Defects Detection Methods For Aeroengine Components

Ceramic-based composite materials are currently recognized internationally as the basic materials of core components in aeroengines of next-generation.Defects detection results are an important basis for qualifying components.Composite materials have nature of heterogeneity and porosity,thus the CT(Computed Tomography)detection images has complex features,contains huge number of suspected defects,make it difficult to identify defects without spatial information.Limited by economic cost,the absolute number of defective samples is small,and proportion of defects and non-defects are severely uneven.To this end,we studied deep learning methods for components defects detection,which can be divided into two categories: anomaly detection and defect segmentation.Using deep anomaly detection algorithms can take advantage of the large amount of nondefective samples in CT slices,which can effectively alleviate the problems of small number of defective samples and unbalanced proportion between defects and non-defects.In image anomaly detection,most anomaly detection algorithms usually model normal samples in train stage and use reconstruction errors to judge anomaly in test stage.However,existing image generative models such as VAE(Variational Auto Encoder)and GAN(Generative Adversarial Networks)have shortages when applied to anomaly detection.For VAE,the main problem is the low quality of generated samples.For GAN,the main problem is the difficulty in reconstructing the input.So we combines the structure of VAE and GAN.By adding an encoder to the generative adversarial network,GAN can reconstruct the inputs directly.As in GAN,adversarial training is used to constrain the coding distribution to Gaussian distribution.Experiments show that our network perform well in reconstructing inputs and generateing complex geometric shapes contained in CT slice.In anomaly detection,we assume that the generative model is trained only on normal samples,thus in test phase reconstruction error on normal samples is lower than on anomaly samples.So we proposes to use reconstruction errors as anomalous score.Samples with defects are consider as anomaly and non-defective samples are consider as normal samples for perfermence evaluation.Results show that our model achieves the best performance among similar algorithms.Analysis of anomalous scores and latent code shows that the assumption that reconstruction error of normal sample is lower than anomaly sample consists with reality.Anomaly detection algorithm can find likely defects in given samples.However on the one hand,we can’t get the shape and location of defects with anomaly detection solely,which make it hard to analyze defects’ distribution in components.On the other hand,anomaly detection algorithms can’t compare to end-to-end networks neither in perfermance nor in compute efficiency.For this reason we further explores end-to-end defect segmentation algorithm.Because existing annotation tools can’t visiualize CT images,in this thesis the ray projection method is used to visualize the 3D volumes of the component CT detection images.We futher visualizes 3D defects samples and its semantic labels using ray cast method.Visualization tools allows us to observe three-dimensional structure of the defect intuitively,which helps to check annotation correctness and improve annotation quality.In order to reduce the workload of semantic labeling,we propose to use anomaly detection to generate annotation suggestion first,which gives locations that may contain defects in CT slice and then perform manual corrections.This simplifies the labeling process and reduces annotation workload to a certain extent.With annotated data,we use U-net to segment the defects in component CT slices,the result of which is qualified.Experiments proved that Unet network is currently the most suitable choice when the number of defect samples is small.At last,we discussed the feasibility of using anomaly error map for spatial attention mechanisms,points out that anomaly error map can be used as saliency map for shallow feature map in segmentation networks.