Direct measurement of input loads for the wind turbine drivetrain under test on a nacelle test bench

Modern wind turbines have some of the highest levels of torque and non-torque loads of all industrial sectors. These high loads present a great challenge for the design of wind turbines. On a nacelle test bench, the wind turbine drivetrain can be tested and validated against the design. It is therefore important to measure the correct level of each load and all its dynamic behaviors during the test. The best way to achieve this is to measure the loads directly in front of the drivetrain. This paper presents a method of direct load measurement on the shaft adapter which connects the drivetrain to the test bench. The technical solution and some important details about the instrumentation on the adapter are also presented. Methods of signal nulling as well as signal conversion from raw signals to the loads are compared as well. The measurements obtained are then compared with the applied loads from the test bench which show good agreement. As a special case, the torque measurement is validated and calibrated up to 5 MN m by means of a state-of-the-art torque transducer.


Introduction
Nacelle test benches with the capability of 6-DOF (degree of freedom) load application for wind turbines have experienced a fast growth in the last decade. Numerous test campaigns have demonstrated their unique abilities in efficiently validating the wind turbine drivetrains under various design load cases. With new types of wind turbines becoming ever larger, the validation on a nacelle test bench tends to become a standard validation task in the development of wind turbine prototypes. Due to high load levels, the load . This can cause losses in accuracy and dynamic behavior in the measurement. In this work we take the "DyNaLab" nacelle test bench at Fraunhofer IWES as an example (Fig. 1). The torque is measured on three link elements of the coupling, the axial force on each link element being measured by means of strain gauge technology. The torque is calculated using three forces and their lever arms with respect to the center of rotation. However, since the torque measured still needs to travel through the moment bearing of the load application unit (LAU) before reaching the DUT, additional uncertainty will be caused by the bearing friction torque, which is affected significantly by the non-torque loads applied and the temperature. The non-torque loads (bending moments, shear forces and thrust force) are calculated from the forces acting on the six hydraulic cylinders which are in turn determined by the hydraulic pressure measurement in each cylinder. The major disadvantages here are the friction forces in the cylinders as well as the dynamic effect of the LAU inertia, which are difficult to take into account in the calculation. dynamical behavior, linearity and accuracy. To address the demand of direct torque measurement with metrological traceability, different approaches have been proposed based on the force-lever principle [1,2]. Another approach based on fiber-optic strain sensors is presented by Gutierrez al. [3] For load measurement in all six DOFs, a design study of multi-component has been carried out by Gnauert al. [4]. At DyNaLab, efforts have been made in a number of test campaigns to measure the input loads directly in front of the DUT, as shown in Fig. 2. The shaft adapter originally responsible only for connecting the DUT to the test bench is selected as the position where the input loads to the DUT are measured. Since the output flange of the test bench cannot fit the input shaft of each DUT, a shaft adapter is almost always necessary for a new test campaign. Usually, the shaft adapter has a structure of hollow cylinder between the connection flanges. This provides a very good condition for load measurement and accessibility for the instrumentation. In context of load measurement, the shaft adapter is also referred to as a measurement adapter.

Design considerations
Originally, the sole function of a shaft adapter is the mechanical adaptation between the DUT and the test bench. In DyNaLab, a new shaft adapter is normally designed and obtained each time a new drivetrain or nacelle is to be tested, since the input shaft of each DUT usually has a unique flange of bolt connection. The factors which need to be considered in the adapter design include a suitable flange connection on each side, the parallel and angular alignment of the connection, as well as the structure strength under all possible load scenarios. To use the adapter as a measurement device, several additional requirements have to be taken into consideration in the design phase. For example, a thinner cylinder wall is desirable to produce a higher signal level and better signal-to-noise ratio. Furthermore, the parallel misalignment between the cylinder and the rotating axis as well as the ovalization of the cylinder have a large  MulƟ-wire cable connecƟng the two posiƟons effect on the measurement and need to be controlled. Even the positions of the welds and bolts play a part in the design. Ultimately, the design of the adapter is always a balance between cost, time, and function considerations. Theoretically, loads in all 6 DOFs can be measured on the adapter. In practice, the loads can be roughly divided into three levels of interest. The load with the highest interest level is the torque, because torque is relevant for almost all types of tests and often has high accuracy and linearity requirements. On the second level are the bending moments and shear forces since they are typically non-torque loads applied by the load application unit and cause relatively large strains on the adapter that can be easily captured by the measurement system. The remaining thrust force has to be categorized into the third and lowest level of interest mainly because the strain signal it generates on the adapter is much smaller than that of the loads in the other 5 DOFs. This small signal is in turn often affected significantly by the boundary condition and by other loads due to the socalled "cross-talk" effect. In this paper, the measurement of loads on the first two levels of interest is studied.

