Wind turbine generator vibration signal anomaly detection method
By combining multi-source synchronous acquisition and deep learning, the problems of low signal-to-noise ratio and spectral ambiguity in wind turbine vibration monitoring systems under complex operating conditions have been solved, achieving high-precision fault detection and early warning, which is suitable for condition monitoring of wind turbines.
Patent Information
- Application Number
- CN202511892851.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-02-27
AI Technical Summary
Existing wind turbine vibration monitoring systems lack the ability to detect sources of disturbance at the foundation level, have difficulty separating structurally conducted noise, have low signal-to-noise ratios, suffer from severe spectral ambiguity, and lack mechanical structure dynamic constraints in data-driven models, making it difficult to adapt to dynamic operating conditions under varying wind conditions.
A multi-source synchronous acquisition method is adopted, combining a triaxial accelerometer and a low-frequency vibration sensor. High-precision signal synchronization is achieved through a GPS timing module. An environmental noise decoupling model and a physical constraint hybrid model are constructed to perform noise reduction and feature extraction of vibration signals. Fault detection is performed using deep learning methods.
It improves the stability and accuracy of the wind turbine vibration monitoring system under complex operating conditions, effectively suppresses false alarms and spectrum spread, provides clear predictions of fault type and severity, and has good practicality and real-time performance.
Smart Images
Figure CN121576231A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power equipment condition monitoring technology, and in particular to a method for detecting abnormal vibration signals in wind turbine generators. Background Technology
[0002] As typical high-speed rotating equipment, wind turbines operate under high loads, strong disturbances, and multiple wind conditions, making their gearbox bearing systems highly susceptible to structural failures such as inner ring pitting, outer ring spalling, and gear tooth breakage. These failures severely impact the overall safety and operational stability of the turbine. To achieve early warning and fault identification of wind turbine operating conditions, current engineering practices commonly employ vibration monitoring by deploying accelerometers at key locations, supplemented by SCADA systems to acquire operating parameters such as wind speed and engine speed, thus aiding in the analysis of fault evolution. Simultaneously, researchers are gradually introducing wavelet analysis, envelope spectrum analysis, and deep learning methods to extract features and classify vibration data, thereby improving the intelligence level of fault detection.
[0003] However, the application of existing technologies in actual wind farm environments still faces several key technical challenges: First, most traditional vibration monitoring systems rely on single-point sensor deployment, lacking the ability to perceive the source of disturbance at the foundation level, making it difficult to separate structurally transmitted noise and resulting in a low signal-to-noise ratio for vibration signals; Second, existing noise reduction methods mostly rely on empirical filters, lacking a fine decoupling mechanism that incorporates the transmission characteristics of physical structures, leading to severe spectral ambiguity; Third, data-driven fault identification models lack constraints on the dynamics of mechanical structures, resulting in insufficient feature interpretability and a tendency to overfit, making it difficult to adapt to the dynamic operating conditions required by wind turbines under varying wind conditions. Summary of the Invention
[0004] This invention provides a method for detecting abnormal vibration signals in wind turbine generators. The method integrates multi-source synchronous acquisition, physical constraint modeling, and deep feature decision-making mechanism to improve the stability, accuracy, and reliability of the monitoring system under complex operating conditions.
[0005] A method for detecting abnormal vibration signals in wind turbine generators includes the following steps: S1: Install a triaxial accelerometer in the gearbox bearing housing and a low-frequency vibration sensor in the tower foundation to synchronously acquire the original vibration signal set; S2: Input the original vibration signal set into the pre-trained environmental noise decoupling model and output the noise-reduced vibration signal; S3: Based on the real-time wind speed data from the SCADA system, the noise-reduced vibration signal is normalized by rotational speed, and the joint time-frequency domain features are extracted to generate a multi-dimensional feature matrix; S4: Input the multidimensional feature matrix into the physical constraint hybrid model and output the anomaly detection result, including the fault type and severity. The physical constraint hybrid model includes a feature extraction layer, a physical constraint layer, and a decision fusion layer.
