BASICS OF VIBRATION MONITORING FOR FAULT DETECTION AND PROCESS CONTROL

Wilfried Reimche, Ulrich Sdmersen, Oliver Pietsch, Christian Scheer, Fiedrich-Wilhelm Bach
University of Hannover, Institute of Materials Science, Department of NDT, Lise-Meitner-Str.1, D-30823 Garbsen, Germany suedmersen@iw.uni-hannover.de Abstract. Recent years have seen the rise of vibration problems associated with structures, which are more delicate and intricate, machines, which are faster and more complex, and production processes, which are automated and interlinked. The occurred problems are direct related with demands of lower investment, running and maintenance costs in coincidence with the requirements of increased productivity and efficiency. Consequently, there has been the necessity for a better understanding of vibration causes and the dynamic response of complex machinery as well as single components. Shown are basics of comprehensive vibration analysis based on an aimed instrumentation of the unit to be supervised and the state of art in monitoring using statistical time values for general signal description and threshold comparison, envelope analysis and spectrum, phase and correlation analysis of multi sensor arrangements. Hints are also given to data reduction using cepstral and hocerence analysis in combination with the extraction of machine or fault specific characteristical patterns as input values of vector or neural classification. Keywords. signal processing, vibration analysis, vector analysis, condition monitoring

1. Introduction The use of vibration analysis as one of the fundamental tools for condition monitoring has been developed extensively over a period of approximately 35 years. With the parallel developments in electronic equipment, transducers, computers and software nowadays machine supervision is almost completely automated. From 1960 to the mid 1970s simple practical methods were used, along with careful watch on the machines behavior, often reinforced by frequent maintenance. Elementary instruments were sometimes used to measure and record the values on which failure detection and concluded maintenance decisions were based. This required highly skilled and experienced maintenance staff to ensure efficient operation and to avoid catastrophic failures. During the 1970s there were basic developments in analogue instrumentation and mainframe computers. Several accelerometers, velocity transducers and displacement pickups were developed and adapted to the requirement of industrial applications. Standards and guidelines give information about their use and present a first knowledge base for failure description and determination. Analogue instrumentation became popular in form of portable vibration measuring and recording meter, FM tape recorders and frequency analyzers. Many of these equipments were heavy and cumbersome but they did prove increased accuracy in recording and analysis. If a company had access to a mainframe computer, data would be kept in store so that first maintenance strategies could be developed, still limited due to the amount of available measurement and analysis channels respectively the data store media. Although some digital instruments were available during the early 1970s, significant developments took place during the late 1970s and the early 1980s due to the availability of new microprocessors. Circuits could be miniaturized, reducing the dimensions and weight of instruments and allowed data to be handled at high speeds. Onboard microprocessors gave instruments the ability to capture data, analyze it via suitable algorithms, then store and display the information. A very significant feature of frequency analysis was efficient computation of the FFT (Fast-Fourier-Transformation) of multi channel measurements within seconds respectively minutes and the ability to store data for further decisions. Long-term data storage became an accepted practice. From the mid 1980s onwards the developments have been associated with the desktop computer, its interfaced equipment and software. Many manufacturers have produced hand-held instruments for the instant measurement, recording and analysis of variables. Information is often directly available from the measurement instrument on a machine component (e.g. standard deviation, Smax). This makes it easier to decide upon maintenance strategies, therefore, the decision maker requires a lower level of skill. Indeed some companies have developed suites of software, which allow the whole process and the actual machine condition to be monitored automatically, giving a complete service for measurement, analysis and problem diagnosis followed by action or maintenance strategies. The 1990s are characterized by minimizing the instrumentation and the data acquisitions and data process units. The instrumentation is simpler to operate, and problems can be analyzed with aid of knowledge-based systems. Sensor systems are combined with electronic data pre-processing and multi-sensor arrangement allow not only to emphasize on single machine components, the whole plant with the interactions and dependencies of all processes becomes transparent. At the mid of the 1990s the remote monitoring was initiated with central control units to interlink the knowledge and experiences of customers, the machine manufacturer and liability insurance companies. In contrast, early condition monitoring systems relied on a few experts having specialist knowledge of the instrumentation and analysis techniques. While vibration analysis was used former mainly to determine faults and critical operation conditions. Nowadays the demands for condition monitoring and vibration analysis are no more limited trying to minimize the consequences of machine failures, but to utilize existing resources more effectively. Some factors which have increased these demands include:

02 a 06 de Junho de 2003 / June 2 to 6 2003 Rio de Janeiro - RJ - Brasil

* * * * *

increased quality expectations reflected in product liability legislation, increased automatization to improve profitability and maintain competitiveness, increased safety and reliability expectations reinforced by legislation, increased cost of maintenance and production due to labor and material cost, and increased market pressure due to globalization of markets.

