A wind turbine tower health monitoring method and system

Abstract

The present application relates to the technical field of wind power tower monitoring, and discloses a wind power tower health monitoring method and system. The method collects structural vibration signals and environmental load data of the wind power tower, extracts dynamic response characteristics of the tower, and establishes a basic monitoring data set of the tower operation state in combination with the environmental load data; the basic monitoring data set is subjected to multidimensional decomposition processing, and the inherent frequency component, damping characteristic parameter and external excitation coupling component of the tower are separated out; a first evaluation index of the tower structure health is constructed according to the correlation between the inherent frequency component and the damping characteristic parameter; historical damage records and material degradation data of the tower are obtained, the historical damage records are subjected to time sequence alignment analysis with the external excitation coupling component, and a second evaluation index of the tower damage evolution is generated; the first evaluation index and the second evaluation index are fused, a comprehensive scoring model of the tower health state is established, the health grade of the tower is divided, and a real-time monitoring report is output.

Description

Technical Field

[0001] This invention relates to the field of wind turbine tower monitoring technology, specifically to a method and system for monitoring the health of wind turbine towers. Background Technology

[0002] As a crucial component of clean energy, the safe and stable operation of wind power facilities is paramount. Wind turbine towers, as key load-bearing structures supporting wind turbine generators, are subjected to complex alternating stresses caused by wind loads, generator vibrations, and changes in the external environment. These factors can lead to fatigue damage, material degradation, and even potential structural failure risks in the tower structure. Traditional tower inspections rely primarily on periodic manual inspections or simple vibration threshold alarms, methods with significant limitations. Manual inspections are time-consuming, costly, and struggle to detect early-stage damage within the structure; while simple vibration over-limit alarms often only trigger when the structural condition has already deteriorated significantly, lacking early warning capabilities.

[0003] The structural health of a tower is closely related to its dynamic response characteristics. Dynamic parameters such as natural frequency and damping ratio are important indicators reflecting the overall stiffness, mass, and energy dissipation capacity of the structure. When cracks appear, connections loosen, or material properties deteriorate, these parameters undergo subtle but measurable changes. However, in actual operating environments, the tower's vibration signal is the result of the combined effects of its own structural dynamic characteristics and external wind loads, turbine operating excitation, and other factors. The signal composition is complex, and effective structural characteristic information is often drowned out by strong environmental noise. Relying solely on single vibration signal analysis makes it difficult to separate changes in the structure's inherent characteristics from the complex environmental response, thus hindering accurate identification of early damage. Furthermore, the health state of a tower is a dynamic evolutionary process, with the current damage state closely related to its historical load history. Existing monitoring methods mostly focus on instantaneous assessment of the current state, lacking the ability to correlate real-time monitoring data with historical damage records and material aging data, thus failing to effectively predict the structure's remaining life and damage development trends. Therefore, an intelligent monitoring method that can deeply integrate real-time dynamic response, environmental load, and historical data is needed to achieve accurate assessment and early warning of the health status of wind turbine tower structures. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring the health of wind turbine towers, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for monitoring the health of wind turbine towers, the method comprising:

[0006] Collect structural vibration signals and environmental load data of wind turbine towers, extract dynamic response characteristics of towers based on structural vibration signals, and establish a basic monitoring dataset of tower operation status by combining environmental load data;

[0007] The basic monitoring dataset is decomposed in multiple dimensions to separate the natural frequency components, damping characteristic parameters and external excitation coupling components of the tower. Based on the correlation between the natural frequency components and damping characteristic parameters, a first assessment index for the structural health of the tower is constructed.

[0008] The historical damage records and material degradation data of the tower are acquired, and the historical damage records are time-series aligned with the external excitation coupling components to generate a second evaluation index for the tower damage evolution.

[0009] By integrating the first and second assessment indicators, a comprehensive scoring model for the health status of the tower is established. Based on the comprehensive scoring model, the health level of the tower is classified and a real-time monitoring report is output.

[0010] Preferably, the dynamic response characteristics of the extraction tower include:

[0011] The energy distribution characteristics and phase shift of the structural vibration signal are extracted by performing a time-frequency joint transformation.

[0012] The amplitude deviation of energy distribution characteristics is corrected by environmental load data, and the local deformation gradient of the tower is calculated by combining the phase offset.

[0013] Based on the relationship between the local deformation gradient and the preset threshold, potential abnormal regions of the tower are marked and a dynamic response feature set is generated.

[0014] Preferably, the first assessment indicator for the health of the constructed tower structure includes:

[0015] Modal parameter identification is performed on the natural frequency components, and the frequency shift rate and mode shape correlation of each mode are calculated.

[0016] The energy dissipation efficiency of the tower structure is quantified based on the attenuation relationship between damping characteristic parameters and frequency offset rate.

[0017] By combining modal correlation and energy dissipation efficiency, a weighted scoring matrix for the first evaluation index is generated.

[0018] Preferably, the second evaluation index for the evolution of tower damage includes:

[0019] Historical damage records are classified according to damage type and matched with the temporal distribution characteristics of external excitation coupling components;

[0020] Analyze the propagation rate of similar damage under external excitation and calculate the linear fitting coefficient of the damage accumulation effect;

[0021] The dynamic correction factor for the second evaluation index is generated based on the product of the linear fitting coefficient and the current external excitation coupling component.

[0022] Preferably, the comprehensive scoring model for establishing the health status of the tower includes:

[0023] The weighted scoring matrix of the first evaluation indicator is normalized to obtain the structural health sub-score;

[0024] The dynamic correction factor of the second evaluation indicator is convolved with the structural health sub-score to generate the initial value of the comprehensive score.

[0025] The weighting of the comprehensive score is adjusted based on the tower's design life parameters and the time decay coefficient of real-time monitoring data.

[0026] Preferably, the classification of the health levels of the tower includes:

[0027] Set a grading threshold for the comprehensive score, and divide the health level range according to the degree of deviation between the initial value and the threshold;

[0028] Extract high-frequency abnormal components from the dynamic response feature set. If the high-frequency abnormal components exceed the upper limit of the level range, trigger a health level recalibration command.

[0029] The parameters of the comprehensive scoring model are updated based on the recalibration command, and a new real-time monitoring report is generated.

[0030] Preferably, the output real-time monitoring report includes:

[0031] Map health level ranges to predefined tower maintenance strategy codes;

[0032] Integrate the temporal characteristics of structural vibration signals, the spatial distribution of environmental load data, and maintenance strategy codes to generate standardized report templates;

[0033] Based on the tower's location identifier and timestamp, the standardized report template is compressed into a binary data stream and stored in a distributed database.

[0034] Preferably, the method further includes:

[0035] Regularly retrieve historical monitoring reports from a distributed database to extract data on the changing trends of health levels;

[0036] A prediction model for the remaining life of the tower is trained using trend data, and the life prediction results are output to the feedback channel of the comprehensive scoring model.

[0037] The fusion weights of the first and second evaluation indicators are dynamically adjusted based on the error rate of the feedback channel.

[0038] Preferably, the method further includes:

[0039] When a potential abnormal area in the tower is detected, a high-precision laser scanning device is activated to acquire three-dimensional deformation data of the tower surface;

[0040] Spatial registration is performed between 3D deformation data and dynamic response feature set to correct the calculation error of local deformation gradient;

[0041] The health level range of the real-time monitoring report is updated based on the corrected deformation gradient.

[0042] Preferably, the present invention also includes a wind turbine tower health monitoring system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described wind turbine tower health monitoring method.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention improves the accuracy and foresight of wind turbine tower structural health status assessment through multi-source data fusion and deep feature analysis. The establishment of a basic monitoring dataset links structural vibration response with environmental load conditions, providing more comprehensive contextual information for analysis. Multi-dimensional decomposition processing effectively separates the natural frequencies and damping parameters characterizing the tower's structural properties, as well as the response components induced by external loads, from complex mixed signals. This separation protects the assessment of the structure's condition from excessive interference from changes in environmental excitation, enhancing the stability of monitoring results. The first assessment index, constructed based on the correlation between natural frequencies and damping characteristics, can sensitively capture subtle changes in the overall structural stiffness and energy dissipation characteristics, making it possible to identify early damage.

