Method and device for monitoring a wind turbine based on operational data

Abstract

The present application relates to the technical field of wind power generator, and particularly relates to a wind power generator fault monitoring method and device based on operation data, comprising the following steps: S1, sensor arrangement and installation: arranging magnetostrictive sensors on rotating parts; S2, data acquisition: acquiring strain signals and deformation data in real time; S3, data preprocessing: preprocessing the strain signals and deformation data; S4, fault feature extraction: extracting fault feature parameters; S5, model construction and training: constructing a wind power generator fault pre-detection model to identify early fault signs and predict the probability of fault occurrence; S6, real-time monitoring and fault early warning: monitoring the strain signals and deformation data of the rotating parts in real time, and generating a fault early warning signal. The present application improves the real-time health state of the key components of the wind power generator, and effectively prevents the occurrence of potential faults.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power generators, in particular to a wind power generator fault monitoring method and device based on operation data. BACKGROUND

[0002] With the rapid development of wind power technology, wind power generators have become an important part of global energy supply. However, key components of wind power generators, such as main shafts, gearboxes and bearings, are prone to fatigue and wear during long-term high-load and high-speed operation. The failure of these components not only leads to a decrease in the operating efficiency of wind power generators, but also can cause serious equipment damage, resulting in high maintenance costs and unpredictable downtime. Therefore, real-time monitoring and fault warning of key components of wind power generators have become an important means to ensure the safe and stable operation of wind power equipment.

[0003] Existing wind power generator fault monitoring technologies mainly rely on monitoring data from vibration sensors and temperature sensors. These methods can only detect problems when a fault has occurred or is about to occur, making it difficult to accurately predict early failures. In addition, existing technologies rely on fixed threshold fault detection mechanisms and lack dynamic adjustment capabilities, making them susceptible to environmental noise and changes in operating conditions, resulting in false positives or false negatives. Moreover, existing fault detection models mostly use simple machine learning algorithms, which cannot fully utilize historical data and complex operating characteristics of wind power generators, resulting in insufficient prediction capabilities of the model under high-dimensional data and nonlinear fault patterns, making it difficult to meet the needs of practical applications.

[0004] The present application aims to overcome the shortcomings of existing technologies and provide a wind power generator fault monitoring method based on operation data to improve the operating reliability and maintenance efficiency of wind power generators and reduce the risk of failure. SUMMARY

[0005] The present application provides a wind power generator fault monitoring method and device based on operation data.

[0006] The wind power generator fault monitoring method based on operation data comprises the following steps:

[0007] S1, sensor arrangement and installation: arranging magnetostrictive sensors on rotating components of the wind power generator to detect small strains and deformations of the rotating components during operation;

[0008] S2, data acquisition: acquiring strain signals and deformation data of the rotating components of the wind power generator in real time through the magnetostrictive sensors;

[0009] S3, data preprocessing: the collected strain signals and deformation data are preprocessed, including signal denoising, outlier rejection and data smoothing processing, and an adaptive filtering algorithm is used to optimize the original strain signals and deformation data;

[0010] S4, fault feature extraction: based on the preprocessed strain signals and deformation data, fault feature parameters are extracted, including frequency domain features, time domain features and time-frequency domain features, and a feature vector for model input is generated;

[0011] S5, model construction and training: based on the constructed feature vector, a wind turbine fault pre-detection model is constructed, and the model is trained in combination with the historical strain signals and deformation data of the wind turbine and the fault case library, to identify early fault signs and predict the probability of fault occurrence, specifically including:

[0012] S51, feature vector input: the generated feature vector is used as the input data of the model, and the input data is standardized and normalized;

[0013] S52, model construction: based on the processed feature vector, a wind turbine fault pre-detection model is constructed;

[0014] S53, model training and verification: the historical strain signals and deformation data of the wind turbine are divided into training set and verification set, the data of the training set is used to train the model through cross-validation and hyperparameter optimization, and the model is verified using the verification set;

[0015] S6, real-time monitoring and fault warning: in the running process of the wind turbine, the strain signals and deformation data of the rotating parts are monitored in real time, and the trained wind turbine fault pre-detection model is used for prediction, when potential fault signs are detected, a fault warning signal is generated, and the fault type, location and development trend are output.

[0016] Optionally, the sensor arrangement and installation in S1 includes:

[0017] S11, sensor arrangement position selection: magnetostrictive sensors are arranged on the surface or inside the predetermined positions of the rotating parts (main shaft, gear box and bearing) of the wind turbine;

[0018] S12, sensor fixing method: the magnetostrictive sensors are fixed on the rotating parts by non-contact or contact installation method, the non-contact installation adopts magnetic support or adhesion technology, and the contact installation adopts mechanical clamp or fastener fixing;

[0019] S13, sensor calibration and adjustment: after the magnetostrictive sensor is installed, preliminary calibration and zero adjustment are carried out to compensate for the displacement or error in the installation process.

[0020] Optionally, the data collection in S2 includes:

[0021] S21, strain signal collection: collecting strain signals of rotating components of the wind turbine in real time through magnetostrictive sensors, including axial strain, radial strain and torsional strain;

[0022] S22, deformation data collection: collecting deformation data of the rotating components through magnetostrictive sensors, including bending deformation, stretching deformation and shearing deformation of the components;

[0023] S23, data collection synchronization and storage: time stamping the collected strain signals and deformation data, and storing them in real time in local or cloud data storage units.

