Offshore wind power tower low frequency swing, settlement tilt early warning method and device

Abstract

The application provides a kind of offshore wind power tower cylinder low frequency swing, settlement tilt early warning method and device, it is related to wind power tower cylinder monitoring technical field.First, the tower cylinder health state benchmark period is defined, and the structure response and environmental load data are collected synchronously, and the abnormal working condition data is strictly cleaned out.Next, the feature vector representing the environmental load and the label vector of the structure response are constructed to form a health sample set.Then, a high-precision mapping model of the environmental load and the structure response is trained using machine learning algorithm.In online monitoring, the residual error between the predicted value and the measured value is calculated to effectively remove environmental interference and extract abnormal signals caused by changes in the structure itself.Finally, trend analysis and multi-level threshold judgment are performed on the residual signals to achieve early and hierarchical warning for tilt acceleration and cumulative deviation.The application can significantly improve the accuracy and early warning capability of offshore wind power tower safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of wind turbine tower monitoring technology, specifically to a method and device for early warning of low-frequency swaying, settlement and tilting of offshore wind turbine towers. Background Technology

[0002] Offshore wind power, as a crucial component of clean energy, is developing towards larger scale and deeper-sea applications. This brings with it the challenge of subjecting the large tower structures supporting the turbines to more complex and severe long-term, cyclical marine environmental loads. These loads can easily induce low-frequency swaying of the tower structure, foundation scour, and even uneven settlement, potentially leading to overall tower tilting. This tilting and settlement is a slow, cumulative process, difficult to detect initially, but once it exceeds a critical point, it will seriously threaten the safe operation of the wind turbine and even cause structural instability. Therefore, real-time and accurate monitoring and early warning of the health status of offshore wind turbine towers, and timely detection of early abnormal changes, are crucial for ensuring safety and economic benefits throughout the entire life cycle.

[0003] Traditional wind turbine structure monitoring methods primarily rely on setting fixed thresholds for structural responses to trigger alarms. However, this approach has significant limitations: the structural response amplitude of offshore wind turbine towers is highly dependent on real-time environmental loads. In severe sea conditions, even if the structure is in a healthy state, its response, such as sway amplitude and instantaneous tilt angle, can increase significantly, easily triggering false alarms. Conversely, in mild weather, even if the structure has experienced early damage or slow tilting, its response may still fall within the threshold range, resulting in missed alarms. This environmental noise greatly masks the true response changes caused by structural degradation, rendering traditional threshold alarm methods ineffective. Existing technologies struggle to separate the influence of environmental excitations from the overall structural response, thus failing to effectively identify early signs of structural performance degradation. A precise method is lacking that can sense environmental loads and, based on this, intelligently assess and provide early warnings of the structural health status.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and device for early warning of low-frequency swaying, settlement and tilting of offshore wind turbine towers, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for early warning of low-frequency oscillation and settlement tilting of offshore wind turbine towers, comprising the following steps:

[0008] Step 1: Determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. Collect relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and remove relevant data from the initial operation period when the tower is in an abnormal power generation state.

[0009] Step 2: Extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and pair environmental feature vectors and response label vectors at the same timestamp to integrate them into a health status sample set;

[0010] Step 3: Perform correlation analysis on the environmental feature vectors and response label vectors in the healthy state sample set to determine the mapping relationship between environmental loads and structural responses during the healthy state period of the wind turbine tower;

[0011] Step 4: Set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then obtain the measured value of the structural response parameter of the short-term evaluation window, compare the measured value of the structural response with the predicted value of the structural response, and obtain the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load;

[0012] Step 5: Based on the changing trend of the residual signal sequence in different time windows, set multi-level warning thresholds, and trigger the corresponding level of warning signal according to the trend of the residual signal sequence and the multi-level warning thresholds.

[0013] Furthermore, the collection of the structural response data and environmental load data specifically includes:

[0014] The no-load commissioning period refers to the stage from the completion of installation of the wind turbine generator set until it is connected to the grid for power generation, during which the wind turbine is in a self-rotating state without power generation load; the initial operation period refers to the continuous and stable operation time of the wind turbine after it is connected to the grid for power generation, and this continuous and stable operation time is at least three months.

[0015] The structural response data includes the tower's triaxial acceleration, biaxial tilt angle, three-dimensional position information of the tower top, and tidal height. The triaxial acceleration is measured by an inertial measurement unit of a microelectromechanical system (MEMS) inside the top compartment of the tower. Similarly, a biaxial tilt sensor is installed in the tower to obtain the tower's tilt angle in the downwind and crosswind directions. A multi-frequency GNSS receiver is installed at the top of the tower to obtain the three-dimensional position information of the tower top center. A tidal meter is installed on the wind turbine foundation platform to obtain the tidal height.

[0016] The aforementioned inertial measurement unit and the tower have a fixed structural coordinate system, namely, the X-axis is defined as the windward axis of the tower, the Y-axis is defined as the crosswind axis of the tower, and the Z-axis is vertically upward;

[0017] The environmental load data includes instantaneous wind speed and direction at the top of the tower, the orientation of the wind turbine nacelle and surface current velocity, as well as the effective wave height and zero-cycle at the location of the wind turbine tower. Wind speed and direction are measured using an ultrasonic anemometer mounted on the tower top anemometer support. Wave radar is installed on the wind turbine foundation platform to continuously measure and record wavefront elevation time-series data at the highest frequency. Fourier transform is used to perform spectral analysis on the wavefront elevation time-series data to calculate the wave energy spectral density function. Spectral moments are calculated based on the energy spectral density function, and the effective wave height is calculated based on the zeroth-order spectral moment. The zero-cycle is also calculated based on the spectral moment. The orientation of the wind turbine nacelle is obtained through the wind turbine's inherent main control system. Relevant sensors are installed on the legs of the wind turbine foundation platform to measure the current velocity profile across the entire water depth, with the data from the uppermost measurement point used as the surface current velocity.

[0018] The specific method of synchronous acquisition is as follows: During the no-load commissioning period and the initial operation period, a short-term evaluation window is preset, and the arithmetic mean of the instantaneous wind speed at each sampling time within the short-term evaluation window is calculated. This arithmetic mean is used as the average wind speed of the short-term evaluation window. When the average wind speed exceeds the preset start threshold, a preset period of data recording window is automatically started. Within this window, all sensors sample at the highest frequency, and the collected data is stored after being stamped with a uniform timestamp. The highest frequency sampling is not lower than 2Hz.

[0019] Furthermore, the logic for removing data related to abnormal power generation during the initial operation period is as follows: read the unit status code from the wind turbine monitoring and data acquisition system, and remove data periods where the status code indicates shutdown, fault, emergency shutdown, manual shutdown, start-up and shutdown processes, and operating power below 5% of rated power; at the same time, for environmental load data, remove data periods where the instantaneous wind speed exceeds the cut-out wind speed of the turbine model and the effective wave height exceeds the safe operating wave height threshold; after removing the relevant data in the above periods, the relevant data retained will be the relevant data under normal power generation conditions.

[0020] Furthermore, the logic for extracting environmental feature parameters and response feature parameters to construct environmental feature vectors and response label vectors is as follows:

[0021] The environmental feature vector is defined as ,in, For environmental feature vectors, The average wind speed, The sine value of the relative wind direction angle. The cosine of the relative wind direction angle. For the effective wave height, For crossing the zero cycle, The surface current velocity is used; the environmental characteristic parameters are obtained as follows: for each timestamp, a time step back is taken to determine the time step. The short-term evaluation window is defined; the arithmetic mean of the instantaneous wind speeds at all sampling times within this time window is calculated as the average wind speed; the absolute difference between the wind direction at this timestamp and the orientation of the wind turbine nacelle is calculated as the relative wind direction angle. Since the wind direction is a periodic angular variable, a sine-cosine transformation is used to convert it into two linear features. The transformation formula is as follows: , ,in, The sine value of the relative wind direction angle. The cosine of the relative wind direction angle. The relative wind direction angle is used; based on the wavefront elevation time series data within the short-term assessment window, the effective wave height and zero-crossing period are calculated by spectral analysis; the arithmetic mean of the surface current velocity at all sampling times within the short-term assessment window is calculated as the surface current velocity at that timestamp.

