A wind power bolt loosening passive monitoring method based on Gaussian process compensation

Abstract

The application discloses a wind power bolt loosening passive monitoring method based on Gaussian process compensation and belongs to the technical field of wind power structure health monitoring. The method acquires multi-source sensing response data through a passive wireless sensing unit; performs joint time-frequency domain filtering processing on the acquired data; synchronously collects environment parameter time series data, constructs a Gaussian process regression model based on the environment parameters, predicts and reversely corrects environment drift components in the filtered characteristic sequence, and acquires bolt state characteristic parameters after environment decoupling; inputs the bolt state characteristic parameters after environment decoupling into a bolt loosening recognition model of a deep forest framework, and outputs loosening degree quantitative indexes and loosening risk grades; when the loosening degree quantitative indexes exceed preset safety threshold values or the loosening risk grades reach early warning grades, early warning information is output. The application realizes passive wireless high-reliability online monitoring of wind power tower drum bolt loosening.

Description

Technical Field

[0001] This invention relates to the field of wind power structural health monitoring technology, and in particular to a passive monitoring method for loosening wind power bolts based on Gaussian process compensation. Background Technology

[0002] With the deepening implementation of the "dual carbon" goals, wind power, as a clean energy source, has experienced rapid development, with continuous growth in installed capacity, unit capacity, and tower height. As a key supporting structure bearing the entire load of the wind turbine, the operational reliability of the wind turbine tower directly affects the safety and service life of the entire unit. High-strength bolts and flanges are typically used to connect the various sections of the tower, as well as between the tower and the foundation and nacelle. Maintaining the preload of these bolts is fundamental to ensuring the overall structural rigidity, fatigue resistance, and overturning resistance of the tower.

[0003] During long-term service, wind turbine tower bolts are subjected to multiple coupled effects, including alternating wind loads, start-stop impact loads, temperature cycles, and structural vibrations. The preload inevitably weakens, leading to bolt loosening. Especially under harsh environmental conditions such as strong winds, large temperature differences (annual temperature differences exceeding 80°C), high humidity (relative humidity varying from 20% to 95%), and rain and snow erosion, the rate of performance degradation of bolt connections accelerates significantly. Bolt loosening causes reduced connection stiffness, localized stress concentration, and changes in structural vibration modes, which can induce fatigue crack propagation and, in severe cases, lead to tower collapse, posing a significant safety hazard. Therefore, timely, accurate, and long-term effective monitoring of the loosening status of wind turbine tower bolts is an urgent engineering requirement for ensuring the safe operation of wind farms and reducing maintenance costs.

[0004] Currently, the inspection of the condition of wind turbine tower bolts in engineering practice mainly relies on regular manual inspections, using torque wrenches for spot checks and re-tightening. This method has the following inherent defects: (1) The inspection efficiency is extremely low. For towers hundreds of meters high and dozens of flange connection points, manual climbing is time-consuming and laborious, and is severely limited by weather conditions; (2) The inspection cycle is long, making it impossible to achieve continuous online monitoring and difficult to capture the early slow evolution process and sudden loosening of bolts; (3) The inspection results depend on the operator's experience and the accuracy of the torque wrench, which is highly subjective and has a high risk of misjudgment and omission; (4) High-altitude operations pose significant personal safety risks. As wind farms develop towards unmanned and intelligent operation and maintenance, the traditional manual mode can no longer meet the needs of refined management.

[0005] In recent years, researchers have attempted to monitor bolt condition using automated sensing technologies such as resistance strain gauges, piezoresistive impedance methods, and acoustic emission methods. While these methods can achieve a certain degree of quantitative characterization of the bolt condition, they all require wired power supply and signal transmission, leading to complex on-site wiring, difficult installation and maintenance, poor long-term reliability, and the cables inside the tower are susceptible to vibration wear and electromagnetic interference, affecting data quality. Although visual image-based bolt loosening detection methods have achieved wireless operation, they are heavily dependent on lighting conditions and image clarity, making it impossible to work stably in all-weather scenarios such as at night, in rain, and fog, and it is difficult to quantify minute loosening angles.

[0006] Passive wireless sensing technology based on radio frequency identification (RFID) offers a new approach to bolt condition monitoring due to its advantages such as no need for battery power, non-contact reading, and low cost. However, existing RFID bolt monitoring solutions still face the following key technical bottlenecks: First, the strain change caused by loosening of bolts is extremely small (usually in the range of 0 to 3000 micro-strains). Conventional RFID tag antennas are not sensitive enough to respond to deformation of this magnitude. The signal characteristics are weak and easily drowned out by the noise of the reading device itself and the background noise of the environment, resulting in an extremely low signal-to-noise ratio and making it difficult to achieve stable sensing.

[0007] Secondly, the wind turbine tower is a fully enclosed metal structure with a small internal space. The metal walls generate strong reflection, coupling, and multipath effects on electromagnetic waves, resulting in severe distortion of the tag's backscattered signal. Traditional single RFID reading methods cannot extract reliable strain information from it, and the identification accuracy drops significantly.

[0008] Third, the tower is exposed to complex meteorological environments for extended periods, and the drastic changes in temperature, humidity, and wind-induced vibration acceleration cause significant drift in the electromagnetic parameters of the tag antenna and patch antenna. Actual measurement data shows that this environmental drift often exceeds the signal change caused by bolt loosening itself. However, there is currently a lack of effective environmental decoupling methods, resulting in a persistently high false alarm rate in monitoring results, which severely restricts the engineering application of the technology.

