Voiceprint detection method for internal abnormity of wind driven generator blade

Abstract

The invention discloses a voiceprint detection method for internal abnormity of a wind driven generator blade, and relates to the technical field of electrical equipment fault detection. According to the voiceprint detection method for the internal abnormity of the wind driven generator blade, the problems of data diversity and fault sample scarcity are solved through a unique mode of implementing corresponding diagnosis for different working condition information in the wind driven generator and an abnormity diagnosis mode based on an auto-encoder. The voiceprint detection method for the internal abnormity of the wind driven generator blade is used for efficiently and accurately detecting the internal abnormity of the wind driven generator blade.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbines, and particularly to a method for acoustic fingerprint detection of abnormalities inside wind turbine blades. Background Art

[0002] In the field of abnormal detection inside wind turbine blades, acoustic fingerprint technology has become a highly potential detection means. Its core principle lies in deeply analyzing the sound signals generated during the operation of the blades, accurately extracting features therefrom, and then effectively diagnosing faults. From the existing technical solutions, the method based on statistical feature quantities is more common. For example, statistical indicators such as root mean square (RMS) and kurtosis are widely used. Specifically, by continuously collecting the sound signals inside the wind turbine blades within a certain period of time and processing these signals to obtain the corresponding statistical feature information. Subsequently, by means of structural trend analysis, potential fault modes are identified based on the change trends of these statistical feature quantities over time. This method can reflect the changes in the operating state of the blades to a certain extent, but there are certain limitations in the diagnostic ability for complex fault situations.

[0003] Another common technical route is the fault identification scheme based on artificial intelligence. In this scheme, the first step is to collect a large number of sample data containing blade defects in advance and use these samples to train a specific artificial intelligence network. Through repeated iterative training, the network can learn the sound signal feature patterns corresponding to different fault types. After the training is completed, the trained network is deployed to the actual monitoring system for real-time identification of faults during the operation of the wind turbine blades. The performance of this artificial intelligence-based method highly depends on the quality and quantity of the training samples, and the training process usually requires a large amount of computing resources and time costs.

[0004] In summary, although these existing technical solutions have achieved certain results in the abnormal detection inside wind turbine blades, they still have their respective deficiencies and urgently need new technical means to further improve the accuracy and reliability of detection. The acoustic fingerprint technical solution proposed by the present invention aims to fill this technical gap and provide a more efficient and accurate solution for the abnormal detection inside wind turbine blades. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for acoustic fingerprint detection of abnormalities inside wind turbine blades to improve the accuracy of abnormal detection inside wind turbine blades.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A method for acoustic fingerprint detection of abnormalities inside wind turbine blades includes:

[0008] S100: A soundprint sensor is installed inside the wind turbine blade and a computing unit is installed inside the nacelle for data collection and analysis;

[0009] S200: processing the collected audio data using wavelet transform technology to obtain wavelet bases under different frequency components;

[0010] S300: Obtain the soundprint signal inside the blade through wavelet transform and reconstruction operations, and then use short-time Fourier transform to transform the soundprint signal into a time-frequency spectrum;

[0011] S400: For the sound pattern period segments of the fan blades under the same working conditions, a two-dimensional dynamic time warping algorithm is used to perform time alignment operations;

[0012] S500: Carry out the training process of unsupervised autoencoder;

[0013] S600: constructing an encoder network and using the encoder network to extract multiple feature vectors, and performing a clustering operation on the feature vectors based on a clustering analysis method.

[0014] Optionally, the unsupervised autoencoding training process in S200 includes setting the periodic segments as data forms of both the input layer and the output layer in the architecture design of the autoencoder, and setting a one-dimensional vector layer in the middle layer.

[0015] Optionally, the S600 constructs an encoder network and uses the encoder network to extract multiple feature vectors. After performing a clustering operation on the feature vectors based on a clustering analysis method, the distance function used in the clustering operation is set to the cosine similarity between two feature vectors. Through the clustering operation, several feature vector categories covered by the normal fan operation state are obtained. Subsequently, for the feature vector extracted from the newly acquired voiceprint fragment, its cosine similarity with the aforementioned normal feature vector category is calculated to determine whether the fan operation state is abnormal.

[0016] Optionally, S200 processes the collected audio data using wavelet transform technology to obtain wavelet bases at different frequency components, and also includes targeted screening based on the frequency band range of interest, and realizes denoising operations by signal reconstruction.