[0022] The response tag vector is defined as ,in, Represents the response feature vector. The standard deviation of the X-axis acceleration. The standard deviation of the acceleration along the Y-axis. The tilt angle along the X-axis. The tilt angle along the Y-axis. The cumulative change in elevation is used to represent the cumulative change in elevation. The methods for obtaining various response characteristic parameters are as follows: High-pass filtering is applied to the collected triaxial acceleration data to remove gravitational acceleration components and extremely low-frequency drift. The standard deviations of the X-axis and Y-axis acceleration data within the short-term evaluation window are calculated. Extremely low frequencies refer to frequencies below 0.1 Hz. The collected downwind and crosswind tilt angles are directly used as the X-axis and Y-axis tilt angles. Based on data under normal power generation conditions, three-dimensional location information collected during the first week after the start of the no-load commissioning period under no-severe weather conditions is selected as reference location information. This three-dimensional location information includes the longitude, latitude, and elevation of the tower top center. Its elevation data is used as the original elevation value. The original elevation value and the tidal height at each time stamp represent the environmental disturbance height. The difference between the elevation at each time stamp and the environmental disturbance height is calculated as the cumulative change in elevation to characterize the foundation settlement phenomenon. Response characteristic parameters are extracted from the structural response data. These response characteristic parameters include the standard deviations of the X-axis and Y-axis accelerations, the X-axis and Y-axis tilt angles, and the cumulative change in elevation.

[0023] After constructing each vector, each data pair with timestamp alignment is defined as a sample point. The data pair includes an environmental feature vector and a response label vector. All sample points are arranged in chronological order to construct a health status sample set.

[0024] Furthermore, by using a trained machine learning model to represent the mapping relationship, the training of the machine learning model specifically includes:

[0025] Before training, the environmental feature vector and response label vector are standardized to convert them into data with a mean of 0 and a standard deviation of 1.

[0026] The health status sample set is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The standardized environmental feature vectors in the health status sample set are used as inputs, and the standardized response label vectors corresponding to the timestamps are used as outputs to construct a standard response prediction model.

[0027] During training, predictions are made on the validation set every 10 iterations, and the loss function on the validation set is calculated. A patience value is preset for each iteration. If the loss function value on the validation set no longer decreases within consecutive patience values, training automatically terminates. The loss function is the mean squared error, which aims to minimize the sum of squared prediction errors across all output dimensions. The model is trained on the training set, and a Bayesian optimization algorithm is used on the validation set to automatically optimize key hyperparameters. The key hyperparameters include at least the learning rate, tree depth, and regularization parameter. The learning rate is set to range from [0.01, 0.3], the tree depth to range from [4, 10], and the weight coefficient of the regularization term in the objective function to range from [0.5, 5].

[0028] After training, the model performance is evaluated on the test set. The evaluation metrics include mean absolute error and coefficient of determination. When the mean absolute error is less than 0.03 and the coefficient of determination is greater than 0.85, the model is considered to have excellent performance; otherwise, the model will be retrained.

[0029] Furthermore, obtaining the residual signal sequence specifically includes:

[0030] The short-term evaluation window set based on the current moment is represented as follows: ,in, Indicates the current moment. Indicates the length of the short-term assessment window;

[0031] There are N sampling times in the short-term evaluation window. Environmental load data are collected at each sampling time within the short-term evaluation window. Environmental feature vectors are extracted at each sampling time within the short-term evaluation window, and then standardized and input into the trained machine learning model to obtain the response label vector corresponding to each sampling time, which is used as the structural response prediction value.

[0032] Next, obtain the structural response data for each sampling time within the short-term evaluation window, calculate the response label vector for the corresponding sampling time, and use it as the measured value of the structural response. Calculate the difference between the measured value of the structural response at each sampling time and the predicted value of the structural response at the corresponding time to obtain the residual signal at that time. Arrange the residual signals of all sampling times in timestamp order to form a residual signal sequence, represented as follows: ,in, This represents the residual signal at the first sampling time point within the short-term evaluation window. This represents the residual signal at the second sampling time point within the short-term evaluation window. This represents the residual signal at the i-th sampling time in the short-term evaluation window. This represents the residual signal at the Nth sampling time within the short-term evaluation window, where i is the short-term evaluation window. The sequential index of the internal sampling time;

[0033] The residual signal contains multiple components, specifically including the standard deviation component of the residual X-axis acceleration, the standard deviation component of the residual Y-axis acceleration, the residual X-axis tilt component, the residual Y-axis tilt component, and the cumulative change component of the residual elevation.

[0034] Furthermore, signal enhancement based on residual signal sequences specifically includes:

[0035] The residual X-axis tilt component and the residual Y-axis tilt component are synthesized into a single residual tilt component representing the overall tilt degree. The synthesis formula is as follows:

[0036] ;

[0037] in, The residual tilt component is the i-th residual signal. The residual X-axis tilt component in the i-th residual signal. Let be the residual Y-axis tilt component in the i-th residual signal.

[0038] Furthermore, a short-term evaluation window and a long-term evaluation window are set. Based on the short-term evaluation window corresponding to the current moment, the residual tilt components of all sampling moments are calculated, and the linear regression slope of the residual tilt components of all sampling moments within the short-term evaluation window corresponding to the current moment is calculated as the linear regression slope of the current moment, which is used to characterize the tilt change rate at the current moment.

[0039] Time window Defined as a long-term evaluation window, in which Calculate the average value of the residual tilt component within the long-term evaluation window at the current moment, and calculate the difference between this average value and the average value of the residual tilt component during the healthy state period of the wind turbine. Use this difference as the medium-to-long-term cumulative deviation at the current moment to characterize the overall offset accumulated by the wind turbine. Here, M represents the length of the long-term evaluation window.

[0040] Furthermore, the specific triggering of the corresponding warning signals includes:

[0041] Set multi-level warning thresholds, including a first-level threshold, a second-level threshold, and a third-level threshold; the first-level threshold is specifically set as follows: The secondary threshold is specifically set as follows: The specific threshold settings for the three levels are as follows: ,in, These are the first-level threshold, the second-level threshold, and the third-level threshold. These are the constant coefficients in the corresponding level thresholds, and , , The mean and standard deviation of the medium- to long-term cumulative bias at all collection times in the healthy state sample set. The mean and standard deviation of the linear regression slope at all sampling times in the health status sample set represent the sum of the slopes.

[0042] When the absolute value of the linear regression slope is greater than the first-level threshold, a first-level warning is triggered, and a signal is issued indicating a risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the second-level threshold, a second-level warning is triggered, and a signal is issued confirming the risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the third-level threshold, a third-level warning is issued, and a signal is issued to take immediate emergency action.

[0043] The present invention also provides a low-frequency oscillation and settlement tilting early warning device for offshore wind turbine towers. This device is used to implement the aforementioned method for early warning of low-frequency oscillation and settlement tilting of offshore wind turbine towers, comprising:

[0044] The health baseline data acquisition module is used to determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. It synchronously collects relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and removes relevant data from the initial operation period when the tower is in an abnormal power generation state.

[0045] The health status sample set construction module is used to extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and to pair environmental feature vectors and response label vectors under the same timestamp to integrate them into a health status sample set.

[0046] The Environment and Response Relationship Analysis Module is used to perform correlation analysis on the environmental feature vectors and response label vectors in the health state sample set to determine the mapping relationship between environmental loads and structural responses during the health state period of wind turbine towers.

[0047] The real-time status monitoring module is used to set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then, it obtains the measured value of the structural response parameter of the short-term evaluation window, compares the measured value of the structural response with the predicted value of the structural response, and obtains the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load;

[0048] The trend analysis and early warning module is used to set multi-level early warning thresholds based on the changing trend of the residual signal sequence in different time windows, and to trigger the corresponding level of early warning signal based on the trend of the residual signal sequence and the multi-level early warning thresholds.

[0049] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0050] This solution effectively addresses the high false alarm and false negative rates of traditional fixed threshold methods by establishing a machine learning model between environmental loads and structural response. Its core advantage lies in its ability to accurately isolate the strong influence of environmental excitations such as wind, waves, and currents on the tower response, thereby extracting residual signals caused solely by changes in the structure's own state, such as foundation settlement and stiffness degradation. This allows the method to keenly detect early, slow-developing anomalies masked by environmental noise, enabling early and accurate warnings of low-frequency tower swaying and settlement tilting. This provides a critical time window for operation and maintenance decisions, significantly improving the reliability and intelligence level of offshore wind turbine tower safety monitoring. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0052] Figure 2 This is a comparison chart of the predicted and measured values ​​of the residual X-axis tilt component of this invention;

[0053] Figure 3 This is a comparison chart of the predicted and measured values ​​of the residual Y-axis tilt component of this invention;

[0054] Figure 4 This is a schematic diagram of the device structure of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0056] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0057] Example:

[0058] Please see Figures 1 to 3 The present invention provides a technical solution:

[0059] A method for early warning of low-frequency oscillation and settlement tilting of offshore wind turbine towers, comprising the following steps:

[0060] Step 1: Determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. Simultaneously collect relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and remove relevant data from the initial operation period when the tower is in an abnormal power generation state. This step is the basis for building the benchmark of the entire early warning system. Its purpose is to obtain a clean, reliable dataset that can represent the behavioral characteristics of the wind turbine tower under the condition of no structural damage.

