Method for monitoring the global posture and stress inversion of a wind power structure by mobile-fixed combination

Abstract

The application discloses a kind of mobile-fixed combination monitoring wind power structure global posture and stress inversion method, belong to wind motor technical field, including offline model training step and online real-time inversion step;Offline model training step includes: S11, obtains offline data;S12, training acceleration conversion model;S13, training acceleration-displacement conversion model;S14, training displacement-bending moment conversion model;Online real-time inversion step includes: S21, on-line reference acceleration data acquisition;S22, according to reference acceleration data, the displacement of wind turbine and the bending moment of target position are calculated;S23, according to the bending moment of target position, stress is calculated.The method of the application, only need to be fixedly arranged a fixed acceleration detection device, a movable acceleration detection device, i.e. wind turbine global posture and stress inversion can be realized, and the hardware cost is low.Low cost, high precision, high fidelity real-time inversion of global dynamic posture and internal stress state of wind turbine can be realized.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine technology, specifically, it relates to a method for inverting the attitude and stress of wind power structures across the entire domain based on a combination of mobile and fixed monitoring. Background Technology

[0002] Wind turbine towers are large, tall, and flexible structures that operate under complex dynamic loads such as wind, waves, currents, and earthquakes. Their structural response, particularly the dynamic displacement and attitude distributed throughout the entire tower, is a core physical quantity for structural safety assessment, fatigue life prediction, operational control strategy optimization, and the construction of digital twins. Therefore, how to economically, accurately, and in real-time acquire the tower's dynamic attitude and stress across its entire range is a long-standing and urgent technical challenge in the wind power industry's operation and maintenance field.

[0003] Currently, the main technological approach to achieving this goal is based on direct measurement and inference using fixed sensor arrays. This is the most traditional method, which involves permanently installing multiple sensors, such as accelerometers, strain gauges, or inclinometers, along the height of the tower. The drawback of this method is that:

[0004] ① Deploying an array of dozens of sensors for each wind turbine would incur enormous costs in terms of hardware, wiring, data acquisition, and long-term maintenance, making it economically unfeasible for large wind farms with over a hundred turbines.

[0005] ② Due to cost constraints, the density of sensor deployment is limited. This "point-to-area" monitoring method is "blind" to the vast area between sensors, and cannot capture non-uniform, high-order vibration modes caused by complex flow fields or local damage, which may miss critical safety hazards (such as fatigue at specific points).

[0006] Furthermore, there are significant limitations to directly installing strain gauges to measure stress and installing displacement gauges to measure displacement, for example:

[0007] ① Strain gauges are difficult to attach or operate in special locations on underwater or even mud-covered pile foundations, areas where wind turbine foundation fatigue is most severe. Furthermore, during strain gauge attachment, the tower or pile foundation experiences initial stress, the value of which is unknown; therefore, the strain gauge measurements do not accurately reflect the actual strain of the tower or pile foundation.

[0008] ② For displacement measurement, the industry typically employs GPS, laser Doppler vibration meter (LDV), or computer vision-based measurement techniques. While these techniques can provide high-precision displacement in certain scenarios, they have inherent limitations in wind turbine tower applications: GPS signals are prone to loss of lock under tower sway and cannot provide information on the tower's relative deformation. LDV and vision-based methods require a stable, unobstructed line of sight and a fixed reference point, which is almost impossible to achieve in vast onshore wind fields, especially offshore wind fields lacking fixed reference points. Adverse weather conditions such as sea fog, rain, snow, and changes in lighting conditions also severely interfere with their measurement accuracy and reliability.

[0009] Another approach is indirect displacement calculation based on acceleration signal integration. Due to the maturity, low cost, and ease of deployment of accelerometer technology, obtaining displacement by measuring acceleration and then integrating has become a widely studied alternative. However, this approach faces a fundamental technical bottleneck—baseline drift. Acceleration signals inevitably contain low-frequency noise, sensor zero-point drift, and unknown initial integration conditions (initial velocity and displacement). These minute errors are rapidly accumulated and amplified during the second integration process, forming a spurious, slowly changing trend term superimposed on the true displacement signal. This results in severely distorted displacement results, failing to reflect the true quasi-static deformation and low-frequency vibrations of the structure, which are crucial for assessing the overall stability and safety of the structure.

[0010] To overcome baseline drift, researchers have proposed various correction algorithms. For example, patent CN115752250A discloses a high-precision bridge displacement monitoring method integrating computer vision and acceleration. This method captures vibration time-history image sequences of measuring points on the bridge structure and simultaneously acquires the measured acceleration responses of these points. The dynamic displacement of the measuring points is obtained by numerically integrating the measured acceleration responses. The displacement time-history curves and dynamic displacements of the measuring points in the image coordinates are bandpass filtered, and a scaling factor is fitted using the least squares method. The two are then fused to obtain the high-precision displacement of the measuring points. This method applies a bandpass filter with a defined cutoff frequency to the integrated displacement signal, directly filtering out low-frequency trend terms and high-frequency noise. However, its drawback is that while filtering out drift, it inevitably filters out the inherent, true low-frequency displacement response components of the structure itself, resulting in information loss. Furthermore, this operation, which requires setting up imaging equipment as a fixed reference point, is difficult to implement at sea. The invention disclosed in CN115062662A proposes a vibration displacement frequency domain reconstruction method based on variational mode decomposition and generalized error control. This scheme decomposes the signal into a series of intrinsic mode components and manually selects intrinsic mode components within a certain frequency range. The drawback of this technique is that the selection of intrinsic mode components requires subjective judgment, and the entire process is difficult to automate.

[0011] Researchers have also proposed many approaches to global response reconstruction. For example, invention patent CN115752250A discloses a method and process for calculating the vibration state of wind turbine towers based on modal superposition. This technology establishes a high-fidelity finite element model of the tower, calculates the truncated mode matrix of the tower using the finite element method, calculates the principal vibration coordinate vector of the tower based on the real-time vibration values ​​at monitoring points, and finally calculates the real-time vibration state of the entire tower using the modal superposition method. This method relies on a high-fidelity finite element model, but the model correction technology required to establish a high-fidelity finite element model presents many challenges. In addition, obtaining the principal vibration coordinate vector also requires modal recognition technology, which necessitates densely arranging sensors on the tower to ensure smooth vibration modes and high spatial resolution, making the prerequisites for this method extremely difficult to achieve.

