A wind power main shaft bearing structure design method based on fatigue life optimization

Abstract

The application relates to the field of wind power generation and discloses a wind power main shaft bearing structure design method based on fatigue life optimization, which comprises the following steps: obtaining rolling body load estimation values and calculating load distribution information entropy; obtaining total cumulative fatigue damage in the whole life cycle by time integration; respectively generating a first control quantity according to a first frequency component and a second control quantity according to a second frequency component; determining a trigger level according to the load distribution information entropy and a relative entropy change rate; generating a virtual information entropy estimation value based on SCADA data to replace the load distribution information entropy when a sensor fails; and iteratively outputting optimal structure parameter values and control strategy parameters. By solving rolling body load distribution based on an observation vector and calculating information entropy, the information entropy is subjected to frequency spectrum decomposition to generate control quantities of an adjusted yaw angle, a common pitch angle and an independent variable pitch angle, so that structure design and active control are combined, and real-time identification and suppression of load deviation are realized.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically to a wind turbine main shaft bearing structure design method based on fatigue life optimization. Background Technology

[0002] The main shaft bearing of a wind turbine generator set is a critical load-bearing component in the transmission chain. It is installed on the main shaft between the wind turbine hub and the gearbox or direct-drive generator, and is used to support the rotation of the main shaft and transmit the mechanical energy captured by the wind turbine. As the single-unit capacity of wind turbine generator sets develops towards larger sizes, the size and load-bearing requirements of the main shaft bearing are constantly increasing. Its design life is usually required to reach more than 20 years. During service, the bearing is subjected to the combined action of radial load, axial load, and overturning moment. Moreover, the load exhibits strong time-varying characteristics with changes in wind speed and direction. This leads to extremely uneven distribution of contact stress between the rolling elements and raceways of the bearing, and premature fatigue spalling is prone to occur in local areas.

[0003] Existing wind turbine main shaft bearing structural design methods typically employ a static optimization strategy based on rated dynamic load. Designers determine key structural parameters such as the number of rolling elements, rolling element diameter, and contact angle according to a standard load spectrum. The optimal parameter combination is obtained using the rated dynamic load or volume as the objective function, and the rated life is calculated based on the Lundberg-Palmgren theory or ISO 281 standard as a design criterion. However, this is a one-time offline optimization. After the bearing is put into operation, its structural parameters cannot be adaptively adjusted according to real-time load changes. This results in the mismatch between the standard load spectrum in the design stage and the actual service load spectrum being uncorrectable. The off-center load condition persists for a long time without active suppression measures, which can easily cause the actual service life of the wind turbine main shaft bearing to be far lower than the design expectation, resulting in a high early failure rate. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a wind turbine main shaft bearing structure design method based on fatigue life optimization, which solves the problem that existing wind turbine main shaft bearing structure design methods easily result in the actual service life of the wind turbine main shaft bearing being far lower than the design expectation and the early failure rate being high.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a wind turbine main shaft bearing structure design method based on fatigue life optimization, comprising the following steps: Obtain structural response data of the main shaft bearing under operating conditions and construct observation vectors. Based on the observation vectors, solve the rolling element load distribution with maximum entropy as a constraint, obtain the estimated value of each rolling element load and calculate the load distribution information entropy. The current damage accumulation rate is calculated based on the power-law relationship between the load distribution information entropy and the fatigue damage accumulation rate, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time. The load distribution information entropy is spectrally decomposed into a first frequency component and a second frequency component. A first control quantity is generated based on the first frequency component and a second control quantity is generated based on the second frequency component. The first control quantity is used to adjust the yaw angle and the common pitch angle, and the second control quantity is used to adjust the independent pitch angle. The trigger level is determined based on the load distribution information entropy and the rate of change of relative entropy, and the corresponding first control quantity and second control quantity are output when the trigger is triggered. In the event of sensor failure, a virtual information entropy estimate is generated based on SCADA data to replace the load distribution information entropy; Using the bearing structural parameters as design variables and minimizing the total cumulative fatigue damage over the entire life cycle as the objective, the optimal structural parameter values ​​and control strategy parameters are iteratively output.

[0006] By adopting the above technical solution, the rolling element load distribution is solved based on the observation vector, and the information entropy is calculated. A power-law relationship between the information entropy and the damage accumulation rate is established to obtain the total cumulative fatigue damage over the entire life cycle. The information entropy is spectrally decomposed to generate control quantities for adjusting the yaw angle, common pitch angle, and independent pitch angle. The control quantities are output according to the trigger level. When the sensor fails, a virtual information entropy replacement is generated. The bearing structural parameters are used as design variables, and the goal is to minimize the total cumulative fatigue damage. The optimal parameters are iteratively output, thereby combining structural design with active control. This achieves real-time identification and suppression of off-center loads and solves the problem that existing wind turbine main shaft bearing structural design methods easily lead to the actual service life of the wind turbine main shaft bearing being far lower than the design expectation and the early failure rate being high.

[0007] Preferably, the step of acquiring structural response data and constructing observation vectors under the operating state of the spindle bearing includes: acquiring the axial strain and tangential strain of four sets of strain rosettes arranged circumferentially on the spindle surface; The three-dimensional vibration acceleration of the bearing housing, the main shaft speed, the generator torque, the pitch angle of the three blades, the yaw angle, the wind speed, and the wind direction are obtained. The above signals are then aligned by time and combined into an observation vector.

