Polyethylene pipelines have become the first choice in the field of gas and oil pipelines due to their advantages in corrosion resistance, plasticity and cost performance. Because of their flammable and explosive characteristics, natural gas and oil are prone to safety hazards, so the quality should be strictly controlled in production and use. PE pipes are usually welded on site during laying. Human factors, long-term weathering and corrosion effects in the welding process may lead to defects in the welded joints of pipes. Therefore, the quality inspection of welded joints of pipes is particularly important. At present, the detection methods of hot-melt joints mainly include ultrasonic testing, infrared thermal imaging detection, ray detection, etc. These detection methods have certain limitations in the actual polyethylene pipeline detection.As a non-destructive, non-ionizing, non-contact new detection technology, Terahertz non-destructive testing technology has a strong penetration of non-polar, non-metallic materials. The non-destructive testing technology based on THz imaging technology can obtain the frequency domain, time domain and spatial information of the tested sample, and can clearly image the defects of the material. With high defect detection and positioning ability, it is gradually gaining prominence in the field of non-destructive testing, which can be used as an effective supplement to traditional testing technology.In order to solve the problems of cold welding, over-welding, non-fusion and inclusion defect detection of PE pipe hot-melt joints, reflective pulsed terahertz time-domain spectroscopy with a frequency range of 0.25~2.5 THz is used to detect PE hot-melt joints. The hot-melt welding between PE100 sheets is used to simulate the actual hot-melt welding of polyethylene pipes. Standard hot-melt welding and cold welding, over welding, incomplete fusion and inclusion hot-melt defect joint samples are made. Through point-by-point scanning of terahertz system, the wave pattern and kurtosis imaging at the joint are obtained. In the wave pattern, there is little difference between the cold welding and over welding defect samples and standard welding samples, with obvious differences between inclusion defects and incomplete fusion defects and standard welding samples. The unfused and mixed hot melt defect samples and standard welding samples have obvious difference in waveform, the surface reflection echo of unfused defect is about 50% of the standard welding, the surface echo of mixed defect is between 20% and 40% of the standard welding. In kurtosis imaging, the weld width of cold welding and over welding defects is larger than that of standard hot fusion welding, and there are defects in the weld. Incomplete fusion defects and inclusion defects can clearly see the rough outline of the defect part. The first kind of wavelet scattering network-convolution neural network defect recognition model is constructed for standard welding, inclusion and incomplete fusion defects. The defect recognition rate can reach more than 98%. The second kind of defect recognition model is constructed for quantitative identification of standard welding and metal inclusion, coarse sand and branch defects. The relative error of defect recognition is less than 7.42%. This study lays a foundation for the application of terahertz non-destructive testing technology in the detection of polyethylene pipe hot-melt joints.