[0012] In summary, existing technologies are either limited by cost and spatial resolution, hindering large-scale application, or suffer from insufficient reliability due to environmental constraints. The seemingly most promising indirect computational path is hampered by the unresolved baseline drift problem, resulting in severely compromised displacement information fidelity. Furthermore, even if displacements at sparse points can be obtained, it is difficult to deduce global dynamic stress as the most direct criterion for structural safety. Traditional stress monitoring relies on strain gauges attached along the tower, which is also an expensive "point-based" measurement that cannot capture the full picture of stress concentration, and the strain gauges themselves are susceptible to environmental influences and may fail.

[0013] Therefore, there is an urgent need in this field for a novel technology that should be able to: 1) achieve "full-area" sensing of the tower at extremely low hardware cost; 2) accurately invert dynamic displacement from easily measurable physical quantities; and 3) further accurately calculate the dynamic stress field across the entire area based on the high-fidelity displacement field. Summary of the Invention

[0014] To address the technical problems of high cost and low accuracy and reliability of conventional inversion methods for obtaining the full-domain dynamic attitude and stress of existing wind turbines, this invention proposes a wind turbine full-domain attitude and stress inversion method, which can solve the above problems.

[0015] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0016] A method for inverting the attitude and stress of a wind turbine structure under a combination of mobile and fixed monitoring, wherein the wind turbine's entire domain includes a monopile and a tower supported on the monopile, and the method includes an offline model training step and an online real-time inversion step.

[0017] Offline model training steps include:

[0018] S11. Acquiring offline data includes: selecting an arbitrary height position on the tower as a reference position, and fixing a fixed acceleration detection device at the reference position. The fixed acceleration detection device continuously collects acceleration data at the reference position, which is offline reference acceleration data; selecting several tower measurement positions along the tower height direction, and inspecting the acceleration detection device moving along the tower height direction. When moving to each tower measurement position in sequence, the device continuously collects acceleration data for a set duration.

[0019] Several single-pile inversion positions are selected along the height direction of the single pile;

[0020] A finite element model of the wind power structure is established, and environmental parameters are input into the finite element model. The finite element model outputs the acceleration simulation data, displacement simulation data, and bending moment simulation data of the tower, as well as the bending moment simulation data of a single pile. The tower includes a reference position and a tower measurement position.

[0021] S12. Construct an acceleration conversion model between a fixed reference position and a measurement position. Take the offline reference acceleration data and the height data of each tower measurement position as input, and the acceleration data collected at each tower measurement position as output, and train the acceleration conversion model.

[0022] S13. Train the acceleration-displacement conversion model by taking the tower's acceleration simulation data as input and displacement simulation data as output.

[0023] S14. Train the displacement-bending moment conversion model by selecting displacement simulation data at several locations on the tower as input, selecting several locations on the tower and / or single pile as target locations, and using the bending moment simulation data at the target locations as output.

[0024] The online real-time inversion steps include:

[0025] S21. Acquire acceleration data continuously collected by the fixed acceleration detection device at the reference position online, which serves as online reference acceleration data;

[0026] S22. Using the acceleration conversion model, acceleration-displacement conversion model, and displacement-bending moment conversion model, calculate the displacement of the wind turbine and the bending moment at the target position based on the online reference acceleration data;

[0027] S23. Calculate the stress based on the bending moment at the target location.

[0028] Compared with existing technologies, the advantages and positive effects of this invention are as follows: The mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion of this invention constructs acceleration conversion models, acceleration-displacement conversion models, and displacement-bending moment conversion models respectively. The acceleration conversion model can invert the acceleration data of the entire tower along the height direction based on the acceleration data of a single fixed position. The acceleration-displacement conversion model can invert the displacement data of that position based on the acceleration data of the tower. The displacement-bending moment conversion model can invert the bending moment data of the entire wind turbine based on the displacement data of several positions on the tower. The displacement data can be used to plot the global attitude of the wind turbine, and the bending moment data can be used to calculate the stress at the target position of the entire wind turbine. This method realizes the inversion of the global attitude and stress of the wind turbine by simply setting up a fixed acceleration detection device at an arbitrarily selected reference position on the tower. The hardware cost of this solution is low.

[0029] The acceleration conversion model is trained by combining a dynamic and static motion sensing concept with two advanced deep learning models. Utilizing spatiotemporal asynchronous response reconstruction technology, it eliminates the reliance on expensive fixed sensor arrays, achieving virtual sensing of the entire acceleration field. Then, using a deep generative model based on probability distribution learning, an end-to-end mapping from acceleration to physical displacement is directly established, fundamentally avoiding the baseline drift problem of traditional integral methods. This method allows for low-cost, high-precision, and high-fidelity real-time inversion of the wind turbine's dynamic attitude and internal stress state across the entire field, achieved simply by deploying one fixed acceleration detection device and one mobile acceleration detection device on the tower.

[0030] Other features and advantages of the present invention will become clearer after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. Attached Figure Description

[0031] Figure 1 This is a flowchart of an embodiment of the wind power structure global attitude and stress inversion method based on mobile-fixed combined monitoring proposed in this invention;

[0032] Figure 2 This is a schematic diagram of the entire wind power structure in one embodiment of the mobile-fixed combined monitoring method for inverting the attitude and stress of the wind power structure proposed in this invention;

[0033] Figure 3 This is a schematic diagram of global acceleration response reconstruction based on spatiotemporal asynchronous signals in one embodiment of the mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion proposed in this invention;

[0034] Figure 4This is a comparison chart of the acceleration inversion effect at the measurement position of tower No. 6 and the actual value in one embodiment of the mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion proposed in this invention;

[0035] Figure 5 This is a schematic diagram of the time history reconstruction of the displacement of key nodes in one embodiment of the mobile-fixed combined monitoring method for inverting the attitude and stress of wind power structures in the whole domain proposed in this invention;

[0036] Figure 6 This is an example of the wind power structure global displacement inversion effect diagram in one embodiment of the mobile-fixed combined monitoring global attitude and stress inversion method proposed in this invention;

[0037] Figure 7 This is an example of the single pile bottom bending moment inversion effect diagram in one embodiment of the mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion proposed in this invention;

[0038] Figure 8 This is a schematic diagram of the U-Net network architecture based on spatiotemporal feature modulation in one embodiment of the mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion proposed in this invention;

[0039] Figure 9 yes Figure 8 Schematic diagram of the mid-temporal feature modulation architecture;

[0040] Figure 10 This is a schematic diagram of the displacement response reconstruction network architecture in one embodiment of the mobile-fixed combined monitoring method for wind power structure global attitude and stress inversion proposed in this invention;

[0041] In the diagram: 110, monopile; 120, tower; 130, fixed acceleration detection device; 140, inspection acceleration detection device. Detailed Implementation

[0042] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0044] It should be noted that in the description of this invention, terms such as "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," indicating directional or positional relationships, are based on the directional or positional relationships shown in the accompanying drawings. These are merely for ease of description and do not indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation; therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0045] Example 1: This invention proposes a method for global attitude and stress inversion of wind turbines, such as... Figure 2 As shown, the entire wind turbine domain includes a monopile 110 and a tower 120 supported on the monopile. The wind turbine domain attitude and stress inversion method includes an offline model training step and an online real-time inversion step.