[0045] A key feature of this method is the introduction of historical damage records and material degradation data, coupled with time-series alignment analysis with external excitation components to generate a second assessment index. This allows the assessment to consider not only the current state but also the cumulative effects and development trends of damage, extending a static health snapshot into a dynamic health evolution map. Finally, a comprehensive scoring model integrating the first and second assessment indices enables a multi-faceted and three-dimensional evaluation of the tower's health status. The health levels defined by this model and the generated monitoring reports provide a clearer picture of the tower's structural safety, helping maintenance personnel identify potential risk areas and understand possible damage development paths. This approach helps drive a shift in tower maintenance strategies from periodic inspections and reactive measures to condition-based predictive maintenance, thereby optimizing maintenance resource allocation and improving the safe operation and life-cycle economic benefits of wind power facilities. Attached Figure Description

[0046] Figure 1 A comparison diagram of vibration and deformation at different heights of the tower;

[0047] Figure 2 Flowchart for extracting the dynamic response characteristics of the tower;

[0048] Figure 3 Flowchart for constructing the first assessment indicator of tower structure health;

[0049] Figure 4 This is a diagram showing the modal correlation between each mode pair. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 This invention provides a method and system for monitoring the health of wind turbine towers. The method includes the acquisition and analysis of integrated structural vibration signals and environmental load data to achieve real-time assessment and prediction of the health status of wind turbine towers. The overall implementation scheme of the method is based on acquiring structural vibration signals and environmental load data of the wind turbine towers. Dynamic response characteristics of the towers are extracted based on the structural vibration signals, and a basic monitoring dataset for the tower's operating status is established by combining the environmental load data. The basic monitoring dataset is decomposed in multiple dimensions to separate the natural frequency components, damping characteristic parameters, and external excitation coupling components of the tower. A first assessment index for the structural health of the tower is constructed based on the correlation between the natural frequency components and the damping characteristic parameters. Historical damage records and material degradation data of the tower are acquired, and time-series alignment analysis is performed on the historical damage records and external excitation coupling components to generate a second assessment index for the tower's damage evolution. The first and second assessment indices are integrated to establish a comprehensive scoring model for the tower's health status. Based on the comprehensive scoring model, the health level of the tower is classified, and a real-time monitoring report is output. This scheme, through data-driven multi-dimensional analysis, ensures the comprehensiveness and accuracy of the monitoring process, providing technical support for maintenance decisions of wind turbine towers.

[0052] Example 1: See Figure 2The structural vibration signals are derived from a piezoelectric accelerometer array installed at key locations on the wind turbine tower. The sensors are uniformly distributed to cover the main mode shapes along the tower's height, with a sampling frequency of 512Hz to capture effective vibration components within the 0-200Hz range. The raw signals are transmitted to a signal conditioning module for anti-aliasing filtering and gain adjustment to eliminate high-frequency noise and DC offset. The time-frequency joint transformation employs a short-time Fourier transform algorithm, dividing the continuously acquired vibration signals into 256-millisecond Hanning windows with a 50% window overlap rate. Each window undergoes a fast Fourier transform to obtain the time-spectrum matrix. Energy distribution characteristics are obtained by calculating the power spectral density integral over time at each frequency point. The phase offset is extracted by constructing an analytical signal using a Hilbert transform to obtain the instantaneous phase angle. Environmental load data are synchronously acquired by an ultrasonic anemometer mounted on the top of the tower and a temperature sensor on the foundation platform, achieving a wind speed resolution of 0.1 m / s and a temperature measurement accuracy of ±0.5℃. To correct for amplitude deviations in energy distribution characteristics, a wind speed-vibration transfer function model is established. The transfer function refers to the transfer function between environmental loads and structural vibration energy, and it adopts the form of a linear time-invariant system transfer function. The transfer function form is as follows:

[0053]

[0054] in, Wind speed power spectral density (unit: m) 2 / s 3 ·Hz), Energy spectral density of structural vibration (unit: m) 2 / s 2 ·Hz), The gain coefficient (unit: s) is determined by the tower height and cross-sectional dimensions. Frequency (unit: Hz) The first natural frequency of the tower (unit: Hz) is obtained from modal parameter identification. The first-order damping ratio of the tower is obtained from the damping characteristic parameter extraction results. The baseline values ​​of vibration energy corresponding to different wind speed ranges are stored as a lookup table, and the residuals between the measured energy values ​​and the baseline values ​​are fitted using the least squares method to obtain the correction coefficient. In the temperature compensation stage, the theoretical frequency drift is calculated using the material's thermal expansion coefficient, and the energy distribution is normalized for temperature. The phase offset is bound to the tower's geometric coordinate system, and the phase propagation vector is calculated using the sensor's position coordinates.

[0055] The calculation of local deformation gradient involves discretizing the tower surface into a finite element mesh, with each mesh node associated with the nearest sensor phase data. The gradient value is calculated using the central difference method, determining the ratio of the phase difference between adjacent nodes to the Euclidean distance. The mesh size is set to 0.5 m × 0.5 m based on the tower diameter. A preset threshold is derived from the maximum allowable dynamic deflection value in the tower design specifications, converted to a phase gradient threshold range of 0.15-0.25 radians / m. Potential anomaly regions are marked using a region growing algorithm, expanding the connected region from the out-of-limit mesh point. The dynamic response feature set is stored in a 3D point cloud format, showing gradient distribution, energy peaks, and phase abrupt change markers. Real-time processing of the time-frequency joint transformation is implemented using a digital signal processor, with the spectral matrix of each time window stored in a circular buffer. The calculation of energy distribution characteristics integrates a spectral moment estimation method, calculating the energy integral for the third-order frequency components (0.5 Hz, 1.0 Hz, and 1.5 Hz) of the main frequency bands. Phase offset extraction employs orthogonal demodulation technology, mixing the vibration signal with a local oscillator to obtain in-phase and quadrature components. An adaptive filtering mechanism is introduced to correct environmental load data, and the nonlinear relationship of vibration energy caused by wind speed changes is modeled by multinomial regression.

[0056] The calculation of local deformation gradients includes a surface fitting step, using bicubic spline interpolation to reconstruct a continuous surface from discrete phase data. A dynamic threshold adjustment mechanism, linked to real-time wind speed data, automatically increases the threshold by 20% under strong wind conditions to reduce false alarm rates. The marking of potential anomaly areas, combined with morphological filtering, eliminates the influence of isolated noise points. The update frequency of the dynamic response feature set is synchronized with data acquisition, and the feature vector includes timestamps, spatial coordinates, and gradient confidence indices. The preprocessing of structural vibration signals includes an outlier removal algorithm, using the Laida criterion to remove data points exceeding three standard deviations. The window function optimization for the time-frequency joint transform is based on the periodic characteristics of the tower's main vibration modes, with the window length designed as an integer multiple of the fundamental frequency period. The frequency band division of energy distribution characteristics corresponds to the first six modal frequencies of the tower, with the energy in each band normalized as a percentage of the total energy. The calculation of phase offset compensates for signal propagation delay, correcting the time difference based on sensor spacing and material sound velocity.

