Method for determining a torque in a wind power plant

Abstract

The present invention relates to a method (10) for determining a torque (84) applied to an input shaft (12) of a wind power plant. The method comprises at least the following steps: - generating (40) a deformation signal (80) depending on a detected deformation of a stationary component (18) of a transmission (14); - generating (42) a speed signal (86) depending on a detected speed of the input shaft (12); - modifying (46) the deformation signal (80) depending on the speed signal (86) in order to generate input data (92) for the neural network; and determining (54) the torque (84) applied to the input shaft (12) as output data depending on the input data (92) by means of the neural network. The invention also relates to a system for determining, by means of the neural network, the torque (84) applied to the input shaft (12) of the wind power plant.

Description

[0001] Method for determining a torque in a wind turbine

[0002] Technical area

[0003] The present invention relates to a method for determining a torque applied to an input shaft of a wind turbine gearbox using a neural network. Furthermore, the invention relates to a system for determining a torque applied to an input shaft of a wind turbine gearbox using a neural network.

[0004] State of the art

[0005] In wind turbines, it is desirable to determine the torque applied to an input shaft of the wind turbine. The torque determined in this way can be used, for example, to control the wind turbine. For example, the angle of attack of each rotor can be adjusted to keep the torque as constant as possible. The applied torque can also correspond to the generated electrical current. By measuring the torque, information can be drawn about wear and tear on the wind turbine in order to optimize maintenance. However, measuring the torque directly at a rotating element of a wind turbine gearbox can be expensive, complex to install, and prone to failure. For example, sliding contacts can be used for data transmission, which can reduce the efficiency of the wind turbine and also wear out.Radio transmission is very expensive and the necessary battery-based power supply requires regular battery replacement.

[0006] Description of the invention

[0007] A first aspect of the invention relates to a method for determining a torque applied to an input shaft of a gearbox of a wind turbine using a neural network. The wind turbine can, for example, have a horizontal or vertical design. The wind turbine can, for example, be designed as a lift rotor. The wind turbine can be designed to extract kinetic energy from wind and convert it into electrical energy. The wind turbine can, for example, feed electrical power into a power grid. The wind turbine can have a generator. The wind turbine can, for example, have a tower, a rotatable nacelle mounted thereon, and a rotor rotatable about a horizontal axis. The rotor can, for example, have three rotor blades whose angle of attack is adjustable. The gearbox can, for example, transmit torque from the rotor to the generator.The gearbox can provide a gear ratio, for example so that the generator rotates faster than the rotor. The gear ratio can be fixed or adjustable, for example two-stage adjustable. The gearbox can, for example, have spur gear stages and / or planetary gear sets. The input shaft can, for example, be permanently connected to the rotor in a rotationally fixed manner. At the input shaft, the rotor can, for example, feed the energy extracted from the wind into the gearbox. The rotor can be held at the input shaft. The gearbox can have an output shaft which is connected to the generator. The applied torque can be the torque with which the rotor drives the input shaft.

[0008] A neural network can, for example, be a mathematical model. The neural network can be created using a computer and / or be designed as an artificial neural network. The neural network can have input nodes, output nodes, and a plurality of intermediate nodes arranged between the input nodes and the output nodes. Respective connections between the nodes can, for example, have a weighting. The input nodes can, for example, be designed as data interfaces via which input data can be fed into the neural network. The output nodes can, for example, be designed as data interfaces via which output data can be output from the neural network. The output data can, for example, relate to the applied torque. The input nodes can be connected to the intermediate nodes, and the intermediate nodes can be connected to one another.The intermediate nodes can be connected to the output nodes. The neural network can, for example, be designed as a convolutional neural network. The method comprises a step of generating a deformation signal depending on a detected deformation of a stationary component of the transmission. The deformation signal can, for example, be an analog or digital signal, which can be received and further processed by the neural network at the input node. The deformation can, for example, be stretching and / or compression due to an applied load. The deformation can, for example, be reversible. The stationary component can, for example, be designed as a transmission housing. The stationary component can, for example, be a fixed rotating element of a planetary gear set of the transmission, such as a ring gear fixed to the housing.For example, one or more deformation sensors can be arranged on the stationary component to measure the deformation of the stationary component. The deformation sensors can generate the deformation signal directly, or an evaluation device can be provided for this purpose. One deformation signal can be generated per deformation sensor. The evaluation can then still be carried out simply by the neural network. However, several sensor values ​​can also be fused into a single deformation signal, for example before the deformation signal is made available to the neural network. In this case, the neural network can be particularly simple. The deformation can be recorded as a curve. The deformation can have a periodic curve, for example, an approximately sinusoidal curve. The periodicity can be caused, for example, by a changing load due to the rotation of the rotor.The deformation signal can also have a periodicity. For example, the deformation signal can have a fundamental frequency.

