Gear Oil Condition Monitoring for Offshore Wind Turbines

Condition monitoring programs for offshore wind turbines typically include time-scheduled oil analysis. This analysis is usually performed by specialized laboratories so that information about impending damage and the best time for an oil change will be discovered.

However, offshore wind turbines are frequently located in difficult-to-access areas, and service is often limited to a specific period of the year. Through the application of online oil monitoring sensors, the sampling interval can be increased from approximately six months to every day. Thus, a more precise trending of important oil quality parameters can be achieved. Accordingly, scheduled maintenance intervals can be changed to wear-dependent intervals. Due to this more precise trending and consequently longer lead time, it is possible to shift maintenance to more moderate seasons.

Linear Variable Filter Mini Spectrometers

Linear variable filter (LVF) based mini spectrometers are a good option for online condition monitoring of gear oil in offshore wind turbines. The LVF mini spectrometer consists of only a few components. Figure 1 illustrates the schematic setup of such a mini spectrometer.

The LVF is a spectral engine realized as a Fabry-Pérot structure that provides virtually constant resolution over the specified wavelength region. The position-dependent thickness of the graded cavity determines the transmission wavelength.

An ideal LVF absorption spectrometer needs a collimated light source. Since pyroelectric detectors are only sensitive to variation in irradiance, a chopper wheel is required for light modulation. Simple electrical modulation of the light source allows frequencies of only a few hertz (Hz). Higher frequencies result in the reduction of modulation depth because of the thermal inertia of the light source and reduce signal amplitude.

Figure 1: A linear variable filter (LVF) spectrometer uses a collimated light source. Because of thermal inertia of the light source, a chopper wheel is required. Depending on the molecules and molecule chains in the sample, light energy will be absorbed at different frequencies. The interrogation unit is given by the combination of LVF mounted atop a pyroelectric detector array.

A measurement system according to Figure 1 can be installed in the bypass of a gear-lubricating system. Using a fluid cell enables a continuous acquisition of infrared (IR) transmission spectra of the oil. A downstream chemometric data processing predicts oil aging by means of important condition parameters.

Case Study

In this study, three common wind turbine gear oils were analyzed by means of infrared analysis. All spectra were examined as used samples taken from different wind power turbines. These types of oil were named oil 1, oil 2 and oil 3. Oil 1 was a mineral lubricating oil, whereas oil 2 and oil 3 were synthetic polyalphaolefin oils. All spectra were acquired by a Fourier transform infrared (FTIR) spectrometer with a spectral resolution of 1 cm^-1 and a spectral range of 4,000 cm^-1 to 500 cm^-1.

 

Figure 2: The fresh oil spectra of three popular wind turbine gear oils.

Figure 2 illustrates the fresh oil spectra of three common wind turbine gear oils. The hatched areas mark unused spectral ranges from 3,067 cm^-1 to 2,758 cm^-1 and 1,500 cm^-1 to 1,330 cm^-1.

During oil aging, many complex reactions of chemical compounds like additive depletion, oil degradation and contamination take place. All of those mechanisms can be monitored in the IR transmission spectra but are hard to resolve even for an experienced analyst. Each of the three analyzed oils showed different aging processes and specific indicators.

For oil 1, the acid number (AN) and oxidation were determined to be significant indicators. Depletion of the major additive (in this case, a phosphorus-based additive), an increasing amount of zinc and a change in viscosity were also used as good aging indicators in a full laboratory analysis but could not be monitored in an adequate quality by means of IR spectroscopy.

For oil 2, the amount of copper and zinc, a change of viscosity and AN were used to check deterioration. Additive depletion was monitored by measuring phosphorus and silicon.

For oil 3, the amount of non-ferrous metals like copper indicated localized flattening in the bearings. A change of viscosity revealed degradation, and a decline of molybdenum and zinc mirrored depletion of the additive package.

In general, IR spectroscopy can be used to obtain information about important lubricant properties. Oxidation and AN levels are directly definable in this way. In the mid-IR range, additive depletion is indicated in an indirect manner. Levels of non-ferrous metals typically cannot be monitored in the mid-IR range but sometimes correlate with spectral changes. For a better overview of these parameters, see Table 1.

