Method, system, medium and equipment for monitoring abnormal state of wind generating set

Abstract

According to the method, the system, the medium and the equipment for monitoring the abnormal state of the wind generating set, a self-adaptive noise reduction mechanism, feature extraction and model classification are combined, so that physical information features are effectively extracted, and accurate state detection is achieved. The method comprises the following steps: firstly, building a signal acquisition system, acquiring an acoustic signal in real time, and executing direct current offset removal preprocessing to eliminate low-frequency noise; secondly, noise reduction of the original sound signal is completed by using a self-adaptive noise reduction method; then, extracting the MFCC characteristics of the noise reduction signal; then, calculating an MFCC inter-frame correlation feature and an aerodynamic-harmonic imbalance factor, and forming a fusion feature vector; and finally, constructing a physical neural network model based on physical information to carry out abnormal state detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and particularly relates to a wind turbine unit abnormal state monitoring method, system, medium and equipment. BACKGROUND

[0002] As the core equipment of renewable energy, the long-term stable operation of the wind turbine unit is crucial to guarantee the safety of the power grid and improve the power generation efficiency. The acoustic signals generated by the unit during operation contain rich state information, which can reflect key operating parameters such as blade aerodynamic characteristics, bearing wear, gear meshing state, etc. Compared with traditional vibration signals, acoustic signals have unique advantages such as non-contact measurement, global perception and flexible installation, and have become a research hotspot in the field of state monitoring.

[0003] However, the operation environment of the wind farm is extremely complex, and the acoustic signal is easily affected by various interferences: the wind flow impact produces wideband background noise, the mechanical resonance causes harmonic interference, the electromagnetic equipment introduces high-frequency spikes, and the environmental factors (such as rain noise and bird chirping) cause transient disturbance. These noises have significant non-stationary characteristics and time-varying statistical characteristics, and are highly overlapped with the target signal in the frequency domain, which leads to serious deterioration of the signal-to-noise ratio and difficulty in feature extraction. In terms of signal denoising, the existing technology has obvious deficiencies. The traditional spectral subtraction method is based on the assumption of stationary noise, and when processing non-stationary noise, it will produce "music noise" and easily weaken the harmonic characteristics of blade faults. Although the Wiener filter optimizes the minimum mean square error, it is not adaptive to the time-varying characteristics of noise. The wavelet threshold denoising has an advantage in multi-scale analysis, but the threshold selection depends on experience, and in the wideband weak amplitude signal, it is easy to cause over-smoothing and lose subtle fault features. Advanced signal decomposition methods such as variational mode decomposition (VMD) can effectively separate signal components, but the parameter adjustment is complex and the calculation cost is high, which is not conducive to real-time application. Deep learning methods such as denoising autoencoder perform well in some scenarios, but require a large amount of labeled data, and have limited generalization ability under specific wind noise.

[0004] At the feature extraction level, existing methods have significant limitations. Traditional feature extraction mainly relies on time-domain statistics (such as root mean square, kurtosis, and peak factor) and frequency-domain features (such as power spectral density and spectral centroid). These features, although simple to calculate, lack clear physical meaning and are not sensitive enough to early faults. For example, the kurtosis index is sensitive to impact-type faults but is easily disturbed by noise, leading to false positives; the spectral centroid can reflect energy distribution changes but cannot distinguish between fault types. Mel-frequency cepstral coefficients (MFCC) have shown excellent performance in speech recognition and have been introduced into mechanical fault diagnosis, but they still have limitations in the application of fan acoustic signals: first, traditional MFCC mainly simulates human auditory characteristics and fails to fully utilize the physical laws of rotating machinery; second, single-frame MFCC features cannot effectively capture the periodic dynamic characteristics of sound signals; third, there is a lack of deep integration with aerodynamic acoustic mechanisms, resulting in insufficient sensitivity to aerodynamic-related faults.

[0005] Traditional machine learning methods (such as support vector machines and random forests) rely on manual feature design and selection and are difficult to handle the temporal dependencies of acoustic signals. Deep learning methods (such as convolutional neural networks and recurrent neural networks) can automatically learn feature representations, but face challenges in fan acoustic monitoring: first, data-driven models lack physical consistency and may produce incorrect judgments that violate physical laws; second, a large amount of labeled data is needed for training, but fault samples are scarce in industrial practice; third, the model decision-making process lacks explainability, affecting the trust of engineers. Physical information neural networks (PINN) have made progress in fluid mechanics and materials science, but their application in fan acoustic monitoring has not been thoroughly explored, especially how to effectively integrate physical knowledge such as aerodynamic acoustic equations and structural dynamics principles remains an open question.

[0006] In summary, existing acoustic monitoring technology for wind turbine generators has three major problems: feature extraction and physical mechanism disconnection: traditional features lack targeted design for specific physical processes of the fan (such as blade passage effect, aerodynamic noise generation, and structural modal coupling), resulting in insufficient fault sensitivity; state recognition models lack physical guidance: data-driven models may learn false correlations and have decreased generalization ability when operating conditions change; noise reduction, feature extraction, and state recognition are isolated: each stage is optimized independently and cannot form an end-to-end collaboration based on physical knowledge.

[0007] These problems severely restrict the practical application of acoustic monitoring technology in wind turbine condition monitoring. Therefore, a new method that can deeply integrate physical knowledge is urgently needed to introduce physical constraints from feature design to model construction, improving the accuracy and robustness of the monitoring system. SUMMARY

[0008] In view of this, the present invention provides a method, system, medium and equipment for monitoring abnormal conditions of wind turbine generator sets, so as to at least solve the problem that the existing acoustic monitoring technology for wind turbine generator sets is difficult to effectively suppress noise and accurately extract features.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for monitoring abnormal conditions of a wind turbine generator set includes the following steps: S1. Real-time acquisition of raw sound signals from wind turbine generator sets during operation. ; S2. A dual-time-scale noise estimation method is used, which involves analyzing the original sound signal. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original sound signal and obtain the noise-reduced acoustic signal. , ,in The signal length; S3. Extract MFCC statistical features, MFCC inter-frame correlation features, and aerodynamic-harmonic imbalance factors from the acoustic signal of the wind turbine generator after noise reduction to obtain a hybrid input feature vector. ; S4. Mix the input feature vectors Anomaly detection is performed by inputting a pre-trained physical information-based neural network PINN, where the physical information-based neural network PINN is trained using denoised acoustic signals.