[0009] By detecting the deformation of the stationary component, the measurement can be particularly simple and cost-effective. For example, respective deformation sensors can simply be attached to the outside of the gearbox housing and their measurement signals transmitted via cable. The deformation of the stationary component can allow conclusions to be drawn about the torque applied to the input shaft, since, for example, individual gearbox shafts are mounted on the stationary component. The deformation can be detected using one or more strain gauges, for example. The deformation can also be detected optically, for example with a laser distance measuring device or a camera with a polarization filter. The deformation can also be detected using a piezo sensor. The deformation can be detected at several different locations and / or on stationary components.This allows for particularly precise detection and / or inference of the applied torque. For example, the neural network can be trained to detect a relationship between the applied torque and multiple measurement signals.

[0010] The method comprises a step of generating a speed signal depending on a detected speed of the input shaft. The speed of the input shaft can, for example, correspond to a rotational speed of the rotor. The speed can, for example, be detected by means of a speed sensor on the rotor, on the generator and / or on a rotating element of the transmission. The speed can also be detected directly on the input shaft. The speed signal can, for example, be an analog or digital signal. The speed signal can be received by the neural network at the input node and further processed. However, the speed signal can also, for example, be used only for pre-processing the deformation signal. The speed sensor can generate the speed signal directly, or an evaluation device can be provided for this purpose.The speed can be a measure of a rotational speed of the input shaft and / or the rotor.

[0011] The method comprises a step of modifying the deformation signal as a function of the speed signal in order to generate input data for the neural network. For example, a periodicity of the speed signal can correspond to the speed. The periodicity can correspond to a fundamental frequency. The fundamental frequency can be a carrier frequency. The fundamental frequency can change during operation and, for example, correspond to a current speed of the input shaft. This periodicity can be adapted, for example, to easily use the neural network for different wind turbines, which, for example, have a different target rotation speed of the rotor. The modification can be a preprocessing of the input data for the neural network.For example, modifying the deformation signal depending on the speed signal can simplify the filtering of disturbances, for example those caused by environmental influences. The modification can be carried out using a digital filter, for example. The determination of the applied torque can be particularly precise as a result of the modification and / or require particularly little computing power. Alternatively or additionally, training the neural network can be particularly simple. For example, the amount of training data required to train the neural network to learn a relationship between the deformation signal and the applied torque can be particularly small. For example, extensive training data no longer needs to be made available for all rotational speeds of the input shaft and taken into account as input data for the neural network.In addition, training data can be generated more easily in a simulation, for example.

[0012] The method comprises a step of determining the torque applied to the input shaft as output data based on the input data using the neural network. The neural network can calculate this determination using a learned relationship between the input data and the applied torque. The output data can be made available, for example, to a control device of the wind turbine. The output data can also be transmitted to a server, for example, for monitoring and power grid control. For example, a brake and / or a blade adjustment of the wind turbine can be controlled based on the determined torque.