[0006] Optionally, S1 includes: S11: Install the triaxial accelerometer at the center of the axial bearing surface of the gearbox bearing housing, ensuring that its X / Y / Z axes coincide with the radial, tangential, and axial directions of the bearing housing, respectively, and collect the bearing vibration signal in real time at a sampling rate of ≥20kHz to generate a triaxial acceleration signal group of the bearing. S12: Four low-frequency vibration sensors are symmetrically installed at the four corners of the tower foundation flange. The Z-axis of the sensors points vertically to the foundation plane. The sensors collect environmental vibration signals at a sampling rate of ≤1kHz to generate a low-frequency vibration signal group for the tower foundation. S13: Send a synchronization trigger pulse to the triaxial accelerometer and low-frequency vibration sensor via the GPS timing module to start all sensors to collect signals under the same time reference and output the time-aligned raw signal set; S14: Align and merge the bearing triaxial acceleration signal group and the tower foundation low-frequency vibration signal group according to the timestamp, and attach real-time wind speed and rotational speed condition labels of the SCADA system to generate the original vibration signal set with spatiotemporal correlation.
[0007] Optionally, S2 includes: S21: Under normal operating conditions of the wind turbine, the original vibration signal set of the historical time period is collected synchronously, and the gearbox health status label under the corresponding working condition is marked to generate a labeled model training dataset. S22: Based on the low-frequency vibration signal group of the tower foundation and the vibration signal of the gearbox bearing housing in the model training dataset, the vibration transfer function is calculated through impact hammer test and least squares system identification algorithm. It characterizes the frequency domain response of the foundation vibration transmitted to the bearing housing; S23: The vibration transfer function As a physical constraint embedded in the convolutional neural network, the environmental noise decoupling model with frozen parameters is trained by minimizing the noise separation loss function, using the model training dataset as input. S24: Input the original vibration signal set into the trained environmental noise decoupling model in real time, based on the vibration transfer function. The vibration coupling component of the tower foundation is eliminated in reverse, and the noise-reduced vibration signal is output.
[0008] Optionally, the specific process of S24 is as follows: The low-frequency vibration signal set of the tower foundation from the original vibration signal set is input into the pre-trained environmental noise decoupling model, and the vibration transfer function is used to... The coupled noise estimate of the bearing housing position is calculated, and the frequency domain Wiener filtering algorithm is used to suppress the noise of the gearbox bearing housing vibration signal.
[0009] Optionally, S3 includes: S31: Read wind speed data from the SCADA system in real time, query the theoretical spindle speed value corresponding to the current wind speed based on the wind turbine power curve characteristic table, and output the theoretical spindle speed value; S32: Based on the theoretical spindle speed, the noise-reduced vibration signal is subjected to order-ratio resampling processing to eliminate the spectral diffusion effect caused by actual speed fluctuations and generate a speed-normalized vibration signal that is independent of the speed. S33: Synchronous compressed wavelet transform and envelope spectrum analysis are performed on the normalized vibration signal of the rotational speed. The three characteristic indicators, namely the time-frequency ridge stability coefficient, the envelope spectrum kurtosis value, and the energy ratio of the fault characteristic frequency band, are calculated respectively and combined to form a three-dimensional feature vector sequence. S34: Stack the three-dimensional feature vector sequence within a continuous time window along the time axis, add real-time wind speed data from the SCADA system as the operating condition dimension, and generate a multi-dimensional feature matrix.
[0010] Optionally, S4 includes: S41: Input the multidimensional feature matrix into the feature extraction layer of the physical constraint hybrid model, and use a three-layer one-dimensional convolutional neural network to perform deep feature learning along the time axis, and output a deep feature vector containing 128-dimensional abstract features. S42: Input the deep feature vector into the physical constraint layer of the physical constraint hybrid model, load the pre-stored bearing fault feature frequency equation library for feature matching verification, discard feature components that deviate from the theoretical value of the bearing fault dynamic equation by more than 5%, and generate a physical verification feature vector that conforms to physical laws. S43: Input the physical verification feature vector into the decision fusion layer of the physical constraint hybrid model, simultaneously perform support vector machine multi-classification and random forest regression analysis, and output a weighted fusion output of four-dimensional classification probability distribution and severity prediction value in the [0,1] interval; S44: Based on the fault type label corresponding to the maximum probability value in the classification probability distribution, and combined with whether the severity prediction value exceeds the threshold, generate an anomaly detection result that includes fault type, severity prediction value, and alarm status.
[0011] Optionally, the fault types include four categories: normal condition, pitting of the bearing inner ring, spalling of the outer ring, and broken gear teeth.
[0012] Optionally, an emergency alarm is triggered when the severity prediction value is above 0.8.
