Guided ultrasonic detection system has the ability to inspect long range and not accessible pipelines. Especially, the T(0,1) mode guided wave was used widely at the detection, because the property of non-dispersive. For rapidly judge common features on pipe, this thesis makes an artificial neural network diagnosis system to separate and recognize the signals on pipeline. In the experimental setup, the torsional mode signal are excited by using an array of transducers distributed around the circumference of the 6-inch standard pipe, and the reflected signals contain flange, weld, elbow, and defect on elbow. These features are extracted and have been further processed to limit the size of the neural network; then, the feature signal classify as axisymmetric called black, non-axisymmetric called red, and dividing between the two called R/B ratio. The research also uses finite element method to simulate the weld by building up different kind of profile to analyze its amplitude and simulate the flange, elbow, and defect on the elbow. Because the reflection waves of simulation are too idealize to be the network data, the training data and validation data are collected from the experimental wave. In the recognition of artificial neural network, the signals were getting from two pipes of industry. One has bitumen on it, which makes signals attenuation. The other has a clear elbow and a notch on elbow. The two-class recognition method successfully separates flange and weld in low frequency; but in high frequency, the weld signal amplitude is close to flange signal, because the signals decay when guided waves pass to bitumen, and this makes the judge become error. Furthermore, the network recognizes defects on elbow, where the signals have 3 peaks and 2 peaks when the elbow has defect on it. The training result shows that the 3 peaks have better convergent than the 2 peaks in the network. Finally, the developed method can recognize those defects on the elbow when the reflection signals have 2 peaks, and when reflection signals have 3 peaks, it could not make a good judge because the network limit by sample data.