[0013] In one embodiment of the method, it is provided that the modification comprises filtering the deformation signal with a frequency filter, in particular a fixed frequency filter. For example, the frequency filter can have a fixed cutoff frequency. For example, the frequency filter can also be adaptive. For example, the frequency filter can set respective cutoff frequencies depending on the rotational speed signal. The frequency filter can, for example, be used to filter respective disturbances from the deformation signal. For example, a curve of the deformation signal can be smoothed by the frequency filter. The frequency filter can be implemented in a digital filter. The filtering can take place after frequency modulation of the deformation signal, which will be described below. In this case, the frequency filter can be static and yet filtering can still take place that is adapted to the current rotational speed.The effort required for adaptive filtering can thus be particularly low. For example, filtering can be performed using a predetermined frequency threshold. The frequency filter can be configured as a high-pass filter, low-pass filter, or band-pass filter. However, the frequency filter can also be adjusted depending on the speed signal; for example, specific frequency thresholds can be set depending on the speed signal.

[0014] In one embodiment of the method, it is provided that the modification of the deformation signal comprises a frequency modulation of the deformation signal as a function of the rotational speed signal. The deformation signal can be modulated such that the modulated deformation signal has a constant fundamental frequency. The deformation signal can thus correspond to a constant rotational speed of the input shaft. For example, a periodicity or curve shape of the deformation signal can have a first frequency at a first rotational speed and a higher second frequency at a different, higher second rotational speed. The frequency modulation makes it possible to stretch the deformation signal at the second rotational speed so that a modified second frequency then corresponds to the first frequency. For this purpose, for example, a time segment of the deformation signal can be compressed or stretched so that the rotational speed corresponds to the constant rotational speed.The constant speed can, for example, correspond to a nominal rotor speed. This means that very few modifications are necessary. Frequency modulation, for example, allows a fixed frequency filter to be used while still adaptively filtering out interference signals. Interference signals can correspond to a deformation of the stationary component that was not caused by the applied torque, but rather by external influences, for example. The deformation signal modified by frequency modulation can be provided directly to the neural network as input data or, for example, can only be demodulated by reversing the frequency modulation.

[0015] In one embodiment of the method, it is provided that the modification of the deformation signal comprises a reversal of the frequency modulation of the deformation signal. The demodulated deformation signal can again have a fundamental frequency corresponding to the rotational speed of the input shaft. The fundamental frequency of the demodulated deformation signal can therefore, for example, again have the fundamental frequency of the originally generated unmodified deformation signal. This reversal can occur, for example, after filtering the deformation signal. The reversal can, for example, comprise a reverse stretching or compression of the deformation signal compared to the previous frequency modulation. This allows the deformation signal to return to its original frequency. The deformation signal can have sections with different frequencies. The deformation signal can thus again correspond to the detected rotational speed of the input shaft.The reversal can, for example, occur before the deformation signal is provided as input data to the neural network. The deformation signal can thus contain information about the rotational speed or correspond to the rotational speed. This allows, for example, the rotational speed signal to be omitted from the input data for the neural network. This allows the neural network to be particularly simple.

[0016] In one embodiment of the method, the input data also includes the speed signal. The speed can be an important parameter for determining the applied torque. The neural network can thus take this parameter into account. Furthermore, reversing the frequency modulation, for example, can be omitted. This makes preprocessing particularly simple.

[0017] In one embodiment of the method, it is provided that the method comprises a step of generating a temperature signal depending on a detected temperature. The method can comprise a step of detecting the temperature. In this way, temperature compensation can take place when determining the applied torque. For example, the temperature can deform the stationary component and / or cause increased resistance in the transmission, which then does not necessarily correspond to the applied torque. The temperature can be detected, for example, by one or more temperature sensors. Multiple detected temperatures or, for example, only an average value of all detected temperatures can be taken into account. One temperature signal can be generated for each detected temperature, or all detected temperatures can be merged into one temperature signal.For example, the detected temperature can be a temperature of the stationary component, such as a housing temperature. The detected temperature can also be, for example, an outside temperature or a transmission oil temperature. In one embodiment of the method, the input data also includes the temperature signal. The neural network can thus directly consider the temperature signal when determining the applied torque. This eliminates the need for analytical evaluation of the temperature signal. Furthermore, the transferability of the neural network between different wind turbines can be improved.