Actually developments still deal with minimization of measurement equipment and analysis techniques implementing worldwide standards for data processing and acquisition, with the possibility of central data acquisition. The supply of more cost-effective monitoring tools has been made possible by technical advances such as: * * * * reduced costs of instrumentation, increased capability of instrumentation such as data pre-acquisition, data storage, radio transmission direct by the sensors with integrated electronical circuits, improved data storage media in combination with low cost computation, and faster and more effective data analysis using specialist software tools.

The exponential increase of computer capacities, calculation speed and the telecommunication allow new analysis techniques to be implemented in modern process monitoring systems like fuzzy logics and neural networks as tools for automatical system control and signal classification. The ultimate form of this approach is an expert system, which eventually may lead to a fully automated condition monitoring system including additional rest life estimations and affirmation referring to the product quality. Nowadays modern vibration monitoring systems often suggest the possibility of faultlessness. As results often failures and alarms occur without giving hints to the original excitation sources. Therefore this paper describes the practical aspect of vibration phenomena and the measurement requirements of a general monitoring system consisting of data collection with data reports in digital manner, followed by the acquisition phase calculating the statistical values and functions in time and frequency domain with integrated data reduction by fault and operational pattern. The more difficult third phase of automatical fault diagnosis is still under development. New techniques like wavelet analysis and neural networks are examined and actually tested. Shown are basics of comprehensive vibration analysis based on an aimed instrumentation of the unit to be supervised and the state of art in monitoring using statistical time values for general signal description and threshold comparison, envelope analysis and spectrum, phase and correlation analysis of multi sensor arrangements. Hints are also given to data reduction using cepstral and hocerence analysis in combination with the extraction of machine or fault specific characteristical patterns as input values of vector or neural classification. In cases of transient vibration signals, e.g. forging, oscillating processes, etc., the traditional methods of data acquisition and processing at rotating machines or stationary processes have to be adapted using certain signal trigger or sample techniques as well as special evaluation routines (e.g. the data acquisition system has to be transformed from an amplitudetime or rotation description to an amplitude-distance or amplitude-event-description. 2. Excitation sources and sensors The aim of all plant operation and monitoring systems should be the threshold of 100% machines availability during operational time by reliable systems and fast reaction fault management. The mission should be the development of integrated strategies, methods and tools of planning and operating production systems by the vision of guaranteed availability and quality. The state of art of investigations for early failure detection emphasizes mainly on single system components of production machines, e.g. rolling element bearings, meshing gears, using vibration and acceleration analysis techniques. Actually extensive research work is done on the field of diagnostics of process data and machine control. Using this basic knowledge, actually, the global consideration of process lines has to be looked at, leading to new ways in the field of transient vibration signal analysis. The knowledge of the machines vibrational signatures and their time dependent behavior are the basics of efficient condition monitoring. A complex production line consists generally as shown in figure (1) of several driving units (e.g. electrical or hydraulic motor, turbine), a transmission unit (e.g. gearbox, coupling), and a processing machine (e.g. generator, compressor, pump, tool, press). Vibrations of machines are the results of the dynamic forces, due to moving parts and structures (e.g. foundation), which are interlinked to the machine and its mechanical properties. Different machine parts will vibrate with various frequencies and amplitudes. All these components generate specific vibration signatures, which are transmitted and reflected in the machines structure. Machine condition, machine faults and on-going damage can be identified in operating machines by fault symptoms, e.g. mechanical vibration, air borne noise, and changes in the process parameters like temperatures and efficiency. In order to optimize fault specific thresholds and determine diagnostic patterns in vibration and acceleration signals at rotating machines, failure simulations at different test benches are performed, e.g. rubbing simulation in turbo machines, investigation of rolling element and journal bearings, centrifugal pump and turbine supervision, as well as cavitation monitoring. The obtained signatures serve as a base to determine fault and operation specific signatures of rotating machines in power plants as steam turbines, boiler feeder pumps, mill fans, steam valves, etc., as well as in industrial application of

quality control systems and production lines. To close the loop of investigation finite element calculations are applied to estimate the vibrational behavior of structure elements or prove excitation sources by fluid flow simulation.