[0024] Optionally, the data preprocessing in S3 includes:

[0025] S31, signal denoising: using wavelet transform to perform signal denoising processing on the collected strain signals and deformation data;

[0026] S32, outlier rejection: identifying and rejecting outliers in the strain signals and deformation data through statistical methods;

[0027] S33, data smoothing processing: using a sliding average algorithm to perform smoothing processing on the signals;

[0028] S34, adaptive filter algorithm optimization: using LMS adaptive filter algorithm to optimize the preprocessed strain signals and deformation data.

[0029] Optionally, the fault feature extraction in S4 includes:

[0030] S41, time domain feature extraction: based on the preprocessed strain signals and deformation data, extracting time domain feature parameters, including mean value, root mean square value (RMS), skewness and kurtosis;

[0031] S42, frequency domain feature extraction: performing Fourier transform on the preprocessed strain signals and deformation data to extract frequency domain feature parameters, including dominant frequency, frequency center, bandwidth and spectral energy;

[0032] S43, time-frequency domain feature extraction: using short-time Fourier transform to extract time-frequency domain features;

[0033] S44, feature vector generation: fusing the extracted time domain, frequency domain and time-frequency domain feature parameters to generate a feature vector, which is used as input for the fault prediction model.

[0034] Optionally, the feature vector input in S51 includes:

[0035] S511, standardization of feature vectors: the generated feature vectors are standardized to eliminate the dimensional differences between the features;

[0036] S512, normalization of feature vectors: the standardized feature vectors are normalized to scale the feature values to a uniform range;

[0037] S513, construction of model input data: the feature vectors V i are input into the wind turbine fault pre-detection model.

[0038] Optionally, the wind turbine fault pre-detection model in S52 adopts a Bayesian neural network (BNN) model, which includes:

[0039] S521, Bayesian inference process: in the fault detection of the wind turbine, a weighted prior distribution is introduced, represented as:

[0040]

[0041] where p(w|D) is the posterior distribution of weight w given data D, p(D|w) is the likelihood function, p(w) is the prior distribution of weight, Z is the normalization constant, α, γ k are adjustment coefficients, w k is the model weight corresponding to the historical fault data, K is the number of historical fault data or reference fault patterns;

[0042] S522, likelihood estimation based on adaptive kernel function: an adaptive kernel function is introduced to adaptively adjust the variance according to the position of the data points, represented as:

[0043]

[0044] where p(D|w) is the likelihood function, is a normal distribution, is a variance, is a reference variance, μ is the mean vector of the data, l is the kernel width parameter, N is the number of data samples, y i is the observation value of the i-th sample;

[0045] S523, prediction distribution calculation: the prediction distribution is calculated by Monte Carlo sampling to approximate the complex data structure and provide prediction of fault occurrence, represented as:

[0046]

[0047] where p(y * |x *D) is the prediction distribution, representing the distribution given input x. * Given data D, output y * The probability, M is the number of samples, w m These are the weights sampled from the posterior distribution p(w|D);

[0048] S524, Fault Detection and Uncertainty Quantification: By calculating the variance of predicted values, a quantitative assessment of uncertainty is provided, expressed as:

[0049]

[0050] Among them, Uncertainty(x) * () is a measure of uncertainty, representing the uncertainty in the input x. * The variance of the time prediction (x * f(x) represents the mean of the predicted values. * ,w m ) represents the m-th sampling weight w m The predicted value below;

[0051] S525, Model Update and Adaptive Adjustment: The model is updated in real time through incremental learning and adaptive adjustment of the learning rate, as shown below:

[0052]

[0053] Among them, w t+1 For the updated weights, w t Let η be the current weight, η be the learning rate, and D be the learning rate. t Here is the data at the current time step t, η0 is the initial learning rate, and λ is the learning rate decay factor.

[0054] Optionally, the model training and validation in S53 includes:

[0055] S531, Dataset partitioning: The historical strain signals and deformation data of wind turbines are divided into training set and validation set according to a predetermined ratio;

[0056] S532, Cross-validation: During model training, k-fold cross-validation is used to divide the training set into k subsets. k-1 subsets are used for training, and the remaining subset is used for validation. This process is repeated k times, with a different subset selected as the validation set each time. The average value is then used as the evaluation standard for model performance.

[0057] S533, Hyperparameter Optimization: Hyperparameter optimization is performed using grid search, represented as:

[0058]

[0059] wherein, θ * is the optimal hyperparameter combination, arg min is to find the parameter value that makes the subsequent expression minimum, θ is the current hyperparameter combination, Θ grid is the predefined hyperparameter grid, k is the number of cross-validation folds, is the loss function, is the output value obtained by using the hyperparameter combination θ to predict on the jth fold validation set, y j is the true label on the jth fold validation set.

[0060] Optionally, the real-time monitoring and failure warning in S6 comprises:

[0061] S61, real-time signal monitoring: during the operation of the wind turbine, the strain signals and deformation data of rotating components (such as main shaft, gearbox, bearing, etc.) are monitored in real time;

[0062] S62, data and model prediction result comparison and analysis: the real-time collected strain signals and deformation data are input into the trained wind turbine failure prediction model, and the current failure state or health risk probability is output;

[0063] S63, failure warning generation and information output: when the real-time prediction result shows that there is a potential failure (the predicted failure probability exceeds the set detection threshold), a failure warning signal is immediately generated, including the type of failure (such as bearing failure, gearbox failure, etc.), the location of failure occurrence, and the development trend of failure (such as the speed of failure deterioration).