[0018] In one embodiment of the method, it is provided that the modification of the deformation signal additionally occurs as a function of the temperature signal. For example, a curve of the deformation signal can be shifted in order to compensate for a constant deformation of the stationary component due to thermal expansion or compression. For example, an amplitude of the deformation signal can be increased or decreased during the modification as a function of the temperature signal. In this way, the deformation signal can be additionally pre-processed, which can simplify the evaluation by the neural network and / or the training of the neural network. If the temperature signal is not made available to the neural network as part of the input data, this can also prevent the neural network from learning the often lower air temperature in strong winds as an incorrect relationship between the temperature signal and the applied torque.

[0019] In one embodiment of the method, it is provided that the input data additionally contain a gear ratio of the transmission in the region of a deformation measurement of the stationary component. For example, the deformation can occur axially in the region of a first planetary gear set or a second planetary gear set. Depending on the gear ratio, the housing of the transmission can be deformed to varying degrees there. The gear ratio can also make it easier to train and / or apply the neural network to different wind turbines. If the deformation measurement is carried out in several areas, an associated gear ratio for each deformation measurement can be included in the input data. If the gear ratio is adjustable, the input data can include a recording of the currently set gear ratio for providing the gear ratio in the region of the deformation measurement of the stationary component.

[0020] In one embodiment of the method, it is provided that the neural network has been trained with training data. The training data can have the torque applied to the input shaft as output data. The training data can have at least modified deformation signals as input data. In addition, the training data can also have the above-described optional further parts of the input data, i.e. the speed signal, the temperature signal and / or the gear ratio of the transmission in the range of the deformation measurement of the stationary component. During training, a relationship between the input data and the output data can be learned in the neural network. The training data can be generated, for example, by determining experimentally generated deformation signals and correspondingly measured applied torques. In addition, the corresponding speed signal is generated, for example, by speed measurement.To generate the training data, the experimentally generated deformation signals are then modified depending on the speed signal to generate the modified deformation signals.

[0021] A second aspect of the invention relates to a system for determining a torque applied to an input shaft of a wind turbine gearbox using a neural network. The system can be configured to perform a method according to the first aspect. Corresponding features and advantages of the first aspect can also form features and advantages of the further aspect, and vice versa.

[0022] The system comprises a detection device. The detection device is designed to detect a deformation of a stationary component of the transmission and a rotational speed of the input shaft. The detection device is also designed to generate a deformation signal and a rotational speed signal. For example, the detection device can comprise at least one rotational speed sensor and at least one deformation sensor. The detection device can comprise an evaluation device. The detection device can comprise, for example, a converter designed to convert an analog sensor signal into a digital signal for generating the deformation signal and / or the rotational speed signal.

[0023] The system comprises a computing device. The computing device can be embodied, for example, as a computer or microprocessor. The computing device can comprise a data transmission device, such as a network card, a data storage device, such as a hard disk, and / or an output device, such as a screen. The computing device is configured to modify the deformation signal as a function of the rotational speed signal in order to generate input data. The computing device is also configured to determine the torque applied to the input shaft as output data using a neural network as a function of the input data. For example, the computing device can comprise a frequency modulator and a frequency filter.

[0024] The method may include a step of training the neural network. For example, the neural network may be trained with synthetic training data, e.g., generated using a simulation, or with experimentally generated training data. During training, the neural network may learn a relationship between the input data and the output data. The input data may be preprocessed for training, e.g., by modifying the deformation signal as a function of the rotational speed signal. A further aspect relates to a corresponding training method for the neural network. Yet a further aspect relates to a computer program product comprising the neural network described here. The computer program product may be stored on a non-volatile data carrier, such as a hard disk or CD.Reading the computer program product can implement the method according to the first aspect on the system according to the second aspect.

