NEURAL NETWORKS USAGE AT CRYSTALLIZERS DIAGNOSTICS Frischer Robert a, David Ji í b, Vro ina Milan c a V B-TU OSTRAVA, 708 33, Ostrava - Poruba, eská republika, E-mailová adresa: robert.frischer@vsb.cz b V B-TU OSTRAVA, 708 33, Ostrava - Poruba, eská republika, E-mailová adresa: j.david@vsb.cz c V B-TU OSTRAVA, 708 33, Ostrava - Poruba, eská republika, E-mailová adresa: milan.vrozina@vsb.cz Abstrakt P ísp vek se zabývá diagnostikou krystalizátoru za ízení plynulého odlévání oceli pomocí softwarové analýzy vibra ních spekter pop. akustické emise, vzniklé p i b hu za ízení nebo p i nenadálých situacích. Princip metody spo ívá v rozboru spektra získaného signálu, jeho zpracováním v programu Matlab a následným porovnáním vzork v programu Statistica. Poslední krok je nejd le it j í, proto e díky n mu jsme schopni s ur itou jistotou íci, v jaké stavu se zkoumaný objekt nachází. Predikcí povrchových vad na plynule litých p edlitcích p i odlévání na za ízení plynulého odlévání oceli s vyu itím data miningových metod a znalostních systém. Pokud výchozí síti poskytneme dobrá po áte ní data, je schopna analyzovat dal í vzorky a s úsp ností okolo 70% íci v jakém stavu se daný objekt nachází, pop. jestli nedo lo ke kritickému stavu (povolené rouby v kritických místech apod.). Data z akcelerometru (mikrofonu) jsou vyhodnocena pomocí programu MATLAB a je na n aplikován filtr, který zaru í zvýrazn ní relevantních dat. Analogový signál je digitalizován pomocí speciálního I/O modulu firmy National Instruments s ozna ením NI PCI 6221. Z výsledk FFT je patrná korelace spektra a stavu zkoumaného materiálu. V p ísp vku bude popsán pr b h e ení této problematiky a sou asné výsledky e ení, p i n m bude pro diagnostiku testována mo nost vyu ití vícevrstvých neuronových sít v prost edí Neuronové sít softwarového systému STATISTICA. Výsledný diagnostický systém, který by m l být výsledkem e ení, by m l výrazn inovovat a zracionalizoval problematiku optimalizace preventivní údr by desek krystalizátor, p i ní by se uplat oval princip ízení na základ maximálního po tu relevantních dat z technologického a údr bá ského procesu. Klí ová slova: diagnostika, FFT, Matlab, spektrum Abstract This paper deal with crystallizers diagnostics, which is part of continuous casting devices. The diagnostics is made by software analysis of vibration spectrums respectively of acoustics emission. These emissions can cause when unexpected situation occur. Principle of this method is analysis of spectrum of gained signal, its processing in Matlab environment and subsequent verification in Statistica software. Last step is most important, because thanks to it we are able to tell, with certain confidence, in which state the examined object occur. If the initial neural network has good starting data, it is able to analyze other samples and with successfulness over 70% tell, in which state the current objet remain, eventually if occur a critical (limiting) state (untighten screws in critical spots and so on). Data from accelerometer (microphone) are evaluated in Matlab environment and special filter is applied on them. Thos filter will highlight relevant data. Analog signal is digitized with special I/O module from National Instrument company. Type prefix is NI PCI 6221. From FFT result is obvious correlation of power spectrum and the inner state of examined object. Far in this paper will be described process of solving this problematic and current result. To achieve sufficient results, we are using multilayer neural networks with help of Statistica program. Resulting diagnostics system, which should be the final solution, should dramatically innovate and rationalize optimization problems of preventive maintenance of crystallizers desks, where principle of control on maximum relevant data basis should take apart. Relevant data are obtained from maintenance department. Key words: diagnostics, FFT, Matlab, spectrum
1. INTORDUCTION This article is focused on continuous steel casting device diagnostics. Main focus is on technical diagnostics of crystallizer. Crystallizer is special device, which can dissipate redundant heat from liquid steel and force it to solidification in predefined profile. In the crystallizer will solidify surface steel layer and feed in the direction of crystallizers output. This process is going together with many unwanted effects, like abrasion of crystallizers walls. Inner wall state is very important quantity, which is closely observed. Too wear crystallizers walls did not dissipate redundant heat well and there is prone to outbreak or to unwanted cracks by heat tension in the surface layer. Technical diagnostics of desks surface is very time consuming process, which require taking out the crystallizer and dismantling it. Our goal was to verify, if there is a possibility to find out the state of crystallizer without need of taking it out of its bearing. Diagnostics has based on vibration spectrum analysis, obtained while crystallizer runs. In the first phase we deal with vibration spectrum analysis of the crystallizer s model. Vibrations were excited artificially, with help of the striker. Single strikes were driven by PLC Siemens. Detail can be seen on (Fig. 1.). Obr. 1. Schéma budící ásti Obr. 