[0064] The wind turbine failure monitoring device based on operation data is used to realize the wind turbine failure monitoring method based on operation data, comprising the following units:

[0065] Sensor arrangement and installation unit: magnetostrictive sensors are arranged on the rotating components of the wind turbine to detect the small strain and deformation of the rotating components during operation;

[0066] Data acquisition unit: real-time acquisition of strain signals and deformation data of rotating components of the wind turbine through magnetostrictive sensors;

[0067] Data preprocessing unit: pre-processing of the collected strain signals and deformation data, including signal denoising, outlier rejection and data smoothing processing, and using an adaptive filtering algorithm to optimize the original strain signals and deformation data;

[0068] Fault feature extraction unit: extracting fault feature parameters based on the pre-processed strain signals and deformation data, including frequency domain features, time domain features and time-frequency domain features, and generating feature vectors for model input;

[0069] The model construction and training unit: based on the constructed feature vector, a wind turbine fault pre-detection model is constructed, and the model is trained in combination with historical strain signals and deformation data of the wind turbine and a fault case library to identify early fault signs and predict the probability of fault occurrence;

[0070] The real-time monitoring and fault early warning unit: in the operation process of the wind turbine, the strain signals and deformation data of the rotating parts are monitored in real time, the trained wind turbine fault pre-detection model is used for prediction, a fault early warning signal is generated when potential fault signs are detected, and the fault type, location and development trend are output.

[0071] The beneficial effects of the present application are:

[0072] The present application can effectively detect the micro-strain and deformation of the rotating parts of the wind turbine during high-load operation by accurately arranging magnetostrictive sensors on the rotating parts of the wind turbine, ensures the stability and measurement accuracy of the sensors in complex operating environments by optimizing the installation position and fixing method of the sensors, ensures the integrity and timeliness of the data by synchronously marking and storing the real-time collected multi-dimensional strain signals and deformation data in combination with a high-precision clock, provides reliable data support for fault monitoring, greatly improves the real-time health status of the key components of the wind turbine, and effectively prevents the occurrence of potential faults.

[0073] The present application greatly improves the signal-to-noise ratio and the accuracy of feature extraction by performing signal noise reduction, outlier rejection, data smoothing processing and adaptive filtering optimization on the collected strain signals and deformation data, comprehensively captures the key fault features in the signals by combining time domain, frequency domain and time-frequency domain feature extraction methods, ensures that the generated feature vector can reflect the potential problems in the operation of the wind turbine, provides high-quality input data for the subsequent fault pre-detection model, significantly improves the detection accuracy and robustness of the model, and can more effectively identify and predict early faults of the wind turbine.

[0074] The present application adopts a Bayesian neural network model, trains the model in combination with historical strain signals and deformation data, and introduces an adaptive kernel function and a dynamic weight update mechanism to enhance the adaptability and prediction ability of the model in processing nonlinear and high-dimensional data, performs hyperparameter optimization and cross-validation through grid search to ensure the generalization ability and stability of the model on different data sets, in actual application, the model performs real-time fault prediction and early warning by monitoring the strain signals and deformation data of the rotating parts in real time, effectively reduces the downtime risk and maintenance cost, and ensures the reliability of the fault early warning based on the detection threshold set based on historical data, and comprehensively improves the operation efficiency and safety of the wind turbine. BRIEF DESCRIPTION OF DRAWINGS

[0075] 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.

[0076] Figure 1 This is a schematic diagram of the monitoring method according to an embodiment of the present invention;

[0077] Figure 2 This is a schematic diagram of the device functional units according to an embodiment of the present invention. Detailed Implementation

[0078] 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.

[0079] 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.

[0080] 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.

[0081] like Figure 1 As shown, the wind turbine fault monitoring method based on operational data includes the following steps:

[0082] S1, Sensor Placement and Installation: Magnetostrictive sensors are placed on the rotating parts of the wind turbine to detect minute strains and deformations of the rotating parts during operation.

[0083] S2, data collection: real-time collection of strain signals and deformation data of rotating parts of wind turbines through magnetostrictive sensors;

[0084] S3, data preprocessing: preprocessing of collected strain signals and deformation data, including signal denoising, outlier removal and data smoothing, and using adaptive filtering algorithm to optimize original strain signals and deformation data to eliminate environmental interference and measurement noise, ensuring the accuracy of subsequent fault analysis;

[0085] S4, fault feature extraction: based on the preprocessed strain signals and deformation data, extract fault feature parameters, including frequency domain features, time domain features and time-frequency domain features, generate feature vectors for model input;

[0086] S5, model construction and training: based on the constructed feature vector, construct a wind turbine fault pre-detection model, and combine the historical strain signals and deformation data of the wind turbine and the fault case library to train the model to identify early signs of failure and predict the probability of failure, including:

[0087] S51, feature vector input: the generated feature vector is used as the input data of the model, and the input data is standardized and normalized to ensure that the feature vector has consistent scale and range before inputting the model, to improve the efficiency and accuracy of model training;

[0088] S52, model construction: based on the processed feature vector, construct a wind turbine fault pre-detection model;

