- 351 views
- 479 downloads
DATA DRIVEN SOFT SENSOR DESIGN: JUST-IN-TIME AND ADAPTIVE MODELS
-
- Author / Creator
- Sharma, Shekhar
-
A number of industrial processes involve variables that cannot be reliably measured
in real time using online sensors. Many such variables are required as inputs in
control schemes to ensure safe and efficient plant operation. Laboratory analysis,
which is a reliable method of measuring these variables, is slow and infrequent. Thus,
mathematical models called soft sensors which can estimate these hard to measure
variables from the abundantly available online process measurements have been used in a number of industrial applications. Among the various soft sensor applications of online prediction, process monitoring, fault detection and isolation, the focus of this
thesis is on online prediction and parameter estimation applications. Just-In-Time (JIT) modeling is a unique framework wherein a local model is created every time a prediction is required. One of the most critical components of JIT models is the similarity criterion which determines the data used in the local models and their associated weights. To handle nonlinear and time varying systems simultaneously under the JIT framework, a new similarity metric which incorporates time, along with the traditional space distance, to evaluate sample weights, is proposed. Further, a query based method to determine the bandwidth of the local models adaptively, as an alternative to the offline global method, is also developed. Next, the distance-angle similarity criterion used in modeling dynamic systems under the JIT technique is studied. An improved weighing scheme is then proposed which enables a more accurate selection of data for local modeling and provides a better interpretation of results. Again, for this proposed weighing scheme also, an alternative to the global bandwidth estimation, called the point-based method, is
proposed. In the field of online soft sensor prediction and parameter estimation applications, adaptive linear regression algorithms such as recursive least squares and moving window least squares are widely used because of their simplicity and ease of implementation. However, these methods are not robust to outlying values. We develop a new robust and adaptive algorithm with a cautious parameter update strategy. The proposed algorithm is also quite flexible and a number of variants are easily formulated. Finally, advantages of the methods are clearly illustrated by applications to numerical examples, experimental data and industrial case studies. -
- Subjects / Keywords
-
- Graduation date
- Fall 2015
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.