[0008] Preferably, obtaining the estimated load values ​​for each rolling element and calculating the load distribution information entropy includes: The observation vector is input into a pre-trained deep neural network to obtain a preliminary estimate of the rolling load output by the deep neural network. The preliminary estimate of the rolling element load is input into the extended Kalman filter joint observer for state recursion to obtain the filtered estimate of the rolling element load. The load distribution information entropy is extracted from the filtered rolling element load estimate.

[0009] Preferably, the step of integrating over time to obtain the total cumulative fatigue damage over the entire life cycle includes: Obtain the dimensional parameters and power exponents corresponding to the bearing structural parameters; Extract the entropy deviation between the theoretical maximum entropy and the load distribution information entropy; The current damage accumulation rate is determined by the power exponent of the scale parameter and the entropy deviation, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time.

[0010] Preferably, the spectral decomposition of the load distribution information entropy includes: Wavelet packet decomposition is performed on the load distribution information entropy time series, and the reconstructed signal with a cutoff frequency below a preset Hertz is taken as the first frequency component; The second frequency component is obtained by separating the first frequency component from the original load distribution information entropy.

[0011] Preferably, the step of generating a first control quantity based on a first frequency component and generating a second control quantity based on a second frequency component includes: Perform proportional-integral calculation on the first frequency component and output the yaw angle correction and common pitch angle increment as the first control quantity. The second frequency component and the current load distribution information entropy are used in a model predictive control algorithm to generate an independent pitch angle increment as the second control variable. The entropy budget is updated based on the current load distribution information entropy, and the second control variable is subjected to amplitude attenuation when the entropy budget is lower than the preset value of the theoretical maximum entropy.

[0012] Preferably, the model prediction control algorithm for combining the second frequency component and the current load distribution information entropy includes: A state-space prediction model is constructed, comprising characteristic values ​​of blade flapping bending moment, characteristic values ​​of wind turbine overturning moment components, and characteristic values ​​of rolling element load deviation. Set the prediction time-domain step number, control the time-domain step number and sampling period, and recursively predict the entropy deviation sequence in the time domain based on the second frequency component and the current load distribution information entropy. Construct an objective function with the sum of the weighted sum of entropy deviations, the sum of squared control increments, and the sum of squared control increment differences as indices, and solve for the control sequence corresponding to the objective function; The independent pitch angle increment of the first sampling step in the control sequence is extracted as the second control output.

[0013] Preferably, determining the trigger level based on the load distribution information entropy and the rate of change of relative entropy includes: Calculate the change in load distribution information entropy over a preset time period and combine it with the theoretical maximum entropy to obtain the relative entropy change rate; The level of silence, warning, or intervention is determined based on the ratio of the current load distribution information entropy to the theoretical maximum entropy and the relative entropy change rate. At the warning level, only the second control variable is output; at the intervention level, both the first and second control variables are output. When the current load distribution information entropy is at the lower end of the preset value of the historical distribution over the past preset hours, the trigger level will be determined as an intervention level.

[0014] Preferably, the step of generating a virtual information entropy estimate based on SCADA data to replace the load distribution information entropy in the event of a sensor failure includes: Calculate the predicted values ​​of each sensor based on the current rolling element load estimate, and compare the residual between the predicted value and the measured value with three times the standard deviation to determine the sensor fault. Using the SCADA data sequence within a preset time period as input, the pre-trained long short-term memory network is invoked to obtain the estimated value of the virtual information entropy output by the long short-term memory network; When all strain gauges and vibration acceleration sensors fail, the load distribution entropy is replaced with a virtual entropy estimate and the control output gain is reduced.

[0015] Preferably, the iterative output of optimal structural parameter values ​​and control strategy parameters includes: Latin hypercube sampling is performed within the design space of bearing structural parameters to generate an initial set of sample points, each of which contains a set of bearing structural parameter values. For each sample point, iterate to obtain the total cumulative fatigue damage over the entire life cycle corresponding to each sample point; Using the Gaussian process as a surrogate model, the mapping relationship between sample points and total cumulative fatigue damage over the entire life cycle is fitted, and sequential sampling is performed with the desired improvement as the acquisition function. After iterative sampling to a preset number, the bearing structural parameter values ​​that minimize the total cumulative fatigue damage over the entire life cycle and their corresponding control strategy parameters are extracted as output.

[0016] This invention provides a wind turbine main shaft bearing structure design method based on fatigue life optimization. It has the following beneficial effects: 1. This invention solves the rolling element load distribution based on observation vectors and calculates information entropy, establishing a power-law relationship between information entropy and damage accumulation rate to obtain the total cumulative fatigue damage over the entire life cycle. It then performs spectral decomposition of the information entropy to generate control quantities for adjusting yaw angle, common pitch angle, and independent pitch angle. Based on the trigger level, the control quantities are output, and a virtual information entropy replacement is generated when a sensor fails. Using bearing structural parameters as design variables and minimizing total cumulative fatigue damage as the objective, the optimal parameters are iteratively output. This combines structural design with active control, achieving real-time identification and suppression of off-center loads. It solves the problem that existing wind turbine main shaft bearing structural design methods often result in actual service life of the wind turbine main shaft bearing being far lower than the design expectation and a high early failure rate.

