A method for detecting small leaks in gas pipelines based on semi-supervised learning is proposed. The proposed method models the time-varying behavior of signals collected from a flexible acoustic emission (AE) sensor that is attached to a gas pipe to distinguish between normal and leak states. Gaussian Hidden Markov Model (GHMM) is adopted to model the temporal behavior of the AE sensor. GHMM is suitable for processing noise-prone AE signals because its probabilistic approach enables discrimination between normal and abnormal states even in the presence of background noise. Furthermore, GHMM can be trained using only normal data in a semi-supervised way to solve the problem related to collecting abnormal data. The proposed method uses a feature extractor that divides the AE signal into multiple segments and then extracts the characteristics of each segment using an ensemble technique. The proposed method is capable of detecting small gas leaks within a pipeline using signals acquired by a flexible AE sensor that is vulnerable to noise. The performance of the proposed method was verified using a dataset collected from a testbed capable of simulating actual gas pipe leaks.