[0057] The environmental load data correction model considers the fluid-structure coupling effect, and the transfer function of wind speed and vibration energy includes a Reynolds number correction term. The temperature compensation stage integrates a material fatigue damage model, and the change in elastic modulus at different temperatures is converted into frequency compensation using empirical formulas. The calculation and verification of local deformation gradients employs a cross-validation method, comparing the differences in gradient calculation results between adjacent sensor groups. Preset threshold grading corresponds to safety standards for different height sections of the tower, with the threshold at the upper part of the tower being 15% higher than that at the foundation. Spatial localization of potential anomaly areas uses a multi-signal classification algorithm, inverting the coordinates of the anomaly source through the phase information of the sensor array. The data structure of the dynamic response feature set includes a header file and a feature matrix. The header file records the tower number, acquisition time, and sensor status. The rows of the feature matrix correspond to the time series, and the columns contain energy values, phase angles, and gradient vectors. The entire processing flow is embedded in a real-time operating system task scheduler, ensuring data processing latency is less than 100 milliseconds.

[0058] The spectral resolution of the time-frequency joint transform is set to 0.5Hz to meet the tower mode frequency separation requirements. The time evolution curve of energy distribution characteristics is tracked using a sliding window, with the window length adapting to the wind speed change rate. Temperature drift compensation for phase offset is achieved using differential measurement with dual temperature sensors to eliminate the influence of ambient temperature gradients. Uncertainty quantification is introduced into the calculation of local deformation gradients, and the impact range of sensor errors on gradient values ​​is evaluated using the Monte Carlo method. An adaptive mechanism for preset thresholds integrates a machine learning classifier, training a threshold adjustment strategy based on historical anomaly data. The severity of potential anomaly areas is graded into three levels: mild, moderate, and severe, based on the duration and spatial range of gradient exceedance. The compressed storage of the dynamic response feature set uses a lossy compression algorithm, reducing storage space by 75% while maintaining feature accuracy. The entire implementation scheme is verified for functional integrity through hardware-in-the-loop testing, with test cases covering different wind speed ranges and typical failure modes. Multidimensional decomposition processing employs a hybrid algorithm combining Variational Mode Decomposition (VMD) and Empirical Mode Decomposition (EMD). Data preprocessing involves detrending the structural vibration signals (sampling frequency 512Hz) in the basic monitoring dataset to eliminate DC offset caused by sensor installation errors. Moving average filtering (window length 100 sampling points) is used to smooth high-frequency noise. During VMD decomposition, VMD parameters are set (number of modes K=6, penalty factor α=2000, noise tolerance τ=1e-7), decomposing the vibration signal into six Intrinsic Mode Functions (IMFs). The first three IMFs correspond to the tower's natural frequency components (first-order bending, second-order bending, and first-order torsional modes), while the latter three IMFs correspond to external excitation coupling components (wind load, unit operating vibration). During EMD secondary purification, the natural frequency components obtained from VMD decomposition are processed using EMD to separate modal aliasing components (such as spurious modes caused by environmental noise), retaining the IMF components that truly reflect the tower's structural characteristics. When extracting damping characteristic parameters, the half-power bandwidth method is used for the purified natural frequency components. The frequency difference between the half-power points on both sides of the resonance peak of each mode is calculated, and the formula is as follows:

[0059]

[0060] in, The modal resonance frequency, , The frequency is the half-power point. Damping characteristic parameters are extracted, and correlation analysis shows that the correlation coefficient between the natural frequency component and the external excitation coupling component is <0.3, verifying the decomposition effect and ensuring that there is no cross-interference between the two types of components.

[0061] The synchronization of structural vibration signal acquisition is ensured by a GPS clock module, with time synchronization errors of less than 1 millisecond for each sensor. Parallel computation of the time-frequency joint transformation utilizes a multi-core processor architecture to process data from eight sensor channels simultaneously. Frequency weighting of energy distribution characteristics considers human perception of vibration, employing the frequency weighting curve from the ISO 2631 standard. Dynamic calibration of phase offset is achieved through a reference sensor; a fixed-point sensor installed on the tower foundation provides the phase reference. Environmental load data quality inspection includes integrity verification, with missing data supplemented using autoregressive interpolation. Visualization of local deformation gradients employs color mapping technology, mapping gradient values ​​to a heatmap superimposed on the tower's 3D model. Seasonal adjustment of preset thresholds considers the thermal expansion and contraction of materials, automatically increasing the threshold by 5% in winter. Tracking and analysis of potential anomaly areas combines image processing technology, using optical flow to calculate the movement trajectory of the anomaly areas. The transmission protocol for the dynamic response feature set adopts the Industrial Internet of Things (IIoT) standard.

[0062] Example 2: See Figure 3 This involves a technical solution that combines historical damage records to generate a second assessment index. The basic monitoring dataset, with its natural frequency components decomposed in multiple dimensions, is used for modal parameter identification, employing a random subspace algorithm to extract modal parameters from the natural frequency components. The Hankel matrix is ​​based on discrete sampling sequences of structural vibration signals. (N is the number of sampling points, taken as 2048) Construction, specific parameters are as follows: Row dimension (delay order): set to 2n (n is the number of predicted tower modes, taken as 6, corresponding to the first 6 main modes), that is, row dimension = 12. Column dimension: set to Calculated The matrix form is as follows:

[0063]

[0064] Hanning windows are used to weight each column of the Hankel matrix to reduce decomposition errors caused by boundary effects. The system order is determined using a stability graph method combined with singular value contribution rates. The steps are as follows: perform singular value decomposition on the Hankel matrix H and extract the singular value sequence. The formula for the Hankel matrix H is as follows:

[0065]

[0066] Where U is the left singular vector matrix, It is a singular value diagonal matrix. It is the transpose of the right singular vector matrix. The formula for calculating the proportion of each singular value to the total singular values ​​is as follows:

[0067]

[0068] in, For the i-th singular value, Let m be the sum of all valid singular values, and m be the row dimension of the Hankel matrix H. Mutation decrease (e.g. from) Down to When the order is initially determined to be 6, the upper limit of the order is initially determined to be 6. The order search range is set (1-20 orders). The modal frequency points identified under different orders are plotted with "order-frequency" as the coordinate axis. Stable modal points with frequency deviation <0.5Hz in 3 or more consecutive orders are marked. The number of stable modal points is consistent with the estimated number of modes (6), and the cumulative proportion of the top 6 singular value contribution rates is >90%. Finally, the system order is determined to be 6. The implementation of the random subspace algorithm is to construct the Hankel matrix to perform matrix decomposition on the vibration data, use singular value decomposition to determine the system order, and obtain the modal frequency and damping ratio through least squares estimation. The frequency offset rate is calculated as the percentage of relative deviation between the current modal frequency and the baseline frequency. The baseline frequency is taken from the modal test report when the tower leaves the factory. The modal correlation is calculated by the modal guarantee criterion. The current mode vector and the reference mode vector are normalized after the dot product operation. The relationship between damping characteristic parameters and frequency deviation rate is established based on a viscous damping model. Energy dissipation efficiency is quantified as the product of damping ratio and frequency deviation rate divided by the critical damping coefficient. The weighted scoring matrix of the first evaluation index is constructed using the analytic hierarchy process (AHP). The weight allocation of the matrix is ​​determined based on the contribution of each mode to the overall stiffness of the tower, and the contribution is evaluated using modal strain energy calculation. The output of the weighted scoring matrix is ​​a six-row, four-column matrix structure, with rows corresponding to the six main vibration modes of the tower, and columns containing frequency score, damping score, mode shape score, and comprehensive score.

