A wind turbine abnormal state detection method and system based on audio data

Abstract

The application discloses a kind of wind turbine abnormal state detection method and system based on audio data, this method includes the following steps: 1) obtain the audio data under different operating conditions of wind turbine;Different operating conditions include cooling fan does not work and fan does not work, cooling fan does not work and fan works, cooling fan works and fan does not work, cooling fan works and fan works;2) according to the sound spectrum graph rule of audio data, filter out the audio part of specific frequency;3) extract the time domain feature and frequency domain feature of filtered audio data, and screening is carried out;4) the time domain feature, frequency domain feature and corresponding operating state after screening are used to train multi-classification model, and the trained multi-classification model is obtained;5) based on the trained multi-classification model, input audio data, obtain output result to realize wind turbine abnormal state detection.The application has the advantages of reducing the operation risk of wind turbine, improving the degree of operation automation and the like.

Description

TECHNICAL FIELD

[0001] The present application mainly relates to the field of wind power technology, and in particular to a wind turbine abnormal state detection method and system based on audio data. BACKGROUND

[0002] The wind turbine is long-term operated in the environment of alternating load, and the blade is prone to crack, corrosion and other faults, which not only reduces the service life of the blade, but also affects the wind energy capture of the wind turbine. Therefore, early diagnosis and early warning of the blade fault have great significance for the safe and healthy operation of the wind turbine. At present, the methods for monitoring the faults of the wind turbine blade include image recognition, thermal imaging, laser detection, vibration signal feature recognition and the like.

[0003] In the current technology for detecting the operation state of the wind turbine based on audio, the characteristic frequency is mainly intercepted, and the characteristic frequency is compared and processed to diagnose the fault. This method is similar to the fault diagnosis based on the vibration signal, but the audio has different propagation modes and is prone to noise interference, which determines the difference in the analysis mode. If the audio signal is simply analyzed in the frequency domain, the detection accuracy of the traditional vibration signal using the same method cannot be achieved, and the periodic characteristics accompanying the wind sweeping of the blade are lost. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a wind turbine abnormal state detection method and system based on audio data, which reduces the operation risk of the wind turbine and improves the automation degree of operation and maintenance.

[0005] To solve the above technical problems, the technical scheme provided by the present application is as follows:

[0006] A wind turbine abnormal state detection method based on audio data, comprising the steps of:

[0007] 1) obtaining audio data of the wind turbine under different operation states; wherein the different operation states include that the cooling fan is not working and the wind turbine is not working, the cooling fan is not working and the wind turbine is working, the cooling fan is working and the wind turbine is not working, and the cooling fan is working and the wind turbine is working;

[0008] 2) filtering the audio data: filtering out the audio part of a specific frequency according to the sound spectrum graph rule of the audio data;

[0009] 3) extracting the time domain features and frequency domain features of the filtered audio data, and screening the time domain features and frequency domain features;

[0010] 4) training a multi-classification model through the screened time domain features, frequency domain features and corresponding operation states to obtain a trained multi-classification model;

[0011] 5) based on the trained multi-classification model, input the audio data, and obtain the output result to realize the abnormal state detection of the wind turbine.

[0012] As a further improvement of the above technical solution:

[0013] In step 3), the time domain features are counted by the waveform data in each data file, specifically including maximum value, minimum value, range, mean value, root mean square difference, variance, skewness, kurtosis, waveform index, pulse index, kurtosis index and margin index; the frequency domain features include short-time zero-crossing rate, spectral centroid, spectral attenuation, frequency domain root mean square, mel frequency cepstral coefficient, chroma frequency, P-order spectral bandwidth, spectral flatness, constant Q spectrum, spectral contrast, spectral graph column coefficient, timbre centroid feature and mel scaling spectrum.

[0014] In step 3), by constructing a multi-classification model, the feature weights of the multi-classification model are extracted by using the eli5 tool library, and the features are selected according to the weights not being 0.

[0015] In step 4), the multi-classification model includes unsupervised learning model and supervised learning model, wherein the construction of the unsupervised learning model is mainly based on the construction of the clustering model, and the elbow rule inflection point is used to objectively select the number of categories of the result.

[0016] The data division in the supervised learning model adopts a random mixed division manner, and a multi-classification logistic regression model is adopted.

[0017] In step 2), the audio data is filtered by Fourier transform; wherein the specific frequency is obtained from the sonogram of the audio data.

[0018] In step 5), the sonogram is generated based on the multi-classification model as a visual criterion for judging whether the howling occurs or whether the drain hole is blocked.

[0019] The application also discloses a wind turbine abnormal state detection system based on audio data, comprising:

[0020] The acquisition unit is used for acquiring audio data under different operating states of the wind turbine; wherein the different operating states include that the cooling fan is not working and the fan is not working, the cooling fan is not working and the fan is working, the cooling fan is working and the fan is not working, and the cooling fan is working and the fan is working.

