Loading...
Non-Linear Profile Monitoring Using Artificial Neural Network Fault Detection
Mohamed, Ahmed Fares
Mohamed, Ahmed Fares
Date
2018-04
Authors
Type
Thesis
Degree
Description
A Master of Science thesis in Engineering Systems Management by Ahmed Fares Mohamed entitled, “Non-Linear Profile Monitoring Using Artificial Neural Network Fault Detection”, submitted in April 2018. Thesis advisor is Dr. Mahmoud Ismail Awad and thesis co-advisor is Dr. Mohammad AlHamaydeh. Soft and hard copy available.
Abstract
In today’s world, the development of technology and industrial systems is becoming much more complex with the ever-demanding need for higher quality. Anomaly detection is the characterization of a normal behavior of a system or process and the identification of any deviation from such normal behavior. Anomaly detection of critical systems provides an important financial and client competitive advantage since it gives decision-makers lead-time and flexibility to manage the health of the system. Structural systems are critical systems that require continuous monitoring of damage accumulation caused by vibrations and other loads that may cause failures of severe consequences. The current research presents a data-driven methodology for the anomaly detection of structural systems using Multivariate Statistical Process Control (MVSPC). In MVSPC, the quality of a system is assumed to be characterized by explanatory variables where one of these variables can be adequately explained as a function of one or more of the other variables, also referred to as a profile or signature. The proposed method is based on modeling the system outputs (displacements or accelerations) as a function of the input (Ground Shaking) using Artificial Neural Networks (ANN). The Hotelling (T squared) technique is then used to identify any shifts in the ANN weights from the healthy state. The results are tested and validated using simulation data that mimic an actual structural system experiencing ground-shaking. The coefficient of determination R2 values exceed 92% that indicating a good fit of the models. In addition, the results indicate a positive false rate range between 0-19% depending on the complication of the system. Overall, the ANN method was able to detect out-of-control Average Run Length (ARL) shifts much faster than the other methods. The methodology presented in this research is scalable and can be applied to a wide range of systems instead of regular inspection checks in order to anticipate and avoid failures. A successful profile monitoring of structural systems will increase safety and reduce cost.