Wind power yaw bearing wear state identification method and system based on multi-sensor fusion

Abstract

This invention discloses a multi-sensor fusion method and system for identifying the wear condition of wind turbine yaw bearings, belonging to the field of wind turbine condition monitoring technology. By utilizing multi-source sensor signals from vibration, acoustic emission, electrical parameters, and environmental perception, the operating conditions are divided into static locking, start-up transition, steady-state rotation, and strong random load scenarios. Sensor channels are dynamically activated and multi-dimensional feature parameters are adaptively extracted in each scenario. A nonlinear product-type comprehensive wear index is constructed, including wear torque increment, vibration envelope peak value, acoustic emission impact count, slip ratio peak value, ambient temperature, and current harmonic distortion rate, to achieve multi-source information fusion diagnosis. The wear depth trend is predicted through a system of differential equations. This invention can adapt to extreme wind farm conditions, achieving accurate identification of early wear and prediction of remaining life, significantly reducing operation and maintenance costs, and has the advantages of being non-invasive, highly robust, and highly interpretable.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation technology, specifically a method and system for identifying the wear status of wind turbine yaw bearings using multi-sensor fusion. Background Technology

[0002] With the continuous transformation of the global energy structure, wind power technology has evolved significantly towards higher power output, longer blades, and adaptability to complex environments. As a core component of wind turbines, enabling automatic wind alignment, ensuring energy capture efficiency, and maintaining load balance, the operational stability of the yaw system directly affects the overall service life and power generation efficiency of the turbine. Yaw bearings typically employ large four-point contact ball bearings or crossed roller slewing bearings, which inevitably experience material loss during long-term service due to fretting wear, rolling contact fatigue, abrasive intrusion, or lubrication failure. This wear evolution not only leads to an abnormal increase in the radial and axial clearances of the bearing but also induces a significant decrease in the contact stiffness between the rolling elements and the raceway, which is reflected in the nonlinear fluctuations of frictional torque during yaw operation. In existing industrial monitoring systems, the identification of wear conditions in large slewing bearings mainly relies on single-dimensional sensor monitoring technology, such as arranging vibration accelerometers on the bearing outer ring or yaw gear surface and using the effective value (RMS) or spectral characteristics of the vibration signal for threshold alarms. Specifically, traditional technical solutions are often based on linear signal processing flows. This involves acquiring the raw signal, then sequentially performing bandpass filtering, Fast Fourier Transform (FFT), and characteristic peak extraction. Once the effective value of the vibration velocity exceeds a preset critical threshold (e.g., 4.5 mm / s), the system triggers an early warning. This approach is applicable to continuously operating industrial equipment in a steady state. Its principle lies in the fact that surface spalling or micro-pitting caused by wear excites high-frequency impact signals in specific frequency bands, and frequency domain analysis can identify wear-related characteristic frequencies. However, as wind farms expand into extreme operating environments such as high altitudes and nearshore areas, existing technologies exhibit severe limitations in monitoring yaw bearing wear, and their inherent technical contradictions are becoming increasingly apparent. First, the yaw action of wind turbines has a significantly discontinuous characteristic; the average number of yaws per day is typically less than twenty, and the duration of each yaw is extremely short. This results in an extremely scarce effective dynamic monitoring window throughout the day. In prolonged static locking states, the microscopic degradation signals on the bearing surface are easily obscured by background noise. Existing all-time linear processing algorithms cannot adaptively switch scenes based on the on / off state of yaw movements, resulting in numerous false alarms due to environmental interference during static periods and missed alarms due to signal aliasing during dynamic high-speed yaw periods. A deeper contradiction lies in the coupling and non-stationarity of the monitoring signals. The wind farm ambient temperature fluctuates significantly between -30°C and 50°C, directly causing temperature drift in bearing material stiffness and sensor sensitivity, rendering the fixed threshold judgment logic ineffective. At the same time, the nacelle load fluctuations caused by sudden wind speed changes during yaw induce low-frequency vibration amplitudes far greater than the weak signals in the early stages of wear. Crucially, existing monitoring schemes generally neglect the nonlinear coupling relationship between yaw motor current, speed, and yaw torque.Wear of yaw bearings significantly alters their internal frictional characteristics, which is reflected in the current characteristics of the drive motor. However, in actual operation, the resistance torque generated by wear is often intertwined with the torque of drastically changing wind loads. Due to the lack of effective decoupling mechanisms and deep integration of multi-source information, technicians struggle to accurately isolate the wear-related components from fluctuating motor current signals. This makes it easy for early-stage wear characteristics (such as the micro-pitting stage) to be masked by instantaneous overload signals, hindering accurate prediction of the bearing's entire lifespan. In summary, overcoming the accuracy bottleneck of single-sensor monitoring under extremely complex operating conditions, achieving adaptive identification of dynamic / static yaw bearing scenarios, and effectively decoupling the complex nonlinear relationship between environmental loads and wear characteristics have become key technical challenges and urgent problems to be solved in the wind power operation and maintenance field. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for identifying the wear condition of wind turbine yaw bearings using multi-sensor fusion, in order to solve the technical problems mentioned in the background art, such as low accuracy and high false alarm rate in identifying the wear condition of wind turbines under discontinuous yaw conditions, due to large environmental load interference, severe sensor temperature drift, and deep coupling between wear characteristics and drive torque.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-sensor fusion method for identifying the wear condition of wind turbine yaw bearings includes the following steps: Step 1: Acquire sensor signals collected by the vibration monitoring unit, acoustic emission monitoring unit, electrical parameter monitoring unit, and environmental perception unit. Based on the yaw motor's enable signal, brake pressure status, and motor current RMS value, divide the operating conditions into static lock scenario, start-up transition scenario, steady-state rotation scenario, and strong random load scenario. In each scenario, dynamically activate or suppress the sampling frequency and signal conditioning circuit of the corresponding sensor channel. Step 2: Under different scenarios, adaptive feature extraction is performed on vibration signals and acoustic emission signals using time-frequency analysis, empirical mode decomposition, envelope spectrum analysis, or wavelet packet decomposition, respectively, to obtain vibration envelope peak value, acoustic emission impact count, high-frequency energy ratio of acoustic emission, and slip ratio peak value. At the same time, electromagnetic torque is calculated through electrical parameters, and dynamic equilibrium equations of the yaw system are established. Extended Kalman filter is used to observe aerodynamic load torque as a disturbance term in real time, and wear friction torque increment is decoupled from electromagnetic torque. Step 3: Assign dynamic weighting coefficients to the vibration envelope peak value, acoustic emission impact count, and wear friction torque increment based on the current scenario. Calculate the comprehensive wear index based on the wear friction torque increment, vibration envelope peak value, acoustic emission impact count, slip ratio peak value, ambient temperature, and motor current harmonic distortion rate. The comprehensive wear index uses a product form to nonlinearly fuse the normalized values ​​of the above parameters and includes an exponential decay term due to temperature influence and a multiplicative correction term for wear based on slip ratio and harmonic distortion rate. Step four: Compare the calculated comprehensive wear index with the preset wear threshold. When the comprehensive wear index exceeds the threshold for a preset number of consecutive times, a wear warning is triggered.

[0005] According to the above technical solution, the comprehensive wear index in step three is defined by the following formula: in, This represents the increase in wear friction torque. The peak value of the vibration acceleration envelope. For acoustic emission impact counting, These are scene-related weighting coefficients. , These are the saturation values ​​of the corresponding features under severe bearing wear conditions. To trigger the peak glide rate in the transition scenario, The threshold for the slippage rate of health status. The slip ratio influence coefficient. The current ambient temperature. For reference temperature, The attenuation coefficient is affected by temperature. The harmonic distortion rate of the motor current. This represents the saturation value of the harmonic distortion rate. This is the clearance correction factor.

[0006] According to the above technical solution, the increase in wear friction torque in step two The decoupling process specifically includes: converting the current in the three-phase stationary coordinate system into direct-axis and quadrature-axis components in the synchronous rotating coordinate system, and calculating the electromagnetic torque by combining the magnetic flux linkage and pole pair number of the motor permanent magnet. Establish the dynamic equilibrium equations for the yaw system: in To vary with wear depth The changing frictional resistance torque, For the inflow wind speed and yaw angle The generated aerodynamic load torque The moment of inertia of the yaw system. Yaw angular velocity, The viscous friction coefficient was determined using a small-amplitude pulse excitation test under static locking conditions. Under no-wind, unloaded conditions, a full-stroke yaw test was performed to obtain the baseline friction torque curve under healthy conditions; then, an extended Kalman filter was used as the observer to measure the aerodynamic load torque. As a system disturbance term, the friction torque increment is extracted from the total electromagnetic torque in real time through iterative calculations of the prediction and update steps.

