Offshore wind turbine structure state monitoring system and method based on artificial intelligence

Abstract

The application relates to the field of intelligent monitoring of offshore wind turbine structure states, and particularly discloses an offshore wind turbine structure state monitoring system and method based on artificial intelligence, which adopts an artificial intelligence-based monitoring method, uses a convolutional neural network model as a feature extractor, generates high-dimensional implicit features of vibration signals excited by blade working characteristics through implicit correlation features of blade rotating speed values at multiple predetermined time points, and performs differential comparison with high-dimensional implicit features of actually detected vibration signals in the predetermined time period, so that an excited vibration signal caused by blade rotation can be extracted from vibration signals of a foundation structure of an offshore wind turbine, the differential feature matrix can more directly reflect a vibration mode of the foundation structure of the offshore wind turbine, and then performance state monitoring of the foundation structure of the offshore wind turbine can be performed based on the vibration mode of the foundation structure of the offshore wind turbine.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent monitoring of offshore wind turbine structure state, and more particularly, to an offshore wind turbine structure state monitoring system based on artificial intelligence and a method thereof. BACKGROUND

[0002] By the end of 2013, the total installed capacity of wind power in the world reached 318GW, of which offshore wind power was 6.8GW, and the total installed capacity of wind power in China was 91.4GW, and the offshore wind power installation was 428MW. With the passage of time, the safety accidents of wind turbine operation also showed an upward trend. In various wind power accidents, structural failure is only second to fire and blade failure, so it is of great significance to monitor the state of the wind turbine structure system.

[0003] Compared with land, the load environment of offshore wind turbine is more complex, and the influence mechanism of various load excitations such as wind, wave, current, even ice, typhoon, earthquake and the like under extreme conditions on the structure is more complex. At the same time, since offshore wind turbines are far away from land, wind farm management personnel cannot regularly evaluate and detect the structure, and the response time for accidents is much longer than that for onshore wind turbines.

[0004] Therefore, an optimized state monitoring scheme for offshore wind turbine structure is expected.

[0005] At present, deep learning and neural networks have been widely used in computer vision, natural language processing, text signal processing and other fields. In addition, deep learning and neural networks have also shown a level close to or even surpassing that of humans in the fields of image classification, object detection, semantic segmentation, text translation and the like.

[0006] In recent years, the development of deep learning and neural networks has provided a solution for the state monitoring of offshore wind turbine structures. SUMMARY

[0007] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide an offshore wind turbine structure state monitoring system based on artificial intelligence and a method thereof, which adopts an artificial intelligence-based monitoring method, uses a convolutional neural network model as a feature extractor, generates high-dimensional implicit features of a vibration signal excited by blade operating characteristics through the implicit correlation features of blade speed values at multiple predetermined time points, and differentially compares the high-dimensional implicit features of the vibration signal actually detected in the predetermined time period, so as to extract the excitation vibration signal caused by the rotation of the blade from the vibration signal of the foundation structure of the offshore wind turbine, so that the differential feature matrix can more directly reflect the vibration mode of the foundation structure of the offshore wind turbine, and then the performance state monitoring of the foundation structure of the offshore wind turbine is performed based on the vibration mode of the foundation structure of the offshore wind turbine.

[0008] According to an aspect of the present application, there is provided an offshore wind turbine structure state monitoring system based on artificial intelligence, comprising:

[0009] a data acquisition module configured to acquire vibration signals collected by a vibration sensor deployed on a foundation structure of an offshore wind turbine within a predetermined time period, and blade rotation speed values at a plurality of predetermined time points within the predetermined time period;

[0010] a rotation speed data encoding module configured to encode the blade rotation speed values at the plurality of predetermined time points within the predetermined time period through a time series encoder comprising a one-dimensional convolution layer to obtain a rotation speed feature vector;

[0011] a waveform generation module configured to generate a generated vibration signal by passing the rotation speed feature vector through a waveform generator;

[0012] a generated waveform encoding module configured to encode a waveform diagram of the generated vibration signal through a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix;

[0013] a detected vibration waveform encoding module configured to encode a waveform diagram of the vibration signals within the predetermined time period through a second convolutional neural network as a feature extractor to obtain a detected vibration feature matrix;

[0014] a difference module configured to calculate a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix;

[0015] a feature correction module configured to correct feature values at each position in the difference feature matrix based on a mean and a variance of a set of feature values of the difference feature matrix to obtain a corrected difference feature matrix; and

[0016] a state monitoring result generation module configured to pass the corrected difference feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the performance of the foundation structure of the offshore wind turbine meets predetermined requirements.