Measurement instrumentation
To compensate the effect of the boundary condition and the strong cross-talk effect between the loads in different directions, redundant measurements on multiple measurement positions have been adopted in DyNaLab throughout different test campaigns, especially for the torque measurement [5][6][7]. An example is shown in Fig. 3, where instrumentation at 8 positions along the circumference of the cylinder is used for the measurement. The measurement adapter is manufactured and instrumented for an industrial test campaign in 2021.
In this case, the measurement positions are selected to avoid the weld seam at 0 ı . Three types of strain gauge full bridges are instrumented evenly distributed along the circumference on the inner surface of the cylinder. The circumference chosen lies in the middle of the surface along the axis of rotation.
Shear strain measuring full bridge at a single position, with the prefix "SHR" in the channel name Bending moment related full-bridge layout with strain gauges at two opposite positions, with the prefix "BM" in the channel name Axial strain measuring full bridge at a single position, with the prefix "AX" in the channel name The configuration of each type of full bridge is illustrated in Fig. 4. The major part of the instrumentation was carried out before assembly, with the adapter lying on the ground. Fig. 5 shows different phases of the instrumentation process. The strain gauges applied are protected against contamination, humidity and crushing with proper cover agent as well as a metal frame. After the adapter is assembled to the test bench, a total of 16 full-bridge channels are connected to a data acquisition (DAQ) system on the rotating part of the test bench. The DAQ system is powered via a slip ring. The data and control communication between the rotating DAQ system and the server is managed by a wireless LAN (WLAN) system.

Load calculation
Although the strain gauge full bridges are intended to measure the corresponding strains caused by specific loads, the raw signals of all the physical channels are electrical signals by nature. In engineering, the unit of the strain gauge raw signals is commonly mV =V . To obtain the load, a relationship between the specific load L and the corresponding raw signal ı needs to be determined. Usually, a linear relationship is assumed. For convenience of analysis, the linear relationship shown in Eq. 1 is adopted in this paper, where the offset parameter b compensates the signal offset at zero The most straightforward and common way in the sensor technology is to calibrate the measured signal with a reference load transducer, whose measurement is metrologically traceable to the national standard. This is shown in Fig. 6(a). However, due to the high levels of the loads that the adapter is supposed to measure, calibrations are usually either technically or economically not feasible. As an alternative, the factor a (the gain or sensitivity parameter) in the relationship can be obtained via analysis of the measurement chain. The parameters of the Wheatstone bridge and the strain gauges are needed to convert the electrical signal first into a strain signal in m=m. Their geometry and material properties are then used to obtain the signal of the load from this strain signal. This process is shown in Fig. 6(b) and is the one commonly used in the DyNaLab. The drawback of this process is the accumulation of uncertainty along the measurement chain. The uncertainty of each parameter in each conversion of the signal makes a contribution to the final measurement uncertainty of the load. Therefore, great care needs to be taken at each step and for each parameter in the process. Another drawback of the process is that the b factor in Eq. 1 (the zero point) cannot be determined through analysis of the measurement chain. This issue can only be solved by a nulling procedure via test.

