A micro-deformation and vibration integrated differential monitoring method for fan screws

Abstract

The present application relates to the technical field of wind power generation equipment state monitoring, and discloses a fan screw differential monitoring method combining micro deformation and vibration, which deploys distributed monitoring nodes on a fan base, collects strain signals and vibration signals based on a synchronous triggering mechanism, extracts deformation variable deviation values and frequency band characteristic components, generates a fusion feature vector by linear combination using a set of weight coefficients, calculates the difference between each node fusion feature vector and the global average feature value, compares the deformation variable deviation values and the difference with multi-level threshold values to determine the loosening state of single screws, extracts the range of each node difference value within the same period, and combines the overall imbalance limit coefficient to evaluate the stress imbalance degree of multiple screws on the base and issue an early warning. The present application uses the common mode rejection principle to strip common environmental interference under complex working conditions, improves the single-point loosening determination accuracy, and realizes early diagnosis of the stress distortion of the overall base structure.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment condition monitoring technology, specifically a differential monitoring method for wind turbine screws that integrates micro-deformation and vibration. Background Technology

[0002] Wind turbine generators operate in complex natural environments for extended periods, and the base connecting bolts bear enormous alternating loads. The tightness of these bolts directly affects the safety of the overall turbine structure. Currently, the industry typically uses strain gauges or vibration sensors to monitor the looseness of the turbine base bolts.

[0003] Existing monitoring methods largely rely on sensing single physical quantities. Under the high-frequency alternating loads generated by wind turbine operation and complex external environmental interference, single strain or vibration signals are easily drowned out by environmental noise. Due to the lack of multi-source data fusion and common-mode interference stripping mechanisms, monitoring systems struggle to effectively distinguish between localized loosening characteristics and global common environmental interference when faced with changes in the overall macroscopic operating conditions of the unit, leading to false alarms or missed alarms, and low accuracy in determining single-point loosening conditions.

[0004] Meanwhile, traditional monitoring logic is often limited to the independent evaluation of the characteristic data of a single screw, typically triggering an alarm only when the characteristic value of a single measuring point exceeds a preset threshold. This approach ignores the overall spatial stress state of the wind turbine base. In actual operating conditions, the wind turbine base often experiences asymmetrical torsional and overturning stresses. At this stage, even if the physical state of a single screw has not yet reached the independent single-point alarm threshold, stress distortion has already occurred in the multi-point fasteners of the base. Existing technology lacks a diagnostic mechanism for the spatial distribution characteristics of multiple nodes and the overall stress balance, making it impossible to achieve early assessment of potential structural imbalances.

[0005] Furthermore, wind turbines are typically deployed in outdoor environments with significant temperature variations, making sensor components susceptible to temperature drift and resulting in nonlinear bias errors. Additionally, due to the large size of the turbine base, existing distributed monitoring nodes generally experience time delays during data communication and acquisition, leading to a lack of strict temporal alignment in the data acquired by each node. This sampling time deviation, coupled with measurement errors caused by extreme high and low temperature environments, distorts the underlying source data, making it difficult to meet the requirements of high-precision spatial difference calculations for data time synchronization and numerical accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a differential monitoring method for wind turbine screws that integrates micro-deformation and vibration. This method solves the problems of existing single monitoring methods, such as blind spots, susceptibility to ambient temperature drift and macroscopic common-mode vibration interference from the wind turbine, which makes it difficult to accurately predict early micro-loosening of a single screw and hidden faults such as uneven force distribution among multiple screws in the base.

[0007] To achieve the above objectives, this invention provides a differential monitoring method for fan screws that integrates micro-deformation and vibration, comprising the following steps:

[0008] Based on multiple distributed monitoring nodes deployed diagonally on the base of the wind turbine, historical deformation data and historical vibration data are obtained by utilizing strain sensors and low-frequency vibration sensors integrated within the distributed monitoring nodes.

[0009] Based on historical deformation data and historical vibration data, the distributed monitoring nodes are controlled to enter the baseline learning mode to construct a multi-dimensional feature baseline library.

[0010] The current strain and vibration signals are acquired using a synchronous triggering mechanism;

[0011] The difference between the current deformation and the benchmark value in the multidimensional feature baseline library is calculated for the data corresponding to the strain signal to obtain the deformation deviation value, and spectral analysis is performed on the data corresponding to the vibration signal to extract frequency band feature components;

[0012] A fused feature vector is generated by linearly combining the deformation deviation value and frequency band feature components using set weighting coefficients.

[0013] Calculate the average eigenvalue of the fused feature vector of all distributed monitoring nodes, and calculate the difference between the fused feature vector of each distributed monitoring node and the average eigenvalue.

[0014] The deformation deviation value and difference value are compared with the set multi-level thresholds to determine the loosening status and severity level of a single screw. Based on the range of the difference values ​​at each node, the degree of uneven force on multiple screws of the wind turbine base is evaluated and an early warning message is output.

[0015] The process of controlling distributed monitoring nodes to enter baseline learning mode and constructing a multi-dimensional feature baseline library specifically includes: for historical deformation data acquired by each distributed monitoring node, performing a moving average calculation to obtain the average deformation baseline value, and simultaneously calculating the sample variance and standard deviation to construct the normal deformation range; for historical vibration data acquired by each distributed monitoring node, using the power spectrum estimation method to divide the continuous data sequence into sub-data segments and performing a fast Fourier transform operation to extract the normal vibration spectrum characteristics of the base under normal operating conditions; and combining the ambient temperature data synchronously recorded by each distributed monitoring node during baseline learning mode to establish a temperature drift curve, and structurally constructing the normal deformation range, normal vibration spectrum characteristics, and temperature drift curve into a multi-dimensional feature baseline library.

[0016] After acquiring the current strain and vibration signals using a synchronous triggering mechanism, the process further includes: sequentially filtering, amplifying, and performing analog-to-digital conversion on the strain and vibration signals to obtain the corresponding high-resolution digital sequences; combining the high-resolution digital sequences with the real-time ambient temperature data to execute a temperature compensation algorithm, specifically based on the calibration-obtained linear and nonlinear temperature compensation coefficients, using the difference between the currently acquired ambient temperature and the calibration reference temperature, and the square of the temperature difference term to perform polynomial fitting operations, eliminating nonlinear bias errors and temperature drift errors under extreme high and low temperature environments, and using this to obtain the true value of the deformation after temperature compensation correction.

[0017] The process of calculating the difference between the current deformation and the benchmark value in the multidimensional feature baseline library to obtain the deformation deviation value for the data corresponding to the strain signal specifically includes: smoothing the discrete measurement sequence corresponding to the strain signal using a sliding time window on the time domain scale, extracting the arithmetic mean of all discrete strain values ​​within the current window as the current deformation; calculating the difference between the current deformation and the average deformation benchmark value in the multidimensional feature baseline library, and performing absolute value calculation on the difference to eliminate the interference of tensile and compressive polarity of the force, as the deformation deviation value.

[0018] The specific steps for extracting frequency band feature components from the vibration signal data include: extracting the data sequence corresponding to the continuous vibration signal, applying a Hanning window for time-domain envelope shaping, and performing a fast Fourier transform to map it to the frequency domain; in the frequency domain, performing local extremum traversal within the ±0.5Hz bandwidth interval of each preset target frequency point associated with the loosening feature, and extracting the maximum amplitude within the interval as an independent feature parameter characterizing the frequency band energy state; merging the independent feature parameters of adjacent preset target frequency points using arithmetic averaging fusion logic to generate a comprehensive energy index reflecting the energy state within the frequency range of the screw loosening feature as the frequency band feature component; wherein, the preset target frequencies are 2Hz and 3Hz.

[0019] The method of generating a fused feature vector by linearly combining deformation deviation values ​​and frequency band feature components using set weighting coefficients specifically includes: normalizing the deformation deviation values ​​and frequency band feature components using preset benchmark values ​​to convert them into dimensionless relative deviation coefficients; linearly combining the normalized relative deviation coefficients using set weighting coefficients, including pre-set strain feature fusion weighting coefficients to reflect the dominance of deformation deviation and corresponding vibration feature fusion weighting coefficients; and using the reciprocal scaling property of the set weighting coefficients to unify the dimensions of physical parameters to generate a fused feature vector that represents a dimensionless scalar. The set weighting coefficients are values ​​obtained through training and optimization with experimental data, the preset benchmark values ​​are the deformation warning threshold and the energy benchmark value, and the strain feature fusion weighting coefficients and vibration feature fusion weighting coefficients are values ​​obtained through training and optimization with experimental data.