[0017] In the offshore wind turbine structure state monitoring system based on artificial intelligence described above, the rotation speed data encoding module comprises: an input vector construction unit configured to arrange the blade rotation speed values at the plurality of predetermined time points within the predetermined time period into a one-dimensional input vector according to a time dimension; and a fully connected encoding unit configured to use a fully connected layer of the time series encoder to perform fully connected encoding on the input vector according to the following formula to extract high-dimensional hidden features of feature values at each position in the input vector: wherein X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, represents matrix multiplication; a one-dimensional convolutional encoding unit configured to use a one-dimensional convolutional layer of the time series encoder to perform one-dimensional convolutional encoding on the input vector to extract high-dimensional hidden correlation features between feature values at respective positions in the input vector according to the following formula:

[0018]

[0019] where a is a width of the convolution kernel in the x direction, F is a parameter vector of the convolution kernel, G is a local vector matrix operated with the convolution kernel function, w is a size of the convolution kernel, and X represents the input vector.

[0020] In the above offshore wind turbine structure state monitoring system based on artificial intelligence, the waveform generator is a generative adversarial network.

[0021] In the above offshore wind turbine structure state monitoring system based on artificial intelligence, the training process of the waveform generator includes: obtaining training data, the training data including training blade speed values at a plurality of predetermined time points in a predetermined time period and a reference waveform image of the predetermined time period; passing the training blade speed values at the plurality of predetermined time points in the predetermined time period in the training data through the time series encoder including the one-dimensional convolutional layer to obtain a training speed feature vector; passing the training speed feature vector through a generator model of the waveform generator to obtain a training waveform generated image; passing the training waveform generated image through a discriminator neural network model of the waveform generator to obtain a first feature map; passing the reference waveform image through the discriminator neural network model of the waveform generator to obtain a second feature map; determining whether values at predetermined positions in the first feature map and the second feature map are the same; in response to the values at the predetermined positions in the first feature map and the second feature map being the same, calculating a negative value of a base-2 logarithm value of the values at the predetermined positions as a first value; in response to the values at the predetermined positions in the first feature map and the second feature map being different, calculating a base-2 logarithm value of the values at the predetermined positions as a second value; calculating a sum of an average value of the first value at the positions with the same value and an average value of the second value at the positions with different values as the discriminator loss function value; and training the waveform generator with the discriminator loss function value and through back propagation of gradient descent.

[0022] In the offshore wind turbine structure state monitoring system based on artificial intelligence, the waveform generation coding module is further used for: each layer of the first convolutional neural network respectively performs the following in the forward transmission of the layer: convolution processing on the input data to obtain a convolution feature map; mean value pooling based on a local channel dimension is performed on the convolution feature map to obtain a pooling feature map; and non-linear activation is performed on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network is the vibration feature matrix generated, and the input of the first layer of the first convolutional neural network is the waveform graph of the generated vibration signal.

[0023] In the offshore wind turbine structure state monitoring system based on artificial intelligence, the vibration waveform detection coding module is further used for: each layer of the second convolutional neural network respectively performs the following in the forward transmission of the layer: convolution processing on the input data to obtain a convolution feature map; mean value pooling based on a local channel dimension is performed on the convolution feature map to obtain a pooling feature map; and non-linear activation is performed on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the second convolutional neural network is the vibration feature matrix detected, and the input of the first layer of the second convolutional neural network is the waveform graph of the vibration signal in the predetermined time period.

[0024] In the offshore wind turbine structure state monitoring system based on artificial intelligence, the feature correction module is further used for: based on the mean value and variance of the feature value set of the difference feature matrix, the feature values of each position in the difference feature matrix are corrected by the following formula to obtain the corrected difference feature matrix.