[0069] Historical damage records are derived from the wind turbine tower operation and maintenance database. Damage type classification employs a rule-based hierarchical classification method, dividing damage into three main categories: fatigue damage, corrosion damage, and loose connections. The temporal distribution characteristics of the external excitation coupling components are extracted using wavelet packet decomposition technology, decomposing the excitation signal into sub-signals of different frequency bands before calculating statistical features. The propagation rate of similar damage under external excitation is calculated using a damage evolution model. For fatigue damage, the Paris law model is adopted, specifically the Paris-Erdogan crack propagation model, adapted to the fatigue characteristics of Q345 steel, the mainstream material for wind turbine towers. The model describes the crack length propagation law with the number of cyclic loads, focusing on key areas prone to fatigue cracks such as welded joints and flange connections. Core formula:

[0070]

[0071] in, Crack propagation rate (unit: mm / cycle). The fatigue constant of the material (Q345 steel under atmospheric conditions) Unit: mm / (MPa·m(1 / 2))m). The fatigue index of the material. The stress intensity factor amplitude (unit: MPa·m^(1 / 2)) is calculated as follows:

[0072]

[0073] in, As a geometric factor, the welded joint is set to 1.8; The alternating stress amplitude is expressed in MPa. Crack length, in meters (m). The Miner linear cumulative damage criterion is used.

[0074]

[0075] in, This represents the cumulative damage value. Let i be the number of cycles under the i-th load segment; Let be the fatigue life under the i-th segment of load, when It was determined to be fatigue failure at that time.

[0076] Corrosion damage was modeled using corrosion kinetics, while loosening of connectors was modeled using stiffness degradation. The linear fitting coefficients for the cumulative damage effect were obtained through multiple linear regression analysis, with the independent variables being the characteristic parameters of the external excitation and the dependent variable being the monthly average increase in damage size. The dynamic correction factor for the second evaluation index was constructed as the dot product of the linear fitting coefficient matrix and the current external excitation characteristic vector. The normalization of the dynamic correction factor used the sigmoid function to map the output value to the 0-1 interval.

[0077] Real-time modal parameter identification employs a recursive random subspace algorithm, updating modal parameter estimates every five minutes. The frequency offset calculation incorporates a temperature compensation mechanism, correcting for temperature-induced frequency drift based on the material's thermal expansion coefficient. Modal shape correlation calculation considers the sensor layout, eliminating errors caused by sensor sparsity through a weighted orthogonality test. Energy dissipation efficiency calculation includes a nonlinear correction term, using a cubic polynomial to fit the damping ratio versus amplitude curve. The weighted scoring matrix update mechanism is linked to the external excitation intensity, automatically increasing the weight of the damping score when wind speed exceeds the rated value. Historical damage record data preprocessing includes anomaly filtering rules, removing damage records that clearly do not conform to physical laws. Temporal feature extraction of external excitation coupling components includes twelve statistical features, covering mean, variance, skewness, kurtosis, waveform factor, and impulse factor parameters.

[0078] Damage propagation rate was calculated using a piecewise linearization method based on time series, dividing the damage development process into multiple linear stages. The significance of the linear fitting coefficients was tested using a t-test, retaining coefficients with a significance level exceeding 95%. Smoothing of the dynamic correction factor employed a moving average filter, with the window length set to 24 hours based on the data sampling frequency. Noise suppression during modal parameter identification utilized correlation function analysis, improving the signal-to-noise ratio through autocorrelation function preprocessing. Long-term trend analysis of frequency offset rate employed a seasonal decomposition method, separating the trend offset caused by aging from the periodic fluctuations caused by load. Spatial compensation for modal shape correlation considered the cross-sectional changes of the tower at different elevations, using a geometric amplification factor to correct the mode shape vector. Environmental correction for energy dissipation efficiency included a humidity influence factor, adjusting the damping ratio calculation results based on relative humidity data. Consistency testing of the weighted scoring matrix was achieved by calculating a consistency index, automatically adjusting matrices that did not meet consistency requirements. Standardization of historical damage records employed a unified damage quantification standard, converting damage dimensions reported by different detection methods into equivalent crack lengths. The feature selection of the external excitation coupling components adopts the principal component analysis method, which reduces the feature dimension while retaining 95% of the original information.

[0079] The confidence interval for damage propagation rate was estimated using the Bootstrap method to generate a probability distribution for damage prediction. The rolling update of linear fitting coefficients employed the forgetting factor least squares method, prioritizing the influence of recent data. Outlier detection for the dynamic correction factor utilized the Isolation Forest algorithm to eliminate anomalous excitation data caused by sensor malfunctions. The entire implementation process was virtually verified using a digital twin platform; the tower's digital model included detailed information on material properties, geometric dimensions, and boundary conditions. The modal parameter identification algorithm ran in a multi-threaded manner within the embedded system, with a real-time thread handling data acquisition and preprocessing, and a background thread performing complex matrix operations. The warning threshold for frequency offset rate was set using a dynamic percentile method, with the 95th percentile of the most recent 30 days as a benchmark. The visualization of modal correlation was presented using 3D animation, with color gradients indicating the degree of modal variation. A trend line warning was set for the long-term monitoring curve of energy dissipation efficiency, triggering a check notification when the efficiency value decreased for ten consecutive days. The weighted scoring matrix was persistently stored using a time-series database structure, retaining complete matrix data for each evaluation. The knowledge graph of historical damage records is constructed using graph database technology to establish relationships between damage types, locations, and sizes. Real-time stream processing of external stimulus coupling components employs a complex event processing engine to achieve millisecond-level feature extraction.

[0080] The damage propagation rate prediction model integrates physical and data-driven models, employing a weighted fusion approach to improve prediction accuracy. Version management of linear fitting coefficients utilizes a Git-style workflow, recording the changes and reasons for each coefficient update. Distributed computation of dynamic correction factors employs a MapReduce architecture, supporting parallel computation of data from multiple towers. Quality assessment of modal parameter identification results includes six validation metrics, covering modal amplitude coherence coefficient, modal confidence factor, and modal phase collision parameter. Spatial visualization of frequency offset rate uses a tower elevation-color mapping map, with gradient colors representing the degree of offset at different heights. Anomaly detection of modal mode correlation employs a control chart method, setting upper and lower control limits for the correlation coefficient. Baseline traceability of energy dissipation efficiency utilizes blockchain technology, recording the timestamp and responsible party for each baseline value change. Remote transmission of the weighted scoring matrix employs data compression technology, using differential encoding to reduce matrix data transmission volume. Image recognition of historical damage records utilizes a convolutional neural network to automatically identify and quantify damage size from detection photographs. Frequency domain analysis of the external excitation coupling components employs order analysis techniques to convert frequency components into harmonics of rotational speed. Reliability assessment of damage propagation rate utilizes Monte Carlo simulation to model damage development paths under different operating conditions. Adaptive learning of linear fitting coefficients employs a reinforcement learning mechanism, automatically adjusting coefficient weights based on prediction errors. The time decay model for the dynamic correction factor uses an exponential decay function, with earlier excitation data contributing a gradually decreasing percentage to the factor. Data flow is asynchronously processed via message queues to ensure system stability under high loads.

[0081] The implementation of the random subspace algorithm includes a data block partitioning strategy, dividing the continuous data stream into fixed-length blocks for batch processing. The Hankel matrix is ​​constructed using a forward-backward approach to improve the accuracy of parameter estimation. The system order is determined using a stability graph method, selecting the optimal order based on the stability of modal parameters at different orders. Confidence intervals for modal parameter estimation are calculated using the covariance matrix, providing an uncertainty measure for each modal frequency and damping ratio. The Paris law model parameters in the damage evolution model are calibrated through fatigue testing, and the corrosion kinetics model includes environmental corrosion factors and material corrosion resistance parameters. The stiffness degradation model considers the combined effects of bolt preload loss and contact surface wear. The update trigger condition for linear fitting coefficients is set to the addition of new damage records or a significant change in the external excitation mode. Real-time calculation of the dynamic correction factor uses a streaming processing architecture, processing continuously arriving excitation data through window functions. The temporal resolution of the factor output is synchronized with the modal parameter identification, ensuring time alignment between the two evaluation metrics.

