European Wind Energy Conference (EWEC 2010)

Detecting Incipient Wind Turbine Gearbox Failure: A Signal Analysis Method for On-line Condition Monitoring
C J Crabtree, Y Feng, P J Tavner New & Renewable Energy Group School of Engineering and Computing Sciences, Durham University, UK. c.j.crabtree@durham.ac.uk, yanhui.feng@durham.ac.uk Quantitative studies of wind turbine (WT) Abstract: reliability from existing public databases have recently been carried out [1][2] and have Condition monitoring of wind turbines is shown that WT gearboxes are a mature gaining importance as turbines become larger technology with a constant or slightly and move to more inaccessible locations, such deteriorating reliability with time. However, as offshore. Condition monitoring based on surveys by WMEP and LWK [3] have shown methods conventionally used in the power that the gearbox exhibits the highest downtime generation industry have been demonstrated per failure among all the onshore WT to work successfully on large wind turbines subassemblies. Figure 1 shows the when attention is paid to data collection. In comparison of the failure rate and downtime of view of the large number of wind turbines different WT subassemblies from these deployed this paper proposes a methodology surveys in the EU. The offshore reliability and for wind turbine condition monitoring that downtime data of Egmond aan Zee wind farm compares conventional condition monitoring [4] has also shown similar results, where the signals with operational signals, such as load gearbox failure rate is not high but the or energy, which could be applied downtime and resultant costs are high. The automatically. A multi-parameter approach, percentage of electricity production lost due to based on comparison of independent signals, the gearbox downtime in that study is the should increase confidence in fault signal highest of all subassemblies. interpretation and alarms generated, potentially reducing the risk of false alarms. This is because the gearbox repair procedure is complex, particularly offshore, requiring not only special equipments such as a support

1 Introduction

Figure 1: Failure/turbine/year and downtime results from European wind turbines

vessel and crane, but the maintenance work may also have to be carried out under unfavourable weather conditions. Public databases have not identified information on WT gearbox failure mechanisms. In the EU funded project ReliaWind, the consortium is developing a systematic and consistent process to deal with the detailed commercial data collected from wind farms, including 10-minute SCADA data, automated fault-logs and O&M reports to identify and understand WT gearbox failure mechanisms in more details [5].

characteristics and WT condition monitoring is more akin to the monitoring of auxiliary plant than of constant speed conventional power generation plant. Oil debris particle count would normally be collected on a time basis [7][8], neglecting WT operating condition, masking trends due to low wind speed turbine stops. Vibration amplitude trends collected on a time basis also appears noisy and disjointed due to high and low load conditions. Experienced WT Operators are able to use conventional analysis methods successfully by taking care in considering the operating conditions during data acquisition. Another issue is CM interpretation when relying on a single signal or analysis method. This can lead to false alarms from the signal monitoring process. Tavner [9] suggested a multi-parameter approach for electrical machinery, where the relationships between different signals could be used to increase detection confidence and reduce false alarms. This paper proposes applying this approach to a WT, using vibration and particle count signals from a gearbox compared with WT cumulative energy produced. An industrial partner, an experienced WT Operator, has given access to CMS signals collected from a WT operating with an incipient gearbox defect. The proposed method has been applied, after the event, to this data. If successful, the method could be developed into an on-line tool to detect incipient gearbox damage automatically, producing an alarm signal, with associated Severity Factor, in which there could be a high degree of confidence, reducing the risk of false alarms.

