A Study on Robust Optimizer and Artificial Intelligence for Optimal Sensor Placement in Pipeline Systems

Abstract

Pipeline systems are extensively utilized across various infrastructure systems, including petrochemical plants, power plants, or ships. Hence, the only advanced-engineered design is required, but also systematic maintenance of the system is also important by employing structural health monitoring (SHM) solutions. The SHM system consists of function of sensor data acquisition, signal processing, and system health assessment. The optimal sensor placement (OSP) has received attention, to address the quality of sensor data, as the first step of SHM. This paper proposes two research thrusts for OSP application. First, a novel optimization algorithm is proposed. The Adam-optimizer is incorporated into the genetic algorithm as a mutation operator, namely Adam-mutated genetic algorithm (AMGA). The Adam can be anticipated to enhance the ability to escaping local minima. To evaluate the performance of the proposed AMGA, two experiments are conducted: (1) benchmark optimization test and (2) OSP application with the reduced scale pipeline system. Two unimodal and three multimodal benchmark problems are employed. The AMGA is evaluated subject to performance metrics for evolutionary algorithm. As the real application of OSP, optimal sensor network is designed based on AMGA. The simulation model of reduced scale pipeline system of naval combat vessel was built and calibrated toward experimental observation acquired from testbed. The statistical database is constructed through Monte Carlo simulation approach. Then, optimal sensor network design is determined by AMGA. To evaluate the performance of the AMGA, both experiments were conducted subject to the standard genetic algorithm (SGA) and compared with each other. As a result, the AMGA outperformed the SGA in terms of ability to escaping local optima (i.e., robustness). A novel OSP framework is proposed by incorporating an eXplainable artificial intelligence (XAI) in pipeline systems as the second research thrust. The novelty is on the second phase of the proposed framework, so-called XAI-based search space reduction method. The statistical database constructed in the first phase of the conventional OSP framework is utilized to train the proposed convolutional neural network (CNN) model. The CNN model is trained to classify the damage scenarios. After AI model train end, Grad-CAM values are extracted. The Grad-CAM values are pre-processed with two steps: (1) peak value extraction by peak finding algorithm, and (2) clustering. As a result, reduced search space (RSS) is obtained against the original search space (OSS). The proposed OSP framework is demonstrated with reduced- and real-scale pipeline system. In the reduced scale pipeline simulation, the feasibility of employing XAI approach is evaluated. In here, the reasonability of using Grad-CAM is justified as a physical criterion. In the real scale pipeline simulation, whole procedure of the proposed framework is conducted. To evaluate the effectiveness of the XAI-based search space reduction method, state-of-the-art (SOTA) meta-heuristic methods are employed including GA-LXPM, DABC, IRGA, AMGA, and GA-TDX. Finally, the effectiveness in solution quality and efficiency in computational costs are evaluated.