[0082] See Figure 4In the modal parameter identification and structural health assessment of wind turbine tower health monitoring, the modal correlation between mode pairs was quantified using the Modal Assurance Criterion (MAC). The horizontal axis represents the MAC value, which measures the similarity between the mode shape vectors of two modes; a value closer to 1 indicates a higher correlation. The figure sets low correlation thresholds (0.8, red dashed line) and high correlation thresholds (0.9, green dashed line) as criteria for determining modal correlation. The vertical axis represents mode pairs, covering pairwise combinations between the main vibration modes of the tower (modes 1-6). In terms of color and length, green bars represent highly correlated mode pairs (such as mode 1-mode 5, mode 1-mode 6), with a criterion value exceeding 0.9, indicating high mode similarity and strong consistency in modal characteristics; blue bars have criterion values ​​between 0.8 and 0.9, indicating moderate correlation; red bars are below 0.8, indicating low mode correlation and significant differences in modal characteristics. This provides a direct basis for the quantitative scoring of mode correlation in the first evaluation index. Combined with indicators such as frequency offset rate and energy dissipation efficiency, a weighted scoring matrix for the health of the tower structure can be constructed, thereby supporting the classification of comprehensive health levels and the generation of real-time monitoring reports.

[0083] Example 3: The comprehensive scoring model generates a quantitative assessment value of the tower's health status by fusing data from the first and second assessment indicators. The weighted scoring matrix of the first assessment indicator originates from the modal parameter identification and weighted scoring process, while the dynamic correction factor of the second assessment indicator originates from the damage evolution analysis process. The normalization of the weighted scoring matrix employs a minimum-maximum scaling method, linearly transforming each element in the matrix to a closed interval between zero and one. The normalization parameters are dynamically calculated based on the scoring data of the most recent thirty natural days, specifically by automatically collecting all data from the weighted scoring matrix elements over the past thirty days at midnight each day, identifying the minimum and maximum values ​​as the normalization benchmark for that day. The calculation of the structural health sub-score is achieved through a weighted summation formula, which expresses the quantitative method for the structural health sub-score:

[0084]

[0085] Where: symbol The structural health sub-score represents the relative health status of the tower structure; a higher value indicates a better structural condition. (Symbol: ) This represents the weight coefficient corresponding to the element in the i-th row and j-th column of the weighted scoring matrix. The weight coefficients are determined by the analytic hierarchy process and satisfy the normalization condition, i.e., the sum of all weight coefficients is one; (symbol) The elements of the normalized scoring matrix are represented, with values ​​ranging from zero to one, obtained from the original scoring values ​​through linear transformation. The subscript *i* ranges from one to six, corresponding to the six main vibration modes of the tower, including the first-order bending mode, the second-order bending mode, and the first-order torsional mode. The subscript *j* ranges from one to four, corresponding to the four scoring dimensions: frequency score, damping score, mode shape score, and comprehensive score. Each dimension reflects different aspects of the tower's dynamic characteristics. The fusion of the dynamic correction factor and the structural health sub-score employs a sliding window convolution operation, with a window width set to twenty-four time steps, each time step corresponding to one hour of monitoring data. The convolution kernel coefficients are generated using a Gaussian function, with a standard deviation of six hours to emphasize the influence of recent data. The specific implementation of the convolution operation involves discretely convolving the time series of the structural health sub-score with the Gaussian kernel. The initial value of the comprehensive score is calculated in the time domain, outputting one score value at each time step. The calculation process involves matrix operations and vector dot products. The tower's design life parameters are obtained from the tower's technical specifications, including a design service life of 25 years, a fatigue life of 100 million cycles, and a material aging coefficient of 0.001 per year. The time decay coefficient of real-time monitoring data is calculated based on the data acquisition timestamp, using an exponential decay model to give higher weight to more recent data. The decay constant is dynamically adjusted according to data quality indicators, including signal-to-noise ratio, data integrity, and sensor health status.

[0086] The dynamic adjustment of the weight allocation ratio adopts an adaptive filtering algorithm. A weight update function is constructed based on the design life parameters and the time decay coefficient of real-time monitoring data. This weight update function is a multivariate function; the inputs are the remaining design life percentage and the time decay coefficient, and the output is the weight allocation ratio between the first and second evaluation indicators. In the comprehensive scoring model, the first evaluation indicator ( ) and the second evaluation indicator (dynamic adjustment factor) The fusion weights of the two indicators were determined using the Analytic Hierarchy Process (AHP) combined with expert scoring to construct a judgment matrix. Five domain experts (three wind power structural engineers and two monitoring technology experts) were invited to perform a pairwise comparison of the two indicators based on the importance of the current structural state and the importance of the damage evolution trend. The judgment matrix was constructed using a 1-9 scaling method. (This indicates that the importance of the second evaluation indicator is twice that of the first evaluation indicator). Calculate the largest eigenvalue of the judgment matrix. Consistency indicators (n=2), random consistency index Consistency ratio This satisfies the consistency requirement. The normalized eigenvectors of the judgment matrix are obtained using the eigenvector method, yielding the basic weights of the first evaluation index. The basic weight of the second evaluation indicator Combined with the service life of the tower (Unit: years, design life 25 years), adjustment factor introduced. ( The value range is 0-1), and the final weight is:

[0087]

[0088]

[0089] in, As the basic weight of the first evaluation indicator, As the basic weight of the second evaluation indicator, To adjust the coefficients, for example: when (New tower) , (Severe injury trend); when At the end of the design life (mid-life), , (Balancing the current state with the trend); when At the end of the design life, , (Revised current state). The weight allocation ratio is recalculated every 24 hours, and the calculation process is executed in a scheduled task in the server background. The output frequency of the comprehensive scoring model is consistent with the data collection frequency, generating a new comprehensive score value every five minutes. The score value is stored in a time series database for subsequent query and analysis. The normalization process of the weighted scoring matrix includes an outlier handling mechanism. When matrix elements exceed the normal range, the Wensoritz method is used for truncation, with a truncation ratio set to 5%. That is, after all data are sorted by size, the data at both ends of the range are replaced with the 95th and 5th quantiles. The calculation of the structural health sub-score introduces a modal confidence factor. Modal scores with low confidence are weighted down. The confidence factor is calculated using the modal amplitude coherence coefficient. The confidence factor of modes with a coherence coefficient lower than 0.8 decreases linearly. The boundary processing of the convolution operation adopts a symmetrical padding method to avoid information loss at both ends of the sequence. The padding length is set to half the window width, i.e., twelve time steps. The padding data is obtained by mirroring the boundary data.

[0090] The standardization of design life parameters converts life parameters in different units into dimensionless life indices. The calculation of the life index considers material fatigue characteristics, environmental corrosion factors, and load history, employing a linear weighted method to synthesize a comprehensive life index. The weighting coefficients are determined through expert scoring. The time decay coefficient is calculated using an exponential decay model, with the decay constant dynamically adjusted based on data reliability indicators, including the signal-to-noise ratio (SNR) and data integrity. The decay constant increases accordingly when the SNR exceeds 20 decibels. The weight allocation ratio is optimized using gradient descent, with the objective function being minimizing the score prediction error. The learning rate is set to 0.01, and the maximum number of iterations is 1000. The real-time implementation of the comprehensive scoring model utilizes a sliding window mechanism with a window length of 720 sampling points, corresponding to 24 hours of data. Upon arrival of each new sampling point, the window slides forward one position, and the comprehensive score is recalculated. The calculation process employs an incremental update algorithm to avoid full recalculation. Model parameters are stored in persistent storage, supporting parameter recovery after system restarts. Parameter backups utilize dual storage media for redundancy, with daily parameter snapshots.