Figure 1. Vibration excitation sources in general and measurement devices To the fulfillment of the demands on comprehensive vibration analysis, an aimed instrumentation of the unit to be supervised is required whereby displacement, velocity and acceleration pick-ups are used. Mathematically these basic descriptive values can be transformed one to the other by integration respectively by differentiation. Practically the choice of sensors depends on the excitations to be determined. Accelerometers show a constant amplitude behavior over a wide frequency range, therefore they can be used to detect low as well as high frequent signal intensities, while velocity transducers show a 1/f (f = excitation frequency) proportional amplitude damping and displacement pick-ups even damp the amplitudes by 1/f. Looking at the balance condition of rotors this effect is favorable due to the automatical reduction of high frequent signal amplitudes due to friction, gearing, etc. which superimpose the lower frequent rotational speed component. In case of rolling element bearing defects, which mainly excite high frequent pulses due to the interaction of rolling elements and raceways, accelerometers fit best. Summarized the sensor choice largely depends on the frequencies to be analyzed, where acceleration in general covers frequencies from 0 to 20 kHz, velocity from 2 Hz to 2kHz, and displacement the range from 0 to 300 Hz. Many techniques exist, which can be used either singly or in combination to identify vibration patterns. Components typically monitored are gears, bearings, couplings, rotors, fans, pumps, turbines, etc. to determine and distinguish several failure causes like imbalance, looseness, misalignment, wear, poor lubrication, general damage, aero/hydrodynamic forces, cavitation, etc. Finally, the response of the vibration monitoring system to the onset of failures and unfortunate operation conditions depends basically on the used instrumentation. The faster the response is required, the more likely a hard-wired system is suggested rather than portable data collectors. Where the consequence of failures is higher, the easier it is to justify the costs, e.g. a big steam turbine normally is monitored permanently while the auxiliary equipment (e.g. vacuum pumps) may be inspected by portable units.

Figure 2. Vibration instrumentation at a steam turbine Dependent on the machine or the single component to be monitored a wide spread of sensor and system applications are available form handheld meters of one or two channels with direct data evaluation to complex multi channel supervision with direct link to the main host process monitoring like one example is shown in Figure (2) by the vibration instrumentation of a 920 MW steam turbine consisting of eddy-probes to determine the relative rotor displacement and expansion,

velocity transducers to monitor the bearing block vibration, and optional accelerometers to measure high frequent excitations during rubbing events. In this case additional accelerometers and displacement pick-ups inside the combined radial/axial bearing between HP-IP-parts should help to solve and prove thermal and expansion related alignment problems. Independent on how sophisticated the equipment may be, the quality of the measurements and their interpretation are until now still dependent on the ingenuity of the operator, whereby the data acquisition of vibration signals is considered in the four sections of: * * * * basic electrical and electronical principles, signal types, interference and transmission, elementary amplifiers and filter design, and vibration signal acquisition and processing on which this article emphasizes on.

3. Data acquisition and data processing State of technology in vibration monitoring of rotating machines is related to the calculation of standard deviation and/or maximum values, their comparison with thresholds and their trend behavior to determine increased wear or changes in the operation conditions. Spectrum analysis with special phase constant averaging routines allows to determine machine specific signatures by magnitude and phase relation. By correlation analysis common information of different vibration signals is evaluated for source localization, and cepstrum analysis is used to quantify periodical information of spectral data.