[0089] S53, model training and verification: divide the historical strain signals and deformation data of the wind turbine into training set and verification set, use the data of the training set to train the model through cross-validation and hyperparameter optimization to improve the generalization ability of the model, ensure the good performance of the model on new data, and use the verification set to verify the model, evaluate its accuracy and stability in fault detection and prediction;

[0090] S6, real-time monitoring and fault warning: during the operation of the wind turbine, real-time monitoring of strain signals and deformation data of rotating parts, and through the trained wind turbine fault pre-detection model for prediction, when detecting potential signs of failure, generate a fault warning signal, and output the fault type, location and development trend, to provide timely maintenance suggestions for maintenance personnel;

[0091] Through the above steps, early detection and warning of wind turbine faults are realized, real-time operation data of key components can be accurately obtained and processed, through efficient fault pre-detection model, timely fault warning is provided in actual operation, thereby improving the operation reliability and maintenance efficiency of wind turbines, and reducing the risk of failure.

[0092] The sensor arrangement in S1 includes:

[0093] S11, sensor arrangement position selection: arrange magnetostrictive sensors on the surface or internal predetermined positions of the rotating components (main shaft, gearbox and bearing) of the wind turbine, the predetermined positions are selected according to the stress concentration areas or fatigue-prone areas of the rotating components, to ensure that the sensors can accurately detect the slight strain and deformation of the rotating components during operation;

[0094] S12, sensor fixing method: the magnetostrictive sensors are fixed on the rotating components by non-contact or contact installation methods, the non-contact installation adopts magnetic support or adhesion technology, and the contact installation adopts mechanical clamp or fastener fixing, to ensure the stability and measurement accuracy of the sensors in high-speed operation environment;

[0095] S13, sensor calibration and adjustment: after the installation of the magnetostrictive sensors, preliminary calibration and zero-point adjustment are carried out to compensate for the displacement or error during installation, to ensure the accuracy and consistency of the measurement data;

[0096] Through the above, it is ensured that the magnetostrictive sensors can stably and accurately detect the slight strain and deformation on the rotating components of the wind turbine, and can maintain the measurement accuracy of the sensors in high-speed and complex operation environment, thereby providing reliable data support for fault monitoring and improving the accuracy of fault detection and the effectiveness of early warning.

[0097] The data collection in S2 includes:

[0098] S21, strain signal collection: real-time collection of strain signals of the rotating components of the wind turbine through magnetostrictive sensors, including axial strain, radial strain and torsional strain, through sensing the strain changes of the rotating components under different working conditions (such as starting, accelerating, decelerating and constant speed operation), the stress state of the components during operation is captured;

[0099] S22, deformation data collection: collect deformation data of the rotating components through magnetostrictive sensors, including bending deformation, stretching deformation and shearing deformation of the components, real-time monitor the deformation of the rotating components under high load and high speed, to ensure that the subtle structural changes caused by fatigue, overload or abnormal operation are captured;

[0100] S23, data collection synchronization and storage: to ensure the synchronization of strain signals and deformation data, high-precision clock is used to mark the strain signals and deformation data with time stamp, and real-time synchronization storage is carried out in local or cloud data storage unit;

[0101] Through the above, the precise monitoring of the rotating parts of the wind turbine under various operating conditions is ensured, and comprehensive structural health information is provided by capturing axial strain, radial strain, torsional strain signals, and bending, stretching, and shearing deformation. High-precision time synchronization and data storage ensure data integrity and timeliness, providing a reliable foundation for fault detection and analysis, and improving the accuracy and early warning capability of fault monitoring.

[0102] The data preprocessing in S3 includes:

[0103] S31, signal denoising: wavelet transform is used to process the collected strain signals and deformation data for signal denoising, represented as:

[0104]

[0105] where d j,k is the detail coefficient, a J is the approximation coefficient, ψ j,k (t) and φ J (t) are the wavelet basis function and scaling function respectively, J is the maximum scale of wavelet decomposition, j is the scale parameter, and k is the translation parameter.

[0106] S32, outlier rejection: statistical methods are used to identify and reject outliers in strain signals and deformation data, using the 3σ principle, i.e. for a set of data X i , calculate its mean μ and standard deviation σ, any value X i deviating from the mean more than 3 times the standard deviation is considered an outlier and is rejected, represented as:

[0107] |X i -μ|>3σ then X i is an outlier;

[0108] S33, data smoothing: sliding average algorithm is used to smooth the signal to eliminate short-term fluctuations in the data, making the signal smoother and more continuous, represented as:

[0109]

[0110] where Y(t) is the smoothed signal, X(t) is the original signal, and N is the size of the sliding window.

[0111] S34, adaptive filter algorithm optimization: LMS adaptive filter algorithm is used to optimize the preprocessed strain signals and deformation data, and the filter coefficients are adjusted according to the real-time collected signals to minimize the noise impact, represented as:

[0112] w n+1 = w n + 2μe(n)X(n);

[0113] wherein w n is a filter coefficient, μ is a step factor, e(n) is an error signal, and X(n) is an input signal;

[0114] Through the above, the signal-to-noise ratio and the feature extraction accuracy are improved, more reliable and accurate data basis is provided for subsequent fault analysis, and the accuracy and real-time performance of the wind turbine fault monitoring are ensured.