[0091] The normalization parameters of the weighted scoring matrix are calculated using dynamic statistics, with the minimum and maximum values ​​calculated based on the scoring data from the most recent thirty days. The statistical window is updated daily via sliding updates, performed during periods of low system load. Smoothing of the structural health sub-score employs Kalman filtering to eliminate random fluctuations caused by measurement noise. The state transition matrix is ​​designed based on the tower's dynamic characteristics, and the process noise covariance and observation noise covariance are calibrated using historical data. The convolution kernel coefficients are determined using frequency domain analysis to highlight the periodic changes in health status. Frequency domain analysis is implemented using Fast Fourier Transform (FFT) with a transform length of 1024 points. The time update of the design life parameters considers the actual usage of the tower. When the tower undergoes major repairs or modifications, the design life parameters are adjusted accordingly, with the adjustment range determined based on component replacement information in the maintenance records. Replacement of major load-bearing components extends the design service life. The adaptive adjustment of the time decay coefficient uses fuzzy logic control, automatically adjusting the decay rate based on data quality indicators. The fuzzy rule base contains twenty control rules, with the input variable being the data quality level and the output variable being the adjustment amount of the decay constant. The weight allocation ratio is constrained to ensure a balance between short-term changes and long-term trends. The upper limit of short-term weights is set to twice that of long-term weights to prevent the model from overemphasizing short-term fluctuations. These constraints are implemented during optimization using the Lagrange multiplier method. The comprehensive scoring model is validated using cross-validation, dividing historical data into training and test sets, with the training set comprising 70% and the test set 30%. Stratified sampling is used to ensure data distribution consistency. Model performance is evaluated using mean squared error (MSE) and coefficient of determination (COD) metrics. The MSE threshold is set to 0.05, and the COD threshold is set to 0.8. Failure to meet either metric triggers model retraining. The model update mechanism includes two modes: online learning and batch learning. Online learning processes real-time data streams and updates model parameters using stochastic gradient descent. Batch learning retrains model parameters weekly, using full data training to ensure model stability.

[0092] Example 4: Health level classification is based on the initial values ​​generated by the comprehensive scoring model. The health level interval classification is based on the tower structure safety factor theory and the dynamic characteristic failure threshold model. The structural safety factor is correlated as follows: the comprehensive score is positively correlated with the ratio of the allowable stress to the actual stress of the tower material (safety factor). Healthy (85-100 points) corresponds to a safety factor ≥1.8, meeting the requirements for long-term stable operation within the design life; Sub-healthy (70-84.9 points) corresponds to a safety factor of 1.5-1.79, requiring increased inspection frequency; Warning (60-69.9 points) corresponds to a safety factor of 1.2-1.49, indicating a risk of local stress exceeding the standard, requiring the development of a maintenance plan; Dangerous (0-59.9 points) corresponds to a safety factor <1.2, indicating that the structure is in a critical state and requires shutdown for maintenance. Dynamic characteristic failure threshold: the upper limit of the high-frequency abnormal component is set based on a 2% deviation threshold of the first-order natural frequency of the tower. The upper limit for the health level is -20dB (corresponding to a stiffness decrease of <2%); the upper limit for the sub-health level is -15dB (corresponding to a stiffness decrease of 2%-5%); the upper limit for the warning level is -10dB (corresponding to a stiffness decrease of 5%-10%); and the upper limit for the danger level is -5dB (corresponding to a stiffness decrease of >10%). The initial value of the comprehensive score is the output of the comprehensive score model. The grading threshold is set using a combination of dynamic threshold algorithm and static benchmark value. The health level range is divided into four levels, corresponding to healthy, sub-healthy, warning, and danger states, with each level range corresponding to a different maintenance response strategy. High-frequency abnormal components in the dynamic response feature set are extracted using a digital filter. The filter is designed as a bandpass filter with a passband frequency range of 10Hz to 50Hz. The amplitude detection of high-frequency abnormal components is implemented using software based on the peak hold circuit principle, recording the maximum amplitude value within each time window. The upper limit of the level range is adjusted based on the linear relationship between the aging coefficient of the tower material and the service life. When the basic upper limit value is determined, the new tower (service life) The basic upper limit of high-frequency abnormal components is set according to the modal frequency deviation threshold: basic upper limit of health level Basic upper limit of sub-health level Warning level base limit Basic upper limit of danger level Aging factor calculation: A linear aging model is used, and the aging factor formula is as follows:

[0093]

[0094] in, For the number of years of operation, Designed for a lifespan of 25 years, The value ranges from 0 to 1. A certain number of operating years. The corresponding level range upper limit 15dB represents the maximum adjustment range within the design life, corresponding to the decline in structural anti-interference capability due to material aging. The upper limit of the level range is set based on the tower structure's dynamic characteristics, and the upper limit value adjusts linearly with the tower's operating years. The triggering condition for the recalibration command includes the determination of the duration of amplitude exceeding the limit; the command is triggered when a high-frequency abnormal component continuously exceeds the upper limit value for a preset duration. The parameter update of the comprehensive scoring model adopts an incremental learning method, adjusting only the model parameters related to the high-frequency abnormal component. The health level range update of the real-time monitoring report adopts a transactional data update mechanism to ensure data consistency. The health level range is mapped to a predefined tower maintenance strategy code, which uses a three-digit coding system. The first digit indicates the urgency of maintenance, the second digit indicates the maintenance type, and the third digit indicates the maintenance scale. The temporal feature extraction of the structural vibration signal includes mean, variance, peak value, and waveform factor. The spatial distribution of environmental load data is represented using a polar coordinate system, recording the load magnitude and direction angle with the tower center as the origin. The standardized report template uses JSON data format and includes three parts: header area, data area, and metadata area. The tower's location identifier uses a globally unified encoding format with millisecond-level timestamp accuracy. Binary data stream compression employs the LZ77 algorithm, and a cyclic redundancy check (CRC) code is added to the compressed data stream. The distributed database uses a sharded storage architecture, partitioning data according to the tower's geographical location. Spectral analysis of high-frequency anomalous components uses Fast Fourier Transform (FFT), with a spectral resolution of 0.1 Hz. Priority management of recalibration commands employs a multi-level queue scheduling algorithm, allowing high-priority commands to interrupt the execution of lower-priority commands.

[0095] Referring to Table 1, the health level intervals are visualized using a heatmap, with different levels corresponding to different color shades. The generation frequency of real-time monitoring reports is synchronized with the data collection frequency, with a new report generated every five minutes. Report transmission employs an asynchronous message queue transmission mode to improve system throughput.

[0096] Table 1: Standards for Defining Health Grade Intervals

[0097]