[0021] The filtering unit is used for filtering the audio data: according to the sonogram rule of the audio data, filtering out the audio part of the specific frequency;

[0022] The feature extraction unit is used for extracting the time domain features and the frequency domain features of the filtered audio data, and screening the time domain features and the frequency domain features.

[0023] a model construction unit configured to train a multi-classification model by screening the time-domain features, the frequency-domain features, and the corresponding operating states, and obtain a trained multi-classification model;

[0024] an anomaly detection unit configured to input the audio data based on the trained multi-classification model, and obtain an output result to realize wind turbine abnormal state detection.

[0025] The application further discloses a wind turbine abnormal state detection device based on audio data, which comprises:

[0026] a WTOD monitoring and triggering collection module configured to monitor the operating state of the wind turbine by accessing the PLC operating data of the wind turbine, and trigger sound collection and collect the sound of the wind turbine and the PLC operating data of the key measurement points of the wind turbine at the same period for a fixed time length when the sound collection condition of the wind turbine is met;

[0027] an audio detection module configured to collect the sound data of the key measurement points of the wind turbine;

[0028] a state analysis and anomaly diagnosis module configured to read the collected sound data of the wind turbine, combine the sound data with the operating data of the wind turbine, perform feature analysis, state recognition and anomaly pattern diagnosis, and output a diagnosis result.

[0029] As a further improvement of the above technical solution:

[0030] The audio detection module comprises one or more detection modules, each of which comprises a microphone, a collector and an intelligent terminal, and the intelligent terminal is configured with a wireless network module to directly connect to a cloud server; the microphone of one of the detection modules is installed near the sound source of the key monitoring position to pick up the audio data of the wind turbine, and the collector is connected to the microphone and the intelligent terminal to convert the audio data into digital signals and send them to the intelligent terminal.

[0031] The microphone of one of the detection modules is installed on the outer wall of the tower drum at the bottom of the tower to collect the sound data of the blades, and the microphone of another detection module is installed in the nacelle to collect the sound data of the gearbox, the generator or the yaw.

[0032] The device further comprises a human-computer interaction module configured to provide query of the diagnosis result and the operating data, playback of the recording, visualization function of the sound spectrogram of the sound data, and basic configuration function of the sound collection and anomaly diagnosis program; and the human-computer interaction module is integrated on the intelligent terminal of the wind turbine or the server of the wind farm.

[0033] The data synchronization module is further configured to synchronize the sound data, WTOD data and diagnosis result data collected and stored in the local terminal to the wind farm server and the cloud server.

[0034] The application further discloses a computer readable storage medium, which stores a computer program, and the computer program performs the steps of the wind turbine abnormal state detection method based on audio data when being run by a processor.

[0035] The application further discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program performs the steps of the wind turbine abnormal state detection method based on audio data when being run by the processor.

[0036] Compared with the prior art, the application has the following advantages:

[0037] The application can find abnormal states such as blade drain hole blockage by acquiring audio data on the wind turbine, filtering the audio data, extracting features, and classifying and identifying, thereby reducing the operation risk of the wind turbine and improving the automation degree of operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 It is a configuration structure diagram of a single unit of the application.

[0039] Figure 2 It is a configuration structure diagram of the synchronous monitoring of a single unit cabin and a tower bottom of the application.

[0040] Figure 3 It is a structure diagram of the batch deployment of a wind farm of the application.

[0041] Figure 4 It is a structure diagram of the device of the application in the embodiment.

[0042] Figure 5 It is a waveform diagram and a Fourier transform spectrum diagram of the "cooling fan not working" in the application.

[0043] Figure 6 It is a waveform diagram and a Fourier transform spectrum diagram of the "cooling fan working" in the application.

[0044] Figure 7 It is a Fourier transform spectrum diagram of the audio data after filtering in the application.

[0045] Figure 8 It is a waveform diagram of the audio data after filtering in the application.

[0046] Figure 9Waveform diagram of audio data after filtering in the present invention.

[0047] Figure 10 Waveform diagram of wavelet transform in the present invention.

[0048] Figure 11 Waveform diagram of digital filter in the present invention.

[0049] Figure 12 Waveform diagram of local polynomial least squares filter in the present invention.

[0050] Figure 13 Waveform diagram of Chebyshev II filter in the present invention.

[0051] Figure 14 Waveform diagram of cooling fan not working in the present invention.

[0052] Figure 15 Waveform diagram of cooling fan working in the present invention.

[0053] Figure 16 Waveform diagram of cooling fan not working in the present invention.

[0054] Figure 17 Waveform diagram of cooling fan not working in the present invention.

[0055] Figure 18 DBSCAN clustering result diagram in the present invention.

[0056] Figure 19 Training result diagram in the present invention.

[0057] Figure 20 Verification result diagram in the present invention.

