Multiobjective Wrapper Sampling Design for Leak Detection of Pipe Networks Based on Machine Learning and Transient Methods
Amir Houshang Ayati, Ali Haghighi
Abstract
This study introduces a novel sampling design (SD) method for hybrid Machine Learning/Transient-Based (ML/TB) leak detection of pipe networks. The proposed technique exploits the hydraulic responses of the network in the frequency domain and the concept of Filter and Wrapper feature selection as a machine learning approach. It also utilizes multiobjective optimization to handle the trade-off between leak detection error and number of sampling nodes. To apply this method, a transient hydraulic simulation model of the network is developed in the time domain. Then, considering a wide range of leak scenarios, the hydraulic responses of the network are calculated at candidate measurement sites. The responses are then transferred into the frequency domain using the Fast Fourier Transform (FFT) and stored as the train and test datasets. To reduce the dimensions of the initial feature vector, a threshold is applied to the responses to filter the very high frequencies. Finally, a classifier based on a Linear Discriminant Algorithm (LDA) coupled to a binary-coded Non-dominated Sorting Genetic Algorithm (NSGA-II) is applied to preprocessed datasets. Four sampling design methods are adopted from the literature and modified in the frequency domain for more investigations and comparisons. Solving two example pipe networks showed that the proposed method outperforms the existing approaches with respect to higher accuracy leak detection with fewer sampling sites.