[0025] The formula is:

[0026]

[0027] wherein m i，j represents the feature value of each position in the difference feature matrix, W and H are the width and height of the difference feature matrix respectively, and μ and σ are the mean value and variance of the feature value set of the difference feature matrix, and α is a weight hyperparameter.

[0028] In the offshore wind turbine structure state monitoring system based on artificial intelligence, the state monitoring result generation module is further used for: the classifier processes the difference feature matrix by the following formula to generate a classification result, wherein the formula is: softmax{(W n , B n ):…:(W1, B1)|Project(F)}, wherein Project(F) represents projecting the difference feature matrix into a vector, W1 to W nare weight matrices of the full connection layers of each layer, B1 to B n are bias matrices of the full connection layers of each layer.

[0029] According to another aspect of the present application, an offshore wind turbine structure state monitoring method based on artificial intelligence comprises:

[0030] obtaining vibration signals of a predetermined time period collected by vibration sensors deployed on the foundation structure of the offshore wind turbine, and blade rotation speed values at a plurality of predetermined time points within the predetermined time period;

[0031] obtaining blade rotation speed values at a plurality of predetermined time points within the predetermined time period through a time series encoder containing a one-dimensional convolution layer to obtain a rotation speed feature vector;

[0032] obtaining a generated vibration signal through a waveform generator from the rotation speed feature vector;

[0033] obtaining a waveform graph of the generated vibration signal through a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix;

[0034] obtaining a waveform graph of the vibration signals of the predetermined time period through a second convolutional neural network as a feature extractor to obtain a detected vibration feature matrix;

[0035] calculating a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix;

[0036] based on the mean and variance of the feature value set of the difference feature matrix, correcting the feature values at each position in the difference feature matrix to obtain a corrected difference feature matrix; and

[0037] obtaining a classification result through a classifier from the corrected difference feature matrix, the classification result being used to indicate whether the performance of the foundation structure of the offshore wind turbine meets predetermined requirements.

[0038] In the above offshore wind turbine structure state monitoring method based on artificial intelligence, obtaining blade rotation speed values at a plurality of predetermined time points within the predetermined time period through a time series encoder containing a one-dimensional convolution layer to obtain a rotation speed feature vector comprises: arranging the blade rotation speed values at a plurality of predetermined time points within the predetermined time period into a one-dimensional input vector according to the time dimension; using the full connection layer of the time series encoder to fully connect and encode the input vector to extract high-dimensional hidden features of the feature values at each position in the input vector according to the following formula, wherein the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, represents matrix multiplication; a one-dimensional convolution layer of the time sequence encoder is used to perform one-dimensional convolution coding on the input vector according to the following formula to extract high-dimensional hidden correlation features between feature values at different positions in the input vector, wherein the formula is:

[0039]

[0040] wherein a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix for the convolution kernel function operation, w is the size of the convolution kernel, and X represents the input vector.

[0041] In the offshore wind turbine structure state monitoring method based on artificial intelligence, the waveform generator is a generative adversarial network.

[0042] In the offshore wind turbine structure state monitoring method based on artificial intelligence, the training process of the waveform generator includes: obtaining training data, wherein the training data includes training blade speed values at a plurality of predetermined time points in a predetermined time period and a reference waveform image of the predetermined time period; passing the training blade speed values at the plurality of predetermined time points in the predetermined time period in the training data through the time sequence encoder including the one-dimensional convolution layer to obtain a training speed feature vector; passing the training speed feature vector through the generator model of the waveform generator to obtain a training waveform generated image; passing the training waveform generated image through the discriminator neural network model of the waveform generator to obtain a first feature map; passing the reference waveform image through the discriminator neural network model of the waveform generator to obtain a second feature map; determining whether the values at predetermined positions in the first feature map and the second feature map are the same; in response to the values at the predetermined positions in the first feature map and the second feature map being the same, calculating the negative value of the base-2 logarithm value of the value at the predetermined position as a first value; in response to the values at the predetermined positions in the first feature map and the second feature map being different, calculating the base-2 logarithm value of the value at the predetermined position as a second value; calculating the sum of the average value of the first value at the same value position and the average value of the second value at the different value position as the discriminator loss function value; and training the waveform generator by the gradient descent backpropagation of the discriminator loss function value.