[0007] According to the above technical solution, the method also includes the step of establishing a dynamic model of yaw bearing wear evolution for trend prediction: The cumulative number of yaw cycles N is defined as the discrete time step, with each yaw cycle corresponding to a complete wind adjustment action; δ(N) is set as the equivalent wear depth, and based on the differential form of Archard's wear law and combined with the dynamics of the yaw system, the wear depth evolution rate equation is derived: in For the comprehensive wear coefficient, and Undetermined coefficients related to material hardness and surface morphology; wear depth and bearing clearance increment. The geometric relationship between them is determined by the contact angle. Decide: slip ratio The evolution equation simultaneously considers the effects of frictional torque increment, clearance increment, and temperature: in For the slippage rate of health status, This is the maximum starting torque of the motor. For correction factor, For reference clearance increment, Using the reference temperature; combining the above equation with an extended Kalman filter, for... and The system performs online condition estimation and predicts the wear depth within a preset number of yaw cycles based on the estimated value. When the predicted equivalent wear depth reaches a preset percentage of the bearing's rated hardened layer depth, the system triggers a high-priority warning.

[0008] According to the above technical solution, the working condition division in step one uses scene confidence vectors and hidden Markov models to achieve nonlinear scene interleaving processing: A four-dimensional scene confidence vector is maintained in real time. Each component represents the probability of belonging to a static locking scene, a start-up transition scene, a steady-state rotation scene, and a strong random load scene at the current moment. This confidence is dynamically updated through a hidden Markov model. The observed variables include the yaw enable signal level, brake pressure value, motor speed and its fluctuation rate, and wind speed pulsation intensity. The scene transition probability matrix is ​​obtained by training through historical operation data. At any given time, feature extraction branches corresponding to multiple scenes are run simultaneously. The computational resource allocation of each branch is proportional to the confidence component. At the same time, based on the current confidence vector and transition matrix, the scene most likely to be entered within a preset time window is predicted. When a high-impact scene is predicted to be entered, a high-speed sampling buffer is activated in advance and the acoustic emission channel gain is increased.

[0009] According to the above technical solution, during the dynamic update of the scene confidence vector and the Hidden Markov Model, online adaptive correction of the scene transition probability matrix is ​​adopted: after each complete yaw cycle, the transition probability matrix is ​​re-estimated using the expectation-maximization algorithm based on the difference between the actual observed scene sequence and the model prediction results, so as to gradually approximate the statistical characteristics of the actual working conditions; at the same time, the system maintains a scene history queue of length L. When the time interval between two adjacent scene switches in the queue is less than a preset threshold, it is determined to be scene jitter. At this time, the system will forcibly flatten the components of the confidence vector and restart the observation accumulation process of the Hidden Markov Model to avoid unstable feature extraction caused by frequent scene switches.

[0010] According to the above technical solution, the dynamic weight coefficient allocation in step three is based on the fuzzy inference system to achieve adaptive adjustment: The input variables of the fuzzy inference system include the current operating condition stability index, the signal-to-noise ratio (SNR) estimates of each sensor channel, and the temperature drift correction confidence level. The operating condition stability index is calculated from the statistical variance of the speed fluctuation rate and the current change rate. The SNR estimates are obtained by comparing the effective value of the signal with a preset background noise level. The temperature drift correction confidence level is determined by the degree to which the ambient temperature deviates from the reference temperature. When the ambient temperature is lower than a preset low temperature threshold, the system automatically reduces the weight of the vibration signal and correspondingly increases the weight of the acoustic emission signal. When the SNR of a sensor is detected to be lower than a preset lower limit, the system proportionally redistributes the weight of that sensor to the other sensors.

[0011] According to the above technical solution, it also includes an adaptive threshold correction step for wear status feedback: the central processing unit stores the comprehensive wear index sequence of historical yaw cycles, establishes a sliding window model of wear evolution trend, and uses the comprehensive wear index calculated at the current moment as prior information to correct the logical threshold of subsequent scenario classification; when the comprehensive wear index is in the early wear stage range, the system automatically lowers the current change rate threshold for switching from the start transition scenario to the steady-state rotation scenario, and the reduction magnitude is positively correlated with the degree to which the comprehensive wear index exceeds the health benchmark; at the same time, the system performs temperature normalization preprocessing on the vibration envelope peak value and acoustic emission impact count based on the real-time output of the temperature correction term in the comprehensive wear index, and the normalization coefficient is obtained by linear interpolation of the pre-calibrated sensor sensitivity temperature curve.

[0012] According to the above technical solution, in the adaptive threshold correction step of wear condition feedback, the positive correlation between the reduction magnitude of the current change rate threshold and the degree to which the comprehensive wear index exceeds the health benchmark is determined by the following piecewise function: when Do not downgrade when the value is less than the first threshold; when When the value is between the first threshold and the second threshold, the downward adjustment magnitude is... The relationship is linear; the reduction is equal to the initial threshold multiplied by the preset proportional coefficient, and then multiplied by... The relative amount exceeding the first threshold; when When the threshold is greater than the second threshold, the reduction range is clamped to the maximum reduction range. This piecewise function ensures that the threshold is adjusted smoothly in the early stage of wear and responds quickly in the period of accelerated wear, while avoiding excessive reduction that could lead to misjudgment of the scenario.

[0013] A multi-sensor fusion wind turbine yaw bearing wear condition identification system includes: The data acquisition hardware layer includes a vibration monitoring unit, an acoustic emission monitoring unit, an electrical parameter monitoring unit, and an environmental sensing unit, which are used to collect vibration signals, acoustic emission signals, electrical parameter signals, and ambient temperature signals of the yaw bearing, respectively. The vibration monitoring unit uses a triaxial MEMS accelerometer, installed near the flange bolts connecting the outer ring of the yaw bearing to the nacelle base. The acoustic emission monitoring unit consists of multiple piezoelectric probes evenly spaced inside the yaw gear ring. Each probe is connected to a pre-amplifier with adjustable gain, and uses two parallel channels with different gains to achieve dynamic range expansion. A field-programmable gate array automatically selects the high-gain or low-gain channel for waveform reconstruction based on the real-time signal amplitude. The electrical parameter monitoring unit is integrated into the DC bus and three-phase output of the yaw driver, using a closed-loop Hall current sensor to monitor the instantaneous current and voltage of the three phases in real time. The field signal processing layer consists of a multi-channel synchronous sampling analog-to-digital converter module and a field-programmable gate array, which is used to perform signal pre-filtering, downsampling and real-time feature operator operations; The communication transmission layer uses an industrial Ethernet bus or CANopen protocol to upload the processed feature vectors to the central processing unit, and uses the wireless LoRa protocol as a backup link. When the industrial Ethernet packet loss rate exceeds a preset threshold, it automatically switches to wireless transmission mode. The edge computing server layer, as the central processing unit, integrates an embedded computing kernel, which runs a real-time operating system and executes the wind power yaw bearing wear state identification method based on multi-sensor fusion as described in any one of claims 1 to 9; the edge computing server layer is also equipped with non-volatile memory for establishing a feature database of the entire bearing life cycle.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a monitoring framework capable of adapting to extreme wind farm conditions by spatiotemporal correlation and dynamic fusion of multi-sensor information. Its core advantages are: on-demand activation of sensor channels through a scene-aware mechanism, effectively reducing overall system power consumption and suppressing environmental noise interference during non-yaw periods; effective decoupling of wear torque and random wind load torque through deep coupling of electrical parameters and dynamic models, solving the industry problem of overload signals obscuring early wear characteristics; and the adoption of a unique comprehensive wear index. By fusing multi-source heterogeneous features with a nonlinear product structure, this invention exhibits higher sensitivity and robustness than traditional linear weighting methods. The wear evolution dynamics model derived in this invention enables interpretable prediction of wear trends and significantly improves early warning capabilities.

[0015] The method in this invention does not require invasive modification of the yaw bearing, has extremely high engineering implementation convenience, and can provide accurate data support for predictive maintenance of wind turbine units. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-sensor fusion method for identifying the wear status of wind turbine yaw bearings according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 like Figure 1As shown, the multi-sensor fusion wind turbine yaw bearing wear condition identification method provided by the present invention is built on a hardware system consisting of a vibration monitoring unit, an acoustic emission monitoring unit, an electrical parameter monitoring unit, an environmental perception unit, and a central processing unit. The method includes the following steps.