[0061] In this embodiment, the collection of the structural response data and environmental load data specifically includes:

[0062] The no-load commissioning period refers to the stage from the completion of wind turbine installation until grid connection for power generation. During this stage, the wind turbine is in a self-rotating state without any power generation load. In this stage, the wind turbine (including blades, nacelle, tower, etc.) has been fully installed and is ready for operation, but has not yet been connected to the grid to transmit electricity. During this stage, the turbine blades can rotate freely under wind conditions, but the generator does not generate load torque. The technical significance of this stage is that the structure mainly bears environmental loads, avoiding additional loads from power generation operation (such as generator torque, grid load changes, etc.). It best reflects the initial health state of the tower foundation and the structure itself, making it an ideal data source for establishing a benchmark model.

[0063] The initial operating period refers to the period of continuous and stable operation of the wind turbine after it is connected to the grid, which must be at least three months. This stage refers to the initial phase of continuous and stable power generation after the wind turbine is successfully connected to the grid. This invention specifies that this stage should be determined within three to six months after the wind turbine generator set begins operation. The technical significance of this stage is that after three months of operation, the initial installation stress of the structure has been redistributed and tends to stabilize, while ensuring that the structure has not suffered significant damage due to long-term fatigue or external scouring. The data from this stage reflects the healthy response of the tower under normal operating loads and is an important supplement and improvement to the benchmark for the no-load commissioning period.

[0064] The structural response data includes the tower's triaxial acceleration, biaxial tilt angle, three-dimensional position information of the tower top, and tidal height. A microelectromechanical system (MES) inertial measurement unit (IMU) is fixedly installed inside the nacelle at the top of the tower. The IMU's coordinate system is fixed to the tower's structural coordinate system, defined as follows: the X-axis is defined as the tower's downwind axis, the Y-axis as the tower's crosswind axis, and the Z-axis is vertically upward. The triaxial acceleration measured by the IMU is used to capture the tower's high-frequency vibration response under high wind loads and other excitations, and to calculate the standard deviation characterizing the vibration intensity. Similarly, a biaxial tilt sensor is installed in the tower, typically in the upper middle part, preferably at the bottom of the nacelle or on an internal platform, to obtain the tower's tilt angle in the downwind and crosswind directions; its technical function is to directly measure the static tilt angle of the tower relative to a gravity reference, providing a direct indicator for monitoring the overall tilt. A multi-frequency GNSS receiver, such as one supporting GPS, GLONASS, or BeiDou systems, is installed at the highest point of the tower. This receiver is typically located on top of the nacelle to acquire the three-dimensional position information of the tower's center. Its technical function is to continuously measure the three-dimensional coordinates (longitude, latitude, and elevation) of the tower's center with centimeter-level or even millimeter-level accuracy using technologies such as real-time dynamic differential carrier phase analysis. Elevation data is the core raw data for calculating foundation settlement. A tidal meter is installed on the wind turbine foundation platform, such as a jacket structure or the transition section of a monopile, to obtain tidal height. The tidal meter can be pressure-type or radar-type. Its technical function is to accurately measure the instantaneous sea level height at the tower's location. This data is crucial for eliminating interference from sea level changes from the absolute elevation measured by GNSS, thereby accurately obtaining the cumulative elevation change of the tower itself caused by foundation settlement.

[0065] The environmental load data includes instantaneous wind speed and direction at the top of the tower, the orientation of the wind turbine nacelle and surface current velocity, as well as the effective wave height and zero-crossing period at the location of the wind turbine tower. Wind speed and direction are measured using an ultrasonic anemometer mounted on a support bracket at the top of the tower; this method has no moving parts, offers high accuracy, and can simultaneously measure three-dimensional wind speed and direction. It samples at the highest frequency, providing instantaneous wind speed and direction data. A wave radar or acoustic Doppler wave height meter is installed on the wind turbine foundation platform. This sensor samples at the highest frequency, continuously measuring and recording the wavefront elevation time series data near the platform. Subsequent data processing involves using Fourier transform to perform spectral analysis on the wavefront elevation time series data, calculating the wave energy spectral density function, calculating the spectral moments based on the energy spectral density function, calculating the effective wave height based on the zeroth-order spectral moment, and similarly calculating the zero-crossing period based on the zeroth and second-order spectral moments; these are all key parameters characterizing wave energy and period. The orientation of the wind turbine nacelle is read through the communication interface of the wind turbine's inherent main control system; the current velocity profile across the entire water depth is measured by installing an acoustic Doppler wave height meter (ADCP) on the legs of the wind power foundation platform, and the data from the uppermost measurement point of the ADCP is taken as the surface current velocity; the uppermost layer is usually located 1-2 meters below the sea surface.

[0066] The specific method of synchronous data acquisition is as follows: During the no-load commissioning period and initial operation period, a short-term evaluation window is preset, preferably 10 minutes. The arithmetic mean of the instantaneous wind speed at each sampling moment within the short-term evaluation window is calculated and used as the average wind speed for that short-term evaluation window. Referring to relevant standards, 10 minutes is recommended as the standard duration for average wind speed. Wind speed is transient, but structural response (especially static tilt and settlement) is a feedback to continuous loads. A time too short (e.g., 1 minute) cannot smooth out turbulence, and the average wind speed is not representative enough; a time too long (e.g., 1 hour) will over-smooth out and may miss some important load events. 10 minutes is generally considered the best compromise that balances instantaneous fluctuations and steady-state trends. When the average wind speed exceeds a preset start-up threshold, the system determines that the current condition is a high load condition and automatically starts a preset period (e.g., 30 minutes) of data recording. The 30-minute recording window is to fully capture a high load event and its resulting complete structural response process. A storm or a large wave group typically lasts from several minutes to twenty minutes. A 30-minute window ensures that, after triggering, not only the peak of the event is recorded, but also the entire dynamic process of the structure from excitation to vibration and then gradual recovery. This is crucial for subsequent analysis of the complete modal response. The trigger threshold can be set with reference to 50% of the rated wind speed. Rated wind speed refers to the minimum wind speed at which the wind turbine reaches its rated power output. Above this wind speed, the turbine stabilizes its power at the rated value through pitch control and other methods, and no longer increases it. This parameter can be determined from the turbine model's technical specifications. 50% of the rated wind speed is usually a critical point, marking the beginning of the load's significant impact on the structure and the starting point where the response becomes apparent and analytically valuable. Collecting data above this threshold allows focusing on the high-load conditions most valuable for structural health assessment, greatly improving dataset quality and modeling efficiency. Within this window, all sensors sample at the highest frequency, and the collected data is stored after being timestamped uniformly; the highest frequency sampling is no less than 2Hz. Using the highest frequency sampling satisfies the Nyquist sampling theorem, and the tower's first natural frequency (the dominant frequency of low-frequency oscillation) is typically between 0.2Hz and 0.5Hz. To accurately reconstruct this signal, the sampling frequency must be at least twice its normal value (i.e., 0.4-1Hz). Using a sampling frequency of at least 2Hz provides ample margin for analyzing low-frequency oscillations below 1Hz, ensuring data accuracy. Furthermore, in addition to the dominant frequency, the structure also exhibits higher-frequency vibration modes. A high sampling frequency helps to provide a more comprehensive understanding of the structure's dynamic characteristics.

