Research On Multi-dimensional Deep Learning Methods For Defect Detection In Thermal Images

Due to the complex produce process,the composite material is prone to defects in production.And most of them are used in important components such as aircraft tail fins and rocket shells,so it has a possibility to cause safety hazards to equipment when there are defects.Therefore,it is important to accurately detect and evaluate the defects of composite materials in the process of preparation and service.Infrared thermography testing(IRT)has been widely used in defect detection of composite materials.However,due to the interference of factors such as background and noise in the original thermal image sequence,the identification of defect characteristics is unsatisfying.Therefore,it is of great practical significance to study suitable and effective thermal image defect detection algorithm which can realize defect feature extraction and enhance defect contrast in IRT of composite materials.This thesis combines multi-dimensional deep learning with IRT.The feasibility of deep learning algorithms is studied respectively in temporal,spatial and spatiotemporal dimensions of IRT data.The main work of this master thesis is as follows:Firstly,the theoretical basis of IRT is described,and the surface temperature field distribution of the tested part under pulse excitation was theoretically deduced.Then,joint scanning thermography(JST)experiments were carried out on the composite specimen with flat bottom hole defects and composite base coating specimen with debonding defects,and the reconstructed thermal images were obtained.COMSOL software was used to finite element simulation study of the composite samples with flat bottom hole defects which were excited by pulse excitation and the surface temperature distribution of simulated specimens was obtained.Secondly,by analyzing the thermal temporal and thermal spatial characteristics of the original data,the sample data of experimental and simulation data were calibrated respectively based on Label Me and MATLAB.The one-dimensional temperature-temporal data sets,the two-dimensional thermal image segmentation data sets and the three-dimensional thermal video segmentation data sets were established.Furthermore,Data Augmentation was used to thermal image data sets and thermal video data sets in order to increase the amount of data.Thirdly,the network structures were designed and modified according to the three dimensions of data and corresponding defect characteristics.1D-CNN network was proposed for feature extraction and regression of temperature-temporal series data because the surface temperature of the specimen has a nonlinear relationship with time and defect burial depth.Aiming at the abnormal distribution of temperature field in thermal images,AG-UNet was proposed with spatial self-attention gate module for extract spatial features of thermal images.3D-UNet network was obtained by 3D convolution module and the temporal convolution module to simultaneously extract thermal temporal and spatial features,which could realize defect segmentation of thermal videos.Finally,the three models were trained by the established data sets.In order to verify the feasibility of these models,they were applied to detect defects of the experimental data.The diameter to depth ratio and the probability of detection(POD)were used to measure the detection results.Compared with traditional algorithms such as principal component analysis(PCA),thermographic signal reconstruction(TSR),fast Fourier transform(FFT),more defects could be detected and defects of smaller diameter to depth ratio could be detected by deep learning models.More significantly,1D-CNN can detect defects of arbitrary shape,which is not affected by the spatial characteristics of defects.