[0009] Fourth, most existing research remains under ideal laboratory conditions and has not formed a complete signal processing and intelligent recognition technology system for the complex working conditions of real wind farms. In particular, it lacks a comprehensive algorithm framework that can simultaneously handle non-stationary noise, multipath interference and environmental drift, resulting in poor engineering applicability.

[0010] To address the aforementioned issues, this invention proposes a passive monitoring method for loose wind turbine bolts based on Gaussian process compensation. By constructing a Gaussian process regression model containing a combined covariance function with squared exponential and periodic terms, it achieves accurate prediction and dynamic compensation for environmental drift. Furthermore, it combines a deep forest architecture for intelligent identification of loosening status, effectively overcoming the shortcomings of existing technologies. Summary of the Invention

[0011] The purpose of this invention is to provide a passive monitoring method for loosening wind turbine bolts based on Gaussian process compensation. This method solves the technical problems of severe sensor signal drift caused by environmental factors, the inability of traditional filtering methods to effectively separate environmental interference from bolt status information, and high false alarm rate in existing passive wireless bolt monitoring methods. It achieves long-term stable and high-precision online monitoring of bolt loosening status under complex meteorological conditions.

[0012] To achieve the above objectives, this invention provides a passive monitoring method for loosening wind turbine bolts based on Gaussian process compensation, comprising the following steps: S1: Data acquisition step, acquiring multi-source sensor response data through passive wireless sensing units deployed on the flange connecting bolts of the wind turbine tower. The multi-source sensor response data includes the backscatter signal parameters of the UHF RFID tag and the electromagnetic scattering echo parameters of the radar-patch antenna unit. The backscatter signal parameters include at least the received signal strength indication value and the phase difference value. The electromagnetic scattering echo parameters include at least the resonant frequency offset, the echo amplitude and the phase spectrum characteristics. S2: Signal preprocessing step, performing joint time-frequency domain filtering on the acquired multi-source sensor response data to eliminate multipath reflection interference and background clutter noise generated by the metal structure of the wind turbine tower, and extracting effective electromagnetic scattering feature sequences; S3: Gaussian process compensation step, synchronously collecting time-series data of environmental parameters at the wind turbine tower service site. This time-series data includes at least temperature, relative humidity, and tower vibration acceleration. Based on the environmental parameter time-series data, a Gaussian process regression model is constructed. This model is used to predict and reverse correct the environmental drift component in the effective electromagnetic scattering characteristic sequence, obtaining the bolt state characteristic parameters after environmental decoupling. The covariance function of this Gaussian process regression model adopts a weighted combination of the squared exponential covariance function and the periodic covariance function. S4: Loosening identification step, inputting the bolt state feature parameters after environmental decoupling into the pre-built bolt loosening identification model for processing, outputting the quantitative index of the loosening degree and loosening risk level of each monitored bolt. This bolt loosening identification model is an intelligent discrimination model built on a multi-task deep forest architecture. This intelligent discrimination model includes cascaded multi-granularity scanning layers and cascaded forest layers, and introduces an adaptive feature selection gating mechanism in the cascaded forest layers; S5: Early warning step. When the quantitative indicator of loosening exceeds the preset safety threshold or the loosening risk level reaches the early warning level, an early warning message containing the loose bolt number, loosening degree and recommended maintenance measures will be output.

[0013] Preferably, the joint time-frequency domain filtering process in step S2 includes: S2-1: Perform empirical mode decomposition on the multi-source sensor response data to obtain several intrinsic mode function components. Determine the boundary order between the effective components and noise components based on the continuous mean square error criterion. Discard the high-frequency noise components and retain the effective intrinsic mode function components for reconstruction to obtain the time-domain filtered signal. S2-2: Perform a short-time Fourier transform on the time-domain filtered signal to obtain the time-frequency spectrum. In this time-frequency spectrum, identify the characteristic distribution area of ​​multipath reflection interference generated by the metal structure of the wind turbine tower in the time-frequency domain, and use the time-frequency masking method to suppress the interference characteristic area. S2-3: Reconstruct the signal after time-frequency masking by performing inverse short-time Fourier transform to obtain the effective electromagnetic scattering characteristic sequence after eliminating multipath reflection interference and background clutter noise.

[0014] Preferably, the specific method for determining the boundary order between the effective component and the noise component based on the continuous mean square error criterion in step S2-1 is as follows: Sort the intrinsic mode function components in descending order of frequency, and assume that the first... The reconstructed signal of each high-frequency component is Calculate its continuous mean square error : ; when When the global minimum first appears, the corresponding The value represents the dividing order; discard the previous value. One high-frequency component is retained, and the remaining low-frequency components are used as effective components for reconstruction.