[0067] In this embodiment, the logic for removing data related to abnormal power generation during the initial operation period is as follows: The unit status code is read from the wind turbine monitoring and data acquisition system, and data periods indicating shutdown, fault, emergency shutdown, manual shutdown, start-up / shutdown processes, and operating power below 5% of rated power are removed (in which case it is considered extremely low load operation, and the state is unstable). Simultaneously, for environmental load data, data periods where the instantaneous wind speed exceeds the cut-out wind speed for this turbine model and the effective wave height exceeds the safe operating wave height threshold are removed. After removing the relevant data for the above periods, the remaining relevant data is considered as relevant data under normal power generation conditions. Under the above-mentioned conditions where status codes are removed, the wind turbine's operating state is not stable. For example, there are huge dynamic loads during start-up / shutdown processes (such as gearbox impact, torque mutation); the structural response is abnormal under fault conditions. These data contradict the definition of health, and if included in training, the model will learn incorrect relationships. The cut-out wind speed is the wind speed at which the wind turbine actively shuts down to protect itself. At this wind speed, the wind turbine has stopped, its state returns to non-power generation mode, and the data is invalid. Data from extreme sea states (maximum wave height) may contain nonlinear or even destructive loads. Health baseline models only need to be established within the environmental limits of normal operation. Extreme data represents anomalies, not healthy states, and should be excluded. Additionally, the rated power here refers to the maximum continuous electrical output power that the wind turbine generator is designed to achieve, which can be directly obtained from the turbine's technical specifications.

[0068] Step 2: Extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and pair environmental feature vectors and response label vectors at the same timestamp to integrate them into a health status sample set;

[0069] In this embodiment, the logic for extracting environmental feature parameters and response feature parameters to construct environmental feature vectors and response label vectors is as follows:

[0070] The environmental feature vector is defined as ,in, For environmental feature vectors, The average wind speed, The sine value of the relative wind direction angle. The cosine of the relative wind direction angle. For the effective wave height, For crossing the zero cycle, This represents the velocity of the surface ocean current.

[0071] The environmental characteristic parameters are obtained as follows: For each timestamp, a backward sequence of length is used. A short-term assessment window is typically set at 10 minutes to align with common wind speed statistics. The arithmetic mean of the instantaneous wind speeds at all sampling times within this short-term assessment window is calculated as the average wind speed; this average wind speed is then squared to obtain the mean wind speed. The direct introduction of this feature term is based on the fundamental principles of wind engineering, where wind load is proportional to the square of wind speed. This aims to allow machine learning models to learn the nonlinear effects of wind load more directly, without requiring the model to fit a square relationship from linear features, thus improving learning efficiency and accuracy. While the square of the average wind speed is crucial here, some complex effects in low-wind-speed regions (such as structural damping and nonlinear aerodynamic effects) may make the relationship between the response and V more pronounced. Providing both the average wind speed and its square provides the model with a set of fundamental functions, allowing the model to automatically weigh and combine them to most accurately fit the complex response relationship across the entire range from low to high wind speeds. The absolute difference between the wind direction at that timestamp and the orientation of the wind turbine nacelle is calculated as the relative wind direction angle. Since the wind direction is a periodic variable from 0° to 360°, directly inputting it as a linear feature into the model would lead to discontinuities between 0° and 360°, severely degrading model performance. To address this issue, this invention uses a sine-cosine transform to convert it into two linear features. The transformation formula is: , ,in, The sine value of the relative wind direction angle. The cosine of the relative wind direction angle. The relative wind direction angle is used; this transformation perfectly preserves all information about the wind direction while eliminating the discontinuities caused by periodic boundaries, making it a standard and effective method for handling angular variables. Introducing the relative wind direction angle into the environmental feature vector considers the directionality of the load; the structural response is highly dependent on the direction of the wind relative to the nacelle's incoming flow. When the wind impacts head-on (relative wind direction angle approximately 0), the response in the downwind direction (X-axis) is the largest; when the wind impacts from the side (relative wind direction angle approximately 90° or 270°), the response in the crosswind direction (Y-axis) may be larger, potentially triggering complex phenomena such as vortex-induced vibration. Based on wavefront elevation time-series data within a short-term assessment window, commonly used spectral analysis methods in this field (such as calculating the energy spectral density function through fast Fourier transform) are employed. The spectral moment is calculated based on the energy spectral density function to obtain the effective wave height and zero-cycle period. Waves exert enormous pressure and impact forces on offshore structural foundations, which are key factors causing structural dynamic responses, especially low-frequency oscillations and foundation moments. The effective wave height represents the energy level of the wave and is a statistical value of the wave height. The significant wave height is directly proportional to the magnitude of the wave load. A larger significant wave height results in a greater impact force on the foundation, thus exacerbating the overall vibration and swaying of the tower. The zero-cycle represents the main frequency range where wave energy is concentrated, and its importance lies in its resonance effect with the structure's natural frequency. If the zero-cycle is close to the first or second natural period of the tower, even a small significant wave height can trigger a severe dynamic amplification effect, leading to a significant increase in response, where response refers to acceleration in the X and Y axes. Therefore, significant wave height and the zero-cycle are introduced into the environmental feature vector. The arithmetic mean of the surface current velocity at all sampling times within the short-term evaluation window is calculated as the surface current velocity at that timestamp. Ocean currents exert stable drag forces and oscillating vortex-induced forces on wind turbine foundations (such as monopiles and jackets). Stable ocean currents generate additional average lateral loads, which manifest as small shifts in the average tilt angle (acceleration in the X and Y axes). When ocean currents at specific velocities flow past a foundation structure, they may generate periodically detached vortices, inducing high-frequency vibrations in the transverse direction. This effect primarily contributes to the standard deviation of acceleration along the Y-axis. Ocean currents alter wave characteristics (such as wavelength and wave height), indirectly affecting wave loads; therefore, surface current velocity is introduced here. In summary, the mean wind speed and the square of the mean wind speed accurately capture the nonlinear nature of wind loads in the construction of this environmental feature vector, the relative wind direction angle clarifies the load direction, and the significant wave height and zero-cycle crossover jointly describe the wave intensity characteristics, which are crucial for evaluating the dynamic response. Surface current velocity supplements the influence of ocean current loads, particularly its potential contribution to flow-induced vibrations near the foundation.

[0072] The response tag vector is defined as ,in, Represents the response feature vector. The standard deviation of the X-axis acceleration. The standard deviation of the acceleration along the Y-axis. The tilt angle along the X-axis. The tilt angle along the Y-axis. This represents the cumulative change in elevation. The various response characteristic parameters are obtained as follows: The raw triaxial acceleration data obtained from the microelectromechanical system inertial measurement unit includes gravitational acceleration components and extremely low-frequency drift noise. To extract the structural vibration signal truly excited by environmental loads, high-pass filtering is required. This invention uses a high-pass digital filter (such as a Butterworth filter) with a cutoff frequency of 0.1Hz to filter the raw acceleration data to remove gravitational acceleration and extremely low-frequency noise, where extremely low frequency refers to frequencies below 0.1Hz. The standard deviation of the filtered X-axis acceleration data within the short-term evaluation window corresponding to each time stamp is calculated; similarly, the standard deviation of the Y-axis acceleration data is calculated. This standard deviation effectively quantifies the vibration intensity of the tower top in the downwind and crosswind directions. The downwind and crosswind tilt angles at each time stamp are directly read from the dual-axis tilt sensor installed on the tower and used directly as the X-axis and Y-axis tilt angles. Based on data from normal power generation, three-dimensional location information collected during the first week of the no-load commissioning period under conditions without severe weather was selected as reference location information. The three-dimensional location information includes the longitude, latitude, and elevation of the tower top center. The elevation data was used as the original elevation value. The original elevation value and the tidal height at each time stamp represent the environmental interference height. The difference between the elevation at each time stamp and the environmental interference height was calculated as the cumulative elevation change to characterize foundation settlement. Response characteristic parameters were extracted from the structural response data. These parameters include the standard deviation of X-axis and Y-axis acceleration, the X-axis and Y-axis tilt angles, and the cumulative elevation change. Tidal fluctuations directly cause changes in mean sea level, while GNSS receivers measure absolute elevation relative to an ellipsoid (WGS-84 coordinate system). Therefore, even if the tower foundation itself does not experience any settlement or uplift, sea level fluctuations will directly cause periodic changes in the GNSS-measured tower top elevation. Without tidal compensation, the elevation data will exhibit significant semi-diurnal or diurnal tidal fluctuations consistent with the tidal cycle. The amplitude of this environmental interference signal (up to several meters) is far greater than the slow trend signal caused by foundation settlement that we are concerned with (which may only be at the centimeter or even millimeter level per year). The weak settlement signal will be completely drowned out by the huge tidal noise and cannot be effectively identified.