[0019] The first step is to establish a dynamic activation and suppression mechanism for sensor channels based on operational condition awareness. The central processing unit acquires the enable signal of the yaw motor, the pressure status feedback of the yaw electromagnetic brake, and the effective current value of the yaw actuator in real time. Since the wind power yaw system does not operate continuously but exhibits intermittent characteristics, if all sensors always operate at full power, it will not only consume a large amount of electrical energy but also collect a large amount of invalid environmental noise during non-yaw periods, increasing the burden on subsequent feature extraction. Therefore, it is necessary to dynamically adjust the sensor activation strategy according to the actual operating state of the yaw system.

[0020] When the system detects a low-level yaw enable signal and a locked brake, it identifies the scenario as static lockout S1. In this scenario, the yaw motor has no current input, the brake rigidly connects the yaw bearing to the nacelle base, and there is no relative movement between the bearing rolling elements and raceways, thus no wear characteristic signals are generated. However, scenario S1 provides a clean window free from wear interference, which can be used for system self-testing, benchmark calibration, and structural health monitoring. The central processing unit issues control commands to shut off the power supply to the high-gain amplifier in the acoustic emission monitoring unit to reduce power consumption, keeping only the low-frequency vibration sampling channel open, with the sampling frequency set to 500Hz to capture structural resonance characteristics caused by tower swaying. Simultaneously, the system records benchmark values ​​such as the current ambient temperature and grease temperature for subsequent temperature debiasing of wear indicators. In addition, in the S1 scenario, the system will also actively use the built-in self-test circuit to verify the electrical connectivity of the vibration sensor and acoustic emission sensor, measure the background noise level of each channel, and determine the sensor or cable fault if the noise exceeds the preset threshold. The system will also record the longitudinal and transverse first-order natural frequencies of the tower and compare them with the design values ​​to diagnose foundation settlement or loose connecting bolts.

[0021] When the yaw enable signal transitions from low to high and the brake pressure decreases, the transition scenario S2 is initiated. At this time, the brake is releasing, the yaw motor begins to overcome the static friction torque, and the rolling element transitions from a stationary state to a rolling state. Although this phase is extremely short, typically 200 to 800 milliseconds, it includes the impact response at the moment of brake release, the slippage phenomenon during the static-dynamic friction transition, and the transient stress wave generated when the rolling element first passes over the wear pit. It is the golden window for detecting early wear. At this time, all channels are activated, and the sampling frequency of the acoustic emission monitoring unit instantly increases to 1MHz, while the sampling frequency of the vibration monitoring unit increases to 20kHz. This high sampling rate ensures that the nanosecond-level transient waveform of the acoustic emission signal and the high-frequency impact component of the vibration signal are completely captured.

[0022] When the yaw motor speed reaches more than 90% of the rated speed and the fluctuation rate is less than 5%, it is determined to be a steady-state rotation scenario S3. In this scenario, the yaw system operates at an approximately constant angular velocity, the load is relatively stable, and the vibration and acoustic emission signals exhibit quasi-periodicity. It is suitable to use the classic time-frequency analysis method to extract the periodic wear characteristics such as the rolling element passage frequency and its harmonics.

[0023] When the fluctuating wind speed collected by the wind speed sensor exceeds 15 m / s and the yaw error angle is greater than 10 degrees, it is classified as a strong random load scenario S4. Under strong wind conditions, the aerodynamic load torque amplitude caused by turbulence is large and changes rapidly, which may completely mask the weak torque fluctuations caused by wear. At the same time, tower vibration and blade vibration under wind load excitation will be transmitted to the yaw bearing through the structure, introducing strong noise into the sensor signal. Therefore, in this scenario, a more robust feature extraction method is needed, and load compensation should be performed based on feedforward wind field information.

[0024] A complete yaw cycle is defined as the entire process from the moment the yaw motor receives the start command and begins to rotate until the yaw action is completed, the motor stops rotating, and the brake relocks. The system uses the yaw cycle as the basic unit of analysis to ensure that wear indicators are comparable across different yaw actions, because the rotation angle and duration corresponding to each yaw cycle may be different, but all are measured on a uniform scale based on a complete wind adjustment.

[0025] Furthermore, the four scenarios mentioned above are not strictly abrupt transitions, but rather have overlapping transition zones. The system introduces a scenario confidence vector, which probabilistically describes the likelihood of belonging to each scenario at the current moment. This confidence is dynamically updated through a Hidden Markov Model, with observed variables including yaw enable signal level, brake pressure value, motor speed and its fluctuation rate, and wind speed pulsation intensity. The scenario transition probability matrix is ​​obtained through training on historical operating data. At any given time, the system simultaneously runs feature extraction branches corresponding to multiple scenarios, but the computational resource allocation for each branch is proportional to the confidence level, thus avoiding the loss of critical information due to misjudgment of a single scenario when operating conditions change rapidly.

[0026] The second step involves adaptive extraction of multi-dimensional feature parameters. Under different judgment scenarios, the central processing unit adjusts the operator parameters of the feature extraction algorithm. The purpose of feature extraction is to transform the raw sensor signal into a quantitative indicator that reflects the wear state. Since the signal characteristics differ significantly under different scenarios, an adaptive strategy must be employed.

[0027] In the steady-state rotation scenario S3, the system employs Short-Time Fourier Transform (SFT) to perform time-frequency analysis on the vibration signal. A Hamming window is selected as the window function, with a window length of 256 points and an overlap rate of 50%. The SFT divides the signal into multiple short time intervals by sliding the window function along the time axis, and performs a Fourier transform on each interval to obtain a two-dimensional time-frequency distribution map. The 256-point window length corresponds to a 64-millisecond time window at a 4kHz sampling rate, which can distinguish the periodic changes in the frequency of the rolling body, typically ranging from 1Hz to 10Hz. The 50% overlap rate ensures a smooth transition along the time axis, and the Hamming window has a low sidelobe level, reducing spectral leakage and thus obtaining high-time-resolution transient impact characteristics.

[0028] Under the strong random load scenario S4, considering the interference of wind-induced vibration noise, the system adjusts the window length of the short-time Fourier transform to 1024 points to improve frequency resolution, and combines it with empirical mode decomposition to decompose the original signal into several intrinsic mode functions. Empirical mode decomposition is an adaptive signal processing method that can decompose nonlinear and non-stationary signals into a series of intrinsic mode functions with different characteristic time scales, each function representing an oscillation mode in the signal. Its decomposition process does not require pre-selection of basis functions, but rather iteratively filters and gradually extracts high-frequency components from the original signal. Subsequently, components containing the wear characteristic frequency band and whose energy proportion is greater than 5% of the total energy are selected for envelope spectrum analysis. The wear characteristic frequency band is 200Hz to 2000Hz. Envelope spectrum analysis first performs a Hilbert transform on the signal to obtain the analytic signal, then takes the modulus to obtain the envelope, and finally performs a Fourier transform on the envelope. The envelope spectrum can highlight the low-frequency modulation information in the high-frequency carrier, which is very effective for detecting the periodic impact generated when a rolling body passes over pitting corrosion.

[0029] Simultaneously, for the acoustic emission signal, the system calculates the impact count, amplitude distribution, and energy rate, and uses wavelet packet decomposition to extract the energy proportion of the high-frequency detail coefficients in the third and fourth layers as sensitive indicators of the early stage of wear. Acoustic emission impact refers to a transient waveform exceeding a threshold; its count reflects the frequency of stress wave events per unit time, while the energy rate characterizes the impact intensity. Wavelet packet decomposition simultaneously decomposes the signal into low-frequency approximations and high-frequency details. The third and fourth layers correspond to the frequency bands of 50kHz to 100kHz and 100kHz to 200kHz, respectively. An increase in the energy proportion within these frequency bands is typically associated with microcrack propagation and pitting initiation.

[0030] The third step is to construct a nonlinear decoupling model of the yaw bearing wear torque and the environmental load torque.

[0031] The central processing unit acquires the three-phase instantaneous current and voltage of the yaw motor through the electrical parameter monitoring unit and calculates the motor's electromagnetic torque. The calculation method is as follows: First, the current in the three-phase stationary coordinate system is transformed into direct-axis and quadrature-axis components in the synchronous rotating coordinate system using the Clarke and Parker transformations. Then, combined with the permanent magnet flux linkage and pole pair number of the motor, the specific value of the electromagnetic torque is calculated. The physical basis of this calculation method is the electromagnetic relationship of the permanent magnet synchronous motor: the electromagnetic torque is proportional to the product of the quadrature-axis current and the permanent magnet flux linkage, and is also affected by the direct-axis current, i.e., the salient-pole effect. The Clarke transformation converts the three-phase stationary quantities into two-phase stationary quantities, and the Parker transformation further converts them into a coordinate system that rotates synchronously with the rotor, thus converting the AC quantity into a DC quantity for easier calculation.