[0098] The standardized report template's header area includes basic information such as tower number, report timestamp, and number of monitoring points. The data area stores the time-domain feature array of structural vibration signals, the spatial distribution matrix of environmental load data, and core data of the health level code. The metadata area records data quality indicators, sensor status, and auxiliary information for calculation parameters. The distributed database index optimization uses a B+ tree structure to improve historical report query efficiency. Data backup employs a multi-site active-active architecture to ensure data security and recoverability. Report integrity verification uses digital signature technology to prevent data tampering. Time-frequency analysis of high-frequency anomaly components uses wavelet transform to simultaneously acquire time and frequency domain features. The recalibration command execution process includes a parameter backup step and supports command rollback operations. Smooth transitions between health level ranges use a hysteresis threshold design to avoid frequent level switching. Maintenance strategy code parsing uses a lookup table approach, mapping digital codes to specific maintenance operation procedures. Multi-language support for real-time monitoring reports uses Unicode encoding to adapt to international deployment requirements. The report archiving strategy is based on a time sliding window, automatically cleaning up expired report data. The calculation of the time-domain features of structural vibration signals uses a streaming processing algorithm to update feature values ​​in real time. The spatial distribution visualization of environmental load data uses contour maps to intuitively display the load distribution. Version management of standardized report templates employs semantic version numbers to ensure backward compatibility. Encrypted transmission of binary data streams uses the AES-256 algorithm to ensure data transmission security. Distributed database query optimization utilizes parallel scanning technology to improve query performance for large datasets. Pattern recognition of high-frequency anomaly components uses a support vector machine classifier to automatically identify anomaly types. Conflict resolution for recalibration commands uses timestamp sorting to ensure command execution order. Adaptive adjustment of health level intervals considers seasonal factors, using different threshold parameters for different seasons. Digital signatures for real-time monitoring reports use asymmetric encryption technology and a public key infrastructure certificate system. Dimensionality reduction of the temporal features of structural vibration signals uses principal component analysis to reduce data storage space. Spatial distribution interpolation of environmental load data uses Kriging to improve the accuracy of the distribution map. Compression optimization of standardized report templates uses dictionary encoding to further improve the compression ratio. Tower location identifier resolution uses reverse geocoding technology to convert coordinates into geographic location descriptions. Timestamp synchronization calibration uses a network time protocol to ensure time consistency across multiple devices. Forward error correction (FEC) codes are used for binary data stream transmission verification to improve data transmission reliability. Distributed database transaction management employs a two-phase commit protocol to ensure data consistency. A multi-level threshold system is used for early warning of high-frequency anomalies, with different thresholds triggering different response levels. Recalibration command logging uses structured logs for easy problem tracking and analysis. Historical tracking of health level ranges utilizes data version control, supporting status replay at any point in time. Access control for real-time monitoring reports employs role-based access management, with different roles having different data access permissions.The report generation system employs redundant deployment in its fault-tolerant design, ensuring that a single node failure does not affect system operation. Cross-validation is used to verify the time-domain characteristics of structural vibration signals, ensuring accurate feature calculations. Spatial distribution standardization of environmental load data utilizes minimum-maximum normalization to eliminate dimensional influences. Standardized report template validation employs schema verification technology to guarantee data format compliance. Tower positioning identifier updates use an event-driven mechanism, automatically updating when the tower's position changes. Timestamp accuracy calibration uses hardware clock synchronization to reduce the impact of system clock drift. Binary data stream cache management uses a least recently used algorithm to improve data access efficiency. Distributed database performance monitoring uses real-time monitoring metrics, including query latency and throughput. Trend analysis of high-frequency anomaly components uses time-series prediction algorithms to predict anomaly development trends. Recalibration command execution monitoring uses a heartbeat detection mechanism to ensure complete command execution. The interactive display of health level ranges uses a web graphical interface, supporting multi-dimensional data drill-down. Real-time monitoring report export supports multiple formats, including PDF and Excel. System operation and maintenance management includes automatic alarm functionality to promptly detect and handle system anomalies. The temporal characteristics of structural vibration signals are stored in a columnar format to improve data analysis efficiency. Spatial distribution compression of environmental load data uses vector quantization technology to reduce storage space usage. The standardized report templates feature a plug-in architecture for easy feature expansion. Tower positioning identifier management uses a centralized registry center to ensure globally unique identifiers. Timestamp timezone processing uses the Coordinated Universal Time (UTC) standard to avoid timezone conversion errors. Flow control of binary data streams uses a token bucket algorithm to ensure network transmission stability. The distributed database expansion design uses a consistent hashing algorithm to support smooth expansion. Association rule mining is used for correlation analysis of high-frequency anomaly components to discover potential connections between anomalies. Priority adjustment of recalibration commands uses dynamic priority scheduling, automatically adjusting based on system load. Audit trails for health level ranges use operation logs to meet compliance requirements. Multi-tenancy support for real-time monitoring reports uses data isolation technology to ensure data security for different users. System performance optimization uses load balancing technology to improve system processing capacity.

[0099] Example 5: The feedback optimization mechanism trains the remaining life prediction model using historical monitoring data and dynamically adjusts the weights of evaluation indicators. Historical monitoring reports are retrieved periodically from a distributed database, with a retrieval cycle set to 24 hours, retrieving monitoring data from the most recent 365 days each time. Trend data extraction employs time series decomposition technology, breaking down health level data into three components: trend, period, and residual. The remaining life prediction model uses a gradient boosting decision tree algorithm, with input features including health level trends, environmental load statistics, and 38 feature variables from maintenance records. Taking the specific operation of tower #B27 in a wind farm as an example, the historical monitoring report contains 430 daily health records, and trend data extraction shows that the health level is slowly declining at a rate of 0.3% per month. The remaining life prediction model is trained using a 10-fold cross-validation method, and the model outputs a remaining life of eight years and seven months for tower #B27. The feedback channel transmits the prediction results to the comprehensive scoring model, which adjusts the scoring weights based on the predicted life. The error rate is calculated using the mean absolute percentage error formula, with the current period's error rate at 6.3%. The weight adjustment adopts an adaptive moment estimation algorithm, which optimizes the fusion coefficient of the first evaluation index and the second evaluation index with the objective function of minimizing the error rate.

[0100] The high-precision laser scanning device is automatically activated when the system detects a potential anomaly area. The determination of a potential anomaly area is based on a local deformation gradient threshold. The laser scanning device uses a phase-detection lidar system with a scanning frequency of 2.4 million points per second, achieving a point cloud density of 1,200 points per square meter. The 3D deformation data is aligned with the tower design model using a point cloud registration algorithm, with registration accuracy controlled within 0.05 millimeters. Taking the anomaly area at a height of 35 meters on tower #B27 as an example, the point cloud data acquired by the laser scan contains 3.2 million 3D coordinate points. Point cloud registration uses a normal distribution transformation algorithm, achieving alignment between the point cloud and the design model after 80 iterations. Deformation gradient calculation error correction uses radial basis function interpolation to unify the spatial coordinates of the laser scanning data and vibration sensor data. The corrected deformation gradient value shows the maximum deformation area gradient as 0.15 radians / meter, a decrease of 0.05 radians / meter compared to the original value. The real-time monitoring report's health level interval update uses a sliding window mechanism, with the window size set to seven days. The updated health level ranges are redefined by comparing and correcting the deformation gradient with the threshold range. In the #B27 tower case, the health level was adjusted from "Caution" to "Good," and the maintenance strategy code was changed accordingly from 201 to 102. The data preprocessing of historical monitoring reports includes a missing value imputation step, using spline interpolation to complete the data points missing due to system maintenance.

[0101] Seasonal adjustments to trend data employ the STL decomposition method to eliminate the impact of seasonal fluctuations on long-term trends. Feature engineering for the remaining life prediction model includes feature selection and transformation, using mutual information to select important features. Data transmission in the feedback channel utilizes a message queue telemetry transmission protocol to ensure timely delivery of prediction results to the comprehensive scoring model. Rolling calculation of the error rate uses an exponentially weighted moving average method, with recent errors receiving higher weights. Constraints on weight adjustments limit the magnitude of weight changes, with a single adjustment not exceeding 5%. The triggering logic of the laser scanning device includes a multi-condition judgment mechanism, requiring simultaneous satisfaction of deformation gradient exceeding limits and abnormal duration conditions. Point cloud data preprocessing includes outlier removal steps, using radius filtering to remove significant noise points. Coordinate transformation of the 3D deformation data employs the Bursa-Taylor seven-parameter transformation model to transform the scanning coordinate system to the tower's global coordinate system. Registration accuracy is evaluated using the Hausdorff distance index, with the maximum distance between the registered point cloud and the design model serving as the criterion for registration quality. Correction of deformation gradient calculation errors uses a particle filtering algorithm, fusing multi-source data to improve gradient calculation accuracy. The update strategy for health level intervals adopts a conservative approach: downgrades are implemented immediately, while upgrades require three consecutive days of meeting the conditions. Changes to maintenance strategy codes trigger the work order generation system, automatically creating corresponding maintenance work orders. Historical monitoring report storage is optimized using a Paquet columnar storage format to improve query efficiency for large datasets. Real-time trend data calculation utilizes the Apache Flink stream processing engine, supporting online updates of trend analysis results. The online learning of the remaining lifespan prediction model employs an incremental update mechanism, locally updating model parameters with each new batch of data. The feedback channel's fault tolerance design uses an exponential backoff retry mechanism, automatically resending data in case of transmission failure. Error rate monitoring is configured with multi-level alarm thresholds; manual intervention is triggered when the error rate exceeds 10%. The laser scanning device's scanning path planning uses a grid coverage algorithm to complete the scanning of the designated area using the shortest path.