Table 1: Important monitoring parameters for the determination of oil quality of three common wind turbine gear oils.

Prediction Models Using FTIR Spectra

The entire mid-IR FTIR spectrum was analyzed using a statistical factor analysis approach with the aid of a software package. A total of 200 data sets per type of oil were filtered out of a database to create multivariate regression models.

Due to the different chemical composition of the three lubricating oils, the number, quality and kind of prediction of important oil condition parameters out of their IR spectrum had to be subdivided to the type of oil.

Figure 3: IR absorption spectra of oil 1 comparing used and unused oil samples.

Oil 1 showed change in the mid-IR range during its aging process. As listed in Table 2, AN, viscosity, oxidation and the level of zinc were good parameters for this lubricant. An overview of this analysis is shown in Table 2.

Table 2: Statistics for factor analysis of oil 1 using FTIR spectra. RMSEP stands for root mean square error of prediction.

Oil 1 contained no ester compounds, so the determination of oxidation (according to DIN 51453) was possible. Applying this method to determine oxidation using LVF spectra was not feasible due to the rather moderate resolution of LVF. This problem was bypassed by employing a partial least square (PLS). For a subsequent comparison of FTIR and LVF PLS results, oxidation of oil 1 was also calculated using PLS and listed in Table 2. AN and oxidation were two parameters that could be predicted in high quality using FTIR spectra and PLS, whereas the prediction of viscosity and zinc performed poorly.

Oil 2 showed nearly no change in the mid-IR range during its aging process. This could be seen when comparing used and unused FTIR spectra, as illustrated in Figure 4.

Figure 4: IR absorption spectra of oil 2 comparing used and unused oil samples.

As shown in Table 1, AN, viscosity and the levels of phosphorus, silicon, zinc and copper were good parameters for this lubricant. The overview of this analysis is illustrated in Table 3.

Table 3: Statistics for factor analysis of oil 2 using FTIR spectra.

The strong ester compounds of oil 2 made it necessary to extend the two unused ranges (3,067 cm^-1 to 2,758 cm^-1 and 1,500 cm^-1 to 1,330 cm^-1) to a third range (1,680 cm^-1 to 1,780 cm^-1). This strong ester compound inhibited the direct measurement of oxidation according to DIN 51453, excluding this quality parameter. Regressions of copper and zinc provided no useful correlation, and the prediction quality of AN, phosphorus, viscosity and silicon was very poor. This was partially associated with the relatively young data set used for the PLS in which no older entries than 40 hours of operation existed.

 

Figure 5: IR absorption spectra of oil 2 comparing used and unused oil samples.

Oil 3 also showed change in the mid-IR range during its aging process. As shown in Table 1, AN, viscosity and the level of zinc, molybdenum and copper were appropriate parameters for this lubricant. The overview of the analysis is listed in Table 4.

Table 4: Statistics for factor analysis of oil 3 using FTIR spectra.

The ester compounds of oil 3 inhibited the direct measurement of oxidation according to DIN 51453, excluding this quality parameter. Regression of copper provided no useful correlation, and the prediction quality of viscosity produced rather poor correlations. However, the prediction of magnesium, molybdenum and zinc generated encouraging results.

All predictable oil parameters using FTIR spectra are summarized in Table 5. The green parameters indicate a high quality of prediction, while the red parameters denote a rather moderate prediction quality. The red parameters can be used for trending purposes but are not suitable for accurate measurements.

Table 5: Oil quality parameters using FTIR spectra and multivariate data analysis. Green parameters indicate a high quality of prediction, and red parameters denote a moderate prediction quality.

Limitations of LVF Spectra

LVF features a linear dependency between position along the length of the filter and its specific bandpass. The full width at half maximum (FWHM) of the specific bandpass is given in percent of its center wavelength.

Figure 6: Intrinsic resolution of an LVF in the spectral range from 2,000 cm^-1 to 715 cm^-1.