[0046] Combination Figure 1 , Figure 2 As shown, the offline model training steps include:

[0047] S11. Acquire offline data, including: selecting an arbitrary height position on the tower as a reference position, with a fixed acceleration detection device 130 fixedly installed at the reference position. The fixed acceleration detection device continuously collects acceleration data at the reference position, which serves as offline reference acceleration data. Several tower measurement positions are selected along the tower height direction. The inspection acceleration detection device 140 moves along the tower height direction, continuously collecting acceleration data for a set duration as it moves to each tower measurement position. This step obtains asynchronous data collected by the fixed acceleration detection device and the moving inspection acceleration detection device.

[0048] Several single-pile inversion positions are selected along the height direction of the single pile.

[0049] A finite element model (FEM) of the wind turbine structure is established. Environmental parameters are input into the FEM, which then outputs simulated acceleration, displacement, and bending moment data for the tower, as well as simulated bending moment data for a single pile. The full-domain acceleration simulation data of the wind turbine structure output by the FEM increases the sample size of acceleration data, solving the problem of low resolution in existing detection data. The output simulated displacement and bending moment data can be obtained without the need for additional detection elements, further addressing the issue of insufficient sample size.

[0050] S12. Construct an acceleration conversion model between a fixed reference position and measurement positions. Use offline reference acceleration data and height data from each tower measurement position as input, and acceleration data collected from each tower measurement position as output, to train the acceleration conversion model. The acceleration conversion model is used to establish the relationship between the acceleration data at the reference position and the global acceleration data of the wind turbine.

[0051] S13. Train the acceleration-displacement conversion model. Use the tower's acceleration simulation data as input and displacement simulation data as output to train the acceleration-displacement conversion model. The acceleration-displacement conversion model is used to establish the relationship between the acceleration at any position on the tower and the displacement at that position.

[0052] S14. Train the displacement-bending moment conversion model. Select simulated displacement data at several locations on the tower as input, select several locations on the tower and / or monopile as target locations, and output simulated bending moment data at the target locations to train the displacement-bending moment conversion model. The displacement-bending moment conversion model is used to establish the relationship between the displacement on the tower and the bending moment over the entire wind power structure.

[0053] The online real-time inversion steps include:

[0054] S21. Acquire acceleration data continuously collected by the fixed acceleration detection device at the reference position online, which serves as online reference acceleration data.

[0055] S22. Using the acceleration conversion model, acceleration-displacement conversion model, and displacement-bending moment conversion model, calculate the displacement of the wind turbine and the bending moment at the target location based on the reference acceleration data.

[0056] Specifically, by inputting online reference acceleration data into the acceleration conversion model, the wind turbine's global acceleration data can be obtained.

[0057] By inputting the wind turbine's global acceleration data into the acceleration-displacement conversion model, the wind turbine's global displacement data can be obtained.

[0058] By inputting the displacement data of the entire wind turbine into the displacement-bending moment conversion model, bending moment data of the entire wind turbine can be obtained. Users can select the bending moment of the location of interest (such as a single pile) for analysis and research as needed, providing direct and key basis for the safety assessment, fatigue life prediction and digital twin of wind turbine structure.

[0059] S23. Calculate the stress based on the bending moment at the target location.

[0060] Taking into account the predicted dynamic bending moment and axial force, the dynamic stress distribution of a single pile section is accurately calculated using formulas from mechanics of materials.

[0061] The wind turbine global attitude and stress inversion method in this embodiment constructs an acceleration conversion model, an acceleration-displacement conversion model, and a displacement-bending moment conversion model. The acceleration conversion model can invert the acceleration data of the entire tower along the height direction based on the acceleration data of a single fixed position. The acceleration-displacement conversion model can invert the displacement data of that position based on the acceleration data of the tower. The displacement-bending moment conversion model can invert the bending moment data of the entire wind turbine based on the displacement data of several positions on the tower. The displacement data can be used to plot the global attitude of the wind turbine, and the bending moment data can be used to calculate the stress at the target position of the wind turbine. This method achieves the inversion of the global attitude and stress of the wind turbine by simply setting up a fixed acceleration detection device at an arbitrarily selected reference position on the tower. The hardware cost of this solution is low.

[0062] The acceleration conversion model is trained by combining a dynamic and static motion sensing concept with two advanced deep learning models. Utilizing spatiotemporal asynchronous response reconstruction technology, it eliminates the reliance on expensive fixed sensor arrays, achieving virtual sensing of the entire acceleration field. Then, using a deep generative model based on probability distribution learning, an end-to-end mapping from acceleration to physical displacement is directly established, fundamentally avoiding the baseline drift problem of traditional integral methods. This method allows for low-cost, high-precision, and high-fidelity real-time inversion of the wind turbine's dynamic attitude and internal stress state across the entire field, achieved simply by deploying one fixed acceleration detection device and one mobile acceleration detection device on the tower.

[0063] In some embodiments, the acceleration data collected at the reference location is continuous in time, while the acceleration data collected at each tower measurement location is only data for a specific time period, and the acceleration data at each tower measurement location is also discontinuous in time. To address the technical problem of discontinuous time, step 12 further includes a step of segmenting the offline reference acceleration data. Based on the start and end times of the acceleration data collected at each tower measurement location, the offline reference acceleration data is segmented along the time axis to obtain several segments of offline acceleration data. The number of segments is consistent with the number of acceleration data segments at the tower measurement locations. In this way, the reference acceleration data segments are aligned with the acceleration data at the tower measurement locations on the time axis.

[0064] Similarly, it also includes the step of cutting the output displacement and bending moment data at each measurement location along the time axis, and training the acceleration-displacement conversion model and the displacement-bending moment conversion model.