[0010] Preferably, after obtaining the original sound signal in S2, preprocessing is performed, specifically including: Obtained through time-frequency transformation The corresponding time-frequency representation matrix The time-frequency transformation methods specifically include: Short Time Fourier Transform (STFT), Fast Fourier Transform (FFT), wavelet transform, or wavelet packet transform. In the time-frequency representation matrix Based on this, the amplitude spectrum is calculated. : .

[0011] Preferably, in S2, the original sound signal is analyzed. corresponding amplitude spectrum statistical distribution along time axis, respectively calculating short-time noise amplitude and long-time noise amplitude The specific content includes: Adopting double-time-scale noise estimation, by analyzing the amplitude spectrum statistical distribution along time axis, respectively calculating short-time noise amplitude and long-time noise amplitude : ; ; In the formula, and respectively represent the first quantile and the second quantile; and are the quantile thresholds of short-time and long-time noise estimation, respectively; reflects the local background level of noise, captures the overall trend and possible sudden changes of noise. Preferably, in S2, the estimated noise amplitude is estimated based on

[0012] and The specific content includes: Introducing dynamic weights to balance the contributions of short-time and long-time estimations: ; ; ; In the formula, is the initial weight, is the time frame index, , is the frequency point index, , , is the number of points of time-frequency transformation, is the imaginary unit, and are the upper and lower bounds of range, satisfying 0≤ < ≤1; Then the basic noise amplitude is: ; Based on the basic noise amplitude , the signal-to-noise amplitude ratio is calculated, which is used to quantify the relative strength of signal and noise:​ ; based on calculating the minimum amplitude threshold : ; wherein, is a scaling factor, controlling the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened; introducing a dynamic noise scaling factor to dynamically adjust the noise amplitude: ; wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; based on the dynamic noise scaling factor calculating the estimated noise amplitude : .

[0013] Preferably, in S2, the sound original signal is reconstructed according to to obtain the acoustic signal after noise reduction The specific content includes: According to and calculate the noise reduction spectrum amplitude : ; Preserve the original phase to restore the complex spectrum after noise reduction : ; Use inverse transform to reconstruct the time domain signal from the noise reduction spectrum : ; wherein, is an overlap-add window function.

[0014] Preferably, the specific content of S3 includes: Mel-frequency conversion is performed on the acoustic signal after noise reduction to obtain the acoustic signal under the Mel frequency : ; Filtering is performed on by a Mel filter bank: ; Extracting MFCC features : ; In the formula, This indicates the index of the filter in the Mel filter bank. The number of filter banks. Indicates the first The center frequency of the Mel filter; Extract within a time window The frame MFCC feature vectors form a matrix: ; in, express The MFCC feature vector of a frame is a The matrix; each frame extracted If there are MFCC features, then the MFCC inter-frame correlation matrix Defined as the Pearson correlation coefficient matrix between MFCC vectors of each frame:

[0015] in, Let be a P×P symmetric matrix, representing the inter-frame correlation matrix of MFCCs. This matrix is ​​used to quantify the linear similarity between MFCC vectors in different frames and provides temporal context information. Based on matrix Calculate each item Reflection Frame and frame Similarity, Represents the element in the matrix, i.e., the first element. Frame and the Pearson correlation coefficient between frame MFCC feature vectors, ranging from [-1, 1]; subscript and These represent the row index and column index, respectively, from 1 to P, where each element... Indicates the first Frame and the Correlation coefficients between frame MFCC feature vectors: ; in, The numerator represents the covariance of two MFCC vectors and is used to measure common variance. These are the indices of the MFCC vector elements, from 1 to L; The i-th frame represents the MFCC vector of the i-th frame. Each element. Indicates the first The mean of the frame MFCC vector is used to center the data; Indicates the first The mean of the frame MFCC vector; Since the correlation matrix is symmetric, i.e. , and diagonal elements are all 1, i.e. , only the elements in the lower triangular part are extracted to construct the feature vector to avoid redundancy: ; Capture the dynamic correlation pattern of the sound signal in multiple time scales, provide the model with more temporal context information than single-frame features; further, the Gaussian weighted order spectrum average method is used to extract the amplitude of the wind turbine order harmonic: ; Where the Gaussian weight function is defined as: ; In the formula, represents the amplitude of the order harmonic, which is used to quantify the strength of the BPF and its nearby harmonics, reflecting the periodic vibration induced by blade rotation and aerodynamic imbalance; represents the harmonic order, indexing the harmonic sequence, where is the maximum order; represents the Gaussian weight, which is a real number in the range [0, 1], used to emphasize the central harmonic and attenuate the edge offset; : represents the order spectrum function, which extracts the periodic components of rotating machinery; represents the blade; represents the offset index, ranging from to ; to define the neighborhood offset within the weighted window, used for smoothing calculation; represents the fundamental frequency of the turbine blade rotation; represents the standard deviation controlling the width of the Gaussian function; Based on the Ffowcs Williams-Hawkings equation model, the broadband noise energy generated by the interaction between airflow and blades of the wind turbine is calculated: ; Where represents the air density, taking into account the influence of medium properties on noise propagation; represents the tip speed, represents the sensor-to-tip distance; Calculate the aerodynamic-harmonic imbalance factor as a normalized deviation: ; By weighted sum and fusion of features, form a hybrid feature vector : ; where, represents the aerodynamic-harmonic imbalance factor, to quantify the deviation of harmonic intensity from the predicted noise, for indicating abnormality; represents the weight of MFCC statistical weight, to balance the contribution; represents the weight of inter-frame correlation weight, to balance the temporal feature; represents the weight of imbalance factor weight; represents the average factor, to normalize the harmonic summation, ensuring scale invariance.

[0016] Preferably, the specific content of S4 includes: The basic network structure of the physical neural network PINN based on physical information adopts a feedforward neural network as the backbone network, for learning the complex mapping from the feature to the health condition; The physical constraint is introduced to the physical neural network PINN based on physical information, realized by a physical loss function , where the loss function is calculated from the feature or its derived quantity, measuring the degree of violation of the network output or intermediate state and the physical principle; Where the physical constraints include: aerodynamic-harmonic conservation constraint : ; where, represents the aerodynamic-harmonic conservation loss, quantifying the deviation of the imbalance factor; represents the health benchmark value; MFCC inter-frame correlation stability constraint : ; where, represents the correlation feature vector of the th time window; represents the total loss function of the correlation feature vector of the previous time window : ; where: represents the weight hyperparameter of the imbalance factor deviation; is the weight hyperparameter of the inter-frame correlation stability constraint; is the supervised learning loss: ; where, represents the true label of the cth class; represents the probability of belonging to the cth class predicted by the network, from the Softmax output; The natural pairs are used to predict probabilities; C represents the total number of categories; During training, the backpropagation algorithm minimizes... At the same time, the weights are adjusted according to the labels and physical constraints. and Adjustments are made to ultimately achieve anomaly monitoring.

