This paper implemented a model using deep learning techniques to early detect pipe leaks that may occur in aging plant pipeline systems. Additionally, considering the difficulty of collecting leakage data in a real environment, we applied an unsupervi...
This paper implemented a model using deep learning techniques to early detect pipe leaks that may occur in aging plant pipeline systems. Additionally, considering the difficulty of collecting leakage data in a real environment, we applied an unsupervised learning technique to effectively detect leaks through learning normal patterns. This study utilized the TadGAN model, a structure that combines an autoencoder and a generative adversarial network. The model consists of four components: an encoder and decoder, which function as a generator, along with two discriminators. All four models were composed of multi-layer perceptrons. Through this approach, the study aimed to reduce the false detection rate that can occur in existing autoencoder models. As a result of the study, the leak detection model based on the TadGAN model achieved a high accuracy performance of 97.35%. Additionally, compared to existing autoencoder models, the performance improved from 4.61% to 11.29%. Through this comparison, it was experimentally confirmed that the false detection rate problem in the existing model was effectively overcome.