Automatic defect recognition in ultrasonic infrared thermography by deep learning of temporal signals

In ultrasonic infrared thermography, ultrasonic waves are used to generate heat in a target being tested. The heat causes temperature variations in the target, which can be captured by an infrared camera. By analyzing the thermal patterns, it is possible to identify defects in the target. We propose a two-stage automatic defect recognition method using temporal signals from ultrasonic infrared thermography. First, the temperature rise thermogram is segmented using Otsu's method to remove most of the noise signals. Second, the heat variation signals are extracted based on the temperature rise thermogram sequences. Then, a deep learning model, termed temporal signal defect identification network (TSDI-Net), is designed to accomplish automatic recognition of defect signals. The TSDI-Net consists of a time-space feature module, a hybrid two stream feature module, a global average pooling module, fully connected layers and a Softmax output layer. To verify the effectiveness of the proposed TSDI-Net, five models in the literature, ACN-LSTM, MACNN, OS-CNN, ResCNN, and XceptionTime, were selected for comparison. Ultrasonic infrared thermography image sequences from 153 components are divided into a training set, a validation set, and a test set with ratios of 75%, 10%, and 15%. Results show that the proposed method outperforms the existing defect recognition deep learning models.