[0017] Preferably, it also includes: S5. Generating alarm signals based on detection results and generating decision suggestions, combined with a real-time feedback mechanism, to achieve closed-loop risk management.

[0018] A wind turbine generator abnormal state monitoring system, comprising: The signal acquisition module is used to acquire the raw sound signals of the wind turbine generator in real time during operation. ; The noise reduction module is used to perform noise estimation using a dual-timescale method by analyzing the original sound signal. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original sound signal and obtain the noise-reduced acoustic signal. , ,in The signal length; The feature extraction module is used to extract MFCC statistical features, MFCC inter-frame correlation features, and aerodynamic-harmonic imbalance factors from the acoustic signal of the wind turbine generator after noise reduction, to obtain a hybrid input feature vector. ; The detection module is used to process the mixed input feature vectors. Anomaly detection is performed by inputting a pre-trained physical information-based neural network PINN, where the physical information-based neural network PINN is trained using denoised acoustic signals.

[0019] Preferably, the noise reduction module analyzes the original sound signal. Corresponding amplitude spectrum Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude The specific content of the noise estimation includes: By analyzing the amplitude spectrum of the noise signal along the time axis, the short-time noise amplitude and the long-time noise amplitude are calculated respectively. where and denote the first quantile and the first quantile respectively; and are the quantile thresholds for the short-time and long-time noise estimation respectively; reflects the local background level of the noise, and captures the overall trend and possible abrupt changes of the noise.

[0020] In the noise reduction module, the specific content of the noise amplitude estimation based on includes: The dynamic weight is introduced to balance the contributions of the short-time and long-time estimations: where is the initial weight, is the time frame index, , is the frequency point index, , , is the number of points in the time-frequency transform, is the imaginary unit, and are the upper and lower bounds of the range, satisfying 0≤ < ≤1; then the basic noise amplitude is: Based on the basic noise amplitude , the signal-to-noise amplitude ratio is calculated to quantify the relative strength of the signal and the noise: Based on , the minimum amplitude threshold is calculated:​​​​​​ ; In the formula, is a scaling factor, controlling the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened; A dynamic noise scaling factor is introduced to dynamically adjust the noise amplitude: ; In the formula, is a scaling factor, controlling the sensitivity of the adjustment factor, is the initial weight; Based on the dynamic noise scaling factor The estimated noise amplitude is calculated : .

[0021] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a wind turbine abnormal state monitoring method as described above.

[0022] An electronic device comprising: a processor and a memory storing one or more programs; when the one or more programs are executed by the processor, a wind turbine abnormal state monitoring method as described above is implemented.

[0023] According to the above technical solution, compared with the prior art, the present disclosure provides a wind turbine abnormal state monitoring method, system, medium and device, which realizes effective suppression of non-stationary noise and accurate detection of abnormal state of wind turbine acoustic signals through the innovative combination of data acquisition system and double-time-scale noise estimation and adaptive noise reduction adjustment model abnormal state detection. This method significantly improves the noise reduction performance and diagnostic accuracy, and supports flexible deployment of local embedded, remote cloud or hybrid hardware solutions: the local solution is suitable for low-latency real-time monitoring of small wind farms, the cloud solution supports big data analysis and remote operation and maintenance of large wind farm groups, and the hybrid solution optimizes power consumption and computing efficiency, ensuring stable transmission and robust diagnosis in harsh wind power environments. Specifically, the following beneficial effects are included: 1. Double-time-scale noise estimation and adaptive adjustment: In the time-frequency analysis stage, the short-time and long-time noise amplitudes are extracted and dynamically fused to calculate the SNMR and optimize the noise reduction threshold, accurately capturing the dynamic distribution of wind turbine sound signal noise, overcoming the limitations of static models, and improving the suppression effect of wide-frequency weak-amplitude noise.

[0024] 2. By innovatively combining MFCC features and physical information vectors, a hybrid feature with more physical meaning and discriminability is constructed, improving the sensitivity to specific faults of wind turbines.

[0025] 3. A physical neural network model based on physical information is designed, and the physical laws of aeroacoustics and rotating machinery are embedded as constraint conditions in network training, so as to ensure that the model output conforms to both data distribution and physical laws, and improve detection precision and reliability. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0027] Figure 1 A flowchart of an abnormal state monitoring method of a wind turbine generator set provided by the present application is shown in the figure. Figure 2 A schematic diagram of a hardware system of a wind turbine generator set provided by the embodiment of the present application is shown in the figure. Figure 3 A time-frequency graph of the abnormal sound of the wind turbine generator set collected by the embodiment of the present application is shown in the figure. Figure 4 A schematic diagram of the denoising result of the method provided by the embodiment of the present application is shown in the figure. Figure 5 A schematic diagram of the extracted MFCC features provided by the embodiment of the present application is shown in the figure. Figure 5 (a) Figure 5 (l) is the change curve of the 12 MFCC features respectively in turn. Figure 6 A schematic diagram of the verification result of the neural network model provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0029] The present application provides an abnormal state monitoring method of a wind turbine generator set, comprising the following steps: S1. Real-time acquisition of sound original signals of a wind turbine generator set during operation ; S2. Using a double-time-scale noise estimation method, analyzing the sound original signals corresponding to the first frame and the first Amplitude spectrum of frequency Statistical distribution along time axis, respectively calculating short-time noise amplitude and long-time noise amplitude , estimating estimated noise amplitude based on and , and then reconstructing the original sound signal according to to obtain the acoustic signal after noise reduction , , wherein is the signal length; S3. Extracting MFCC statistical features, MFCC inter-frame correlation features and aerodynamic-harmonic imbalance factors from the acoustic signal after noise reduction of the wind turbine unit to obtain a mixed input feature vector ; S4. Inputting the mixed input feature vector into the trained physical neural network PINN based on physical information for abnormal state detection, wherein the physical neural network PINN based on physical information is trained by the acoustic signal after noise reduction.