[0058] Figure 21 ROC curve diagram in the present invention

[0059] Figure 22 Waveform diagram of blade abnormality in the present invention.

[0060] Figure 23 Waveform diagram of blade abnormality in the present invention.

[0061] Figure 24 Short-time energy numerical diagram in the present invention.

[0062] Figure 25 Waveform diagram of different rotating speeds of whistling in the present invention.

[0063] Figure 26 Flowchart of the method of the present invention in the embodiment. DETAILED DESCRIPTION

[0064] The application will be further described in conjunction with the accompanying drawings and specific embodiments of the application.

[0065] As shown in the figure, the wind turbine abnormal state detection method based on audio data of the embodiment comprises the steps of: Figure 1

[0066] 1) acquiring audio data of the wind turbine under different operating states; wherein the different operating states include that the cooling fan is not working and the fan is not working, the cooling fan is not working and the fan is working, the cooling fan is working and the fan is not working, and the cooling fan is working and the fan is working;

[0067] 2) filtering the audio data: filtering out the audio part of a specific frequency according to the sound spectrum graph rule of the audio data;

[0068] 3) extracting the time domain features and frequency domain features of the filtered audio data, and screening the time domain features and frequency domain features;

[0069] 4) training a multi-classification model through the screened time domain features, frequency domain features and corresponding operating states to obtain a trained multi-classification model;

[0070] 5) inputting the audio data based on the trained multi-classification model to obtain an output result to realize wind turbine abnormal state detection.

[0071] The application acquires audio data on the wind turbine, filters the audio data, extracts features, and classifies and identifies to form a set of feasible abnormal state detection methods based on wind turbine audio data, which can find abnormal states such as blade drain hole blockage, reduce the operation risk of the wind turbine, and improve the automation degree of operation and maintenance.

[0072] In a specific embodiment, in step 3), the time domain features are counted by waveform data in each data file, specifically including maximum value, minimum value, range, mean value, root mean square difference, variance, skewness, kurtosis, waveform index, pulse index, kurtosis index and margin index; the frequency domain features include short-time zero-crossing rate, spectral centroid, spectral attenuation, frequency domain root mean square, mel frequency cepstral coefficient, chroma frequency, P-order spectral bandwidth, spectral flatness, constant Q spectrum, spectral contrast, spectrogram column coefficient, timbre centroid feature and mel scaling spectrum.

[0073] In a specific embodiment, in step 3), the feature weights of the multi-classification model are extracted by constructing a multi-classification model using the eli5 tool library, and the features are screened according to the weights not being 0.

[0074] ​In a specific embodiment, in step 4), the multi-classification model comprises an unsupervised learning model and a supervised learning model, wherein the unsupervised learning model is constructed mainly by constructing a clustering model, and a DBSCAN clustering algorithm is used to objectively select the number of categories of the result according to the inflection point of the elbow rule.

[0075] In a specific embodiment, the data division in the supervised learning model adopts a random mixed division manner, and a multi-classification logistic regression model is used.

[0076] In a specific embodiment, in step 2), the audio data is filtered by Fourier transform; and the specific frequency is obtained from the spectrogram of the audio data.

[0077] In a specific embodiment, in step 5), the spectrogram is generated based on the multi-classification model as a visual criterion for judging whether the howling occurs or whether the drain hole is blocked.

[0078] The above application will be further described below in combination with a complete specific embodiment:

[0079] 1, sample collection

[0080] In order to meet the requirements of the machine learning algorithm on the number of samples, and in order to distinguish the running state of the fan under different backgrounds (working conditions), a period of time sample is collected, and the sample is selected and labeled artificially by human ear.

[0081] The collected audio data has a time span of about half a year, and is sampled every 1 hour, and each sampling is 30 seconds. For these audios, it is found by human ear that the working or not of the cooling fan has a great influence on the identification of the working condition of the fan blade, and therefore the audio is divided into four types of "cooling fan not working, fan not working", "cooling fan not working, fan working", "cooling fan working, fan not working", "cooling fan working, fan working", and classified and identified by a multi-classification machine learning model.

[0082] 2, data exploration

[0083] The collected audio data is visually explored in time domain and frequency domain.

[0084] 2.1, filtering

[0085] 2.1.1, Fourier transform

[0086] According to the rule discovered from the spectrogram, the audio under the working state of the fan is analyzed by filtering:

[0087] The cooling fan is not working, and there is no strong interference in the background sound of the fan, and the scanning tower sound is mixed with environmental noise. Figure 5This provides the waveform and Fourier transform spectrum of a given audio audio file. The spectrum shows the result when the horizontal axis is mapped to a range of 20,000, with the number of intermediate points equal to half the length of the audio sample points. From... Figure 5 As can be seen from the waveform diagram, the environmental noise has little impact on the tower sweeping sound, and the periodicity of the tower sweeping is very obvious. Similarly, the high-frequency part of the spectrum shows the frequency regularity, while the low-frequency part is mixed with the environmental noise frequency, but overall, it has no significant impact on the tower sweeping frequency.