[0043] In the offshore wind turbine structure state monitoring method based on artificial intelligence, the waveform diagram of the generated vibration signal is input into a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix, each layer of the first convolutional neural network respectively performs the following in the forward propagation of the layer: performing convolution processing on the input data to obtain a convolution feature map; performing mean pooling based on the local channel dimension on the convolution feature map to obtain a pooling feature map; and performing nonlinear activation on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the first convolutional neural network is the generated vibration feature matrix, and the input of the first layer of the first convolutional neural network is the waveform diagram of the generated vibration signal.

[0044] In the offshore wind turbine structure state monitoring method based on artificial intelligence, the waveform diagram of the vibration signal in the predetermined time period is input into a second convolutional neural network as a feature extractor to obtain a detection vibration feature matrix, each layer of the second convolutional neural network respectively performs the following in the forward propagation of the layer: performing convolution processing on the input data to obtain a convolution feature map; performing mean pooling based on the local channel dimension on the convolution feature map to obtain a pooling feature map; and performing nonlinear activation on the pooling feature map to obtain an activated feature map; wherein the output of the last layer of the second convolutional neural network is the detection vibration feature matrix, and the input of the first layer of the second convolutional neural network is the waveform diagram of the vibration signal in the predetermined time period.

[0045] In the offshore wind turbine structure state monitoring method based on artificial intelligence, based on the mean and variance of the feature value set of the difference feature matrix, the feature values of each position in the difference feature matrix are corrected to obtain a corrected difference feature matrix, including: based on the mean and variance of the feature value set of the difference feature matrix, the feature values of each position in the difference feature matrix are corrected to obtain the corrected difference feature matrix by the following formula:

[0046] Wherein, the formula is:

[0047]

[0048] Where m i，j represents the feature value of each position in the difference feature matrix, W and H are the width and height of the difference feature matrix respectively, and μ and σ are the mean and variance of the feature value set of the difference feature matrix, and α is a weight hyperparameter.

[0049] In the offshore wind turbine structure state monitoring method based on artificial intelligence, the corrected difference feature matrix is input into a classifier to obtain a classification result, including: the classifier processes the difference feature matrix according to the following formula to generate a classification result, wherein the formula is: softmax{(W n , B n ):…:(W1, B1)|Project(F)}, wherein Project(F) represents projecting the difference feature matrix into a vector, W1 to W n are weight matrices of each layer of full connection layer, and B1 to B n are bias matrices of each layer of full connection layer.

[0050] Compared with the prior art, the offshore wind turbine structure state monitoring system and method based on artificial intelligence provided by the application adopt an artificial intelligence-based monitoring method, take a convolutional neural network model as a feature extractor, generate high-dimensional implicit features of a vibration signal excited by blade operating characteristics from implicit correlation features of blade rotation speed values at multiple predetermined time points, and perform difference comparison with high-dimensional implicit features of the vibration signal actually detected in the predetermined time period, so as to extract an excitation vibration signal caused by blade rotation from the vibration signal of the foundation structure of the offshore wind turbine, so that the difference feature matrix can more directly reflect the vibration mode of the foundation structure of the offshore wind turbine, and then the performance state of the foundation structure of the offshore wind turbine is monitored based on the vibration mode of the foundation structure of the offshore wind turbine. BRIEF DESCRIPTION OF DRAWINGS

[0051] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, but do not limit the present application. In the drawings, the same reference numerals refer to the same components or steps throughout the drawings.

[0052] Figure 1 is an application scenario diagram of the offshore wind turbine structure state monitoring system based on artificial intelligence according to embodiments of the present application.

[0053] Figure 2 is a block diagram of the offshore wind turbine structure state monitoring system based on artificial intelligence according to embodiments of the present application.

[0054] Figure 3 is a flowchart of the offshore wind turbine structure state monitoring method based on artificial intelligence according to embodiments of the present application.