[0030] In order to further implement the above technical solutions, after the original sound signal is obtained in S2, the original sound signal is preprocessed, and the specific content includes: obtaining the corresponding time-frequency representation matrix , wherein the time-frequency transformation method specifically includes: short-time Fourier transform STFT, fast Fourier transform FFT, wavelet transform or wavelet packet transform; on the basis of the time-frequency representation matrix , the amplitude spectrum is calculated: .

[0031] It should be noted that: the input signal is the acoustic and vibration signal of the wind turbine unit collected, wherein is the signal length. In the initial stage, the signal is preprocessed to remove the DC offset, so as to effectively eliminate the DC component interference, and obtain the preprocessed time domain signal . Removing the DC offset can be realized by various technical means, including but not limited to traditional mean subtraction, high-pass filtering method to suppress low-frequency DC component, and adaptive filtering technology.

[0032] In order to further implement the above technical solutions, in S2, the corresponding amplitude spectrum of the original sound signal is analyzed, and the statistical distribution along the time axis is calculated, respectively calculating short-time noise amplitude and long-term noise amplitudes The specific content includes: Adopting double-time-scale noise estimation, the short-term noise amplitude and the long-term noise amplitude are calculated respectively by analyzing the amplitude spectrum and the statistical distribution along the time axis: ; ; In the formula, and respectively represent the first quantile and the second quantile; and are the quantile thresholds of the short-term and long-term noise estimation respectively; reflects the local background level of the noise, and captures the overall trend and possible sudden changes of the noise. In order to further implement the above technical solutions, the specific content of estimating the estimated noise amplitude in S2 based on

[0033] and includes: Introducing dynamic weights to balance the contributions of the two estimations of short-term and long-term: ; ; ; In the formula, is the initial weight, is the time frame index, , is the frequency index of the time-frequency representation matrix, , , is the number of points of the time-frequency transform, is the imaginary unit, and are the upper and lower bounds of the range, satisfying 0≤ < ≤1; Then the basic noise amplitude is: ; Based on the basic noise amplitude , the signal-to-noise amplitude ratio is calculated, which is used to quantify the relative intensity of the signal and the noise: ; Based on​ Computing the minimum amplitude threshold : ; wherein, is a scaling factor, controlling the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened; Introducing a dynamic noise scaling factor to dynamically adjust the noise amplitude: ; wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight, calculated according to the above formula ; Computing the estimated noise amplitude based on the dynamic noise scaling factor : .

[0034] It should be noted that: Since the noise of the wind turbine sound signal has the characteristics of short-time burstiness and long-time stationarity coexisting, a single scale noise estimation is difficult to accurately capture its dynamic distribution. To solve this problem, the present method uses double-time scale noise estimation.

[0035] In calculating the short-time and long-time noise amplitudes, the value range can be dynamically adjusted to adapt to different noise environments. and are short-time and long-time noise estimation thresholds, respectively, which can be calculated based on statistical methods, or other methods, including but not limited to mean plus multiple of standard deviation or K-means clustering method, to optimize the noise estimation performance. reflects the local background level of the noise, captures the overall trend and possible burstiness of the noise. To fuse the short-time and long-time noise characteristics, a dynamic weight is introduced to balance the contributions of the two estimates. The weight reflects the average intensity of the signal relative to the long-time noise, used to adjust the contribution ratio of the short-time and long-time estimates.

[0036] In order to effectively denoise, the present application calculates the signal noise amplitude ratio SNMR based on the basic noise amplitude, which is used to quantify the relative intensity of the signal and the noise. In order to avoid excessive inhibition of the signal during the denoising process, the minimum amplitude threshold is further calculated based on the SNMR.

[0037] In order to further implement the above technical solutions, the sound original signal is reconstructed according to in S2, to obtain the denoised acoustic signal Details include: According to And Calculate the amplitude of the noise reduction spectrum : ; Retain the original phase, restore the complex spectrum after noise reduction : ; Using the inverse transform to reconstruct the noise reduction spectrum into a time domain signal : ; In the formula, The overlap window function.

[0038] In order to further implement the above technical solutions, the specific content of S3 includes: The acoustic signal after noise reduction Mel frequency conversion is carried out to obtain the acoustic signal under the Mel frequency : ; Filtering by Mel filter bank : ; Extract MFCC features : ; In the formula, Indicates the index of the filter in the Mel filter bank, The number of filter banks, Indicates the center frequency of the Mel filter; In a time window, extract Frame MFCC feature vectors to form a matrix: ; Where, Indicates the Frame MFCC feature vector, which is a Matrix; Each frame extracts MFCC features, so the MFCC inter-frame correlation matrix Is defined as the Pearson correlation coefficient matrix between the frame MFCC vectors:

[0039] Where, is a P x P symmetric matrix, denoting the inter-frame MFCC correlation matrix, which is used to quantify the linear similarity between different frames of MFCC vectors, providing temporal context information, based on the matrix is calculated, each term reflects the similarity between frames and , and denotes the element in the matrix, i.e., the Pearson correlation coefficient between the MFCC feature vector of the th frame and the th frame, with the range [-1, 1]; the subscripts and denote the row index and the column index, respectively, from 1 to P, where each element denotes the correlation coefficient between the MFCC feature vector of the th frame and the th frame: ; wherein denotes the covariance of two MFCC vectors, the numerator part, which is used to measure the common variation; is the index of the MFCC vector element, from 1 to L; denotes the th element of the th frame MFCC vector. denotes the mean of the th frame MFCC vector, which is used to center the data; denotes the mean of the th frame MFCC vector. Since the correlation matrix is symmetric, i.e., , and the diagonal elements are all 1, i.e., , only the elements in the lower triangular part are extracted to construct the feature vector to avoid redundancy: ; captures the dynamic correlation pattern of the sound signal in multiple frames of time scale, providing the model with temporal context information beyond the single-frame features; further, a Gaussian weighted order spectrum average method is adopted to extract the amplitude of the wind turbine order harmonics: ; wherein the Gaussian weight function is defined as: ; In the formula, denotes the The amplitude of the harmonic order is used to quantify the strength of the BPF and its nearby harmonics, which reflects the periodic vibration induced by blade rotation and aerodynamic unbalance; it can be used as a core component of physical features to help detect abnormal states. denotes the harmonic order, indexing the harmonic sequence, e.g. is the fundamental, is the second harmonic, where is the maximum order; denotes the Gaussian weight, a real number ranging in [0, 1], used to emphasize the central harmonics and decay the edge offset; denotes the order spectrum function, extracting the periodic components of the rotating machinery; denotes the blade; denotes the offset index, ranging from to ; to define the neighborhood offset within the weighted window, used for smoothing calculation; denotes the fundamental frequency of the unit blade rotation; denotes the standard deviation controlling the width of the Gaussian function; Based on the model of the Ffowcs Williams-Hawkings equation, the wideband noise energy generated by the interaction of airflow and blades of the wind turbine is calculated: ; where, denotes the air density, to consider the influence of medium properties on noise propagation; denotes the tip speed, denotes the sensor-to-tip distance; The aerodynamic-harmonic unbalance factor is calculated as a normalized deviation: ; By weighted and fused features, a hybrid feature vector is formed: ; where, denotes the aerodynamic-harmonic unbalance factor, to quantify the deviation of harmonic strength and predicted noise, used to indicate abnormalities; denotes the weight of the MFCC statistical weight, to balance the contribution; denotes the weight of the inter-frame correlation weight, to balance the timing features; denotes the weight of the unbalance factor weight; denotes the average factor, normalizing the harmonic summation to ensure scale invariance.