[0032] A dynamic equilibrium equation for the yaw system is established, which expresses the electromagnetic torque as the sum of four terms: frictional drag torque varying with wear depth, aerodynamic load torque generated by the inflow wind speed and yaw angle, inertial torque, and viscous friction torque. This equation originates from rigid body rotational dynamics: the driving torque equals the drag torque plus the inertial torque. The frictional drag torque is the core variable to be monitored in this invention; it increases with bearing clearance, raceway pitting, or deterioration of lubrication. The aerodynamic load torque is the result of the eccentric load generated by the wind acting on the impeller, transmitted through the tower to the yaw bearing; its magnitude is proportional to the square of the wind speed and the sine of the yaw error angle. The inertial torque exists only when the yaw angular velocity changes. The viscous friction torque is proportional to the first power of the angular velocity. The viscous friction coefficient is calibrated through a small-amplitude pulse excitation test under static scenario S1: a known pulse torque is applied, the angular velocity decay curve is measured, and the coefficient is calculated using the logarithmic decay method. This calibration method utilizes the relationship between the logarithmic decay rate and the damping ratio of free vibration. Once the moment of inertia is known, the coefficient of viscous friction can be calculated.

[0033] A baseline friction torque curve was obtained by conducting an unloaded yaw test in a windless environment. Unloaded yaw refers to a yaw test conducted under conditions where the blades are removed or the impeller is locked to prevent wind load. In this state, the aerodynamic load torque is zero, and the electromagnetic torque is entirely used to overcome frictional resistance and inertial forces, thus allowing direct extraction of the baseline friction torque curve under healthy conditions. Then, an extended Kalman filter (EPF) algorithm was used to observe the wear friction torque in real time. This algorithm treats the aerodynamic load torque as a system disturbance term and, through iterative prediction and update steps, extracts the friction torque increment caused by increased bearing clearance and raceway pitting from the total electromagnetic torque. The EPF is a standard method for nonlinear system state estimation. Its prediction step uses the state estimate from the previous moment and the system model to deduce the prior estimate for the current moment; the update step uses the measured value at the current moment to correct the prior estimate, obtaining the posterior estimate. In this invention, the aerodynamic load torque is considered a known disturbance input, obtained from lidar feedforward or wind speed estimation. The EPF can stably estimate the wear friction torque even in noisy environments.

[0034] Furthermore, to more accurately characterize the relationship between wear and frictional torque, this invention introduces a geometric constraint model for bearing clearance. When wear leads to material loss on the raceway surface, the clearance between the rolling elements and the raceway increases, causing a shift in the load distribution angle of the rolling elements, which in turn alters the waveform quality of the motor current. Specifically, in addition to the fundamental component, the current waveform will exhibit harmonics of specific orders, and there is an approximate quadratic function relationship between the harmonic distortion rate and the clearance value. By pre-calibrating this functional relationship through offline experiments, the clearance increment can be calculated based on the current harmonic distortion rate during operation, and this can be used as an auxiliary verification for estimating wear frictional torque.

[0035] The fourth step is to implement multi-source information fusion diagnosis based on scenario confidence. The system assigns dynamic weight coefficients to the feature parameters of different sensors according to the current operating scenarios S2, S3, and S4. In the static locking scenario S1, the bearing has no relative motion and does not generate wear characteristics; therefore, S1 does not participate in wear fusion diagnosis but is only used for benchmark calibration and structural health monitoring. The principle of weight allocation is: in each scenario, the sensor feature most sensitive to wear characteristics and least affected by environmental noise is assigned a higher weight. For example, in the start-up transition scenario, the acoustic emission signal is most sensitive to transient impacts, so it has the highest weight; in the steady-state rotation scenario, the decoupled wear torque increment directly reflects the change in frictional resistance and is not affected by the amplitude saturation of vibration and acoustic emission, so it has the highest weight; in the strong random load scenario, the vibration signal may be submerged by wind noise, while the high-frequency components of acoustic emission are less affected by wind load, so the weights of the two are roughly equal. The specific weight allocation is as follows: In the initial transition scenario S2, the weight of the acoustic emission energy characteristic value is 0.6, and the weight of the vibration impact pulse amplitude is 0.4; in the steady-state rotation scenario S3, the weight of the decoupled wear torque increment is 0.5, the weight of the vibration acceleration envelope peak value is 0.3, and the weight of the acoustic emission impact count is 0.2; in the strong random load scenario S4, the weight of the decoupled wear torque increment is 0.4, the weight of the envelope spectrum peak value of the characteristic frequency band after empirical mode decomposition is 0.4, and the weight of the proportion of high-frequency acoustic emission energy is 0.2.

[0036] The above weighting is the basic configuration. In a preferred embodiment, the weighting coefficients are further fine-tuned based on the real-time signal-to-noise ratio and temperature drift confidence level. For example, when the ambient temperature is below -20°C, the sensitivity of the vibration sensor may decrease, while the piezoelectric element of the acoustic emission sensor performs relatively stably at low temperatures. In this case, the system will automatically increase the proportion of acoustic emission weighting.

[0037] The core originality of this invention lies in the establishment of a comprehensive wear index. This index integrates wear torque increment, vibration envelope peak value, acoustic emission impact count, slip ratio anomaly peak value, bearing clearance harmonic distortion rate, and the influence of ambient temperature. Its expression is as follows: Among them, parameters This is the increase in wear friction torque, measured in Newton-meters, obtained through decoupling via extended Kalman filtering in the third step. Physically, it represents the increase in the frictional resistance torque of the yaw bearing relative to its healthy baseline value. The saturation value is determined through bench testing: it is measured when the pitting area of ​​the bearing raceway reaches 5% of the total area or the wear depth reaches 80% of the hardened layer depth. Value. For a typical 5MW wind turbine, It is approximately 50 to 80 Newton-meters.

[0038] parameter This represents the peak value of the vibration acceleration envelope, expressed in meters per second squared. The peak value reflects the impact amplitude generated when the rolling element passes over a wear pit. This is the saturation value, corresponding to the peak envelope when the bearing is severely worn, typically 3 to 5 meters per second squared.

[0039] parameter Acoustic emission impact count is defined as the number of events exceeding a threshold within a unit of time (typically 1 second). (Healthy bearing) It is close to 0, and gradually increases with the appearance of micropitting. This is the saturation value, corresponding to the count during severe peeling, typically between 200 and 500.

[0040] parameter These are scenario-related weighting coefficients, satisfying that the sum of the three is 1. For specific values, please refer to the aforementioned allocation rules.

[0041] parameter This represents the peak slip ratio measured during the initial transition scenario. Slip ratio is defined as the theoretical speed minus the actual speed, divided by the theoretical speed. The theoretical speed is calculated from the motor current frequency and the number of pole pairs. The peak slip ratio of a healthy bearing is typically less than 0.05; when wear causes rolling element resistance, the peak slip ratio can increase to 0.12 to 0.20. The health threshold is set to 0.05. This is the slip ratio influence coefficient, obtained through experimental regression, with a typical value of 0.8. This multiplicative factor... The value ranges from 1.0 (no slip anomaly) to approximately 2.2 (severe blockage).

[0042] parameter The current ambient temperature is expressed in degrees Celsius. For reference temperature, take 20°C. This is the temperature-dependent attenuation coefficient. This coefficient originates from the viscosity-temperature characteristics of the grease: at low temperatures, viscosity increases, leading to a false increase in frictional torque, but this false increase should decrease exponentially with increasing temperature. For low-temperature greases specifically designed for wind power applications, The typical value is 0.002℃. -1 Therefore, the exponent term At low temperatures, it is less than 1, for It has a suppressive effect, that is, it subtracts the pseudo-increase caused by low temperature; at high temperature, it is greater than 1, because the lubricating grease becomes thinner and the friction torque decreases at high temperature, so the wear index needs to be adjusted upward to maintain consistency.

[0043] parameter The harmonic distortion rate of motor current is defined as the square root of the sum of the squares of the effective values ​​of all harmonics divided by the effective value of the fundamental frequency, and is usually expressed as a percentage. (Healthy bearings) Approximately 3% to 5%; as the clearance increases, It could rise to 15% to 25%. The saturation value is set to 25%. This is the clearance correction factor, taken as 0.5. This multiplicative factor... The value range is from 1.0 to 1.5.