[0088] Therefore, there is no need to filter when the cooling fan is not working and the blower is working.

[0089] The operation of the cooling fan caused strong interference with the noise from the fan sweeping the tower, therefore Figure 6 The waveform diagram shows a chaotic and jagged state, with the sound of the tower sweeping being drowned out.

[0090] Meanwhile, the amplitude information of the high-frequency part of the Fourier transform spectrum is overwhelmed by the amplitude information of the low-frequency part, making the characteristics of the high-frequency part indistinct. Since fan noise is mainly concentrated in the low-frequency part, it is necessary to perform high-pass filtering on the low-frequency part to extract the characteristics of the high-frequency part. Specifically:

[0091] 1) Filtering. After some experimentation, frequencies below 1kHz were filtered out, but this only reduced the perceived frequency of the "cooling fan working" sound in terms of timbre and decibels. Figure 7 The spectrum diagram.

[0092] from Figure 7 As can be seen, after filtering out the low-frequency components (note that not all background noise was filtered out), the high-frequency components exhibit a clear periodicity. This indicates that the filtering effect is significant.

[0093] 2) The waveform and spectrogram after filtering. For example... Figure 8 As shown, although the waveform has many spikes, the periodicity of the tower sweep can be clearly shown, and a similar pattern can be seen from the spectrogram.

[0094] 2.1.2 Wavelet Transform

[0095] Wavelet transform was performed on the audio of "cooling fan working, fan working". Four methods were used for filtering: wavelet hard thresholding, wavelet soft thresholding, wavelet compromise thresholding, and wavelet weighted average filtering. The waveform after filtering is shown in Figure 10.

[0096] pass Figure 10 The waveforms after the four wavelet transforms show that the filtering effect is poor, with many spikes, and the periodic characteristics of the sweeping tower cannot be clearly reflected. Further optimization can be achieved by adjusting the different thresholds of the four wavelet transform methods.

[0097] 2.1.3, Digital filtering

[0098] Using a digital filter, the "cooling fan working, fan working" audio is band-pass filtered (selecting the part between 1000-4096Hz), using IIR or FIR filter to filter data along one dimension, the original audio and the sound spectrogram after digital filtering are shown in Figure 11 .

[0099] As can be seen from Figure 11 , the audio frequency generated by "fan working" is indeed mainly concentrated between 1kHz and 4kHz, and the sound spectrogram is more obvious between them.

[0100] 2.1.4, Forward-backward filtering:

[0101] Using a forward-backward filter, the "cooling fan working, fan working" audio is band-pass filtered (selecting the part between 1000-4096Hz), using IIR or FIR filter to filter data along one dimension, the original audio and the sound spectrogram after forward-backward filtering are similar to Figure 11 .

[0102] 2.1.5, Local polynomial least squares filtering

[0103] Using a filter based on local polynomial least squares fitting in time domain, combined with the weighted average method of moving window, the weighted coefficients of the given high-order polynomial are fitted in the sliding window, and the filtering is performed on the basis of ensuring the shape and width of the signal, the filtering result is shown in Figure 12 .

[0104] As can be seen from Figure 12 , the filtering result is not as good as the previous filter, and the possible reason is that the window size and the order of the fitted polynomial need to be adjusted constantly.

[0105] 2.1.6, Elliptic filtering

[0106] Design an N-order digital or analog elliptic filter, use a cascaded second-order part to filter along one-dimensional data, use the digital IIP filter defined by SOS to filter the audio data, and return the filter coefficients, the result is similar to Figure 11 .

[0107] 2.1.7, Chebyshev filtering

[0108] By designing an N-order digital or analog Chebyshev I or II type filter, and returning the filter coefficients, the sound spectrogram result is as follows Figure 13The Chebyshev I filter result is better than the Chebyshev II filter result.

[0109] From the filtering results of the above various filters, it can be seen that the Fourier transform, digital filtering, forward-backward filtering, elliptical filtering and Chebyshev I filtering have similar results. According to the "Occam razor" principle, the most commonly used and most easily understood Fourier transform is selected as the filtering tool when the cooling fan works.

[0110] 2.2, audio waveform diagram

[0111] In order to observe the shape of the data in each audio file, the time sequence waveform curve is drawn, such as Figure 14 and 15 .

[0112] Figure 14 The audio waveform diagram of the working and non-working fan under the condition that the cooling fan does not work is shown. It can be seen that under the condition that there is no strong external interference, whether the fan works or not, the waveform diagram can be well displayed.

[0113] Figure 15 The audio waveform diagram of the working and non-working fan under the condition that the cooling fan works is shown. It can be seen that the noise generated by the working cooling fan causes strong interference to the fan blade sweep tower sound, and there are many burrs in the waveform diagram. Generally, the blade sweep tower sound is submerged, but the periodicity can be seen.