[0040] In order to further implement the above technical solutions, the specific content of S4 includes: The basic network structure of the physical neural network PINN based on physical information adopts a feedforward neural network as the backbone network, which is used to learn the complex mapping from the feature to the health condition; The physical constraints are introduced into the physical neural network PINN based on physical information, and are realized by a physical loss function , wherein the loss function is calculated by the feature or its derived quantity, which measures the degree of violation of the network output or intermediate state and the physical principle; The physical constraints include: aerodynamic-harmonic conservation constraint : ; wherein, represents the aerodynamic-harmonic conservation loss, which quantifies the deviation of the imbalance factor; represents the health benchmark value, which is usually set to 0, indicating the ideal state of no imbalance; MFCC inter-frame correlation stability constraint : ; wherein, represents the correlation feature vector of the i-th time window; represents the correlation feature vector of the previous time window; The total loss function of the correlation feature vector is : ; wherein: represents the weight hyperparameter of the imbalance factor deviation; is the weight hyperparameter of the inter-frame correlation stability constraint; is the supervised learning loss: ; wherein, represents the true label of the c-th class; represents the probability of belonging to the c-th class predicted by the network, which is output from the Softmax; represents the natural pair applied to the predicted probability; C represents the total number of classes; During the training process, the back propagation algorithm minimizes at the same time, adjusts the weights according to the labels, and adjusts according to the physical constraints and , and finally realizes the anomaly monitoring.

[0041] In order to further implement the above technical solutions, it also includes: S5. Based on the detection result, an alarm signal is generated, and a decision suggestion is generated, combined with a real-time feedback mechanism to realize closed-loop risk management.

[0042] An abnormal state monitoring system of a wind turbine generator set, comprising: A signal acquisition module for acquiring real-time sound original signals of the wind turbine generator set in operation ; A noise reduction module for using a double-time-scale noise estimation method to calculate short-time noise amplitude and long-time noise amplitude by analyzing the amplitude spectrum of the corresponding first frame and the amplitude spectrum of the corresponding first frequency point along the time axis, respectively, and estimate the estimated noise amplitude based on and , and then reconstruct the sound original signal according to to obtain the noise-reduced acoustic signal , , wherein is the signal length; A feature extraction module for extracting MFCC statistical features, MFCC inter-frame correlation features and aerodynamic-harmonic imbalance factors from the noise-reduced acoustic signal of the wind turbine generator set to obtain a hybrid input feature vector ; A detection module for inputting the hybrid input feature vector into a trained physical neural network PINN based on physical information for abnormal state detection, wherein the physical neural network PINN based on physical information is trained by the noise-reduced acoustic signal.

[0043] It should be noted that: The signal acquisition module includes a signal acquisition device, a signal acquisition card, and a sensor deployment area near the bottom of the wind turbine tower (including three points 1, 2 and 3 to cover multi-angle sound signals) to install the signal acquisition device. The original signal is amplified and filtered by a conditioning circuit to adapt to a wide frequency noise environment and improve signal quality. Then, the signal acquisition card is used to digitize and collect the filtered sound signal, and the ring network optical fiber is used to transmit the signal to the upper computer in the control room of the power plant booster station.

[0044] ​The audio sensor and the synchronous data acquisition card constitute a signal acquisition device, which is installed on the wind turbine and flexibly sets the number and position of the sensor according to the on-site collection requirement. The upper computer as a data storage and processing unit is installed in the control room of the booster station, and mainly completes the functions of subsequent signal filtering, feature calculation and fault online detection. The signal acquisition module mainly ensures the accuracy, real-time and robustness of signal acquisition, supports multi-channel fusion to optimize the signal-to-noise ratio, and transmits data to the upper computer in the control room of the booster station through the optical fiber transmission and network switching mechanism, realizes the edge-cloud collaborative processing.

[0045] In order to further implement the above technical solutions, the noise reduction module analyzes the original sound signal The corresponding amplitude spectrum Along the time axis, the short-time noise amplitude And the long-time noise amplitude are calculated respectively, and the specific content of the estimated noise amplitude And is estimated based on The specific content includes: Adopting double-time-scale noise estimation, the short-time noise amplitude And the long-time noise amplitude are calculated respectively by analyzing the statistical distribution of the amplitude spectrum Along the time axis: ; ; In the formula, And respectively represent the first Quantile and the first Quantile; And are the quantile thresholds of short-time and long-time noise estimation; Reflects the local background level of noise, Captures the overall trend and possible sudden changes of noise; The dynamic weight Is introduced to balance the contribution of the short-time and long-time estimations: ; ; In the formula, Is the initial weight, Is the time frame index, , Is the frequency point index, , , Is the number of points of time-frequency transformation, Is the imaginary unit, and for the upper and lower bounds of the range, satisfying 0≤ < ≤1; then the basic noise amplitude is: ; Based on the basic noise amplitude , the signal noise amplitude ratio is calculated, which is used to quantify the relative strength of signal and noise: ; Based on , the minimum amplitude threshold is calculated: ; wherein, is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the key features of the signal are not weakened; A dynamic noise scaling factor is introduced to dynamically adjust the noise amplitude: ; wherein, is a scaling factor that controls the sensitivity of the adjustment factor, is the initial weight; Based on the dynamic noise scaling factor , the estimated noise amplitude is calculated: .

