CNG composite gas cylinders are subject to alternating loads for a long time, making them prone to fatigue damage and internal defect formation. These defects decrease strength and cause safety concerns during use. Composite gas cylinders containing internal defects show no obvious macroscopic deformation, and thus, it is difficult to directly detect them. The current gas cylinder testing program lacks an effective means to rapidly detect internal defects in the winding layer, which may result in missed inspections of cylinders containing internal defects. This paper proposes a defect detection scheme based on a combination of the existing key problems associated with CNG composite gas cylinder testing and the existing gas cylinder testing standards and processes. The proposed defect detection scheme uses surface thermal imagery of the steam cleaning process inside the gas cylinder. The scheme uses the steam in the cylinder flushing process as the internal thermal excitation of the cylinder. Using the transient temperature distribution of the cylinder surface recorded by an infrared camera, an artificial neural network is applied to locate and quantitatively identify the defects in the cylinder winding layer. This experimental research shows that the artificial neural network can accurately locate and quantitatively identify the defects in the cylinder winding layer with high recognition efficiency, which is suitable for online detection of gas cylinders.