Leak detection for natural gas gathering pipelines under corrupted data via assembling twin robust autoencoders

Timely leak detection is vital to guarantee the safe and reliable operation of natural gas gathering pipelines, and the data-driven methods become a prospective tool with their widespread installation of sensors. However, these methods face several challenges such as the corrupted normal data, deficient representations learning and their insufficient utilization, low identification accuracy induced by situation without labeled leak data. Nevertheless, previous approaches mostly focused on addressing only one or two of these challenges. To collaboratively solve the above challenges, this paper proposes an unsupervised leak detection method based on twin robust autoencoders (T-RAEs) for natural gas gathering pipelines. First, a fresh robust autoencoders (RAEs) approach is developed to deal with various outliers of the corrupted normal data for multivariate time series so as to learn distinct latent representations. Next, based on the developed RAEs approach, an unsupervised T-RAEs framework is presented to jointly build the normal models of given pipelines, which considers not only the learning of diverse dependency patterns but also the dispose of various outliers. Specifically, the robust long short-term memory autoencoder (R-LSTM-AE) is employed to discover long-term dependency patterns while coping with the unstructured outliers, and the robust one-dimensional convolutional autoencoder (R-1D-CAE) is utilized to capture the short-term dependency patterns while managing with the structured outliers. Unlike the reconstruction errors of R-LSTM-AE in input space, and the errors for R-1D-CAE are computed in both input and hidden spaces to fully exploit its learned hierarchical information. Then, an integration strategy is put forward to integrate the obtained reconstruction errors of T-RAEs for the calculation of their global leak scores. Afterward, to scale the diverse magnitudes of integrated errors and eliminate their correlations induced via correlated neurons across layers, the minimum covariance determinant (MCD) method is employed as a robust normalized aggregation method to aggregate these errors along the pathway. Finally, the efficacy of the proposed leak detection method is verified by experiment results on real-world datasets obtained from natural gas gathering pipelines.