Defect Detection System and Method for Onshore Wind Turbines

Abstract

This application relates to the field of defect detection for onshore wind turbines, specifically disclosing a defect detection system and method for onshore wind turbines. It employs a deep learning convolutional neural network model to perform deep feature mining on multiple vibration signals of the onshore wind turbine tower structure. Based on the global implicit correlation feature information of the vibration signals of the onshore wind turbine tower, structural defects are determined. Furthermore, a wave function representation of vectors is introduced to aggregate the feature expressions of vibration information at each position of the vibration feature vector in a complex function domain-like manner. This compensates for the negative impact of differences in the distribution phase between the feature positions of each vector on the probabilistic aggregation, thereby improving the accuracy of defect detection for the onshore wind turbine tower structure. This allows for early risk warning and reduces maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of defect detection for onshore wind turbines, and more specifically, to a defect detection system and method for onshore wind turbines. Background Technology

[0002] Currently, wind energy, as one of the most promising and technologically mature renewable energy sources, has become one of the fastest-growing new energy sources. Under the guidance of national and local policies, onshore wind turbine resources are being fully developed and utilized.

[0003] The two main components of a wind turbine are the tower and the foundation. The wind turbine tower is generally a conical structure, with the nacelle and blades mounted on top. During normal operation, the wind turbine tower must withstand the combined effects of various loads, including wind loads, the weight of the nacelle and blades, the inertial forces generated by tower vibrations, and the operating loads of the wind turbine control system. Due to the randomness of wind loads, the tower will inevitably deform and vibrate under these combined loads. This vibration will create additional stress on the tower structure, thus affecting the overall stability of the wind turbine.

[0004] Since the tower structure is the foundation of onshore wind power systems, damage to it could lead to the entire turbine overturning. Therefore, there is a need for an optimized defect detection system for onshore wind turbine tower structures to detect defects, provide early warnings of risks, and reduce maintenance costs. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a defect detection system and method for onshore wind turbines. This system employs a deep learning convolutional neural network model to perform deep feature mining on multiple vibration signals of the onshore wind turbine tower structure. Based on the global implicit correlation feature information of the vibration signals of the onshore wind turbine tower, structural defects are determined. Furthermore, a wave function representation of vectors is introduced to aggregate the feature expressions of vibration information at each position of the vibration feature vector in a complex function domain-like manner. This compensates for the negative impact of differences in the distribution phase between the feature positions of each vector on the probabilistic aggregation, thereby improving the accuracy of defect detection for the onshore wind turbine tower structure. This allows for early risk warning and reduces maintenance costs.

[0006] According to one aspect of this application, a defect detection method for onshore wind turbines is provided, comprising:

[0007] Multiple vibration signals were acquired from multiple vibration sensors deployed on the tower structure of the onshore wind turbine;

[0008] Each vibration signal among the multiple vibration signals is subjected to Gram angle and field transformation to obtain multiple vibration Gram angle and field images;

[0009] Each vibration Gram angle and field image in the plurality of vibration Gram angle and field images is passed through a first convolutional neural network as a filter to obtain a plurality of vibration time-series feature vectors;

[0010] Each vibration time sequence feature vector in the plurality of vibration time sequence feature vectors is corrected to obtain a plurality of corrected vibration time sequence feature vectors;

[0011] Using the mean vector of the plurality of corrected vibration time series feature vectors as a reference feature vector, the covariance matrix between each corrected vibration time series feature vector and the reference feature vector is calculated to obtain a plurality of covariance matrices;

[0012] The plurality of covariance matrices are arranged into an input tensor and then passed through a second convolutional neural network, which acts as a feature extractor, to obtain a classification feature map; and

[0013] The classification feature map is passed through a classifier to obtain a classification result, which is used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0014] In the aforementioned defect detection method for onshore wind turbines, each vibration Gram angle and field image in the plurality of vibration Gram angle and field images is passed through a first convolutional neural network (CNN) acting as a filter to obtain a plurality of vibration time-series feature vectors. This includes: each layer of the first CNN performing the following operations on the input data during forward propagation: convolution processing on the input data to obtain a convolutional feature map; mean pooling based on the local feature matrix on the convolutional feature map to obtain a pooled feature map; and nonlinear activation on the pooled feature map to obtain an activation feature map. The output of the last layer of the first CNN is the plurality of vibration time-series feature vectors, and the input of the first layer of the first CNN is each vibration Gram angle and field image in the plurality of vibration Gram angle and field images.

[0015] In the above-mentioned defect detection method for onshore wind turbines, each vibration time-series feature vector in the plurality of vibration time-series feature vectors is corrected to obtain a plurality of corrected vibration time-series feature vectors, including: correcting each vibration time-series feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector to obtain the corrected vibration time-series feature vector.

[0016] In the above-mentioned defect detection method for onshore wind turbines, the vibration time-series feature vectors are corrected based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector to obtain the corrected vibration time-series feature vector. This includes: correcting each vibration time-series feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector using the following formula to obtain the corrected vibration time-series feature vector.