[0046] A computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements a wind turbine abnormal state monitoring method as described above.

[0047] An electronic device, comprising: a processor and a memory for storing one or more programs; when the one or more programs are executed by the processor, a wind turbine abnormal state monitoring method as described above is implemented.

[0048] The application will be further described below through specific experiments: As Figure 2 shown, the multi-channel distributed high-resolution acoustic signal acquisition and analysis scheme of the wind turbine mainly includes audio sensors, signal acquisition cards and host computers. Analog signals are collected by sensors, and then digital signals are output through conditioning circuits and acquisition cards, and transmitted to the host computer in the booster station for processing and analysis. In this example, the main parameter configuration is: window length Set to 256 samples, overlap length set to 128 samples, FFT point number set to 512, short-time noise amplitude and long-time noise amplitude Percentile is 25% and 75% respectively, Set to 0.5 as the lower limit of weight calculation, Set to 0.9 as the upper limit of weight calculation, Set to 0.005 as the base value of the minimum amplitude threshold, Set to 1.5 as the dynamic noise scaling factor parameter.

[0049] Figure 3 The frequency spectrum analysis of the collected abnormal sound of the wind turbine is shown. The horizontal axis represents time (unit: seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (unit: Hz), ranging from 0 to 16000 Hz. The amplitude (unit: dB) gradually changes from a higher amplitude (about -10 dB) to a lower amplitude (about -40 dB) to show the intensity distribution of the sound signal. This abnormal sound is caused by abnormal blade bearing of the unit, resulting in abnormal frequency band and sound characteristics, among which the abnormal frequency band is marked in the figure, which is in sharp contrast with the background noise.

[0050] The sampling rate of the dynamic signal is 32000 Hz, and the total duration is about 10 seconds (320000 sampling points in total). After normalization and removal of DC offset, a time-frequency representation matrix is generated, and the amplitude spectrum is further calculated, resulting in an amplitude spectrum size of (257, 2501), corresponding to 257 frequency components and 2501 time frames, reflecting the time-frequency characteristics of about 10 seconds of signal.

[0051]

[0052] By analyzing the 10-second amplitude spectrum, the short-time and long-time noise amplitudes are calculated using a multi-scale method, and the short-time noise and long-time noise matrices are generated based on the 25% and 75% percentiles, respectively, with a matrix size of (257, 1). Then, based on the formula , the is 0.1620, and its output is shown in Tables 2 and 3.

[0053]

[0054]

[0055] Using the basic noise matrix and the amplitude spectrum to calculate SNMR, calculate the minimum amplitude threshold, calculate the dynamic noise scaling factor, and calculate the final noise amplitude, the output is shown in Tables 4 and 5.

[0056]

[0057]

[0058] With the final noise amplitudes, the final amplitude spectrum matrix is calculated, the complex spectrum matrix is reconstructed, and the denoised signal is reconstructed, whose output is shown in Table 6.

[0059]

[0060] Figure 4 The sound spectrum analysis after applying the denoising method is shown. The horizontal axis represents time (unit: seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (unit: Hz), ranging from 0 to 16000 Hz. The amplitude range (unit: dB) is consistent with Figure 3 . As can be seen by comparison with Figure 3 , the method of the present application significantly suppresses the background noise band (manifested as an increase in the large area of lower amplitude region in the figure), while effectively preserving the abnormal characteristic frequency band caused by the blade bearing (manifested as a more prominent higher amplitude region). This indicates that the denoising method maintains the relative intensity of the key abnormal characteristics while removing the interfering noise.

[0061] As shown in Figure 5 , MFCC feature extraction is performed on the denoised signal. These time series plots show the changes in 12 MFCC features.

[0062] Next, the MFCC inter-frame correlation matrix R is calculated, and the lower triangular part is extracted to obtain the correlation feature vector. Among them, part of the feature vector is: R=[-0.2470,0.0719,0.0527,0.0529,-0.0085,-0.1529,0.2629,-0.0475,-0.0043,0.5560].

[0063] This vector captures the inter-frame dynamic correlation pattern, and the correlation decreases (e.g., the negative value increases) under abnormal conditions, reflecting the increase in signal non-stationarity.

[0064] For physical information features, the vector includes the BPF and its harmonic amplitudes (H=3): =[5.2,3.1,1.8] (unit: dB), extracted from the order spectrum using Gaussian weighting , These amplitudes are low (<3 dB) under healthy conditions, but are elevated in this abnormal instance, indicating the aerodynamic imbalance caused by the blade bearing failure. The energy of the aerodynamic noise band = 12.5 (unit: dB·Hz), broadband noise (band 1000-5000 Hz) is predicted based on the Ffowcs Williams-Hawkings model.

[0065] Innovative fusion of aerodynamic-harmonic imbalance factor, amplification of coupling effect: under abnormal blade bearing, BPF harmonic and trailing edge noise are enhanced at the same time, leading to a significant deviation of the imbalance factor from the health benchmark (<20). Finally, a mixed input feature vector is obtained , which is used as input for the PINN model.

[0066] To realize the abnormal state detection of the acoustic signal of the wind turbine generator, the method constructs a physics-informed neural network (PINN) based on physical information, taking the mixed input feature vector (containing MFCC statistical features, inter-frame correlation features, and aerodynamic-harmonic imbalance factors) as input, which is suitable for pattern recognition of non-stationary time series signals. Among them, the weight of the MFCC statistical feature fusion weight is , the weight of the inter-frame correlation weight is , and the weight of the imbalance factor weight is .

[0067] The model enhances generalization and physical consistency by incorporating physical constraints (such as residual minimization of the aeroacoustic model), and efficiently captures dynamic context information. Its architecture is as follows: Input layer: receives the mixed feature sequence , with dimensions (batch size, sequence length, feature dimension), and transposes it to (batch size, feature dimension, sequence length) to adapt to one-dimensional convolution operations.

[0068] Time series feature extraction module: PINN-TDNN layer 1: one-dimensional convolution (kernel size 5, padding 2), input channel number is feature dimension (e.g. 24, based on MFCC statistics, inter-frame correlation, and aerodynamic-harmonic imbalance factor), output channel number 128; followed by batch normalization (BatchNorm1d), ReLU activation, and Dropout (probability 0.5) to prevent overfitting.