[0073] The absence of severe weather requires the following quantifiable conditions to be met: First, wind speed conditions: the average wind speed during this period should consistently be lower than the cut-in wind speed of the wind turbine. For example, this can be defined as a 10-minute moving average wind speed consistently below 3 m / s. This is because when the wind speed is below the cut-in wind speed, the wind turbine is in a shutdown state, the blade pitch angle is fixed, the nacelle's yaw activity is infrequent or ceases, the wind load on the tower is minimal and most stable, with almost no dynamic swaying, and the GNSS antenna is in its most static state. Second, wave height conditions: the effective wave height during this period should be lower than a preset calm sea state threshold. For example, this can be defined as an effective wave height consistently below 0.5 meters. Low wave heights indicate a calm sea surface, minimizing overall tower swaying caused by wave impact on the foundation, slight foundation swaying, and the multipath effect of sea surface reflection on GNSS signals. Third, operational status conditions: querying the wind turbine monitoring and data acquisition system, the turbine's status code should consistently indicate standby or normal shutdown during this period. This ensures that the wind turbine is not performing any operations that could cause vibration, such as generator grid connection, pitch control, or active yaw. Finally, there is the duration requirement; the time period should be long enough, for example, at least 6 consecutive hours. A longer stable time period allows for extensive sampling of GNSS elevation data and calculation of averages. Random noise can be effectively offset by averaging filtering, resulting in a very accurate reference value.

[0074] The reason for collecting data under the aforementioned non-adverse weather conditions is that GNSS measurements themselves contain centimeter-level noise. Under adverse weather conditions, the swaying of the tower, the slight shaking of the foundation, and signal errors significantly amplify this noise, making single-point measurements or short-time averages completely unrepresentative of the structure's true static elevation.

[0075] After constructing each vector, each data pair with timestamp alignment is defined as a sample point. The data pair includes an environmental feature vector and a response label vector. All sample points are arranged in chronological order to construct a health status sample set.

[0076] Step 3: Perform correlation analysis on the environmental feature vectors and response label vectors in the healthy state sample set to determine the mapping relationship between environmental loads and structural responses during the healthy state period of the wind turbine tower;

[0077] In this embodiment, the mapping relationship is represented by a trained machine learning model. The training of the machine learning model specifically includes:

[0078] Before training, the environmental feature vector and response label vector are standardized, also known as Z-score normalization, to convert them into data with a mean of 0 and a standard deviation of 1. This operation aims to eliminate the adverse effects of differences in feature units and numerical ranges on model training, accelerate model convergence, and improve the final performance of the model.

[0079] To conduct effective model training, tuning, and performance evaluation, the complete health status sample set needs to be randomly shuffled and divided into three mutually exclusive subsets: a training set, a validation set, and a test set, with a ratio of 6:2:2. The training set is used to train the machine learning model and adjust its weight parameters. The validation set is used to evaluate model performance during training, thereby optimizing hyperparameters and performing early detection to prevent overfitting. The test set is used to finally evaluate the model's generalization ability after training is complete.

[0080] This invention preferably utilizes a gradient boosting decision tree model family, such as XGBoost, LightGBM, or CatBoost. These models, by integrating multiple weak learners (decision trees), offer advantages such as high prediction accuracy, effective handling of nonlinear relationships, and insensitivity to missing features, making them highly suitable for modeling the complex multidimensional input-output mapping relationships in this invention. The standardized environmental feature vectors from the healthy state sample set are used as input, and the corresponding timestamp-standardized response label vectors are used as output to construct a standard response prediction model. Through training, the model can predict the expected structural response of a healthy tower based on the input environmental conditions.

[0081] During training, predictions are made on the validation set every 10 iterations, and the loss function on the validation set is calculated. A patience value is preset for each iteration. If the loss function value on the validation set no longer decreases within consecutive patience values, training automatically terminates. This patience value can be set to 50 iterations. The training process of complex models such as gradient boosting trees may not be monotonically decreasing. The loss function may experience brief plateaus or even small fluctuations during the global downward trend. If the patience value is set too small (e.g., 5 or 10), the model may stop prematurely due to temporary performance fluctuations before finding a better solution, resulting in insufficient training and the model's capabilities not being fully utilized. Conversely, if it is set too large, it can easily lead to overfitting. 50 iterations is an empirical value that has been widely validated in practice and can well balance sufficient exploration and timely stopping. When training automatically terminates, it rolls back to the model state with the lowest loss on the validation set. The loss function chosen is mean squared error, which aims to minimize the sum of squared prediction errors across all output dimensions. The model is trained on the training set, and a Bayesian optimization algorithm is used on the validation set to automatically optimize key hyperparameters. This algorithm is more efficient than grid search or random search. The key hyperparameters include at least the learning rate, tree depth, and regularization parameter. The learning rate is set to a range of [0.01, 0.3]. Searching within this range controls the contribution weight of each tree; a smaller value results in a more robust model but slower training. If the learning rate is below 0.01, the model converges very slowly, requiring a large number of iterations to achieve good performance, significantly increasing training time and computational cost, making it inefficient in practical engineering applications. If the learning rate is too large (e.g., greater than 0.3), the step size of each iteration is too large, potentially causing the loss function to oscillate around the optimal value or even diverge, failing to converge stably. 0.3 is a sufficiently large value, allowing the model to learn quickly without making the training process unstable. This range ensures that the optimization algorithm can find robust solutions while maintaining training efficiency. The tree depth ranges from [4, 10], and the search is conducted within this range to control the model's complexity and fitting ability. A tree depth less than 4 results in an overly simple model (referred to as a weak learner), with limited learning capacity. This model cannot capture the complex nonlinear relationships and interactions between environmental loads and structural responses, leading to underfitting and insufficient model accuracy. Conversely, an excessively large tree depth (e.g., greater than 10) results in a highly complex single decision tree capable of extremely fine-fitting the training data, including noise. This easily leads to overfitting, where the model performs exceptionally well on the training set but poorly on unseen validation and test sets, exhibiting poor generalization ability. For the problem with seven input features in this application, trees with depths between 4 and 10 are sufficient to capture all meaningful feature relationships. The weight coefficients of the regularization term (L1 or L2 regularization term) in the objective function range from [0.5, 5], and the search is conducted within this range to penalize model complexity and prevent overfitting.Regularization terms (L1 or L2) penalize model complexity (such as leaf node weights and the number of leaves), and are a core method for preventing overfitting. A small regularization weight (e.g., less than 0.5) means the penalty is weak, and the model may still tend to become complex, increasing the risk of overfitting. An excessively large regularization weight (e.g., greater than 5) over-penalizes model complexity, forcing the model to become too simple, thereby impairing its ability to learn true patterns in the data and leading to underfitting.

[0082] After training, the model performance is evaluated on the test set. Evaluation metrics include mean absolute error (MAE) and coefficient of determination (COD). A MAE less than 0.03 and a COD greater than 0.85 indicate excellent model performance, accurately representing the mapping relationship under healthy conditions, and suitable for subsequent real-time early warning. If this standard is not met, data quality, feature engineering, or adjustments to the model architecture and optimization strategy are required. MAE represents the average absolute deviation between the model's predicted and actual values. For example, a prediction error of 0.03 degrees for tilt angle features; an error of 0.03 m / s² for acceleration standard deviation. This order of magnitude error is far smaller than the expected structural response change caused by structural damage (such as foundation scour or cracks). This means that the model's prediction error itself will not mask the abnormal signals we truly want to detect, meeting the basic accuracy requirements of the early warning system for the baseline model. The COD measures the model's interpretability of the target variable's variance. A COD greater than 0.85 means the model can capture and explain more than 85% of the response data changes. A COD greater than 0.8 is generally considered to indicate strong correlation.

[0083] The final machine learning model, once trained and accepted, has its internal weights and structure fixed, representing a quantitative mapping relationship between environmental loads and structural responses of the offshore wind turbine tower during its healthy state period.

[0084] Step 4: Set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then obtain the measured value of the structural response parameter of the short-term evaluation window, compare the measured value of the structural response with the predicted value of the structural response, and obtain the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load;

[0085] This step compares the real-time monitored structural response with the predicted values ​​of the health status model, removes the influence of environmental loads, and extracts the abnormal response signal caused purely by changes in the structure's own state. The signal is then enhanced to highlight its trend characteristics.

[0086] In this embodiment, obtaining the residual signal sequence specifically includes:

[0087] The short-term evaluation window set based on the current moment is represented as follows: ,in, Indicates the current moment. This indicates the length of the short-term evaluation window; the length of this window is a configurable system parameter, and its selection requires a trade-off between the real-time nature of monitoring and the stability of the signal. As a non-restrictive parameter, a preferred value for T is 10 minutes. Shorter windows (such as 10 minutes) respond quickly and can capture changes more rapidly.