[0013] The beneficial effects of this invention are: This invention, by deploying triaxial accelerometers and low-frequency vibration sensors on the gearbox bearing housing and tower foundation respectively, and combining them with a GPS timing module for high-precision triggering and synchronization, can simultaneously acquire microscopic structural vibration signals and macroscopic foundation disturbance signals. Furthermore, it aligns all sampled data with timestamps and fuses operating condition labels (including wind speed and spindle speed), thereby constructing an original vibration signal set with spatiotemporal consistency and operational status correlation. This provides comprehensive and accurate data support for subsequent modeling and diagnosis, outperforming traditional sampling methods that rely on single-point signals.
[0014] This invention constructs an environmental noise decoupling model that integrates vibration transfer function and deep neural network. By combining impact hammer testing and least squares system identification algorithm, the vibration transfer function from the tower foundation to the bearing housing is accurately obtained and embedded into the network structure to achieve frequency domain filtering of physical constraints. Frequency domain Wiener filtering enables directional noise reduction of the bearing housing vibration signal, effectively suppressing false alarms or spectral spread caused by wind-induced interference from the foundation, improving frequency domain clarity, and significantly outperforming traditional noise reduction methods based solely on filters or thresholding.
[0015] This invention proposes a physical constraint hybrid model that introduces bearing fault dynamics equations as a physical verification mechanism on top of the traditional data-driven model. This ensures that the frequency components in the deep features must meet the characteristic frequency matching conditions of a specific bearing model, thereby effectively filtering out pseudo-features without a physical basis. Furthermore, the model integrates a support vector machine multi-classifier and a random forest regressor to achieve multi-class discrimination of fault types and continuous value prediction of severity. Based on classification probabilities and prediction thresholds, it generates clear alarm status signals, exhibiting good practicality, real-time performance, and interpretability, making it suitable for the deployment and application of wind power operation and maintenance monitoring systems. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the S3 process in an embodiment of the present invention. Detailed Implementation
[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.
[0019] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.
[0020] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.
[0021] like Figures 1-2 As shown, a method for detecting abnormal vibration signals in wind turbine generators includes the following steps: S1: Install a triaxial accelerometer in the gearbox bearing housing and a low-frequency vibration sensor in the tower foundation to synchronously acquire the original vibration signal set, specifically: S11, Bearing Triaxial Accelerometer Signal Acquisition: A triaxial accelerometer is securely installed at the center of the axial bearing surface of the wind turbine gearbox bearing housing. To ensure that the acquired signals accurately characterize the bearing vibration state, the three measuring axes of the sensor must be strictly aligned with the radial direction (X-axis), tangential direction (Y-axis), and axial direction (Z-axis) of the bearing housing, respectively. Calibration can be performed during installation using a laser alignment instrument or a three-dimensional coordinate positioning device.
[0022] The acquisition system is configured with a sampling rate of 20kHz to ensure coverage of the mid-to-high frequency signal range involved in common rolling bearing faults (such as inner race fault frequency, rolling element pass frequency, etc.), facilitating subsequent effective time-frequency analysis. The data acquired by the sensors are named acceleration signals according to their respective channels. This constitutes a complete triaxial acceleration signal group for the bearing.
[0023] S12, Low-frequency vibration signal acquisition for tower foundation: To monitor the overall vibration response between the wind turbine tower and the foundation, four low-frequency vibration sensors are symmetrically installed at the four corners of the tower foundation flange. During installation, the sensors should be positioned so that their Z-axis points perpendicularly to the foundation plane to capture environmental vibration characteristics along the vertical direction. The signal channel of each sensor is named... These sensors are used to characterize the response at different foundation locations. The sampling rate of these sensors is set to no more than 1 kHz, which is sufficient to cover the low-frequency range related to structural response (typically in the 0.5 Hz-200 Hz range), while reducing the burden of data storage and processing. This constitutes a low-frequency vibration signal set for the tower foundation.
[0024] S13, Time-step triggering for raw signal acquisition: To ensure that the triaxial accelerometer and low-frequency vibration sensor acquire signals under the same time reference, the system is equipped with a GPS timing module that sends high-precision synchronization trigger pulses to all sensors. The timing module uses the PPS (Pulse Per Second) signal format, with a timing accuracy better than 1μs, which can meet the time consistency requirements of high-frequency vibration data.
