Research On Detection Algorithm Of Nuclear Fuel Pellets Surface Defects Based On Dual-Modal Fusion

Nuclear energy is a high-efficiency and low-carbon energy,which can alleviate global energy shortages and climate change under reasonable application.Nuclear power is one of the most important usage of nuclear energy.At present,the development of nuclear power is in an important period of opportunities,but it is also facing challenges.Security is one of the most notable issues.Nuclear fuel pellets are not only the core components of nuclear reactors but also the first crucial barrier to ensure the safe operation of nuclear reactors.Therefore,controlling the surface quality of nuclear fuel pellets strictly and getting rid of unqualified pellets timely are significant guarantees to minimize nuclear accidents and realize the safe use of nuclear energy.Currently,single-modal information based on 2D grayscale images or 3D point cloud data is widely used in the industry for surface defect detection.However,since a single modality cannot fully describe objects,the detection results are prone to miss and false.To improve the reliability and speed of online defect detection,this thesis first converted the distance value of the point cloud data into the gray value of the image and generated it into a 2D depth image to reduce redundant data.Then,a lightweight deep learning network architecture was used to fuse the plane grayscale and spatial depth of the pellets.By fusing complementary information,high-precision and intelligent real-time detection of three typical defects on the fuel pellet surface can be realized.The proposed research can also be broadly applied to the quality inspection of other industrial products.The main contents of this thesis are as follows:(1)Establish a dual-modal defect dataset.First,a 3D snapshot sensor was used to obtain a pseudo-color image of the fuel pellet surface,and decomposed the depth image(red channel)representing the spatial position and the grayscale image(green channel)reflecting the surface brightness.After that,using an annotation tool(Semantic Segmentation Editor)to gain label images and established a dual-modal defective pellet dataset which provided the data basis for the subsequent research.(2)Verify the feasibility of the detection algorithm of nuclear fuel pellets surface defects based on dual-modal fusion.First,to avoid network overfitting due to the consistent features in the industrial image,a simplified model based on a single-modal segmentation network Seg Net was built combining depth images with gray images.Afterwards,a dual-modal fusion defect detection network named SFuse Net was established.Experiments showed that the method of dual-modal fusion could improve segmentation accuracy and was suitable for small industrial datasets.However,the fusion model consumed more storage resources and computing resources.(3)Design a lightweight detection network based on dual-modal fusion.A lightweight model can make the algorithm more flexible and efficient.In the thesis,a lightweight single-modal ENet was used as the branch of the lightweight dual-modal fusion detection network HAEFuse Net.To improve the feature reconstruction of the network,replaced the upsampling residual structure in the original ENet network with a hybrid dilated convolution structure.Furthermore,a dual-feature fusion module based on the spatial attention mechanism was proposed,which integrated deep branch and gray branch,as well as the shallow features and deep features of the network,making the segmentation of small targets and edges more effective.Experiments showed that the average intersection of union for the proposed HAEFuse Net was 80.53%,which was better than Sfuse Net.In addition,the number of parameters of HAEFuse Net was7 times less than that of SFuse Net,and the segmentation speed was 0.4 times faster,reaching 0.053 s/time.(4)Set up a defect detection system.The pre-trained HAEFuse Net detection model was deployed to build the software platform and we also developed a software visual interface to simulate real-time online detection.Based on the engineering requirements of the defective area ratio,the upper limit value of the rejection was set to be 10% to classify the pellets and test the performance of the system.Experiments showed that the classification accuracy of the system reached 99.49%,and the speed on med-to-high end equipment was no more than 0.3 s/time,which satisfied the accuracy and efficiency requirements of industrial real-time online inspection.