Expression of raw signal
A total of 20 physical full-bridge channels are available. This provides a good opportunity to find a combination of several channels to provide the raw signal expression ı of each specific load L. Greater numbers of channels help to reduce the effect of unfavorable boundary condition and the cross-talk effect on the raw signal of the load.
The simplest case is the raw signal of the bending moment. Since the bending moment is the dominant load regarding the level of strain, the corresponding physical channel can be directly chosen as the raw signal, as expressed in Eq. 2.
In the equation, ı M 20 and ı M 110 denote the orthogonal bending moments in alignment with the positions of strain gauges. C H bm20 and C H bm110 are the physical channels of "BM_020" and "BM_110", respectively, which are shown in Fig. 3.
Before examining the raw signals for the torque and the shear forces, it is necessary to discuss the contributors to the shear strain at a certain position, because the measurements of both the torque and shear forces are based on the shear strain channels. Fig. 7 shows that torque, shear force and bending moment all contribute to the local shear strain, albeit not always in the same direction. With the help of the calculation shown in Eq. 3, the component due to the torque can be separated from the components of the other two contributors. Using the same principle, the raw signal of the torque ı T can be expressed as the mean value of all the 8 physical channels for shear strain, as in Eq. 4.
Shear strain as a result of torque, shear force and bending moment ı T = .C H shr20 + C H shr65 + ::: A similar approach has also been adopted in a previous application of torque measurement on the main shaft of a wind turbine, which is reported in another publication [5].
According to Eq. 3, the summation of shear strains caused by the shear force and the bending moment can be determined with the shear strains measured at two positions 180 ı apart. Similarly, the corresponding raw signal in mV/V, denoted as ı FM , can be expressed as shown in Eq. 5.
In practice, it is found that ı FM determined with only two measurements is accompanied by high levels of noises. A possible reason for the noise is the non-symmetrical boundary condition and load application due to the bearing and structural deformations on the nearby structures. To reduce the noise, a similar method as for the torque measurement can be adopted, namely, to use measurements at more positions. Since the shear force causes shear strains of different levels at difference positions, a compound shear strain can be obtained with measurements at different positions being properly weighted and combined together. Eq. 6 shows an example of the compound raw signal for the raw signal ı FM 20 . The raw signal of ı FM 110 can be obtained in a similar way.
To subtract the raw signal related purely to the shear force, the signal of the bending moment ı M has to be used to compensate the component caused by the bending moment out of ı FM , as shown in Eq. 7. The coefficient k FM denotes the influence of the bending moment. In this case, the coefficient is determined from the measurement of a specific test where the test bench applies bending moments in different levels and keeps all other loads stable.

Signal nulling
Before calculating loads from the raw signals, the offset in the raw signals needs to be compensated by determining the b parameter. This is called a nulling process. Owing to the imbalance in the full bridges during installation, the raw signals are normally not zero without the loads being applied. The aim of a nulling process is to determine the signal level at zero load. Since signal drift is a common phenomenon with the strain gauge measurement, the nulling process should be repeated on a regular basis depending on the speed of the signal drift. A straightforward way of nulling is to measure the raw signal before assembly when the measurement adapter is still lying on the ground or is suspended freely on the test bench without being connected to the DUT. Although this method is also used in practice, it has a crucial drawback in that the influence of bolt connection is not considered during the nulling process, which could cause a significant discrepancy later on, due to the high level of pretension forces on the bolts. Additionally, it is obviously not feasible to be carried out during a test campaign after assembly.
To carry out the nulling after assembly, a more practical method is often adopted at DyNaLab: the test bench is run slowly first in the positive and then in the negative direction for a couple of rotations each way. The mean value of the loads in the two directions can be determined as the level of zero load. The process can be described as follows: The DUT is first driven in the normal rotational direction by the test bench idling at a very low speed, in this case at 1 rpm. This part of the test is termed as Run I. After a stop, the test bench then drives the DUT at the same speed but in the opposite rotational direction. The DUT idles again without electrical power. This part of the test is termed as Run II. The measurement of two whole rotations in each run is used to calculate the mean value of that particular raw signal. The average of the mean values in the two runs (Eq. 8) is then adopted as the signal level for the corresponding zero load.
An example is shown in Fig. 9, where the raw signal of torque experiences a clear level change after the rotational direction is changed. Measurements from at least two whole rotations (highlighted in the figure) in each run are available for the nulling. The raw signals of the bending moment and shear force are also shown in the same figure. It can be noticed that these two signals also experience a slight change in the mean signal level, while remain unchanged in the signal oscillation. Considering that both signals are measured on the rotating part and therefore refer to the rotational coordination system, it is expected that the change in their signal level is not a result of external load change, which would rather lead to the change in the oscillation amplitude. One possible assumption for this is the internal load adjustment of the whole test setup following a re-positioning of the setup after the rotational direction is changed.
However, the exact reason of this phenomenon remains in this case unclear. It would be an interesting point to have further study and observation in further test campaigns or on other nacelle test benches. Nevertheless, the amplitude of the signal change in the bending moment and shear force here is very small, well below 100 kN m and 100 kN, and is considered to have limited impact on the torque nulling.
The process is especially suitable for the nulling of torque. During the idling run, the torque transferred through the adapter is only used to turn the DUT. The amplitude of this idling torque is assumed to be the same in both runs, given that the idling speed is kept the same in both runs. Since the DUT is driven in opposite directions of rotation between the two runs, the torque experienced by the adapter also changes its direction. This relationship can be expressed as given in Eq. 9.
T idling;I + T idling;II = 0 Since the idling torques are functions of the raw signal ı, the offset parameter b can be obtained as given in Eq. 11 when substituting the terms from Eq. 10 into Eq. 9.
T idling;I = a .ı idling;I + b/ T idling;II = a .ı idling;II + b/ (10) b = −.ı idling;I + ı idling;II /=2 Actually the process is generally also a very good way to measure the idling torque of the DUT, as shown in Eq. 12. The torque difference between the two runs corresponds to twice the idling torque. Since the measurement focuses on the signal difference, it is unaffected by the zero-drift of the torque measurement.
T idling = .T idling;I − T idling;II /=2 (12) For bending moments and shear forces, the loads are measured on the basis of the rotational coordination system (CSYS) because the adapter is a rotating part. With constant loads applied in the stationary coordinate system, the measurement of bending moments and shear forces in rotating CSYS will theoretically produce a sinusoidal sig- nal around the zero. Owing to the cross-talk effect and the influence of the boundary condition, the signal may not be perfectly sinusoidal, but still exhibit a periodic behavior, as shown in Fig. 9. It is assumed that the mean level of the periodic signal still corresponds to zero load at low speed. This assumption is accompanied by the uncertainty that the misalignment of the drivetrain during assembly may cause a constant component on the loads in rotating CSYS which shifts the mean level of the loads away from zero. Two courses of action can be taken to reduce the uncertainty. The first one is obviously to reduce the misalignment in the design and assembly, while the second is to carry out K Fig. 12 Load comparison of bending moments and shear forces in stationary CSYS the test for nulling at a slower rotating speed, so that the dynamic effect of the LAU inertia can be reduced.
The process described here cannot be used for the nulling of thrust force, which should be carried out before the adapter is installed or with the help of the thrust measurement of the test bench.