[0115] The fault feature extraction in S4 includes:

[0116] S41, time domain feature extraction: based on the preprocessed strain signal and deformation data, time domain feature parameters are extracted, the time domain features include mean, root mean square (RMS), skewness and kurtosis, and are expressed as:

[0117]

[0118] wherein Mean is the mean, X i represents the i-th signal sample, and N is the sample number;

[0119]

[0120] wherein RMS is the root mean square;

[0121]

[0122] wherein μ is the mean, σ is the standard deviation, and Skewness is the skewness;

[0123]

[0124] wherein Kurtosis is the kurtosis;

[0125] S42, frequency domain feature extraction: Fourier transform is performed on the preprocessed strain signal and deformation data to extract frequency domain feature parameters, including main frequency, frequency center, bandwidth and spectral energy, and is expressed as:

[0126]

[0127] wherein X(f) is a frequency domain signal, x(t) is a time domain signal, f is a frequency, and j is an imaginary unit;

[0128]

[0129] wherein f c is the frequency center, and |X(f)| is the spectral amplitude;

[0130] S43, time-frequency domain feature extraction: short-time Fourier transform is used to extract time-frequency domain features, denoted as:

[0131]

[0132] where X(t,f) is the time-frequency domain signal, w(t-τ) is the window function, and τ is the time delay;

[0133] S44, feature vector generation: the extracted time domain, frequency domain and time-frequency domain feature parameters are fused to generate a feature vector, which is used as the input of the fault prediction model, denoted as:

[0134] V=[T1,T2,...,T m ,F1,F2,...,F n ,SF1,SF2,...,SF p ];

[0135] where T1,T2,...,T m represent the extracted time domain feature parameters, F1,F2,...,F n represent the extracted frequency domain feature parameters, and SF1,SF2,...,SF p represent the extracted time-frequency domain feature parameters;

[0136] Through the above, the all-around information of the signal can be captured, ensuring that the key fault features are not missed, thereby improving the accuracy and robustness of the fault detection model. Combined with multiple feature extraction methods, the early faults of the wind turbine can be more effectively identified and predicted, improving the reliability and safety of the system.

[0137] The feature vector input in S51 includes:

[0138] S511, standardization processing of feature vector: the generated feature vector is standardized to ensure that different features have the same mean and standard deviation, so as to eliminate the dimensional difference between features, denoted as:

[0139]

[0140] where V i is the i-th feature in the feature vector, μ i is the mean of the feature V i , σ i is the standard deviation of the feature V i , and V i ' is the standardized feature value;

[0141] S512, normalization processing of feature vector: the standardized feature vector is normalized to scale the feature values to a unified range, so as to enhance the training effect of the model, denoted as:

[0142]

[0143] wherein V i is the normalized feature value, min(V i ′) and max(V i ′) are the minimum and maximum values of the feature V i ′, respectively;

[0144] S513, construction of model input data: input the normalized and standardized feature vector V i ′ into the wind turbine fault pre-detection model, to ensure that the model can be trained and predicted based on the cleaned and optimized data;

[0145] Through the above, the consistency and comparability of the model input data are ensured, the deviation caused by the dimensional difference between different features is effectively avoided, the training efficiency and stability of the model are enhanced, and the model can perform well in learning ability and fault detection precision on diversified input data, thereby improving the overall reliability and precision of the wind turbine fault monitoring.

[0146] The wind turbine fault pre-detection model in S52 adopts a Bayesian neural network (BNN) model, which includes:

[0147] S521, Bayesian inference process: in the fault detection of the wind turbine, in order to better capture the historical fault patterns, the Bayesian inference process is improved to introduce a weighted prior distribution, so that the model is more biased towards known fault patterns, and the prediction ability in complex environments is improved, which is expressed as:

[0148]

[0149] wherein p(w|D) is the posterior distribution of the weight w given the data D, p(D|w) is the likelihood function, indicating the probability of the data D given the weight w, p(w) is the prior distribution of the weight, Z is the normalization constant, ensuring that the prior distribution is a probability distribution, and a, g k are adjustment coefficients for adjusting the relationship between the weight and the historical data weight, w k is the model weight corresponding to the historical fault data, and K is the number of historical fault data or reference fault patterns;

[0150] S522, likelihood estimation based on adaptive kernel function: an adaptive kernel function is introduced to adaptively adjust the variance according to the position of the data points, so that the model can more accurately process complex fault data, which is expressed as:

[0151]

[0152] Where p(D|w) is the likelihood function, representing the probability of data D under weight w. If it follows a normal distribution, then the observed value y is... i In the predicted value f(x) i The probability of being near w). For variance, Let y be the baseline variance, μ be the mean vector of the data, l be the kernel width parameter, N be the number of data samples, and y be the mean vector of the data. i Let be the observation value of the i-th sample;

[0153] S523, Predictive Distribution Calculation: The predictive distribution is approximated using Monte Carlo sampling, handles complex data structures, and provides predictions of fault occurrence, expressed as:

[0154]

[0155] Where p(y) * |x * D) is the prediction distribution, representing the distribution given input x. * Given data D, output y * The probability, M is the number of samples, w m These are the weights sampled from the posterior distribution p(w|D);

[0156] S524, Fault Detection and Uncertainty Quantification: By calculating the variance of predicted values, a quantitative assessment of uncertainty is provided, expressed as:

[0157]

[0158]

[0159] Among them, Uncertainty(x) * ) is a measure of uncertainty, representing the uncertainty in the input x * Variance of time prediction (x * f(x) represents the mean of the predicted values. * ,w m ) represents the m-th sampling weight w m The predicted value below;

[0160] S525, Model Update and Adaptive Adjustment: The model is updated in real-time through incremental learning and adaptive adjustment of the learning rate to maintain its efficiency, as shown below:

[0161]

[0162] Among them, w t+1 For the updated weights, wt Let η be the current weight, η be the learning rate, and D be the adaptive weight over time step t. t Here is the data at the current time step t, where η0 is the initial learning rate and λ is the learning rate decay factor.