[0015] Preferably, the method for constructing the Gaussian process regression model in step S3 includes: S3-1: Based on temperature relative humidity and tower vibration acceleration Constructing a three-dimensional environment state vector The environmental parameters of all historical moments are mapped to the three-dimensional environmental state space to form a sequence of state trajectory points. S3-2: Establish sensor response drift With this three-dimensional environment state vector Gaussian process regression model between: ; in, To obey the nonlinear function of the Gaussian process prior, , It is a mean function. Let covariance function be used. The observed noise term follows a zero-mean Gaussian distribution, i.e. ; S3-3: The covariance function The weighted combination of the squared exponential covariance function and the periodic covariance function is adopted: ; in, The signal variance is the squared exponential term. For the first The feature length scale parameter of the dimensional environmental features The variance of the periodic term, For periodic parameters, For the length scale parameter of the periodic term, For the Kronecker function, This is the noise variance parameter; S3-4: Given historical observation data Under these conditions, the hyperparameter set of the Gaussian process regression model is learned by maximizing the marginal log-likelihood function. ; S3-5: For the current environment state vector Based on the learned Gaussian process regression model, the posterior mean of the sensor response drift at the current moment is predicted. and posterior variance ,in The covariance matrix among historical observation samples, Let the covariance vector between the current environmental state vector and each historical observation sample be the posterior mean. The effective electromagnetic scattering feature sequence is reverse-corrected, and the confidence level after correction is evaluated using the posterior variance.

[0016] Preferably, in step S3-1, the three-dimensional environment state vector Temperature in relative humidity and tower vibration acceleration The sampling frequency is not less than 10 Hz, and the sampling frequency of the multi-source sensor response data in step S1 is not less than 100 Hz. The environmental parameter time series data and the multi-source sensor response data are aligned on the timestamp to form a synchronous observation sample.

[0017] Preferably, the marginal log-likelihood function in step S3-4 is: ; in, This is the vector of historical drift observations. for The covariance matrix, whose elements are determined by the covariance function in step S3-3; This set of hyperparameters The marginal log-likelihood function is obtained by maximizing the conjugate gradient method.

[0018] Preferably, the posterior mean in steps S3-5 and the posterior variance The calculation formula is: ; ; in, for A column vector of dimension, whose ? The elements are ; The specific method of this reverse correction is as follows: Starting from the effective electromagnetic scattering characteristic value at the current moment... Subtract the posterior mean from the middle The bolt state characteristic parameters after environmental decoupling are obtained. .

[0019] Preferably, the method for constructing the bolt loosening identification model based on a multi-task deep forest architecture in step S4 includes: S4-1: Construct a training dataset, which contains multi-source sensor response samples and their corresponding bolt status labels collected under different preload states and environmental conditions. S4-2: Multi-granularity scanning windows are used to extract features from the training data. At least three different scales of sliding windows are set to capture short-term abrupt changes, medium-term gradual changes, and long-term trend features during the bolt loosening process, respectively. The feature vectors generated by each scale window are pooled and then concatenated to form the input feature vector of the cascaded forest layer. ; S4-3: The cascaded forest layer adopts a multi-layer cascaded structure. Each layer of the cascaded forest contains several random forests and several completely random forests. Each forest outputs a probability estimation vector of the category. This probability estimation vector is concatenated with the original input features of the cascaded forest of that layer and then passed to the next layer of the cascaded forest. S4-4: An adaptive feature selection gating mechanism is introduced into this cascaded forest layer. This gating mechanism assigns learnable weight coefficients to the output probability of each forest and dynamically adjusts the contribution of different forest outputs in the hierarchical transmission through attention calculation to suppress redundant feature transmission. S4-5: The model training adopts a layer-by-layer cascade growth strategy. After each new layer of cascaded forest is added, the overall performance is evaluated on the validation set. If the performance improvement is lower than the preset threshold, the layer growth is stopped, and the final bolt loosening recognition model is output.

[0020] Preferably, in step S4-4, the weight coefficients of each forest output probability in the adaptive feature selection gating mechanism are calculated in the following way: The first in the current hierarchical forest The output probability vector of each forest The concatenated feature vector passed from the previous cascaded forest Perform feature interaction and calculate the first Attention score for each forest: ; in, The weight matrix is ​​a learnable matrix. For learnable attention vectors, For bias terms, This represents a vector concatenation operation; The attention scores of all forests are normalized to obtain the normalized weight coefficients for each forest: ; in, This represents the total number of forests in this cascaded forest hierarchy; The output probability vectors of each forest are weighted and summed using their normalized weight coefficients to obtain the weighted fused concatenated output: ; This Input features of the cascaded forest above The features are concatenated and used as input features for the next layer of the cascaded forest.

[0021] Preferably, the method for calculating the quantitative index of loosening degree in step S4 is as follows: Based on the bolt state characteristic parameters after environmental decoupling, the first strain-sensitive feature vector of the radio frequency identification channel and the second strain-sensitive feature vector of the radar channel are extracted respectively. The first strain-sensitive feature vector and the second strain-sensitive feature vector are input into a pre-trained strain regressor, which is built based on a support vector regression architecture, and outputs the current micro-strain estimate of the bolt. ; Compare the current estimated microstrain value with the reference microstrain value under the initial preload condition. Compare and calculate quantitative indicators of loosening degree : ; in, The preset limit micro-strain value is the ultimate micro-strain value corresponding to the complete loosening of the bolt, which is obtained through calibration tests. This quantitative indicator of looseness The value range is 0~1, according to Values ​​are used to classify the risk level of loosening: when This is normal, when It was slightly loose at the time, when At that time, it was moderately loosened. It was severely loose at the time.

