Fuzzy Logic, Neural Networks and Statistical Classifiers for the Detection and Classification of Control Valve Blockages
Open Access
- Author:
- Roemer, Michael Richard
- Graduate Program:
- Mechanical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- March 20, 2017
- Committee Members:
- Christopher Rahn, Thesis Advisor/Co-Advisor
Karl Reichard, Thesis Advisor/Co-Advisor - Keywords:
- neural network
fuzzy logic
statistical classifier
condition-based maintenance
cbm
support vector machines
discriminant analysis
k-nearest-neighbor
vibration
diagnostics
rotary valve
process control valve
control valve
blockage
valve blockage - Abstract:
- Given recent advances in data processing, information technology and machine learning techniques, condition-based maintenance (CBM) has increasingly been used on high-cost assets as a way to improve product reliability and decrease long-term maintenance costs. While long established, CBM is often applied in a primitive manner using simple threshold monitoring and trending techniques. As such, there are opportunities for expanding this work using more intelligent methods which are now viable in practical settings as sensor and IT solutions decrease in cost. Mechanical systems worth monitoring include the control valve, which is used in a myriad of industries, including both commercial and military applications, and is particularly valuable in many settings. As an example, control valves are critical components in submarine subsystems, such as carbon removal units, which can require precise valve movements. In this and related applications, ensuring that movements are fully realized is crucial to the control process in question. This thesis explores a variety of techniques for classifying valve movements with the goal of detecting two specific types of blockages in control valves. Objects with different material properties can produce vastly different dynamic responses in a control valve when blocking the valve stroke. Therefore, it is the goal of this work to detect two types of these blockages in order to cover a broad range of potential obstructive sources. To do this, seeded-fault testing is performed to simulate these behaviors with the resulting dynamic responses captured by both external sensor instrumentation and inherent control system outputs. Both data sources are used separately for analysis. Classification using these data is tested using fuzzy inference systems (FIS’s), artificial neural networks (ANN) and a variety of geometric and instance-based statistical classification methods. Discriminating features are identified, extracted and subsequently input to these classifiers to test for overall accuracy. Furthermore, multiple approaches to processing these data are used, including real-time monitoring simulations and post-hoc movement analysis. In most cases, accurate blockage detection and classification is achieved from these analyses. As a result, it is hoped that this work can serve as the foundation for future research in the area of fault detection and identification on control valves which govern flow loops in important process control applications.