[0163] The above-mentioned features enhance the model's adaptability and accuracy in handling nonlinear and high-dimensional data. They also provide quantification of the uncertainty of prediction results, which helps to make more reliable decisions when detecting faults in the early stages. The model can dynamically adjust itself to cope with constantly changing fault modes, improving the robustness of the system and the accuracy of fault early warning. This provides strong support for the efficient operation and maintenance of wind turbines.

[0164] Model training and validation in S53 includes:

[0165] S531, Dataset partitioning: The historical strain signals and deformation data of wind turbines are divided into training set and validation set according to a predetermined ratio;

[0166] S532, Cross-validation: During model training, k-fold cross-validation is used to divide the training set into k subsets. K-1 subsets are used for training, and the remaining subset is used for validation. This process is repeated k times, each time selecting a different subset as the validation set. The final average is used as the evaluation criterion for model performance, expressed as:

[0167]

[0168] Where k is the number of folds in the cross-validation, with a value of 5 or 10, and accuracy... j Let be the model accuracy on the j-th validation set;

[0169] S533, Hyperparameter Optimization: Hyperparameter optimization is performed using grid search, represented as:

[0170]

[0171] Where, θ * For the optimal combination of hyperparameters, arg min finds the parameter values ​​that minimize the subsequent expressions, θ represents the current combination of hyperparameters, and Θ... grid The predefined hyperparameter grid contains all combinations of hyperparameters to be evaluated, and k is the number of folds for cross-validation. The loss function is used to measure the model's predicted values. Compared with the true value y j The difference between them (mean squared error), The output value y is obtained by predicting on the j-th fold validation set using the hyperparameter combination θ. j The true label on the j-th fold validation set;

[0172] The above methods enable a comprehensive evaluation of all possible hyperparameter combinations, avoiding the omission of optimal configurations. At the same time, cross-validation effectively prevents overfitting, thereby ensuring stable model performance on different datasets and improving the reliability and prediction accuracy of the model in practical applications. This provides strong technical support for early fault detection and prevention of wind turbines.

[0173] Real-time monitoring and fault warning in S6 include:

[0174] S61, Real-time signal monitoring: During the operation of the wind turbine, the strain signals and deformation data of rotating components (such as main shaft, gearbox, bearings, etc.) are monitored in real time. These signals are continuously collected by sensors to ensure timely acquisition of dynamic stress and deformation information of rotating components under different operating conditions.

[0175] S62, Comparison and analysis of data and model prediction results: Input the strain signals and deformation data collected in real time into the trained wind turbine fault pre-detection model, and output the probability of the current fault state or health risk.

[0176] S63, Fault Warning Generation and Information Output: When the real-time prediction results show that there is a potential fault (the predicted fault probability exceeds the set detection threshold), a fault warning signal is immediately generated, including the type of fault (such as bearing fault, gearbox fault, etc.), the location of the fault, and the development trend of the fault (such as the speed of fault deterioration).

[0177] The above measures reduce downtime and maintenance costs, extend equipment lifespan, and ensure the reliability and accuracy of early warnings through real-time prediction model comparison with preset thresholds, making wind turbine operation more efficient, intelligent, and safe.

[0178] The detection threshold is set based on statistical analysis of historical data, specifically including:

[0179] Calculate the mean and standard deviation of the characteristics: Extract strain signals and deformation data from historical data, calculate the mean and standard deviation, and use them to describe the characteristic distribution under normal operating conditions;

[0180] Setting a probability threshold: Based on the mean and standard deviation, determine the probability distribution of feature values ​​under normal conditions, and identify anomalies by setting a probability threshold, as follows:

[0181] T = μ + z α / 2 ·σ;

[0182] Where α is the confidence level, with a value of 0.05 (corresponding to a 95% confidence interval) or 0.01 (corresponding to a 99% confidence interval), and z α / 2The critical value of the standard normal distribution, T is a detection threshold, and the characteristic value exceeding the threshold will be regarded as an anomaly.

[0183] As Figure 2 The wind turbine fault monitoring device based on operation data is used to implement the wind turbine fault monitoring method based on operation data, and comprises the following units:

[0184] A sensor arrangement and installation unit: a magnetostrictive sensor is arranged on a rotating component of a wind turbine to detect the slight strain and deformation of the rotating component during operation;

[0185] A data acquisition unit: real-time acquisition of strain signals and deformation data of the rotating component of the wind turbine through the magnetostrictive sensor;

[0186] A data preprocessing unit: preprocessing of the collected strain signals and deformation data, including signal denoising, outlier rejection and data smoothing processing, and using an adaptive filtering algorithm to optimize the original strain signals and deformation data;

[0187] A fault feature extraction unit: extracting fault feature parameters based on the preprocessed strain signals and deformation data, including frequency domain features, time domain features and time-frequency domain features, and generating a feature vector for model input;

[0188] A model construction and training unit: constructing a wind turbine fault pre-detection model based on the constructed feature vector, and training the model in combination with historical strain signals and deformation data of the wind turbine and a fault case library to identify early fault signs and predict the probability of fault occurrence;

[0189] A real-time monitoring and fault warning unit: real-time monitoring of the strain signals and deformation data of the rotating component during the operation of the wind turbine, predicting through the trained wind turbine fault pre-detection model, generating a fault warning signal when potential fault signs are detected, and outputting the fault type, location and development trend.