[0102] Point cloud data compression and transmission employs an octree-based point cloud compression algorithm to reduce the amount of data transmitted over the network. Visualization of 3D deformation data utilizes point cloud slicing technology to display the tower deformation distribution in three dimensions. The registration algorithm is accelerated using a CUDA parallel computing architecture, leveraging a graphics processor to improve registration speed. Verification of deformation gradient calculation error correction employs leave-one-out cross-validation, dividing sensor data into training and test sets. Audit logs for health level interval updates record detailed reasons and data basis for each change. The mapping table for maintenance strategy codes is updated regularly to reflect the latest maintenance procedures and standards. The historical monitoring report retrieval interface supports multi-condition composite queries, allowing queries by time range, health level, and tower number. Anomaly detection for trend data uses a local outlier factor algorithm to automatically identify abnormal trend points. The interpretative analysis of the remaining life prediction model uses the LIME method to quantify the contribution of each feature to the prediction results. Data security in the feedback channel employs end-to-end encryption using a transport layer security protocol to prevent tampering of prediction results during transmission. The error rate trend analysis employs the Savitzky-Gore filtering method to smooth random fluctuations and highlight long-term changes. Historical tracking of weight adjustments utilizes a Git version control system to retain a complete record of each adjustment. The laser scanning device's status is monitored in real-time, automatically switching to backup equipment upon detecting anomalies. Point cloud data quality assessment uses a comprehensive evaluation system encompassing point cloud density and noise level. 3D deformation data storage employs a hierarchical detail structure, supporting rapid rendering and detailed analysis. Registration accuracy is manually verified periodically, with automatic registration results checked through manual annotation. Deformation gradient calculation error correction optimization uses a Bayesian optimization algorithm to automatically find the optimal combination of correction parameters. The health level interval update push mechanism uses a WebSocket publish-subscribe pattern, with relevant systems receiving level change notifications in real-time. Historical monitoring report data archiving uses a layered storage architecture, with recent hot data stored in high-speed storage and historical cold data transferred to an object storage system. The trend data prediction function integrates a long short-term memory network model, providing health level predictions for the next thirty days. The remaining lifespan prediction model is deployed using Docker containerization technology for rapid scaling and resource isolation.

[0103] The load balancing of the feedback channel employs a minimum connection count strategy, evenly distributing data transmission tasks across multiple server nodes. Root cause analysis of the error rate utilizes a random forest model to automatically analyze the primary causes of errors. Simulation testing of weight adjustments uses a historical data playback mechanism to verify the effectiveness of new weights before practical application. Calibration and maintenance of the laser scanning device are performed monthly to ensure scanning accuracy meets requirements. Point cloud data backup employs a three-copy storage strategy to prevent data loss. The 3D deformation data analysis tool provides various measurement functions, including deformation, curvature, and volume change measurements. The registration algorithm optimization utilizes fast point feature histogram feature matching for acceleration, first matching feature points and then fine-tuning the registration. Real-time deformation gradient calculation error correction achieves processing speeds of 200 measurement points per second through algorithm optimization. A rollback mechanism for health level interval updates ensures rapid recovery to the previous state in case of erroneous updates. Comprehensive monitoring of system operation status includes multi-dimensional monitoring indicators for hardware status, software performance, and data quality.

[0104] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0105] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind turbine tower health monitoring method, characterized by, include: Collect structural vibration signals and environmental load data of wind turbine towers, extract dynamic response characteristics of towers based on structural vibration signals, and establish a basic monitoring dataset of tower operation status by combining environmental load data; The basic monitoring dataset is decomposed in multiple dimensions to separate the natural frequency components, damping characteristic parameters and external excitation coupling components of the tower. Based on the correlation between the natural frequency components and damping characteristic parameters, a first assessment index for the structural health of the tower is constructed. The historical damage records and material degradation data of the tower are acquired, and the historical damage records are time-series aligned with the external excitation coupling components to generate a second evaluation index for the tower damage evolution. By integrating the first and second assessment indicators, a comprehensive scoring model for the health status of the tower is established. Based on the comprehensive scoring model, the health level of the tower is classified and a real-time monitoring report is output. The comprehensive scoring model for establishing the health status of the tower includes: The weighted scoring matrix of the first evaluation indicator is normalized to obtain the structural health sub-score; the dynamic correction factor of the second evaluation indicator is convolved with the structural health sub-score to generate the initial value of the comprehensive score; the weight allocation ratio of the comprehensive score is adjusted according to the tower's design life parameters and the time decay coefficient of real-time monitoring data. The health levels of the tower are classified as follows: Set a grading threshold for the comprehensive score, and divide the health level intervals according to the degree of deviation between the initial value and the threshold; extract high-frequency abnormal components from the dynamic response feature set, and trigger a health level recalibration command if the high-frequency abnormal components exceed the upper limit of the level interval; update the parameters of the comprehensive score model based on the recalibration command and regenerate the real-time monitoring report.

2. The wind tower health monitoring method of claim 1, wherein, The dynamic response characteristics of the extraction tower include: The energy distribution characteristics and phase shift of the structural vibration signal are extracted by performing a time-frequency joint transformation. The amplitude deviation of energy distribution characteristics is corrected by environmental load data, and the local deformation gradient of the tower is calculated by combining the phase offset. Based on the relationship between the local deformation gradient and the preset threshold, potential abnormal regions of the tower are marked and a dynamic response feature set is generated.

3. The wind tower monitoring method of claim 2, wherein, The first assessment indicator for the health of the constructed tower structure includes: Modal parameter identification is performed on the natural frequency components, and the frequency shift rate and mode shape correlation of each mode are calculated. The energy dissipation efficiency of the tower structure is quantified based on the attenuation relationship between damping characteristic parameters and frequency offset rate. By combining modal correlation and energy dissipation efficiency, a weighted scoring matrix for the first evaluation index is generated.

4. The wind tower monitoring method of claim 3, wherein, The second evaluation metric for the evolution of damage to the generation tower includes: Historical damage records are classified according to damage type and matched with the temporal distribution characteristics of external excitation coupling components; Analyze the propagation rate of similar damage under external excitation and calculate the linear fitting coefficient of the damage accumulation effect; The dynamic correction factor for the second evaluation index is generated based on the product of the linear fitting coefficient and the current external excitation coupling component.

5. The wind tower monitoring method of claim 4, wherein, The output real-time monitoring report includes: Map health level ranges to predefined tower maintenance strategy codes; Integrate the temporal characteristics of structural vibration signals, the spatial distribution of environmental load data, and maintenance strategy codes to generate standardized report templates; Based on the tower's location identifier and timestamp, the standardized report template is compressed into a binary data stream and stored in a distributed database.

6. The wind tower monitoring method of claim 5, wherein, Also includes: Regularly retrieve historical monitoring reports from a distributed database to extract data on the changing trends of health levels; A prediction model for the remaining life of the tower is trained using trend data, and the life prediction results are output to the feedback channel of the comprehensive scoring model. The fusion weights of the first and second evaluation indicators are dynamically adjusted based on the error rate of the feedback channel.

7. The wind tower monitoring method of claim 6, wherein, Also includes: When a potential abnormal area in the tower is detected, a high-precision laser scanning device is activated to acquire three-dimensional deformation data of the tower surface; Spatial registration is performed between 3D deformation data and dynamic response feature set to correct the calculation error of local deformation gradient; The health level range of the real-time monitoring report is updated based on the corrected deformation gradient.

8. A wind turbine tower health monitoring system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the wind turbine tower health monitoring method as described in any one of claims 1 to 7.

Images