[0022] Therefore, the passive monitoring method for loose wind turbine bolts based on Gaussian process compensation using the above-mentioned structure has the following beneficial effects: (1) The present invention adopts two passive sensing schemes: RFID and radar-patch antenna. It does not require batteries and signal cables, and overcomes the problems of complex installation, difficult maintenance and poor long-term reliability of traditional wired sensing. It is particularly suitable for engineering scenarios where the internal space of wind turbine towers is limited and high-altitude operations are difficult.

[0023] (2) This invention uses temperature, humidity, and vibration acceleration as three-dimensional environmental state vectors to construct a Gaussian process regression model with a covariance function combining squared exponential and periodic terms. This model can accurately capture the nonlinear influence of environmental factors on the sensing response and the diurnal periodicity. Through posterior mean back-correction, the signal fluctuation amplitude caused by environmental drift is reduced from ±18% to within ±4%, effectively solving the long-standing problem of environmental false alarms in this field.

[0024] (3) This invention utilizes two independent channels, namely RFID strain sensing and radar-patch antenna electromagnetic scattering sensing, and through joint time-frequency domain filtering, effectively suppresses multipath reflection of tower metal and background clutter interference. The two channels are redundant to each other, which improves signal quality and recognition robustness.

[0025] (4) Compared with deep neural networks, deep forests can still achieve good performance with limited sample size, and the cascaded structure can adaptively determine the model complexity. An adaptive feature selection gating mechanism (attention weighting) is introduced to dynamically filter important forest outputs, suppress the propagation of redundant features, and improve the model's generalization ability and computational efficiency.

[0026] (5) This invention is entirely based on commercial UHF RFID readers and FMCW radar platforms, with controllable costs; each step can be trained offline and executed quickly online, meeting real-time monitoring requirements. Field verification shows that the system has a detection rate of 94.7% for moderate to severe loosening, with a false alarm rate of only 3.2%, which can reduce manual inspection costs by more than 30% and improve maintenance efficiency by about 40%, showing broad prospects for promotion and application.

[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process of a passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to the present invention. Figure 2 This is a schematic diagram of the joint time-frequency domain filtering process in this invention; Figure 3 This is a schematic diagram of the Gaussian process compensation model in this invention, showing the construction of the environmental state space, the combination of covariance functions, and the drift prediction and correction process; Figure 4 This is a schematic diagram of the bolt loosening identification model based on deep forest architecture in this invention, including a multi-granularity scanning layer, a cascaded forest layer, and an adaptive gated attention module. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Unless otherwise defined, the technical or scientific terms used in this invention shall 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.

[0031] Example like Figure 1 As shown, this invention provides a passive monitoring method for loosening wind turbine bolts based on Gaussian process compensation, and the specific steps are as follows: S1. Sensor unit layout and data acquisition: The following example illustrates the process using a 2.5 MW wind turbine tower at an onshore wind farm. The tower is 90 m high and consists of three sections connected by flanges. M36 high-strength bolts are evenly distributed around the circumference of each flange. The flanges connecting the tower bottom to the foundation and the flanges in the middle section of the tower are selected as key monitoring targets, totaling 64 bolts.

[0032] An integrated passive wireless sensor patch is attached to the end of the screw (near the nut) of each bolt to be monitored. This patch contains a UHF RFID tag antenna (using a composite structure of a bent dipole coupled to the patch, with a center frequency of 915 MHz) and a radar patch antenna unit (with a resonant frequency designed to be 2.45 GHz). The patch uses a flexible polyimide substrate, 0.2 mm thick, and is fixed with epoxy resin adhesive. Before attachment, the screw end is cleaned: acetone is used to remove oil, sandpaper is used to increase surface roughness, and anhydrous ethanol is used for cleaning and drying.

[0033] Meanwhile, a miniature environmental monitoring node is installed on the inner wall of the tower 1.5 m away from the flange, which includes a PT100 temperature sensor (accuracy ±0.2 ℃), a humidity-sensitive capacitive sensor (accuracy ±1.5 %RH) and a piezoelectric accelerometer (sensitivity 100 mV / g).

[0034] An integrated reader / writer unit is installed on the side wall of the tower foundation. It contains a UHF RFID reader module (output power 30 dBm, gain 8 dBi circularly polarized antenna) and an FMCW radar module (sweep range 2.4 GHz–2.6 GHz, transmit power 10 dBm). The reader / writer communicates with the wind farm's SCADA system via Ethernet.

[0035] After the system is powered on, the initialization process is executed: the reader self-tests and calibrates the transmit power and receive gain; environmental nodes handshake to confirm that the data link is normal; the EPC codes of each sensor tag are read, and a bolt number-tag EPC mapping table is established. After initialization is completed, the reader performs data acquisition tasks according to a preset cycle (once every 10 minutes).

[0036] The specific data acquisition process is as follows: The RFID reader sends Query commands sequentially in a polling manner. Each tag responds 3 times, and the average value is taken. The RSSI value (unit: dBm, quantization accuracy: 0.5 dB) and phase difference value (unit: degrees, quantization accuracy: 0.1°) are recorded. The FMCW radar module transmits linear frequency modulated detection signals to each patch antenna element in sequence. The received echo signals are mixed to obtain intermediate frequency signals, sampled by a 12-bit ADC, and then sent to the FPGA for real-time FFT processing to extract the resonant frequency offset, echo amplitude, and phase spectrum characteristics. The environmental monitoring node synchronously collects temperature, humidity, and vibration acceleration data. All data is timestamped with GPS and then uploaded to the edge computing gateway.

