This paper proposes a scalable and unsupervised feature engineering method that uses vibration imaging and deep learning. For scalability, a vibration imaging approach is devised that incorporates data from systems with various scales, such as small testbeds and real field-deployed systems. Moreover, a deep learning approach is proposed for unsupervised feature engineering. The overall procedure includes three key steps: 1) vibration image generation; 2) unsupervised feature extraction; and 3) fault classifier design. To demonstrate the validity of the proposed approach, three case studies are conducted using an RK4 rotor kit and a power plant journal bearing system. By incorporating smaller-system data as well as real-system data, the proposed approach can substantially increase the applicability of the fault diagnosis method while maintaining good accuracy. Moreover, the time and effort needed to develop a diagnostic approach for other rotor systems can be reduced considerably.

Pipelines in critical engineered facilities, such as petrochemical and power plants, conduct important roles of fire extinguishing, cooling, and related essential tasks. Therefore, failure of a pipeline system can cause catastrophic disaster, which may include economic loss or even human casualty. Optimal sensor placement is required to detect and assess damage so that the optimal amount of resources is deployed and damage is minimized. This paper presents a novel methodology to determine the optimal location of sensors in a pipeline network for real-time monitoring. First, a lumped model of a small-scale pipeline network is built to simulate the behavior of working fluid. By propagating the inherent variability of hydraulic parameters in the simulation model, uncertainty in the behavior of the working fluid is evaluated. Sensor measurement error is also incorporated. Second, predefined damage scenarios are implemented in the simulation model and estimated through a damage classification algorithm using acquired data from the sensor network. Third, probabilistic detectability is measured as a performance metric of the sensor network. Finally, a detectability-based optimization problem is formulated as a mixed integer non-linear programming problem. An Adam-mutated genetic algorithm (AMGA) is proposed to solve the problem. The Adam-optimizer is incorporated as a mutation operator of the genetic algorithm to increase the capacity of the algorithm to escape from the local minimum. The performance of the AMGA is compared with that of the standard genetic algorithm. A case study using a pipeline system is presented to evaluate the performance of the proposed sensor network design methodology.