[0065] Since intermittent measurements continuously acquire acceleration data at any location on the tower, special techniques are needed to reconstruct the acceleration signals at unmeasured locations using the measured locations. In this embodiment, deep learning is used to train a global acceleration reconstruction model, also known as an acceleration conversion model.

[0066] In this embodiment, when training the acceleration conversion model, the offline acceleration data of each segment and the height data of each tower measurement position are used as inputs, and the acceleration data collected at each tower measurement position is used as outputs. The acceleration conversion model is trained using a spatiotemporal feature modulation network.

[0067] Using the continuous acceleration data collected at the reference position in step S11 and the spatiotemporally asynchronous acceleration data collected during movement, along with their corresponding spatial location information, the first deep learning model is trained. This enables it to output an acceleration response at a corresponding height based on the input continuous acceleration signal and any specified height coordinates.

[0068] The acceleration conversion model is a deep learning model employing a spatiotemporal feature modulation (SMT) U-Net architecture. The U-Net architecture includes encoder-decoder structures of varying sizes. Each encoder module contains a residual connection layer, a SMT layer, and an attention mechanism layer with a random deactivation layer. Adjacent encoder modules are connected via a max-pooling layer. Each decoder module contains a residual connection layer, a SMT layer, and an attention mechanism layer with a random deactivation layer. Adjacent encoder and decoder modules, as well as adjacent decoder modules, are connected via an upsampling layer. Decoders and encoders at symmetrical positions are skipped connections. The SMT layer receives intermediate features output from the residual connection layer. and height data of each tower measurement location Mapping a one-dimensional height value to a scale vector and an offset vector The inverted acceleration data is output through element-wise affine transformation.

[0069] In some embodiments, step S11 includes: establishing a finite element model of the wind power structure based on the design parameters of the wind turbine.

[0070] The simulation of the structural response of the wind turbine under different environmental parameters includes randomly selecting environmental parameters from the environmental load spectrum of the operating sea area and inputting the environmental parameters into the finite element model, thereby simulating the structural vibration response of the wind power structure under different environmental conditions. The sampling frequency of the finite element model output data is consistent with the sampling frequency of the fixed acceleration detection device and the inspection acceleration detection device. The environmental parameters include any combination of wind speed, wind direction, wave height, and wave direction.

[0071] In some embodiments, the acceleration conversion model is a deep learning model, employing a spatiotemporal feature modulation U-Net architecture.

[0072] In some embodiments, the acceleration-displacement conversion model and the displacement-bending moment conversion model adopt a deep generative model based on continuous normalized flow (CNF), and flow matching technology is used during training. The objective function is:

[0073] .

[0074] in, and These are the model's input and output, respectively. The parameters of the neural network, This refers to a stage in the flow process. Representation phase The output value of the neural network at that time. This indicates the calculation of the expected value.

[0075] In some embodiments, step S23 uses a curve fitting method to connect the displacements of all virtual measuring points at the same time in space and draw the deformation posture curve at that time.

[0076] In some embodiments, step S23 further includes using a preset function fitting method to fit the deformation posture curve to generate a continuous deformation posture curve covering the entire tower.

[0077] In some embodiments, step S11 further includes a step of normalizing the acceleration data collected at the tower measurement location and the data output by the finite element model, converting the original physical scale signal into... Signals within an interval.

[0078] In some embodiments, step S22 includes:

[0079] S221, Acceleration reconstruction steps, including:

[0080] Several virtual measuring points are selected from the single pile and the tower along the height direction.

[0081] The online reference acceleration data is segmented on the time axis, with the segment length being the same as the acquisition duration of the acceleration data collected at the tower measurement location during the offline model training step, resulting in several segments of online acceleration data. The start time of each segment of online acceleration data is recorded.

[0082] The online acceleration data for each segment, along with the reference acceleration data and the height h of each virtual measuring point, are input into the acceleration conversion model. The output value of the acceleration conversion model is then compared with the first acceleration normalization factor. Multiplying these results yields the acceleration at each virtual measurement point, representing the inversion acceleration at each virtual measurement point within its corresponding time range. The first acceleration normalization factor is then applied. The same normalization factor was used when training the acceleration conversion model.

[0083] The original data needs to be normalized before training, but the normalized data has no actual physical meaning. Therefore, the output data after the acceleration-displacement conversion model is dimensionless and needs to be multiplied by the corresponding acceleration normalization factor to obtain acceleration with physical meaning.

[0084] S222, Displacement Inversion and Attitude Synthesis Steps, including:

[0085] The inversion acceleration of each virtual measuring point is input into the acceleration-displacement conversion model. The output value of the acceleration-displacement conversion model is compared with the second acceleration normalization factor. Multiplying these values ​​yields the displacement of each virtual measuring point, representing the inverted displacement of each virtual measuring point within its corresponding time range. Similarly, the output values ​​of the acceleration-displacement conversion model also need to be converted into physically meaningful displacements. Second acceleration normalization factor. The same normalization factor was used when training the acceleration-displacement conversion model.

[0086] Connect the inversion displacements of all virtual measurement points at the same moment in space to plot the deformation attitude curve at that moment.

[0087] S223. Dynamic stress field calculation steps: Select the inversion displacement of several virtual measuring points and input it into the displacement-bending moment conversion model. The output value of the displacement-bending moment conversion model and the displacement normalization factor are... Multiplying these values ​​yields the global inversion bending moment within the corresponding time range. Users can then select the bending moment at the target output location from the global inversion bending moment. Similarly, the output values ​​of the displacement-bending moment conversion model also need to be converted into physically meaningful bending moments.

[0088] Displacement normalization factor The same normalization factor was used when training the displacement-moment conversion model.

[0089] S224. The inversion acceleration, inversion displacement, deformation attitude curves and inversion bending moment in each time range are spliced ​​together on the time axis to obtain time-continuous inversion data.

[0090] In some embodiments, the method for calculating stress in step S25 includes:

[0091] S231. Establish a rectangular coordinate system with the front-to-back direction of the wind turbine as the X-axis, the left-to-right direction as the Y-axis, the central axis of the tower as the Z-axis, and the center of the single pile as the origin.

[0092] S232, Stress at target position g for:

[0093] .

[0094] in, and Let be the bending moments at the target location g in the X and Y directions, respectively. The height of the target position g The above structural self-weight, For height The cross-sectional area of ​​the tower at that location, For height The radius of the tower section at that location, For height Moment of inertia of the tower section at the location, Let g be the angle between the line connecting the target position g and the central axis of the tower and the X-axis.