[0025] All acquisition devices start data recording through a unified clock synchronization mechanism, so that signals from different sources have a unified timestamp, which can be directly used for subsequent joint feature analysis and multimodal modeling.
[0026] S14, Generation of the original vibration signal set: The triaxial acceleration signal set of the bearing acquired in S11 and the low-frequency vibration signal set of the tower foundation acquired in S12 are aligned by timestamp and merged to form a unified multi-source vibration signal set. Simultaneously, wind speed data at the corresponding time points are acquired in real time from the SCADA system. and spindle speed data This is added as a working condition label to each group of vibration samples.
[0027] The final output format of the original vibration signal set is as follows: ; This dataset possesses time alignment, spatial distribution, and traceability of operating conditions, providing a complete input foundation for subsequent vibration signal denoising, rotational speed normalization, and multidimensional feature extraction.
[0028] S2: Input the original vibration signal set into the pre-trained environmental noise decoupling model, and output the denoised vibration signal, specifically: S21, Constructing the training dataset for the environmental noise decoupling model: Under the premise that the wind turbine is in normal operation and without any structural faults, select multiple historical time periods, collect raw vibration signal sets according to the method in S1, and attach a gearbox health status label (such as "normal", "minor abnormality", "no fault observed", etc.) generated by the SCADA system to each sample to form a labeled model training dataset. The sample size is no less than 1000 sets, covering different wind speeds, loads and environmental disturbance conditions to ensure the diversity and generalization ability of the training data.
[0029] S22, Calculate the vibration transfer function from the tower foundation to the bearing housing: To characterize the coupling effect of the tower foundation vibration signal on the gearbox bearing housing signal, the vibration transfer function is obtained through the following procedure. : Impact hammer test setup: With the wind turbine stopped, a standard impact hammer (with a triaxial force sensor) is used to apply pulse excitation to the tower foundation flange, while simultaneously acquiring the triaxial acceleration response at the bearing housing; Data processing: Extract the time-domain waveforms of the impact signal and response signal, and perform Fast Fourier Transform (FFT) on them respectively, denoted as... and , Least squares system identification: The frequency domain least squares method is used to solve for the system's transfer characteristics and calculate the vibration transfer function. ; Function stability test: Verify that the phase delay and amplitude characteristics of the obtained transfer function differ by less than 5% during different test cycles. If the condition is not met, repeat the test.
[0030] The vibration transfer function This reflects how foundational disturbances affect the vibration characteristics of the bearing housing through structural paths in the frequency domain, and is the core of the physical constraints of subsequent models.
[0031] S23, Training the Environmental Noise Decoupling Model: Using the model training dataset constructed in step S21 as input, a signal denoising model with time-frequency feature extraction capability is constructed using a convolutional neural network (CNN). A vibration transfer function is introduced into the network structure. This is embedded as a regularization physical constraint into the intermediate layer output of the neural network.
[0032] The noise separation loss function is defined as follows: ; in, The clean vibration signal is artificially calibrated. For the model's predicted output, The propagation mapping is automatically learned within the model. These are the physical constraint regularization coefficients. This represents the Frobenius norm.
[0033] After training, the model parameters are frozen to form a pre-trained environmental noise decoupling model, which has the ability to automatically filter out structural conducted noise from actual vibration signals.
[0034] S24, Perform joint environment-physics noise reduction: Input the original vibration signal set into the pre-trained environment noise decoupling model, and perform the following joint noise reduction operation: 1. Signal Conversion: Convert the triaxial acceleration signal of the bearing housing. Vibration signal group of tower foundation Perform a Fast Fourier Transform to obtain the following results: and ; 2. Noise estimation: Input model, combined with pre-embedded Calculate and estimate the noise component at the bearing housing; 3. Frequency Domain Filtering: The frequency domain Wiener filtering algorithm is applied to reduce noise in the bearing vibration signal. The specific expression is as follows: ; in, This is the frequency domain signal after noise reduction. This is an adaptive adjustment factor that is adjusted in real time according to the intensity of the disturbance to the wind turbine foundation. 4. Signal restoration: For Perform an inverse Fourier transform to recover the time-domain signal, obtaining the final denoised vibration signal. .