[0017] 2. This invention introduces the power-law relationship between load distribution information entropy and fatigue damage accumulation rate into the bearing design process, establishing a direct quantitative correlation between the degree of information entropy deviation and local fatigue damage. This allows bearing design optimization to no longer rely on indirect rated dynamic load indicators, but instead directly use the total cumulative fatigue damage over the entire service life as the optimization target. This improves the accuracy of fatigue life prediction and the pertinence of design optimization, making the design results closer to the damage evolution law under actual service conditions.

[0018] 3. This invention achieves decoupled control of continuous off-center load bias and instantaneous off-center load fluctuation by performing spectral decomposition on the real-time information entropy and generating slow-scale control quantities and fast-scale control quantities respectively. The slow-scale controller eliminates long-term load asymmetry by adjusting the yaw angle and common pitch angle, while the fast-scale controller quickly suppresses transient off-center load caused by turbulence through independent pitch control. The two work together to maintain the load distribution of the bearing rolling elements in a uniform state in real time, delaying early fatigue spalling of the raceway surface and substantially improving the actual service life of the bearing.

[0019] 4. This invention combines sensor fault detection with virtual information entropy estimation based on SCADA data, enabling seamless switching to virtual sensors to maintain system operation when the main sensor fails, while reducing control gain to ensure safety. This redundant design avoids control failure or system shutdown caused by single-point sensor failure, improving the reliability and environmental adaptability of the entire bearing fatigue life optimization system, and is especially suitable for scenarios with limited maintenance conditions, such as remote wind farms. Attached Figure Description

[0020] Figure 1 This is a flowchart of a wind turbine main shaft bearing structure design method based on fatigue life optimization proposed in this invention; Figure 2 This is an architecture diagram of a wind turbine main shaft bearing structure design system based on fatigue life optimization proposed in an embodiment of the present invention. Figure 3This is a schematic diagram of information entropy spectrum decomposition proposed in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the trigger level determination proposed in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the sensor fault and virtual entropy switching process proposed in an embodiment of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: In the first embodiment of the present invention, the present invention provides a wind turbine main shaft bearing structure design method based on fatigue life optimization, such as... Figure 1 As shown, it includes the following steps: Obtain structural response data of the main shaft bearing under operating conditions and construct observation vectors. Based on the observation vectors, solve the rolling element load distribution with maximum entropy as a constraint, obtain the estimated value of each rolling element load and calculate the load distribution information entropy. Furthermore, structural response data of the spindle bearing under operating conditions is acquired and observation vectors are constructed, including: Obtain the axial and tangential strains of four sets of strain rosettes arranged circumferentially on the main shaft surface; The three-dimensional vibration acceleration of the bearing housing, the main shaft speed, the generator torque, the pitch angle of the three blades, the yaw angle, the wind speed, and the wind direction are obtained. The above signals are then aligned by time and combined into an observation vector.

[0023] Furthermore, the estimated load values ​​for each rolling element are obtained, and the load distribution information entropy is calculated, including: The observation vector is input into a pre-trained deep neural network to obtain a preliminary estimate of the rolling load output by the deep neural network. The preliminary estimate of the rolling element load is input into the extended Kalman filter joint observer for state recursion to obtain the filtered estimate of the rolling element load. The load distribution information entropy is extracted from the filtered rolling element load estimate.

[0024] Specifically, in this embodiment, in order to realize online estimation of the load distribution of the rolling elements of the main shaft bearing and calculation of information entropy, it is first necessary to obtain the original data reflecting the service state of the bearing and perform structured processing on the data. During operation, the wind turbine main shaft bearing is subjected to time-varying radial load, axial load and overturning moment, and the response signals at different positions contain rich structural dynamic information.

[0025] Four sets of strain gauges are arranged at equal intervals along the circumference of the main shaft surface. Each set of strain gauges includes axial strain gauges and tangential strain gauges to obtain the bending strain and torsional strain of the main shaft at different azimuth angles. At the same time, the existing CMS system of the wind turbine is used to collect the triaxial vibration acceleration signal of the bearing housing, the real-time rotational speed is obtained using the main shaft speed encoder, and the generator torque, the three blade pitch angles, the yaw angle, the wind speed, and the wind direction are read using the SCADA system. These signals come from sensors with different sampling frequencies and need to be time aligned first so that the measurement values ​​of each sensor at the same time form a complete observation vector. The strain gauge signals, vibration acceleration signals, rotational speed signals, generator torque signals, blade pitch angle signals, yaw angle signals, wind speed signals, and wind direction signals are linearly interpolated and resampled according to the acquisition timestamp, with a unified sampling frequency of 100-100 Hz. Then, all signal values ​​at the same time are spliced ​​into an observation vector, and each element of the observation vector corresponds to the measurement value of a sensor.