[0095] In Example 2, in step S11, a selection is made on a single pile. The inversion position for each single pile is selected on the tower. Each tower section was measured at a specific location and numbered from bottom to top. The corresponding height is The selected measurement locations should include at least the location of the wind turbine foundation mud surface, the tower foundation connection location, the location of each tower section, and the location of the top of the tower.

[0096] Measurement location from the tower Select a reference position Record the measured position and height. A fixed acceleration detection device is installed at this location to continuously collect acceleration signals from the reference position. The sampling frequency during the test is... A signal with a frequency not lower than 20Hz, acquired by a fixed acceleration detection device, is called a continuous acceleration signal, denoted as . ,in, Indicates the length as an acceleration signal, subscript This indicates that the signal comes from a fixed acceleration detection device, indicated by the superscript. Indicates a reference position.

[0097] The fixed acceleration detection device continuously collects acceleration data at the reference position, which is offline reference acceleration data.

[0098] Measurements from different locations on the tower First, select at least one position. Intermittent measurements are performed. The mobile inspection acceleration detection device collects the tower's acceleration signal when passing through selected locations during ascent or descent, with a measurement duration of not less than [number missing]. ,in A positive integer, sampling frequency Simultaneously record the height of the measurement location. After the measurement is completed, the movable inspection acceleration detection device continues to move along the tower axis to the designated position. Continue measuring for no less than Time, sampling frequency Simultaneously record the measured position and height. This intermittent measurement process is repeated until all or most of the tower measurement locations are covered, forming a height set of intermittent measurement locations. Acceleration signals obtained from intermittent measurements exhibit spatiotemporal asynchrony, meaning they are spatially dense and discrete but temporally segmented and continuous; this is called a spatiotemporal asynchrony acceleration signal, denoted as . ,in, Indicates acceleration signal, subscript Indicates that the signal originates from a moving sensor, indicated by the superscript. Indicates the location of intermittent measurements.

[0099] In some embodiments, the fixed acceleration detection device 130 can be implemented using an acceleration sensor, while the inspection acceleration detection device can be a magnetically attached wall-climbing robot equipped with a high-precision acceleration sensor. This robot inspects the outer wall of the tower according to a preset path. The inspection mode involves the robot moving to a preset position and stopping to measure, thereby obtaining a set of spatially dense but temporally discontinuous inspection acceleration signals. These signals record the acceleration response at different time ranges and locations.

[0100] When segmenting the acceleration data acquired at the reference location, a sliding window method is used to divide each segment of the spatiotemporal asynchronous acceleration signal. Cut out Duration signal fragments And record the start and end times of each signal segment. and ,in and For the first The start and end times of each signal segment This constitutes the intermittent acceleration measurement dataset. Simultaneously, the time measured intermittently. and For continuous acceleration signals Perform appropriate cutting to obtain the length. signal fragments Furthermore, these data points coincide with the start and end times of the intermittent measurement signal segments, forming an offline acceleration dataset. .

[0101] Finite element model simulation of wind turbine output under different environmental parameters, each simulation duration is not less than ,in The value is a positive integer, and the reference position is also output. and intermittent measurement locations Acceleration and displacement signals, all measurement positions The bending moment signal, the sampling frequency of the output data is .

[0102] In some embodiments, the method further includes using a sliding window to segment the displacement and bending moment data at each measurement location to form... Duration The acceleration, displacement, and bending moment signal segments. Based on this, the reference position... Construct a continuous acceleration simulation dataset. Continuous displacement simulation dataset Continuous bending moment simulation dataset Measurement location Construct an acceleration simulation dataset Displacement simulation dataset With bending moment simulation dataset .

[0103] The acceleration simulation signals at the reference position and the measurement position are normalized respectively to form a normalized continuous acceleration measurement dataset. With normalized acceleration measurement dataset The superscript wavy line indicates the normalized data. and As input, As output, train the acceleration-acceleration conversion model. This enables it to output the corresponding intermittent measurement position acceleration based on the continuous measurement position acceleration and the intermittent measurement position height.

[0104] Normalization is the process of transforming signals at the original physical scale into... A normally distributed signal within an interval.

[0105] When constructing the displacement-moment transformation model, select the key stress locations that need to be inverted. ,high From the dataset and Select the bending moment data subset corresponding to this location. Normalize it. At the same time, from the dataset... and Select Data subset corresponding to each measurement location , respectively and After normalization, we get and .Will As input, As output, the displacement-moment transformation model is trained. This enables it to output the bending moment at the key stress position based on the displacement of the measured position.

[0106] Acceleration conversion model For deep learning models, such as Figure 8 As shown, a spatiotemporal feature modulation (SMT) U-Net architecture is adopted. This embodiment uses encoder-decoder structures of different sizes and skip connections. Each encoder module includes a residual connection layer, a SMT layer, and an attention mechanism layer containing a random deactivation layer. Adjacent encoder modules are connected through a max-pooling layer. As the depth of the U-Net architecture increases, the width and height of the input data of the encoder module gradually decrease, while the number of channels increases. Each decoder module includes a residual connection layer, a SMT layer, and an attention mechanism layer containing a random deactivation layer. Adjacent encoder and decoder modules, as well as adjacent decoder modules, are connected through an upsampling layer. Decoders and encoders at symmetrical positions are skip connections. Finally, the encoder output is processed by a convolutional layer to generate the inverted acceleration. .

[0107] Spatiotemporal feature modulation layer, such as Figure 9 As shown, the intermediate features output by the residual connection layer For one of the inputs, copy the intermediate features Then, each copied intermediate feature is fed into an average pooling layer branch of different sizes. Each branch also contains a convolutional layer and a nearest neighbor upsampling layer to obtain the intermediate features. , , , , Then and , , , , The aggregated features are obtained after passing through a convolutional layer and a GELU activation layer. The spatiotemporal feature modulation layer also includes height data from the measurement locations of each tower. As another input, map the one-dimensional height value to a scale vector. and an offset vector Finally, the normalized measured position height is output through an element-wise affine transformation. Inversion acceleration characteristics:

[0108] .