Figure 3. Vibration excitation sources and their statistical time values A general overview about exemplarily processed statistical time values dependent on the shape of the vibration signal is visualized in Figure (3). Aim of these values is to quantify the visual impression of the time signals as demonstrated here

for local and distributed faults at gears and rolling element bearings. Based on the large statistical variation width in the pure maximum value an integral description of the dynamical signal information is given by the effective value also called standard deviation (STD), squared the variance presents the dynamic signal energy. The Crest-factor describes the peak intensity in the signal progress, calculated by the relation of the maximal value to the standard deviation. Through the fourfold exponential weighting of short-time pulse excitation against the energy contained in the signal, the relative load independence of the Kurtosis-factor () arises, in contrast to the characteristical values described above. By definition the Kurtosis-factor reaches for purely sinusoidal excitation a value of 1.5 (e.g. pure unbalance excitation), in gaussian rustling a value of 3.0 (e.g. stochastical mechanical/flow friction), while for short time pulses within the time set (e.g. short time rubbing, local rolling element bearing damage, cavitation) values up to 100 can appear. The form-factor, as ratio of the quadratic average (RMS, root mean square value) to the linear average, applies as an analysis tool to share harmonics from coupled information in the signal, with a value of 1.11 for purely harmonical signal components. Stepping up additional damage specific non-harmonical related signal components in the observation period, or the harmonically components shift themselves, this reflects as changes of the form factor. The single calculation algorithms are to be taken from special literature. As a further remedy to the comparison of conditions or signal progresses the correlation coefficient  is available as applied here to describe changes of probability functions for the two shown faults. It describes whether in the mean of a signal order and/or an observation time interval two time signals are independent from each other, if covariance or contravariance exists. Calculated from the linear average of the product from both observed time signals, normalized by the product of the accompanying standard deviation, it compares obviously the changes in the signals shape. To the premature determination of faults and damages of single machine components, the trend setting of certain statistical time values of the vibration signals suits itself. So the initiating and growing of defects in rolling element bearings of pumps, compressors, and drive units in most cases are clearly detected by using statistical values as should be demonstrated by the above presented examples. Due to the regarded results the first impression might lead to conclude that time domain analysis by trend setting of statistical values is sufficient to determine faults successfully. Near problems with source localization in complex machinery, the example shows that the signatures for the gear faults here are similar to that ones of the bearing fault and superimposed structure resonances often influence the pure time signal analysis.

Figure 4. Vibration excitation sources of gear wheels (time signals, amplitude spectra and frequency values) The single frequency components contained in the time signals can only be determined by amplitude and phase relations of frequency analysis using spectra, coherences and cepstra. To the use of automated routines for failure determination and diagnostic purposes the occurring data quantities have to be reduced to single values as the amplitudes of

characteristcal frequency components, like the narrow banded spectra information of rotational related excitation as unbalance, misalignment, gear mesh, etc. The cepstrum, the inverse FFT of the spectrum, summarizes all periodical information within one value, e.g. the amplitude and frequency modulations of bearing, gear, and blade faults become visible at increased gamnitudes of the speed rahmonics. If faults at rotational speed related machine components occur, these modulations are visible at these rahmonics. The combination of excited center frequency, for the fault example the teeth mesh (TM), and the cepstrum gamnitudes (Cn) at the 1st speed rahmonic characterize the fault in Figure (4). Broad banded information of system and fluid resonances are monitored by FFT-integrals in certain frequency ranges. Friction related information is represented in the spectra ground level (Ilog). Multi sensor systems in combination with correlation techniques avoid measurement errors and determine excitation sources. Similar to the cepstrum analysis the hocerence (inverse FFT of the coherence) enables to summarize all periodically related information obtained in two signals. Likely as in case of the statistical time values different faulty operation conditions could have similar shapes of spectra. Especially for complex machine arrangements with multiple step gears, several rolling element bearings, the difficulty is to separate clearly the different speed related excitations from structure or fluid induced components. A common method to identify speed related information is time averaging of the vibration signals using external signal triggering (e.g. by strobe light, laser, decoder). The influence of an external trigger with time domain averaging to the vibration signals shows Figure (5) at the example of three different fault patterns. After time averaging the signal amplitudes of acceleration are reduced significantly, all stochastical signal components are not visible anymore. Only speed related information is obtained, emphasizing the fault pattern for eccentricity and defect gearing. The apparent periodical excitations of local bearing faults are proved as non harmonical speed related. Therefore, time domain averaging in combination with statistical time values permit a first classification of fault patterns.