[0069] PINN-TDNN layer 2: one-dimensional convolution (kernel size 5, padding 2), input channel number 128, output channel number 128; followed by batch normalization and ReLU activation.

[0070] Physical embedding layer: innovatively introduces physical constraints, calculates physical loss through aerodynamic-harmonic imbalance factor health benchmark.

[0071] Adaptive pooling layer: AdaptiveAvgPool1d is used to compress the time dimension to a fixed length of 1, generating a fixed-size feature vector.

[0072] Classifier module: Fully connected layer: input 128 dimensions, output 2 classes (normal / abnormal), implicitly process classification probability through Softmax.

[0073] The PINN model is implemented under the Python framework (PyTorch) and runs on the CPU, which is suitable for real-time anomaly monitoring. In this example, to verify the effectiveness of the model, 2100 wind turbine unit acoustic data samples (1590 normal samples and 510 abnormal samples) are used as input, based on MFCC features and physical information features (such as aerodynamic-harmonic imbalance factors). The dataset is divided into a training set (1680 samples, 1272 normal and 408 abnormal), a validation set (210 samples, 159 normal and 51 abnormal), and a test set (210 samples, 159 normal and 51 abnormal) in the ratio of 8:1:1 to ensure the model's generalization ability. The optimizer uses Adam with an initial learning rate of 0.001, weight decay of 1e-4, and learning rate step scheduling (decrease to 0.9 times every 5 cycles).

[0074] As shown in Figure 6 , the model's accuracy on the test set is 98.10%, showing high efficiency in distinguishing between normal and abnormal states. The low false positives (2 cases) and false negatives (2 cases) further verify the model's robustness. Batch normalization and Dropout effectively prevent overfitting, learning rate scheduling optimizes the convergence speed, and physical constraints improve the generalization performance across operating conditions (>98%).

[0075] The above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing examples, or make equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A wind turbine abnormal condition monitoring method, characterized by, Comprising the following steps: S1. Collecting the sound original signal of the wind turbine operation in real time ; S2. Adopting double time-scale noise estimation method, by analyzing sound original signal Corresponding to the frame, the amplitude spectrum of frequency point Along the time axis statistical distribution, respectively calculate short-time noise amplitude And long-time noise amplitude Based on And Estimate noise amplitude , and then according to Reconstruct the sound original signal, get the noise reduction acoustic signal , , wherein Signal length; S3. Extract the MFCC statistical features, MFCC inter-frame correlation features and the aerodynamic-harmonic imbalance factor from the acoustic signal after noise reduction of the wind turbine generator set to obtain a mixed input feature vector ; S4. combine the input feature vectors inputting the trained physics-informed neural network (PINN) for anomaly condition detection, wherein the physics-informed neural network (PINN) is trained based on the denoised acoustic signals.

2. A wind turbine abnormal condition monitoring method according to claim 1, characterised in that, S2. The acquired sound raw signal is pre-processed, including the following steps: obtained by time-frequency transform corresponding time-frequency representation matrix wherein the method by time-frequency transform specifically comprises: short-time Fourier transform (STFT), fast Fourier transform (FFT), wavelet transform or wavelet packet transform; On the basis of the time-frequency representation matrix the amplitude spectrum is calculated: 。 3. A wind turbine abnormal condition monitoring method according to claim 1, characterised in that, S2 in which the sound original signal is analyzed the corresponding amplitude spectrum the statistical distribution along the time axis, respectively calculating the short-time noise amplitude and the long-time noise amplitude The specific content includes: Using a double-time-scale noise estimation, the short-time noise amplitude and the long-time noise amplitude are calculated by analyzing the amplitude spectrum of the signal along the time axis ； ； wherein and denote the first decile and the first decile, respectively; and are the decile thresholds for the short-time and long-time noise estimates, respectively; reflects the local background level of the noise, captures the overall trend and possible abrupt changes in the noise.

4. A wind turbine abnormal condition monitoring method according to claim 1, characterised in that, S2 based on and to estimate the estimated noise amplitude include: Introducing dynamic weights to balance the contribution of both short and long estimates: ； ； wherein is an initial weight, is a time frame index, , is a frequency point index, , , is a number of points of the time-frequency transform, is an imaginary unit, and is upper and lower bounds of a range, satisfying 0≤ <1 ; the base noise amplitude is: ； Based on the base noise amplitude Computing the signal noise amplitude ratio for quantifying the relative strength of signal and noise: ； Based on Computing a minimum amplitude threshold : ； wherein is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the signal's key features are not attenuated; Introducing dynamic noise scaling factor to dynamically adjust the noise amplitude: ； wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; Dynamic noise scaling factor Computing an estimated noise amplitude : 。 5. A wind turbine abnormal condition monitoring method according to claim 4, characterised in that, S2 in accordance with reconstructing the sound original signal to obtain a noise-reduced acoustic signal The specific content includes: In accordance with and computing the noise reduction spectrum magnitude : ； Preserving original phase, restoring complex spectrum after noise reduction : ； reconstructing the noise-reduced spectrum into a time-domain signal using an inverse transform : ； In the formula, is a superimposed window function.