[0044] The structure of this equation has the following originality. First, it is a product form rather than a nonlinear weighting: when only one of the three factors—wear torque increment, peak vibration, and acoustic emission count—deviates from the normal value... It shows a moderate increase; if both deviate simultaneously, the product effect causes The increase is greater than the sum of the two; if all three deviate simultaneously, This will increase dramatically. This aligns with the physical law of accelerated wear deterioration; the simultaneous appearance of multiple independent symptoms indicates that wear has entered a rapid development phase. Second, exponential temperature correction: Traditional methods, using linear subtraction or simple table lookup, cannot accurately reflect the exponential change in grease viscosity with temperature. The exponential decay form of this invention is consistent with the theoretical curve of the viscosity-temperature formula, resulting in higher correction accuracy. Third, multiplicative correction of slip ratio and harmonic distortion: Abnormal slip ratio and current distortion are indirect symptoms of wear, but they have a synergistic relationship with direct symptoms. Treating them as multiplicative factors rather than additive terms means that when direct symptoms are already obvious, the presence of indirect symptoms will be further amplified. This allows for earlier warnings; conversely, if direct signs are weak but indirect signs are obvious, It will also rise moderately, allowing for the detection of early anomalies.

[0045] This comprehensive wear and tear index The entire calculation process is completed in the central processing unit. Real-time values ​​of each parameter are obtained through online measurement or extended Kalman filtering estimation, while saturation values ​​are pre-stored in the database through offline bench tests. Weighting coefficients are dynamically selected based on the scenario confidence level. The wear threshold was exceeded for three consecutive yaw cycles. When this occurs, a wear warning is triggered. Threshold The initial value is set to 0.65, which is higher than the traditional linear weighting threshold of 0.6, to match the sensitivity of the product structure.

[0046] The fifth step is to establish an adaptive threshold correction mechanism for wear state feedback. The central processing unit stores historical wear evolution trends, and the estimated wear state value calculated at the current moment will serve as prior information, influencing the logical threshold for scene classification at the next moment. The purpose of this step is to enable the system to "remember" historical wear levels and automatically adjust the sensitivity of subsequent judgments based on wear development. For example, when a bearing has entered the micropitting stage, its current change rate and speed fluctuation characteristics may differ from those in the healthy state. The threshold previously used to distinguish between S2 and S3 may no longer be applicable and needs to be appropriately lowered to maintain the accuracy of scene classification.

[0047] If the current condition is identified as early stage of wear, that is Between 0.3 and 0.5, the system will automatically lower the current change rate threshold for switching from the initial transition scenario S2 to the steady-state rotation scenario S3. The initial threshold for the current change rate is defined as 50 amperes per second, calibrated based on 5% of the yaw motor's rated current. The reduction increment is 5%, but never lower than 25 amperes per second. The physical basis for this reduction logic is that in the early stages of wear, the bearing friction torque increases, causing the current rise rate during motor startup to slow down, i.e., the current change rate decreases. If the original high threshold is maintained, the system may mistakenly believe that the motor has not yet reached its rated speed, thus delaying the scene switch and affecting the accuracy of subsequent feature extraction. By adaptively lowering the threshold, it ensures that the scene classification always matches the actual dynamic response.

[0048] Meanwhile, the system is based on the aforementioned comprehensive wear index The temperature correction term in the code compensates for the sensor signal in real time. In fact, The exponential temperature term has already had the temperature effect deducted, requiring no additional processing. However, in the calculation... Previously, all original features (such as vibration envelope peak values) should also be temperature normalized to eliminate the influence of sensor sensitivity drift with temperature. The normalization method uses a pre-calibrated sensitivity-temperature curve and is achieved through linear interpolation.

[0049] The sixth step involves establishing a kinetic model for the wear evolution of the yaw bearing for trend prediction. This invention proposes a kinetic model for wear evolution to predict the trend of wear depth with the number of yaw cycles, thus providing an early warning before wear reaches a critical value. This model is based on Arcard's wear law and incorporates the aforementioned measurable parameters (friction torque increment). slip ratio clearance increment (etc.), established wear depth yaw period number The differential equation.

[0050] definition Let be the cumulative number of yaw cycles (dimensionless), where each yaw cycle corresponds to one complete yaw adjustment maneuver. The equivalent wear depth (mm) evolves according to the wear energy dissipation rate. Based on the differential form of Arcard's law and combined with the dynamics of the yaw system, the equation for the wear depth evolution rate is derived: in The overall wear coefficient can be calibrated through bench testing; These are undetermined coefficients related to material and surface morphology. The equation reveals a positive correlation between wear rate and the increase in frictional torque and slip ratio, while the denominator contains... and The term reflects the nonlinear effect of increased surface roughness on the coefficient of friction.

[0051] Wear depth and bearing clearance increment The geometric relationship between them can be approximated as follows: ,in This refers to the contact angle. Increased clearance leads to higher current harmonic distortion. Ascending according to the quadratic polynomial relation: in Harmonic distortion rate in a healthy state The fitting coefficients are denoted as .

[0052] slip ratio The evolution of friction torque depends not only on the increase in friction torque, but also on clearance and temperature, and its expression is: in For the slippage rate of health status, This is the maximum starting torque of the motor. For correction factor, For reference clearance increment, The reference temperature is 293K.

[0053] Combining the above equations, we obtain the following set of equations for wear evolution dynamics: in This is the basic friction increment function without considering the clearance effect (which can be obtained through pure wear tests). is the clearance coupling coefficient.

[0054] All parameters of this dynamic model can be identified through offline bench tests or field data. In actual operation, the central processing unit uses an extended Kalman filter to analyze the state variables. and Perform online estimations and recursively predict the future based on the current estimates. Wear depth within one yaw cycle. When predicted When the bearing's rated hardened layer depth is expected to reach 15% within the next maintenance cycle, the system issues a high-priority warning and provides a predicted remaining service life and its confidence interval. Compared to purely data-driven black-box methods, this model offers stronger interpretability and extrapolation capabilities, making it particularly suitable for scenarios where samples are scarce in the early stages of wear.

[0055] Furthermore, the vibration monitoring unit employs a triaxial MEMS accelerometer, mounted near the flange bolts connecting the outer ring of the yaw bearing to the nacelle base. The sensor has a sensitivity of 100mV / g and a linear amplitude-frequency response range covering 0.5Hz to 12kHz. The analog signal output from the sensor is transmitted to the signal conditioning circuit via a twisted-pair shielded cable with a moisture-proof shielding layer. The signal conditioning circuit includes a fourth-order Butterworth low-pass filter with a cutoff frequency set at 10kHz to eliminate high-frequency aliasing noise. The Butterworth filter is characterized by a flat, ripple-free amplitude-frequency response within the passband and an acceptable nonlinear phase response. The fourth-order filter is implemented by cascading two second-order sections, with an attenuation slope of -80dB per decibel. The 10kHz cutoff frequency is higher than the wear characteristic frequency band (maximum 2kHz) and lower than half the sampling frequency (if the sampling frequency is 20kHz, the Nyquist frequency is 10kHz), perfectly meeting the anti-aliasing requirements.

[0056] Furthermore, the acoustic emission monitoring unit consists of four piezoelectric acoustic emission probes evenly spaced inside the yaw gear ring, with a center frequency of 150kHz, fixed by magnetic adsorption bases. Each probe is connected to a preamplifier with adjustable gain, amplification ranging from 20dB to 60dB. Feature extraction of the acoustic emission signal includes calculating the variance, skewness, and peak factor of the waveform envelope, used to characterize the impact dispersion of the rolling element as it passes through pitting. Skewness describes the waveform asymmetry: burst-type acoustic emission signals typically exhibit a rapid rise and slow fall pattern, with a positive skewness; continuous noise has a skewness close to zero. Peak factor is the ratio of peak value to RMS value; burst-type impacts have a high peak factor (greater than 5), while continuous noise has a low peak factor (less than 3). These two parameters can distinguish between burst signals generated by pitting and continuous friction signals caused by poor lubrication.

[0057] Furthermore, the electrical parameter monitoring unit is integrated into the DC bus and three-phase output terminals of the yaw drive, employing a closed-loop Hall effect current sensor to monitor the starting current waveform of the yaw motor. By calculating the slope of the rising edge of the starting current and the delay in the appearance of the peak current, the abnormal increase in static friction torque due to bearing wear is identified. The rising slope of the starting current is inversely proportional to the static friction torque: when wear causes an increase in static friction torque, the motor requires more time to overcome static friction, the rising current becomes slower, and the appearance of the peak current is delayed. The system uses the slope and delay under healthy conditions as a benchmark and monitors their relative changes in real time.

[0058] Furthermore, the decision-making logic for switching from static scenario S1 to the start-up transition scenario S2 in the first step also includes: after the yaw driver receives the yaw command pulse issued by the host computer, the central processing unit activates the high-speed sampling buffer within 100ms, with a pre-trigger depth set to 50ms, to fully capture the structural vibration response of the brake release transient. Pre-triggering means that data acquisition and storage in the buffer begins before the rising edge of the enable signal arrives. Since the transient response of the brake release may occur within a few milliseconds after the enable signal transition, without pre-triggering, this transient information will be lost. The 50ms pre-trigger depth ensures complete capture.