Figure 5. Influence of time averaging If those trigger units are not installed, in case of multi shaft arrangements, or even for just in time measurements phaseenhanced data processing is applied as shown in Figure (6) at the example of source localization at the complex rolling element bearing arrangement at fresh air fans. The auto power spectral density of a non-averaged time signal includes the amplitude information of all frequency components, while the phase relation is lost. Fault specific patterns as the teeth mesh frequency with modulation, the blade rotation sound as well as the characteristical bearing frequencies are characterized by constant phase relations. Choosing one of these machine specific components as reference, relating all phase relations to this component, the new calculated averaged spectrum only includes the related phase constant frequency components. The upper diagram in Figure (6) shows the amplitude spectra of acceleration measured at the casing of a fresh air fan with a complex bearing arrangement. The averaged spectrum includes many narrow banded speed related information as well as broad banded system related excitations. To separate all these information becomes quite difficult. If phase enhanced spectrum calculation is applied using the blade rotation sound (number of blades x speed = BRS) all driving speed related information is obtained, just as would be seen by time averaging. Applied here, the signal ground level drops down, while also the narrow banded speed related modulations are reduced, hinting to other main excitation sources than pure speed related ones (lower left spectrum in Figure (6)). If one of the bearing characteristic frequencies is chosen, e.g. the ball rotating frequency outer race (BPFO) from the cylindrical bearing NU348 similar results are obtained (upper right spectrum in Figure (6). But choosing the BPFO of the rolling element bearing FAG7252B nearly the original shape of spectra is obtained, proving the main excitation source, as well as for the broad banded friction related peaks visible at the signal ground level, as for the narrow banded speed

related modulation in the lower right spectrum in Figure (6). Further data reduction by calculating the corresponding cepstra allows to determine excitation patterns.

Figure 6. Applied phase enhanced data processing 3.1 Determination of fault patterns and their classification In order to detect all mentioned failure mechanism and their effects to the characteristical values a further step is the calculation of vectors. Simplified a matrix of single values, e.g. the statistical time values, certain machine specific frequency components (e.g. teeth mesh, blade rotation sound, etc.), and the corresponding cepstrum components like shown in the next figures. A cross-link with a weighting function can emphasize certain fault patterns. For some defected gear wheels with fracture and pitting examples of the input values for vector analysis are visualized in Figure (7) for different fault conditions, normalized to an averaged reference. Changes in the operational behavior become directly visible and experienced operational staff is able to determine different classes of defect. Related to complex machinery the information has to be compressed further, here several single components interact and the obtained information is not only related to one sensor signal.

Figure 7. Failure descriptive values for gear faults According to the type of defect, the vectors produce the largest correlation for failure and grade of damage. The distance and angle between the vectors present an easy classification and consider the correlation between the measuring vectors and the prototype or reference vector. The simplest type of classification is reached by the minimal analysis of distance, magnitude and angle between each of them. As result an actual condition based cycle is obtained in Figure (8), which classifies the actual machines condition, the resistance against wear, and the distance to unfortunate operation. Certain more complex algorithms for classification have to be applied, if the single failure classes are not clearly separated. This principle of signal classification is also used in the field of eddy current measurements for failure detection and material classification, ultrasonic testing or automatical image processing of x-ray inspection. Also here the correlation with process parameters increases the probability of secured fault detection and description. In a further step the vector itself might fit as an input of a neural classification and interactive pattern recognition system which is able, if sufficient samples of reference and failure classes are available, to determine faults automatically and initiate fast control action. A further step to determine different fault signatures automatically is the implementation and development of expert systems using neural networks. Historically, the interest in neural networks has two roots: the desire to understand the

principles the human brain works with, and the wish to build machines that are capable of performing complex tasks with automatical learning.