[0020] The calculation of the difference between the fused feature vector and the average feature value of each distributed monitoring node specifically includes: obtaining the fused feature vectors of all distributed monitoring nodes within the same synchronous time slice, calculating the arithmetic mean of the fused feature vectors as the average feature value representing the overall state of the base at the current moment; calculating the algebraic difference between the fused feature vector and the average feature value of each distributed monitoring node, separating the difference value that retains positive and negative polarity information; organizing the difference values ​​of each node into a multidimensional state vector, and mapping the multidimensional state vector to spatial state coordinate points in the multidimensional feature space, constructing a three-dimensional spatial difference model to represent the overall degree of force imbalance and abnormal physical location of the base.

[0021] The determination of a single screw's loosening status and severity level involves comparing the deformation deviation value and the difference value with a set multi-level threshold. Specifically, this includes: calculating the relative change ratio of the frequency band characteristic components relative to the normal energy benchmark in the multi-dimensional characteristic baseline library to obtain the energy deviation ratio of the current node; setting a multi-dimensional fault judgment benchmark for comparing the deformation deviation value and the difference value; the set multi-level thresholds include a yellow warning deformation threshold, a red warning deformation threshold, a yellow warning energy deviation ratio range, and a red warning energy deviation ratio limit, with the red warning deformation threshold being greater than the yellow warning deformation threshold; in the comprehensive grading assessment, by performing a comparison, when the determined deformation deviation value is greater than the value representing initial loosening... When a screw shows initial signs of loosening and triggers a single-point yellow warning command, it is determined that the deformation threshold for a yellow warning is less than or equal to the deformation threshold for a red warning indicating severe loosening, or the energy deviation ratio is within the set range for the yellow warning energy deviation ratio. When the deformation deviation value is greater than the deformation threshold for a red warning indicating severe loosening, or the energy deviation ratio exceeds the set range for the red warning energy deviation ratio, a screw is determined to be in a severely loose state and triggers a single-point red warning command. When the deformation deviation value is less than or equal to the deformation threshold for a yellow warning, and the energy deviation ratio is less than the lower limit of the set range for the yellow warning energy deviation ratio, a screw is determined to be in a normal state.

[0022] The multi-level thresholds are boundary parameters extracted and calibrated based on the statistical regularities of wind turbine bench destructive test data and historical real fault sample database. The deformation threshold for yellow warning is 10 micrometers, the deformation threshold for red warning is 15 micrometers, the energy deviation ratio range for yellow warning is 10% to 15%, and the energy deviation ratio limit for red warning is more than 15%.

[0023] The assessment of the uneven stress distribution of multiple screws on the wind turbine base based on the range of the difference values ​​of each node and the output of early warning information specifically includes: extracting the maximum and minimum values ​​from the fusion feature vector sets corresponding to each distributed monitoring node within the current synchronization cycle, and calculating the difference between the two as the range span reflecting the spread of the feature space, i.e., the range of the difference values ​​of each node; when the range span is determined to be greater than the product of the set overall imbalance limit coefficient and the average feature value, it is determined that the wind turbine base is experiencing asymmetrical abnormal torsional stress, causing stress distortion in multiple screws, triggering a multi-point imbalance alarm independent of the single-point early warning, and then generating an early warning data packet carrying the early warning level identifier, timestamp, abnormal physical location node number, and the underlying feature parameters of the triggered alarm transient as early warning information and pushing it to the monitoring terminal; when the range span is determined to be less than or equal to the product of the set overall imbalance limit coefficient and the average feature value, it is determined that the overall stress on multiple screws on the wind turbine base is in a balanced state; whereby the set overall imbalance limit coefficient is a proportional limit extracted and calibrated based on the statistical regularity of wind turbine bench destructive test data and historical real fault sample library.

[0024] The synchronous triggering mechanism for acquiring current strain and vibration signals specifically includes: generating a synchronous triggering command through an internal timer and sending it to all distributed monitoring nodes via a wireless communication link; upon receiving the synchronous triggering command, each distributed monitoring node synchronously initiates the data conversion process based on a triggering accuracy constraint better than 1 millisecond; during the data conversion process, the acquisition cycle of each distributed monitoring node is locked to a sampling rate benchmark of 100Hz, thereby accurately capturing the instantaneous force and vibration state of different physical locations of the base under the same macroscopic time slice and eliminating spatial difference calculation errors caused by time delay, thus obtaining strain and vibration signals that are strictly aligned in the time dimension.

[0025] This invention provides a differential monitoring method for fan screws that integrates micro-deformation and vibration. It has the following beneficial effects:

[0026] 1. This invention integrates static micro-deformation and dynamic low-frequency vibration data, generates a dimensionless fusion feature vector using normalization and weight allocation, and calculates the difference between single-node features and global average feature values. By utilizing the common-mode suppression principle, it effectively removes common environmental interferences of wind turbines under complex variable load conditions, overcomes the problem of false alarms or missed alarms caused by the influence of external noise in traditional single physical quantity monitoring, and improves the accuracy of determining the loose state of single screws.

[0027] 2. This invention constructs a range assessment mechanism based on multi-node fusion feature vectors. By extracting the range span of the difference values ​​of each node within the same synchronization period and comparing it with the set overall imbalance limit coefficient, it can directly identify the asymmetric torsional and overturning stress borne by the base as a whole before a single-point feature triggers an independent warning threshold. This enables early diagnosis of stress distortion of multi-point fasteners and the overall structural imbalance of the base, making up for the shortcomings of conventional single-point independent monitoring in reflecting the hidden dangers of the overall structure.

[0028] 3. This invention employs a hardware-level synchronous triggering mechanism with an accuracy better than 1 millisecond, combined with a polynomial temperature compensation algorithm incorporating first-order linear and second-order nonlinear coefficients to process the underlying signal. The synchronous triggering mechanism strictly aligns the sampling time of multiple nodes, eliminating spatial difference calculation errors caused by communication delays; the polynomial temperature compensation eliminates nonlinear bias and temperature drift of the sensor under extreme high and low temperature environments. The combination of these two methods ensures the authenticity of the underlying source data under harsh field conditions, providing reliable data support for subsequent spatial difference calculations. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the system architecture provided in an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram of the process for a differential monitoring method for fan screws that integrates micro-deformation and vibration, provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions in 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.

[0032] See attached document Figure 1 This invention provides a differential monitoring method for wind turbine screws that integrates micro-deformation and vibration. The system that this method relies on includes distributed monitoring nodes, a data processing unit, and an early warning display terminal.

[0033] Distributed monitoring nodes are deployed in the load-bearing area where the wind turbine base and screws connect, and are used to collect micron-level deformation data and low-frequency vibration signals at the screw fixing position.

[0034] The data processing unit communicates with each distributed monitoring node to receive, store, and process the collected deformation data and vibration signals, and to perform feature extraction, multi-point differential calculation, and loosening determination logic.

[0035] The early warning display terminal is connected to the data processing unit for receiving and outputting early warning information and monitoring status.

[0036] See attached document Figure 2 This invention provides a differential monitoring method for fan screws that integrates micro-deformation and vibration, comprising the following steps:

[0037] S10, Multiple distributed monitoring nodes are deployed at the diagonal position of the wind turbine base to acquire historical deformation data and historical vibration data of each distributed monitoring node;

[0038] S20, Based on the historical deformation data and the historical vibration data, construct a multidimensional feature baseline library;

[0039] S30, acquire strain signal and vibration signal, generate deformation deviation value based on strain signal and multidimensional feature baseline library, and extract frequency band feature components based on vibration signal;

[0040] S40, Based on preset weighting coefficients, the deformation deviation value and the frequency band feature components are linearly combined to generate a fused feature vector;

[0041] S50, calculate the average feature value of the fusion feature vector of all distributed monitoring nodes, and determine the difference between the fusion feature vector and the corresponding average feature value;

[0042] S60, determine the loosening state of a single screw based on the deformation deviation value and the difference value, and determine the degree of uneven force distribution on multiple screws of the fan base based on the loosening state.

[0043] In one possible implementation, multiple distributed monitoring nodes are deployed diagonally along the wind turbine base, utilizing strain sensors and low-frequency vibration sensors integrated within these nodes to detect structural deformation and vibration status. In this embodiment, this step specifically includes the implementation of the following sub-steps.