[0059] Furthermore, the calculation of frictional resistance torque in the third step also incorporates a geometric constraint model for the radial and axial clearances of the bearing. When wear causes an increase in clearance, the rolling element load distribution angle shifts. The system establishes a mapping relationship between the current harmonic characteristics and the bearing clearance value by calculating the harmonic distortion rate of the motor current at a specific mechanical angle. This mapping relationship is achieved through a quadratic polynomial obtained from offline experiments. The specific method for establishing the mapping relationship is as follows: on a test bench, the bearing clearance is gradually increased from zero to the failure value, while simultaneously recording the current harmonic distortion rate under different clearances. The result is obtained through least squares fitting. A quadratic polynomial of the form, where This represents the clearance increment. Measured during operation. Substituting this into the polynomial allows us to deduce the gap increment.

[0060] Furthermore, the comprehensive wear index in step four The calculation process incorporates built-in temperature debiasing, achieved through an exponential temperature decay term. The system calls upon data from the Pt100 temperature sensor in the environmental sensing unit in real time; without additional table lookups, it directly substitutes the data into the temperature correction term to automatically eliminate the pseudo-torque increase component caused by the increased viscosity of the grease at low temperatures. This exponential decay term is mathematically consistent with the viscosity-temperature formula for grease, but the coefficient... It was obtained through regression analysis of field data, which better reflects the characteristics of the lubricating grease in actual units.

[0061] Furthermore, the method also includes a step of predicting the long-term evolution trend of the identification results, namely the sixth step of the wear evolution dynamics model. The central processing unit uses this model to analyze the comprehensive wear index. and wear depth Perform time series forecasting. When the 90th percentile of the predicted remaining useful life is below a safety threshold, such as 500 yaw cycles, the system issues a high-priority warning.

[0062] This invention also provides a multi-sensor fusion system for identifying the wear condition of wind turbine yaw bearings. The system comprises: a data acquisition hardware layer, including a vibration acceleration sensor, an acoustic emission probe, a Hall current / voltage module, and a temperature probe; a field signal processing layer, consisting of a high-precision multi-channel synchronous sampling ADC module and a field-programmable gate array, responsible for pre-filtering, downsampling, and real-time feature operator operations; a communication transmission layer, using an industrial Ethernet bus or CANopen protocol to upload the processed feature vectors to a central processing unit; and an edge computing server layer, which acts as the central processing unit to execute the aforementioned wear condition identification method.

[0063] The edge computing server layer integrates an embedded computing kernel, which runs a real-time operating system, ensuring that the execution latency of commands switching between operating conditions is less than 10ms. The server layer is also equipped with non-volatile memory for building a feature database of the entire bearing lifecycle.

[0064] Example 2 In a preferred embodiment of the present invention, the triaxial accelerometer in the vibration monitoring unit employs MEMS technology, with a full-scale range of ±50g and a nonlinearity of less than 0.1%FS. The sensor integrates a digital compensation circuit, capable of real-time offsetting zero-bias shifts caused by changes in the direction of the gravity vector. In static scenario S1, the system utilizes the low-frequency characteristics of this sensor to monitor the longitudinal and lateral first-order natural frequencies of the wind turbine tower. If the frequency deviation exceeds ±5%, it is determined that there is an anomaly in the foundation or tower connection. At this time, the system reduces the confidence level of the wear diagnosis to avoid structural vibration mixing with bearing wear characteristics. Specifically, after a structural anomaly is confirmed, the system multiplies the calculated result of the comprehensive wear index $\Psi$ by an attenuation factor less than 1, such as 0.5, and simultaneously notes in the output report that "structural anomalies may affect the accuracy of wear diagnosis," reminding maintenance personnel to prioritize addressing structural issues.

[0065] In a preferred embodiment of the present invention, the conditioning circuit of the acoustic emission monitoring unit employs dynamic range extension technology. When the yaw bearing is in the middle or late stage of wear, the amplitude of the sudden acoustic emission signal generated is huge, which can easily lead to saturation of traditional circuits. The present invention samples simultaneously through two parallel channels with different gains, where the low-gain channel is 10dB and the high-gain channel is 40dB. The field-programmable gate array automatically reconstructs the waveform based on the real-time amplitude of the signal, ensuring that the weak signal generated by micro-pitting and the large-amplitude signal generated by severe peeling can be restored without distortion. The specific logic of waveform reconstruction is as follows: after the two signals are aligned with the same time delay, the amplitudes of the two signals at the same moment are compared. If the high-gain channel is not saturated, that is, the amplitude is less than 95% of the full scale, the output of the high-gain channel is used; if the high-gain channel is saturated, the switch is made to the low-gain channel, and the output of the low-gain channel is multiplied by the corresponding gain compensation coefficient, that is, amplified by the factor corresponding to 10dB, to restore the original signal amplitude.

[0066] In a preferred embodiment of the present invention, in the steady-state rotation scenario S3, the feature extraction algorithm further incorporates cepstral analysis to identify periodic interference composed of the gear meshing frequency and its harmonics of the yaw reducer. The cepstral analysis is specifically implemented by first calculating the logarithm of the vibration signal power spectrum, then performing an inverse Fourier transform to obtain the cepstral spectrum. The gear meshing frequency and its harmonics appear as high-amplitude narrow-band peaks at specific cepstral frequencies in the cepstral spectrum. By setting a narrow-band notch filter in this cepstral frequency range, the influence of gearbox vibration on bearing wear identification can be eliminated, thus highlighting the energy peak of the bearing rolling element passing frequency. The physical significance of the cepstral spectrum lies in its conversion of the periodic components in the spectrum into single spectral lines in the cepstral spectrum, making it highly suitable for separating gear meshing vibrations containing the fundamental frequency and multiple harmonics. Setting the width of the notch filter to twice the peak half-width and the notch depth to -20dB effectively suppresses gear interference without damaging bearing information in adjacent cepstral frequency regions.

[0067] In a preferred embodiment of the present invention, the nonlinear decoupling model in the third step incorporates wind data from a lidar sensor located on the top of the nacelle for feedforward compensation. The lidar sensor pre-acquires the inflow wind field distribution 100 meters in front of the unit and calculates the eccentric load torque acting on the impeller using an impeller aerodynamic model. This allows the dynamic model to predict the changing trend of the aerodynamic load torque before yaw occurs. This feedforward information, combined with the yaw motor current feedback, enables the decoupling algorithm to maintain stable extraction of the wear drag torque even during sudden wind speed changes. The principle of feedforward compensation is that extended Kalman filtering typically relies solely on feedback, i.e., measured values, to correct state estimates. When external disturbances, i.e., sudden changes in aerodynamic loads, occur, the feedback lags. By introducing lidar feedforward, the system can input the disturbance estimate into the state prediction equation before the disturbance actually affects the yaw system, thereby adjusting the state estimate in advance and reducing errors.

[0068] In a preferred embodiment of the present invention, the fusion weight coefficient allocation in the fourth step is based on a fuzzy inference system. The inputs to this system are "current operating condition stability," "sensor signal-to-noise ratio estimate," and "temperature drift correction confidence level." When the sensor is at an extreme low temperature of -30°C, the system automatically identifies that the temperature feedback from the platinum resistance thermometer is in the nonlinear sensitive region. At this time, it automatically lowers the weight of the vibration signal in the fusion model and correspondingly increases the weight of the acoustic emission signal. An example of the rule base of the fuzzy inference system is as follows: if the temperature is extremely low and the vibration signal-to-noise ratio is low, the vibration weight is reduced and the acoustic emission weight is increased. The input quantities are fuzzified using a membership function, and then fuzzy inference and defuzzification are performed to obtain the precise weight adjustment amounts.

[0069] In a preferred embodiment of the present invention, the wear condition identification method executed by the system also has a self-healing calibration function. During the low wind speed period each month, i.e., when the wind speed is below 3 m / s, the system automatically triggers a full-stroke no-load yaw diagnosis, with a yaw angle range of 0° to 360°. The friction torque extracted at this time is automatically updated with the vibration reference value to eliminate long-term drift caused by grease aging or grease redistribution, ensuring that the online monitoring system is always in an accurate calibration state. In addition to timed triggering, the self-healing calibration can also be triggered manually by maintenance personnel remotely, or automatically triggered after a long period of no warning but a significant change in environmental conditions, such as grease replacement. During the calibration process, the system temporarily switches to "calibration mode," ignoring wear warning outputs, and automatically resumes normal monitoring after calibration is completed.