2 Condition Monitoring of Gearbox

High replacement costs, complex repair procedures and revenue loss due to long downtime are therefore the major reasons for concern about gearbox reliability amongst WT Operators. An effective condition monitoring (CM) programme can detect significant health condition changes of WT subassemblies at an early stage. Gearbox monitoring has become an essential part of on-line WT CM. Vibration and gearbox oil debris particle count monitoring are broadly accepted in the WT industry and specified as essential measurements in some certification guidelines[6] and experienced WT Operators have successfully used these techniques to detect incipient gearbox damage, avoiding gearbox failure. In order to apply conventional time and frequency domain signal analysis methods successfully, for the detection of WT gearbox faults, it is important to ensure the signals are recorded under consistent conditions. This is necessary because of differences between WTs and conventional rotating machines, such as steam or gas turbines. First a WT is subject to large, rapid load fluctuations, second it may operate at variable speed and third there may be substantial differences between the behaviour of WTs of nominally the same type at different sites. Therefore, WT condition monitoring signals collected on a continuous basis exhibit large variation and non-stationary

3 Background
The WT under study was a two speed, active stall machine equipped with a SKF WindCon condition monitoring system (CMS) recording operational and CM signals including WT load, four gearbox vibration signals and a gearbox oil debris particle count sensor. The paper focuses on one particular gearbox axial vibration signal, from the high speed end, and the 50-100m ferrous particle count.

It should be noted that similar trends to those presented here had been observed in other vibration signals and particle counts from this WT but for conciseness are not included here. In this incident the Operator had successfully used the SKF WindCon system to detect an incipient fault, using fast Fourier transform vibration analysis and oil debris particle counts as conventionally used in the power generation industry, and had consequentially rectified a bearing problem in the WT gearbox. This paper uses one segment of the CMS signals before the incident to develop the proposed method, after the event, with the intent of helping the Operator to improve failure detection accuracy in the future.

Figures 2(b) and 2(c) show the next two consecutive periods and we observe an increase in vibration amplitude and variability. The wider scatter in vibration data suggests the presence of a gearbox disturbance, transmitted to the gear case through a strong transmission path. Figure 2(c) in particular indicates possible gearbox damage. As we move further forward in time we see, in Figure 2(d), a further change in the vibrationload characteristic, a decrease in gear case vibration, albeit with variable data. This vibration reduction continues through Figures 2(e) and 2(f). This decrease does not necessarily indicate an improvement in gearbox health but may suggest a worsening. In this case the damage has become more serious to the point that the transmission path has been degraded and the gear case sees reduced vibration amplitude, despite increased damage in the bearing. The WT under discussion continued to operate for a short while longer before being shut down by the Operator for inspection, based on the vibration amplitude results. Upon inspection it was discovered that the gearbox had developed damage to an intermediate shaft bearing. The bearing was replaced during the 700-800MWh energy production period, and in Figure 2(g) a change in the vibration level can clearly be seen. Figure 2(h) shows the vibration against load characteristic following maintenance, which is useful as it shows the relationship between vibration and load for a healthy machine and could be used as a benchmark for other WTs which may have incipient damage. It can also be seen that in the light of the bearing replacement the WT had been operating with minor damage for some time. This is indicated by the level of vibration shown in Figure 2(a) which is much higher than for the machine with the new bearing, Figure 2(h). However, diagnosis can be difficult on the basis of a single signal alone. Multi-parameter monitoring could give an Operator more confidence in the observation of deterioration or change of state, which may be valuable when managing maintenance programmes.

4 Method and Results

The CMS data used in this analysis was recorded from the SKF WindCon system on this turbine during a 5 month period spanning the date of the fault repair. In order to provide a step by step demonstration of progressive gearbox damage, the high speed end, axial vibration data was split into periods of 100MWh energy production. By doing this the effect of no-load periods were eliminated and data clarity increased. The changing relationship between gearbox vibration and energy production can be observed. Eight consecutive 100MWh energy production periods are shown in Figures 2(a) to 2(h) beginning at the start of the data set. By plotting vibration against energy generated we get small, directly comparable data sets, containing similar quantities of operational data. If data points were plotted in periods of time, for example a week of operation, it would be more difficult to observe relevant changes in relationships between signals. In Figure 2(a) we see two distinct traces for vibration against load resulting from the twospeed nature of the machine. The higher amplitude points are for operation at 1500rev/min and the lower at 1000rev/min.

The gearbox of this WT was fitted with a particle counter to measure debris present in the lubrication oil. This measurement is independent of machine vibration but debris will be produced within the gearbox when damage causing vibration occurs. To remove the effects of no load from the cumulative count of ferrous particles in oil, it was plotted against approximate cumulative energy

produced, derived from the mean power signal and time, Figure 3. Any change in condition should be clearly observable from the rate of particle production. To aid comparison the axial vibration was also plotted against the same variable, Figure 3. Figure 3 shows steadily increasing ferrous particle production with energy produced during period A. This corresponds directly to

Enveloped Gearbox (High Speed End) Axial Vibration against Turbine Load
(a)
Axial Vibration (g)

Enveloped Gearbox (High Speed End) Axial Vibration against Turbine Load -100 - 200 Period 100-200MWh Period: Generation MWh
0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

(b)
Axial Vibration (g)

Enveloped Gearbox (High Speed End) Axial Vibration against Turbine Load - Generation Period 200-300MWh Period: 200 - 300 MWh
0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

Load (MW)

Load (MW)

Enveloped Gearbox (High Speed End) Axial Vibration Period: 300 - 400 MWh against Turbine Load - Generation Period 300-400MWh

Enveloped Gearbox (High Speed End) Axial Vibration Period: 400 - 500 MWh against Turbine Load - Generation Period 400-500MWh

(c)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

(d)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

Load (MW)

Load (MW)

Enveloped Gearbox (High Speed End) Axial Vibration Period: 500 - 600 MWh against Turbine Load - Generation Period 500-600MWh

Enveloped Gearbox (High Speed End) Axial Vibration Period: Generation MWh against Turbine Load -600 - 700 Period 600-700MWh

(e)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

(f)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

Load (MW)

Load (MW)

Enveloped Gearbox (High Speed End) Axial Vibration Period: Generation MWh against Turbine Load -700 - 800 Period 700-800MWh

Enveloped Gearbox (High Speed End) Axial Vibration Period: 800 - 900 MWh against Turbine Load - Generation Period 800-900MWh

(g)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

(h)
Axial Vibration (g)

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

Load (MW)

Load (MW)

Figure 1: Enveloped gearbox axial vibration against turbine load for eight Figure 2: Enveloped gearbox axial vibration against turbine load for eight consecutive periods of consecutive periodsenergy productionproduction 100MWh of 100MWh energy

the steadily increasing axial vibration amplitude in period A Figure 4. By observing simultaneous changes in two independent signals greater confidence is gained that there is a genuine deterioration in machine health.
Gearbox Oil Debris (Ferrous 50-100um) against Cumulative Energy Generated
01/08/2008 01/09/2008 01/10/2008
1000 900 800

Figure 5 shows the timings of detection with the proposed method, potentially giving up to 2 months warning before maintenance when the incipient fault could be detected. For on-line detection, the method would have to be developed into a Severity Factor to give the Operator an indication of WT gearbox health condition, which could reduce the risks of false alarms from the separate signals.

Particle Count

700 600 500 400 300 200 100 0 0 100

200

300

400

500

600

700

800

900

Energy Generated (MWh)

Figure 3: Gearbox oil debris particle count against cumulative energy production
Enveloped Gearbox (High Speed End) Axial Vibration against Cumulative Energy Generated
01/08/2008 01/09/2008 01/10/2008
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 100 200 300 400 500 600 700 800 900

Figure 5: Time to detection using the proposed method In this case the Operator had inspected the gearbox in the field, based on conventional vibration analysis methods, confirmed an intermediate shaft bearing fault and successfully replaced the component. This early detection and maintenance action forestalled a possible gearbox failure, allowing the Operator to save more than 100k before consideration of any installation or consequential downtime losses more than justifying the CMS capital expenditure. The proposal in this paper is that the multiparameter approach described here could be programmed into the CMS system to provide a reliable and timely automatic warning of the incipient fault.

Vibration (g)

Energy Generated (MWh)

Figure 4: Gearbox axial vibration against cumulative energy production The amplitude of vibration decreased as energy production progressed into period B Figure 4, as previously observed in Figures 2(d) and 2(e), corresponding to an increase in particle production observed in Figure 3. Increased particle production simultaneous with decreased vibration corresponds to a weakened vibration transmission path to the gear case, due to further bearing deterioration. Marker point X on each plot indicates the point at which maintenance was carried out by the Operator. After inspection of the gearbox the intermediate shaft bearing was replaced. Immediately following this maintenance action we see a marked decrease in the rate of oil debris particle production and in axial vibration.

5 Conclusions
In summary, the following conclusions can be drawn: A multi-parameter method is proposed for the analysis of online CM signals from a WT gearbox. Comparison of independent CM gearbox signals, such as vibration and oil debris particle count, against energy produced, can give early detection of incipient gearbox damage.

The method has been applied, after the event, to CMS data from an operating WT with an incipient gearbox fault and has been shown to give a significant warning, up to 2 months, of damage, which could have led to gearbox failure. This fault had already been detected by the Operator using conventional methods on the SKF WindCon system prior to the start of the data set used here. The proposed method could produce a signal leading to a Severity Factor for such an incipient fault. The Severity Factor would require calibration against other incidents to raise confidence but it has been shown to be effective in this incident. Such a Severity Factor, once calibrated, could reduce the risks of false alarms from the signals. The method could therefore allow a WT Operator to take appropriate and cost-effective maintenance decisions before an impending failure. The method could be programmed into a commercial CMS applied to WT gearboxes which provided the CM signals.

References
[1] Tavner P.J., Xiang J.P., Spinato F.
Reliability analysis for wind turbines, Wind Energy, 10(1), July 2006.

[2] Spinato F., Tavner P.J., van Bussel

G.J.W., Koutoulakos E. Reliability of wind turbine subassemblies, IET Renewable Power Generation, Vol. 3, Iss. 4, pp. 115, 2009.

[3] Faulstich, S. et al.: Windenergie Report

Deutschland 2008, Institut fr solare Energieversorgungstechnik (Hrsg.), Kassel, 2008.

[4] Various Authors, Operations Report 2007,

Document No. OWEZ_R_000_20081023, Noordzee Wind, October 2008.

[5] Wilkinson M R et al. Methodology and

results of the ReliaWind reliability field study, Scientific Track Proceedings, EWEC 2010, Warsaw, Poland.

[6] Germanischer Lloyd Industrial Services

GmbH, Wind Energy Guideline for the Certification of Condition Monitoring Systems for Wind Turbines, 2007.

Future work will focus on testing different gearbox faults in a WT test environment using this and other methods programmed into an SKF WindCon system so that a proposed Severity Factor could be calibrated. The method also needs to be applied to field data, ideally on-line, requiring further automation for detection.

[7] CONMOW, State of the Art Report:

Condition Monitoring for Wind Turbines, 2002-2006.

[8] Hameed, Z., Hong, Y.S., Cho, Y.M. Ahn,

S.H., Song, C.K., Condition monitoring and fault detection of wind turbines and related algorithms: A review, Renewable and Sustainable Energy Reviews, Volume 13, Issue 1, January 2009, Pages 1-39.

Acknowledgements
This work was funded by the UK EPSRC Supergen Wind Energy Technologies Consortium EP/D034566/1 and the EU FP7 Project ReliaWind 212966. The authors would like to thank RELIAWIND and Supergen Wind project partners for their, discussion on practical experiences and for their suggestions on relevant work.

[9] Tavner P.J., A review of condition

monitoring of rotating machines, IET Electrical Power Applications, 2008, Vol. 2, No. 4, pp. 215247.