[0017] The formula is as follows:

[0018]

[0019] Where V represents each vibration time-series feature vector in the various vibration time-series feature vectors, v max -1 represents the reciprocal of the maximum value among all eigenvalues ​​at all positions in each of the vibration time-series eigenvectors, and ⊙ represents the dot product by position.

[0020] In the above-mentioned defect detection method for onshore wind turbines, the multiple covariance matrices are arranged into an input tensor and then passed through a second convolutional neural network as a feature extractor to obtain a classification feature map. This includes: using each layer of the second convolutional neural network as a feature extractor to perform convolution, pooling, and activation processing on the input data during the forward propagation of the layers to generate the classification feature map from the last layer of the second convolutional neural network, wherein the input of the first layer of the second convolutional neural network is the input tensor.

[0021] In the aforementioned defect detection method for onshore wind turbines, the classification feature map is processed by a classifier to obtain a classification result. This classification result indicates whether a structural defect exists in the tower structure of the onshore wind turbine. The method includes: the classifier processes the classification feature map using the following formula to generate the classification result, where the formula is: softmax{(W n B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the classification feature map into a vector, W1 to W n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias matrix of each fully connected layer.

[0022] According to another aspect of this application, a defect detection system for onshore wind turbines is provided, comprising:

[0023] A vibration signal acquisition model is used to acquire multiple vibration signals collected by multiple vibration sensors deployed on the tower structure of an onshore wind turbine;

[0024] The Gram angle and field transformation module is used to perform Gram angle and field transformation on each of the multiple vibration signals to obtain multiple vibration Gram angle and field images.

[0025] The filtering module is used to pass each vibration Gram angle and field image in the plurality of vibration Gram angle and field images through a first convolutional neural network as a filter to obtain a plurality of vibration time-series feature vectors;

[0026] The correction module is used to correct each vibration time sequence feature vector in the plurality of vibration time sequence feature vectors to obtain a plurality of corrected vibration time sequence feature vectors;

[0027] The covariance matrix calculation module is used to calculate the covariance matrix between each of the multiple corrected vibration time-series feature vectors and the reference feature vector, using the mean vector of the multiple corrected vibration time-series feature vectors as the reference feature vector, to obtain multiple covariance matrices.

[0028] The feature extraction module is used to arrange the multiple covariance matrices into an input tensor and then pass them through a second convolutional neural network, which acts as a feature extractor, to obtain a classification feature map; and

[0029] The classification module is used to pass the classification feature map through a classifier to obtain a classification result, which is used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0030] In the aforementioned defect detection system for onshore wind turbines, the filtering module is further configured to: perform the following operations on the input data during the forward propagation of the first convolutional neural network: convolution processing on the input data to obtain a convolutional feature map; mean pooling based on the local feature matrix on the convolutional feature map to obtain a pooled feature map; and nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the plurality of vibration time-series feature vectors, and the input of the first layer of the first convolutional neural network is each vibration Gram angle and field image in the plurality of vibration Gram angle and field images.

[0031] In the above-mentioned defect detection system for onshore wind turbines, the correction module includes: a maximum value correction unit, used to correct each vibration time sequence feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time sequence feature vector to obtain the corrected vibration time sequence feature vector.

[0032] In the above-mentioned defect detection system for onshore wind turbines, the maximum value correction unit is further used to: correct each vibration time sequence feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time sequence feature vector, using the following formula to obtain the corrected vibration time sequence feature vector;

[0033] The formula is as follows:

[0034]

[0035] Where V represents each vibration time-series feature vector in the various vibration time-series feature vectors, v max -1 represents the reciprocal of the maximum value among all eigenvalues ​​at all positions in each of the vibration time-series eigenvectors, and ⊙ represents the dot product by position.

[0036] In the aforementioned defect detection system for onshore wind turbines, the feature extraction module is further configured to: use each layer of the second convolutional neural network, which serves as the feature extractor, to perform convolution, pooling, and activation processing on the input data during the forward propagation of the layers to generate the classification feature map from the last layer of the second convolutional neural network, wherein the input of the first layer of the second convolutional neural network is the input tensor.

[0037] In the aforementioned defect detection system for onshore wind turbines, the classification module is further configured to: process the classification feature map using the following formula to generate a classification result, wherein the formula is: softmax{(W n B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the classification feature map into a vector, W1 to W n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias matrix of each fully connected layer.

[0038] According to another aspect of this application, a computer-readable medium is provided having computer program instructions stored thereon, which, when executed by a processor, cause the processor to perform the defect detection method for onshore wind turbines as described above.