6. A wind turbine abnormal condition monitoring method according to claim 1, characterised in that, S3. The specific content includes: the noise-reduced acoustic signal performing a mel-frequency transform to obtain the acoustic signal in mel-frequency : ； filtering by a mel filter bank filtering: ； extracting mfcc features : ； In the formula, This indicates the index of the filter in the Mel filter bank. The number of filter banks. Indicates the first The center frequency of the Mel filter; extracting, within a time window, frame MFCC feature vectors, constituting a matrix: ； wherein denotes the MFCC feature vector of a frame is a matrix of dimension ; each frame extracts MFCC features, the MFCC inter-frame correlation matrix is defined as the matrix of Pearson correlation coefficients between the MFCC vectors of the frames: ； wherein, is a P x P symmetric matrix, representing the MFCC inter-frame correlation matrix, used to quantify the linear similarity between different frames of MFCC vectors, providing temporal contextual information, is based on the matrix is calculated, each term reflects the similarity between the frame and the frame , denotes an element in the matrix, i.e. the Pearson correlation coefficient between the MFCC feature vector of the frame and the frame , ranging from [-1, 1]; the subscripts and denote the row index and the column index, respectively, from 1 to P, where each element denotes the correlation coefficient between the MFCC feature vector of the frame and the frame , ； in, The numerator represents the covariance of two MFCC vectors and is used to measure common variance. These are the indices of the MFCC vector elements, from 1 to L; Indicates the first The first frame of the MFCC vector Each element. Indicates the first The mean of the frame MFCC vector is used to center the data; Indicates the first The mean of the frame MFCC vector; Since the correlation matrix is symmetric, i.e. , and the diagonal elements are all 1, i.e. , only the elements of its lower triangular part are extracted to construct the feature vector to avoid redundancy: ； The dynamic correlation mode of the sound signal in multiple frame time scales is captured, and the timing context information beyond the single frame feature is provided for the model. Further, the Gaussian weighted order spectrum average method is adopted to extract the amplitude of the wind turbine order harmonic: ； Wherein the Gaussian weight function is defined as: ； wherein, represents the amplitude of the th harmonic, used to quantify the strength of the BPF and its nearby harmonics, reflecting the periodic vibration induced by blade rotation and aerodynamic unbalance; represents the harmonic order, indexing the harmonic series, where is the maximum order; represents the Gaussian weight, a real number ranging in [0, 1], used to emphasize the central harmonics and decay the edge offsets; : represents the order spectrum function, extracting the periodic components of the rotating machinery; represents the blade; represents the offset index, ranging from to ; to define the neighborhood offsets within the weighted window, used for smoothing calculation; represents the fundamental frequency of the rotating machinery; represents the standard deviation controlling the width of the Gaussian function; Based on the model of Ffowcs Williams-Hawkings equation, the broadband noise energy generated by the interaction of airflow and blades of the wind turbine is calculated: ； wherein represents the air density to take into account the effect of the medium properties on the noise propagation; represents the tip speed, represents the sensor to tip distance; The aerodynamic-harmonic unbalance factor is calculated as a normalized deviation: ； By weighting and fusing the features, a hybrid feature vector is formed : ； wherein, represents a pneumatic-harmonic imbalance factor to quantify the deviation of harmonic strength from the prediction noise for indicating anomalies; represents a weight for the MFCC statistical weight to balance the contribution; represents a weight for the inter-frame correlation weight to balance the temporal features; represents a weight for the imbalance factor weight; represents an averaging factor to normalize the harmonic sum to ensure scale invariance.

7. A wind turbine abnormal condition monitoring method according to claim 6, wherein, S4. The specific content includes: The basic network structure of the physical neural network (PINN) based on physical information adopts a feedforward neural network as a backbone network for learning a complex mapping from features to health conditions; The physical neural network (PINN) based on physical information introduces physical constraints, and is realized through a physical loss function wherein the loss function is calculated by a feature or a derived quantity thereof, which measures the degree of violation of the physical principle by the network output or intermediate state. Wherein the physical constraints include: Pneumatic-harmonic conservation constraint : ； wherein, represents the aerodynamic-harmonic conservation loss, quantifying the deviation of the imbalance factor; represents the health benchmark value; MFCC inter-frame dependency stability constraint : ； wherein, represents a correlation feature vector of a first time window; represents a correlation feature vector of a previous time window total loss function ； where: is a weight hyperparameter representing the imbalance factor bias; is a weight hyperparameter of the inter-frame correlation stability constraint; is a supervised learning loss: ； wherein, represents the true label of the cth class; represents the network-predicted probability of belonging to class c, from the Softmax output; represents the natural pair-wise application to the predicted probabilities; C represents the total number of classes. During the training process, the back propagation algorithm adjusts the weights according to the labels while minimizing and adjusts according to the physical constraints and to ultimately achieve anomaly monitoring.

8. A wind turbine abnormal condition monitoring method according to claim 1, characterized in that, Also including: S5. Based on the detection result, an alarm signal is generated, and a decision suggestion is generated, combined with a real-time feedback mechanism to realize closed-loop risk management.

9. A wind turbine generator system abnormal state monitoring system characterized by, Comprising: The signal acquisition module is used for collecting sound original signals of the wind turbine generator set in real time ; The noise reduction module is used to perform noise estimation using a dual-timescale method by analyzing the original sound signal. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original sound signal and obtain the noise-reduced acoustic signal. , ,in The signal length; The feature extraction module is configured to extract MFCC statistical features, MFCC inter-frame correlation features and a harmonic-to-aerodynamic imbalance factor from the acoustic signal after noise reduction of the wind turbine generator set to obtain a hybrid input feature vector. ; The detection module is used to process the mixed input feature vectors. Anomaly detection is performed by inputting a pre-trained physical information-based neural network PINN, where the physical information-based neural network PINN is trained using denoised acoustic signals.

10. A wind turbine generator system abnormal state monitoring system according to claim 9, wherein In the noise reduction module, the original sound signal is analyzed The corresponding amplitude spectrum The statistical distribution along the time axis is calculated respectively Short-time noise amplitude and long-time noise amplitude Based on And Estimate the specific content of the estimated noise amplitude ​ A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ； ； wherein and denote the first decile and the first decile, respectively; and are decile thresholds for short-time and long-time noise estimates, respectively; reflects the local background level of noise, captures the overall trend and possible abrupt changes in noise.

11. A wind turbine generator system abnormal state monitoring system according to claim 9, wherein In the noise reduction module based on and To estimate the noise amplitude The specific content includes: Introducing dynamic weights to balance the contribution of both short and long estimates: ； ； wherein is an initial weight, is a time frame index, , is a frequency point index, , , is a number of points of the time-frequency transform, is an imaginary unit, and is upper and lower bounds of a range, satisfying 0≤ <1 ≤1; the base noise amplitude is: ； Based on the base noise amplitude Computing the signal noise amplitude ratio For quantifying the relative strength of signal and noise: ； Based on Computing a minimum amplitude threshold : ； wherein is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the signal's key features are not attenuated; Introducing dynamic noise scaling factor to dynamically adjust the noise amplitude: ； wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; Dynamic noise scaling factor Computing an estimated noise amplitude : 。 12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the wind turbine abnormal state monitoring method of any one of claims 1-8.

13. An electronic device comprising: A processor and a memory, the memory is used to store one or more programs; characterized in that when the one or more programs are executed by the processor, a wind turbine abnormal state monitoring method as claimed in any one of claims 1-8 is realized.

Images