[0109] Where ⊙ represents element-wise multiplication, For acceleration features, inverted acceleration is generated through convolutional layers. Here, each branch utilizes average pooling operations at different scales, acting as a bandpass filter to separate dynamic information across different frequency bands. The large-scale pooling branch, by averaging over a larger time window, effectively filters out high-frequency noise and rapid fluctuations, thereby capturing low-frequency, global trend features related to lower-order modes in the signal. The small-scale pooling branch retains more temporal details, focusing on capturing high-frequency, rapidly changing feature components related to higher-order modes or local vibrations. This summary information is transformed by its respective convolutional layer and then upsampled to the original sequence length. Subsequently, these K "feature spectra" containing multi-scale information are concatenated along the channel dimension and integrated through a final fusion convolutional layer.

[0110] Acceleration-displacement transformation model and displacement-moment conversion model All are deep learning models, and both use the U-Net architecture. Taking the acceleration-displacement transformation model as an example, such as... Figure 10 As shown, encoder-decoder structures of different sizes and skip connections are used, arranged in a symmetrical configuration. Each encoder module contains an embedded flow stage. The U-Net architecture consists of a positional encoding layer, an attention mechanism layer, and adjacent encoder modules are connected by a max-pooling layer. As the depth of the U-Net architecture increases, the width and height of the input data for the encoder modules gradually decrease, while the depth increases. Each decoder module contains an embedding flow stage. The system consists of a positional encoding layer, an attention mechanism layer, and two adjacent decoder modules connected by an upsampling layer. The dimensions of the decoder are exactly the opposite of the corresponding encoders, and the decoders and corresponding encoders at symmetrical positions are connected in a skip connection.

[0111] During the response inversion phase, the moving sensor can be directly removed, and only the fixed reference accelerometer can be used to continue measuring the reference position. The acceleration signal is continuously measured. Therefore, if this method directly uses the acceleration sensor of the SCADA system, no additional sensor needs to be installed, which is a major advantage of this technology.

[0112] Continue to monitor the reference position using a fixed acceleration detection device. The acceleration signal is continuously measured at intervals of time. The continuous acceleration signal is divided into segments with a duration of [duration value missing] using a sliding window method. For each signal segment, perform the normalization process as described above, and record the standard deviation of the signal during normalization, i.e., the first acceleration normalization factor. This constitutes a continuous acceleration measurement dataset. .

[0113] Use a trained acceleration-acceleration conversion model Acceleration-displacement conversion model and continuous acceleration measurement dataset The overall attitude of the wind power structure is then retrieved. The specific process is as follows:

[0114] S211, Transfer the continuous acceleration measurement dataset Height of intermittent measurement position Input a pre-trained acceleration-accelerometer conversion model Generate intermittent measurement locations The corresponding dimensionless acceleration dataset multiplied by Get the actual acceleration .

[0115] S212. Normalize the acceleration signals at all altitudes, and simultaneously record the second acceleration normalization factor during signal normalization. ,Will Input a pre-trained acceleration-displacement conversion model Generate a dimensionless displacement dataset Multiply by , to obtain continuous and intermittent measurement positions inversion displacement .

[0116] S213. For onshore wind turbines, the bottom of the tower is taken as the zero point of displacement; for offshore wind turbines, the seabed position is taken as the zero point of displacement. A curve fitting method is used to... Fitting is performed to obtain a continuous and smooth deformation posture curve covering the entire wind turbine foundation and tower.

[0117] Using a trained displacement-moment conversion model Inversion of key stress locations The corresponding bending moment, the specific process is as follows:

[0118] S231, From the set of displacements Select Displacement dataset corresponding to each measurement location Normalization is performed on the data to obtain a normalized displacement dataset. Simultaneously record the displacement normalization factor. .

[0119] S232, will Input into the trained displacement-moment conversion model Inversion of key stress locations Corresponding dimensionless bending moment Multiply by To obtain the key stress locations inversion bending moment .

[0120] S233. Determine the forward / backward and left / right directions of the wind turbine based on its yaw. Establish a coordinate system with the forward / backward direction as the X-axis, the left / right direction as the Y-axis, the tower centerline as the Z-axis, and the tower base center as the origin. According to the formula...

[0121]

[0122] Calculate the location of critical stress The stress. Among them, and These are the positions retrieved in step S132. The bending moments in the X and Y directions, For height The above refers to the structural self-weight (including the total structural weight above the critical sections of the wind turbine foundation, such as the nacelle, impeller, and upper tower). For height The cross-sectional area of ​​the tower at that location, For height The radius of the tower section at that location, For height Moment of inertia of the tower section at the location, The angle between the line connecting the accelerometer and the central axis of the tower and the X-axis.

[0123] Example 3: Numerical calculation to verify the example.

[0124] 1. Simulation platform and model establishment

[0125] This simulation example uses the OpenFAST simulation platform developed by NREL and its NREL 5MW offshore wind turbine benchmark model. This example assesses the stress at the tower base; therefore, it only analyzes the acceleration-displacement-bending moment transformation of the tower. The wind turbine tower model was discretized into 49 beam elements, containing a total of There are nodes, numbered from bottom to top. The top of the tower (node ​​50) was selected as the reference position, and a dual-axis accelerometer was installed there. Its acceleration response was measured using... Indicates other locations To measure the intermittent locations, a magnetically attached wall-climbing machine with a dual-axis accelerometer was used to measure the movement of these locations.

[0126] Wind and wave loads are the main environmental loads acting on offshore wind turbines. In this embodiment, the IECKaimal model is used to simulate turbulent wind conditions. Wind speed is assumed to follow a normal distribution N(11.4,2) with a mean of 11.4 m / s and a standard deviation of 2 m / s, and wind direction is assumed to follow a uniform distribution U(-20,20) ranging from -20 degrees to 20 degrees. Furthermore, random wave loads are generated based on the environmental load spectrum, where wave height and wave period are modeled as random variables following distributions N(3.8,0.5) m and N(5.5,0.2) s, respectively. Wave direction is assumed to be consistent with the wind direction.

[0127] The simulation included the wind turbine's blade pitch, nacelle yaw, and rotor rotation. Initial conditions were set as follows: blade pitch angle 0°, nacelle yaw angle 0°, rotor speed 12.1 rad / min, and uniform node distribution on the tower. In this embodiment, 50 independent random environmental condition simulations were performed. Under each condition, the wind turbine simulation lasted 30 minutes, with a data sampling frequency of 50 Hz, generating 90,000 data points in both the X and Y directions for each node. Environmental conditions (such as average wind speed and significant wave height) for each condition were independently and randomly sampled from a preset probability distribution, ensuring the diversity of conditions. Under the combined action of environmental and operational loads, the tower's acceleration, displacement, and bending moment responses were recorded for verification of the inversion method.