[0044] S101, Determine the spatial deployment topology strategy for the monitoring nodes. Considering the complex alternating loads experienced by the wind turbine base, a 4-node configuration is adopted for the basic deployment. Four distributed monitoring nodes are installed at the fixed screw locations at the four corners of the wind turbine base, forming a diagonally covered topology to capture changes in the overall overturning moment of the base during operation. As a preferred implementation, to further improve the diagnostic accuracy of multi-point stress imbalance states, the spatial deployment topology is expanded to a 6-node configuration. In addition to deploying one node at each of the four corners of the wind turbine base, one node is added at the midpoint of each of the two long sides of the base, forming a hexagonal distribution pattern. The nodes are installed using bolt fixing or welding, ensuring they fit snugly against the load-bearing area where the base connects to the screws. This rigid mechanical connection ensures that the micron-level deformation caused by changes in screw stress is transmitted to the sensor probe without attenuation, avoiding stress relaxation errors caused by traditional flexible adhesives.

[0045] S102, Constructing the physical protective enclosure for the monitoring nodes. Considering the environmental factors commonly found inside wind turbine towers, such as salt spray, oil contamination, and drastic temperature and humidity fluctuations, the distributed monitoring node's enclosure is made of high-strength metal materials and features a modular design for easy replacement and maintenance of internal components. The enclosure's protection rating is set at IP67 to prevent the intrusion of moisture and dust in harsh working environments. The enclosure internally encapsulates strain sensors, low-frequency vibration sensors, signal preprocessing circuitry, and a wireless communication module, forming a highly integrated independent sensing unit.

[0046] The S103 is equipped with dual-modal sensing elements to acquire structural physical quantities. High-precision strain gauges are used as strain sensors, which are laser-welded or cured with high-temperature rigid epoxy resin onto the mounting surface where the housing contacts the fan base. Based on the resistance strain measurement principle of a Wheatstone bridge, when the screw preload decreases, causing a small displacement in the base connection area, the strain gauge undergoes synchronous mechanical deformation, resulting in a linear change in its resistance value. The strain gauge has a resolution better than 0.1 micrometers. Based on the inherent physical properties of conventional constantan alloy material, its sensitivity coefficient is set to 2.0, and the nominal resistance value is configured to 120 ohms or 350 ohms, with a measurement accuracy better than 0.01%FS (full-scale output). In addition to capturing static deformation, a microelectromechanical system (MEMS) accelerometer is used to collect the vibration state at the screw fixing position. Its measurement range is configured to ±2g, with a sensitivity better than 0.01mg. To fully cover the low-frequency swaying characteristics of the wind turbine body and the screw loosening characteristic frequency required for subsequent diagnosis, the frequency response range of the accelerometer is set to cover 0.1Hz to 50Hz, with the corresponding noise density controlled within 10μg / √Hz, ensuring that the sensor outputs a raw analog signal with a sufficient signal-to-noise ratio when operating in the low-frequency range.

[0047] S104, a signal preprocessing circuit is constructed to perform analog signal conditioning. Based on the weak and easily interfered characteristics of the dual-modal sensor output signal, the signal preprocessing circuit independently conditions the analog signals output by the two sensors. The circuit's front-end is configured with an instrumentation amplifier as a low-noise amplifier, with an input noise voltage density below 1nV / √Hz. The filtering stage uses a fourth-order Butterworth filter, with a passband frequency range set from 0.1Hz to 50Hz. The Butterworth filter is chosen primarily because of its extremely flat amplitude-frequency response within the passband, preventing phase or amplitude distortion of key low-frequency characteristics at the hardware level, thereby filtering out high-frequency noise and power frequency interference in the wind turbine operating environment. For the specific circuit hardware topology implementation of the fourth-order Butterworth filter, those skilled in the art can use conventional cascaded active low-pass and high-pass filter operational amplifier circuits; its underlying circuit design is well-known in the field and will not be elaborated here. The conditioned analog signal is input to a 24-bit resolution Sigma-Delta analog-to-digital converter, which integrates a digital filter. According to the Nyquist sampling theorem, to reproduce the upper frequency signal of 50Hz without aliasing, the system converts the analog signal into a high-resolution digital signal at a sampling rate of 100Hz. To ensure the effectiveness of subsequent multimodal data fusion, the analog-to-digital converter adopts a multi-channel synchronous sampling mechanism to ensure strict time-domain alignment of the strain signal and vibration signal at the same moment.

[0048] S105 is configured with a wireless communication module for data uploading and interaction. The wireless communication module receives the digital signal output from the analog-to-digital converter and wirelessly transmits it to the data processing unit. To adapt to different types of wind turbine structures, the wireless communication module supports both Bluetooth Low Energy (BLE) and LoRa wireless communication protocols. In line-of-sight transmission scenarios inside the wind turbine tower, BLE with a communication distance of 10 to 30 meters and a transmission rate of 1 Mbps is used. In scenarios requiring cross-tower or long-distance centralized data collection, the LoRa protocol with a communication distance of several kilometers is switched to. The operating current of the communication module is controlled below 10mA, supports a sleep / wake-up mechanism, and has an average power consumption of less than 1mW. During the data packetization stage, the communication module adds a globally unified high-precision timestamp to each group of synchronously sampled data, solving the data synchronization problem between multiple nodes in a distributed spatial layout and providing time consistency assurance for subsequent spatial differential calculations. For the data encryption transmission process involved in wireless communication, those skilled in the art can use the standard AES symmetric encryption algorithm to encrypt the data packets. The encryption logic is well-known in the field and will not be elaborated here.

[0049] In this embodiment, the distributed monitoring nodes are controlled to enter baseline learning mode, continuously collecting historical deformation and vibration data within a set time period, thereby constructing a multi-dimensional feature baseline library including normal deformation range, normal vibration spectrum characteristics, and temperature drift curves. This step aims to provide a reliable individual reference benchmark for subsequent loosening state determination through a power-on multi-dimensional feature self-learning mechanism that eliminates the need for manual on-site calibration. The system executes the following specific implementation logic for this baseline self-learning process.

[0050] S201 executes the system power-on trigger and raw data buffering logic. After the monitoring system powers on and completes the self-test of each monitoring node sensor and the establishment of the communication link, it automatically triggers the entry into baseline learning mode. To ensure that the collected historical data can completely cover at least one complete stress change cycle of the wind turbine from yaw to wind, startup to stable operation, the continuous operation period of the baseline learning mode is fixed at 30 minutes. During this period, at least four monitoring nodes deployed at the diagonal position of the wind turbine base continuously collect historical deformation data and historical vibration data at a constant sampling rate of 100Hz. To ensure the temporal consistency of subsequent multi-point differential calculations, the data processing unit adds a global high-precision timestamp to each sampled data packet through a synchronous trigger mechanism to ensure strict alignment of multi-source heterogeneous data in the time dimension. The data processing unit is internally configured with a non-volatile storage module with a capacity of not less than 1GB to completely cache the 30 minutes of raw time series data, and this storage module is configured to meet the cyclic storage requirements of at least 30 days of historical data, thereby providing sufficient data sources for subsequent statistical analysis and anomaly backtracking. To prevent the misclassification of a loosened, abnormal state as a normal baseline during a wind turbine restart due to a fault, the data processing unit, after acquiring 30 minutes of baseline data, must compare it with the factory-set initial health baseline locked in internal non-volatile memory. Overwriting and generating the current multidimensional feature baseline library is only permitted when the evaluation deviation is within the set safety tolerance range, or when an authorized update instruction is received from the host computer after maintenance.

[0051] S202, extracting the normal deformation baseline and fluctuation range based on the moving average algorithm. After completing the above-mentioned raw data buffering, to obtain the static force baseline of the screw under normal preload, the data processing unit performs a moving average calculation on the buffered historical deformation data. Based on the physical purpose of filtering out gust impact interference caused by transient wind loads, the calculation window width for the moving average is set to 1 minute, and the window update step size is set to 1 second. The specific calculation formula for the normal deformation baseline of the corresponding monitoring node is as follows:

[0052] ;

[0053] In the formula, Indicates the first The average deformation baseline value of each monitoring node; The node number is determined by a value from 1 to 4. This represents the total number of data points within the sliding window, based on a 100Hz sampling rate and a 1-minute window length. The constant value is set to 6000. This fixed non-zero integer setting directly avoids the singularity problem of the denominator approaching 0 during the calculation process. Indicates the first The node at the th Instantaneous deformation measurement values ​​at each sampling time.