[0114] 2.3, sound spectrum waveform diagram

[0115] From the waveform diagram, the characteristics of different working conditions can be clearly seen, and the basic frequency information of different working conditions can be seen from the sound spectrum diagram, such as Figure 16 and 17 .

[0116] Figure 16 The audio spectrum diagram of the working and non-working fan under the condition that the cooling fan does not work is shown. It can be seen that under the condition that there is no strong external interference, whether the fan works or not, the sound spectrum diagram displays the periodic characteristics, and the audio frequency when the fan works is concentrated between 1k-4kHz.

[0117] Figure 17The audio spectrogram of the working state of the cooling fan is shown. It can be seen that the noise frequency generated by the cooling fan is about 0-2 kHz, and the color of the spectrogram between 0-1 kHz is darker, and the display energy is larger, and the color of the spectrogram between 1k-2kHz is lighter, and the display energy is smaller, which provides an idea for subsequent filtering, that is, the audio part below 1kHz can be filtered out, so that the blade sweep tower sound is relatively clear. The sound frequency of the fan working state is mainly concentrated in 1k-4kHz, and the periodic characteristic is obvious. It can also be seen from the spectrogram that there is an energy enhancement feature (may be a howling).

[0118] 3、Feature engineering

[0119] The feature construction, feature evaluation, feature selection and other methods based on time domain and frequency domain are described, and the features are reselected according to the feature importance.

[0120] 3.1、Feature construction

[0121] In order to identify the different types of fan audio of running state, it is necessary to construct a classification model. From Figures 5 to 8 It can be seen that there are obvious characteristics between different types of audio. Therefore, from the two angles of time domain and frequency domain, the features are extracted to construct the classification model.

[0122] Among them, the time domain features are mainly described by the waveform data in each data file. The statistical method is introduced as follows:

[0123] Maximum value:

[0124] x max =max(x i )

[0125] Minimum value:

[0126] x min =min(x i )

[0127] Range:

[0128] x peak-peak =max(x i )-min(x i )

[0129] Mean:

[0130]

[0131] Root mean square error:

[0132]

[0133] Variance:

[0134]

[0135] Skewness:

[0136]

[0137] Kurtosis:

[0138]

[0139] Waveform:

[0140]

[0141] Peak:

[0142]

[0143] Impulse:

[0144]

[0145] Kurtosis:

[0146]

[0147]

[0148] wherein

[0149]

[0150] Margin:

[0151]

[0152] wherein:

[0153]

[0154] wherein the content of the frequency domain feature extraction is as follows:

[0155] 1) Short-time zero-crossing rate: measures the number of times through zero points within each frame of signal;

[0156] 2) Spectral centroid: first moment of spectrum;

[0157] 3) Spectral roll-off: measure the shape of the signal;

[0158] 4) Frequency domain root mean square: root mean square after framing;

[0159] 5) Mel-frequency cepstral coefficient: describes the overall shape of the spectral envelope;

[0160] 6) Chroma frequency: represents the chroma of the audio;

[0161] and other some frequency domain features, such as P order spectrum bandwidth, spectral flatness, constant Q color spectrum, spectral contrast, spectral graph column coefficient, timbre centroid feature, mel scale spectrum, etc.

[0162] The above features are obtained by Librosa library, which needs to be processed by frame, and the default is 512 frames. After extracting the above frequency domain features, further calculate the mean, sum, median, 25% quantile, 75% quantile, maximum, minimum, etc. Time domain features, a total of 1365 features.

[0163] 3.2, Feature selection:

[0164] In the feature construction process, 1365 features are generated, among which there must be features with zero contribution to the model (i.e. redundant features), so they need to be screened.

[0165] By constructing a multi-classification logistic regression model (supervised learning), the feature weight of the classification model is extracted using the eli5 tool library, and the features are selected according to the weight not being 0, and 264 features are obtained. Reconstruct the classification model with these 264 features.

[0166] The naming rules of the measuring points are shown in Table 1.

[0167] Table 1 Measuring point naming rules

[0168]

[0169]

[0170] 4, Model training and verification

[0171] The data samples and working condition labels are divided into training set and verification set, the multi-classification model is trained, and the accuracy of the model is verified.

[0172] 4.1, Model construction of unsupervised learning

[0173] The model construction of unsupervised learning is mainly to construct a clustering model, and the commonly used clustering algorithms include Kmeans, DBSCAN, AP clustering, hierarchical clustering, Birch clustering, GMM clustering, etc. After trying, it is found that the result of DBSCAN is acceptable. At the same time, in order to evaluate the effect of the clustering model and simplify the model label, the "fan working" or not is artificially used as the label (clustering result) concerned by the model.