Figure 8. Definition of vectors and possibility of graphical visualization Therefore, neural networks could be defined as algorithms for cognitive tasks, such as learning and optimization, which are in a loose sense based on concepts derived from research of the nature of the brain. Simplified, they consist of several nodes in certain layers which are interlinked (the links are called synapses), back propagation may exist. The transfer function for each node is either a discontinuous step function or a smoothly increasing generalization known as a sigmoidal function. Nodes without links towards them are called input neurons, output neurons are those with no link leading away from them. The idea is to use the vibration signal as input of the neural network, which then tells the action the operator has to do, or even perform the action itself. In these cases the neural network is used like a black box, learning different failure patterns during operation. Here is the start of the problem, due to the fact that enough failure patterns must be available during the time of learning of the neural network, which is quite time intensive. Otherwise the network will fail. With the complexity of machine structures, and that all machines show their own significance until now only limited practical application is done of these technique for failure detection. In regard of the automatic failure detection and diagnosis of the neural network, timesavings and high precision are achieved, if the network is operating properly. As a new analysis technique to determine fault patterns the Wavelet analysis provides a method of decomposing a recorded signal into a family of component parts. For general signal analysis, the objective is usually to extract frequency data from the signal and learn how its frequency composition changes with time. The wavelet method provides a good mean of doing this because each wavelet has a particular frequency content and is located at a particular position on the time axis. There are many ways of choosing the form of the analyzing wavelet. The theory is quite complicated. In general this analysis method could be looked at as a type of filtering signal information by certain window functions to emphasize occurring high frequent and low frequent excitation with the same amplification to determine even lowest signal intensities in the analyzed signals, even if they only arise for a short time. This technique is applied successfully for the evaluation of video signals with low signal to noise ratio. The application for the determination of transients in vibration monitoring and acoustic emission is under development. 4. Case study: motor drive bearing fault of a fresh water pump Important for failure source localization is the correlation of the vibration signatures with the process parameters of the machines in operation, due to the fact that the mechanical property of structures changes significantly with the change of temperature, the flow rates are direct correlated to the friction and interaction of blade arrangements, as well as mechanical forces influence the interaction of gear steps. The example demonstrates the necessity of including process parameters for source localization to determine the source of bearing faults at fresh water pumps. The condition monitoring system of the six 2.2 MW fresh water pumps of a 1450 MW nuclear power plant showed several times an increased standard deviation of vibration velocity at the electrical engines for certain pump arrangements. It could be stated that wear of the rolling element bearings was the excitation source, at bearing operating hours quite less than the nominal ones. The vibration monitoring system always determined the increasing faulty operation conditions in time to operate the pump until the next planned shut down (normal operation cycle one year). Additional acceleration measurements at several positions at the pumps and the foundations should prove the motors or pumps as source of increased vibration. By determination of the characteristic vibration signatures of each system component, the vibration transmission between the motors, the gears and the pumps themselves, and the interaction between the operating pumps should be investigated. The evaluation of the vibration signals was performed with a 16 channel data-acquisition and processing computer system using time domain, frequency domain and correlation analysis with trend settings.

Figure 9. Spectra of acceleration at motor drive unit Temperature dependent expansions of the motor foundation due to uncontrolled ventilation superposed with resonance excitation were stated as source of the increased vibration amplitudes. A schematical drawing of the pump arrangement consisting of electrical motor, planetary gear, and one stage centrifugal pump is presented in Figure (9). The vibration measurements were carried out at each pump and foundation for several frequency ranges up to 50 kHz. First investigations showed a high amount of sidebands with the distance of twice the motor slip frequency at the harmonics of the rotational speed in the spectra of acceleration at some motor bearings, mainly at VC60, proved by cepstrum analysis. Observing the trend of these signatures no stationary results were obtained for constant load conditions, in some cases the sidebands occur in others not. Additionally, significant differences in the signal ground level could be stated characterizing the amount of mechanical friction. Misalignment between the corresponding electrical engine and the gear due to temperature gradients in the 33 m long foundation could be investigated as main influence to the vibration amplitudes (increased second harmonic of rotation). Depending on the direction of wind, climate conditions and the fresh-water temperature the foundations are influenced by the convection as demonstrated in Figure (10). The middle and lower foundation showed always nearly homogenous temperature profiles, while the motor foundation is characterized by significant changes in the temperature gradient. This is the source of relative expansion and misalignment exciting the slip frequencies as seen in Figure (9). A second source of exceeded vibrational excitation is a system resonance (RE) of the concrete motor foundation, which is in resonance with the second harmonic of rotation, also depending on the mechanical restraining forces. To reduce the vibration excitations the radial flexibility of the coupling between motor and gear at one pump was increased during the next planned shut-down. Additional support of the motor foundation is suggested to shift the system resonance of the concrete motor foundation.