[0054] After calculating and obtaining the average deformation baseline, to further define the random fluctuation boundary under normal operating conditions, the system simultaneously calculates the sample variance and standard deviation of the deformation within this window. The system employs a sample variance calculation logic incorporating Bessel correction to eliminate unbiased estimation errors. The specific formula is as follows:

[0055] ;

[0056] In the formula, Indicates the first The variance of historical deformation data of each node This represents the corresponding standard deviation.

[0057] Based on the standard deviation results of the above output, the data processing unit uses Laida's rule to construct the upper and lower limits of the normal deformation variable, that is, to define the normal interval as... Any minute deformation exceeding this range will be flagged by the system as an early sign of suspected preload decay.

[0058] S203, Extract the baseline of normal vibration spectrum characteristics. In addition to extracting the static deformation baseline, a reference model also needs to be established for the background vibration of the wind turbine base under normal operating conditions to isolate common-mode vibration interference caused by the rotation of the wind turbine nacelle. Based on the technical objective of reducing the variance of power spectrum estimation using the classical periodogram method, the data processing unit uses the Welch method to estimate the power spectrum of 30 minutes of historical vibration data. The system divides the 30-minute continuous data sequence into 15 macroscopic data segments, each 2 minutes long. For each 2-minute macroscopic data segment, the system further uses a sliding window to divide it into multiple sub-data segments of 10.24 seconds each. A 50% overlap rate is set between adjacent sub-data segments to reduce variance fluctuations caused by signal truncation. Within each data segment, a Hanning window function is used for windowing, and a 1024-point Fast Fourier Transform (FFT) operation is performed. Since the system sampling rate is 100Hz, the frequency resolution corresponding to this FFT operation is approximately 0.098Hz, effectively covering the low-frequency target band from 0.1Hz to 50Hz. The formula for calculating the baseline characteristic of the normal vibration spectrum is as follows:

[0059] ;

[0060] In the formula, Representing frequency point The baseline of the spectral amplitude at that location; The total number of sub-data segments is set to a constant of 15 here; Indicates the first After FFT operation, the windowed data segments are located at frequency points. The amplitude component at that frequency represents the accumulated normal vibration energy of the base at that frequency during the baseline learning period. This represents the index of the macro data segment, with values ​​ranging from 1 to... .

[0061] The butterfly decomposition and complex multiplication-addition logic involved in the Fast Fourier Transform can be implemented by conventional algorithm code in standard digital signal processing libraries by those skilled in the art. Its underlying operation mechanism is a well-known technology in this field and will not be described in detail here.

[0062] S204 configures the data structure and dynamic mapping relationship of the multi-dimensional feature baseline library. To achieve long-term management and environmental adaptation of the aforementioned multi-dimensional baseline data, the data processing unit establishes a temperature drift curve by combining the ambient temperature data recorded synchronously during baseline learning. The system uses a global timestamp as a unique index to align the average deformation baseline value extracted in S202 with the synchronously acquired ambient temperature data in the time domain, thereby constructing a two-dimensional discrete point set reflecting the correspondence between temperature and sensor output.

[0063] As a preferred implementation, the data processing unit structures the calculation results of all dimensions to construct a multi-dimensional feature baseline library. This baseline library consists of a strain baseline sub-library, a spectral baseline sub-library, and a temperature drift baseline sub-library. Each sub-library independently contains three data fields: a data array, calibration parameters, and an expiration date identifier. The data array is stored in single-precision floating-point format, and its length is strictly defined as the number of valid data points during baseline learning. The calibration parameters include the initial gain coefficient and bias coefficient used for subsequent sensor output correction. The expiration date identifier is set as a high-precision timestamp of the baseline establishment time. By checking this expiration date identifier, the system can determine whether it is necessary to re-trigger the baseline learning mode to update the baseline library when the wind turbine undergoes a major overhaul or experiences severe structural aging during long-term operation.

[0064] In this embodiment, after the baseline self-learning mode finishes running, the monitoring system seamlessly switches to real-time data acquisition mode, executing synchronous acquisition preprocessing and full-temperature-range temperature compensation logic. This process aims to acquire real-time multi-dimensional sensing data without temporal phase difference and free from environmental temperature drift interference, providing accurate numerical input for subsequent feature fusion and spatial difference analysis. This step specifically includes the implementation of the following sub-steps.

[0065] The S301 executes synchronous acquisition triggering logic based on a high-precision global clock. The core physical premise of implementing the multi-point differential algorithm lies in the strict alignment of the sensing data from each node along the diagonal of the base in the time dimension. To avoid phase deviation caused by independent sampling, the data processing unit generates a synchronous trigger command through an internal timer and sends it to all distributed monitoring nodes via a wireless communication link. Based on a trigger accuracy constraint better than 1 millisecond, monitoring nodes distributed in different areas of the wind turbine base synchronously initiate the data conversion process upon receiving the trigger command. The acquisition cycle of each node is strictly locked to a 100Hz sampling rate benchmark, thereby accurately capturing the instantaneous force and vibration states of different physical locations on the base within the same macroscopic time slice, eliminating spatial differential calculation errors caused by time delay.

[0066] S302 performs customized filtering and amplification preprocessing for multi-source heterogeneous signals. To address the environmental noise and inherent baseline drift issues embedded in the original analog signals, the system configures independent signal preprocessing channels within the monitoring nodes. For micron-level strain signals, considering their susceptibility to background electromagnetic noise, a low-noise amplifier with an instrumentation amplifier architecture is connected at the beginning of the signal path, with its input noise voltage density strictly controlled below 1nV / √Hz. The amplifier's gain is adjustable from 100 to 1000 times, with the specific gain value determined by matching the output range of the strain sensor with the input full-scale range of the subsequent analog-to-digital converter. The amplified signal then passes through a low-pass filter with a cutoff frequency of 10Hz to physically block high-frequency electromagnetic environmental noise interference, while fully preserving the DC (0Hz) and extremely low-frequency static micro-deformation signals characterizing the true physical displacement of the structure.

[0067] In parallel, for low-frequency vibration signals, the system employs a fourth-order Butterworth bandpass filter architecture to extract the effective frequency band from 0.1Hz to 50Hz. The frequency-domain shaping network implemented based on this fourth-order digital filter meets the stringent technical requirements of less than 0.1dB ripple in the passband and greater than 60dB attenuation in the stopband, effectively eliminating high-frequency electromagnetic noise and power frequency interference while fully preserving the effective signal components within the screw loosening characteristic frequency range. The vibration channel signal is conditioned by an amplifier with a gain set between 10 and 50 times, the gain of which is also adaptively configured based on the output sensitivity of the MEMS accelerometer. Finally, these two independent analog signals are synchronously fed into a Sigma-Delta analog-to-digital converter with 24-bit resolution for digitization. As a preferred approach, this analog-to-digital converter integrates a digital filter to further suppress quantization noise and clock jitter interference that occur during signal discretization.

[0068] S303, based on polynomial fitting, features full-temperature-range adaptive temperature compensation. The operating environment of wind turbine generators experiences significant diurnal temperature variations. These drastic temperature changes inevitably cause thermal expansion and contraction of the strain sensor substrate material, leading to severe zero-point drift. To restore the sensor's true stress response characteristics, the data processing unit performs adaptive temperature compensation simultaneously during the data preprocessing stage. The system acquires each set of deformation data while simultaneously reading the ambient temperature value from the built-in temperature sensor in real time. To ensure the reliability of the compensation benchmark, the measurement accuracy of this built-in temperature sensor is set to ±0.5℃, with a resolution better than 0.1℃. The temperature compensation calculation formula for the raw strain measurement data is as follows:

[0069] ;

[0070] In the formula, This represents the true value of the deformation after temperature compensation correction. This value directly represents the pure mechanical strain after filtering out thermal stress interference. This represents the raw deformation measurement value output by the analog-to-digital converter at the node's front end; Defined as the difference between the current real-time ambient temperature and the calibration reference temperature; This represents the linear temperature compensation coefficient. This represents the quadratic nonlinear temperature compensation coefficient. By introducing a quadratic polynomial operational architecture, this algorithm can effectively approximate and eliminate the nonlinear bias error of the sensor under extreme high and low temperature environments. The term representing the square of the temperature difference is used to compensate for the second-order nonlinear effect of the thermal expansion coefficient of the sensor substrate material.