To simulate the performance of an LVF mini spectrometer, the data sets of the three oils were band limited, and the spectral resolution was reduced to the specific LVF. Therefore, two LVF data sets were generated. The first LVF data set was calculated to a filter resolution of 1 percent and a spectral range of 1,850 cm^-1 to 925 cm^-1. A second data set was calculated to the same filter resolution, but the spectral range was extended from 1,428 cm^-1 to 715 cm^-1. The specifications of these filters were chosen with respect to two commercially available LVFs.

Table 6: Specifications of used LVF.

Figure 7: A comparison of the original FTIR spectrum to the calculated spectra of an LVF mini spectrometer.

Prediction Models Using LVF Spectra

To simulate the performance of LVF-based mini spectrometers, the downscaled LVF data sets were used for a PLS. Data sets for LVF1, LVF2 and a combination of both were performed. The prediction qualities were compared with the previous FTIR analysis. Two exemplary predicted vs. measured plots using LVF1 are illustrated in Figure 8 and Figure 9.

Figure 8: Predicted vs. measured plot for a set of used oil samples of oil 1.

Figure 8 shows the prediction of AN based on spectra of the LVF1 data set. Compared to the ideal FTIR data set, prediction quality barely degraded.

Figure 9: Predicted vs. measured plot for a set of used oil samples of oil 3.

Figure 9 demonstrates the prediction of molybdenum levels based on spectra of the LVF1 data set. Compared to the ideal FTIR data set, prediction quality only slightly degraded.

Table 7: Statistics for factor analysis of oil 1 using LVF1 and LVF2 spectra.

Table 7 illustrates PLS statistics of oil 1, limiting prediction to identified quality parameters using FTIR spectra, while Table 8 shows the same statistics for oil 3.

Table 8: Statistics for factor analysis of oil 3 using LVF1 and LVF2 spectra.

The Results

Regressions based on the LVF data sets showed promising prediction accuracy. A comparison of the FTIR and LVF results is shown in Table 9 and Table 10 for oil 1 and oil 3, respectively. Oil 2 was excluded because of rather moderate prediction accuracy even when using FTIR spectra.

Table 9: Statistics for factor analysis of oil 1 comparing FTIR, LVF1, LVF2 and a combination of LVF1 and LVF2 spectra.

In comparing FTIR and LVF statistics of oil 1, it was determined that the prediction of AN was affected by the band limitation of LVF spectra. A combination of LVF1 and LVF2 offered results similar to the FTIR results. Predicting oxidation showed an increase in accuracy with LVF1. This phenomenon can be explained by the cut-off of unnecessary spectral range and thus a reduction of noise components. By using LVF2, the detection accuracy dropped, as the oxidation band was no longer contained in the spectral range of this filter. This decline was small but observable. It seems that the limitation of resolution compared to the limitation of spectral range had a rather low impact on the prediction accuracy.

Table 10: Statistics for factor analysis of oil 3 comparing FTIR, LVF1, LVF2 and a combination of LVF1 and LVF2 spectra.

In comparing FTIR and LVF statistics of oil 3, it was observed that the prediction of molybdenum was lightly affected by the band and resolution limitation of LVF spectra. When considering the evaluation of LVF1, the model accuracy produced nearly similar results as the FTIR model. LVF2 slightly reduced the accuracy, while a combination of LVF1 and LVF2 was comparable to the FTIR results. The prediction quality of zinc was moderate but stable in all four cases.

Conclusion

In this study, the influence of band and resolution limitation of IR spectra with application to oil condition monitoring was analyzed. Several important oil quality parameters were predicted using multivariate statistics and IR spectra. Three popular wind turbine gear oils were used to validate the practicality of LVF mini spectrometers for condition monitoring purposes. LVF mini spectrometers in combination with multivariate data processing were shown to be a promising possibility for an oil condition monitoring sensor. In addition, it was determined that band limitation plays a more important role than the resolution limitation. A combination of two LVFs with an increase of covered spectral range in some cases produced similar results as those models based on FTIR spectra.