[0026] After obtaining the observation vector, it is necessary to solve for the rolling element load distribution based on the maximum entropy principle, and denote the number of single-row rolling elements in the main shaft bearing as... The normal contact loads borne by each rolling element constitute a load vector to be estimated. Under conditions of limited information, the probability density function of the load distribution is unknown. The maximum entropy principle provides an effective way to infer the load distribution from limited measurement data without introducing additional assumptions. In this embodiment, the following maximum entropy estimation problem is constructed: using the information entropy of the load proportion of each rolling element as the objective function, and the total radial force balance, overturning moment balance, and the mapping relationship between the observation vector and the load as constraints, the load distribution that maximizes entropy is solved.

[0027] The total radial force and overturning moment in the above constraints can be calculated from the strain gauge signal using a simply supported beam model of the main shaft. The mapping relationship between the observation vector and the load is pre-calibrated through finite element simulation, and is expressed as a linear superposition of the measurement value of each sensor and the contribution of each rolling element load. Solving this constraint optimization problem yields the maximum entropy estimate of the rolling element load.

[0028] However, maximum entropy estimation is sensitive to observation noise. To improve the robustness of the estimation, this embodiment further introduces a joint observer of extended Kalman filtering and maximum entropy estimation, using a deep neural network as a surrogate model for maximum entropy estimation. The input of this deep neural network is the observation vector, and the output is the maximum entropy estimate of the rolling element load. The training data of the deep neural network comes from the sample set generated by finite element simulation, covering typical load ranges, including various combinations of radial load, overturning moment, and eccentric load direction. During the online operation phase, the real-time observation vector is input to the trained deep neural network to obtain a preliminary estimate of the rolling element load. Subsequently, this preliminary estimate is used as the observation input for extended Kalman filtering. The state equation of extended Kalman filtering describes the dynamic evolution of the load over time, and its state transition matrix is ​​determined according to the bearing quasi-static model. The observation equation links the output of the deep neural network with the actual load. Through time updates and measurement updates of extended Kalman filtering, the filtered rolling element load estimate is obtained. This estimate has lower variance and better continuity compared to the direct output of the deep neural network.

[0029] Finally, the load distribution information entropy is calculated based on the filtered rolling element load estimates. Let the load estimate of the jj-th rolling element be... The total radial load is Then the load ratio Load distribution information entropy Defined as , Take 32, when all rolling element loads are equal. At this point, the information entropy reaches its maximum value. When the load is completely concentrated on a single rolling element, the information entropy approaches 0. The real-time value of this information entropy directly reflects the degree of off-center loading. The larger the entropy value, the more uniform the load distribution; the smaller the entropy value, the more severe the off-center loading.

[0030] The current damage accumulation rate is calculated based on the power law relationship between the load distribution information entropy and the fatigue damage accumulation rate, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time. Furthermore, integrating over time yields the total cumulative fatigue damage over the entire lifespan, including: Obtain the dimensional parameters and power exponents corresponding to the bearing structural parameters; Extract the entropy deviation between the theoretical maximum entropy and the load distribution information entropy; The current damage accumulation rate is determined by the power exponent of the scale parameter and the entropy deviation, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time.

[0031] Specifically, in order to directly correlate load distribution information entropy with bearing fatigue life, a quantitative conversion model between entropy deviation and damage accumulation rate needs to be established. The more uniform the load distribution on the rolling elements, the closer the information entropy is to the theoretical maximum value, and the slower the fatigue damage accumulation. Conversely, the more severe the off-center loading, the smaller the entropy value, and the faster the damage accumulation rate. The deviation from entropy satisfies a power-law relationship: ,in The current load distribution information entropy, For the theoretical maximum entropy, For scale parameters, The power exponent, and Related to bearing structural parameters, these are obtained through a pre-built response surface model.

[0032] After calculating the current damage accumulation rate, it is necessary to integrate over time to obtain the total cumulative fatigue damage over the entire life cycle. Let the design life be... The total cumulative fatigue damage Depend on Given. As an option, the theoretical maximum entropy is... With current information entropy The difference is defined as the entropy deviation. The current rate of damage accumulation can then be expressed as: Integrating over time yields... In this embodiment, Taking 20 years, this is equivalent to 6.307 × 10⁻⁶. 8 Seconds. By monitoring information entropy in real time. And calculate the entropy deviation. This allows us to obtain the current damage accumulation rate, and then obtain the cumulative fatigue damage through numerical integration. This damage value serves as the optimization target, guiding the iterative optimization of the bearing structural parameters.

[0033] Please see the appendix Figure 3 The load distribution information entropy is decomposed into a spectrum into a first frequency component and a second frequency component. A first control quantity is generated based on the first frequency component and a second control quantity is generated based on the second frequency component. The first control quantity is used to adjust the yaw angle and the common pitch angle, and the second control quantity is used to adjust the independent pitch angle. Furthermore, the load distribution information entropy is subjected to spectral decomposition, including: Wavelet packet decomposition is performed on the load distribution information entropy time series, and the reconstructed signal with a cutoff frequency below a preset Hertz is taken as the first frequency component; The second frequency component is obtained by separating the first frequency component from the original load distribution information entropy.

