Open Access Repository

A hybrid failure diagnosis and prediction framework for large industrial plants

Downloads

Downloads per month over past year

Kim, D 2018 , 'A hybrid failure diagnosis and prediction framework for large industrial plants', PhD thesis, University of Tasmania.

[img]
Preview
PDF (Whole thesis)
Kim_whole_thesi...pdf | Download (6MB)

| Preview

Abstract

Modern industrial plants contain large number of facilities interacting with thousands of sensors and control. A single failure in a facility can produce inconsistent outcomes, which can be affected to core part of industrial plant, and it becomes a critical industrial disaster. Therefore, it is crucial to find and apply the best solution for maintaining facilities and preventing industrial disasters. The early-stage solution was the regular maintenance but this approach cannot be a perfect solution to prevent most industrial disasters. This is because regular maintenance is not effective for all but only few facilities, and is spent too much time and cost to afford. The recent trend of industrial plant maintenance focuses on two main factors, alarms and human expertise.
The system collects the status of different types of facilities from the sensors, which are attached, on each facility. If there is any specific symptom detected from sensors, the alarm will be ringed. The collected alarm data are sent to the human experts in the real time. The human experts have experienced various types of industrial disasters so they have sufficient knowledge in diagnosing and treating failures.
In this dissertation, I studied how to use alarm data and expert knowledge together with these characteristics. In this study, I constructed knowledge using failure reports reflecting alarm data, expert knowledge, which are significant knowledge resources of the industrial field and proposed a method to continuously manage and use such knowledge.
This dissertation can be divided mainly into three parts of subjects of researches.
In the first study, I propose a hybrid knowledge engineering method based on machine learning- expert knowledge, which enables machine learning and domain experts to generate and update knowledge together.
First of all, after constructing a knowledge base by applying real-time alarm data and machine learning, the expert can directly update the knowledge continuously, thereby enabling knowledge creation and management in a fast and efficient manner.
After constructing a knowledge base by applying real-time alarm data and machine learning, the expert can directly update the knowledge continuously, thereby enabling knowledge creation and management in a fast and efficient manner.
Second, I propose a methodology for constructing causal knowledge as overall conditions and treatment actions for failure. Failure report includes the cause-and-effect relationship and its order of occurrence. The proposed methodology analyzes the failure reports written by domain experts using natural language processing techniques and helps to organize the cause-effect and treatments for the failures into a network form.
Finally, the knowledge constructed by the hybrid knowledge engineering method and the causal failure knowledge are fused and applied to the fault diagnosis and prediction.
As a result of the performance analysis, the proposed framework is superior to the other methodologies regarding failure diagnosis and prediction. The proposed decision support method in this dissertation can evolve the two types of knowledge required in the field gradually. Thus it was able to solve the knowledge management difficulties, and using the two knowledge together complementarily; knowledge management efficiency has been achieved. Moreover, it showed superior performance compared to existing methods based on data.

Item Type: Thesis - PhD
Authors/Creators:Kim, D
Keywords: Sensor Data Mining, Machine learning, Knowledge-based System, Expert system, Knowledge Engineering, Knowledge Reuse, Failure Detection, Failure Prediction
DOI / ID Number: 10.25959/100.00028813
Copyright Information:

Copyright 2018 the author

Item Statistics: View statistics for this item

Actions (login required)

Item Control Page Item Control Page
TOP