[0017] Optionally, S300 obtains the soundprint signal inside the fan through transformation and reconstruction operations, and then uses short-time Fourier transform to transform the soundprint signal into a time-frequency spectrum, and accurately marks the start and end time of the first cycle in the time-frequency spectrum to obtain a training sample.

[0018] Optionally, after accurately marking the start and end times of the first cycle in the time-frequency spectrum to obtain a training sample, and after sufficient training samples, the blade cycle truncation neural network is trained so that it can accurately truncate a complete cycle from the audio signal inside each blade. After obtaining the truncated cycle signal, the working conditions are classified by the fan model or clustering algorithm.

[0019] Compared with the prior art, in the method for detecting voiceprint anomalies inside the blades of a wind turbine provided by the present invention, the cycle screening network is used to accurately extract the complete cycle signal of the fan from the audio, reducing the difficulty of signal extraction. At the same time, the error brought during the signal extraction process is effectively reduced, improving the reliability and accuracy of the entire diagnosis system. Additionally, based on the DTW algorithm, the signals of adjacent cycles are compared, which is conducive to accurately identifying the minute changes and anomalies in the signals, thereby enabling the timely detection of potential faults in the fan. Furthermore, the anomaly diagnosis model based on the autoencoder trains a suitable encoder by collecting a large number of normal samples, greatly reducing the dependence on fault samples. The learning method based on normal samples not only reduces the workload and cost of sample collection but also improves the generalization ability of the model, thus better adapting to the anomaly diagnosis of fans under different working conditions and effectively solving the problems of difficult model training and poor diagnosis effect caused by insufficient fault samples in traditional anomaly detection. Description of the Drawings

[0020] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0021] Figure 1 It is a flowchart of the method for detecting voiceprint anomalies inside the blades of a wind turbine provided by an embodiment of the present invention;

[0022] Figure 2 It is a diagram of the wavelet transform formula for step S300 in an embodiment of the present invention;

[0023] Figure 3 It is a diagram of the short-time Fourier transform formula for step S300 in an embodiment of the present invention. Detailed Embodiments

[0024] In order to make the technical problems, technical solutions, and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0025] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. The meaning of "several" is one or more, unless otherwise clearly and specifically defined.

[0026] Since the operating conditions of wind turbines are extremely complex, the sound signals they generate are interfered by many factors, such as unstable wind speed, aging of blades, changes in ambient temperature and humidity, etc. These interference factors will cause the sound signals to show highly nonlinear and non-stationary characteristics. Methods based on statistical features often find it difficult to effectively capture these complex and changeable signal characteristics. They can only analyze from a limited statistical dimension and cannot fully explore the deep information in the sound signal, which greatly reduces the accuracy and reliability of fault diagnosis. When faced with some more hidden or early faults, this method may miss or misdiagnose due to the inability to accurately identify subtle signal changes, which in turn poses a potential threat to the safe and stable operation of wind turbines.

[0027] In practical applications, it is extremely difficult to obtain fault samples of wind turbine blades, because faults are often random and low-frequency, and obtaining fault samples usually requires high costs, including shutdown detection, disassembly of equipment, and professional human and material investment. The limited number of fault samples will lead to insufficient training data, which in turn makes the trained artificial intelligence model have poor generalization ability and cannot adapt well to complex and changeable actual working conditions. Secondly, the training process of artificial intelligence models requires powerful computing resources and a lot of time, which to a certain extent limits the widespread application and rapid deployment of this technology, especially for some wind farms located in remote areas with relatively scarce computing resources. There are many difficulties in implementation.

[0028] See also Figures 1 - 3 The embodiment of the present invention provides a method for detecting abnormalities in the internal part of a wind turbine blade, including:

[0029] S100: A soundprint sensor is installed inside the wind turbine blade and a computing unit is installed inside the nacelle for data collection and analysis;

[0030] S200: processing the collected audio data using wavelet transform technology to obtain wavelet bases under different frequency components;

[0031] S300: Obtain the voiceprint signal inside the blade through wavelet transform and reconstruction operations, and then use the short-time Fourier transform to transform the voiceprint signal into a time-frequency spectrum;

[0032] S400: For the voiceprint cycle segments of the fan blades under the same working conditions, perform time alignment operations using the two-dimensional dynamic time warping algorithm;

[0033] S500: Carry out the training process of the unsupervised autoencoder;

[0034] S600: Construct an encoder network and use the encoder network to extract multiple feature vectors, and perform clustering operations on the feature vectors based on the clustering analysis method.