[0034] Furthermore, generating a first control quantity based on a first frequency component and a second control quantity based on a second frequency component respectively includes: Perform proportional-integral calculation on the first frequency component and output the yaw angle correction and common pitch angle increment as the first control quantity. The second frequency component and the current load distribution information entropy are used in a model predictive control algorithm to generate an independent pitch angle increment as the second control variable. The entropy budget is updated based on the current load distribution information entropy, and the second control variable is subjected to amplitude attenuation when the entropy budget is lower than the preset value of the theoretical maximum entropy.

[0035] Furthermore, a model predictive control algorithm is applied to the second frequency component and the current load distribution information entropy, including: A state-space prediction model is constructed, comprising characteristic values ​​of blade flapping bending moment, characteristic values ​​of wind turbine overturning moment components, and characteristic values ​​of rolling element load deviation. Set the prediction time-domain step number, control the time-domain step number and sampling period, and recursively predict the entropy deviation sequence in the time domain based on the second frequency component and the current load distribution information entropy. Construct an objective function with the sum of the weighted sum of entropy deviations, the sum of squared control increments, and the sum of squared control increment differences as indices, and solve for the control sequence corresponding to the objective function; The independent pitch angle increment of the first sampling step in the control sequence is extracted as the second control output.

[0036] Specifically, in this embodiment, in order to achieve active suppression of off-center load, it is necessary to separate the components of load distribution information entropy at different time scales and design control strategies for each. The off-center load sources of wind turbine main shaft bearings can be divided into two categories: one is the continuous off-center load bias caused by long-term asymmetry of wind resources, yaw error or changes in average wind speed, which changes relatively slowly; the other is the instantaneous off-center load fluctuation caused by turbulence, gusts or blade passing frequency, which changes relatively rapidly.

[0037] To apply control to these two different types of off-center loads, it is first necessary to perform spectral decomposition on the real-time information entropy, and denote the time series of load distribution information entropy as follows: The frequency band was divided using wavelet packet decomposition, with Daubechies-4 wavelet as the basis function and 5 decomposition levels. After wavelet packet decomposition, the low-frequency reconstructed signal with a cutoff frequency below 0.02 Hz was extracted as the first frequency component. This component reflects the persistent bias. Then, the first frequency component is separated from the original information entropy to obtain the second frequency component. This component reflects the instantaneous off-center load fluctuation; through this decomposition, the two time scales can be decoupled, providing independent input signals for subsequent slow-scale control and fast-scale control.

[0038] For the first frequency component A slow-scale controller is designed to eliminate persistent bias load, and a proportional-integral controller is used to process the first frequency component. An error signal is defined. ,in The proportional-integral controller outputs the yaw angle correction, which is the theoretical maximum entropy. and common pitch angle increment The calculation formula is: ; ; In this embodiment, the scaling factor Take 0.35 as the integral coefficient. Take 0.02; Take 0.28, The value is set to 0.015. The output amplitude of the yaw angle correction is limited to ±3 degrees, and the output amplitude of the common pitch angle increment is limited to ±2 degrees. The first control variable is... and Together they constitute.

[0039] For the second frequency component A fast-scale controller was designed to suppress instantaneous off-center load fluctuations, and a model predictive control algorithm was employed to handle the second frequency component and the current load distribution entropy. The core of model predictive control is to establish a state-space predictive model, whose state vector includes blade flapping moment eigenvalues, turbine overturning moment component eigenvalues, and rolling element load deviation eigenvalues. The prediction time-domain step number was set. Control the number of time-domain steps Sampling period Seconds. At each control moment, based on the second frequency component and current load distribution information entropy Recursive prediction of entropy deviation sequence in the time domain Construct the objective function: ;in For three independent pitch angle increments, , , In this embodiment, the weighting coefficient is used. Take 1000, Take 1, Take 0.1; The power exponent determined in step S4 is subject to constraints including an independent pitch angle increment amplitude not exceeding ±5°, a change rate not exceeding ±5° per second, and a blade root bending moment not exceeding the design limit. The control sequence corresponding to the objective function is solved, and the independent pitch angle increment of the first sampling step is extracted as the second control output.

[0040] During the control process, it is also necessary to dynamically adjust the output amplitude of the second control variable based on the current load distribution information entropy, and define the entropy budget. When the entropy budget is less than one-tenth of the theoretical maximum entropy, i.e. At the same time, the second control quantity is subjected to amplitude attenuation processing, for example, the output amplitude is multiplied by an attenuation factor of 0.5 to avoid over-adjustment causing system instability; through the above dual time scale control, the slow-scale controller continuously corrects long-term off-center load bias, and the fast-scale controller quickly suppresses instantaneous off-center load fluctuations. The two work together to keep the load distribution of the bearing rolling elements close to uniform.

[0041] Please see the appendix Figure 4 The trigger level is determined based on the load distribution information entropy and the rate of change of relative entropy, and the corresponding first and second control quantities are output when the trigger is triggered. Furthermore, the trigger level is determined based on the load distribution information entropy and the rate of change of relative entropy, including: Calculate the change in load distribution information entropy over a preset time period and combine it with the theoretical maximum entropy to obtain the relative entropy change rate; The level of silence, warning, or intervention is determined based on the ratio of the current load distribution information entropy to the theoretical maximum entropy and the relative entropy change rate. At the warning level, only the second control variable is output; at the intervention level, both the first and second control variables are output. When the current load distribution information entropy is at the lower end of the preset value of the historical distribution over the past preset hours, the trigger level will be determined as an intervention level.