[0071] The calibration coefficients mentioned above and The data was obtained by fitting high and low temperature alternating calibration test data before the equipment left the factory. Specifically, the calibration test was carried out in a constant temperature chamber. Multiple discrete temperature points were set within the operating temperature range of -40℃ to +85℃. The output offset of the sensor at each temperature point was measured, and the compensation coefficient was obtained by fitting based on the least squares method, thereby ensuring that the sensor's detection accuracy remains above 95% within this wide temperature range.

[0072] S304 executes standardized data encapsulation and circular buffer storage logic. Discrete time-series data, after preprocessing and temperature compensation correction, needs to be uploaded to the main control unit for feature fusion analysis. The microcontroller inside the monitoring node packages the discrete data according to a predetermined communication protocol, constructing a high-density batch standardized data packet containing 10 consecutive synchronous sampling points. Its length is dynamically set to 128 bytes (or a length greater than 128 bytes to adapt to the protocol's upper limit). This data packet structure not only carries the core strain and vibration data fields but also forcibly encapsulates the node's unique number and a high-precision timestamp field, thus ensuring the spatiotemporal traceability of multi-source data at the protocol level. The wireless communication module sends information to the data processing unit at a stable rate of 10 data packets per second. The receiving end of the data processing unit continuously performs cyclic redundancy check and unpacking operations. Verified valid data is sequentially pushed into the circular buffer constructed internally by the system. Considering the physical requirements of the time window length for subsequent feature extraction algorithms, the effective depth of this circular buffer is set to be able to completely cache the raw continuous measurement data of the most recent 10 minutes. When the continuous data flow causes the buffer to reach its capacity limit, the system automatically executes a first-in-first-out circular overwrite logic to evict old data, thereby maintaining the uninterrupted dynamic data queue required by the real-time differential computing engine within the limited hardware memory capacity.

[0073] In this embodiment, after the data processing unit continuously receives and buffers the preprocessed data stream from the monitoring nodes, the system enters the multi-source heterogeneous feature extraction and dimensionless fusion execution stage. This stage aims to extract fault features from the sensed signals from both the time and frequency domains, and eliminate the observation limitations of a single sensing mechanism through cross-weighting. Based on the general technical principles of multi-dimensional feature mapping, this process specifically includes the implementation logic of the following sub-steps.

[0074] S401 performs time-domain deformation feature extraction based on a sliding time window. For the micro-deformation data channel, based on the fundamental principles of elasticity, the decay of screw preload directly leads to microscopic relative displacement at the base connection interface. The data processing unit quantifies the physical displacement deviation of the screw fixing point by comparing the real-time measurement sequence with the baseline state. Considering that transient wind load impacts and mechanical vibrations are easily superimposed with random high-frequency jitter on single-point sampling values, the system uses a sliding time window to smooth the data in the time domain. The width of this sliding time window is set to 10 seconds. Combined with the system's 100Hz underlying sampling rate configuration, this window stably contains 1000 discrete sampling points, effectively filtering out transient stress interference and obtaining effective measurement values ​​representing the current static preload state. As the data queue is continuously updated, the system triggers calculation logic once per second, extracting the arithmetic mean of all discrete strain values ​​within the current window as the input variable for the current deformation state. Based on the extracted mean parameter, the system calculates the deformation deviation value:

[0075] ;

[0076] In the formula, Defined as the deformation deviation value, its unit is micrometer; This represents the average deformation at the current moment after smoothing via a sliding window. This represents the baseline value of the average deformation of the corresponding node established and stored locally during the baseline self-learning phase; This represents the absolute value operation function.

[0077] The purpose of introducing absolute value calculation here is to eliminate numerical sign interference caused by differences in sensor assembly orientation and different polarities of screw force, and to strictly focus the evaluation on the absolute magnitude of structural displacement deviating from the normal benchmark.

[0078] S402 performs time-frequency domain spatial mapping and feature frequency point capture for low-frequency vibration channels. Operating in parallel with the time-domain extraction logic, the data processing unit performs frequency domain feature extraction based on Fast Fourier Transform (FFT) within the vibration data channel. The basic principle of time-frequency conversion is to map the discrete acceleration sequence that fluctuates with time to the frequency domain, thereby separating the specific resonant energy hidden in the wideband random response. The system extracts a continuous vibration sequence of the most recent 10.24 seconds from the circular buffer as the input source. Multiplying this time window length by the 100Hz sampling rate yields precisely 1024 valid sampling points, directly satisfying the underlying computational constraint of the FFT algorithm for input sequence lengths raised to integer powers of 2. To suppress the energy leakage effect caused by finite-length sequence truncation at the frequency domain edges, the original discrete-time sequence is subjected to a Hanning window for time-domain envelope shaping before being input into the FFT calculation engine.

[0079] The transformed amplitude spectrum has a frequency domain resolution of approximately 0.098 Hz. Based on the generated frequency domain distribution data, the system performs a local peak search for specific frequency points. Given that real-time fluctuations in the wind turbine's main shaft speed can cause slight shifts in the characteristic resonant frequency of the mechanical structure, the system does not employ a static extraction method that extracts the amplitude of a single absolute frequency point. Instead, the algorithm performs local extremum traversal within ±0.5 Hz bandwidth intervals of the two preset target frequencies associated with loosening features, 2 Hz and 3 Hz, extracting the maximum amplitude within these closed intervals as independent feature parameters characterizing the energy state of that frequency band. The extracted values ​​are labeled as follows: and The dimensions of all of them are milligravitational acceleration (mg).

[0080] S403, constructing a comprehensive energy index for a specific frequency band characterizing the loosening state of the structure. After acquiring the discrete frequency band characteristics, the data processing unit further synthesizes evaluation components sensitive to screw loosening faults. Based on previous dynamic tests, the 2Hz to 3Hz frequency band is confirmed to be a typical characteristic frequency range for base screw loosening. When initial loosening of local bolts causes a decrease in connection stiffness, the vibration energy in this frequency band will increase. Therefore, the system extracts parameters for this range and calculates the characteristic components of the 2 to 3Hz frequency band:

[0081] ;

[0082] In the formula, The comprehensive energy characteristic components representing the target's sensitive frequency band; and These are the peak amplitudes within the neighborhood of the 2Hz and 3Hz frequency points captured in the previous steps. The division operation included in this formula uses a fixed constant of 2 in the denominator to avoid singular anomalies where the denominator approaches 0. The arithmetic average fusion logic is employed, its underlying principle being to smooth out sporadic measurement noise and random disturbances that may exist at a single center frequency by merging the energy densities of adjacent preset target frequencies, thus providing a more stable frequency band energy assessment benchmark for subsequent judgments.

[0083] S404 performs dimensionless weighted fusion of multi-source heterogeneous parameters. After independently acquiring time-domain deformation deviation and frequency-domain vibration energy indices, the system establishes a cross-modal data fusion framework. Single-dimensional sensor data has observation blind spots; micro-deformation measurements are sensitive to static preload but easily affected by local material inhomogeneities, while acceleration measurements reflect dynamic responses but are easily mixed with broadband background noise. To ensure the physical rationality of cross-modal fusion, the system relies on a synchronization triggering mechanism with better than 1 millisecond accuracy in the underlying hardware architecture, ensuring that the strain data and vibration data input to the fusion formula are strictly aligned in the time dimension. Based on synchronously acquired multi-source data, the system calculates the fused feature vector:

[0084] ;

[0085] In the formula, This represents the generated fusion feature vector; The extracted deformation deviation value; The comprehensive energy characteristic components of the characteristic frequency band; Defined as the strain feature fusion weighting coefficient; Defined as the vibration characteristic fusion weighting coefficient.

[0086] Directly adding parameters from different physical domains can lead to dimensional conflicts, therefore the weighting coefficients... and In addition to allocating information contribution proportions, its numerical setting also possesses a scaling attribute based on the reciprocal of the corresponding physical parameter dimensions. That is, through coefficient multiplication, it forcibly converts the micrometer and milligravity acceleration dimensions into normalized values, ensuring the output feature vector... It is a dimensionless scalar.

