The Study Of Active Thermography Data Analysis Based On Neural Networks

Active thermography is a nondestructive testing technique for defect detection and material evaluation.Its data feature information is mainly reflected by spatio-temporal continuous evolution law,and feature design is difficult.Machine learning is widely used in machine vision tasks,and the form of visual data is similar to that of thermal imaging data.Several key techniques of neural networks applied to active thermal imaging data analysis are studied in this paper.(1)Analyze the active thermal imaging and determine the key problems in the application of neural networks.The difficulties in detecting cracks and evaluating thermal conductivity in laser thermal imaging were analyzed by neural networks.Aiming at the problem of small sample training,the model parameters,transfer learning and simulation data are studied.(2)Analyze the effectiveness of model parameters and simulation data training for small sample training.Convolutional neural networks with different parameter scales were constructed,and the results of experimental data training and testing were compared with those of simulation data training and experimental data testing.The results show that the simulation data is effective for solving the problem of small sample training,but the network parameters are not enough to achieve satisfactory results.(3)Analyze the effectiveness of migration learning and simulation data for small sample training.Based on the vgg-16 model of pre-training of ImageNet training set,direct transfer learning of experimental data,simulation data transfer learning and experimental data testing were conducted respectively,and the experimental data were verified in three aspects: transfer learning of simulation data pre-training model.The results show that using other machine vision task pre-training models for transfer learning is effective for solving small sample training problems,and after using simulation data pre-training,using experimental data transfer learning method is better,and the model miss rate is as low as 1.7%.(4)BP neural network was used to analyze the relationship between thermalsignal and thermal conductivity.The simulation data are used to train and evaluate the model,and the logarithmic data preprocessing method is proposed to solve the problem of large relative error of low thermal conductivity prediction.Experiments show that this method can effectively reduce the relative errors of low thermal conductivity data.