[0025] Short description of the characters

[0026] Fig. 1 illustrates a method for determining a torque applied to an input shaft of a wind turbine gearbox using a neural network. Fig. 2 schematically illustrates a system for determining a torque applied to an input shaft of a wind turbine gearbox using a neural network.

[0027] Fig. 3 illustrates a deformation signal of a stationary component of the wind turbine gearbox.

[0028] Fig. 4 illustrates a speed signal of the input shaft of the transmission.

[0029] Fig. 5 illustrates a frequency-modulated strain signal.

[0030] Fig. 6 illustrates the deformation signal after filtering with a frequency filter and after reversing the frequency modulation.

[0031] Detailed description of embodiments

[0032] Fig. 1 illustrates a method 10 for determining a torque applied to an input shaft 12 of a gearbox 14 of a wind turbine using a neural network. The gearbox 14 and the input shaft 12 are shown in Fig. 2, with Fig. 2 illustrating parts of the wind turbine. A rotor 16 is permanently and non-rotatably attached to the input shaft 12. The rotor 16 has two or three adjustable rotor blades, which are not shown in Fig. 2. The gearbox 14 has a gearbox housing 18 as a stationary component. A strain gauge 20 and a temperature sensor 22 are attached to the gearbox housing 18.

[0033] The method 10 comprises a step of generating 40 a deformation signal 80 using a detection device 24 as a function of a deformation of the transmission housing 18 detected by the strain gauge 20. The deformation signal 80 is shown in Fig. 3 in a graph plotted against a time curve. An amplitude corresponds to a magnitude of the deformation of the transmission housing 18. A second curve 82 illustrates a fundamental frequency of the deformation signal 80. This is superimposed with high-frequency disturbance variables. In Fig. 3, the corresponding torque applied to the input shaft 12, which is to be determined using the method 10, is additionally illustrated by a curve 84.

[0034] The method 10 includes a step of generating 42 a speed signal 86 as a function of a detected speed of the input shaft 12 by means of the detection device 24. The speed signal 86 is shown in Fig. 4 in a graph over time, which corresponds to Fig. 3. For this purpose, the detection device 24 has a speed sensor (not shown).

[0035] The method 10 comprises a step of generating 44 a temperature signal as a function of a temperature of the transmission housing 18 detected by the temperature sensor 22 by means of the detection device 24.

[0036] The method 10 comprises a step of modifying 46 the deformation signal 80 as a function of the rotational speed signal 86 by means of a computing device 26 in order to generate input data for the neural network. For this purpose, a frequency modulation of the deformation signal 80 as a function of the rotational speed signal 86 takes place in a sub-step 48. In Fig. 3 and Fig. 4 it can be seen that the rotational speed is approximately constant in three sections. In a first section, which has a time period T, the fundamental frequency of the second curve 82 is, for example, F and the rotational speed of the input shaft 12 is D. In a second section, the time period is 0.5T and the fundamental frequency is twice as high and thus 2F. Likewise, the rotational speed of the input shaft 12 is twice as high and is 2D, as can be seen in Fig. 4. In a third section, the time period is 2T and the fundamental frequency is half as high and thus 0.5F.Likewise, the speed of the input shaft 12 is half as high and amounts to 0.5D. By means of frequency modulation 48, the result of which is shown in Fig. 5, the deformation signal 80 was frequency modulated so that it corresponds to a constant speed of the input shaft 12 with the amount D. In Fig. 5, this was represented by a change in the time axis, which is now no longer linear. All three sections now have a time duration of T, which means that a speed would always be D. A frequency-modulated deformation signal represented by curve 88 in Fig. 5 thus has a fundamental frequency with the amount F in all sections, which is represented by a second curve 90, the fundamental frequency of the frequency-modulated deformation signal 88. Likewise, the high frequencies corresponding to interference signals have also been modulated. Here, too, it must be taken into account that the time axis is scaled differently for all three ranges.

