Abstract: Honeycomb structures are widely valued for their light weight and resistance to high-temperature, corrosion, and impact, making them important weight reduction materials in aerospace and other fields. However, various types of defects can arise during the manufacturing and service stages of honeycomb structures, particularly internal corrosion caused by moisture ingress and defects such as adhesive layer aging and debonding. Therefore, a honeycomb-structure defect classification and recognition method based on a bidirectional long short-term memory network (Bi-LSTM) is proposed by combining infrared thermal wave imaging technology and recurrent neural networks. The Bi-LSTM model was used to identify and judge the logarithmic temperature–time curve of infrared sequence images. The focal loss and 5-point connection method improved the effective recognition accuracy of the system by effectively solving the problem of imbalanced positive and negative samples which can lead to inaccurate background recognition and low recognition accuracy. Experimental analysis showed that the Bi-LSTM model had an accuracy of 99% in distinguishing water and glue in honeycomb structures and can thus be effectively applied for their detection and recognition.