[0128] 2. Establishment of training and validation datasets

[0129] (1) Data output mode one, used to train and construct an acceleration conversion model between the reference position and the intermittent measurement position. The data extracted from this simulation reflects the following actual test conditions: Assuming movement is detected by the accelerometer sensor on the wall-climbing robot, and each movement of the robot constitutes one working condition, then:

[0130] The first working condition simulates and outputs the synchronous acceleration / displacement data of the tower top and node numbered 1.

[0131] The second working condition simulates and outputs the synchronous acceleration / displacement data of the tower top and node numbered 1.

[0132] This process continues until the 50th operating condition, at which point the simulation outputs the synchronous acceleration data of the tower top and the node numbered 50.

[0133] The resulting data is a "spatiotemporally asynchronous" signal. While the data is time-synchronized within a single set of conditions (e.g., between the tower top and node numbered i), the data between different sets of conditions (e.g., between node numbered i and node numbered j) is completely asynchronous in both time and space. This design faithfully simulates the actual scenario of a wall-climbing robot measuring different positions on the tower under different times and conditions. Figure 3 As shown.

[0134] To this end, displacement data for 50 nodes and 50 sets of working conditions are output in both the X and Y directions. The motion measurement lasts for 2 minutes, and all data are processed according to... The data points are divided into 17250 signal segments, resulting in int(2×60×50 / 256)×15×50=17250 segments. 80% of these segments, totaling 13800 segments, are selected for further analysis. The training was carried out, and the remaining 20%, a total of 3450 signal segments, were used to test the trained model.

[0135] (2) Data output mode two, used for training acceleration-displacement conversion model Output acceleration and displacement data for 50 nodes and 50 sets of working conditions in both the X and Y directions. Arrange all data according to... The data points are divided into 877,550 signal segments, resulting in an int(30×60×50 / 256)×50×50 = 877,550 segments. 80% of these segments, totaling 702,000, are selected for further analysis. The training was carried out, and the remaining 20%, a total of 175,500 signal segments, were used to test the trained model.

[0136] (3) Data output mode three, used for training displacement and moment conversion model Select displacement data for five nodes (numbered 50, 40, 30, 20, and 10) in both the X and Y directions, and output the bending moment data at the tower base (number 1). Then, process all data according to... The data points are divided into 17550 signal segments, resulting in int(30×60×50 / 256)×50=17550 segments. 80% of these segments, totaling 14040 segments, are selected for further analysis. The training was carried out, and the remaining 20%, a total of 3510 signal segments, were used to test the trained model.

[0137] 3. Model Training and Validation Strategies

[0138] 3.1. Model Training and validation

[0139] The model training and validation results are determined by the coefficient of determination (R²) between the inverted acceleration and the simulated acceleration. Since there are 49 moving measurement locations, the R² values ​​from the acceleration inversion at these 49 locations were averaged during the test. After validation on 13,800 acceleration signal segments in data output mode one, the resulting R² value was 0.9975, almost close to 1, indicating that on the training dataset, the acceleration at any location on the tower can be inverted from the measured acceleration at the top of the tower, achieving the training objective.

[0140] After the model training is complete, the acceleration at the top of the tower and the height values ​​at measurement locations 1-49 from the remaining 3450 acceleration signal segments will be input. This generates the acceleration response at the moving measurement location. Figure 4 The comparison between the inversion results of the measured position of tower No. 6 and the simulation results shows that they are almost completely consistent, proving that... It can invert the acceleration at any position on the tower by measuring the acceleration at the top of the tower.

[0141] The above studies verified the model's performance under ideal simulation data using numerical models. However, in real-world engineering applications, sensor signals are inevitably contaminated by multiple factors such as environmental noise and electromagnetic interference. Therefore, whether the model has practical deployment value requires evaluating its performance stability under noisy input conditions. To address this, Gaussian white noise of varying intensities was artificially injected into the reference acceleration signal at the top of the tower to simulate a real-world scenario where the reference sensor itself is contaminated. The noise intensity was quantified by the signal-to-noise ratio (SNR); a lower SNR indicates more severe noise contamination. Five typical noise levels—25 dB, 20 dB, 15 dB, 10 dB, and 5 dB—were tested and compared with the baseline performance under noise-free conditions.

[0142] Table 1 shows the reconstruction accuracy of the model under different signal-to-noise ratios:

[0143] No noise 25dB 20dB 15dB 10dB 5dB <![CDATA[R 2 ]]> 0.996443 0.991021 0.989777 0.981861 0.962045 0.919902

[0144] A comprehensive analysis of the above results reveals that when the noise level is low (≥25dB), the model's performance is almost unaffected, with R² values ​​consistently above 0.99. Even under a relatively strong noise environment of 15dB, R² remains above 0.98, maintaining high inversion capability. Under strong noise conditions of 10dB, while the R² value decreases somewhat, it still remains above 0.96. This is because Gaussian noise is often a high-frequency component, while structural vibration acceleration is relatively lower frequency than noise. U-Net's multi-layer encoder-decoder structure can effectively filter out high-frequency random noise and reconstruct the signal during downsampling.

[0145] 3.2. Model Training and validation

[0146] Import 702,000 acceleration signal segments from data output mode two. The trained dataset yielded an R-value of 0.9967, indicating that the displacement at the corresponding position can be inverted from the measured acceleration on the training dataset, achieving the training objective. Subsequently, the remaining 175,500 acceleration signal segments were imported into the trained dataset. This generates the inverted displacement. For example... Figure 5 As shown, by The displacement time history at position 6, obtained through inversion, perfectly matches the simulated true value, and the predicted R² value reaches 0.988, proving its ability to accurately handle arbitrary acceleration inputs and invert displacement. Furthermore, a curve fitting method is used to fit the real-time displacements at 50 measurement positions to obtain the discrete attitude of the tower. For example... Figure 6 As shown, the deformation attitude curves of the wind power foundation structure at different times are consistent with the deformation form of the structure, confirming the rationality of the displacement inversion.

[0147] 3.3. Model Training and validation

[0148] Import 17,550 acceleration signal segments from data output mode 3. The trained R² value was 0.9943. Then, the remaining 3510 displacement signal segments were imported into the trained... The bending moment at the bottom of the tower is generated, and the circumferential stress is calculated at the same time. Figure 7 The results of the bending moment inversion at the tower base section were compared with the simulation values. The two results were highly consistent, with an average R² of 0.982.