[0037] It should be noted that in a real example, the rotational speed of the input shaft 12 can also fluctuate within a section. The corresponding calculations or modifications can then be performed based on an average fundamental frequency. In addition, a moving time window can be used for the frequency modulation 48 of the deformation signal 80. For example, a time overlap of time sections for each frequency modulation 48 can be 2.5 to 5 seconds. The overlap can be fixed so that the neural network can easily take it into account when determining the applied torque. However, non-overlapping time sections can also be used as input data for the neural network. The frequency modulation 48 can compress or stretch the deformation signal 80 accordingly.A time period of the deformation signal 80, which is frequency modulated, can be, for example, 12 seconds.

[0038] In a sub-step 50 of the modification step 46, the frequency-modulated deformation signal 88 is filtered by the computing device 26 using a fixed frequency filter. This filters out the high-frequency fluctuations of the frequency-modulated deformation signal 88. This filtering 50 occurs due to the frequency modulation 48 as a function of the rotational speed of the input shaft 12, although the frequency filter with its frequency limits is fixed. The frequency-modulated and filtered deformation signal thus generated corresponds to the second curve 90 in Fig. 5, but can also be different. For actual measured values, a target frequency can be specified by a bandpass filter, around which filtering is carried out, for example, with a bandwidth of ±1 Hz, ±0.5 Hz, or even just ±0.1 Hz. The narrower the bandwidth, the more accurately the filtered signal can correspond to the applied torque.With a larger bandwidth, however, fluctuations in the rotational speed within the respective time period that was frequency-modulated can be taken into account to a certain extent by the neural network when determining the applied torque. This can reduce the computational effort, as larger time periods can be uniformly frequency-modulated without significantly compromising the accuracy of the determination. The target frequency can correspond to the frequency-modulated and thus essentially constant fundamental frequency. The actual fundamental frequency of the deformation signal 80, however, depends on the current rotational speed and the transmission ratio within the measuring range and is therefore variable.

[0039] In a sub-step 52, the frequency modulation of the deformation signal 80 and thus of the second curve 90 in Fig. 5 is reversed. This is shown in Fig. 6 by curve 92, which corresponds to part of the input data or to the entire input data for the neural network. The three different sections now again have different lengths and the time axis of the diagram in Fig. 6 is again shown linearly. Accordingly, the frequency of the deformation signal 92 demodulated in this way is also different in the different sections and corresponds to that in Fig. 3. The input data preprocessed in this way allow the neural network to be applied to different wind turbines and to determine the applied torque via a deformation measurement of the gear housing 18 with high precision.

[0040] Optionally, the speed signal 86 can also form part of the input data for the neural network. The temperature signal can be used in two ways. It can form part of the input data for the neural network. Alternatively or additionally, the temperature signal can also be used for further preprocessing of the deformation signal 80. The modification 46 of the deformation signal 80 can therefore also take place depending on the temperature signal. For example, before or after one of the sub-steps 48, 50, 52, a curve can be shifted up or down depending on the temperature signal. This shift can result from a comparison with a standard temperature. For example, the curve 92 of the demodulated and filtered deformation signal 80 can be shifted downwards overall if the temperature was above the standard temperature in all three sections.This allows for compensation for any underlying deformation of the transmission housing 18 due to its temperature. In real-world examples, the temperature may also vary within a section or differ between sections, so that subsections of the respective curves may be shifted differently.

[0041] The method 10 includes a step 54 of determining the torque applied to the input shaft 12 as output data as a function of the input data using the neural network implemented in the computing device 26. The output data can then correspond to the torque 84 applied to the input shaft 12, as illustrated in Fig. 3. The output data can be used, for example, to control a blade pitch of the wind turbine.

[0042] The detection device 24 with its sensors and the computing device 26 together form a system for determining the torque applied to the input shaft 12 of the gearbox 14 of the wind turbine by means of the neural network.