Sensitivity parameter
After the raw signal is defined and nulled, the sensitivity parameter a of the linear relationship (Eq. 1) between a specific raw signal ı and the corresponding load can be obtained by analyzing the measurement chain shown in Fig. 6. When analyzing the relationship between the strain and the load, both manual calculation and FEM simulation of the adapter can be used. In this case, the results of the manual calculation are adopted and the FEM results are used for verification.

Measurement of non-torque loads
As mentioned above, the measurement adapter shown in Fig. 5 has been instrumented for an industrial test campaign. An industrial multi-MW wind turbine has been tested for more than half a year, providing a number of opportunities to verify the load measurement on the adapter.
Since the adapter rotates together with the main shaft, the signals of loads in all the directions are measured in the rotating CSYS. With the help of additional measurement of the angular position, the loads in the rotating CSYS can be converted into loads in the stationary CSYS. In this case, a MEMS-based (Micro-Electro-Mechanical Systems) inclinometer is deployed on the drivetrain to measure the adapter's angular position. An inclinometer measures the position with respect to the direction of gravity and is therefore very convenient to be instrumented. Unlike the rotary encoder, the inclinometer is deployed on the rotating part and therefore can be connected to the same DAQ system for the load measurement.
Apart from the torque and the thrust force, which can be adopted directly without change, bending moments and shear forces in orthogonal directions along 20 ı and 110 ı are used in the rotational CSYS to calculate the loads in the stationary CSYS, shown in Fig. 10. An example of the bending moments before and after the transformation is given in Fig. 11. For comparison, the loads measured by the test bench's hydraulic system are also plotted in the figure. It can be seen that the bending moments measured by the adapter and by the test bench are in very good agreement. To protect the commercial interest of the turbine manufacture, the measured loads shown in the figures of this paper are often normalized. Since the loads in each figure or subfigure are normalized using the same scale, the scientific characteristics of the figures are maintained. Similarly, the comparison can also be carried out for loads in other directions. Fig. 12 shows another measurement where bending moments and shear forces change simultaneously or alternately during the course of the test. The loads are only compared in the stationary CSYS. The results also show a good match between the measurement on the adapter and the measurement from the test bench.