[0037] The sampling frequency of the multi-source sensor response data was set to 200 Hz, and the sampling frequency of the environmental parameter time series data was set to 20 Hz. The environmental parameter time series data and the multi-source sensor response data were aligned on the timestamp to form a synchronous observation sample.

[0038] S2. Signal preprocessing and joint time-frequency domain filtering: like Figure 2 As shown, after the edge computing gateway receives the original data frame and completes the CRC check, it performs the joint time-frequency domain filtering process in step S2. This step is illustrated using the RSSI sequence of a certain RFID channel as an example; the processing methods for other channels (phase difference, radar amplitude, radar phase) are exactly the same.

[0039] S2-1, Empirical Mode Decomposition Filtering, assuming the original RSSI sequence is... , length is Points (corresponding to 10 seconds, sampling rate 200 Hz). Empirical mode decomposition (EMD) is performed on this sequence, yielding 8 intrinsic mode function components. And a residual term. The frequency distribution characteristics of each component are: These are high-frequency components (frequency > 10 Hz), mainly corresponding to electronic noise and rapid random fluctuations; These are low-frequency components (frequency < 10 Hz), mainly corresponding to bolt status information and slowly changing interference.

[0040] The continuous mean square error criterion is used to determine the boundary order between the effective component and the noise component. Assume... The reconstructed signal of each high-frequency component is Calculate its continuous mean square error: ; Where the definition When CMSE(k) first reaches its global minimum, the corresponding... The value represents the boundary order. In this embodiment, the calculated value is... It is the global minimum, therefore the boundary order is... ,give up ,reserve The time-domain filtered signal is obtained by reconstructing the signal with the residual term.

[0041] S2-2. Short-Time Fourier Transform and Time-Frequency Masking: A short-time Fourier transform is performed on the time-domain filtered signal to obtain the time-frequency spectrum. A Hanning window with a length of 128 points and an overlap rate of 50% is used. The characteristic distribution areas of multipath reflection interference generated by the metal structure of the wind turbine tower are identified in the time-frequency spectrum. Based on the pre-calibrated multipath characteristics of the tower metal (background spectrum collected under unloaded and bolt-free conditions), multipath reflection interference manifests as a persistent broad-spectrum energy band within the frequency range of 40–120 Hz, with an energy amplitude typically 15–20 dB higher than the background noise.

[0042] Constructing time-frequency masking templates ,when When falling into the interference feature region ,otherwise The time-frequency masking method is used to suppress interference feature regions. ; S2-3. Signal Reconstruction: The time-frequency masked signal is reconstructed using an inverse short-time Fourier transform to obtain the effective electromagnetic scattering characteristic sequence after eliminating multipath reflection interference and background clutter noise. After the above joint time-frequency domain filtering, the amplitude fluctuation range of the RSSI characteristic sequence is reduced from ±2.5 dB to ±0.6 dB, and the fluctuation range of the phase difference sequence is reduced from ±5° to ±1.2°.

[0043] S3. Construction and online calibration of the Gaussian process compensation model, as detailed below: like Figure 3 As shown in Figure S3-1, the environmental state space is constructed, and historical data (1008 samples) collected over 7 consecutive days is used as the training set. Temperature is used as the training factor. relative humidity and tower vibration acceleration Constructing a three-dimensional environment state vector In this embodiment, the temperature range is -8.5 ℃ to 36.2 ℃, the humidity range is 22.3 %RH to 92.7%RH, and the vibration acceleration range is 0.08 m / s². 2 ~4.7 m / s 2 .

[0044] Sensor response drift Defined as the characteristic value measured under the current environmental conditions compared to the reference environment ( , , The difference between the measured values ​​under the baseline condition. The sensor response value under the baseline condition is obtained through interpolation: samples near the baseline environment are selected from historical data (satisfied with...) , , ), take their average as .

[0045] For each dimension of the five-dimensional feature vector (RFID_RSSI, RFID_Phase, Radar_FreqOffset, Radar_Amplitude, Radar_Phase), the drift is calculated separately, and each channel is modeled independently.

[0046] S3-2. Construction of Gaussian process regression model, establishing sensor response drift. With the three-dimensional environment state vector Gaussian process regression model between: ; in , It is a mean function (set to zero mean). To observe the noise term, it follows the... .

[0047] The covariance function is a weighted combination of the squared exponential covariance function and the periodic covariance function: ; The first term is the squared exponential covariance function, used to capture the smooth and continuous influence of environmental factors on the sensing response; the second term is the periodic covariance function (periodic). The first term is set to 24 hours to capture diurnal periodic variations; the second term is the noise variance term.

[0048] S3-3, Hyperparameter learning, given historical observation data Under the given conditions, the hyperparameter set is learned by maximizing the marginal log-likelihood function. : ; in for covariance matrix The conjugate gradient method is used to optimize the likelihood function. Taking the RSSI channel as an example, the optimal hyperparameters learned are: , (Temperature characteristic length scale). (Humidity characteristic length scale) (Vibration acceleration characteristic length scale). , Minutes (approximately 24 hours, closely matching the day-night cycle). .

[0049] S3-4, Online drift prediction and reverse correction, for the current environmental state vector Based on the learned Gaussian process regression model, the posterior mean of the sensor response drift at the current moment is predicted: ; And posterior variance: ; in for A column vector of dimension, whose ? The elements are .