[0174] 4.1.1, Elbow rule

[0175] It is known that the clustering model can artificially set the number of categories of the result, but the artificial setting has strong limitations and subjectivity, therefore, the elbow rule (according to the inflection point of the elbow rule) is used to objectively select the number of categories of the result, and when the result is clustered into 3 categories, the inflection point of the elbow appears, and the effect is the best.

[0176] 4.1.2, DBSCAN effect diagram

[0177] The DBSCAN clustering model is constructed, and the number of categories is set to 3 categories, and the clustering result is as shown in Figure 18 It can be found that although the clustering result is 3 categories, clustering into 2 categories should be the most ideal result, and the reason is that the audio data of the clustering result is a circle, the wind speed is low, which leads to slow speed, and thus the blade sweep tower sound is low, so false positives occur, so the clustering model can be further optimized.

[0178] In order to better identify different working conditions, optimizing the clustering model is one aspect, but it is ultimately an unsupervised method, while the present application can distinguish audio through the human ear to produce more specific and accurate labels, so a supervised learning model will be further constructed.

[0179] 4.2, model training and verification of supervised learning

[0180] 4.2.1 data division and label

[0181] Data division refers to the division of training data and validation data, which is related to the basis of model training and generalization. In order to cover as much audio data as possible in different time periods, a random mixed division method is used, and the ratio of the training set to the validation set is 3:1. The data division and label are shown in Table 2.

[0182] Table 2 Data division and label

[0183]

[0184] 4.2.2, algorithm selection

[0185] The construction of the classification model mainly uses a multi-classification logistic regression model, and the model hyperparameters are shown in Table 3:

[0186] Table 3 Algorithm parameter setting

[0187]

[0188] 4.2.3, model result

[0189] Training result: By calculating the accuracy, confusion matrix and classification result of the training data, as shown in Figure 19 It can be seen that the accuracy result is 0.992, and the main difference is the training process of the fourth category result.

[0190] Validation results: By calculating the accuracy, confusion matrix, and classification results of the training data, such as... Figure 20 As shown, it exhibits good generalization performance.

[0191] The ROC curves for multi-class classification results are as follows: Figure 21 As shown, the AUC value under the ROC curve is 1.

[0192] 5. Exploration of Leaf Anomaly Models

[0193] like Figure 22 As shown in the spectrogram, the energy within the red box is relatively strong, indicating a fairly obvious whistling sound, and it is concentrated on a single blade, possibly due to a blocked drainage hole. Therefore, we attempted to explore this abnormal pattern of blade whistling.

[0194] It is worth noting that the howling is accompanied by the "Doppler effect," and the process of the blades sweeping the tower is accompanied by a change in pitch from deep to thin. Because the energy in the spectrogram is stronger during the howling, in order to more clearly calculate the short-time energy, the frame length is increased. An attempt is made to use 44,100 sampling points per second as one frame, calculating the short-time energy value for each frame, such as... Figure 23 and 24 As shown.

[0195] Depend on Figure 23 and 24 As shown in the spectrogram, the short-time energy values ​​are also relatively high during the periods when whistling occurs, and the time points are relatively consistent. However, the current method still has certain shortcomings and limitations. The main reasons are the uncertainty of background noise and the different wind turbine speeds, which lead to different intensity of tower sweeping noise. Therefore, the calculated short-time energy values ​​are different, and it is impossible to determine a unified threshold for judgment. Figure 25 The spectrograms of the whistling sound at different speeds are shown.

[0196] Therefore, as Figure 25 As shown, a sound spectrum can be generated through the model as a visual criterion to determine whether a howling sound has occurred, so as to determine whether there are abnormalities such as blocked drain holes.

[0197] This invention also discloses a wind turbine abnormal state detection device based on audio data, comprising:

[0198] The WTOD monitoring and trigger acquisition module is used to monitor the operating status of the unit through the access unit PLC operating data, and trigger sound acquisition when the unit sound acquisition conditions are met, acquiring unit sound for a fixed duration and PLC operating data of key unit measurement points during the same period;

[0199] The audio detection module is used to collect sound data from key measurement points of the unit;

[0200] The state analysis and abnormality diagnosis module is configured to read the collected unit sound data, and in combination with the unit operation data, perform feature analysis, state recognition and abnormality pattern diagnosis, and output a diagnosis result.

[0201] In a specific embodiment, the audio detection module includes one or more detection modules, which include a microphone, a collector and an intelligent terminal, all of which are installed on the wind turbine unit. The microphone and the collector are configured to convert the operation sound of the unit into a digital signal, and the intelligent terminal is configured to collect the sound signal, collect key operation data of the unit PLC, perform operation state analysis and abnormality diagnosis program, store and send data to an upper server, and provide a man-machine interaction function.