[0039] Compared with existing technologies, the onshore wind turbine defect detection system and method provided in this application employs a deep learning convolutional neural network model to perform deep feature mining on multiple vibration signals of the onshore wind turbine tower structure. This allows for structural defect judgment based on the global implicit correlation feature information of the vibration signals of the onshore wind turbine tower. Furthermore, a wave function representation of vectors is introduced to aggregate the feature expressions of vibration information at each position of the vibration feature vector in a complex function domain-like manner. This compensates for the negative impact of differences in the distribution phase between the feature positions of each vector on the probabilistic aggregation, thereby improving the accuracy of defect detection for the onshore wind turbine tower structure. This enables early risk warning and reduces maintenance costs. Attached Figure Description

[0040] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0041] Figure 1 This is an application scenario diagram of the defect detection method for onshore wind turbines according to an embodiment of this application;

[0042] Figure 2 This is a flowchart of a defect detection method for an onshore wind turbine according to an embodiment of this application;

[0043] Figure 3 This is a schematic diagram of the system architecture of a defect detection method for onshore wind turbines according to an embodiment of this application;

[0044] Figure 4 This is a block diagram of a defect detection system for an onshore wind turbine according to an embodiment of this application. Detailed Implementation

[0045] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0046] Scene Overview

[0047] As mentioned earlier, wind energy, as one of the most promising and technologically mature renewable energy sources, has become one of the fastest-growing new energy sources. Under the guidance of national and local policies, onshore wind turbine resources are being fully developed and utilized.

[0048] The two main components of a wind turbine are the tower and the foundation. The wind turbine tower is generally a conical structure, with the nacelle and blades mounted on top. During normal operation, the wind turbine tower must withstand the combined effects of various loads, including wind loads, the weight of the nacelle and blades, the inertial forces generated by tower vibrations, and the operating loads of the wind turbine control system. Due to the randomness of wind loads, the tower will inevitably deform and vibrate under these combined loads. This vibration will create additional stress on the tower structure, thus affecting the overall stability of the wind turbine.

[0049] Since the tower structure is the foundation of onshore wind power systems, damage to it could lead to the entire turbine overturning. Therefore, there is a need for an optimized defect detection system for onshore wind turbine tower structures to detect defects, provide early warnings of risks, and reduce maintenance costs.

[0050] Currently, deep learning and neural networks have been widely applied in fields such as computer vision, natural language processing, and speech signal processing. Furthermore, deep learning and neural networks have demonstrated near-human or even surpassed human-level performance in areas such as image classification, object detection, semantic segmentation, and text translation.

[0051] In recent years, the development of deep learning and neural networks has provided new ideas and solutions for defect detection in the tower structure of onshore wind turbines.

[0052] Accordingly, considering that the two main components of a wind turbine are the tower and the foundation, and that the tower structure is the foundational structure of an onshore wind power system, and that the tower will inevitably deform and vibrate under the combined loads of other components, this vibration will cause additional stress on the tower structure, thus affecting the stability of the entire wind turbine. Therefore, in the technical solution of this application, it is desired to detect defects in the tower structure of the onshore wind turbine by collecting multiple vibration signals from multiple vibration sensors deployed on the tower structure. This is essentially a classification problem; that is, using the multiple vibration signals collected by the multiple vibration sensors as input data, deep feature mining is performed based on the deep neural network model, and then a classifier is used to classify and determine whether there are structural defects in the tower structure of the onshore wind turbine. This ensures the normal operation of the onshore wind turbine while providing risk warnings and reducing maintenance costs.

[0053] Specifically, in the technical solution of this application, firstly, multiple vibration signals are collected by multiple vibration sensors deployed on the tower structure of the onshore wind turbine. It should be understood that, since the Gramian angular field (GAF) is based on the Gram principle, it can transfer the time series in the classical Cartesian coordinate system to the polar coordinate system for representation. GAF can well preserve the dependencies and correlations of the original acoustic emission time series signal, possessing similar time series characteristics to the original acoustic emission signal. Depending on the trigonometric functions used for encoding, GAF can yield the Gramian angular sum field (GASF) and the Gramian angular difference field (GADF). Since the GADF transformation is irreversible, in the technical solution of this application, the GASF transformation method, which allows for inverse transformation, is selected for encoding the acoustic emission signal. That is, each vibration signal among the multiple vibration signals is subjected to Gramian angular sum field transformation to obtain multiple vibration Gramian angular sum field images. In a specific example, the encoding steps from the vibration signal to the GASF image are as follows: For a time series with C dimensions = {Q1, Q2, ..., QC}, where each dimension contains n sampling points Qi = {qi1, qi2, ..., qin}, the data for each dimension is first normalized. Then, all values ​​in the data are integrated into the range [-1, 1]. After integration, the normalized values ​​are replaced by trigonometric function values ​​(Cos values), and polar coordinates are used instead of Cartesian coordinates, thus preserving the absolute temporal relationship of the sequence.

[0054] Then, since the convolutional neural network model has excellent performance in extracting local hidden features of images, each vibration Gram angle and field image in the multiple vibration Gram angle and field images is further processed by the first convolutional neural network as a filter to extract the local high-dimensional hidden feature distribution of each vibration Gram angle and field image in the multiple vibration Gram angle and field images, thereby obtaining multiple vibration time-series feature vectors.