[0050] The specific method of reverse correction is as follows: Starting from the effective electromagnetic scattering characteristic value at the current moment... Subtract the posterior mean from the middle The bolt state characteristic parameters after environmental decoupling are obtained: ; The above correction process was performed on each dimension of the five-dimensional eigenvector. To verify the compensation effect, statistical analysis was performed on data from 30 consecutive days of operation: without compensation, the daily fluctuation of the RSSI eigenvalues ​​was approximately ±18%; after compensation, the daily fluctuation was reduced to within ±4%. Posterior variance. Simultaneously output for confidence assessment: when When this happens, the system automatically marks the confidence level of the correction result as low.

[0051] S4. Deep Forest Model Construction and Field Validation, as detailed below: like Figure 4 As shown in S4-1, the training dataset was constructed on a tower bolt simulation platform built in the laboratory. Standard M36 bolts (performance grade 10.9, rated preload 400 kN) were used as the test object. The preload was gradually unloaded from 100% of the rated value to 20% in 5% increments, for a total of 17 preload levels. Ambient temperature (-10℃, 0℃, 10℃, 20℃, 30℃) and relative humidity (30%, 50%, 70%, 90%) were controlled using a temperature and humidity chamber, resulting in 20 combinations of environmental conditions. Each condition was repeated 10 times, yielding 3400 laboratory samples. Simultaneously, 600 samples were obtained from the wind farm site. After merging, the samples underwent the same preprocessing and compensation procedures as in steps S2 and S3, resulting in 4000 valid samples. These were then divided into a training set (2800 samples), a validation set (400 samples), and a test set (800 samples) in a 7:1:2 ratio.

[0052] Each sample group has a 5-dimensional feature vector, and the labels include bolt micro-strain values. (Measured by strain gauges) and loosening level (four categories: normal, slightly loose, moderately loose, and severely loose).

[0053] S4-2, Multi-granularity scanning feature extraction: Three scale sliding windows (window sizes of 10, 20, and 40 sampling points respectively) are used to perform multi-granularity scanning on the training data. Taking a window as an example, the feature dimension within each window is... The 10-dimensional feature vector is generated after max pooling. The three scales are then concatenated to form the input feature vector of the cascaded forest layer. The dimension is 22.

[0054] S4-3, Cascaded Forest Layers and Adaptive Feature Selection Gating Mechanism: Each cascaded forest layer contains four random forests and four fully random forests, with each forest containing 100 decision trees. In each layer, each forest outputs a 4-dimensional class probability vector. .

[0055] In the adaptive feature selection gating mechanism, the attention score of each forest is calculated as follows: ; Normalized weighting coefficients: ; Weighted fusion output: ; Will and Concatenation serves as the input features for the next layer of the cascaded forest. Compared to traditional splicing methods, adaptive gating compresses the dimensional increment of each layer from 32 dimensions to 4 dimensions.

[0056] S4-4. Layer-by-layer cascaded growth: The model training adopts a layer-by-layer cascaded growth strategy: The initial 1-layer validation set accuracy is 83.2%, which increases to 86.5% after adding the 2nd layer (an improvement of 3.3%, which is greater than the threshold of 1%, so growth is accepted), and 87.1% after adding the 3rd layer (an improvement of 0.6%, which is lower than the threshold, so growth stops). Finally, a 2-layer cascaded structure is adopted.

[0057] S4-5. Quantitative Indicators of Loosening Degree and Risk Classification: For real-time input environmental decoupling characteristics, after multi-granularity scanning and cascaded forest forward propagation, the average output of all 800 regression trees is used as the micro-strain estimate. Calculate quantitative indicators of the degree of loosening: ; in (Rated preload condition). (Complete loosening limit). According to Risk levels are determined by the value: normal, Slightly loose, Moderately loose, It is severely loose.

[0058] S4-6. Field verification results: During three consecutive months of field operation, the system outputs the data for each bolt every 10 minutes. Value and risk level. During this period, manual torque sampling inspections (once every two weeks, with a sampling rate of 20%) will be arranged for comparative verification.

[0059] The results showed that: (1) the system detected 94.7% of moderate to severe loosening events (a total of 19 moderate to severe loosening events were detected, 18 of which were manually confirmed, and 1 was missed); (2) the false alarm rate was 3.2% (the system issued 25 warnings, of which 23 were manually confirmed and 1 was a false alarm); (3) the average advance warning time for slight loosening was about 7 days.

[0060] Following a period of sustained high winds (maximum wind speed 22 m / s), the system detected [something] on bolt number 5. The bolt preload suddenly increased from 0.21 to 0.58 (moderate looseness), triggering an orange alert. Maintenance personnel climbed the tower to inspect and found that the bolt preload had dropped from 390 kN to approximately 250 kN (a decrease of about 35%), consistent with the system monitoring results. It was promptly retightened. After retightening... The value was restored to 0.18.

[0061] Compared to adjacent units that did not have this system installed, the test group experienced zero unplanned shutdowns due to bolt issues within six months, while the control group experienced two. The frequency of manual inspections in the test group decreased from once per quarter to once every six months, resulting in a reduction of approximately 36% in labor costs for inspections.

[0062] S5, Extreme Operating Condition Verification: When the looseness quantification index output by step S4 When the system exceeds the preset safety threshold (in this embodiment, D_L ≥ 0.6, which means reaching moderate looseness or above) or the loosening risk level reaches the warning level (moderate or above), the system generates and outputs warning information.