[0190] The present application encompasses any alternative, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0191] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as falling within the protection scope of the present application.

Claims

1. A method for monitoring wind turbine faults based on operational data, characterized in that, Includes the following steps: S1, Sensor Placement and Installation: Magnetostrictive sensors are placed on the rotating parts of the wind turbine to detect minute strains and deformations of the rotating parts during operation. S2, Data Acquisition: Real-time acquisition of strain signals and deformation data of rotating components of wind turbines using magnetostrictive sensors; S3, Data Preprocessing: The acquired strain signals and deformation data are preprocessed, including signal noise reduction, outlier removal and data smoothing, and an adaptive filtering algorithm is used to optimize the original strain signals and deformation data. S4, Fault Feature Extraction: Based on the preprocessed strain signal and deformation data, extract fault feature parameters, including frequency domain features, time domain features and time-frequency domain features, and generate feature vectors for model input; S5, Model Building and Training: Based on the constructed feature vectors, a wind turbine fault pre-detection model is built. This model is then trained using historical strain signals and deformation data from the wind turbine, along with a fault case library, to identify early fault signs and predict the probability of fault occurrence. Specifically, this includes: S51, Feature Vector Input: The generated feature vectors are used as input data for the model, and the input data is standardized and normalized. S52, Model Construction: Based on the processed feature vectors, construct a wind turbine fault pre-detection model; S53, Model Training and Validation: The historical strain signals and deformation data of the wind turbine are divided into training set and validation set. The model is trained using the data in the training set through cross-validation and hyperparameter optimization, and the model is validated using the validation set. S6, Real-time monitoring and fault early warning: During the operation of the wind turbine, the strain signals and deformation data of the rotating components are monitored in real time, and the fault pre-detection model of the wind turbine is used for prediction. When signs of potential faults are detected, a fault early warning signal is generated and the fault type, location and development trend are output.

2. The wind turbine fault monitoring method based on operational data according to claim 1, characterized in that, The sensor arrangement and installation in S1 includes: S11, Sensor placement selection: The magnetostrictive sensor is placed at a predetermined position on the surface or inside the rotating parts of the wind turbine. S12, Sensor fixing method: The magnetostrictive sensor is fixed to the rotating part by non-contact or contact mounting. Non-contact mounting uses magnetic brackets or adhesive technology, while contact mounting uses mechanical clamps or fasteners. S13, Sensor Calibration and Adjustment: After the magnetostrictive sensor is installed, preliminary calibration and zero-point adjustment are performed to compensate for displacement or errors during the installation process.

3. The wind turbine fault monitoring method based on operational data according to claim 1, characterized in that, The data acquisition in S2 includes: S21, Strain signal acquisition: Real-time acquisition of strain signals of rotating components of wind turbine generator through magnetostrictive sensors, including axial strain, radial strain and torsional strain; S22, Deformation data acquisition: Deformation data of the rotating component is acquired through a magnetostrictive sensor, including bending deformation, stretching deformation and shear deformation of the component; S23, Data Acquisition Synchronization and Storage: The acquired strain signals and deformation data are timestamped and stored in real time in local or cloud data storage units.

4. The wind turbine fault monitoring method based on operational data according to claim 1, characterized in that, The data preprocessing in S3 includes: S31, Signal Denoising: Wavelet transform is used to perform signal denoising processing on the acquired strain signal and deformation data; S32, Outlier Removal: Identifying and removing outliers in strain signals and deformation data using statistical methods; S33, Data smoothing: The signal is smoothed using a moving average algorithm; S34, Adaptive Filtering Algorithm Optimization: The preprocessed strain signal and deformation data are optimized using the LMS adaptive filtering algorithm.

5. The wind turbine fault monitoring method based on operational data according to claim 1, characterized in that, The fault feature extraction in S4 includes: S41, Time-domain feature extraction: Based on the preprocessed strain signal and deformation data, extract time-domain feature parameters, including mean, root mean square value, skewness and kurtosis. S42, Frequency Domain Feature Extraction: Perform Fourier transform on the preprocessed strain signal and deformation data to extract frequency domain feature parameters, including main frequency, frequency center, bandwidth and spectral energy. S43, Time-frequency domain feature extraction: Short-time Fourier transform is used to extract time-frequency domain features; S44, Feature Vector Generation: The extracted time-domain, frequency-domain, and time-frequency-domain feature parameters are fused to generate a feature vector, which is used as input to the fault pre-detection model.