[0055] It should be understood that, since each of the vibration time-series feature vectors corresponds to the vibration signal obtained by the vibration sensor at a certain location, the difference in its periodic vibration will be reflected in the distribution of eigenvalues ​​at various locations of the vibration time-series feature vectors, thereby perturbing the multiple vibration time-series feature vectors and causing the feature representation to diverge. Even if the covariance matrix is ​​calculated using the mean vector as a reference, it is not easy to completely eliminate the negative impact of this feature aggregation.

[0056] Therefore, in the technical solution of this application, after obtaining the vibration time-series feature vector, the wave function characterization aggregation of the vibration time-series feature vector V is first performed, as follows:

[0057]

[0058] Where V represents each vibration time-series feature vector in the various vibration time-series feature vectors, v max -1 represents the reciprocal of the maximum value among all eigenvalues ​​at all positions in each of the vibration time-series eigenvectors, and ⊙ represents the dot product by position.

[0059] In this way, by introducing a wave function representation of the vector onto the feature distribution of the vibration time-series feature vector V, i.e., amplitude representing intensity information and phase representing periodic position information, the feature representation of vibration information at each position of the vibration time-series feature vector V is aggregated in a complex function domain. Thus, when aggregating features under a predetermined class probability based on the covariance matrix, the negative impact caused by the difference in the distribution phase between the feature positions of each vector in the class probability aggregation is compensated (that is, in-phase reinforcement and out-of-phase cancellation based on the wave function principle), thereby improving the classification ability of the classification feature map obtained after arranging the multiple covariance matrices as input tensors, and thus improving the classification accuracy.

[0060] Furthermore, in order to determine the defects of the onshore wind turbine tower structure based on the global implicit correlation information of various vibration signals of the tower structure and the overall structure of the onshore wind turbine tower, the technical solution of this application further uses the mean vector of the multiple corrected vibration time-series feature vectors as a reference feature vector, and calculates the covariance matrix between each corrected vibration time-series feature vector and the reference feature vector to obtain multiple covariance matrices. Then, the multiple covariance matrices are arranged into an input tensor to integrate the various feature information, and then implicit correlation features are extracted through a second convolutional neural network as a feature extractor to obtain a classification feature map. In this way, the classification feature map can be passed through a classifier to obtain a classification result indicating whether there are structural defects in the tower structure of the onshore wind turbine.

[0061] Based on this, this application proposes a defect detection method for onshore wind turbines, comprising: acquiring multiple vibration signals collected by multiple vibration sensors deployed on the tower structure of the onshore wind turbine; performing Gram angle and field transformation on each of the multiple vibration signals to obtain multiple vibration Gram angle and field images; passing each of the multiple vibration Gram angle and field images through a first convolutional neural network as a filter to obtain multiple vibration time-series feature vectors; and correcting each of the multiple vibration time-series feature vectors to obtain multiple... The corrected vibration time-series feature vectors are used as a reference feature vector. The covariance matrix between each corrected vibration time-series feature vector and the reference feature vector is calculated to obtain multiple covariance matrices. The multiple covariance matrices are arranged into an input tensor and then passed through a second convolutional neural network as a feature extractor to obtain a classification feature map. The classification feature map is then passed through a classifier to obtain a classification result, which is used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0062] Figure 1 The illustration shows an application scenario of the defect detection method for onshore wind turbines according to an embodiment of this application. For example... Figure 1 As shown, in this application scenario, firstly, by deploying onshore wind turbines (e.g., such as...) Figure 1 The tower structure of F shown (e.g., as shown in the figure) Figure 1 Multiple vibration sensors (e.g., as shown in D) Figure 1 Multiple vibration signals are collected (T1-Tn as shown in the diagram). These multiple vibration signals are then input to a server (e.g., a server with a defect detection algorithm deployed on the onshore wind turbine) that is equipped with such an algorithm. Figure 1 In the S shown, the server is able to process the acquired multiple vibration signals using a defect detection algorithm for onshore wind turbines to generate a classification result indicating whether there are structural defects in the tower structure of the onshore wind turbine.

[0063] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0064] Exemplary methods