[0035] Specifically, first, deploy high-precision voiceprint sensors inside the fan blade. At the same time, configure a computing unit in the nacelle to achieve the functions of data acquisition and analysis. For the collected audio data, use wavelet transform technology for processing to obtain wavelet bases under different frequency components. The wavelet transform formula is as follows (as Figure 2 shown):

[0036]

[0037] where j and k are the coefficients of the wavelet basis, representing different discrete scales and translation positions.

[0038] On this basis, perform targeted screening according to the concerned frequency band range, and achieve the denoising operation through signal reconstruction, thereby effectively improving the signal quality and providing a more reliable data basis for subsequent data analysis and processing. Through wavelet transform and reconstruction operations, the denoised voiceprint signal inside the fan blade is successfully obtained. Subsequently, use the short-time Fourier transform (STFT) to convert the voiceprint signal into a time-frequency spectrum. The transformation formula is as follows (as Figure 3 shown):

[0039]

[0040] Since the fan blade maintains a periodic rotation state, correspondingly, the time-frequency spectrum will also show periodic change characteristics. On this basis, accurately mark the start and end times of the first cycle in the time-frequency spectrum, and then obtain a training sample.

[0041] When a sufficient number of training samples are accumulated, use them to train the blade cycle truncation neural network so that it can accurately intercept a complete cycle from the audio signal inside each blade. After obtaining the truncated cycle signal, the working conditions can be classified by the fan model or the clustering algorithm.

[0042] In this embodiment, according to the actual characteristics of the fan operation, its operating conditions are subdivided into three gears: low speed, medium speed and high speed. Specifically, the low speed operating condition is defined as the operating state where the fan rotation period reaches 10 seconds or more; the medium speed operating condition corresponds to the state where the fan rotation period is in the range of 8 seconds to 10 seconds; and the high speed operating condition specifically refers to the state where the fan rotation period is less than 8 seconds.

[0043] In the subsequent data analysis process, in order to effectively eliminate the complex interference factors introduced by different working conditions, a strategy of in-depth analysis was adopted only for data collected under the same working conditions. In this way, the discreteness and uncertainty of the data caused by differences in working conditions were significantly reduced, thereby greatly improving the accuracy and reliability of data analysis. For the periodic fragments of the soundprint of the fan blades under the same working conditions, the two-dimensional dynamic time warping algorithm was used to implement the time alignment operation to facilitate the subsequent feature extraction and processing. The training process of the unsupervised autoencoder was then carried out, with the aim of effectively extracting the feature vectors in the future.

[0044] Specifically, in the architecture design of the autoencoder, the periodic fragment is set as the data form of the input layer and the output layer at the same time, and a one-dimensional vector layer is set in the middle layer. After completing the entire training process, only the one-dimensional vector layer is retained, thereby successfully obtaining a feature vector that can highly condense the key information of the periodic fragment. After completing the construction of the encoder network, the fan in the normal state is allowed to continue to run for a specific period of time. In this process, multiple feature vectors are extracted using the constructed encoder network. Then, these feature vectors are clustered based on the clustering analysis method, in which the distance function used is set as the cosine similarity between two feature vectors. Through the clustering process, several feature vector categories covered in the normal fan operation state can be obtained, and these categories represent the typical feature distribution under the normal operating mode. In the subsequent monitoring link, the cosine similarity between the special diagnosis vector extracted from the newly acquired voiceprint fragment and the normal feature vector category determined above is calculated. Once the cosine similarity is less than the preset threshold compared with all normal categories, it can be determined that the operation state of the fan is abnormal at this time.

[0045] Through the above specific implementation process, it can be known that the soundprint detection method for internal abnormalities of wind turbine blades provided by the present application provides a periodic screening network, which can accurately extract the complete periodic signal of the wind turbine from a segment of audio. In actual working conditions, the operating environment of the wind turbine is complex and there are many interference factors, and the periodic screening network can effectively solve the impact of complex working conditions on signal extraction. By accurately extracting periodic signals, a high-quality data foundation is provided for subsequent analysis and diagnosis, avoiding diagnostic errors caused by inaccurate or incomplete signals, thereby improving the reliability and accuracy of the entire diagnostic system.

[0046] Secondly, the signals of adjacent cycles are compared based on the DTW algorithm, which further enhances the accuracy and effectiveness of diagnosis. The DTW algorithm can perform dynamic time adjustment on signals of different cycles in the time series. Even when there is a certain time offset or distortion in the periodic signal, it can accurately find the optimal matching path between signals, thereby more precisely identifying the subtle changes and abnormalities in the signals, providing strong support for timely detection of potential faults in the fan.

[0047] The anomaly diagnosis model based on the autoencoder has unique advantages. In traditional fault diagnosis methods, a large number of fault samples are required for model training. However, in practical applications, it is often very difficult and costly to obtain fault samples. This model only needs to collect a large number of normal samples to train a suitable encoder, greatly reducing the dependence on fault samples.

[0048] Through learning a large number of normal samples, the autoencoder can automatically extract the feature representation of the fan in the normal operating state and establish a benchmark model for the normal state. During the actual diagnosis process, when the input signal has a large difference from the normal benchmark model, it can be judged as an abnormal state. This learning method based on normal samples not only reduces the workload and cost of sample collection, but also improves the generalization ability of the model, enabling it to better adapt to the fan anomaly diagnosis under different working conditions, effectively solving the problems of difficult model training and poor diagnosis effect caused by insufficient fault samples.

[0049] In the description of the above embodiments, the specific features, structures, materials or characteristics can be combined in a suitable manner in any one or more embodiments or examples.

[0050] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for detecting abnormalities in the internal soundprint of a wind turbine blade, characterized in that: include: S100: A soundprint sensor is installed inside the wind turbine blade and a computing unit is installed inside the nacelle for data collection and analysis; S200: processing the collected audio data using wavelet transform technology to obtain wavelet bases under different frequency components; S300: Acquire the soundprint signal inside the fan blade through wavelet transform and reconstruction operation, and then transform the soundprint signal into a time-frequency spectrum by short-time Fourier transform; S400: For the sound pattern period segments of the fan blades under the same working conditions, a two-dimensional dynamic time warping algorithm is used to perform time alignment operations; S500: Carry out the training process of unsupervised autoencoder; S600: constructing an encoder network and using the encoder network to extract multiple feature vectors, and performing a clustering operation on the feature vectors based on a clustering analysis method.

2. The method for detecting abnormalities in the internal parts of a wind turbine blade according to claim 1, characterized in that: The unsupervised autoencoding training process of S500 includes setting the periodic segments as data forms of both the input layer and the output layer in the architecture design of the autoencoder, and setting a one-dimensional vector layer in the middle layer.

3. The method for detecting abnormalities in the internal parts of a wind turbine blade according to claim 1, characterized in that: The S600 constructs an encoder network and uses the encoder network to extract multiple feature vectors. After performing a clustering operation on the feature vectors based on a clustering analysis method, it also includes setting the distance function used in the clustering operation to the cosine similarity between two feature vectors. Through the clustering operation, several feature vector categories covered by the normal fan operation state are obtained. Subsequently, for the feature vector extracted from the newly acquired voiceprint fragment, its cosine similarity with the aforementioned normal feature vector category is calculated to determine whether the fan operation state is abnormal.

4. The method for detecting abnormalities in the internal parts of a wind turbine blade according to claim 1, characterized in that: S200 uses wavelet transform technology to process the collected audio data to obtain the wavelet basis of different frequency components. It also includes targeted screening based on the frequency band range of concern and realizes denoising operations through signal reconstruction.

5. The method for detecting abnormalities in the internal parts of a wind turbine blade according to claim 1, characterized in that: S300 obtains the soundprint signal inside the wind turbine blade through wavelet transform and reconstruction operations, and then uses short-time Fourier transform to transform the soundprint signal into a time-frequency spectrum. It then accurately marks the start and end time of the first cycle in the time-frequency spectrum to obtain a training sample.

6. The method for detecting abnormalities in the internal parts of a wind turbine blade according to claim 5, characterized in that: After accurately marking the start and end time of the first cycle in the time-frequency spectrum and obtaining a training sample, the blade cycle interception neural network is trained after a sufficient number of training samples, so that it can accurately intercept a complete cycle from the audio signal inside each blade. After obtaining the truncated periodic signal, the operating condition is classified by the fan model or clustering algorithm.

Images