Figure 10. Temperature profiles at motor and pump foundation

5. Summary The experiences obtained by monitoring several machine arrangements in power plants as well as in production industries prove the successful use of aimed vibration monitoring for fast failure source localization and process optimization at machines in operation. Most installed vibration monitoring systems serve for threshold comparison and alarm monitoring. To get the alarm is state of art, to quantify and classify the obtained information is the second step implemented in modern process control systems. To fix the excitation source and to develop reaction strategies with short time delay proving the success of actions still requires certain expert knowledge. As proved by several industrial projects the interlink between vibration analysis and the process parameters represent a fast and reliable tool for condition-based description of machines in operation. Aim of all investigations is to realize a partial share for the economic/ecologic production within an optimized production process and high performance availability. Therewith, the plant management system, the possibilities of supervision and the condition-based maintenance must be interlinked, combining the tools of advanced data processing, evaluation, documentation and data visualization as one of the most important point of view for system acceptance by the operational staff. On-line process monitoring is beneficial for maintaining high quality products at high production rates and low costs fixed by the terms of availability, operational reliability and service life. Therefore the main task of a vibration diagnostic system is the processing of necessary measured data to identify, or at least to limit, the cause of damage and disturbances in rotating machines and production lines, whereby two methods of fulfillment for this task are offered: the manual diagnostic, where the specialist has to extract significant information from the vibration signal, his knowledge is of main importance for the correct trend setting of the results, conclusions about fault patterns and actions to save the machines life. the automatical diagnosis, where the knowledge of the specialist has been successfully transferred into an expert system to realize an automatic interpretation of the measured data. The task of the expert system is not to replace the specialist, but rather to provide support to speed-up his diagnosis. It could help for automatical determination of simple problems displaying possible causes of damage and their probabilities. It can recognize and avoid critical operational conditions when short reaction times are necessary and there is no time to consult experts. And finally, it provides assistance to an expert in solving complex problems. Practice has shown that vibration monitoring is, from an economic point of view, a profitable strategy. The investment costs can be amortized already after the prevention of only one unscheduled downtime of the machine. New technical requirements will grow out of this demand and will lead to further developments in present concepts for hardware and software modules which can be subdivided among others by the following concepts: * * * improved expert systems with higher level of diagnostic authoritativeness and greater diagnostic accuracy, improved infrastructure by more powerful performance information and control systems to interlink planning, start-up, production, and maintenance, build up of computerized information centers to interlink the experiences and investigations of machine manufacturer, the operator in industrial production, also including the related insurance companies,

with the ideal threshold of a planned 100% machines availability during the fabrication cycle. 6. References Barber, A., 1992, Handbook of noise and vibration control, 6th edition, Elsevier Advanced Technology Publications, UK Sdmersen, U.; Pietsch, O.; et.al, 1998, Advanced monitoring of power plants by vibration analysis, CETIM98, Senlis, France Newland, D.E., 1993, An introduction to random vibrations, spectrum and wavelet-analysis, New York, Jon Wiley, USA Barschdorff, D.; Kronmller, M., 1995, Acoustic quality test and fault detection of automobile gearboxes using artificial neural networks and fuzzy logic, 18. ISATA, Stuttgart, Germany Barron, R., 1996, Engineering condition monitoring, practice, methods and applications, University of Strathcycle, Glasgow, Longman Kuo, R.J., 2000, Multi-sensor integration for on-line tool wear estimation through artificial networks and fuzzy neural networks, Engineering Applications of Artificial Intelligence, Vol.13, part 3 Sdmersen, U.; Pietsch, O.; Scheer, C.; et.al, 2002, Failure source localisation at turbo sets by combined vibration anlysis and process parameter analysis, ISROMAC 9, Honolulu, Hawaii, USA

7. Copyright Notice The author is the only responsible for the printed material included in his paper.