| Internal defects recognition is the key link of quality inspection of titanium alloy castings,which directly affects the service performance of aircraft engines and other important equipment.Currently,aviation titanium alloy casting internal defect recognition is usually completed by artificial in the darkroom through visual.The high strength work is easy to fluctuate the staff’s physical and mental state,and there are differences in the experience and professional level of different staff,which would lead that the reliability and stability of defect identification can not be guaranteed,and even affect the safety of aircraft.Deep learning and digital X-ray image driven automatic recognition of internal defects in aerospace titanium alloy castings is a developing trend.However,the scale of defects(the area of defects)in digital X-ray images of aerospace titanium alloy castings is widely distributed,and there is small difference among some defect classes and difference within some defect classes,and the defect sample size is small,which leads to three problems in automatic defect recognition.(1)It is difficult to segment and locate multiscale defects in X-ray images.(2)It is difficult to extract the highly discriminative features of defect X-ray images and to and classify defect X-ray images.(3)The simulation-generation of defects X-ray images is difficult.In view of the above problems,the key technology for automatic recognition of internal defects in aerospace titanium alloy castings were studied from three aspects: multi-scale defects segmentation,extraction of highly discriminative features of defect images and classification of defect images,and simulation-generation of defect images.The main work and achievements are as follows.(1)How to achieve defect segmentation and localization in high-resolution X-ray images of aerospace titanium alloy castings with high accuracy when multi-scale defects coexist was studied.The efficient and lightweight attention module was embedded in the skip connections at all levels of Unet and the feature fusion of each adjacent two levels of Unet to form an attention-enhanced multi-level feature pyramid,and the ASCUnet(Ameliorated Skip Connection Unet,which improves the skip connection.Unet)model was constructed.The data augmentation strategy suitable for multi-scale defects segmentation in aerospace titanium alloy casting X-ray images was proposed.The impact of the efficient and lightweight attention module on the performance of the Unet series models in the task of this dissertation was studied.The results show that the efficient and lightweight attention module improved the model’s MIo U(Mean Intersection over Union,the average intersection ratio,used to quantitatively describe the segmentation accuracy)by 4.5%.By comparing ASCUnet with four other state-of-the-art models,the effect of model structure on segmentation performance was studied.The principle of better performance of ASCUnet in defect segmentation task dominated by small-scale defects was further revealed.The fusion of low-level high-resolution features containing spatial information has a more significant effect on the model segmentation performance than selection of the backbone network.Defects segmentation test was carried out on 185sub-images.ASCUnet achieved 90.4% MIo U and an average speed of 1.736 FPS(Frames Per Second,frames per second),and the defect segmentation and localization in the case of coexistence of multi-scale defects was realized relatively quickly and accurately.(2)How to extract the high discriminative features of X-ray images of internal defects and realize defect classification with high accuracy was studied.The feature extraction part of the 121-layer densely connected neural network was employed as the backbone,and the extracted features were used as the two inputs of the bilinear pooling layer.The BX-Net(Bilinear Xray Image Classification Neural Network)model was constructed.A data augmentation strategy for X-ray images of internal defects was proposed.BX-Net was trained by data augmentation strategy and transfer learning.The extracted features were mapped into a two-dimensional space by using the data dimension reduction methods.The two-dimensional spatial mapping distribution of features extracted by BX-Net and other eight methods were compared and studied.The results show that only the two-dimensional spatial mapping distribution of the features extracted by BX-Net agreed with the two-dimensional spatial distribution of the expected highly discriminative features,which indicated that the features extracted by BX-Net were highly discriminative.Defects classification test were performed on 100 images,BX-Net achieved the classification of images of defects with a small amount of parameters(1.22e7),high recall and accuracy(above 90%).(3)How to generate new and effective images of defects to provide more samples of defects for automatic recognition scheme was studied.The deep generative model framework was constructed to study the effect of different convolution kernel sizes and the resolution of the generated images on the visual quality of the generated defects X-ray images.The results show that when the size of the convolution kernel is 5 × 5 and the resolution of the generated images is 128 × 128 pixels,the visual quality of the generated X-ray images of defects was the highest.The deep generative models were constructed according to the above convolution kernel size and the resolution of the generated image to generate four types of X-ray images of internal defects with high visual quality that are not included in the real data set.BX-Net was used to estimate the eigen-distribution of the generated image set containing four types of defects,and to and classify the generated images of four types of defects.The eigen-distribution was the same as the eigen-distribution of real dataset,which verified the usability of the deep generative models constructed in this dissertation and the effectiveness of the generated images.(4)Based on the ASCUnet and the BX-Net fine-tuned on the generated data set,the software for automatic defect identification in X-ray images of aerospace titanium alloy castings was developed.Tests were carried out on a set of X-ray images of intake ducts with variable cross-section complex structures.Compared with the results recognized by professionals,the software finally completed the test with a missed detection rate of 2.4%,a false alarm rate of 6.7%,a recognition accuracy of 95.8%,and a processing speed of 5～6seconds per image(the GPU of the computing platform was RTX 2070 Super).The software greatly improves the efficiency and automation level of internal defect recognition of aerospace titanium alloy castings and is expected to promote the process of internal defect recognition of aerospace titanium alloy castings towards digital management. |
PDF Full Text Request