2. Fig. 1. Driving part of the model and model itself The striker was driven by magnetic field with 2s period. Total hit count was set to 9. Purpose of this was to measure response of the system to Dirac impulse. Strike generated by us, has far away to ideal form, but it was sufficient to our purpose. Main disadvantage were undesirable bounce and unequal running of single strikes. To eliminate these disadvantages, we did several strikes in series. From nine measurements, there were at least four very similar. Other measurements show extreme deviations or periodic errors. Theoretically the every impulse should be a source of vibrations with representation of every part of the entire spectrum (white noise generator). In our case, there were represented frequencies from zero to few khz and with erratic representation of single parts. Responses to the pulses were vibrations, which goes throw the crystallizers model. The model acts as selective band pass filter, which some frequencies suppress and the others pass throw without any major changes. We use a special accelerometer from Analog Devices. For data digitalization we used a National Instruments PCI 6221 card. Sampling frequency was set to 100kHz. We expect frequencies in range from 0kHz to 10kHz. So the samples are sufficiently oversampled. As a result of this part was series of amplitude envelopes. Data analysis in the time domain
has no sense, with respect of its course. That is why the data were transformed into the frequency domain with the help of FFT (Fast Fourier Transform). 2. THE ANALYSIS IN FREQUENCY DOMAIN AREA The goal of this part was to convert and analyze single pulses in frequency domain. All of the calculations were done in Matlab environment. Basically, all of the work starts here. It was necessary to process records with the help of fast Fourier transform, separate irrelevant data and find similarities in single records. For this purpose a complex filter were made and will be presented further in this paper. Data were registered continuously in one block. 9 strikes with 2s interval (Fig. 2). In the first step was necessary to isolate single strikes and extract them into the alternative matrix (Fig. 2). On every single data record was applied an FFT. As a result we obtained power spectral density (PSD) of each pulse. Every PSD was further stored into another matrix. Some vibrations on the specific frequencies goes throw without any major changes, but the others are highly suppressed. In the spectrum were so included many relevant data, which was necessary to filter off. We focused on the frequencies, which present highest amplitude in its area. Designed filter pass through the record of every impulse and search for specific values of single spectra component. For another PSD analysis is the position of the peaks very important, more precisely frequencies, which belongs to them. Location of these anomalies is relatively difficult, because signal spectrum has odd curve. After the application of classic algorithm for finding maximum (f (x-1) < max > f (x+1) ), were obtained hundreds of data, which was necessary to evaluate. Some peaks were very close to each other, other on the contrary dominate some specific empty area and has very low amplitude. It was so necessary to develop filter, which could consider amplitude of the peak regarding to its surroundings. Newly designed filter is capable these anomalies eliminate. It passes through the record and for the highest amplitude in current area, a local maximum, is capable to verify, if is really the highest. This interval is optional and its value is inversely proportional to maximum count in the record. This interval is able to express as a insensitiveness. It value tell us, on which interval has to be a local maximum valid, in order to be admit and stored its position. Inspiration for this algorithm was human perception. When we take a look at a graph of PSD (Fig. 2. Left down), there are obviously evident particular maximums. When some simplification, we can say, that perception of particular extremes depends on its position compared with other values. If we have two sharp local maxims side by side, it is possible to ignore them, even if their amplitudes are very high. On the other side if the local maximum is alone and even if has much lower amplitude with compare to global maximum, is perceived sharply. Obtaining relevant peaks from PSD is key process. It is only its position, which guarantees correct learning of artificial neural network, eventually back detection and indication of the result. In present time is to every local maximum assigned a flag, which tells how long current maximum is a maximum. In future the algorithm will be extend into possibility to work in narrow band and choose local maximums more precisely than now. It is important to remove ubiquity noise from signal record. To make this possible, local maximums with amplitude lower than 7% (empirically discovered value) of global maximum were removed (Fig. 2). After this final touch, the picture become clear and readable and could be process to another processing with the help of neural network. Used filter is still improving and shortly will be able to reach even better results.
