Intelligent fault finding of the gas turbine of the power plant using the event analysis method

Electrical Engineering Master Thesis

Control Orientation

Abstract

New methods of identifying errors in systems such as using qualitative analysis of events can lead to tangible and understandable results for everyone. The gas turbine system is one of the systems in which there is a possibility of many errors. And sometimes even modeling them with mathematical methods is difficult. The method of qualitative analysis of events seeks to solve this issue in a simple and expressive language. This method was first implemented by Orban in 1972 for fault detection in gas turbine. But it remained silent due to the weakness in error identification.

The basis of the method of qualitative analysis of events is to show a language for events that take place in a series of intertwined and related processes. In this method, extracting, fragmenting and sorting connected events is based on seven geometric shapes. The distinction and unique characteristics of these forms are in their first and second derivatives. which is done by using the method of minimum error sum of squares to identify and measure the similarity between them. In this way, the polynomial function (up to order 2) is fitted on the whole data. With the result of inappropriate adjustment, the data set is divided into two equal parts and the bisection process continues until the appropriate adjustment is achieved. In the meantime, the measure of error acceptability is done by the F test method. And finally, the example corresponding to the fitted curve of the geometric shape is stored in a library by assigning the name related to the geometric shape as one of the detected pieces of the event. The next pieces are placed next to each other in the same way according to time intervals. The result of this process is the formation of a time signal in character language (transformation into a language based on basic geometric shapes). With this method, the error signal is characterized as a pattern. and using fuzzy logic, the similarities between stored patterns and new data from sensors that are unknown; are evaluated. And if the similarity with the error patterns is observed, the error identification process is performed.

Key words: Bisection of intervals, qualitative analysis of events, gas turbine, fuzzy logic

Chapter One

1-1 Introduction:

Knowledge progress and emergence New technologies in the field of industrial automation and precision instruments have effectively narrowed the field for human presence to control and monitor processes. In such a way that the unprecedented growth of electronic circuits in recent years has caused; The scope for many possibilities related to data management and interpretation will increase. Advanced microprocessor systems are one of them. With their faster speed, a significant amount of large-scale data has been made available to users at the same time. This shows the platform for using artificial intelligence for monitoring, control and diagnosis in a wide range of processes. In important industries such as the power plant industry, the existence of the gas turbine system as a complex multidimensional system that consists of different subsystems with non-linear parameters, the existence of data with a variety of numbers and scales, makes the possibility of mistakes in the decisions of the operators in many cases. For this purpose, efforts are made to optimally exploit and minimize operator errors from new technologies that can automatically detect and identify the error by storing and interpreting data; Employed Among these methods is the use of the method of qualitative analysis of events[1]. In this method, the events or signals of a process in different time intervals that are a combination of several consecutive events. By splitting and storing each of them in the form of mathematical characteristics that are qualitatively and quantitatively classified. It provides conditions for comparison with the main stored patterns from a real signal.

1-2 Problem definition

Using complex differential equations to identify the system and its modeling has always been associated with problems.. And sometimes due to the use of simplifying hypotheses or calculation errors, it will not be suitable for accurate modeling [3]. For this purpose, today, the use of methods that are far from complex mathematical equations and calculations. It has been given more attention. Qualitative analysis is one of the methods that has made this possible widely in various fields.

The basis of event analysis consists of two important components:

A language to represent events

A method to identify the event [2]

how to map events to operating conditions and environment is shown in Figure (1-1). In this figure, the language of events using seven geometric patterns that are related to the first and second derivatives of the event; Figure (1-2) is considered. To extract these seven shapes, it is necessary first. The whole data is fitted with a polynomial in an iterative process from degree zero to degree two. Then, using the F test method, a criterion for measuring the structure or the evaluated equation is obtained, and this procedure is subjected to qualitative analysis or event identification with the technique of halving the distance until the whole data is subjected to the acceptance criterion of the degree. The data obtained after the operation of two halves, adjustment and benchmarking are extracted and stored in a library called primary patterns resulting from qualitative analysis as the main database. Then, with the help of fuzzy logic and using the similarity matrix, stored patterns with unknown data obtained from sensors. And by performing the extraction process and all the steps that are used for the sample signals, start comparing with the database patterns that correspond to the fuzzy rules (then-if) and are already considered for the events of this system; It is executed. And finally, the most dominant condition will be identified in terms of quantity and quality [9].

Converting a time-varying signal into character format or determining the language of the event was presented in 1995 in detail. But since 2001, this issue has been seriously pursued to the extent that the difficult problem of patterns extracted in noisy processes [3], which caused the lack of correct identification in diagnosis, was reviewed and improved. The noise characteristics included the noise level and the changing of the shapes of the patterns by changing the sampling speed and scales. In order to extract

events automatically and reach a suitable solution[4] in the noise process, the bisection method of intervals[5] was suggested[6] [8]. In this approach, parameterizing the data in the form of a sequence of patterns and shapes with its repetition will lead to the best fit compared to noise [9]. Not verified. Start by halving the data distance from each other and repeat the halving process until success is achieved in the model. The fuzzy inference method for trend [7] in this thesis relies on the use of automatic trend extraction in noisy processes [9].

Introduction

A good system in the field of error detection[8] should be able to. It can be separated in two safe and unsafe spaces. And even with the complexity of the conditions, the system can recognize the boundary between the safe and unsafe state by assuming the noisy conditions in various states such as transient and permanent, and separate these two dimensions from the extracted data with the least cost in time and at high speed. In such systems, artificial intelligence can be used to achieve these goals. In these methods, the basis of work is defining and organizing database data [9]. Among them, we can mention normalizing and converting them into vectors. Or the network under contract training, that is, from different methods, it randomly selects the available sets for training and another set for testing. In addition, it evaluates its network in noisy conditions. to have a proper reaction with data that has not been seen so far. Since in most industries, data is sent to the control center as continuous and time-varying strings.