[0070] In a preferred embodiment of the present invention, the communication transmission layer uses the LoRa wireless protocol as a backup link. When physical damage or electromagnetic interference causes the packet loss rate to exceed 5% in the industrial Ethernet, the system automatically switches to a narrowband low-power wireless transmission mode. At this time, the edge computing server layer automatically compresses the feature vector, transmitting only the highest priority comprehensive wear index and key alarm codes to ensure uninterrupted operation of the monitoring service. The compression strategy is to compress the feature vector, which originally required hundreds of bytes and included statistics on all vibration, acoustic emission, and electrical parameters, into a single vector containing only... The message is an ultra-short message consisting of a 4-byte floating-point value, a 1-byte warning flag, and a 4-byte timestamp, to accommodate the narrow bandwidth limitations of LoRa, which are typically from a few hundred bps to tens of kbps.

[0071] As a preferred embodiment of the present invention, the comprehensive wear index With wear depth The quantitative relationship between them is calibrated using a pre-constructed nonlinear mapping function based on offline bench test data. This mapping function employs a radial basis function neural network to... Ambient temperature and yaw speed Input: Equivalent wear depth Radial basis function (RBF) neural networks are a type of feedforward network where hidden layer neurons use Gaussian radial basis functions as activation functions, and the output layer uses linear weighting. During network training, the centers of the radial basis functions are first determined through clustering, and then the output weights are calculated using the least squares method. Because RBF networks have local approximation capabilities, their fitting effect on nonlinear mappings is superior to traditional multilayer perceptrons. In actual operation, the system uses this mapping function to inversely calculate the equivalent wear depth. If the equivalent wear depth reaches 15% of the bearing's rated hardened layer depth, the bearing is determined to have entered a dangerous stage of fatigue spalling, and the system will immediately output a high-priority maintenance recommendation form with a geographic location tag and equipment number to the central control center. The equivalent wear depth is a virtual quantity; it is not directly equal to the actual pitting depth, but rather... This is mapped to a percentage of the wear process from 0% to 100%. The rated hardened layer depth is a bearing design parameter, typically 2mm to 4mm. 15% of this corresponds to 0.3mm to 0.6mm, at which point pitting visible to the naked eye has appeared on the raceway surface, requiring replacement.

[0072] Example 3 The multi-sensor fusion wind turbine yaw bearing wear condition identification system of this invention consists of a data acquisition hardware layer, a field signal processing layer, a communication transmission layer, and an edge computing server layer in terms of hardware architecture. Specifically, the data acquisition hardware layer includes vibration monitoring units, acoustic emission monitoring units, electrical parameter monitoring units, and environmental sensing units arranged at key physical nodes of the yaw system. The vibration monitoring unit uses a triaxial MEMS accelerometer, which is installed near the flange bolts connecting the outer ring of the yaw bearing to the nacelle base, and has high sensitivity response capabilities in the radial, axial, and tangential directions.

[0073] Furthermore, the acoustic emission monitoring unit consists of four piezoelectric probes equally spaced inside the yaw gear ring, with the probe center frequency calibrated at 150kHz. Each probe is equipped with a pre-gain adjustable amplifier and employs a two-way parallel dynamic range extension technology.

[0074] In the on-site signal processing layer, the system utilizes a high-precision multi-channel synchronous sampling ADC module to digitize the analog signal. The vibration signal passes through a fourth-order Butterworth low-pass filter with its cutoff frequency precisely set at 10kHz. Simultaneously, the electrical parameter monitoring unit acquires the three-phase instantaneous current and voltage signals of the motor in real time through a closed-loop Hall current sensor.

[0075] The execution logic of this invention is first built upon a dynamic activation mechanism for sensor channels based on operational condition perception. The central processing unit, through real-time polling of the yaw enable signal, brake pressure feedback, and the effective value of the motor current, divides the operating conditions into four scenarios: static lockout (S1), start-up transition (S2), steady-state rotation (S3), and strong random load (S4). In S1, the system maintains only the low-frequency vibration channel and records the baseline value. In S2, all channels operate at the highest sampling rate. In S3 and S4, the time-frequency analysis parameters are adjusted according to a preset strategy.

[0076] A key technical aspect of this invention lies in the nonlinear decoupling of the yaw bearing wear torque and the environmental load torque. The central processing unit calculates the electromagnetic torque based on electrical parameters. To extract the frictional torque increment caused by wear from the complex total torque, this invention establishes a complete dynamic equilibrium equation for the yaw system. Aerodynamic load torque is predicted by introducing lidar-feedforward wind field data, and this aerodynamic load is observed in real time as an external disturbance using an extended Kalman filter algorithm, ultimately calculating the wear frictional torque increment. In the specific implementation of this algorithm, the state vector includes the wear frictional drag torque and the yaw angular velocity, and the optimal estimate is obtained through iterative prediction and updating.

[0077] During the multi-source information fusion diagnosis phase, the system assigns dynamic weights based on the current operating conditions (S2, S3, and S4) and calculates the unique comprehensive wear index of this invention. Taking a 5MW offshore wind turbine as an example, under steady-state rotation scenario S3, the system measured the increase in wear torque. =12 N·m, saturation value is taken as 50 N·m; vibration envelope peak value =0.8m / s 2 The saturation value is taken as 3.0 m / s. 2 Acoustic emission impact counting =45, saturation value is 200; peak slip ratio =0.12, threshold =0.05; temperature =-10℃; Harmonic distortion rate =8%, saturation value is 25%. Weighting is... =0.5, =0.3, =0.2. Therefore, the calculation yields: The value exceeded the threshold of 0.65, triggering a wear warning. Actual endoscopic examination confirmed significant pitting on the raceway, verifying the accuracy of the invention.

[0078] To achieve predictive maintenance, the central processing unit also integrates a sixth-step wear evolution dynamics model. This model utilizes an extended Kalman filter to measure wear depth. and slip ratio Perform online estimations and recursively predict the future based on the current estimates. Wear trend within a yaw cycle. When predicted The system issues a high-priority warning when the bearing's rated hardened layer depth will reach 15% during the next maintenance cycle.

[0079] In a preferred embodiment of the present invention, the system has a self-healing calibration function. During the low wind speed period each month, the system automatically performs a full-range no-load yaw diagnosis and updates the baseline values ​​of friction torque and vibration.

[0080] To verify the technical superiority of this invention, a six-month field test was conducted on a 5MW offshore wind turbine. The comparison group included a single vibration threshold method and a fixed-weight multi-source fusion method. Test results showed that the accuracy of the initial wear identification of this invention reached 96.4%, the false alarm rate under strong wind load scenarios was only 1.2%, and the wear depth estimation error was ±0.03mm, significantly better than the comparison schemes.

[0081] In summary, this invention utilizes a unique comprehensive wear index. The wear evolution dynamics model of the multi-scenario adaptive fusion framework enables accurate quantitative identification and advanced prediction of the wear state of wind turbine yaw bearings, which has significant innovation and industrial applicability.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0083] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying the wear condition of wind turbine yaw bearings using multi-sensor fusion, characterized in that: Includes the following steps: Step 1: Acquire sensor signals collected by the vibration monitoring unit, acoustic emission monitoring unit, electrical parameter monitoring unit, and environmental perception unit. Based on the yaw motor's enable signal, brake pressure status, and motor current RMS value, divide the operating conditions into static lock scenario, start-up transition scenario, steady-state rotation scenario, and strong random load scenario. In each scenario, dynamically activate or suppress the sampling frequency and signal conditioning circuit of the corresponding sensor channel. Step 2: Under different scenarios, adaptive feature extraction is performed on vibration signals and acoustic emission signals using time-frequency analysis, empirical mode decomposition, envelope spectrum analysis, or wavelet packet decomposition, respectively, to obtain vibration envelope peak value, acoustic emission impact count, high-frequency energy ratio of acoustic emission, and slip ratio peak value. At the same time, electromagnetic torque is calculated through electrical parameters, and dynamic equilibrium equations of the yaw system are established. Extended Kalman filter is used to observe aerodynamic load torque as a disturbance term in real time, and wear friction torque increment is decoupled from electromagnetic torque. Step 3: Assign dynamic weighting coefficients to the vibration envelope peak value, acoustic emission impact count, and wear friction torque increment based on the current scenario. Calculate the comprehensive wear index based on the wear friction torque increment, vibration envelope peak value, acoustic emission impact count, slip ratio peak value, ambient temperature, and motor current harmonic distortion rate. The comprehensive wear index uses a product form to nonlinearly fuse the normalized values ​​of the above parameters and includes an exponential decay term due to temperature influence and a multiplicative correction term for wear based on slip ratio and harmonic distortion rate. Step four: Compare the calculated comprehensive wear index with the preset wear threshold. When the comprehensive wear index exceeds the threshold for a preset number of consecutive times, a wear warning is triggered.

2. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 1, characterized in that: The comprehensive wear index in step three is defined by the following formula: in, This represents the increase in wear friction torque. The peak value of the vibration acceleration envelope. For acoustic emission impact counting, These are scene-related weighting coefficients. These are the saturation values ​​of the corresponding features under severe bearing wear conditions. To trigger the peak glide rate in the transition scenario, The threshold for the slippage rate of health status. The slip ratio influence coefficient. The current ambient temperature. For reference temperature, The attenuation coefficient is affected by temperature. The harmonic distortion rate of the motor current. This represents the saturation value of the harmonic distortion rate. This is the clearance correction factor.

3. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 1, characterized in that: Increment of wear friction torque in step two The decoupling process specifically includes: converting the current in the three-phase stationary coordinate system into direct-axis and quadrature-axis components in the synchronous rotating coordinate system, and calculating the electromagnetic torque by combining the magnetic flux linkage and pole pair number of the motor permanent magnet. Establish the dynamic equilibrium equations for the yaw system: in To vary with wear depth The changing frictional resistance torque, The inflow wind speed and yaw angle The generated aerodynamic load torque The moment of inertia of the yaw system. Yaw angular velocity, The viscous friction coefficient was determined using a small-amplitude pulse excitation test under static locking conditions. Under no-wind, unloaded conditions, a full-stroke yaw test was performed to obtain the baseline friction torque curve under healthy conditions; then, an extended Kalman filter was used as the observer to measure the aerodynamic load torque. As a system disturbance term, the friction torque increment is extracted from the total electromagnetic torque in real time through iterative calculations of the prediction and update steps.

4. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 3, characterized in that: The process also includes establishing a dynamic model of yaw bearing wear evolution for trend prediction: defining the cumulative number of yaw cycles N as the discrete time step, with each yaw cycle corresponding to a complete wind adjustment action; and letting δ(N) be the equivalent wear depth. Based on the differential form of Archard's wear law and combined with the dynamics of the yaw system, the wear depth evolution rate equation is derived: in For the comprehensive wear coefficient, and Undetermined coefficients related to material hardness and surface morphology; wear depth and bearing clearance increment. The geometric relationship between them is determined by the contact angle. Decide: slip ratio The evolution equation simultaneously considers the effects of frictional torque increment, clearance increment, and temperature: in For the slippage rate of health status, This is the maximum starting torque of the motor. For correction factor, For reference clearance increment, Using the reference temperature; combining the above equation with an extended Kalman filter, for... and The system performs online condition estimation and predicts the wear depth within a preset number of yaw cycles based on the estimated value. When the predicted equivalent wear depth reaches a preset percentage of the bearing's rated hardened layer depth, the system triggers a high-priority warning.

5. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 1, characterized in that: The work condition division in step one uses scene confidence vectors and hidden Markov models to achieve nonlinear scene interleaving processing: A four-dimensional scene confidence vector is maintained in real time. Each component represents the probability that the current moment belongs to a static locking scene, a start-up transition scene, a steady-state rotation scene, and a strong random load scene. This confidence is dynamically updated through a hidden Markov model. The observed variables include yaw enable signal level, brake pressure value, motor speed and its fluctuation rate, and wind speed pulsation intensity. The scene transition probability matrix is ​​obtained by training through historical operation data. At any given time, multiple feature extraction branches corresponding to different scenarios are run simultaneously, with the computational resource allocation of each branch proportional to the confidence component. Meanwhile, based on the current confidence vector and transition matrix, the scenario most likely to be entered within a preset time window is predicted, and when a high-impact scenario is predicted to be entered, the high-speed sampling buffer is activated in advance and the acoustic emission channel gain is increased.

6. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 5, characterized in that: During the dynamic update of the scene confidence vector and the Hidden Markov Model, an online adaptive correction of the scene transition probability matrix is ​​adopted: after each complete yaw cycle, the transition probability matrix is ​​re-estimated using the expectation-maximization algorithm based on the difference between the actual observed scene sequence and the model prediction results, so as to gradually approximate the statistical characteristics of the actual working conditions; at the same time, the system maintains a scene history queue of length L. When the time interval between two adjacent scene switches in the queue is less than a preset threshold, it is determined to be scene jitter. At this time, the system will forcibly flatten the components of the confidence vector and restart the observation accumulation process of the Hidden Markov Model to avoid unstable feature extraction caused by frequent scene switches.

7. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 1, characterized in that: The dynamic weight coefficient allocation in step three is based on a fuzzy inference system to achieve adaptive adjustment: The input variables of the fuzzy inference system include the current operating condition stability index, the signal-to-noise ratio (SNR) estimate of each sensor channel, and the temperature drift correction confidence level. The operating condition stability index is calculated from the statistical variance of the speed fluctuation rate and the current change rate. The SNR estimate is obtained by comparing the effective value of the signal with the preset background noise level. The temperature drift correction confidence level is determined by the degree to which the ambient temperature deviates from the reference temperature. When the ambient temperature is lower than the preset low temperature threshold, the system automatically reduces the weight of the vibration signal and correspondingly increases the weight of the acoustic emission signal. When the signal-to-noise ratio of a sensor is detected to be lower than a preset lower limit, the system will redistribute the weight of that sensor proportionally to the other sensors.

8. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 7, characterized in that: It also includes an adaptive threshold correction step for wear status feedback: the central processing unit stores the comprehensive wear index sequence of historical yaw cycles, establishes a sliding window model of wear evolution trend, and uses the comprehensive wear index calculated at the current moment as prior information to correct the logical threshold of subsequent scenario classification; when the comprehensive wear index is in the early wear stage range, the system automatically lowers the current change rate threshold for switching from the start transition scenario to the steady-state rotation scenario, and the reduction magnitude is positively correlated with the degree to which the comprehensive wear index exceeds the health benchmark; at the same time, the system performs temperature normalization preprocessing on the vibration envelope peak value and acoustic emission impact count based on the real-time output of the temperature correction term in the comprehensive wear index, and the normalization coefficient is obtained by linear interpolation of the pre-calibrated sensor sensitivity temperature curve.

9. The multi-sensor fusion method for identifying the wear state of wind turbine yaw bearings according to claim 8, characterized in that: In the adaptive threshold correction step of wear condition feedback, the positive correlation between the magnitude of the reduction in the current change rate threshold and the degree to which the comprehensive wear index exceeds the health benchmark is determined by the following piecewise function: when Do not downgrade when the value is less than the first threshold; when When the value is between the first threshold and the second threshold, the downward adjustment magnitude is... The relationship is linear; the reduction is equal to the initial threshold multiplied by the preset proportional coefficient, and then multiplied by... The relative amount exceeding the first threshold; when When the threshold is greater than the second threshold, the reduction range is clamped to the maximum reduction range. This piecewise function ensures that the threshold is adjusted smoothly in the early stage of wear and responds quickly in the period of accelerated wear, while avoiding excessive reduction that could lead to misjudgment of the scenario.

10. A multi-sensor fusion wind turbine yaw bearing wear condition identification system, characterized in that: include: The data acquisition hardware layer includes a vibration monitoring unit, an acoustic emission monitoring unit, an electrical parameter monitoring unit, and an environmental sensing unit, which are used to collect vibration signals, acoustic emission signals, electrical parameter signals, and ambient temperature signals of the yaw bearing, respectively. The vibration monitoring unit uses a triaxial MEMS accelerometer, installed near the flange bolts connecting the outer ring of the yaw bearing to the nacelle base. The acoustic emission monitoring unit consists of multiple piezoelectric probes evenly spaced inside the yaw gear ring. Each probe is connected to a pre-amplifier with adjustable gain, and uses two parallel channels with different gains to achieve dynamic range expansion. A field-programmable gate array automatically selects the high-gain or low-gain channel for waveform reconstruction based on the real-time signal amplitude. The electrical parameter monitoring unit is integrated into the DC bus and three-phase output of the yaw driver, using a closed-loop Hall current sensor to monitor the instantaneous current and voltage of the three phases in real time. The field signal processing layer consists of a multi-channel synchronous sampling analog-to-digital converter module and a field-programmable gate array, which is used to perform signal pre-filtering, downsampling and real-time feature operator operations; The communication transmission layer uses an industrial Ethernet bus or CANopen protocol to upload the processed feature vectors to the central processing unit, and uses the wireless LoRa protocol as a backup link. When the industrial Ethernet packet loss rate exceeds a preset threshold, it automatically switches to wireless transmission mode. The edge computing server layer, as the central processing unit, integrates an embedded computing kernel, which runs a real-time operating system and executes the wind power yaw bearing wear state identification method based on multi-sensor fusion as described in any one of claims 1 to 9. The edge computing server layer is also equipped with non-volatile memory for building a feature database of the bearing's entire life cycle.

Images