[0063] The specific format of the warning information is: {timestamp} + {fan number} + {flange number} + {bolt number} + {current D_L value} + {risk level} + {recommended maintenance measures}. For example: "2026-03-15 14:32:18|WT-05|F2|B5|0.58|moderately loose|recommend re-tightening within 72 hours".

[0064] The warning information is output through the following two channels: (1) it is pushed to the monitoring screen of the wind farm control room via Ethernet, and is highlighted in the alarm list of the SCADA system, while triggering an audible and visual alarm; (2) it is pushed to the mobile APP of the maintenance team leader via 4G wireless router, and is notified in the form of pop-up window and SMS, so as to ensure that the maintenance personnel receive it in time.

[0065] The warning information is also stored in the local database and archived by bolt number and time index for use in historical trend analysis and operation and maintenance decision support.

[0066] In one actual early warning incident (the aforementioned bolt #5), the system issued an alert at 3:12 AM. The maintenance team leader received and confirmed the alert at 8:30 AM that same day, and completed the tower inspection and re-tightening at 10:00 AM the following morning. From the issuance of the alert to the closed-loop handling of the problem, the total time was approximately 31 hours, far shorter than the discovery cycle that might take weeks or even months under the traditional manual inspection model, effectively preventing the potential hazard from worsening.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A passive monitoring method for loosening wind turbine bolts based on Gaussian process compensation, characterized in that, Includes the following steps: S1: Data acquisition step, acquiring multi-source sensor response data through passive wireless sensing units deployed on the flange connecting bolts of the wind turbine tower. The multi-source sensor response data includes the backscatter signal parameters of the UHF RFID tag and the electromagnetic scattering echo parameters of the radar-patch antenna unit. The backscatter signal parameters include at least the received signal strength indication value and the phase difference value. The electromagnetic scattering echo parameters include at least the resonant frequency offset, the echo amplitude and the phase spectrum characteristics. S2: Signal preprocessing step, performing joint time-frequency domain filtering on the acquired multi-source sensor response data to eliminate multipath reflection interference and background clutter noise generated by the metal structure of the wind turbine tower, and extracting effective electromagnetic scattering feature sequences; S3: Gaussian process compensation step, synchronously collecting time-series data of environmental parameters at the wind turbine tower service site. This time-series data includes at least temperature, relative humidity, and tower vibration acceleration. Based on the environmental parameter time-series data, a Gaussian process regression model is constructed. This model is used to predict and reverse correct the environmental drift component in the effective electromagnetic scattering characteristic sequence, obtaining the bolt state characteristic parameters after environmental decoupling. The covariance function of this Gaussian process regression model adopts a weighted combination of the squared exponential covariance function and the periodic covariance function. S4: Loosening identification step, inputting the bolt state feature parameters after environmental decoupling into the pre-built bolt loosening identification model for processing, outputting the quantitative index of the loosening degree and loosening risk level of each monitored bolt. This bolt loosening identification model is an intelligent discrimination model built on a multi-task deep forest architecture. This intelligent discrimination model includes cascaded multi-granularity scanning layers and cascaded forest layers, and introduces an adaptive feature selection gating mechanism in the cascaded forest layers; S5: Early warning step. When the quantitative indicator of loosening exceeds the preset safety threshold or the loosening risk level reaches the early warning level, an early warning message containing the loose bolt number, loosening degree and recommended maintenance measures will be output.

2. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 1, characterized in that: The joint time-frequency domain filtering process in step S2 includes: S2-1: Perform empirical mode decomposition on the multi-source sensor response data to obtain several intrinsic mode function components. Determine the boundary order between the effective components and noise components based on the continuous mean square error criterion. Discard the high-frequency noise components and retain the effective intrinsic mode function components for reconstruction to obtain the time-domain filtered signal. S2-2: Perform a short-time Fourier transform on the time-domain filtered signal to obtain the time-frequency spectrum. In this time-frequency spectrum, identify the characteristic distribution area of ​​multipath reflection interference generated by the metal structure of the wind turbine tower in the time-frequency domain, and use the time-frequency masking method to suppress the interference characteristic area. S2-3: Reconstruct the signal after time-frequency masking by performing inverse short-time Fourier transform to obtain the effective electromagnetic scattering characteristic sequence after eliminating multipath reflection interference and background clutter noise.

3. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 2, characterized in that: The specific method for determining the boundary order between the effective component and the noise component based on the continuous mean square error criterion in step S2-1 is as follows: Sort the intrinsic mode function components in descending order of frequency, and assume the first... The reconstructed signal of each high-frequency component is Calculate its continuous mean square error : ； when When the global minimum first appears, the corresponding The value represents the dividing order; discard the previous value. One high-frequency component is retained, and the remaining low-frequency components are used as effective components for reconstruction.

4. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 1, characterized in that: The method for constructing the Gaussian process regression model in step S3 includes: S3-1: Based on temperature relative humidity and tower vibration acceleration Constructing a three-dimensional environment state vector The environmental parameters of all historical moments are mapped to the three-dimensional environmental state space to form a sequence of state trajectory points. S3-2: Establish sensor response drift With this three-dimensional environment state vector Gaussian process regression model between: ； in, To obey the nonlinear function of the Gaussian process prior, , It is a mean function. Let covariance function be used. The observed noise term follows a zero-mean Gaussian distribution, i.e. ; S3-3: The covariance function The weighted combination of the squared exponential covariance function and the periodic covariance function is adopted: ； in, The signal variance is the squared exponential term. For the first The feature length scale parameter of the dimensional environmental features The variance of the periodic term, For periodic parameters, For the length scale parameter of the periodic term, For the Kronecker function, This is the noise variance parameter; S3-4: Given historical observation data Under these conditions, the hyperparameter set of the Gaussian process regression model is learned by maximizing the marginal log-likelihood function. ; S3-5: For the current environment state vector Based on the learned Gaussian process regression model, the posterior mean of the sensor response drift at the current moment is predicted. and posterior variance ,in The covariance matrix among historical observation samples, Let the covariance vector between the current environmental state vector and each historical observation sample be the posterior mean. The effective electromagnetic scattering feature sequence is reverse-corrected, and the confidence level after correction is evaluated using the posterior variance.

5. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 4, characterized in that: The three-dimensional environment state vector in step S3-1 Temperature in relative humidity and tower vibration acceleration The sampling frequency is not less than 10 Hz, and the sampling frequency of the multi-source sensor response data in step S1 is not less than 100 Hz. The environmental parameter time series data and the multi-source sensor response data are aligned on the timestamp to form a synchronous observation sample.

6. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 4, characterized in that: The marginal log-likelihood function in step S3-4 is: ； in, This is the vector of historical drift observations. for The covariance matrix, whose elements are determined by the covariance function in step S3-3; This set of hyperparameters The marginal log-likelihood function is obtained by maximizing the conjugate gradient method.

7. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 4, characterized in that: The posterior mean in step S3-5 and the posterior variance The calculation formula is: ； ； in, for A column vector of dimension, whose ? The elements are ; The specific method of this reverse correction is as follows: Starting from the effective electromagnetic scattering characteristic value at the current moment... Subtract the posterior mean from the middle The bolt state characteristic parameters after environmental decoupling are obtained. .

8. The passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 1, characterized in that: The method for constructing the bolt loosening identification model based on a multi-task deep forest architecture in step S4 includes: S4-1: Construct a training dataset, which contains multi-source sensor response samples and their corresponding bolt status labels collected under different preload states and environmental conditions. S4-2: Multi-granularity scanning windows are used to extract features from the training data. At least three different scales of sliding windows are set to capture short-term abrupt changes, medium-term gradual changes, and long-term trend features during the bolt loosening process, respectively. The feature vectors generated by each scale window are pooled and then concatenated to form the input feature vector of the cascaded forest layer. ; S4-3: The cascaded forest layer adopts a multi-layer cascaded structure. Each layer of the cascaded forest contains several random forests and several completely random forests. Each forest outputs a probability estimation vector of the category. This probability estimation vector is concatenated with the original input features of the cascaded forest of that layer and then passed to the next layer of the cascaded forest. S4-4: An adaptive feature selection gating mechanism is introduced into this cascaded forest layer. This gating mechanism assigns learnable weight coefficients to the output probability of each forest and dynamically adjusts the contribution of different forest outputs in the hierarchical transmission through attention calculation, thereby suppressing redundant feature transmission. S4-5: The model training adopts a layer-by-layer cascade growth strategy. After each new layer of cascaded forest is added, the overall performance is evaluated on the validation set. If the performance improvement is lower than the preset threshold, the layer growth is stopped, and the final bolt loosening recognition model is output.

9. A passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 8, characterized in that: In step S4-4, the weight coefficients of each forest output probability in this adaptive feature selection gating mechanism are calculated in the following way: The first in the current hierarchical forest The output probability vector of each forest The concatenated feature vector passed from the previous cascaded forest Perform feature interaction and calculate the first Attention score for each forest: ； in, The weight matrix is ​​a learnable matrix. For learnable attention vectors, For bias terms, This represents a vector concatenation operation; The attention scores of all forests are normalized to obtain the normalized weight coefficients for each forest: ； in, This represents the total number of forests in this cascaded forest hierarchy; The output probability vectors of each forest are weighted and summed using their normalized weight coefficients to obtain the weighted fused concatenated output: ； This Input features of the cascaded forest above The features are concatenated and used as input features for the next layer of the cascaded forest.

10. A passive monitoring method for loose wind turbine bolts based on Gaussian process compensation according to claim 8, characterized in that: The calculation method for the quantitative index of loosening degree in step S4 is as follows: Based on the bolt state characteristic parameters after environmental decoupling, the first strain-sensitive feature vector of the radio frequency identification channel and the second strain-sensitive feature vector of the radar channel are extracted respectively. The first strain-sensitive feature vector and the second strain-sensitive feature vector are input into a pre-trained strain regressor, which is built based on a support vector regression architecture, and outputs the current micro-strain estimate of the bolt. ; Compare the current estimated microstrain value with the reference microstrain value under the initial preload condition. Compare and calculate quantitative indicators of loosening degree : ； in, The preset limit micro-strain value is the ultimate micro-strain value corresponding to the complete loosening of the bolt, which is obtained through calibration tests. This quantitative indicator of looseness The value range is 0~1, according to Values ​​are used to classify the risk level of loosening: when This is normal, when It was slightly loose at the time, when At that time, it was moderately loosened, when It was severely loose at the time.

Images