[0055] Figure 4An architecture schematic diagram of a method for offshore wind turbine structural state monitoring based on artificial intelligence according to an embodiment of the present application. DETAILED DESCRIPTION

[0056] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part but not all of the embodiments of the present application, and the present application can be implemented in many different ways. Therefore, the contents described herein should be considered as illustrative rather than limiting the present application.

[0057] Scenario overview

[0058] As mentioned before, by the end of 2013, the total installed capacity of wind power worldwide reached 318 GW, of which offshore wind power was 6.8 GW, and the total installed capacity of wind power in China was 91.4 GW, and the offshore wind power installed capacity was 428 MW. With the passage of time, the safety accidents of wind turbine operation also showed an upward trend. Among various wind power accidents, structural failure is only second to fire and blade failure, so it is of great significance to monitor the state of the wind turbine structure system.

[0059] Compared with onshore, the load environment of offshore wind turbine is more complex, and the influence mechanism of various loads such as wind, wave, current, even ice, typhoon, earthquake and other load excitations in extreme conditions on the structure is more complex. At the same time, since offshore wind turbines are far away from land, wind farm management personnel cannot regularly evaluate and detect the structure, and the response time for accidents is much longer than that of onshore wind turbines.

[0060] Therefore, an optimized state monitoring scheme for offshore wind turbine structure is expected.

[0061] At present, deep learning and neural networks have been widely used in computer vision, natural language processing, text signal processing and other fields. In addition, deep learning and neural networks in image classification, object detection, semantic segmentation, text translation and other fields have also shown a level close to or even beyond human level.

[0062] In recent years, the development of deep learning and neural networks provides a solution for offshore wind turbine structural state monitoring.

[0063] Correspondingly, since the traditional modal identification method is assumed to have or approximate random white noise characteristics of external excitation, and for the fan due to the existence of periodic load mixed by blade rotation, it often cannot meet the assumption requirements of random white noise, so for the natural frequency and the harmonic are more close, especially when the harmonic component in the response is large, these methods cannot distinguish true and false modal, resulting in low identification accuracy. Therefore, in the technical scheme of the present application, it is expected that the vibration signal excited by the blade working characteristics is generated through the implicit correlation characteristics of the blade speed value at multiple predetermined time points, and compared with the vibration signal actually detected in the predetermined time period, so as to extract the excitation vibration signal caused by the blade rotation in the vibration signal of the offshore wind turbine foundation structure, so that the difference feature matrix can more intuitively reflect the vibration mode of the offshore wind turbine foundation structure, and then the performance state monitoring of the offshore wind turbine foundation structure is carried out based on the vibration mode of the offshore wind turbine foundation structure.

[0064] Specifically, in the technical scheme of the present application, first, the vibration sensor deployed in the offshore wind turbine foundation structure acquires the vibration signal in a predetermined time period, and the blade speed value at multiple predetermined time points in the predetermined time period. It can be understood that when the performance state of the offshore wind turbine foundation structure is good, the vibration signal generated by the normal operation of the offshore wind turbine has certain implicit rules, and when the state is abnormal, the vibration signal of the offshore wind turbine foundation structure will change, so the vibration signal of the offshore wind turbine foundation structure is used as data input to represent the state performance characteristics of the offshore wind turbine foundation structure.

[0065] Then, considering that the blade speed value has dynamic correlation rules in the time sequence dimension, therefore, in order to fully extract the dynamic implicit correlation characteristics, the blade speed value at multiple predetermined time points in the predetermined time period is encoded through a time sequence encoder containing a one-dimensional convolution layer, to extract the change characteristics of the blade speed in the time sequence dimension, so as to obtain a speed feature vector. In one example of the present application, the time sequence encoder is composed of alternately arranged fully connected layers and one-dimensional convolution layers, which extracts the correlation of the blade speed value in the time sequence dimension through one-dimensional convolution encoding and extracts the high-dimensional implicit features of the blade speed value through fully connected encoding.