[0035] S3: Based on the real-time wind speed data from the SCADA system, the noise-reduced vibration signal is normalized by rotational speed, and joint time-frequency features are extracted to generate a multi-dimensional feature matrix. Specifically: S31, Theoretical spindle speed extraction: Wind speed data is read in real time from the SCADA system via an industrial communication interface (such as OPC or Modbus protocol). And based on the calibrated wind turbine power curve characteristic table, the theoretical main shaft speed value corresponding to the current wind speed is retrieved. This characteristic table is obtained by fitting wind tunnel experimental data with field operation data, and has the following mapping relationship: ; in, Wind speed data collected by the SCADA system This represents the theoretical spindle speed at the corresponding wind speed.
[0036] Theoretical spindle speed value It can serve as a reference benchmark for the ideal operating conditions of wind turbine units and be used for subsequent frequency normalization calculations.
[0037] S32, Generation of Speed-Normalized Vibration Signal: Due to instantaneous speed fluctuations during wind turbine operation, direct spectral analysis results in spectral diffusion, leading to ambiguity of characteristic frequencies. Therefore, it is necessary to perform order-ratio resampling processing on the noise-reduced vibration signal. The specific steps are as follows: 1. Obtain the theoretical spindle speed value for the time period corresponding to the current vibration signal. ; 2. Reduce the noise of the vibration signal. Angle-normalized resampling is performed to convert it into an order-ratio domain signal. ,in: ; For mechanical rotation angle, This represents the actual spindle speed, derived from SCADA data. 3. The result after resampling This is a normalized vibration signal based on spindle speed, which is insensitive to changes in spindle speed and is beneficial for extracting stable fault characteristic frequency components.
[0038] S33, Vibration Feature Extraction and 3D Feature Vector Generation: Parallel synchronous compressed wavelet transform and envelope spectrum analysis are performed on the rotational speed normalized vibration signal obtained in step S32 to extract joint diagnostic features in the time-frequency and envelope domains. The process is as follows: Synchronous Compressed Wavelet Transform (SST): Utilizing Morlet mother wavelet pairs Perform time-frequency transformation to generate a time-frequency energy map; extract the dominant frequency ridge from the time-frequency energy map and calculate its time-frequency ridge stability coefficient, which is used to measure the continuity of the frequency trajectory and the sensitivity to disturbances.
[0039] Envelope spectrum analysis: A Hilbert transform is performed on the normalized vibration signal to obtain the envelope signal. An FFT is then performed on the envelope signal to extract: 1. Envelope spectrum kurtosis value: reflects the intensity of the impact component; 2. Fault characteristic frequency band energy ratio: such as the ratio of the energy in the frequency band corresponding to the inner ring fault frequency and the rolling element passage frequency.
[0040] Finally, by integrating the above three types of features, a three-dimensional feature vector sequence in the following format is formed: ; in Indicates the first The feature vectors of each time window correspond to the time-frequency ridge stability coefficient, envelope kurtosis value, and fault feature frequency band energy proportion in three dimensions.
[0041] S34, Multidimensional Feature Matrix Construction: Multiple three-dimensional feature vector sequences within a continuous time window (e.g., every 5 seconds) are stacked along the time axis, and wind speed data from the SCADA system is appended to each time window. This forms a four-dimensional feature term that includes the working condition dimension. Finally, a multi-dimensional feature matrix with the following structure is constructed: ; in, Number of time windows; Each row's feature is [ .
[0042] The resulting multidimensional feature matrix has a clear time series structure, complete feature information, and working condition correlation, which serves as the input basis for subsequent anomaly detection in the physical constraint hybrid model.
[0043] S4: Input the multidimensional feature matrix into the physical constraint hybrid model, and output the anomaly detection results, including the fault type and severity. The physical constraint hybrid model includes a feature extraction layer, a physical constraint layer, and a decision fusion layer, specifically: S41, Process the multidimensional feature matrix through the feature extraction layer: The multidimensional feature matrix generated in step S3... The input physical constraint hybrid model's feature extraction layer consists of three one-dimensional convolutional neural networks, specifically designed for deep abstract representation of time-series features. The structure is configured as follows: The first layer contains 64 convolutional kernels, each with a kernel size of 5, a stride of 1, an activation function of LeakyReLU, and a leakage rate of 0.1. The second layer contains 32 convolutional kernels, each with a kernel size of 3, and uses the same activation function as above. The third layer contains 16 convolutional kernels, each with a kernel size of 3, and uses the same activation function as above. Each convolutional operation is followed by a max pooling layer of size 2 to reduce the temporal dimension and preserve local peak features.