Torque measurement and calibration
Directly after the industrial test campaign mentioned above, the same DUT and adapter have been used again for an EU research project named WindEFCY [8]. This project has provided a unique opportunity to also measure the torque with a calibrated highly accurate 5 MN m torque transducer [9,10] (referred to as torque transducer below) from the German national metrology institute Physikalisch-Technische Bundesanstalt (also known as PTB). The behavior and accuracy of the torque measurement on the measurement adapter can therefore be studied with measurements of the transducer as the reference.
As shown in Figs. 13 and 14, the torque transducer is integrated into the test assembly between the measurement adapter and the DUT. Several additional adaptation structures are used to connect the transducer to the assembly. The load application unit of the test bench is shut down during the whole project to protect the torque transducer, which is not designed to withstand large bending moments and shear forces at the level found in a wind turbine.
To demonstrate why it is necessary to measure the torque at multiple positions on the adapter, the measurement using all 8 SHR channels is compared with the measurements with only 4 and 2 channels of the SHR full bridges installed on the adapter. Positions are always selected to be equally distributed on the circumference. For the 4-position torque signal, SHR20, SHR110, SHR200 and SHR290 are chosen, while for the 2-position torque signal, SHR20 and SHR200 are used. In Fig. 15 Comparing the 8-position torque measurement on the adapter with the measurement of the torque transducer, the signal quality of the torque from the adapter can be checked and the torque measurement can be calibrated with the PTB transducer as the reference.
In Fig. 16, the torque measurements of the same profile as in Fig. 15 are shown. The difference in the signal levels in the upper part of the figure indicates a sensitivity error in the measurement of the adapter, which corresponds to the factor a in Eq. 1. Calibration based on a series of designed profiles similar to the one shown in the figure has been carried out. According to this calibration, the sensitivity parameter a should be increased by 4.3%. This magnitude of error in a is a result of all the uncertainty contributions in the analysis of the whole measurement chain shown in Fig. 6. The calibration also gives an offset correction of the measurement, and as a result the calibrated torque meas-urement on the adapter T 0 can be expressed as a function of the original measurement T : The factor k addresses the sensitivity error of the measurement and has a value of k 1.043. With the calibration repeated on different days and under different conditions, it is found that the k factor is very stable, although a slight effect of the temperature can be observed. Throughout the calibrations studied, the k factor has been observed to drift by a maximum of 0.002 in line with the temperature changes. Other than the stable k, the offset parameter d is more prone to drift with temperature change. A total drift range for d of about 200 kN m was observed in the calibrations studied. It is worth pointing out that the drift happens slowly over a longer period of time and can be effectively reduced by regularly carrying out the nulling process. However, to study the drift behavior of the measurement, no nulling process has been carried out during the WindEFCY project, since a drift-free torque signal was already available on the reference torque transducer of the PTB. For future work, there are motivations to compensate the signal drift in the Fig. 16 Part of a torque calibration test carried out following designed profile. The PSD is created based on the measurements on the highest torque level in the profile. The Spectrum is plotted against harmonic orders of the rotation frequency measurement from the outset. The most important reason is to reduce the drift during the long-term tests, which can last for several hours. Another reason is to reduce the number of repetitions of the nulling process to give more test time. Concrete actions are planned to compensate the signal drift with the help of additional temperature sensors installed on the adapter.
To check the dynamic behavior of the adapter torque measurement, the signal is compared with the torque transducer measurement in a form of PSD analysis. This is shown in the lower part of Fig. 16. It is seen that the PSD of the torque measured by the adapter (based on 8 positions) has the same or similar levels as that of the torque transducer on all the first 20 harmonic orders of the rotation frequency. Therefore, the dynamic behavior of the torque signal from the measurement adapter can be considered to reach a level similar to that of the torque transducer in the examined frequency. In line with the analysis in Fig. 15, it has proven necessary to perform the torque measurement with strain gauges at 8 positions.

Conclusions
This paper discussed the challenge and necessity of direct load measurement on nacelle test benches. A detailed so-lution based on strain gauge instrumentation on the shaft adapter has been presented. A campaign has provided chances for the application and verification of the solution. It is shown in the paper that the direct measurement of torque and non-torque loads together on the adapter in front of the device under test (DUT) is feasible. Using strain gauges in different configurations and at multiple positions on the measurement adapter, signals of the transferred loads in almost all directions except for the thrust force can be successfully picked up and separated from each other. The relationships between loads and raw signals are determined through analysis of the whole measurement chain including mechanical property of the adapter and the strain gauge instrumentation. A signal nulling method based on specially designed test procedures is presented. Measurements of bending moments and shear forces show a good match with the corresponding measurements from the hydraulic system of the test bench. Moreover, the torque measurement is checked and calibrated up to 5 MN m with a stateof-the-art torque transducer. A sensitivity discrepancy of around 4.3% and an additional temperature-dependent offset drift have been found in the torque measurement on the adapter. Analysis of the dynamic behavior in the torque signals has shown that measurement using strain gauges at multiple positions can effectively reduce the noise in the signal. The torque measurement with 8 positions can achieve a similar low noise level to that of the torque transducer. Measurements with 4 positions can be a good compromise for similar industrial applications. Further research will focus on reducing the temperature-related signal drift in all the load measurements with the help of locally installed temperature sensors.