Detection of internal defects in glass fiber reinforced polymer (GFRP) is vital for aviation safety. In this work, we report a novel approach to predict the defect depths in GFRP based on terahertz time-domain spectroscopy signal analysis with neural networks. We exploit three neural network models, which can be trained and tested with the collected terahertz time-domain signals or the corresponding spectral signals, to carry out defect detection and classification based on the depth. Results show that in general the one-dimension convolutional neural network model outperforms the long-short term memory recurrent neural network (LSTM-RNN) and the bidirectional LSTM-RNN models. The corresponding recall rates are above 0.85 and can even reach 0.97, and the macro F1 score is larger than 0.91. Based on the automatic defect detection and classification, a terahertz image showing the locations and depths of defects can be efficiently reconstructed. We envision this work will advance the development of automatic nondestructive defect detection based on terahertz techniques and neural networks.

检测玻璃纤维增​​强聚合物（GFRP）的内部缺陷对于航空安全至关重要。在这项工作中，我们报告了一种基于神经网络的太赫兹时域光谱信号分析来预测GFRP缺陷深度的新颖方法。我们利用三个神经网络模型，可以使用收集的太赫兹时域信号或相应的频谱信号进行训练和测试，以基于深度进行缺陷检测和分类。结果表明，一维卷积神经网络模型通常优于长期记忆递归神经网络（LSTM-RNN）和双向LSTM-RNN模型。相应的召回率高于0.85，甚至可以达到0.97，并且宏F1得分大于0.91。基于自动缺陷检测和分类，可以有效地重建显示缺陷位置和深度的太赫兹图像。我们认为这项工作将推动基于太赫兹技术和神经网络的自动无损缺陷检测的发展。