Intelligent Predictive Maintenance for Itapebi Hydro-Generator

Transcription

1 Intelligent Predictive Maintenance for Itapebi Hydro-Generator LUIS CLAUDIO RIBEIRO, MARCO AURÉLIO M. GUTIERREZ, ELIAS G. DA SILVA Itapebi - Neoenergia Power Plant Co. Av. Edgar Santos, Bloco A4, 1º andar, Salvador, , BA BRAZIL {lribeiro, mgutierrez, LUIZ EDUARDO BORGES DA SILVA, ERIK L. BONALDI, LEVY ELY DE L. DE OLIVEIRA, JONAS GUEDES BORGES DA SILVA E GERMANO LAMBERT-TORRES CGTI Centro de Gestão de Tecnologia e Inovação R. dos Expedicionários, 1325, Campinas, , SP BRAZIL {leborgess, erik.bonaldi, levy.oliveira, jonas.borges, germanoltorres}@gmail.com Abstract: - This paper presents the development of a supervision system for predictive maintenance and diagnosis of hydro-generators. The aim of the developed system is to verify the degradation conditions of Itapebi generators. Initially, a system for extracting features of the hydro-generator operational database has been developed to detect possible problems that cause premature fails. The system has been divided in two parts. The first one is a data acquisition system directly connected to the generator in order to sample some operational variables. The second part concerns an intelligent data mining, into the database involving the supervision system variables, to use the existing historic data to perform analysis of the problems and possible causes. Key-Words: - Electrical measurements, maintenance, data mining, intelligent systems, hydro-generators. 1 Introduction The generators are the most important equipments in the energy generation process. The power system reliability, power system supply and power system stability are indexes directly affects for the generator operational conditions. For this reason, protection and monitoring equipment are increasingly employed in order to prevent fails.one of the technologies that can be employed within the purpose of predicting failures is the electric signature analysis (ESA) [1], which consists of a set of methods and techniques that monitor the condition of electric machines by identifying patterns and deviations. It is detected by processing and analysis of voltage and current signals acquired machinery under monitoring. These techniques based on electrical signatures can be applied from the generator and primary source until the motor and load coupled. They may be based on: (a) invasive methods, such as the electric circuit analysis (with static analysis and non-energized machine, also referred to as offline analysis and therefore invasive), or (b) non-invasive methods, such as ESA (dynamic analysis, i.e. with the machine in operation, also referred to as online analysis) [1].For a more comprehensive monitoring of the generator, it is important to the application of invasive and non-invasive methods, based not only on the signature electric as well as other monitoring techniques such as vibration analysis. It is recommended the application of invasive techniques in shutdowns, while non-invasive techniques should be applied periodically during the operating cycle of the machine. This project aims to develop a methodology for the detection and dynamic analyses of online monitoring of the condition of hydro-generators based on acquisition, processing and analysis of voltage and current signals. The main fails such as short-circuit in stator and rotor windings, fails in excitement system, misalignment and eccentricity of spinning field have been studied by electric signature analysis. The paper presents the developed system and some practical results in a Itapebi Power-Plant, located in Itapebi, northeast part of Brazil. ISBN:

2 2 Electric Signature Analysis Electric Signature Analysis (ESA) is the term used for all evaluations of voltage and current signals of electric machines. The most common analysis transforms the voltage and current signal to the frequency domain where they are analyzed. The analysis is based on two fundamental assumptions: (a) the signature of a machine with failure is different from the signature of a machine in perfect state of operation and (b) the failures are repeated with regular patterns, causing failure patterns, which can be identified and related parts of the machine. These techniques can be applied in electric motors and generators. It is important to note that the Voltage Signature Analysis (VSA) is related to an upstream analysis, i.e. toward the generator; and the Current Signature Analysis (CSA) is related to a downstream, i.e. toward the motor. In this project, CSA and the Extended Park Vector Approach (EPVA) are the methods used in this development because they have more features applicable to electric generators. Also these methods have been applied in electric motors, but not to generators yet. 2.1 Current and Voltage Signature Analysis CSA Current Signature Analysis or VSA Voltage Signature Analysis techniques are used to generate analyses and trend of electric machines dynamically. They aim to detect predictive problems in a rotating electric machine, such as: problems in the stator winding, rotor problems, problems on the engagement, problems in bound load, efficiency and system load; problems in the bearing, among others. It may initially cause a certain astonishment that the electrical signals contain information in addition to the electrical characteristics of the machine under supervision, but they work for mechanical defects as a transducer, allowing the electrical signals (voltage and/or current) can carry information of electrical and mechanical problems until the power panel of the machine. The signs of current and/or voltage of one or three phases of the machine produce, after analyzed, the signature of machine, i.e., its operating pattern. This signature is composed of magnitudes of frequencies of each individual component extracted from their signs of current or voltage. This fact isolated by itself is already a gain, as it allows the monitoring of the evolution of the magnitudes of the frequencies, which can denote some sort of evolution of operating conditions of the machinery. The response that the user of such a system needs to know is whether your machine is "healthy" or not, and that part of the machine the failure might occur. This analysis (diagnosis) is not something easy to be done, because it involves a set of comparisons with previously stored patterns and own "history" of the machine under analysis. In this instant, normally a specialist is called to produce the final diagnosis, generating the command when stopping the machine. Thus, the system developed in this project for automatic diagnosis that combines the history of turbo-generator, expert knowledge and failure patterns can be quite useful for the company. The Fast Fourier Transform (FFT) is the main tool employed, however some systems employ in conjunction with other techniques to increase the ability of fault detection signal from acquisition, through processing, to the diagnostic step. 2.2 Approach by Park Vector Extended The EPVA technique is used in the project mainly to check electrical stator imbalances, however it can be employed in the detection of other types of failure [1].The central idea of the technique is checking breakdowns by the distortion circle Park, that is, the more distorted the greater is the imbalance of electric machine. The current components of the vector Park are described by i D and i Q : i D 2 i 3 A 1 ib 6 1 ic 6 and i Q 1 ib 2 1 ic 2 Where the currents i A, i B and i C are the three phases. Ideally we have: 6 6 i D cos( ) 2 im t i Q ( ) 2 im sen t and ISBN:

3 To normal conditions, as shown in Fig. 1, the circle of the Park is centered at the origin of coordinates. (a) Fig. 1. Balanced signals: (a) signals in time and (b) circle of Park. (b) The circle Park passes to suffer distortions when there are abnormal conditions of operation or when failures of electrical or mechanical sources arise. However, these distortions in the circle of Park are not easy to be examined or measured, hence the proposition of the EPVA method extension, observing the spectrum of Park vector module. The EPVA technique combines the robustness analysis of circle of Park and the flexibility of spectral analysis [1]. An important point of this process is the transformation of Park causes a fundamental component of signals analyzed is withdrawn. This fact allows the components characteristics of failure to appear with greater prominence. And more, because it is a method that simultaneously covers the three phases, flaws that can only be detected through this joint analysis, such as the electric stator unbalance, are also covered by. 3 Electrical Signal Processing 3.1 Composition of the digital signal The purchasing system used makes the sampling of analog signals through an analog-to-digital converter (ADC). This way, an analog signal is sampled at a fixed rate for a fixed period and then stored and streamed to the analysis software. This way, each signal is encoded in a vector of integers, whose values represent its amplitude in different moments. Knowing of the transformation ratio of the sensor being used and the ADC input range, the software gets the signal amplitudes to current values in Amperes and voltage values in Volts. For example, assuming a current transformer ratio of transformation is of 100 mv/a and a sampling system of 16-bit input range ± 5V, a sign of 10 to be measured as a level 1 V on the input of the ADC and converted to an integer value If the entry is +5 V, this value would be or It is for the software to interpret these integer values, interpret them and reconstruct the signs of current and voltage. However, the signs used are now discrete (digital) and all the calculations and algorithms are employed the techniques of digital signal processing. 3.2 Conditioning and digital signal processing For the characteristic extraction of digital signals, there is a pre-conditioning process and then some paths to compute the values of each variable. Different parameters are obtained in the time domain and frequency domain. In Fig. 2, a flowchart of the used techniques is shown. The blocks with grey background represent a processed signal that can be viewed or used for the extraction of parameters or characteristics. The blocks represent a blank or algorithm applied to digital signal processing. The following is brief commentary of each of the processing blocks: Composition of signs: Converts data sent by the system of acquisition in digital signals whose amplitudes represent actual values of current and voltage. Preconditioning: process that eliminates the early part of the signal to avoid samples obtained during the transitional period of the filters. It is then eliminated the average value of each signal whose nature is toggled. Park Transformed: when there is a three-phase electrical system, comprising three strands (I A, I B and I C ) and three voltages (V AB, V BC and V CA ), applies the Park to get the vector of Park, consisting of the components I Q, I D ISBN:

4 and I 0. In some cases, it is also used the spectrum of the module of this vector to get the electric unbalance the system. Hilbert Transform: when applied to a signal, returns the magnitude (envelope) and instantaneous phase of the same. Filter RMS: knowing the fundamental frequency of the signal, the RMS filter returns the snapshot of the signal RMS value throughout the sampling period, resulting in the so-called RMS Curve. Windowing: filter applied to a signal in time to reduce the effect of "leakage" in the spectrum of frequencies of the same, due to process a signal with beginning and end. There are several types of windowing (Blackman, Hamming, Hanning, etc.). Fourier transform: used to transform a time domain signal into the frequency domain, the discrete Fourier transform (DFT) returns a vector of complex where obtained the spectral ranges and their corresponding stages. Fig. 2. Block diagram of the algorithms of signal conditioning and processing. In the process of windowing, was selected the Blackmann window. This window facilitates the identification of peaks, since it presents slightly wider lobes and lower "leakage" on their side bands than other Windows. To accelerate the achievement of the DFT, we used an algorithm called FFT (Fast Fourier Transform "). However, it is a requirement that the number of samples of the input signal is a power of 2. In the case where the acquired signals do not satisfy this rule, the technique of "zero-padding", or inclusion of zeros, fills the vector with null values until the number of samples is a power of 2. When the "zero-padding" is applied, proportional correction is made in the amplitudes of the spectrum of frequencies for which the true values are obtained. For example, in a signal with zeros are inserted samples to your end so that the signal has the total of samples, i.e samples. This technique can increase the spectral resolution, but it should be clear that there is more information than a spectrum obtained from a sign without the "zero-padding". 4 Description of the Data Mining Algorithm This section introduces expeditiously the data mining algorithm used to perform comparisons between the processed signals, the database signals and failures patterns. The used technique was based on the Rough Set Theory. This technique aims to extract a set of rules (or conditions) from a database through two hyper-sets, called upper approximation set and lower approximation set. The set of rules contains the lower approximation set and is contained by the upper approximation set. The central idea of the algorithm is to reduce the number of elements in the upper approximation set and to increase the number of elements in the lower approximation set. In an ideal condition, these two sets would become only one set that would be the required set. This set is ISBN:

5 represented by the set of production rules. The used algorithm has six main steps, they are: 1. initialization; 2. remove equal examples; 3. remove of dispensable attributes; 4. compute the core set; 5. compute the reduce set; and 6. merge rules. 5 Illustrative Example The Itapebi hydroelectric plant is operated by Itapebi Power Generation Co., which is a publicly traded corporation, controlled by holding Neoenergia Co.. It is located on the Rio Jequitinhonha Valley, in the currencies of the States of Bahia and Minas Gerais, Brazil. After an analysis of three generating units of 150 MW each, we opted to make this installation in the generator # 3, who would suffer a stop for scheduled maintenance within a few months. The plant began operating in Some Features about the Data Acquisition Figure 3 shows the layout of the installation monitoring system, while Fig. 4 shows how the system was actually installed in the Panel chosen. Fig. 3. Layout of data acquisition system. Fig. 4. Photo of the installation of data acquisition System. The monitoring system makes their analysis based on an analysis of current and voltage signals measured from hydro-generator. Each measurement is called "acquisition" and contains one or more current and voltage signals sampled from hydro-generator, via data acquisition device. The time of each acquisition is determined by the sampling frequency and number of samples that the user has defined. The analysis of acquisitions is an important step in the verification and confirmation of fault diagnostics issued by the system. Through dialogue of analysis, you can view the current and voltage signals acquired and their signatures (VSA, CSA, EPVA), all in a friendly and easy to use. Presented in graphical form and organized into "tabs", the analysis Dialog allows the user to show parameters and failure patterns extracted from the acquisition. It has five views, organized in tabs above. They are: Sign in time: shows the signs of current and voltage at the time, such which were sampled. In the upper right is shown the phasorial diagram of the electrical system-phase related to the measured signals; CSA + VSA: in this view are shown signs of current and voltage in the frequency domain. Through the analysis of the current signature (VSA) and voltage (VSA), allows the user to show various patterns of hydrogenerator failures; EPVA: exposes the vector of Park of current and voltage of the system through graphics module of vector Park and Park circle. Possesses diagnostic aid standards analysis EPVA; ISBN:

6 5.2 Computational Package The developed computational package is composed to two main parts: (a) data acquisition control and (b) feature extraction and signal processing. The first part contains parameters for data acquisition process. Special window interfaces have been developed to transfer all system control to the operator. Examples of this interface are shown in Fig. 5. These interfaces are in Portuguese language. The first figure shows an example of the supervision interface with the data acquisition control information; and the second figure is one of the analysis procedures. As illustrated in Figure 5a, the supervisory dialog is organized through the following regions: Header: top that indicates the handle of hydro-generator, your general condition and the condition of the parties which compose it. The general condition is represented by the circle in the right part of the identifier; Trends: central part of the dialogue that allows monitoring of key parameters and patterns of failure through trend curves; Acquisition: located at the bottom, allows the user to control how measurements are made. Furthermore, it allows presents a description of recent acquisitions and the history of attempts to purchase. Figure 5b illustrates the messages of the history of attempts to purchase. Through this history it is possible to identify problems with the purchasing system. (a) (b) Fig. 5. (a) Supervisory dialog screen and (b) Supervisory dialog screen-purchase history (interfaces in Portuguese language). 6 Conclusions This paper shows a complete development of a supervision system with an intelligent data signal processing based on feature extraction using Rough Set Theory. The feature extraction relates processed current and voltage signals from the hydro-generator by VSA, CSA and EPVA techniques, hydro-generator electrical and mechanical parameters, and typical types of failures existing in this kind of machine. Hardware and software have been developed to acquire and treat the electrical signals in a non-invasive process. It means, the operational condition of the generator is verified without any type of disturbance in the machine or in its control. The electrical signals are taken out of the machine, more specifically in the secondary of instrument transformers (CT and PT) in the panel control. This system is currently in full operation at Itapebi Power Plant, in Brazil. Reference: 1. Bonaldi, E.L.: Failure Predictive Diagnostic in Three-Phase Induction Motors with MCSA and Rough Set Theory. Ph.D. Thesis, Itajuba Federal School of Engineering, Itajuba Brazil (2006) - in Portuguese. ISBN: