The development of an online condition-monitoring system for detecting damage to wind turbine dynamic mechanical components is essential for improved safety and reduced operational costs.  A condition-monitoring system will help achieve predictive, condition-based maintenance that improves safety, decreases maintenance costs and increases system availability.  During the past few years, major monitored objects on wind turbines have been drivetrain components and blades since they can cause long downtime and are extremely costly to repair or replace.  For blades, it is typically termed as structural health monitoring  and is out of the scope of this study. This paper will focus on wind turbine drivetrain components,  specifically gearboxes.
There have been various techniques explored for wind turbine gearbox condition monitoring (CM). Examples include vibration or acoustic emission analysis, oil-debris analysis and temperature monitoring. The data analyzed is either collected by the turbine supervisory control and data acquisition (SCADA) system or a dedicated CM system. In the works of Zaher et al.,  an online wind turbine fault detection framework was presented on the basis of the use of SCADA system data. One key component of the framework is specific diagnostic models targeting various turbine subsystems, such as the gearbox and generator, or turbine performance. By examining the deviations of test data sets from models developed with training data sets collected from turbine SCADA systems and defining normal turbine subsystem behavior, the successful detection of gearbox failures, on the basis of cooling oil and bearing temperature data, and generator failures, on the basis of generator winding temperature, was demonstrated. Another physics of failure approach to wind turbine condition-based maintenance was presented in the work of Gray and Watson . It also used SCADA data. Targeting one failure mode, a damage model was developed that describes the relationship between the turbine operating environment, applied loads and the rate at which damage accumulates. By using measured SCADA data and the developed damage model, a reliability value that reflects how possible the monitored gearbox may fail under the targeted failure mode can be obtained. Such a value could be further converted into a failure probability, which can be an input for the turbine operator to plan maintenance activities. As pointed out by the authors, for other failure modes seen in wind turbines, the development of additional damage models is needed so as to make the proposed approach beneficial to the industry. There is no doubt that it is advantageous in using SCADA data for wind turbine gearbox CM since the data is typically readily available, and purchase of a dedicated CM systems is not required, assuming the turbine did not come with an integrated CM system from the manufacture. However, the SCADA data may not be detailed enough to pinpoint the specific location of a damaged component inside a wind turbine gearbox, such as annulus gear tooth damage, as the collected data is typically averaged over a 10 min time interval. Such gaps in data acquisition can be filled by the installation of dedicated CM systems. A brief overview of dedicated CM techniques used in wind turbine gearbox CM was given in the work of Sheng and Veers . Two of the identified areas of research are to (i) improve accuracy and reliability of diagnostic decisions and (ii) automate the ‘expert’ in CM system data interpretation. This research is an attempt to target these two research areas by the exploration of data fusion techniques. Studies show that using data fusion methods—integrating several different diagnostic tools—result in a diagnostic system with improved detection rates and fewer false alarms, as compared with individual diagnostic tools. [9, 10] Automated data fusion processes can aid decision making by refining and reducing the quantity of information that the wind turbine operators must examine, thus enabling a timely, robust and relevant assessment of the situation.
The objective of this research is to demonstrate vibration-based health-monitoring techniques, oil-debris analysis techniques and wind turbine operational parameters combined to provide improved detection and decision-making capabilities as compared with using only individual diagnostic tools. This hypothesis is evaluated using data collected from two wind turbine gearboxes tested at the National Renewable Energy Laboratory (NREL) dynamometer test stand. The gearboxes were designed and rebuilt identically except for manufacturing variances. Vibration and oil-debris data were collected along with torque and speed data. Baseline data was collected on the dynamometer test stand from a ‘healthy’ test gearbox, which was rebuilt with new internal components and did not have any operational experience. Data then was collected from the dynamometer re-test of an identical gearbox after its internal components had sustained damage from its field test. The damaged test gearbox was also originally rebuilt with new internal components, which could be considered having a ‘healthy’ start. It first finished run-in in the NREL dynamometer and then was sent to a wind farm close by the NREL wind site in Colorado, USA for field test. The test gearbox was installed on a three-blade, stall-regulated, upwind wind turbine with a rated power of 750 kW and a rated wind speed of 16 m s − 1. The wind turbine has a double speed generator, which runs at 1200 rpm in low wind conditions or 1800 rpm in high wind conditions. During the field test, two oil loss events occurred and led to some damage to gears and bearings inside the test gearbox. This test gearbox will be considered as the damaged gearbox in this paper. The gearbox was then removed from the field and re-tested under controlled conditions in the NREL dynamometer test stand, where the data used in this study and corresponding to the damaged gearbox was collected. Details of the NREL dynamometer test stand, analysis of the vibration and oil-debris data and a fusion framework for the two monitoring techniques and operational conditions are discussed in this paper.
2 WIND TURBINE DRIVETRAIN TESTS
The wind turbine drivetrain tests were performed in the dynamometer test facility located at NREL. The test facility was developed to conduct performance and reliability tests on wind turbine drivetrain prototypes and commercial machines. [11, 12] The facility is capable of providing static, highly accelerated life and model-in-the-loop tests. The prime mover of the dynamometer is composed of a 2.5 MW induction motor, a three-stage epicyclical reducer and a variable-frequency drive with full regeneration capacity. The rated torque provided by the dynamometer to a test article can be up to 1.4 MNm, with speeds varying from 0 rpm to 16.7 rpm. Non-torque loading actuators, rated up to 440 kN for radial load and 156 kN for thrust load, also can be utilized in the dynamometer to apply thrust, bending and shear loads similar to those typically generated by a wind turbine rotor. Figure 1 provides a diagram of the test facility, and Figure 2 is a photo of the test stand with the test article installed.
Figure 3 illustrates the internal configuration of the wind turbine gearbox that was tested. The analysis provided herein focuses on the high-speed shaft (HSS) and intermediate-speed shaft (ISS) gear sets located in the gearbox of the wind turbine (between the rotor hub and the generator). The damage observed on the HSS gear set was severe scuffing. The ISS gear set also had scuffing and fretting corrosion on the teeth. Photos of the damage to the HSS and ISS gear set teeth are provided in Figures 4 and 5.
The tests of the gearboxes in the wind turbine drivetrain test stand were run at three operating conditions (listed in Table 1). These tests were performed on both the ‘healthy’ gearbox and the damaged gearbox. For the ‘healthy’ gearbox, the tests were conducted at the 1200 rpm, 25% torque and 1800 rpm, 50% torque test conditions. Each test was run for approximately 75 min. For the damaged gearbox, the tests were conducted using all three test conditions, and each test was run for about 15 min. Vibration data was measured by accelerometers placed at various locations along the drivetrain. The locations of two sensors used in this study—AN6 and AN7—are shown in Figure 6. Sensor AN6 mainly monitors the radial vibration of gears and bearings on the ISS. Sensor AN7 mainly monitors the radial vibration of gears and bearings on the HSS. Both of these sensors are integrated-circuit piezoelectric-type accelerometers and both have a sensitivity of 100 mV g − 1 (1 g = 9.8 m s − 2).
Oil debris generated was measured by an inductance-type oil-debris sensor installed in the main gearbox lubrication supply line. The oil-debris sensor has a 38 mm bore diameter. It measures the change in a magnetic field that is caused by passage of a metal particle larger than 350 μ and counts the number of particles that pass. Filters located downstream of the oil-debris sensor capture the debris after it is measured by the sensor. The oil-debris sensor was engaged for the duration of each test case, and the particle counts increased whenever it detected any particles.
Additionally, torque was measured by four strain gages arranged as a full Wheatstone bridge and mounted on the input shaft to the test gearbox. The HSS rotations per minute were measured by an encoder—generating 500 pulses per revolution—mounted on the non-drive end of the generator. The torque and HSS rotations per minute were sampled at 100 Hz continuously on the ‘healthy’ gearbox and at 40 kHz continuously on the damaged gearbox. Two different sampling rates were used because, for the ‘healthy’ gearbox, a data-acquisition system was developed for collecting the test gearbox dynamics data, and the configuration used a continuous sampling rate of 100 Hz. For the damaged gearbox, a dedicated CM data-acquisition system was used, and a sampling rate of 40 kHz was configured for all of its measurement channels so that the test schedule could be met.
3 VIBRATION ANALYSIS
Numerous diagnostic techniques have been developed from vibration data to detect damage and abnormal conditions of dynamic mechanical components. The rationale is that damaged components produce specific fault patterns in the accelerometer vibration signatures. Condition indicator (CI) refers to the vibration characteristics extracted from these signatures and is used to reflect the health of the component. Gear CIs are generated from fault patterns produced in vibration signatures when damaged components interact with their environment. The CI used is selected based on its ability to detect the specific component fault under investigation while minimizing false alarms. This requires determining both ‘healthy’ and faulted signatures of the component under investigation and also defining a threshold for the CI.
Gears produce vibration signals synchronous with shaft speed; therefore, gear CIs typically are calculated from time synchronous averaged (TSA) data. ‘TSA’ refers to the extracting of periodic waveforms from additive noise in the time domain by averaging signals over several revolutions of the shaft, thus reducing the noise in the signal. Employing TSA data requires using a tachometer or a sensor that generates a pulse per shaft revolution to re-sample the vibration data for each revolution of the shaft. This type of processing based on the encoder output was not performed during testing. Additionally, during tests conducted in the test stand, minor variances were observed in the speed data. For this reason, a gear CI was selected that could be calculated from spectrum (frequency) data without necessitating TSA. Note that, in the field, the rotor speed varies on the basis of wind conditions. This, in turn, causes varying loads and speeds on the gearbox. Depending on the variances in speed, sample rates and sample durations used by the wind turbine CM system, a new TSA technique might need to be developed for this application.
Table 1. Gearbox test conditions.
Input shaft torque
(% of rated torque)
HSS = high-speed shaft.
The CI selected to detect damage on HSS and ISS gear sets is referred to as sideband index (SI).  An SI is a measure of local gear faults and is defined as the average amplitude of the first-order sidebands of the fundamental gear meshing frequency. An increase in the magnitude of the sidebands of the fundamental gear meshing frequency (number of teeth × shaft speed) indicates a local gear fault such as a tooth damage. The sidebands are the two frequencies calculated as the total number of gear teeth minus 1 multiplied by the gear rotations per minute and the total number of gear teeth plus 1 multiplied by the gear rotations per minute. Table 2 provides details of the two gear sets and frequencies on the basis of shaft speeds. Because of the minor speed variances during testing and the lack of a signal pulse per shaft revolution, the frequency range used to calculate the sidebands was increased to include the speed variation. The maximum amplitude within the two sideband frequency ranges was then determined. The average amplitude of the two sidebands within this range was then equal to the SI.
Table 2. Gear dimensions.
Number of teeth
Gear meshing frequency (Hz)
ISS = intermediate-speed shaft; HSS = high-speed shaft.
The spectrum data used to calculate SI for the HSS and the ISS gear sets was collected at a sampling rate of 5 kHz, for 1.6 s every 1.5 min, from the ‘healthy’ gearbox and 40 kHz for continuously 10 min from the damaged gearbox. Only 15 spectra were analyzed for conditions 2a and 2c for the ‘healthy’ gearbox and 8 spectra for the 2a to 2c damage conditions from the two accelerometers (AN6 and AN7). Spectrum data from the 2b ‘healthy’ gearbox condition was unavailable; therefore, only the SI obtained from 2a and 2c is discussed. For illustrative purposes, all of the spectra for this analysis from the two accelerometers were averaged and plotted in Figures 7 and 8 respectively. Each figure contains two plots, which obtained under (a) test condition 2a and (b) test condition 2c. A rough review of the plots reveals that spectra from a gearbox with damaged components have richer frequency contents and relatively higher amplitude than those from a ‘healthy’ gearbox. Although scales can be expanded to compare the change in amplitudes for different conditions, visually analyzing this data is a time-consuming process. CIs enable reducing each spectrum to only one parameter for each acquisition. Damage progression could be determined by monitoring the CI along with torque, speed and debris generated.
Figures 9 and 10 are plots of the SI for the gear sets at the 2a and 2c conditions for the ‘healthy’ and damaged gearboxes. The readings represent an elapsed acquisition time of about 70 min for the ‘healthy’ gearbox and 10 min for the damaged gearbox. Figure 9 is a plot of the SI for the HSS gear set calculated from vibration data measured with accelerometer AN7. On the basis of this small data set, a threshold could be set at 0.0125, which is the mean of the upper SI limit under healthy condition, 0.01, and the lower SI limit under damaged condition, 0.015, to indicate damage to this gear set for both test conditions. The values were greatest for 2c, the increased speed and load condition.
Figure 10 shows a plot of the SI for the ISS gear set calculated from vibration data measured with accelerometer AN6. On the basis of this small data set, it is impossible to set up a limit by following the same principle as used in Figure 9 to indicate damage to this gear set at the greater load and speed conditions. One possible reason is that the SI is more sensitive to local cracks although the real damage is somewhat global, but future investigation on new CIs is needed to evaluate such a projection. In addition, the magnitude of the sidebands measured with the AN6 accelerometer at sidebands for the ISS gear mesh frequencies also were significantly less than those measured for the HSS gear with AN6. The SI values for the ISS gear using the AN7 accelerometer also were comparable with AN6 SI values. One possible reason for such a signal strength reduction is the floating Sun gear configuration of the gearbox, but it needs a comparison against fixed Sun gear configuration to verify. The test gearbox disassembly results showed that, in addition to spotty fretting corrosion and scuffing, all teeth on ISS gear set has polishing wear, which may be another factor. The difficulty in differentiating between the damaged and ‘healthy’ gears at the lesser load conditions indicates that many factors affect the response of a sensor to a specific fault. Factors that affect vibration response can include accelerometer type, location, mounting and signal processing. Gear speed, applied torque and system structural dynamics also can have an effect. The relationship between all of these variables is complex and not well defined. Determining factors having the greatest effect (variability) on the CI response requires further study.
4 OIL-DEBRIS ANALYSIS
The baseline oil-debris data was collected from the gearbox after its run-in was completed and under ‘healthy’ conditions. The oil-debris sensor was located in the main lubrication loop of the test gearbox. It monitored the lubrication oil for ferrous particles larger than 350 μ. Downstream of the sensor, the oil system had a 50 μ and 10 μ 2-stage filtration system to prevent debris from being recirculated within the lubrication system. The oil system also had a 3 μ offline filtration system that could run continuously 24 h a day, 7 days a week, even when the gearbox is not being tested. The ‘healthy’ gearbox was run for 4 h; during that period, a maximum of 11 particles were detected by the oil-debris sensor. The 11 particles detected were used as the baseline particle count.
When the damaged gearbox was tested at the three conditions, debris was detected by the oil-debris sensor. A plot of the change in debris during testing of the three conditions is shown in Figure 11. The assumption that the debris was generated only by the damaged ISS and HSS gears during testing could not be verified. The scanning electron microscope analysis of the filter cloth used to drain the gear oil during the test gearbox disassembly indicated that the major particulate constituents in the specimen were steel, iron oxide, brass and zinc. It implied potential gears and bearings damage. Additional damage propagation tests are required to correlate the tooth damage to oil debris generated. For the purposes of this analysis, it was assumed that the debris measured by the oil-debris sensor came from the gears during testing. The period in which this data was collected for the three conditions also is listed in the Figure 11 legend. The collection periods are longer than the vibration data collection periods for the damaged gearbox tests, as shown in Figures 9 and 10. The reason for the difference in collection period duration is that each test involved start-up and coast-down processes; during which, the oil-debris sensor was functioning, and the vibration spectra were only extracted when the test gearbox settled into stable operational states. The readings are comparable with minutes. If the gear-damage level and progression could be verified during the period when oil-debris data was collected, then a threshold for the oil-debris sensor could be set based on the number of particles detected.
Comparing the results obtained by vibration analysis and oil-debris particle counts under different testing conditions, e.g. Figures 9 and 11, it is seen that vibration monitoring technique is capable of identifying the location of potential damage on the basis of gear meshing frequencies or bearing characteristic frequencies, and oil-debris particle counts have much higher generation rates for wind turbine gearboxes with damaged components. However, it is hard for oil debris monitoring technique to tell exactly where the damage is located without further particle elemental analysis. For both techniques, extensive data is needed to define accurate thresholds so that reliable diagnostics can be reached.
5 DATA FUSION
Prior to review of the test results, data fusion analysis techniques were to be applied to the vibration and oil-debris data, mapping the oil debris generated and the vibration CI to gear damage. The intent is to combine the capability of pinpointing damage location by the vibration CI with the damage severity estimated by the oil-debris analysis. Multisensor data fusion works similarly to the human brain; it integrates data from multiple sources and uses it to make decisions. One objective of CM is to make it easier to make a decision about—or take action on—the current state of the system. Data fusion is one method that combines historical knowledge of the system to improve a maintenance person's ability to make decisions regarding the health of the system. Decision-level fusion was used to integrate these inputs because it does not limit the fusion process to a specific feature, thus enabling different features to be used without changing the entire analysis. This requires assessing the performance of the individual measurement technologies prior to their integration. If the individual CIs have low detection rates and high false alarm rates, then little is gained by integrating the technologies.
An understanding of the strengths, weaknesses and constraints of each measurement technology is required before the strengths can be capitalized upon via data fusion. A review of the data at the two damage conditions indicates that the vibration CI detected the damage to the HSS gear set at both load and speed conditions and detected damage to the ISS gear set at the high load and speed conditions. The poor performance of the CI for this gear set could be due to many environmental factors; this warrants further investigation. Other gear CIs also could be investigated. The small amount of data collected by the oil-debris sensor made it difficult to obtain inferences on its performance. Further analysis of the individual CIs is required before a fusion model can be developed.
For illustration purposes, a data fusion model applied to rotorcraft spiral bevel gears, Figure 12, is briefly discussed. The core of the model is decision-level fusion, which is a method that can be used to integrate multiple inputs. Various methods, [14-16] such as Voting techniques, Bayesian Inference, Fuzzy Inference Systems, Neural Networks and Dempster–Shafer Theory, can be used to implement decision-level fusion. Among these options, Fuzzy inference is adopted because of its advantages in better reflecting real world applications and its simplicity and flexibility in terms of implementation. A detailed description of the process used to define the membership functions can be found in Dempsey . Input values are defined, such as oil-debris particle counts, vibration CI SI, component (HSS gear), torque and speed with thresholds that correlate to damage levels. By using rules and membership functions, outputs per damage level can be defined as damage low, damage medium and damage high. These damage levels can be correlated to specific maintenance actions such as no action, inspect and repair. After more data becomes available, a data fusion model such as that shown in Figure 12 can be developed for the wind turbine drivetrain application.
On the basis of the respective results obtained by vibration analysis and oil-debris particle counts under different testing conditions, Figures 9– 11, it is reasonable to conclude that (i) from vibration analysis, the HSS gear set has damage, and the ISS gear set is doubtful to have damage and (ii) from oil-debris particle counts, the tested gearbox has gear or bearing damages. If the results from both vibration analysis and oil-debris particle counts are considered, it is pretty confident to conclude that both the HSS and ISS shaft gear sets of the tested gearbox have damage.
The objective of this research was to demonstrate that when combined, vibration-based health-monitoring techniques, oil-debris analysis techniques and wind turbine operational parameters provide improved detection and decision-making capabilities as compared with the capabilities of individual diagnostic tools. Data was collected from a ‘healthy’ gearbox and a damaged gearbox—both tested on the NREL dynamometer test stand. Vibration and oil-debris data were collected along with torque and speed data. Results indicate that the vibration-based CI SI can be used to indicate damage to gear teeth of the HSS gear set. Damage to the ISS gear set only could be detected at the greater speed and load test conditions. Investigation on alternative CIs has been planned in the future work following this research. Oil-debris particle counts also increased when the gearbox with damaged components was tested. It is clear that when wind turbine gearbox has damaged components, symptoms can be observed through the analysis of both vibration and oil debris data. The fusion of these two monitoring techniques with turbine operational conditions has shown to increase the confidence level in diagnosis results and to reduce the chances of false alarms, which is critical for owner and operator to improve their operation and maintenance practices and reduce the cost of wind energy. Further research is required to improve the performance of the individual health-monitoring tools. This can be accomplished by defining thresholds for individual measurement technologies. The thresholds definition warrants additional data collection throughout the life cycle of wind turbine gearboxes and better approaches in handling the timing across different sensing technologies. This will enable detection rates and false alarm rates to be quantified for individual measurement technologies. Quantifying the strengths, weaknesses and constraints of each measurement technology currently used to monitor wind turbine health—and then capitalizing on these strengths via data fusion—is the key to the development of future health-monitoring systems.
The authors thank the US Department of Energy for its support of this work. We also acknowledge and appreciate the support given by the NREL condition-monitoring partners.