Defect Detection of Electrical Insulating Materials Using Optically Excited Transient Thermography and Deep Autoencoder

This article presents an automatic and unsupervised internal damage detection technique for electrical insulating materials using optically excited transient thermography and a deep autoencoder (DAE). For this purpose, an optically excited thermography system is built, and the thermal images of the specimen surface during the cooling period are collected for the training of the DAE model. The reconstruction error of the DAE model is used as a feature to generate the feature image. The final binary detection image is obtained by applying adaptive threshold image segmentation. To avoid unwanted errors, small connected components and isolated pixels are removed as they may be falsely considered damages in the feature image. To improve the detection accuracy and reliability, the architecture and hyperparameters of the DAE network are optimized. The featured image and segmented image obtained using the DAE are compared with those obtained using three other traditional methods [manual time-domain image selection, frequency-domain feature analysis, and principal component analysis (PCA)]. Based on the defined evaluation metrics, the proposed DAE framework exhibits the best visual effects and the highest defect detection accuracy. The test results also show that this method is effective in detecting real composite insulators with air-gap defects under different detection interferences.