[0042] Specifically, in this embodiment, to avoid frequent and meaningless actions of the control system and improve regulation efficiency, it is necessary to dynamically determine whether to initiate control and what level of control to initiate based on the current load distribution information entropy and its changing trend. When the information entropy is close to the theoretical maximum value and changes slowly, it indicates that the bearing is under uniform load and no intervention is required; when the information entropy decreases rapidly or falls below the threshold, control needs to be gradually intervened, defining the past... =Change in load distribution information entropy within 10 seconds The relative entropy change rate rH(t) is given by the following formula: ,in This represents the theoretical maximum entropy. In this embodiment, when... When this occurs, it is determined to be at the silent level, and no control input is output. When this condition is met, it is determined to be a warning level, and only the second control variable is output. When the threshold is reached, an intervention level is determined, and both the first and second control variables are output. Additionally, the historical load distribution entropy information of the past 72 hours is recorded, and an empirical cumulative distribution function is established. ,when When the trigger level is determined to be the intervention level, the trigger level is forcibly determined. Through this graded triggering mechanism, only fast independent pitch adjustment can be used when the off-center load is slight, and slow yaw correction can be activated when the off-center load is severe, thereby reducing the wear of the yaw mechanism and improving the stability of the control system.

[0043] Please see the appendix Figure 5 In the event of sensor failure, a virtual information entropy estimate is generated based on SCADA data to replace the load distribution information entropy. Furthermore, in the event of sensor failure, a virtual information entropy estimate is generated based on SCADA data to replace the load distribution information entropy, including: Calculate the predicted values ​​of each sensor based on the current rolling element load estimate, and compare the residual between the predicted value and the measured value with three times the standard deviation to determine the sensor fault. Using the SCADA data sequence within a preset time period as input, the pre-trained long short-term memory network is invoked to obtain the estimated value of the virtual information entropy output by the long short-term memory network; When all strain gauges and vibration acceleration sensors fail, the load distribution entropy is replaced with a virtual entropy estimate and the control output gain is reduced.

[0044] Specifically, to ensure the control system continues to function normally in the event of sensor failure, a virtual sensor needs to be trained using SCADA data from a healthy state to replace the failed physical sensor. Generally, strain gauges or vibration acceleration sensors may experience faults such as zero-point drift, signal interruption, or abnormal noise during long-term operation. Based on the current rolling element load estimate, the predicted values ​​for each sensor are calculated using a forward mapping model. Compare the predicted values ​​with the actual measured values residual and three standard deviations Comparison. When If the failure persists for more than 55 seconds, the sensor is considered faulty; during the healthy operation phase, SCADA data sequences and corresponding information entropy are collected. The network is trained using a long short-term memory network. The input to this network is a sequence of SCADA data from the past 60 seconds, including spindle speed, generator torque, three-blade pitch angle, yaw angle, wind speed, and wind direction. The output is a virtual information entropy estimate. When all strain gauge sensors and vibration acceleration sensors are faulty, use Replacement load distribution information entropy Simultaneously, the output gain of the first and second control variables is reduced to half of their original values. After any major sensor recovers and stabilizes for 30-30 seconds, the system switches back to the measured information entropy and restores the original gain. This redundancy design improves the system's reliability in harsh environments. Using bearing structural parameters as design variables and minimizing total cumulative fatigue damage over the entire life cycle as the objective, the system iteratively outputs the optimal structural parameter values ​​and control strategy parameters.

[0045] Furthermore, the iterative output of optimal structural parameter values ​​and control strategy parameters includes: Latin hypercube sampling is performed within the design space of bearing structural parameters to generate an initial set of sample points, each of which contains a set of bearing structural parameter values. For each sample point, iterate to obtain the total cumulative fatigue damage over the entire life cycle corresponding to each sample point; Using the Gaussian process as a surrogate model, the mapping relationship between sample points and total cumulative fatigue damage over the entire life cycle is fitted, and sequential sampling is performed with the desired improvement as the acquisition function. After iterative sampling to a preset number, the bearing structural parameter values ​​that minimize the total cumulative fatigue damage over the entire life cycle and their corresponding control strategy parameters are extracted as output.

[0046] Specifically, in this embodiment, in order to obtain the optimal combination of structural parameters for bearing fatigue life, the bearing structural parameters are used as design variables, and iterative optimization is performed with the goal of minimizing the total cumulative fatigue damage over the entire life cycle. The bearing structural parameters include bearing clearance. , Roller groove curvature radius coefficient Contact angle Roller logarithmic trimming amount c and roller convexity amount .

[0047] A Bayesian optimization method is used for global search. First, Latin hypercube sampling is performed within the design space to generate an initial set of sample points. Each sample point corresponds to a set of bearing structural parameter values. For each sample point, the closed-loop simulation in the previous steps is performed to obtain the corresponding total cumulative fatigue damage over the entire life cycle. Using a Gaussian process as a surrogate model, the sample points are fitted with... The mapping relationship between them. The covariance function of the Gaussian process uses the Matérn 5 / 2 kernel, and its form is: ,in and There are two sets of structural parameter vectors. Length scale The initial setting is one-fifth of the length of each parameter range.