[0065] Figure 2 The diagram illustrates a flowchart of a defect detection method for onshore wind turbines. For example... Figure 2As shown, the defect detection method for onshore wind turbines according to an embodiment of this application includes: S110, acquiring multiple vibration signals collected by multiple vibration sensors deployed on the tower structure of the onshore wind turbine; S120, performing Gram angle and field transformation on each of the multiple vibration signals to obtain multiple vibration Gram angle and field images; S130, passing each of the multiple vibration Gram angle and field images through a first convolutional neural network as a filter to obtain multiple vibration time-series feature vectors; S140, correcting each of the multiple vibration time-series feature vectors to obtain... Multiple corrected vibration time-series feature vectors; S150, using the mean vector of the multiple corrected vibration time-series feature vectors as a reference feature vector, calculating the covariance matrix between each corrected vibration time-series feature vector and the reference feature vector to obtain multiple covariance matrices; S160, arranging the multiple covariance matrices into an input tensor and passing it through a second convolutional neural network as a feature extractor to obtain a classification feature map; and S170, passing the classification feature map through a classifier to obtain a classification result, the classification result being used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0066] Figure 3 The illustration shows a schematic diagram of the architecture of a defect detection method for onshore wind turbines according to an embodiment of this application. Figure 3 As shown, in the network architecture of the defect detection method for onshore wind turbines, firstly, the obtained multiple vibration signals (e.g., such as...) are processed... Figure 3 Each vibration signal in P1 (as shown) is subjected to Gram angle and field transformation to obtain multiple vibration Gram angle and field images (e.g., such as...). Figure 3 (as shown in P2); then, each vibration Gram angle and field image in the plurality of vibration Gram angle and field images is passed through a first convolutional neural network (e.g., as shown in P2) as a filter. Figure 3 The CNN1 shown is used to obtain multiple vibrational temporal feature vectors (e.g., such as...). Figure 3 The VF1 shown in the figure is then corrected to obtain multiple corrected vibration time-series feature vectors (e.g., VF1). Figure 3 (as shown in VF2); then, using the mean vector of the plurality of corrected vibration time-series feature vectors as a reference feature vector, the covariance matrix between each corrected vibration time-series feature vector and the reference feature vector is calculated to obtain a plurality of covariance matrices (e.g., as shown in VF2); Figure 3 The MF shown in the diagram is then arranged into an input tensor (e.g., as shown in the diagram). Figure 3The T shown in the diagram is then passed through a second convolutional neural network (e.g., as shown in the diagram) which acts as a feature extractor. Figure 3 The CNN2 shown is used to obtain classification feature maps (e.g., such as...). Figure 3 The FC shown in the diagram); and finally, the classification feature map is passed through a classifier (e.g., such as...). Figure 3 The classifier shown is used to obtain a classification result, which is used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0067] In steps S110 and S120, multiple vibration signals are acquired by multiple vibration sensors deployed on the tower structure of the onshore wind turbine. Gram angle and field transformations are then performed on each of the multiple vibration signals to obtain multiple vibration Gram angle and field images. As mentioned earlier, it should be understood that considering the two main components of the wind turbine are the tower and the foundation, and the tower structure is the foundation structure of the onshore wind power system, and considering other loads, the tower will inevitably deform and vibrate. This vibration will cause additional stress on the tower structure, thereby affecting the stability of the entire wind turbine. Therefore, in the technical solution of this application, it is desirable to use multiple vibration signals acquired by multiple vibration sensors deployed on the tower structure of the onshore wind turbine to detect defects in the tower structure of the onshore wind turbine. This is essentially a classification problem. That is, multiple vibration signals collected by multiple vibration sensors are used as input data, and deep feature mining is performed on them based on the deep neural network model. Then, a classifier is used to classify and judge whether there are structural defects in the tower structure of the onshore wind turbine. In this way, while ensuring the normal operation of the onshore wind turbine, risk warnings are given and maintenance costs are reduced.

[0068] Specifically, in the technical solution of this application, firstly, multiple vibration signals are collected by multiple vibration sensors deployed on the tower structure of the onshore wind turbine. It should be understood that, since the Gramian angular field (GAF) is based on the Gram principle, it can transfer the time series in the classical Cartesian coordinate system to the polar coordinate system for representation. GAF can well preserve the dependencies and correlations of the original acoustic emission time series signal, possessing similar time series characteristics to the original acoustic emission signal. Depending on the trigonometric functions used for encoding, GAF can yield the Gramian angular sum field (GASF) and the Gramian angular difference field (GADF). Since the GADF transformation is irreversible, in the technical solution of this application, the GASF transformation method, which allows for inverse transformation, is selected for encoding the acoustic emission signal. That is, each vibration signal among the multiple vibration signals is subjected to Gramian angular sum field transformation to obtain multiple vibration Gramian angular sum field images.

[0069] Accordingly, in a specific example, the encoding steps from the vibration signal to the GASF image are as follows: For a time series with C dimensions = {Q1, Q2, ..., QC}, where each dimension contains n sampling points Qi = {qi1, qi2, ..., qin}, the data for each dimension is first normalized. Then, all values ​​in the data are integrated into the range [-1, 1]. After integration, the normalized values ​​are replaced by trigonometric function values ​​(Cos values), and polar coordinates are used instead of Cartesian coordinates, thus preserving the absolute temporal relationship of the sequence.

[0070] In step S130, each vibration Gram angle and field image in the plurality of vibration Gram angle and field images is passed through a first convolutional neural network acting as a filter to obtain multiple vibration temporal feature vectors. That is, since convolutional neural network models have excellent performance in extracting local latent features of images, in the technical solution of this application, each vibration Gram angle and field image in the plurality of vibration Gram angle and field images is further processed by the first convolutional neural network acting as a filter to extract the local high-dimensional latent feature distribution of each vibration Gram angle and field image in the plurality of vibration Gram angle and field images, thereby obtaining multiple vibration temporal feature vectors.