[0087] As a preferred implementation method, the specific values ​​of the weighting coefficients are obtained through training and optimization using a large amount of experimental data in the early stages. Before weighting, the deformation warning threshold (e.g., 15 micrometers) and the energy benchmark value are used to calculate the weighting coefficients. and A division normalization process is performed to convert it into a dimensionless relative deviation coefficient between 0 and 1. Then, a weighting method is applied. , A linear combination of the two normalization coefficients is performed. This asymmetric numerical ratio reflects the dominant role of deformation deviation in assessing preload decay, indicating that static deformation deviation at the material physics level has a more direct causal relationship with failure than structural dynamic response. The generated dimensionless fused eigenvector is continuously output along the time axis and serves as the basic reference node data in subsequent steps for constructing a three-dimensional spatial difference model and performing multi-point difference calculations.

[0088] In this embodiment, after acquiring the dimensionless fused feature vectors of each monitoring node, the data processing unit constructs a multi-point difference model. Based on the dynamic stress characteristics of the wind turbine structure, temperature fluctuations in the natural environment or overall low-frequency vibrations during wind turbine operation typically generate common-mode responses with similar directions and amplitudes on the base structure. In contrast, the preload decay of a single fastening screw manifests as differential-mode distortion at a local node. Based on the general technical principle of common-mode suppression mentioned above, this embodiment compares the relative deviations between local features and the global benchmark to isolate common interferences from macroscopic operating conditions. The specific implementation logic includes the following sub-steps.

[0089] S501, calculates the global average characteristic benchmark of the monitoring node group. To establish a dynamic reference line representing the overall state of the current wind turbine base, the data processing unit extracts data from all nodes within the same synchronization time slice. Taking a basic topology with four monitoring nodes deployed diagonally on the wind turbine base as an example, the expression for calculating the average characteristic value is:

[0090] ;

[0091] In the formula, This represents the average characteristic value of the four monitoring nodes within the current synchronization period; , , and These represent the fused feature vectors input in real time from monitoring nodes 1 to 4, respectively. Since the denominator of this operation is a fixed constant of 4 based on the number of physical nodes, the singularity anomaly of division by zero is directly avoided at the algorithm's underlying level. The calculated... It reflects the overall stress level and comprehensive vibration energy of the fastener group at the current moment, and serves as a benchmark reference for subsequent differential comparison.

[0092] S502 performs spatial difference calculations for a single node relative to the global benchmark. Based on the established dynamic mean benchmark, the data processing unit calculates the algebraic difference between the feature vector and the benchmark for each independent monitoring node, thereby separating out the specific changes at local locations. The mathematical expression of the difference calculation is:

[0093] ;

[0094] In the formula, For the first The difference value of each node; For the first The fused feature vector of each monitoring node; variables The nodes are numbered using natural numbers 1, 2, 3, and 4 in this embodiment. The difference calculation result directly retains the positive and negative polarity information. When the result is positive, it indicates that the characteristic strength of the node is higher than the overall level of the base, indicating abnormal stress concentration in the local area or low-frequency vibration amplification due to decreased connection stiffness; if the result is negative, it indicates that the response of the area is lower than the overall average. The absolute magnitude of the difference value directly quantifies the severity of the node's deviation from the normal equilibrium stress state.

[0095] S503 constructs a three-dimensional spatial state mapping model. After obtaining the relative deviations of discrete individuals, the system structurally reorganizes these parameters in the spatial dimension. The data processing unit organizes the difference values ​​of each node into a multi-dimensional state vector, specifically represented as follows: In the geometric mapping of the four-dimensional feature space, this vector corresponds to a spatial state coordinate point. In an ideal state, where all screws on the wind turbine base maintain standard preload and are subjected to perfectly symmetrical forces, the difference values ​​at each node approach zero, and this state coordinate point is located at the origin of the coordinate system. When the preload of a screw decays, disrupting the force balance, the corresponding difference value drifts significantly, causing the coordinate point to move directionally along the associated coordinate axis. The spatial distance from this point to the origin characterizes the overall degree of force imbalance in the base, while the distribution pattern of its movement direction precisely maps the physical location of the anomaly.

[0096] S504, extract auxiliary statistical features of the difference distribution. To support subsequent detection of hidden faults caused by uneven stress on multiple screws, the data processing unit simultaneously calculates the statistical parameters of the difference value set within the current period. The data processing unit calculates the standard deviation of the difference values. The formula is as follows:

[0097] ;

[0098] In the formula, The total number of nodes is 4 here. Since the arithmetic mean of the difference set is theoretically always zero, the above standard deviation calculation is equivalent to a normalized expression of the distance from the spatial coordinate point to the origin, which is used to quantify the overall dispersion of the difference distribution of each node.

[0099] Simultaneously, the system extracts the maximum and minimum difference values ​​for the current calculation cycle:

[0100] ;

[0101] ;

[0102] In the formula, and These are the maximum and minimum values, respectively. and This represents the extreme value search function. This set of parameters is used to capture the most extreme non-equilibrium states on the base surface, serving as the basis for calculating the range span in subsequent multi-point non-equilibrium logic (i.e., directly through...). (Calculated directly from the input parameters).

[0103] In addition, to determine the dominant mode of fault deviation, the data processing unit statistically analyzes the sign distribution of the difference values. Specifically, the system determines the polarity of each difference value through a sign function and accumulates the number of positive and negative deviation nodes respectively, thereby determining whether the current base anomaly is dominated by local positive distortion or negative collapse.

[0104] The above-mentioned multi-point differential calculation and feature extraction steps are driven by the underlying high-precision synchronous trigger clock and are continuously executed at a frequency of 1Hz. The differential parameters output are updated in real time and pushed into the subsequent early warning judgment module.

[0105] As a preferred extended implementation, for the expanded stress area of ​​the base of large wind turbine units, the system monitoring topology can be adjusted to a hexagonal distribution of 6 nodes. Based on the same common-mode suppression principle described above, those skilled in the art can simultaneously adjust the denominator and the number of feature variables in the mean calculation formula to 6, and correspondingly extract 6 independent difference values ​​to obtain the isomorphically extended six-dimensional state vector. This spatial dimension expansion, while maintaining the underlying differential logic, further improves the geometric resolution for locating latent faults on complex contact surfaces.

[0106] In this embodiment, after acquiring the feature extraction data of each monitoring node and the output of the three-dimensional spatial multi-point difference model, the data processing unit enters the multi-level early warning judgment and fault diagnosis stage. This judgment mechanism integrates the micro-deformation index characterizing the static state, the low-frequency vibration energy deviation reflecting dynamic characteristics, and the multi-point force balance characteristics characterizing spatial distribution. Through the joint diagnosis of the above parameters, the system can comprehensively assess the condition of individual fasteners and the health status of the entire wind turbine base bearing system. The specific implementation process includes the following sub-step logic.

[0107] S601 sets multi-dimensional fault judgment benchmarks and dynamic thresholds. To meet the engineering requirements for early warning of loose screws on wind turbine bases, the system pre-sets multi-dimensional reference boundaries for condition classification. The judgment thresholds and boundary parameters at each level are extracted and calibrated based on statistical patterns from destructive test data of wind turbine benches and historical real fault sample libraries. For static micro-deformation characteristics, the system sets two levels of deformation deviation limits. Combining the typical mechanical stiffness decay law of wind turbine bases, the yellow warning deformation threshold representing slight loosening of initial preload is set at 10 micrometers; the red warning deformation threshold representing loss of preload is set at 15 micrometers. For dynamic low-frequency vibration characteristics, the system sets threshold ranges based on the energy deviation ratio of characteristic frequency bands. Physically, when the stiffness of a local screw decreases, the vibration damping in the corresponding area within a specific frequency band decreases, causing an abnormal amplification of the vibration energy in that frequency band relative to the normal baseline level. The system sets the corresponding yellow warning energy deviation ratio range to 10% to 15%, and the red warning energy deviation ratio limit to exceed 15%. To ensure the overall stress balance of the base, the system sets a limit on the overall imbalance ratio for evaluating the span of local stress distortion. This value is set to 30% in the basic configuration of this embodiment. The above-mentioned multi-dimensional benchmark parameters are pre-written into the storage medium of the data processing unit through the system configuration file and serve as the benchmark for state comparison during operation.