6. The wind turbine generator fault monitoring method based on operational data according to claim 5, characterized in that, The feature vector input in S51 includes: S511, Standardization of feature vectors: Standardize the generated feature vectors to eliminate the dimensional differences between features. S512, Normalization of eigenvectors: Normalize the standardized eigenvectors to scale the eigenvalues ​​to a uniform range. S513, Constructing the model input data: The feature vector V, after standardization and normalization... i "Input into the wind turbine fault pre-detection model." 7. The wind turbine fault monitoring method based on operational data according to claim 6, characterized in that, The wind turbine fault pre-detection model in S52 adopts a Bayesian neural network model, which includes: S521, Bayesian inference process: In wind turbine fault detection, a weighted prior distribution is introduced, represented as: Where p(w|D) is the posterior distribution of the weights w given the data D, p(D|w) is the likelihood function, p(w) is the prior distribution of the weights, Z is the normalization constant, and α, γ k w is the adjustment coefficient. k The model weights are those corresponding to historical fault data, where K is the number of historical fault data or reference fault modes. S522, Likelihood Estimation Based on Adaptive Kernel Function: An adaptive kernel function is introduced to adaptively adjust the variance according to the location of the data points, expressed as: Where p(D|w) is the likelihood function. It follows a normal distribution. For variance, Let y be the baseline variance, μ be the mean vector of the data, l be the kernel width parameter, N be the number of data samples, and y be the mean vector of the data. i Let be the observation value of the i-th sample; S523, Predictive Distribution Calculation: The predictive distribution is approximated using Monte Carlo sampling, handles complex data structures, and provides predictions of fault occurrence, expressed as: Where p(y) * |x * D) is the prediction distribution, representing the distribution given input x. * Given data D, output y * The probability, M is the number of samples, w m These are the weights sampled from the posterior distribution p(w|D); S524, Fault Detection and Uncertainty Quantification: By calculating the variance of predicted values, a quantitative assessment of uncertainty is provided, expressed as: Among them, Uncertainty(x) * ) is a measure of uncertainty, representing the uncertainty in the input x * Variance of time prediction f(x) is the mean of the predicted values. * ,w m ) represents the m-th sampling weight w m The predicted value below; S525, Model Update and Adaptive Adjustment: The model is updated in real time through incremental learning and adaptive adjustment of the learning rate, as shown below: Among them, w t+1 For the updated weights, w t Let η be the current weight, η be the learning rate, and D be the learning rate. t Here is the data at the current time step t, η0 is the initial learning rate, and λ is the learning rate decay factor.

8. The wind turbine fault monitoring method based on operational data according to claim 7, characterized in that, The model training and validation in S53 includes: S531, Dataset partitioning: The historical strain signals and deformation data of wind turbines are divided into training set and validation set according to a predetermined ratio; S532, Cross-validation: During model training, k-fold cross-validation is used to divide the training set into k subsets. k-1 subsets are used for training, and the remaining subset is used for validation. This process is repeated k times, with a different subset selected as the validation set each time. The average value is then used as the evaluation standard for model performance. S533, Hyperparameter Optimization: Hyperparameter optimization is performed using grid search, represented as: Where, θ * For the optimal combination of hyperparameters, arg min finds the parameter values ​​that minimize the subsequent expressions, θ represents the current combination of hyperparameters, and Θ... grid For a predefined hyperparameter grid, k is the number of folds for cross-validation. For loss function, The output value y is obtained by predicting on the j-th fold validation set using the hyperparameter combination θ. j Let be the true label on the j-th fold validation set.

9. The wind turbine fault monitoring method based on operational data according to claim 1, characterized in that, The real-time monitoring and fault early warning in S6 include: S61, Real-time signal monitoring: During the operation of the wind turbine, the strain signals and deformation data of the rotating components are monitored in real time. S62, Comparison and analysis of data and model prediction results: Input the strain signals and deformation data collected in real time into the trained wind turbine fault pre-detection model, and output the probability of the current fault state or health risk. S63, Fault Warning Generation and Information Output: When the real-time prediction results show that there is a potential fault, a fault warning signal is immediately generated, including the type of fault, the location of the fault, and the development trend of the fault.

10. A wind turbine generator fault monitoring device based on operational data, used to implement the wind turbine generator fault monitoring method based on operational data as described in any one of claims 1-9, characterized in that, include: Sensor placement and installation unit: Magnetostrictive sensors are placed on the rotating parts of the wind turbine to detect minute strains and deformations of the rotating parts during operation; Data acquisition unit: Real-time acquisition of strain signals and deformation data of rotating components of wind turbine generator through magnetostrictive sensors; Data preprocessing unit: preprocesses the acquired strain signals and deformation data, including signal noise reduction, outlier removal and data smoothing, and uses an adaptive filtering algorithm to optimize the original strain signals and deformation data; Fault feature extraction unit: Extracts fault feature parameters based on preprocessed strain signals and deformation data, including frequency domain features, time domain features and time-frequency domain features, and generates feature vectors for model input; Model building and training unit: Based on the constructed feature vectors, a wind turbine fault pre-detection model is built, and the model is trained by combining historical strain signals and deformation data of wind turbines and fault case library to identify early fault signs and predict the probability of fault occurrence. Real-time monitoring and fault early warning unit: During the operation of the wind turbine, the strain signal and deformation data of the rotating parts are monitored in real time. The fault prediction is performed by a trained wind turbine fault pre-detection model. When signs of potential faults are detected, a fault early warning signal is generated and the fault type, location and development trend are output.

Images