[0088] There are N sampling times in the short-term evaluation window. Environmental load data is collected at each sampling time within the short-term evaluation window. Environmental feature vectors are extracted at each sampling time within the short-term evaluation window, and then standardized before being input into the trained machine learning model to obtain the response label vector corresponding to each sampling time, which is used as the structural response prediction value. This step is crucial. It is essential to ensure that the data preprocessing methods in the prediction stage are completely consistent with those in the training stage; otherwise, the prediction results will be distorted.

[0089] Next, acquire the structural response data for each sampling time within the short-term evaluation window, calculate the response label vector for the corresponding sampling time, and use it as the measured value of the structural response. This vector represents the structural response that the tower should exhibit under the current environmental conditions in a healthy state. In parallel, for the same sampling time i, synchronously acquire the raw data of the structural response at the sampling time, and calculate the response label vector for this time using the exact same algorithm as in step 2.

[0090] The difference between the measured structural response and the predicted structural response at each sampling time is calculated to obtain the residual signal at that time. The residual signals from all sampling times are then arranged in timestamp order to form a residual signal sequence, denoted as follows: ,in, This represents the residual signal at the first sampling time point within the short-term evaluation window. This represents the residual signal at the second sampling time point within the short-term evaluation window. This represents the residual signal at the i-th sampling time in the short-term evaluation window. This represents the residual signal at the Nth sampling time within the short-term evaluation window, where i is the short-term evaluation window. The sequential index of the internal sampling time;

[0091] The residual signal contains multiple components, specifically including the standard deviation component of the residual X-axis acceleration, the standard deviation component of the residual Y-axis acceleration, the residual X-axis tilt component, the residual Y-axis tilt component, and the cumulative change component of the residual elevation.

[0092] In this embodiment, signal enhancement based on the residual signal sequence specifically includes:

[0093] The residual X-axis tilt component and the residual Y-axis tilt component are synthesized into a single residual tilt component representing the overall tilt degree. The synthesis formula is as follows:

[0094] ;

[0095] in, The residual tilt component is the i-th residual signal. The residual X-axis tilt component in the i-th residual signal. Let be the residual Y-axis tilt component in the i-th residual signal.

[0096] Combining the tilt angles from two directions into a single overall residual tilt component provides a more comprehensive characterization of the overall health of the structural foundation. Foundation anomalies in offshore wind turbine towers (such as uneven settlement, foundation scour, and soil softening) strictly only cause tilting in the downwind or crosswind direction. More often, it is a spatial, vector-like tilt. Monitoring only the X or Y direction may miss the true overall risk. Vector synthesis reflects a more realistic displacement; for example, foundation settlement in the southeast direction may cause the tower to tilt in the northwest direction. This true tilt direction and magnitude are a combined result of the tilts in both the X and Y directions. Through the synthesis formula, we obtain a scalar representing the magnitude or amplitude of the total tilt deviation at the current moment, objectively reflecting the severity of the tilt, unaffected by specific directions. Furthermore, it simplifies monitoring indicators to some extent and improves the robustness of the early warning system. If only two independent early warning channels are set up for X and Y, an early warning may be triggered by a momentary disturbance in one direction (such as the impact of an unusually large wave). After synthesizing an index, it is necessary to continuously judge the overall tilt amplitude, which can effectively filter out instantaneous noise in a single direction, allowing the system to focus more on overall and continuous abnormal changes, thereby reducing the false alarm rate.

[0097] Step 5: Based on the changing trend of the residual signal sequence in different time windows, set multi-level warning thresholds, and trigger the corresponding level of warning signal according to the trend of the residual signal sequence and the multi-level warning thresholds.

[0098] In this embodiment, a short-term evaluation window and a long-term evaluation window are set. Based on the short-term evaluation window corresponding to the current moment, the residual tilt components of all sampling moments within that window are calculated. The linear regression slope of the residual tilt components of all sampling moments within the short-term evaluation window corresponding to the current moment is then calculated as the linear regression slope for the current moment, used to characterize the rate of tilt change at the current moment. This slope is a signed quantity, and its absolute value directly quantifies the instantaneous rate of change in the overall tilt of the tower within that short-term window. A positive slope indicates that the tilt is increasing, and a negative slope indicates that it is recovering. Subsequent warnings primarily focus on the magnitude of its absolute value.

[0099] Time window Defined as a long-term evaluation window, in which Calculate the average value of the residual tilt component within the long-term assessment window at the current moment, and calculate the difference between this average value and the average value of the residual tilt component during the healthy state period of the wind turbine. This difference is used as the medium-to-long-term cumulative deviation at the current moment to characterize the overall accumulated offset of the wind turbine. Here, M represents the length of the long-term assessment window, set to 7 to 30 days, preferably 15 days, to assess the cumulative effect over a longer period. This value is a scalar that quantifies the extent of irreversible offset that the overall tilt state of the current tower has undergone relative to the healthy baseline. A positive value indicates that tilt has accumulated in a certain direction.

[0100] In this embodiment, triggering the corresponding warning signal specifically includes:

[0101] Set multi-level warning thresholds, including a first-level threshold, a second-level threshold, and a third-level threshold; the first-level threshold is specifically set as follows: This threshold is used to determine whether the short-term tilt rate is abnormal. When the recent tilt rate is significantly faster than the typical rate fluctuation during a healthy period, an early warning is triggered even if the cumulative tilt amount is not large. This is an empirical multiplier used to adjust the sensitivity of the early warning system. Its value ranges from 3 to 6, with 4 being preferred. This means that when the short-term rate of change exceeds four standard deviations of the healthy period's fluctuation range, the trend is considered abnormal. In the field of structural health, changes in response rate are often precursors to the initiation or development of damage. For example, in the early stages of foundation scour, the tilting rate of the tower may increase slightly but continuously, but the accumulated absolute tilt may not yet exceed the historical fluctuation range. The condition for setting this primary threshold is precisely to keenly capture this accelerating trend. The primary threshold here concerns both the rate and direction of change; regardless of whether the trend is positive or negative, if its absolute value is too large, it signifies an anomaly. Therefore, it is necessary to capture bidirectional anomalies by considering the absolute value. The secondary threshold is specifically set as follows: This threshold is used to confirm that the cumulative tilt deviation has exceeded the normal fluctuation range of a healthy state, indicating a clear and statistically significant abnormal deviation. The value range is 2 to 4, with 2.5 being preferred. This means that when the cumulative deviation exceeds the baseline of the healthy period plus 2.5 times the standard deviation, an anomaly is confirmed. When the medium-to-long-term cumulative deviation exceeds the secondary threshold, it means that after removing environmental disturbances, the static position of the structure has undergone a statistically significant and non-negligible permanent change. It confirms that the problem suspected in the primary warning does exist and has already had a substantial impact. The specific setting of the tertiary threshold is as follows: This threshold sets a higher safety margin, indicating that the tilt has reached a dangerous level that may jeopardize structural safety, requiring immediate intervention. The value ranges from 4 to 8, and must satisfy the following conditions: Level 5 is preferred. This indicates a risk level far exceeding the typical anomaly confirmation level. This level no longer focuses on trends, but only on the severity of the consequences. When the long-term cumulative deviation exceeds the higher Level 3 threshold, it indicates that the structural condition has deviated drastically from the healthy baseline, possibly approaching the design limits or the critical point of structural failure, posing an immediate risk.

[0102] in, These are the first-level threshold, the second-level threshold, and the third-level threshold. These are the constant coefficients in the corresponding level thresholds, and , , The mean and standard deviation of the medium- to long-term cumulative bias at all collection times in the healthy state sample set. This represents the mean and standard deviation of the linear regression slope at all sampling times in the health status sample set.

[0103] Data fluctuations in a healthy population follow a normal distribution (or an approximately normal distribution). According to the principles of statistical significance testing (such as 95% or 99% confidence intervals), to determine that a data point does not belong to the normal fluctuations of a healthy population, it must deviate sufficiently from the mean. It is this sufficiently large quantification that... This is the fundamental premise of this judgment. The secondary threshold is the anomaly boundary; exceeding it means there is a very high probability of an anomaly. The tertiary threshold, on the other hand, is the danger boundary or safety limit boundary. In design, every structure has a theoretical safety margin, setting the hazard prediction level much higher than the anomaly threshold to establish a buffer zone between confirming an anomaly and facing danger. Therefore, if the tertiary threshold is slightly higher than or roughly the same as the secondary threshold, a crisis may immediately occur upon confirming the anomaly, defeating the purpose of early warning. This ensures the tiered nature of early warnings and the operability of actions. And setting... The significance lies in the fact that only when the observed rate change is significantly greater than this background noise is it considered a genuine danger signal, rather than a random fluctuation. This is entirely consistent with the principle of setting detection thresholds in radar systems to distinguish between real targets and background noise.