[0044] After three layers of convolution and pooling operations, the original feature sequence is transformed into a fixed-length depth feature vector with dimension 128. This vector comprehensively expresses the dynamic evolution pattern of the input signal, structural response characteristics, and factors influencing wind speed conditions.
[0045] S42, Verifying the bearing failure dynamics equations through a physical constraint layer: This involves using deep feature vectors... The physical constraint layer in the input physical constraint hybrid model is loaded with a built-in bearing failure dynamics equation library to verify whether the frequency information contained in the feature vector conforms to the physical mechanism of common bearing failure modes.
[0046] The dynamic equation library covers 12 common bearing types and includes the following formulas (examples): Inner ring failure frequency: ; Outer ring failure frequency: ; Frequency of rolling element passage: ; The parameters include, Main spindle speed For the number of rolling elements, For the rolling element diameter and pitch circle diameter, Contact angle, The system calls the feature verification module to transform and compare each frequency-related component in the deep feature vector; If the frequency of a component deviates from the theoretical value of the dynamic formula by more than 5%, it is identified as an abnormal component and is removed. The output retains only the physical verification feature vectors that satisfy the physical constraints. ,in .
[0047] This step can effectively filter out overfitting or physically unfounded feature representations in neural networks, improving model interpretability and reliability.
[0048] S43, Generate classification probability distribution through decision fusion layer: This involves physically verifying feature vectors. The synchronous input decision fusion layer undergoes parallel inference through two independent models: 1. Support Vector Machine (SVM) multi-classifier: Kernel function: Radial basis function (RBF); Output: A classification probability distribution vector of length 4. These correspond to the following four types of fault type labels: Normal condition, pitting of the bearing inner ring, peeling of the outer ring, broken gear teeth 2. Random Forest Regressor (RF): Configuration: Includes 50 trees at a depth of [missing information]. Decision tree; Output: Predicted severity of the current fault. The closer it is to 1, the more serious the condition.
[0049] This fusion layer combines the advantages of classification and regression models, providing continuous quantitative indicators while clearly defining the fault type, which is beneficial for subsequent level early warning and maintenance priority ranking.
[0050] S44, Parse and generate anomaly detection results: Finally, based on the above two outputs, generate anomaly detection results as follows: Fault type label generation: from classification probability distribution Select the label corresponding to the highest probability as the final fault type, for example: ; Alarm status determination: If the severity prediction value If the alarm flag is set to 1, an emergency alarm state is triggered; otherwise, the normal operation state is maintained and the alarm flag is set to 0. Output content structure: The anomaly detection results are output in the form of a structure, which includes the anomaly detection results of fault type, severity prediction value and alarm status.
[0051] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0052] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for detecting abnormal vibration signals in wind turbine generators, characterized in that, Includes the following steps: S1: Install a triaxial accelerometer in the gearbox bearing housing and a low-frequency vibration sensor in the tower foundation to synchronously acquire the original vibration signal set; S2: Input the original vibration signal set into the pre-trained environmental noise decoupling model and output the noise-reduced vibration signal; S3: Based on the real-time wind speed data from the SCADA system, the noise-reduced vibration signal is normalized by rotational speed, and the joint time-frequency domain features are extracted to generate a multi-dimensional feature matrix; S4: Input the multidimensional feature matrix into the physical constraint hybrid model and output the anomaly detection result, including the fault type and severity. The physical constraint hybrid model includes a feature extraction layer, a physical constraint layer, and a decision fusion layer.
2. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 1, characterized in that, S1 includes: S11: Install the triaxial accelerometer at the center of the axial bearing surface of the gearbox bearing housing, ensuring that its X / Y / Z axes coincide with the radial, tangential, and axial directions of the bearing housing, respectively, and collect the bearing vibration signal in real time at a sampling rate of ≥20kHz to generate a triaxial acceleration signal group of the bearing. S12: Four low-frequency vibration sensors are symmetrically installed at the four corners of the tower foundation flange. The Z-axis of the sensors points vertically to the foundation plane. The sensors collect environmental vibration signals at a sampling rate of ≤1kHz to generate a low-frequency vibration signal group for the tower foundation. S13: Send a synchronization trigger pulse to the triaxial accelerometer and low-frequency vibration sensor via the GPS timing module to start all sensors to collect signals under the same time reference and output the time-aligned raw signal set; S14: Align and merge the bearing triaxial acceleration signal group and the tower foundation low-frequency vibration signal group according to the timestamp, and attach real-time wind speed and rotational speed condition labels of the SCADA system to generate the original vibration signal set with spatiotemporal correlation.
3. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 2, characterized in that, S2 includes: S21: Under normal operating conditions of the wind turbine, the original vibration signal set of the historical time period is collected synchronously, and the gearbox health status label under the corresponding working condition is marked to generate a labeled model training dataset. S22: Based on the low-frequency vibration signal group of the tower foundation and the vibration signal of the gearbox bearing housing in the model training dataset, the vibration transfer function is calculated through impact hammer test and least squares system identification algorithm. It characterizes the frequency domain response of the foundation vibration transmitted to the bearing housing; S23: The vibration transfer function As a physical constraint embedded in the convolutional neural network, the environmental noise decoupling model with frozen parameters is trained by minimizing the noise separation loss function, using the model training dataset as input. S24: Input the original vibration signal set into the trained environmental noise decoupling model in real time, based on the vibration transfer function. The vibration coupling component of the tower foundation is eliminated in reverse, and the noise-reduced vibration signal is output.
4. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 3, characterized in that, The specific process of S24 is as follows: The low-frequency vibration signal set of the tower foundation from the original vibration signal set is input into the pre-trained environmental noise decoupling model, and the vibration transfer function is used to... The coupled noise estimate of the bearing housing position is calculated, and the frequency domain Wiener filtering algorithm is used to suppress the noise of the gearbox bearing housing vibration signal.
5. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 4, characterized in that, S3 includes: S31: Read wind speed data from the SCADA system in real time, query the theoretical spindle speed value corresponding to the current wind speed based on the wind turbine power curve characteristic table, and output the theoretical spindle speed value; S32: Based on the theoretical spindle speed, the noise-reduced vibration signal is subjected to order-ratio resampling processing to eliminate the spectral diffusion effect caused by actual speed fluctuations and generate a speed-normalized vibration signal that is independent of the speed. S33: Synchronous compressed wavelet transform and envelope spectrum analysis are performed on the normalized vibration signal of the rotational speed. The three characteristic indicators, namely the time-frequency ridge stability coefficient, the envelope spectrum kurtosis value, and the energy ratio of the fault characteristic frequency band, are calculated respectively and combined to form a three-dimensional feature vector sequence. S34: Stack the three-dimensional feature vector sequence within a continuous time window along the time axis, add real-time wind speed data from the SCADA system as the operating condition dimension, and generate a multi-dimensional feature matrix.
6. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 5, characterized in that, The physical constraint hybrid model sequentially extracts deep feature vectors through a feature extraction layer, matches bearing fault dynamics equations through a physical constraint layer, and generates classification probability distributions through a decision fusion layer, outputting anomaly detection results.
7. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 6, characterized in that, S4 includes: S41: Input the multidimensional feature matrix into the feature extraction layer of the physical constraint hybrid model, and use a three-layer one-dimensional convolutional neural network to perform deep feature learning along the time axis, and output a deep feature vector containing 128-dimensional abstract features. S42: Input the deep feature vector into the physical constraint layer of the physical constraint hybrid model, load the pre-stored bearing fault feature frequency equation library for feature matching verification, discard feature components that deviate from the theoretical value of the bearing fault dynamic equation by more than 5%, and generate a physical verification feature vector that conforms to physical laws. S43: Input the physical verification feature vector into the decision fusion layer of the physical constraint hybrid model, simultaneously perform support vector machine multi-classification and random forest regression analysis, and output a weighted fusion output of four-dimensional classification probability distribution and severity prediction value in the [0,1] interval; S44: Based on the fault type label corresponding to the maximum probability value in the classification probability distribution, and combined with whether the severity prediction value exceeds the threshold, generate an anomaly detection result that includes fault type, severity prediction value, and alarm status.
8. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 7, characterized in that, The fault types include four categories: normal condition, pitting of the bearing inner ring, peeling of the outer ring, and broken gear teeth.
9. The method for detecting abnormal vibration signals in a wind turbine generator according to claim 8, characterized in that, An emergency alarm is triggered when the severity prediction value is above 0.8.