[0108] S602 performs single-point state determination based on the fusion of deformation and spectral characteristics. Based on the deterioration characteristics of mechanical structures, fastener loosening not only causes micron-level distortion of the static physical gap at the connection surface but also induces synchronous changes in the dynamic stiffness of the local system. Therefore, the data processing unit evaluates the absolute deformation deviation and relative frequency band energy fluctuation amplitude of each independent monitoring node in parallel. The system calculates the energy deviation ratio of this node within the characteristic frequency band of screw loosening, using the following mathematical model:

[0109] ;

[0110] In the formula, Defined as the percentage of energy deviation of the current node in the 2Hz to 3Hz frequency band; This refers to the comprehensive energy index of this frequency band extracted in real time. This is the normal energy reference for the frequency band corresponding to this node established by the system during the baseline self-learning phase; This represents the absolute value function, used to extract the relative absolute amplitude of fluctuations; A very small constant (e.g., with a value of 10) is set to prevent the denominator from being zero. -6 ).

[0111] The introduction of this minimal constant mathematically avoids the singularity calculation anomalies caused by the divisor approaching zero under extremely low baseline energy conditions, ensuring the numerical stability of the algorithm's underlying structure. This formula, by calculating the relative rate of change, eliminates the systematic error caused by the inherent differences in vibration amplitude between different nodes in their initial installation state.

[0112] Based on the calculated parameters, the judgment logic comprehensively classifies the state of a single node according to the conservative early warning principle. The data processing unit obtains the deformation deviation value of the current node. When it is determined Larger than 10 micrometers and less than or equal to 15 micrometers, or energy deviation ratio When the value is between 10% and 15%, the system determines that the fasteners in the monitored area of ​​that node show initial signs of loosening, triggering the generation of a single-point yellow warning command. Larger than 15 micrometers, or energy deviates from the ratio When the rate exceeds 15%, the system directly determines that the fastener at that location is severely loose or on the verge of falling off, triggering the generation of a single-point red warning command. This dual-dimensional union judgment mechanism not only ensures the capture of slow deformation evolution in pure statics, but also covers specific working conditions where small deformations have not yet become apparent but the dynamic stiffness of the connection part has undergone a sudden change, minimizing the risk of missed fault reports caused by fluctuations in a single sensed physical quantity or local failure of a sensor.

[0113] S603 performs multi-point implicit imbalance diagnosis based on spatial feature vectors. In addition to absolute threshold interception for discrete nodes, the system further utilizes a pre-established multi-point difference model to identify overall base stress imbalance faults caused by uneven distribution of preload on multiple screws. The data processing unit extracts the range of spatial difference values ​​from multiple monitoring nodes within the current synchronization cycle and verifies it relative to the global average benchmark. The mathematical condition triggering this type of anomaly determination is expressed as:

[0114] ;

[0115] In the formula, and These represent the maximum and minimum values ​​in the fused feature vector set of each node within this calculation cycle; the difference between the two constitutes the range parameter, which reflects the maximum span of the spatial distribution of the features of each node. The overall imbalance limit coefficient set for the system is a constant of 30%. It is the arithmetic mean of the fused feature vectors of all monitoring nodes.

[0116] When the above inequality condition is met, it indicates that the distribution span of the eigenvector in the physical space of the base has exceeded the tolerance limit of the normal overall stress state. Based on this, the system determines that the wind turbine base is subjected to asymmetrical abnormal torsional or overturning stress and immediately triggers a multi-point imbalance alarm, which is set to an orange state independent of the single-point warning. The triggering of this diagnostic condition means that even if no single point touches the absolute threshold of loosening, the load-bearing components of the foundation have already undergone stress distortion, providing a time window for early intervention against potential fatigue cracking risks in the base.

[0117] S604 generates an early warning data packet and distributes it to multi-level monitoring terminals. Once any dimension of diagnostic logic touches the alarm boundary, the data processing unit immediately performs structured encapsulation of the current operating condition data to generate an early warning data packet. This data packet carries a precise early warning level identifier, a system clock timestamp with second-level accuracy, a monitoring node number indicating the physical location of the anomaly, and low-level characteristic parameters that trigger the alarm transient. These characteristic parameters fully record the deformation deviation value, characteristic frequency band amplitude components, and spatial differential offset value that trigger the diagnostic judgment, providing high-confidence data support for subsequent manual review by maintenance personnel, cause tracing, and fault evolution trend analysis.

[0118] The early warning data packet is synchronously pushed to distributed display and monitoring terminals via the underlying communication bus. After receiving the data, the local display panel located at the bottom of the wind turbine tower or inside the control cabinet displays a distribution map of abnormal node locations and their transient parameters through an LCD interface, providing intuitive guidance for on-site maintenance personnel to confirm the location and perform targeted tightening operations. The wireless communication module on the data processing unit simultaneously sends the early warning data frame to an external remote network. For the basic logic interaction implementation of the cloud platform remote monitoring system—including receiving early warning information, storing structured databases, providing multi-dimensional historical data query services, and pushing early warning reminders to designated terminals via application clients—those skilled in the art can use mature IoT message subscription and distribution middleware and relational databases. The underlying network architecture and data flow are well-known technologies in the field and will not be elaborated upon here.

[0119] As a preferred extended implementation, when the underlying monitoring node topology is upgraded from a basic diagonal distribution of 4 nodes to a hexagonal distribution of 6 nodes, in order to match the increased coverage area of ​​the base stress region with the increased number of nodes, the system will adjust the overall imbalance limit coefficient in the aforementioned imbalance diagnosis triggering conditions. The threshold was adjusted from 30% to 25% based on the baseline. This adaptive threshold adjustment strategy for topology scale effectively compensates for the statistical trend that the range span in multi-node datasets is easily diluted, ensuring that the multi-point judgment model can still maintain a high sensitivity to identify hidden faults with uneven stress on multiple screws in complex load-bearing surface monitoring scenarios.

[0120] Specific application examples:

[0121] Assume that the screw on the left side of the rear of the wind turbine base (corresponding to monitoring node 3 in the system) is slowly loosened throughout the entire monitoring process.

[0122] Normal operation phase (status indicator: green):

[0123] Characteristics of monitoring data:

[0124] Deformation deviation value Stable at 2 to 5 micrometers;

[0125] Eigencomponents The deviation rate is close to 0%;

[0126] Fusion feature vectors with average eigenvalues Basically the same;

[0127] Difference Approaching 0.

[0128] System analysis: The equipment is operating stably, and the system status indicator light remains green.

[0129] Initial loosening stage (triggers yellow alert):

[0130] Monitoring data characteristics: As the screw preload begins to decrease, Slowly climbing to 10 micrometers, while The deviation rate reached 8%.

[0131] System analysis: The data meets the triggering conditions for a yellow alert.

[0132] System Action: The warning terminal outputs an initial loosening alert.

[0133] Severe loosening stage (triggered red alert)

[0134] Monitoring data characteristics: If no manual intervention is implemented initially, the loosening worsens. The system captures... It increased dramatically to 18 micrometers (breaking the absolute safety threshold of 15 micrometers). The deviation rate surged to 18%, and the feature vectors and differences were fused. A step amplification occurs simultaneously.

[0135] System analysis: The data processing unit immediately determined that node 3 was severely loose.

[0136] System action: Output a single-point red warning and accurately locate the physical location of the anomaly (left rear) on the display terminal.

[0137] Operation and maintenance guidance results (achieving closed loop):

[0138] Intervention effect: Through the system's precise positioning and hierarchical alarm, it successfully guided on-site maintenance personnel to go directly to the abnormal point to complete targeted tightening operations before serious failures such as cracking of the wind turbine base occurred, thus eliminating safety hazards.

Claims

1. A differential monitoring method for fan screws integrating micro-deformation and vibration, characterized in that, Includes the following steps: Multiple distributed monitoring nodes are deployed at the diagonal position of the wind turbine base to acquire historical deformation data and historical vibration data of each distributed monitoring node; Based on the aforementioned historical deformation data and historical vibration data, a multidimensional feature baseline library is constructed; Acquire strain signals and vibration signals, generate deformation deviation values ​​based on the strain signals and the multidimensional feature baseline library, and extract frequency band feature components based on the vibration signals; Based on preset weighting coefficients, the deformation deviation value and the frequency band feature components are linearly combined to generate a fused feature vector; Calculate the average eigenvalue of the fused feature vector of all distributed monitoring nodes, and determine the difference between the fused feature vector and the corresponding average eigenvalue; The loosening state of a single screw is determined based on the deformation deviation value and the difference value, and the degree of uneven force distribution on multiple screws of the fan base is determined based on the loosening state.

2. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, Controlling the distributed monitoring nodes to enter baseline learning mode and constructing a multi-dimensional feature baseline library specifically includes: For the historical deformation data obtained by each of the distributed monitoring nodes, a moving average calculation is performed to obtain the average deformation baseline value, and the sample variance and standard deviation are calculated simultaneously to construct the normal deformation range. For the historical vibration data acquired by each of the distributed monitoring nodes, the power spectrum estimation method is used to divide the continuous data sequence into sub-data segments and perform fast Fourier transform operation to extract the normal vibration spectrum characteristics of the base under normal working conditions. A temperature drift curve is established by combining the ambient temperature data synchronously recorded by each of the distributed monitoring nodes during the baseline learning mode, and the normal deformation range, the normal vibration spectrum characteristics, and the temperature drift curve are structured to form the multidimensional feature baseline library.

3. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, After acquiring the current strain and vibration signals using a synchronous triggering mechanism, the following is also included: The strain signal and the vibration signal are sequentially filtered, amplified, and converted from analog to digital to obtain the corresponding high-resolution digital sequence. The high-resolution digital sequence is combined with the real-time read ambient temperature data to perform a temperature compensation algorithm. Specifically, based on the calibration-obtained linear temperature compensation coefficient and the quadratic nonlinear temperature compensation coefficient, a polynomial fitting operation is performed using the difference between the currently collected ambient temperature and the calibration reference temperature and the square of the temperature difference to eliminate nonlinear bias error and temperature drift error under extreme high and low temperature environments, in order to obtain the true value of the deformation after temperature compensation correction.

4. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, Extracting the current deformation based on the strain signal and calculating the difference between the current deformation and the benchmark value in the multidimensional feature baseline library to obtain the deformation deviation value specifically includes: In the time domain, a sliding time window is used to smooth the discrete measurement sequence corresponding to the strain signal, and the arithmetic mean of all discrete strain values ​​in the current window is extracted as the current deformation. Calculate the difference between the current deformation and the average deformation baseline value in the multidimensional feature baseline library, and perform an absolute value operation on the difference to eliminate the interference of tensile and compressive polarity, which is then used as the deformation deviation value.

5. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, Performing spectral analysis on the vibration signal to extract frequency band feature components specifically includes: The data sequence corresponding to the continuous vibration signal is extracted, Hanning window is applied to perform time-domain envelope shaping, and fast Fourier transform is performed to calculate the mapping to the frequency domain; In the frequency domain, local extreme value traversal is performed within the ±0.5Hz bandwidth interval of each preset target frequency point associated with loosening characteristics, and the maximum amplitude within the interval is extracted as an independent feature parameter characterizing the energy state of the frequency band. The independent feature parameters of adjacent preset target frequency points are merged by arithmetic averaging fusion logic to generate a comprehensive energy index that reflects the energy state within the frequency range of screw loosening characteristics as the frequency band feature component. The preset target frequencies are 2Hz and 3Hz.

6. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, The process of generating a fused feature vector by linearly combining the deformation deviation value and the frequency band feature components using set weighting coefficients specifically includes: The deformation deviation value and the frequency band characteristic component are normalized using a preset reference value and converted into a dimensionless relative deviation coefficient. Using the set weighting coefficients, which include a pre-set strain feature fusion weighting coefficient to reflect the dominant role of deformation deviation and a corresponding vibration feature fusion weighting coefficient, the normalized relative deviation coefficients are linearly combined, and the physical parameter dimensions are unified by the reciprocal scaling property of the set weighting coefficients, in order to generate the fusion feature vector that is expressed as a dimensionless scalar. Wherein, the set weight coefficient is a value obtained through training and optimization with experimental data, the preset benchmark value is the deformation warning threshold and the energy benchmark value, and the strain feature fusion weight coefficient and the vibration feature fusion weight coefficient are values ​​obtained through training and optimization with experimental data.

7. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, Calculating the difference between the fused feature vector of each of the distributed monitoring nodes and the average feature value specifically includes: Obtain the fusion feature vectors of all the distributed monitoring nodes within the same synchronous time slice, and calculate the arithmetic mean of the fusion feature vectors as the average feature value representing the overall state of the base at the current moment; Calculate the algebraic difference between the fused feature vector and the average feature value of each of the distributed monitoring nodes, and separate the difference value that retains the positive and negative polarity information; The difference values ​​of each node are organized into a multidimensional state vector, and the multidimensional state vector is mapped to spatial state coordinate points in the multidimensional feature space in the spatial dimension. A three-dimensional spatial difference model is constructed to characterize the overall degree of force imbalance and abnormal physical location of the base.

8. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, The deformation deviation value and the difference value are compared with the set multi-level thresholds to determine the loosening state and severity level of a single screw. Specifically, this includes: Calculate the relative change ratio of the frequency band feature components relative to the normal energy reference in the multidimensional feature baseline library, and obtain the energy deviation ratio of the current node; A multi-dimensional fault judgment benchmark is set for the comparison of the deformation deviation value and the difference value. The set multi-level thresholds include a yellow warning deformation threshold, a red warning deformation threshold, a yellow warning energy deviation ratio range, and a red warning energy deviation ratio limit, and the red warning deformation threshold is greater than the yellow warning deformation threshold. In the comprehensive rating assessment, by performing the comparison, when it is determined that the deformation deviation value is greater than the yellow warning deformation threshold representing initial loosening and less than or equal to the red warning deformation threshold representing severe loosening, or when the energy deviation ratio is within the set yellow warning energy deviation ratio range, it is determined that the single screw has initial loosening signs and triggers a single-point yellow warning command. When the deformation deviation value is determined to be greater than the red warning deformation threshold indicating severe loosening, or the energy deviation ratio exceeds the set red warning energy deviation ratio limit, the single screw is determined to be in a severely loose state and a single-point red warning command is triggered. When the deformation deviation value is less than or equal to the yellow warning deformation threshold and the energy deviation ratio is less than the lower limit of the set yellow warning energy deviation ratio range, the single screw is determined to be in a normal state. The multi-level thresholds are boundary parameters extracted and calibrated based on the statistical regularities of wind turbine bench destructive test data and historical real fault sample database. The yellow warning deformation threshold is 10 micrometers, the red warning deformation threshold is 15 micrometers, the yellow warning energy deviation ratio range is 10% to 15%, and the red warning energy deviation ratio limit is more than 15%.

9. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, Based on the range of the difference values ​​at each node, the assessment evaluates the degree of uneven force distribution on multiple screws of the wind turbine base and outputs early warning information, specifically including: Extract the maximum and minimum values ​​from the fusion feature vector set corresponding to each of the distributed monitoring nodes within the current synchronization period, and calculate the difference between the two as the range span reflecting the feature space scattering span, i.e. the range of the difference values ​​of each node; When it is determined that the range is greater than the product of the set overall imbalance limit coefficient and the average characteristic value, it is determined that the wind turbine base is subjected to asymmetrical abnormal torsional stress, which causes stress distortion in multiple screws, triggering a multi-point imbalance alarm independent of the single-point warning. Subsequently, a warning data packet carrying the warning level identifier, timestamp, abnormal physical location node number and triggering alarm transient low-level characteristic parameters is generated and pushed to the monitoring terminal as the warning information. When it is determined that the range is less than or equal to the product of the set overall imbalance limit coefficient and the average characteristic value, it is determined that the overall force on the multiple screws of the fan base is in a balanced state. The set overall imbalance limit coefficient is a proportional limit extracted and calibrated based on the statistical regularity of wind turbine bench destructive test data and historical real fault sample database.

10. The method for differential monitoring of fan screws integrating micro-deformation and vibration according to claim 1, characterized in that, The synchronous triggering mechanism is used to acquire the current strain and vibration signals, specifically including: A synchronization trigger command is generated through an internal timer and sent to all the distributed monitoring nodes via a wireless communication link; Upon receiving the synchronization trigger command, each of the distributed monitoring nodes synchronously initiates the data conversion process based on a trigger accuracy constraint of better than 1 millisecond. During the data conversion process, the acquisition cycle of each of the distributed monitoring nodes is locked at a sampling rate of 100Hz. This allows for the accurate capture of the instantaneous stress and vibration states at different physical locations of the base under the same macroscopic time slice, eliminating spatial difference calculation errors caused by time delay, and obtaining the strain signal and vibration signal that are strictly aligned in the time dimension.

Images