Research And System Development Of Surface Defect Detection Technology For Small Aluminum Castings Turbines Based On Machine Vision

Aluminum castings are widely used in industry because of their advantages of high strength,light weight and forming complex parts.However,appearance defects will appear in the casting and machining process,which will affect the quality of products.Aluminum casting turbines used for small engines are complex special-shaped parts with small size and often appear a variety of appearance defects on the surface.At present,artificial visual detection is still adopted,which leads to low detection efficiency,poor consistency and missing detection.Taking small aluminum casting turbines as an example,this thesis carried out research on intelligent detection methods and key technologies for surface defects of small aluminum castings based on machine vision,and developed a turbine surface defect detection system based on machine vision.The main research contents are as follows:(1)Turbine surface defect analysis and visual detection scheme design.Six kinds of defects in turbine forming and machining were studied: top blade bend,bottom blade bend and broken blade,side sand hole and surface nodules,top sand hole and surface nodules,top eccentricity and bottom sand hole.The shape characteristics and distribution law of various defects are analyzed,the design of optical detection scheme and automatic detection scheme is completed,and the hardware models such as light source,camera and electrical components are determined.Three visual detection stations are designed,which are respectively the detection stations on the side of the turbine,the detection stations on the top of the turbine and the detection stations on the bottom of the turbine.The self-writing machine software and PLC are used to jointly control the rotating table and the linear module and other moving components to realize the transmission and defect detection between the turbine stations.(2)Research on turbine surface defect detection technology based on traditional image processing algorithms,and development of defect visual detection algorithms for turbine blade top bend,blade bottom bend and broken blade,top circle eccentricity and bottom sand hole.For the top blade bend,the detection area was reduced by positioning and cutting.After the blade edge was obtained by Canny algorithm,the line segment of the top blade edge was extracted by Ramer segmentation algorithm,and a straightness detection algorithm was proposed for detection.For the bottom blade bending and blade breaking defects,the mean filtering with the core size of 11*11 was selected after the filtering comparison experiment,and the blade breaking was determined according to the area screened by the threshold value.The Hoff line detection was performed for the non-broken blade defects,and the fitting line Angle was calculated to determine whether the blade was bent.For the eccentric defects of the top circle,the gray threshold was screened to get the area to be examined,and the least square method was selected to get the center of the circle.The Euclidean distance between the two centers was calculated to judge the eccentric defects of the circle.For the bottom sand hole defect,the gray threshold is used to screen out the region after noise removal by means of mean filtering,and the kernel with radius of 1000 is used for open operation and image difference to further obtain the region of interest.The existence of sand hole is determined by the pixel area of the region after gray threshold screening.(3)In view of the small target defects such as sand holes and surface nodules on turbine surface,which are characterized by various forms,random distribution and weak comparison with background,the deep learning detection algorithm based on the addition of small sample data is carried out.Aiming at the problem of defect data set being small samples,the Mosaic data enhancement method is optimized to avoid fuzzy defects of small targets.The mainstream target detection algorithm is analyzed and experiments are carried out.The YOLOv5 detection network is selected for the target detection in this thesis.In view of the small target defects of sand holes and surface nodules on turbines,the k-means ++ clustering algorithm was used to generate appropriate anchor frame size for the defect data set,and the relevant parameters of the detection network were compared to determine the depth and width of YOLOv5 and the input image resolution of 960*960.The network structure was improved,and the three attention mechanism modules were compared.The detection accuracy was increased to 95.2% by adding coordinate attention mechanism,and then a small target detection branch was added to the network.After overall optimization and improvement,the detection accuracy was increased to 97.8%.C++ language was used to realize the deployment of the model in the system.After comparing five deployment modes,Tensor RT was selected to conduct model deployment reasoning.The reasoning speed reached 4ms/ page under laboratory conditions.(4)The turbine defect detection software was developed using MFC interface framework based on C++ language on Window10 platform,and the turbine defect detection system was tested on site.A total of 1000 products,300 defective products and 200 qualified products were randomly selected from five batches to test the stability and detection accuracy of the system.Finally,the comprehensive defect detection rate is 98.2%,the false detection rate is 3.8%,and the average detection time of a single turbine is 5-6 seconds,which meets the requirements for the appearance defect detection of small aluminum castings turbines in this thesis.