[0071] Specifically, in the embodiments of this application, the process of passing each vibration Gram angle and field image in the plurality of vibration Gram angle and field images through a first convolutional neural network as a filter to obtain a plurality of vibration time-series feature vectors includes: each layer of the first convolutional neural network performing the following on the input data during the forward propagation of the layer: performing convolution processing on the input data to obtain a convolutional feature map; performing mean pooling based on the local feature matrix on the convolutional feature map to obtain a pooling feature map; and performing nonlinear activation on the pooling feature map to obtain an activation feature map; wherein, the output of the last layer of the first convolutional neural network is the plurality of vibration time-series feature vectors, and the input of the first layer of the first convolutional neural network is each vibration Gram angle and field image in the plurality of vibration Gram angle and field images.

[0072] In step S140, each vibration time-series feature vector in the plurality of vibration time-series feature vectors is corrected to obtain a plurality of corrected vibration time-series feature vectors. It should be understood that since each vibration time-series feature vector corresponds to a vibration signal obtained from a vibration sensor at a specific location, the difference in its periodic vibration will be reflected in the distribution of eigenvalues ​​at various locations of the vibration time-series feature vector, thereby perturbing the plurality of vibration time-series feature vectors and causing feature expression divergence. Even if the covariance matrix is ​​calculated using the mean vector as a reference, it is not easy to completely eliminate the negative impact of this feature aggregation. Therefore, in the technical solution of this application, after obtaining the vibration time-series feature vectors, wave function characterization aggregation of the vibration time-series feature vectors V is first performed. Accordingly, in a specific example, each vibration time-series feature vector is corrected based on the reciprocal of the maximum value among the eigenvalues ​​at all locations of each vibration time-series feature vector to obtain the corrected vibration time-series feature vector.

[0073] Specifically, in this embodiment of the application, the process of correcting each vibration time-series feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector to obtain the corrected vibration time-series feature vector includes: correcting each vibration time-series feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector using the following formula to obtain the corrected vibration time-series feature vector;

[0074] The formula is as follows:

[0075]

[0076] Where V represents each vibration time-series feature vector in the various vibration time-series feature vectors, v max -1The expression represents the reciprocal of the maximum value among all eigenvalues ​​at all positions in each of the vibration time-series feature vectors, and ⊙ represents the dot product by position. It should be understood that by introducing a wave function representation of the vector onto the feature distribution of the vibration time-series feature vector V—that is, amplitude representing intensity information and phase representing periodic position information—the feature representation of vibration information at each position of the vibration time-series feature vector V is aggregated in a complex function domain. This compensates for the negative impact of the difference in phase distribution between the feature positions of each vector on the class probability aggregation when aggregating features under a predetermined class probability based on the covariance matrix (i.e., in-phase reinforcement and out-of-phase cancellation based on the wave function principle). This improves the classification ability of the classification feature map obtained after arranging the multiple covariance matrices as input tensors, thereby improving classification accuracy.

[0077] In steps S150 and S160, the mean vector of the plurality of corrected vibration time-series feature vectors is used as a reference feature vector. The covariance matrix between each corrected vibration time-series feature vector and the reference feature vector is calculated to obtain multiple covariance matrices. These multiple covariance matrices are then arranged as input tensors and passed through a second convolutional neural network, which acts as a feature extractor, to obtain a classification feature map. It should be understood that, further, in order to determine the defects of the onshore wind turbine tower structure based on the global implicit correlation information of each vibration signal of the onshore wind turbine tower structure and the overall tower structure of the onshore wind turbine, the technical solution of this application further uses the mean vector of the plurality of corrected vibration time-series feature vectors as a reference feature vector, and calculates the covariance matrix between each corrected vibration time-series feature vector and the reference feature vector to obtain multiple covariance matrices. Then, the multiple covariance matrices are arranged into an input tensor to integrate the various feature information, and then the implicit correlation features are extracted through a second convolutional neural network, which acts as a feature extractor, to obtain a classification feature map.

[0078] Accordingly, in a specific example, the process of arranging the plurality of covariance matrices into an input tensor and then passing them through a second convolutional neural network as a feature extractor to obtain a classification feature map includes: using each layer of the second convolutional neural network as a feature extractor to perform convolution, pooling, and activation processing on the input data during the forward pass of the layer to generate the classification feature map by the last layer of the second convolutional neural network, wherein the input of the first layer of the second convolutional neural network is the input tensor.

[0079] In step S170, the classification feature map is processed by a classifier to obtain a classification result, which indicates whether there are structural defects in the tower structure of the onshore wind turbine. Accordingly, in a specific example, the classifier processes the classification feature map using the following formula to generate the classification result, where the formula is: softmax{(W n B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the classification feature map into a vector, W1 to W n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias matrix of each fully connected layer.

[0080] In summary, the defect detection method for onshore wind turbines according to this application is explained. It employs a deep learning convolutional neural network model to perform deep feature mining on multiple vibration signals of the onshore wind turbine tower structure. Based on the global implicit correlation feature information of the vibration signals of the onshore wind turbine tower, structural defects are determined. Furthermore, in this process, the wave function representation of vectors is introduced to aggregate the feature expressions of vibration information at each position of the vibration feature vector in a complex function domain-like manner. This compensates for the negative impact of differences in the distribution phase between the feature positions of each vector on the probabilistic aggregation, thereby improving the accuracy of defect detection for the onshore wind turbine tower structure. This allows for early risk warning and reduces maintenance costs.

[0081] Exemplary System

[0082] Figure 4 The diagram illustrates a block diagram of a defect detection system for an onshore wind turbine according to an embodiment of this application. Figure 4As shown, the defect detection system 400 for onshore wind turbines according to an embodiment of this application includes: a vibration signal acquisition model 410, used to acquire multiple vibration signals collected by multiple vibration sensors deployed on the tower structure of the onshore wind turbine; a Gram angle and field transformation module 420, used to perform Gram angle and field transformation on each of the multiple vibration signals to obtain multiple vibration Gram angle and field images; a filtering module 430, used to pass each of the multiple vibration Gram angle and field images through a first convolutional neural network as a filter to obtain multiple vibration time-series feature vectors; and a correction module 440, used to correct each of the multiple vibration time-series feature vectors. The system performs row correction to obtain multiple corrected vibration time-series feature vectors; a covariance matrix calculation module 450 is used to calculate the covariance matrix between each corrected vibration time-series feature vector and the reference feature vector using the mean vector of the multiple corrected vibration time-series feature vectors as a reference feature vector to obtain multiple covariance matrices; a feature extraction module 460 is used to arrange the multiple covariance matrices into an input tensor and then pass them through a second convolutional neural network as a feature extractor to obtain a classification feature map; and a classification module 470 is used to pass the classification feature map through a classifier to obtain a classification result, the classification result being used to indicate whether there are structural defects in the tower structure of the onshore wind turbine.

[0083] In one example, in the aforementioned defect detection system 400 for onshore wind turbines, the filtering module 430 is further configured to: perform the following on the input data during the forward propagation of the first convolutional neural network: convolution processing on the input data to obtain a convolutional feature map; mean pooling based on the local feature matrix on the convolutional feature map to obtain a pooled feature map; and nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the plurality of vibration time-series feature vectors, and the input of the first layer of the first convolutional neural network is each vibration Gram angle and field image in the plurality of vibration Gram angle and field images.

[0084] In one example, in the above-mentioned defect detection system 400 for onshore wind turbines, the correction module 440 includes: a maximum value correction unit 441, which is used to correct each vibration time sequence feature vector based on the reciprocal of the maximum value among the feature values ​​of all positions in each vibration time sequence feature vector to obtain the corrected vibration time sequence feature vector.

[0085] In one example, in the above-mentioned defect detection system 400 for onshore wind turbines, the maximum value correction unit 441 is further configured to: correct each vibration time-series feature vector based on the reciprocal of the maximum value among the feature values ​​at all positions in each vibration time-series feature vector, using the following formula to obtain the corrected vibration time-series feature vector;

[0086] The formula is as follows:

[0087]

[0088] Where V represents each vibration time-series feature vector in the various vibration time-series feature vectors, v max -1 represents the reciprocal of the maximum value among all eigenvalues ​​at all positions in each of the vibration time-series eigenvectors, and ⊙ represents the dot product by position.

[0089] In one example, in the aforementioned onshore wind turbine defect detection system 400, the feature extraction module 460 is further configured to: use each layer of the second convolutional neural network, which serves as a feature extractor, to perform convolution, pooling, and activation processing on the input data during the forward pass of the layers to generate the classification feature map by the last layer of the second convolutional neural network, wherein the input of the first layer of the second convolutional neural network is the input tensor.

[0090] In one example, in the aforementioned defect detection system 400 for onshore wind turbines, the classification module 470 is further configured to: process the classification feature map using the following formula to generate a classification result, wherein the formula is: softmax{(W n B n ):…:(W1,B1)|Project(F)}, where Project(F) represents projecting the classification feature map into a vector, W1 to W n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias matrix of each fully connected layer.

[0091] Here, those skilled in the art will understand that the specific functions and operations of each unit and module in the aforementioned onshore wind turbine defect detection system 400 have been referenced above. Figures 1 to 3 The method for detecting defects in onshore wind turbines is described in detail here, and therefore, its repeated description will be omitted.

[0092] As described above, the onshore wind turbine defect detection system 400 according to embodiments of this application can be implemented in various terminal devices, such as servers for onshore wind turbine defect detection algorithms. In one example, the onshore wind turbine defect detection system 400 according to embodiments of this application can be integrated into a terminal device as a software module and / or a hardware module. For example, the onshore wind turbine defect detection system 400 can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the onshore wind turbine defect detection system 400 can also be one of many hardware modules of the terminal device.

[0093] Alternatively, in another example, the defect detection system 400 of the onshore wind turbine and the terminal device can also be separate devices, and the defect detection system 400 of the onshore wind turbine can be connected to the terminal device via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0094] Exemplary computer program products and computer-readable storage media

[0095] In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps of the functions in the defect detection methods for onshore wind turbines according to various embodiments of this application described in the "Exemplary Methods" section of this specification.

[0096] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0097] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the defect detection method for onshore wind turbines described in the "Exemplary Methods" section of this specification.

[0098] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0099] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0100] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0101] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0102] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0103] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method of defect detection of an onshore wind turbine, characterized in that, The method comprises: acquiring a plurality of vibration signals collected by a plurality of vibration sensors deployed on a tower structure of an onshore wind turbine; respectively performing Gram angle and field transformation on each of the plurality of vibration signals to obtain a plurality of vibration Gram angle and field images; respectively passing each of the plurality of vibration Gram angle and field images through a first convolutional neural network serving as a filter to obtain a plurality of vibration time series feature vectors; correcting each of the plurality of vibration time series feature vectors to obtain a plurality of corrected vibration time series feature vectors; calculating a covariance matrix between each of the plurality of corrected vibration time series feature vectors and a reference feature vector which is a mean vector of the plurality of corrected vibration time series feature vectors to obtain a plurality of covariance matrices; arranging the plurality of covariance matrices into an input tensor and passing the input tensor through a second convolutional neural network serving as a feature extractor to obtain a classification feature map; and passing the classification feature map through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the tower structure of the onshore wind turbine has a structural defect. The method further comprises: correcting each of the plurality of vibration time series feature vectors based on a reciprocal of a maximum value among feature values at all positions in each of the plurality of vibration time series feature vectors to obtain the corrected vibration time series feature vector. The method further comprises: performing, by each layer of the first convolutional neural network, convolution processing, pooling processing and activation processing on input data in a forward propagation of the layer to generate the classification feature map by a last layer of the second convolutional neural network, wherein an input of a first layer of the second convolutional neural network is the input tensor. wherein represents the inverse of the maximum of the eigenvalues of all positions in each of the vibration time series feature vectors, represents the inverse of the maximum of the eigenvalues of all positions in each of the vibration time series feature vectors, represents a point-wise multiplication by position; The method further comprises: performing, by each layer of the first convolutional neural network, the following on input data in a forward propagation of the layer:

2. The method for detecting defects of a land-based wind turbine according to claim 1, characterized in that performing convolution processing on the input data to obtain a convolution feature map; performing local feature matrix based mean pooling on the convolution feature map to obtain a pooling feature map; and performing non-linear activation on the pooling feature map to obtain an activation feature map. The method further comprises: passing the classification feature map through the classifier to obtain the classification result, wherein the classification result is used to indicate whether the tower structure of the onshore wind turbine has a structural defect. The method further comprises:

3. The method of defect detection of a land-based wind turbine according to claim 2, characterized in that ​ The classifier processes the classification feature map to generate a classification result with a formula as follows: wherein denotes projecting the classification feature map into a vector, to is a weight matrix of each layer of fully connected layers, to denotes a bias matrix of each layer of fully connected layers.

4. A defect detection system of a land-based wind turbine for carrying out the method of defect detection of a land-based wind turbine according to claim 1, characterized in that ​ The vibration signal acquisition model is configured to acquire a plurality of vibration signals collected by a plurality of vibration sensors deployed on a tower structure of an onshore wind turbine; The Gram angle and field transformation module is configured to perform Gram angle and field transformation on each of the plurality of vibration signals to obtain a plurality of vibration Gram angle and field images. The filtering module is configured to pass each of the plurality of vibration Gram angle and field images through a first convolutional neural network as a filter to obtain a plurality of vibration time series feature vectors. The correction module is configured to correct each of the plurality of vibration time series feature vectors to obtain a plurality of corrected vibration time series feature vectors. The covariance matrix calculation module is configured to calculate a covariance matrix between each of the plurality of corrected vibration time series feature vectors and a reference feature vector, which is a mean vector of the plurality of corrected vibration time series feature vectors, to obtain a plurality of covariance matrices. The feature extraction module is configured to pass the plurality of covariance matrices arranged as an input tensor through a second convolutional neural network as a feature extractor to obtain a classification feature map. The classification module is configured to pass the classification feature map through a classifier to obtain a classification result, which is used to indicate whether the tower structure of the onshore wind turbine has a structural defect. The filtering module is further configured to perform the following operations in the forward propagation of each layer of the first convolutional neural network:

5. The defect detection system of a land-based wind turbine according to claim 4, characterized in that perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on a local feature matrix on the convolution feature map to obtain a pooling feature map; and perform nonlinear activation on the pooling feature map to obtain an activated feature map. The output of the last layer of the first convolutional neural network is the plurality of vibration time series feature vectors, and the input of the first layer of the first convolutional neural network is each of the plurality of vibration Gram angle and field images. ​ ​

Images