[0048] Using the desired improvement as the acquisition function, the next sampling point is calculated on the surrogate model, and a closed-loop simulation is performed on that sampling point to obtain new results. And update the sample set; repeat the sampling and update steps until the number of iterations reaches 500; finally extract the samples. The minimum bearing structure parameter values ​​and their corresponding control strategy parameters are taken as the final output. Through this optimization process, the high-dimensional design space can be searched efficiently within limited computing resources to obtain the globally optimal bearing structure design scheme.

[0049] Example 2: In the second embodiment of the present invention, the present invention provides a wind turbine main shaft bearing structure design system based on fatigue life optimization, such as... Figure 2 As shown, it includes the following modules: Data acquisition module: used to acquire structural response data of the main shaft bearing under operating conditions and construct observation vectors. Based on the observation vectors, the rolling element load distribution is solved with maximum entropy as a constraint to obtain the estimated value of each rolling element load and calculate the load distribution information entropy. Damage calculation module: used to calculate the current damage accumulation rate based on the power law relationship between load distribution information entropy and fatigue damage accumulation rate, and to obtain the total cumulative fatigue damage over the entire life cycle by integrating over time. Control decomposition module: used to perform spectral decomposition on the load distribution information entropy, dividing it into a first frequency component and a second frequency component, generating a first control quantity based on the first frequency component and a second control quantity based on the second frequency component, respectively. The first control quantity is used to adjust the yaw angle and common pitch angle, and the second control quantity is used to adjust the independent pitch angle. Trigger determination module: used to determine the trigger level based on the load distribution information entropy and the rate of change of relative entropy, and output the corresponding first control quantity and second control quantity when triggered; Redundancy fault-tolerant module: used to generate a virtual information entropy estimate based on SCADA data to replace the load distribution information entropy when the sensor fails; The optimization iteration module is used to iteratively output the optimal structural parameter values ​​and control strategy parameters with bearing structural parameters as design variables and the goal of minimizing the total cumulative fatigue damage over the entire life cycle.

[0050] In large-scale wind farms, multiple megawatt-class wind turbines operate under complex wind conditions characterized by high turbulence and frequent yaw. The main shaft bearings commonly exhibit off-center loading due to time-varying combined loads, leading to early fatigue spalling of the raceway surface and an actual service life far below the design value. Traditional offline structural design methods cannot adapt to real-time operating condition changes and lack the ability to actively suppress off-center loading, resulting in high operation and maintenance costs. To address these issues, this invention employs a wind turbine main shaft bearing structural design system based on fatigue life optimization, the architecture of which is as follows: Figure 2 As shown. The specific implementation process of this system is as follows: First, the data acquisition module acquires the structural response data of the main shaft bearing under operating conditions and constructs an observation vector. Based on the observation vector, the rolling element load distribution is solved with the maximum entropy as a constraint to obtain the estimated value of each rolling element load and calculate the load distribution information entropy. Subsequently, the damage calculation module calculates the current damage accumulation rate based on the power law relationship between the load distribution information entropy and the fatigue damage accumulation rate, and integrates it over time to obtain the total cumulative fatigue damage over the entire life cycle. The control decomposition module performs spectral decomposition on the load distribution information entropy, dividing it into a first frequency component and a second frequency component. It generates a first control quantity based on the first frequency component and a second control quantity based on the second frequency component. The first control quantity is used to adjust the yaw angle and the common pitch angle, and the second control quantity is used to adjust the independent pitch angle. The trigger determination module determines the trigger level based on the load distribution information entropy and the rate of change of relative entropy, and outputs the corresponding first control quantity and second control quantity when triggered; The redundant fault-tolerant module generates a virtual information entropy estimate based on SCADA data to replace the load distribution information entropy when the sensor fails. Finally, the optimization iteration module uses the bearing structural parameters as design variables and aims to minimize the total cumulative fatigue damage over the entire life cycle, iteratively outputting the optimal structural parameter values ​​and control strategy parameters. Through the collaborative work of the above modules, the system can adaptively adjust the control strategy under different wind conditions and continuously optimize the bearing structural parameters, thereby effectively suppressing off-center loading and extending the actual fatigue life of the bearing.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for designing the structure of a wind turbine main shaft bearing based on fatigue life optimization, characterized in that, Includes the following steps: Obtain structural response data under the operating state of the main shaft bearing and construct an observation vector. Based on the observation vector, solve the rolling element load distribution with maximum entropy as a constraint, obtain the estimated value of each rolling element load and calculate the load distribution information entropy. The current damage accumulation rate is calculated based on the power-law relationship between the load distribution information entropy and the fatigue damage accumulation rate, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time. The load distribution information entropy is spectrally decomposed into a first frequency component and a second frequency component. A first control quantity is generated based on the first frequency component and a second control quantity is generated based on the second frequency component. The first control quantity is used to adjust the yaw angle and the common pitch angle, and the second control quantity is used to adjust the independent pitch angle. The trigger level is determined based on the load distribution information entropy and the rate of change of relative entropy, and the corresponding first control quantity and second control quantity are output when the trigger is triggered. In the event of sensor failure, a virtual information entropy estimate is generated based on SCADA data to replace the load distribution information entropy; Using the bearing structural parameters as design variables and minimizing the total cumulative fatigue damage over the entire life cycle as the objective, the optimal structural parameter values ​​and control strategy parameters are iteratively output.

2. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The process of acquiring structural response data under the operating state of the spindle bearing and constructing observation vectors includes: Obtain the axial and tangential strains of four sets of strain rosettes arranged circumferentially on the main shaft surface; The three-dimensional vibration acceleration of the bearing housing, the main shaft speed, the generator torque, the pitch angle of the three blades, the yaw angle, the wind speed, and the wind direction are obtained. The above signals are then aligned by time and combined into an observation vector.

3. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The process of obtaining estimated load values ​​for each rolling element and calculating load distribution information entropy includes: The observation vector is input into a pre-trained deep neural network to obtain a preliminary estimate of the rolling load output by the deep neural network. The preliminary estimate of the rolling element load is input into the extended Kalman filter joint observer for state recursion to obtain the filtered estimate of the rolling element load. The load distribution information entropy is extracted from the filtered rolling element load estimate.

4. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The total cumulative fatigue damage over the entire life cycle obtained by integrating over time includes: Obtain the dimensional parameters and power exponents corresponding to the bearing structural parameters; Extract the entropy deviation between the theoretical maximum entropy and the load distribution information entropy; The current damage accumulation rate is determined by the power exponent of the scale parameter and the entropy deviation, and the total cumulative fatigue damage over the entire life cycle is obtained by integrating over time.

5. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The spectral decomposition of the load distribution information entropy includes: Wavelet packet decomposition is performed on the load distribution information entropy time series, and the reconstructed signal with a cutoff frequency below a preset Hertz is taken as the first frequency component; The second frequency component is obtained by separating the first frequency component from the original load distribution information entropy.

6. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The steps of generating a first control quantity based on a first frequency component and generating a second control quantity based on a second frequency component include: Perform proportional-integral calculation on the first frequency component and output the yaw angle correction and common pitch angle increment as the first control quantity. The second frequency component and the current load distribution information entropy are used in a model predictive control algorithm to generate an independent pitch angle increment as the second control variable. The entropy budget is updated based on the current load distribution information entropy, and the second control variable is subjected to amplitude attenuation when the entropy budget is lower than the preset value of the theoretical maximum entropy.

7. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 6, characterized in that: The model prediction control algorithm that combines the second frequency component and the current load distribution information entropy includes: A state-space prediction model is constructed, comprising characteristic values ​​of blade flapping bending moment, characteristic values ​​of wind turbine overturning moment components, and characteristic values ​​of rolling element load deviation. Set the prediction time-domain step number, control the time-domain step number and sampling period, and recursively predict the entropy deviation sequence in the time domain based on the second frequency component and the current load distribution information entropy. Construct an objective function with the sum of the weighted sum of entropy deviations, the sum of squared control increments, and the sum of squared control increment differences as indices, and solve for the control sequence corresponding to the objective function; The independent pitch angle increment of the first sampling step in the control sequence is extracted as the second control output.

8. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The determination of the trigger level based on the load distribution information entropy and the rate of change of relative entropy includes: Calculate the change in load distribution information entropy over a preset time period and combine it with the theoretical maximum entropy to obtain the relative entropy change rate; The level of silence, warning, or intervention is determined based on the ratio of the current load distribution information entropy to the theoretical maximum entropy and the relative entropy change rate. At the warning level, only the second control variable is output; at the intervention level, both the first and second control variables are output. When the current load distribution information entropy is at the lower end of the preset value of the historical distribution over the past preset hours, the trigger level will be determined as an intervention level.

9. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The step of generating a virtual information entropy estimate based on SCADA data to replace the load distribution information entropy in the event of sensor failure includes: Calculate the predicted values ​​of each sensor based on the current rolling element load estimate, and compare the residual between the predicted value and the measured value with three times the standard deviation to determine the sensor fault. Using the SCADA data sequence within a preset time period as input, the pre-trained long short-term memory network is invoked to obtain the estimated value of the virtual information entropy output by the long short-term memory network; When all strain gauges and vibration acceleration sensors fail, the load distribution entropy is replaced with a virtual entropy estimate and the control output gain is reduced.

10. The wind turbine main shaft bearing structure design method based on fatigue life optimization according to claim 1, characterized in that: The iterative output of optimal structure parameter values ​​and control strategy parameters includes: Latin hypercube sampling is performed within the design space of bearing structural parameters to generate an initial set of sample points, each of which contains a set of bearing structural parameter values. For each sample point, iterate to obtain the total cumulative fatigue damage over the entire life cycle corresponding to each sample point; Using the Gaussian process as a surrogate model, the mapping relationship between sample points and total cumulative fatigue damage over the entire life cycle is fitted, and sequential sampling is performed with the desired improvement as the acquisition function. After iterative sampling to a preset number, the bearing structural parameter values ​​that minimize the total cumulative fatigue damage over the entire life cycle and their corresponding control strategy parameters are extracted as output.

Images