[0202] Considering a simplest configuration and deployment scheme, a single wind turbine unit is monitored based on the blade sweep tower sound for blade structure health and related abnormal conditions, and one set of microphone, collector and intelligent terminal are configured, as shown in FIG. 1. The microphone is installed on the outer wall of the tower at the bottom of the tower, the collector and the intelligent terminal are installed inside the tower at the bottom of the tower, the intelligent terminal is configured with a wireless network module, and the cloud server is directly connected through a 3G / 4G / 5G network. Figure 1

[0203] If the sound in the nacelle is monitored at the same time, another set of microphone, collector and intelligent terminal are added in the nacelle based on the above scheme, as shown in FIG. 2. The microphone is installed near the sound source (considering the gearbox, generator, yaw) to be monitored, and the intelligent terminal in the nacelle can use a simplified version, which only bears the functions of collecting and forwarding digital signals, and sends the sound data collected in the cabin to the intelligent terminal at the bottom of the tower for subsequent processing through a local area network. Figure 2

[0204] When a batch of wind turbines are deployed in a wind farm, the data of each intelligent terminal of the wind turbine is collected through the wind farm network to the server in the central control room of the wind farm, and then sent to the cloud server. Considering the cost and complexity, no wireless network is used; at the same time, the man-machine interaction function of the system is transferred from the intelligent terminal of a single wind turbine to the wind farm server for deployment.

[0205] In a specific embodiment, a man-machine interaction module is further included, which is configured to provide functions such as query of diagnosis results and operation data, playback of recorded sound, visualization of sound spectrum diagram of sound data, and basic configuration functions of sound collection and abnormality diagnosis programs. The module can be deployed on the intelligent terminal on the wind turbine side, or on the wind farm server, to realize cluster monitoring of sound abnormalities of all units in the wind farm.

[0206] In addition, there is a data synchronization module, which is responsible for synchronizing the sound data, WTOD data and diagnosis result data collected and stored in the local intelligent terminal to the wind farm server and the cloud server. ​​

[0207] The cloud-deployed system software contains all the functions of the local human-computer interaction module, while providing richer interfaces, more data query and analysis functions, larger storage and stronger computing power to analyze results and assist decision-making in more dimensions.

[0208] The application further discloses a wind turbine abnormal state detection system based on audio data.

[0209] The acquisition unit is configured to acquire audio data under different operating states of the wind turbine, wherein the different operating states include that the cooling fan is not working and the fan is not working, the cooling fan is not working and the fan is working, the cooling fan is working and the fan is not working, and the cooling fan is working and the fan is working.

[0210] The filtering unit is configured to filter the audio data according to the sonogram rule of the audio data, and filter out audio parts of specific frequencies.

[0211] The feature extraction unit is configured to extract time domain features and frequency domain features of the filtered audio data, and screen the time domain features and the frequency domain features.

[0212] The model construction unit is configured to train a multi-classification model by using the screened time domain features, the frequency domain features and corresponding operating states, and obtain the trained multi-classification model.

[0213] The abnormality detection unit is configured to input the audio data based on the trained multi-classification model, obtain an output result, and realize wind turbine abnormal state detection.

[0214] The application further discloses a computer readable storage medium, which stores a computer program, and the computer program executes steps of the wind turbine abnormal state detection method based on audio data when the computer program is run by a processor. The application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program executes steps of the wind turbine abnormal state detection method based on audio data when the computer program is run by the processor.

[0215] The present application realizes all or part of the processes in the above-mentioned embodiment methods, and can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium, and the computer program can realize the steps of the above-mentioned various method embodiments when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. The memory can be used to store computer programs and / or modules, and the processor realizes various functions by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device, etc.

[0216] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application belongs to the protection scope of the present application. It should be noted that some improvements and refinements made by ordinary skilled in the art without departing from the principles of the present application should be considered as the protection scope of the present application.

Claims

1. A method for detecting abnormal states of wind turbine generators based on audio data, characterized in that, Including the following steps: 1) Acquire audio data of the wind turbine under different operating conditions; the different operating conditions include cooling fan not working and wind turbine not working, cooling fan not working and wind turbine working, cooling fan working and wind turbine not working, and cooling fan working and wind turbine working. 2) Filtering audio data: Based on the spectrogram patterns of the audio data, filter out audio components at specific frequencies; 3) Extract the time-domain and frequency-domain features of the filtered audio data, and then filter the time-domain and frequency-domain features; 4) The multi-classification model is trained using the selected time-domain features, frequency-domain features, and corresponding running states to obtain a trained multi-classification model; 5) Based on the trained multi-classification model, input audio data and obtain output results to realize the abnormal state detection of wind turbine units; In step 3), the time-domain features are statistically analyzed using waveform data in each data file, specifically including maximum value, minimum value, range, mean, root mean square error, variance, skewness, kurtosis, waveform index, impulse index, kurtosis index, and margin index. The frequency-domain features include short-time zero-crossing rate, spectral centroid, spectral attenuation, frequency-domain root mean square, Mel frequency cepstral coefficients, chromaticity frequency, P-order spectral bandwidth, spectral flatness, constant Q chromatogram, spectral contrast, spectrogram coefficients, sound quality centroid features, and Mel scaling spectrum.