Obr. 3. Postup p i zpracování ulo ených dat Obr. 4. Fig. 2. Progress when processing stored data On figure 2. we can see progress when processing gained data. At the beginning can be seen a series of captured vibrations. These are further extracted into the sections with same length, where the beginning is defined by start of the strike. There was used a simple algorithm, which passed through the record and search for first value, which overtop the stable value of the record. Algorithm is presented in (1). if abs (f (x-1) - f (x) ) > 1 (1) If the condition is come true, the values x are representing strike s starts and are starting value when further data processing. After the starts extraction conversion from time domain into frequency domain has passed. Power spectrum density of pulses can be seen on figure 2. (down left). One disadvantage is, that this PSD contain abundance of relevant data. These data were removed with the help of algorithm (2), where represent lower resolution limit. if PSD (x) < maximum (PSD) * (2)
Suitable selection of variable we can affect number of peaks (Fig. 2., right down), which serve as a learning group for subsequent neural network. Measuring chain scheme can be seen on figure 3. Vibrations are sampled with the help of accelerometer from Analog Devices with type labeling ADXL001. This sensor excels with his high resonant frequency, which exceeding 20kHz. There are plenty of other sensors on the market which are cheaper, but most of them are suitable only for sampling frequencies about 600Hz max, so they are intended for mobile applications. These accelerometers are for our purposes absolutely unsuitable. Obr. 5. Schéma m ícího et zce Obr. 6. Fig. 3 Measuring chain scheme Modified values from each pulse are saved into the file and present to artificial neural network as a learning pattern, eventually as a data suitable for detection. To neural network realization, was used software Statistica by StatSoft. This analytic software is primary intended for data mining, so to obtain potentially useful data from data file. Potential and possibilities of neural network are used to make this possible. Data suitable for learning are presented to network and it will use propped method (backpropagation, metoda sdru ených gradient, Levenbergova-Marquardtova metoda and so on) to learning. The learning is basically a setting of weights of particular neurons. After this are presented to network data patterns, and it react by back propagation algorithm, to which group the data patterns belong. CONCLUSION When back data evaluation is running, the learned network reacts with accuracy about 70 80%.So it is able to tell with relatively high accuracy in which state the model is. The processing filter is still improving. We are constantly increasing number of relevant data and increasing signal to noise ratio. This motivated us to measure real industry based operation data. In present time we are working with 3 hours record, which indeed contain high noise ratio. Huge data amount and consequently high count of power spectral densities (thousands) require new progressions and diagnostics methods. Nevertheless this type of non-destructive diagnostics appear as perspective and with consecutive improving of evaluation algorithms and filters we ll be able to evaluate current stat of the crystallizer eventually overcome to creation of emergency conditions.
This publication paper originated thanks to financial aid of Ministry of Industry and Trade of the Czech Republic grant-project solution, TIP registration number FR-TI1/319 Development of New Progressive Tools and systems of Dependability Control Support of Primary Cooling on Slab Device of Continuous Casting for Duality Improvement of Demanding Flat Products. BIBLIOGRAPHY [ 1 ] DAVID, J., HEGER, M., VRO INA, M., VÁLEK, L. Visualisation of Data Fields. Archives of Metalurgy and Materials, volume 55, issue 3/2010, p.795-801, ISSN 1733-3490. [ 2 ] DAVID, J., FRISCHER, R. Development of New Progresive Tools and Systems for Quality Impoving of Dependable Flat Products on Slab Device of Continuous Casting. Acta metallurgica slovaca konference, Ko ice, vol 1., 2010, p.162-166. ISSN 1338 1660. [ 3 ] VRO INA, M. a kol. Vyu ití znalostních systém v ízení údr by metalurgických za ízení se zapojením pr b né diagnostiky do e ení. Záv re ná zpráva grantového projektu 106/05/2596 za období 2005-2007, V B-TU Ostrava, 2008 [ 4 ] DAVID, J. a kol. Etapa 2 - Identifikace diagnostických veli in a vývoj diagnostického systému (pr b ný materiál o e ení etapy projektu v programu TIP).V B-TU Ostrava, 2010/12.