A multi-feature parameter neural network analysis technique based on terahertz nondestructive testing for the analysis of the adhesion quality of the heat-resistant composites was proposed. A film (thickness is 0.1mm) extraction method was put forward to simulate the bonding defect of the heat-resistant composites. The terahertz time-domain spectroscopy nondestructive testing technology was used to detect the multilayer high temperature resistant composite bonding defects. Compared the similarities and differences of terahertz time-domain and the frequency domain information between the upper debond defect and the lower debond defect, the multi characteristic parameters were proposed for the adhesive quality, such as upper debond parameter, lower debond parameter and centroid absorption parameter for frequency domain. The characteristic parameters were optimized as the input of back propagation neural network to recognize the upper and the lower debond defects. Based on the back propagation neural network training test, the identification was realized for the 0.1mm thickness upper debond defect and the 0.1mm thickness lower debond defect.