2. The method for detecting abnormal states of wind turbine generators based on audio data according to claim 1, characterized in that, In step 3), a multi-classification model is constructed, and the feature weights of the multi-classification model are extracted using the eli5 tool library. Features are then selected based on the fact that the weights are not zero.

3. The method for detecting abnormal states of wind turbine generators based on audio data according to claim 2, characterized in that, In step 4), the multi-classification model includes an unsupervised learning model and a supervised learning model. The construction of the unsupervised learning model mainly focuses on building a clustering model, using the DBSCAN clustering algorithm, and objectively selecting the number of categories based on the inflection point of the elbow rule.

4. The method for detecting abnormal states of wind turbine generators based on audio data according to claim 3, characterized in that, The data partitioning in the supervised learning model adopts a random mixed partitioning method and uses a multi-class logistic regression model.

5. The method for detecting abnormal states of wind turbine generators based on audio data according to any one of claims 1 to 4, characterized in that, In step 2), the audio data is filtered by Fourier transform; the specific frequencies are obtained from the spectrogram of the audio data.

6. The method for detecting abnormal states of wind turbine generators based on audio data according to any one of claims 1 to 4, characterized in that, In step 5), a spectrogram is generated based on a multi-classification model as a visualization criterion to determine whether a whistling sound has occurred or whether a drain hole is blocked.

7. A wind turbine abnormal state detection system based on audio data, used to perform the steps of the wind turbine abnormal state detection method based on audio data as described in any one of claims 1-6, characterized in that, include: The acquisition unit is used to acquire audio data of the wind turbine under different operating conditions; The different operating states include cooling fan not working and fan not working, cooling fan not working and fan working, cooling fan working and fan not working, and cooling fan working and fan working. The filtering unit is used to filter audio data: based on the spectrogram pattern of the audio data, it filters out audio components of specific frequencies; The feature extraction unit is used to extract the time-domain and frequency-domain features of the filtered audio data, and to filter the time-domain and frequency-domain features. The model building unit is used to train the multi-classification model using the selected time-domain features, frequency-domain features, and corresponding running states to obtain a trained multi-classification model. The anomaly detection unit is used to detect abnormal states of wind turbine units by taking audio data as input and obtaining output results based on a trained multi-classification model.

8. A wind turbine abnormal state detection device based on audio data, used to perform the steps of the wind turbine abnormal state detection method based on audio data as described in any one of claims 1-6, characterized in that, include: The WTOD monitoring and trigger acquisition module is used to monitor the operating status of the unit through the access unit PLC operating data, and trigger sound acquisition when the unit sound acquisition conditions are met, acquiring unit sound for a fixed duration and PLC operating data of key measurement points of the unit during the same period; The audio detection module is used to collect sound data from key measurement points of the unit; The status analysis and anomaly diagnosis module is used to read the collected unit sound data, combine it with the unit operation data, perform feature analysis, status identification and anomaly pattern diagnosis, and output the diagnosis results.

9. The wind turbine abnormal state detection device based on audio data according to claim 8, characterized in that, The audio detection module includes one or more detection modules. Each detection module includes a microphone, a data acquisition unit, and a smart terminal. The smart terminal is equipped with a wireless network module to directly connect to the cloud server. The microphone of one detection module is installed near the key monitored sound source to pick up the audio data of the unit. The data acquisition unit is connected to the microphone and the smart terminal respectively to convert the audio data into digital signals and send them to the smart terminal.

10. The wind turbine abnormal state detection device based on audio data according to claim 9, characterized in that, One set of the detection module's microphones is installed on the outer wall of the tower at the bottom of the tower to collect sound data from the blades; the other set of the detection module's microphones is installed inside the nacelle to collect sound data from the gearbox, generator, or yaw.

11. The wind turbine abnormal state detection device based on audio data according to claim 10, characterized in that, It also includes a human-computer interaction module, which provides functions such as querying diagnostic results and operating data, recording playback, visualization of sound data spectrograms, and basic configuration functions for sound acquisition and abnormal diagnosis programs; the human-computer interaction module is integrated into the smart terminal on the wind turbine side or the wind farm server.

12. The wind turbine abnormal state detection device based on audio data according to claim 11, characterized in that, It also includes a data synchronization module, which is used to synchronize the collected and stored audio data, WTOD data, and diagnostic result data to the wind farm server and cloud server.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the wind turbine abnormal state detection method based on audio data as described in any one of claims 1 to 6.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is run by the processor, it executes the steps of the wind turbine abnormal state detection method based on audio data as described in any one of claims 1 to 6.

Images