligent Detection Method Of Internal Defects Of Casting Based On Ultrasonic Image

With the wide application of castings in mechanical products,it is recently found that the internal defects of castings seriously affect the quality of products,which even can cause unpredictable losses.Thus,the quality detection of castings has been a key part in industrial production.As a kind of nondestructive testing method,ultrasonic testing has several special advantages,including simple operation,high sensitivity and no pollution to human and environment.Given that most of ultrasonic fault detectors can directly show the scanning image of defects,it is feasible and necessary to directly detect the internal defects of castings by image recognition.Deep learning technology has made remarkable achievements in the field of image recognition,and the application of convolutional neural network in defect recognition has also attracted lots of scholars' attention.Since there are some problems in internal defect detection method,such as slow speed in detection and low accuracy of defect identification,it makes a research on the intelligent detection method of internal defects,which is based on ultrasonic image.Firstly,with the help of A-type pulse reflecting ultrasonic fault detector,it tries to obtain the image of internal defects.Due to the limitation of the equipment,it causes the low contrast ratio and complex background noise on collected original images.Therefore,it uses digital image processing technology to deal with the original images,which can remove the pseudo noise and improve the contrast ratio of images.Then,the accuracy of defect identification will be improved.In order to solve the problem of small samples,it adopts the image enhancement technique to expand the number of samples and avoid over-fitting phenomenon.Secondly,it tries to set up the basic convolutional neural network through MATLAB simulation software,and the preprocessed image is taken as the input of the model,so as to realize network training.With the help of test set,it tries to test the simulation experiment,and takes the accuracy,loss ratio and detection time of defect identification as the performance indicators of the evaluation model.Finally,it tries to improve the convolutional neural network model from two perspectives,namely,structure and function.In terms of structure,it plans to increase perceptual structure.Through different sizes of weight parameters,it extracts various characteristic values,so as to avoid the shortage of expressiveness in single feature image and improve the accuracy of defect identification.As for function,it introduces sparse constrained functions to let network weight tend to be sparse and establish light-weight convolutional neural network.Then,the trained network will have certain sparsity and pretty good generalization ability,which will strengthen the running speed of network model.According to experimental results,it is found that the improved network model has better recognition effect when it is compared with other existing excellent detection methods.Meanwhile,the feasibility and effectiveness of the improved convolutional neural network model is also verified.