[0149] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for monitoring the global posture and stress inversion of a wind power structure by mobile-fixed combination, characterized in that, The wind power structure includes a monopile and a tower supported on the monopile. The method includes an offline model training step and an online real-time inversion step. Offline model training steps include: S11. Acquiring offline data includes: selecting an arbitrary height position on the tower as a reference position, and fixing a fixed acceleration detection device at the reference position. The fixed acceleration detection device continuously collects acceleration data at the reference position, which is offline reference acceleration data; selecting several tower measurement positions along the tower height direction, and inspecting the acceleration detection device moving along the tower height direction. When moving to each tower measurement position in sequence, the device continuously collects acceleration data for a set duration. Several single-pile inversion positions are selected along the height direction of the single pile; A finite element model of the wind power structure is established, and environmental parameters are input into the finite element model. The finite element model outputs the acceleration simulation data, displacement simulation data, and bending moment simulation data of the tower, as well as the bending moment simulation data of a single pile. S12. Construct an acceleration conversion model between a fixed reference position and a measurement position. Take offline reference acceleration data and height data of each tower measurement position as input, and acceleration data collected at each tower measurement position as output, and train the acceleration conversion model. S13. Train the acceleration-displacement conversion model by taking the tower's acceleration simulation data as input and displacement simulation data as output. S14. Train the displacement-bending moment conversion model by selecting displacement simulation data at several locations on the tower as input, selecting several locations on the tower and / or single pile as target locations, and using the bending moment simulation data at the target locations as output. The online real-time inversion steps include: S21. Acquire acceleration data continuously collected by the fixed acceleration detection device at the reference position online, which serves as online reference acceleration data; S22. Using the acceleration conversion model, acceleration-displacement conversion model, and displacement-bending moment conversion model, calculate the displacement of the wind turbine and the bending moment at the target position based on the online reference acceleration data; S23. Calculate the stress based on the bending moment at the target location.

2. The method of claim 1, wherein, Step 12 also includes the step of segmenting the offline reference acceleration data. Based on the start and end times of the acceleration data collected at each tower measurement location, the offline reference acceleration data is segmented on the time axis to obtain several segments of offline acceleration data. The number of segments is consistent with the number of acceleration data segments at the tower measurement location. When training the acceleration conversion model, the offline acceleration data of each segment and the height data of each tower measurement position are used as inputs, and the acceleration data collected at each tower measurement position is used as outputs. The acceleration conversion model is trained using a spatiotemporal feature modulation network.

3. The method of claim 1, wherein, Step S11 includes: establishing a finite element model of the wind power structure based on the design parameters of the wind turbine; The simulation of the structural response of the wind turbine under different environmental parameters includes randomly selecting environmental parameters from the environmental load spectrum of the operating sea area, inputting the environmental parameters into the finite element model, and the sampling frequency of the output data of the finite element model being consistent with the sampling frequency of the fixed acceleration detection device and the inspection acceleration detection device. The environmental parameters include any combination of wind speed, wind direction, wave height, and wave direction.

4. The method of claim 1, wherein, The acceleration conversion model is a deep learning model that uses a spatiotemporal feature modulation U-Net architecture. It receives intermediate features from the neural network and height data from the measurement positions of each tower, maps the one-dimensional height value to a scale vector and an offset vector, and outputs the inverted acceleration data through element-wise affine transformation.

5. The method according to claim 1, characterized in that, The acceleration-displacement conversion model and displacement-bending moment conversion model adopt a deep generative model based on continuous normalized flow (CNF). Flow matching technology is used during training, and the objective function is: ； in, and These are the model's input and output, respectively. The parameters of the neural network, This refers to a stage in the flow process. Representation phase The output value of the neural network at that time. This indicates the calculation of the expected value.

6. The method according to claim 1, characterized in that, In step S23, the displacements of all virtual measuring points at the same moment are connected in space using a curve fitting method to draw the deformation attitude curve at that moment.

7. The method according to claim 6, characterized in that, Step S23 also includes using a preset function fitting method to fit the deformation posture curve to generate a continuous deformation posture curve covering the entire tower.

8. The method according to claim 1, characterized in that, Step S11 also includes a step of normalizing the acceleration data collected at the tower measurement location and the data output by the finite element model, converting the original physical scale signal into... Signals within an interval.

9. The method according to claim 1, characterized in that, Step S22 includes: S221, Acceleration reconstruction steps, including: Select several virtual measuring points along the height direction from the single pile and the tower; The online reference acceleration data is segmented on the time axis, with the segment length being the same as the acquisition duration of the acceleration data collected at the tower measurement location in the offline model training step, to obtain several segments of online acceleration data. The start time of each segment of online acceleration data is recorded. The online acceleration data of each segment and the height h of each virtual measuring point are input into the acceleration conversion model. The output value of the acceleration conversion model is multiplied by the first acceleration normalization factor to obtain the acceleration of each virtual measuring point, which is the inversion acceleration of each virtual measuring point within the corresponding time range. S222, Displacement Inversion and Attitude Synthesis Steps, including: The inversion acceleration of each virtual measuring point is input into the acceleration-displacement conversion model. The output value of the acceleration-displacement conversion model is multiplied by the second acceleration normalization factor to obtain the displacement of each virtual measuring point, which is the inversion displacement of each virtual measuring point within the corresponding time range. Connect the inversion displacements of all virtual measuring points at the same moment in space to plot the deformation attitude curve at that moment; S223, Dynamic stress field calculation steps: Select the inversion displacement of several virtual measuring points and input it into the displacement-bending moment conversion model. Multiply the output value of the displacement-bending moment conversion model with the displacement normalization factor to obtain the global inversion bending moment within the corresponding time range. Select the inversion bending moment of the target position from the global inversion bending moment. S224. The inversion acceleration, inversion displacement, deformation attitude curves and inversion bending moment of each time range are spliced ​​together on the time axis.

10. The method according to claim 1, characterized in that, The stress calculation method in step S23 includes: S231. Establish a rectangular coordinate system with the front-to-back direction of the wind turbine as the X-axis, the left-to-right direction as the Y-axis, the central axis of the tower as the Z-axis, and the center of the single pile as the origin. S232, Stress at target position g for: ； in, and Let be the bending moments at the target location g in the X and Y directions, respectively. The height of the target position g The above structural self-weight, For height The cross-sectional area of ​​the tower at that location, For height The radius of the tower section at that location, For height Moment of inertia of the tower section at the location, Let g be the angle between the line connecting the target position g and the central axis of the tower and the X-axis.

Images