[0043] Reference symbol

[0044] 10 procedures

[0045] 12 Input shaft

[0046] 14 gearboxes

[0047] 16 Rotor

[0048] 18 Gearbox housing

[0049] 20 strain gauges

[0050] 22 Temperature sensor

[0051] 24 Detection device

[0052] 26 Calculating device

[0053] 40 Step Generate Deformation Signal

[0054] 42 Step Generate speed signal

[0055] 44 Step Generate Temperature Signal

[0056] 46 Step Modify Deformation Signal

[0057] 48 Substep Frequency Modulation Deformation Signal

[0058] 50 Substep Filtering the frequency-modulated deformation signal

[0059] 52 Substep Reversing the frequency modulation of the deformation signal

[0060] 54 Step Determine torque using a neural network

[0061] 80 deformation signal

[0062] 82 Fundamental frequency deformation signal

[0063] 84 torque at input shaft

[0064] 86 Speed ​​signal

[0065] 88 frequency-modulated deformation signals

[0066] 90 fundamental frequency frequency-modulated deformation signal

[0067] 92 demodulated and filtered deformation signal /

[0068] Curve for input data of the neural network

Claims

Patent claims 1 . Method (10) for determining a torque (84) applied to an input shaft (12) of a gearbox (14) of a wind turbine by means of a neural network, the method (10) comprising at least the following steps: - generating (40) a deformation signal (80) as a function of a detected deformation of a stationary component (18) of the transmission (14); - generating (42) a speed signal (86) as a function of a detected speed of the input shaft (12); - modifying (46) the deformation signal (80) in dependence on the speed signal (86) to generate input data (92) for the neural network; and - Determining (54) the torque (84) applied to the input shaft (12) as output data as a function of the input data (92) by means of the neural network.

2. Method (10) according to claim 1, wherein the modifying (46) comprises filtering (50) the deformation signal (80) with a frequency filter, in particular a fixed frequency filter.

3. The method (10) according to claim 1 or 2, wherein the modifying (46) of the deformation signal (80) comprises a frequency modulation (48) of the deformation signal (80) as a function of the rotational speed signal (86), in particular wherein the modulated deformation signal (88) has a constant fundamental frequency.

4. The method (10) according to claim 3, wherein the modifying (46) of the deformation signal (80) comprises a reversal (52) of the frequency modulation (48) of the deformation signal (80), in particular wherein the demodulated deformation signal (88) again has a fundamental frequency corresponding to the rotational speed of the input shaft.

5. Method (10) according to one of the preceding claims, wherein the input data (92) additionally comprise the speed signal (86).

6. Method (10) according to one of the preceding claims, wherein the method (10) comprises a step of generating (44) a temperature signal as a function of a detected temperature.

7. The method (10) according to claim 6, wherein the input data additionally comprises the temperature signal.

8. Method (10) according to claim 6 or 7, wherein the modification (46) of the deformation signal (80) is additionally carried out in dependence on the temperature signal.

9. Method (10) according to one of the preceding claims, wherein the input data additionally comprises a gear ratio of the transmission (14) in the range of a deformation measurement of the stationary component (18).

10. Method (10) according to one of the preceding claims, wherein the neural network has been trained with training data which has the torque applied to the input shaft (12) as output data and at least modified deformation signals as input data in order to learn a relationship between the input data and the output data.

11. System for determining a torque (84) applied to an input shaft (12) of a gearbox (14) of a wind turbine by means of a neural network, in particular wherein the system is designed to carry out a method (10) according to one of the preceding claims 1 to 9, wherein the system has a detection device (24), wherein the detection device (24) is designed to detect a deformation of a stationary component (18) of the gearbox (14) and a rotational speed of the input shaft (12) and to generate (40, 42) a deformation signal (80) and a rotational speed signal (86), and a computing device (26), wherein the computing device (26) is designed to modify (46) the deformation signal (80) as a function of the rotational speed signal (86) in order to generate input data (92) and to determine the torque applied to the input shaft (12) as output data by means of a neural network as a function of the input data (92).

Images