[0104] When the absolute value of the linear regression slope is greater than the first-level threshold, a first-level warning is triggered, and a signal is issued indicating a risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the second-level threshold, a second-level warning is triggered, and a signal is issued confirming the risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the third-level threshold, a third-level warning is issued, and a signal is issued to take immediate emergency action.

[0105] In this embodiment, predicted values, measured values, and residual tilt component data were acquired at 35 sampling times. The predicted values ​​of the residual tilt components include predicted values ​​of the residual X-axis tilt component and the residual Y-axis tilt component; the measured values ​​of the residual tilt components include measured values ​​of the residual X-axis tilt component and the residual Y-axis tilt component, as shown in the table below:

[0106] Table 1: Comparison of predicted, measured, and residual tilt component values.

[0107]

[0108] In this embodiment, the preferred setting is a short-term evaluation window of 10 minutes and a long-term evaluation window of 15 days; respectively set , , , , , , Thus, the first-level threshold is calculated. Degree / Window, Secondary Threshold Degree, Level 3 Threshold For example, if the short-term evaluation window contains seven sampling points, and the current sampling time is 26, then the short-term window sequence is [20, 21, 22, 23, 24, 25, 26], and the residual tilt component sequence is [0.33, 0.46, 0.18, 0.01, 0.35, 0.49, 0.36]. If the absolute value of the calculated linear regression slope is less than the first-level threshold, then the first-level warning will not be triggered.

[0109] Please see Figure 4 The present invention also provides a low-frequency oscillation and settlement tilting early warning device for offshore wind turbine towers. This device is used to implement the aforementioned method for early warning of low-frequency oscillation and settlement tilting of offshore wind turbine towers, comprising:

[0110] The health baseline data acquisition module is used to determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. It synchronously collects relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and removes relevant data from the initial operation period when the tower is in an abnormal power generation state.

[0111] The health status sample set construction module is used to extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and to pair environmental feature vectors and response label vectors under the same timestamp to integrate them into a health status sample set.

[0112] The Environment and Response Relationship Analysis Module is used to perform correlation analysis on the environmental feature vectors and response label vectors in the health state sample set to determine the mapping relationship between environmental loads and structural responses during the health state period of wind turbine towers.

[0113] The real-time status monitoring module is used to set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then, it obtains the measured value of the structural response parameter of the short-term evaluation window, compares the measured value of the structural response with the predicted value of the structural response, and obtains the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load;

[0114] The trend analysis and early warning module is used to set multi-level early warning thresholds based on the changing trend of the residual signal sequence in different time windows, and to trigger the corresponding level of early warning signal based on the trend of the residual signal sequence and the multi-level early warning thresholds.

[0115] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0116] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0117] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers, characterized in that, The specific steps include: Step 1: Determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. Collect relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and remove relevant data from the initial operation period when the tower is in an abnormal power generation state. Step 2: Extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and pair environmental feature vectors and response label vectors at the same timestamp to integrate them into a health status sample set; Step 3: Perform correlation analysis on the environmental feature vectors and response label vectors in the healthy state sample set to determine the mapping relationship between environmental loads and structural responses during the healthy state period of the wind turbine tower; Step 4: Set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then obtain the measured value of the structural response parameter of the short-term evaluation window, compare the measured value of the structural response with the predicted value of the structural response, and obtain the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load; Step 5: Based on the changing trend of the residual signal sequence in different time windows, set multi-level warning thresholds, and trigger the corresponding level of warning signal according to the trend of the residual signal sequence and the multi-level warning thresholds.

2. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 1, characterized in that, The collection of the structural response data and environmental load data specifically includes: The no-load commissioning period refers to the stage from the completion of installation of the wind turbine generator set until it is connected to the grid for power generation, during which the wind turbine is in a self-rotating state without power generation load; the initial operation period refers to the continuous and stable operation time of the wind turbine after it is connected to the grid for power generation, and this continuous and stable operation time is at least three months. The structural response data includes the tower's triaxial acceleration, biaxial tilt angle, three-dimensional position information of the tower top, and tidal height. The triaxial acceleration is measured by an inertial measurement unit of a microelectromechanical system (MEMS) inside the top compartment of the tower. Similarly, a biaxial tilt sensor is installed in the tower to obtain the tower's tilt angle in the downwind and crosswind directions. A multi-frequency GNSS receiver is installed at the top of the tower to obtain the three-dimensional position information of the tower top center. A tidal meter is installed on the wind turbine foundation platform to obtain the tidal height. The aforementioned inertial measurement unit and the tower have a fixed structural coordinate system, namely, the X-axis is defined as the windward axis of the tower, the Y-axis is defined as the crosswind axis of the tower, and the Z-axis is vertically upward; The environmental load data includes instantaneous wind speed and direction at the top of the tower, the orientation of the wind turbine nacelle and surface current velocity, as well as the effective wave height and zero-cycle at the location of the wind turbine tower. Wind speed and direction are measured using an ultrasonic anemometer mounted on the tower top anemometer support. Wave radar is installed on the wind turbine foundation platform to continuously measure and record wavefront elevation time-series data at the highest frequency. Fourier transform is used to perform spectral analysis on the wavefront elevation time-series data to calculate the wave energy spectral density function. Spectral moments are calculated based on the energy spectral density function, and the effective wave height is calculated based on the zeroth-order spectral moment. The zero-cycle is also calculated based on the spectral moment. The orientation of the wind turbine nacelle is obtained through the wind turbine's inherent main control system. Relevant sensors are installed on the legs of the wind turbine foundation platform to measure the current velocity profile across the entire water depth, with the data from the uppermost measurement point used as the surface current velocity. The specific method of synchronous acquisition is as follows: During the no-load commissioning period and the initial operation period, a short-term evaluation window is preset, and the arithmetic mean of the instantaneous wind speed at each sampling time within the short-term evaluation window is calculated. This arithmetic mean is used as the average wind speed of the short-term evaluation window. When the average wind speed exceeds the preset start threshold, a preset period of data recording window is automatically started. Within this window, all sensors sample at the highest frequency, and the collected data is stored after being stamped with a uniform timestamp. The highest frequency sampling is not lower than 2Hz.

3. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 2, characterized in that, The logic for removing data from abnormal power generation states during the initial operation period is as follows: read the unit status code from the wind turbine monitoring and data acquisition system, and remove data periods indicating shutdown, fault, emergency shutdown, manual shutdown, start-up and shutdown processes, and operating power below 5% of rated power; at the same time, for environmental load data, remove data periods where the instantaneous wind speed exceeds the cut-out wind speed of the turbine model and the effective wave height exceeds the safe operating wave height threshold; after removing the relevant data from the above periods, the remaining relevant data is regarded as the relevant data under normal power generation state.

4. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 3, characterized in that, The logic for extracting environmental feature parameters and response feature parameters to construct environmental feature vectors and response label vectors is as follows: The environmental feature vector is defined as follows: ,in, For environmental feature vectors, The average wind speed, The sine value of the relative wind direction angle. The cosine value of the relative wind direction angle. For the effective wave height, For crossing the zero cycle, The surface current velocity is used; the environmental characteristic parameters are obtained as follows: for each timestamp, a time step back is taken to determine the time step. A short-term evaluation window is defined; the arithmetic mean of the instantaneous wind speeds at all sampling times within this short-term evaluation window is calculated as the average wind speed; the absolute difference between the wind direction at this timestamp and the orientation of the wind turbine nacelle is calculated as the relative wind direction angle. Since the wind direction is a periodic angular variable, a sine-cosine transformation is used to convert it into two linear features. The transformation formula is as follows: , ,in, The sine value of the relative wind direction angle. The cosine value of the relative wind direction angle. The relative wind direction angle is used; based on the wavefront elevation time series data within the short-term assessment window, the effective wave height and zero-crossing period are calculated by spectral analysis; the arithmetic mean of the surface current velocity at all sampling times within the short-term assessment window is calculated as the surface current velocity at that timestamp. The response tag vector is defined as ,in, Represents the response feature vector. The standard deviation of the X-axis acceleration. The standard deviation of the acceleration along the Y-axis. The tilt angle along the X-axis. The tilt angle along the Y-axis. The cumulative change in elevation is used to represent the cumulative change in elevation. The methods for obtaining various response characteristic parameters are as follows: High-pass filtering is applied to the collected triaxial acceleration data to remove gravitational acceleration components and extremely low-frequency drift. The standard deviations of the X-axis and Y-axis acceleration data within the short-term evaluation window are calculated. Extremely low frequencies refer to frequencies below 0.1 Hz. The collected downwind and crosswind tilt angles are directly used as the X-axis and Y-axis tilt angles. Based on data under normal power generation conditions, three-dimensional location information collected during the first week after the start of the no-load commissioning period under no-severe weather conditions is selected as reference location information. This three-dimensional location information includes the longitude, latitude, and elevation of the tower top center. Its elevation data is used as the original elevation value. The original elevation value and the tidal height at each time stamp represent the environmental disturbance height. The difference between the elevation at each time stamp and the environmental disturbance height is calculated as the cumulative change in elevation to characterize the foundation settlement phenomenon. Response characteristic parameters are extracted from the structural response data. These response characteristic parameters include the standard deviations of the X-axis and Y-axis accelerations, the X-axis and Y-axis tilt angles, and the cumulative change in elevation. After constructing each vector, each data pair with timestamp alignment is defined as a sample point. The data pair includes an environmental feature vector and a response label vector. All sample points are arranged in chronological order to construct a health status sample set.

5. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 1, characterized in that, The training of a machine learning model specifically includes: Representing the mapping relationship using a fully trained machine learning model. Before training, the environmental feature vector and response label vector are standardized to convert them into data with a mean of 0 and a standard deviation of 1. The health status sample set is divided into a training set, a validation set, and a test set in a ratio of 6:2:

2. The standardized environmental feature vectors in the health status sample set are used as inputs, and the standardized response label vectors corresponding to the timestamps are used as outputs to construct a standard response prediction model. During training, predictions are made on the validation set every 10 iterations, and the loss function on the validation set is calculated. A patience value is preset for each iteration. If the loss function value on the validation set no longer decreases within consecutive patience values, training automatically terminates. The loss function is the mean squared error, which aims to minimize the sum of squared prediction errors across all output dimensions. The model is trained on the training set, and a Bayesian optimization algorithm is used on the validation set to automatically optimize key hyperparameters. The key hyperparameters include at least the learning rate, tree depth, and regularization parameter. The learning rate is set to range from [0.01, 0.3], the tree depth to range from [4, 10], and the weight coefficient of the regularization term in the objective function to range from [0.5, 5]. After training, the model performance is evaluated on the test set. The evaluation metrics include mean absolute error and coefficient of determination. When the mean absolute error is less than 0.03 and the coefficient of determination is greater than 0.85, the model is considered to have excellent performance; otherwise, the model will be retrained.

6. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 5, characterized in that, Obtaining the residual signal sequence specifically includes: The short-term evaluation window set based on the current moment is represented as follows: ,in, Indicates the current moment. Indicates the length of the short-term assessment window; There are N sampling times in the short-term evaluation window. Environmental load data are collected at each sampling time within the short-term evaluation window. Environmental feature vectors are extracted at each sampling time within the short-term evaluation window, and then standardized and input into the trained machine learning model to obtain the response label vector corresponding to each sampling time, which is used as the structural response prediction value. Next, obtain the structural response data for each sampling time within the short-term evaluation window, calculate the response label vector for the corresponding sampling time, and use it as the measured value of the structural response. Calculate the difference between the measured value of the structural response at each sampling time and the predicted value of the structural response at the corresponding time to obtain the residual signal at that time. Arrange the residual signals of all sampling times in timestamp order to form a residual signal sequence, represented as follows: ,in, This represents the residual signal at the first sampling time point within the short-term evaluation window. This represents the residual signal at the second sampling time point within the short-term evaluation window. This represents the residual signal at the i-th sampling time in the short-term evaluation window. This represents the residual signal at the Nth sampling time within the short-term evaluation window, where i is the short-term evaluation window. The sequential index of the internal sampling time; The residual signal contains multiple components, specifically including the standard deviation component of the residual X-axis acceleration, the standard deviation component of the residual Y-axis acceleration, the residual X-axis tilt component, the residual Y-axis tilt component, and the cumulative change component of the residual elevation.

7. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 6, characterized in that, Signal enhancement based on residual signal sequences specifically includes: The residual X-axis tilt component and the residual Y-axis tilt component are synthesized into a single residual tilt component representing the overall tilt degree. The synthesis formula is as follows: ； in, The residual tilt component is the i-th residual signal. The residual X-axis tilt component in the i-th residual signal. Let be the residual Y-axis tilt component in the i-th residual signal.

8. The method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 7, characterized in that, Set short-term evaluation window and long-term evaluation window. Based on the short-term evaluation window corresponding to the current time, calculate the residual tilt components of all sampling times within it, and calculate the linear regression slope of the residual tilt components of all sampling times within the short-term evaluation window corresponding to the current time, which is used as the linear regression slope of the current time to characterize the tilt change rate at the current time. Time window Defined as a long-term evaluation window, in which Calculate the average value of the residual tilt component within the long-term evaluation window at the current moment, and calculate the difference between this average value and the average value of the residual tilt component during the healthy state period of the wind turbine. Use this difference as the medium-to-long-term cumulative deviation at the current moment to characterize the overall offset accumulated by the wind turbine. Here, M represents the length of the long-term evaluation window.

9. A method for early warning of low-frequency oscillation, settlement, and tilting of offshore wind turbine towers according to claim 8, characterized in that, The specific triggering of the corresponding warning signals includes: Set multi-level warning thresholds, including a first-level threshold, a second-level threshold, and a third-level threshold; the first-level threshold is specifically set as follows: The secondary threshold is specifically set as follows: The specific threshold settings for the three levels are as follows: ,in, These are the first-level threshold, the second-level threshold, and the third-level threshold. These are the constant coefficients in the corresponding level thresholds, and , , The mean and standard deviation of the medium- to long-term cumulative bias at all collection times in the healthy state sample set. The mean and standard deviation of the linear regression slope at all sampling times in the health status sample set represent the sum of the slopes. When the absolute value of the linear regression slope is greater than the first-level threshold, a first-level warning is triggered, and a signal is issued indicating a risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the second-level threshold, a second-level warning is triggered, and a signal is issued confirming the risk of low-frequency oscillation and settlement tilting. When the medium- and long-term cumulative deviation is greater than the third-level threshold, a third-level warning is issued, and a signal is issued to take immediate emergency action.

10. A low-frequency oscillation, settlement, and tilting early warning device for offshore wind turbine towers, characterized in that, The aforementioned low-frequency swaying, settlement, and tilting early warning device for offshore wind turbine towers is used to implement the low-frequency swaying, settlement, and tilting early warning method for offshore wind turbine towers as described in any one of claims 1-9, comprising: The health baseline data acquisition module is used to determine the health status period of the wind turbine tower, including the no-load commissioning period and the initial operation period. It synchronously collects relevant data of the wind turbine tower during the health status period, including structural response data and environmental load data, and removes relevant data from the initial operation period when the tower is in an abnormal power generation state. The health status sample set construction module is used to extract environmental feature parameters and response feature parameters from the removed relevant data to construct environmental feature vectors and response label vectors, and to pair environmental feature vectors and response label vectors under the same timestamp to integrate them into a health status sample set. The Environment and Response Relationship Analysis Module is used to perform correlation analysis on the environmental feature vectors and response label vectors in the health state sample set to determine the mapping relationship between environmental loads and structural responses during the health state period of wind turbine towers. The real-time status monitoring module is used to set a short-term evaluation window based on the current time, monitor the environmental load data of the short-term evaluation window and extract the environmental feature vector, and obtain the predicted value of the structural response parameter under the short-term evaluation window by combining the mapping relationship; then, it obtains the measured value of the structural response parameter of the short-term evaluation window, compares the measured value of the structural response with the predicted value of the structural response, and obtains the residual signal sequence to characterize the structural response change caused by the change of the wind turbine tower's own state after removing the influence of environmental load; The trend analysis and early warning module is used to set multi-level early warning thresholds based on the changing trend of the residual signal sequence in different time windows, and to trigger the corresponding level of early warning signal based on the trend of the residual signal sequence and the multi-level early warning thresholds.

Images