[0066] Further, the rotating speed feature vector is input into a waveform generator to generate a generated vibration signal excited by the working characteristic of the blade, that is, to generate the vibration signal. In particular, the waveform generator includes a generator model and a discriminator neural network model. The generator model of the waveform generator is capable of generating a waveform generated image representing the working characteristic of the blade by using the dynamic features of the rotating speed value of the blade in the time sequence dimension. The discriminator neural network model of the waveform generator is capable of making the feature information extracted from the waveform generated image and a reference waveform image similar by means of “cheating” the discriminator, so as to make the generated vibration signal capable of representing the vibration signal excited by the working characteristic of the blade.

[0067] In this way, the waveform image of the generated vibration signal and the waveform image of the vibration signal in the predetermined time period are respectively input into a convolutional neural network model as a feature extractor to extract local high-dimensional hidden feature distribution information of the waveform image of the generated vibration signal and the waveform image of the vibration signal respectively, so as to obtain a generated vibration feature matrix and a detected vibration feature matrix.

[0068] Then, a difference between the generated vibration feature matrix and the detected vibration feature matrix in the high-dimensional feature space is calculated to obtain a difference feature matrix, that is, the excitation vibration signal caused by the rotation of the blade is extracted from the vibration signal of the offshore wind turbine foundation structure, so that the difference feature matrix can more intuitively reflect the vibration mode of the offshore wind turbine foundation structure, thereby improving the accuracy of subsequent classification.

[0069] It can be understood that, when the rotating speed feature vector is input into the waveform generator to obtain the generated vibration signal and the generated vibration feature matrix is obtained by feature extraction, the generated vibration signal obtained by the waveform generator may have a deviation from the internal distribution of the source data, so that the difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix may have an abnormal instance containing abnormal vibration information features, which affects the classification effect of the difference feature matrix.

[0070] Therefore, in the technical solution of the present application, the difference feature matrix is preferably optimized, specifically as follows:

[0071]

[0072] wherein m i，j is the feature value of each position of the difference feature matrix, W and H are the width and height of the difference feature matrix, and λ and σ are the mean and variance of the feature value set of the difference feature matrix, and α is a weight hyperparameter.

[0073] Herein, the above optimization takes the feature set as an adaptive instance, utilizes the intrinsic prior information of the statistical characteristics to dynamically generate information normalization of a single feature value, and takes the normalized module length information of the feature set as a bias to serve as an invariance description in the set distribution domain, so that the feature optimization that shields the disturbance distribution of abnormal instances as much as possible is realized, the classification effect of the difference feature matrix is improved, and the classification accuracy is further improved.

[0074] Based on this, the application provides an offshore wind turbine structure state monitoring system based on artificial intelligence, which comprises: a data acquisition module, configured to acquire vibration signals in a predetermined time period collected by a vibration sensor deployed on a foundation structure of an offshore wind turbine, and blade rotation speed values at a plurality of predetermined time points in the predetermined time period; a rotation speed data encoding module, configured to obtain a rotation speed feature vector by inputting the blade rotation speed values at the plurality of predetermined time points in the predetermined time period into a time series encoder comprising a one-dimensional convolution layer; a waveform generation module, configured to obtain a generated vibration signal by inputting the rotation speed feature vector into a waveform generator; a generated waveform encoding module, configured to obtain a generated vibration feature matrix by inputting a waveform diagram of the generated vibration signal into a first convolutional neural network serving as a feature extractor; a detected vibration waveform encoding module, configured to obtain a detected vibration feature matrix by inputting a waveform diagram of the vibration signals in the predetermined time period into a second convolutional neural network serving as a feature extractor; a difference module, configured to calculate a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix; a feature correction module, configured to correct feature values at each position in the difference feature matrix based on the mean and variance of a feature value set of the difference feature matrix to obtain a corrected difference feature matrix; and a state monitoring result generation module, configured to obtain a classification result by inputting the corrected difference feature matrix into a classifier, wherein the classification result is used to indicate whether the performance of the foundation structure of the offshore wind turbine meets a predetermined requirement.

[0075] Figure 1 The application scenario diagram of the offshore wind turbine structure state monitoring system based on artificial intelligence is shown in FIG. 1. Figure 1 As shown in FIG. 1, in the application scenario, first, a vibration sensor (for example, V shown in FIG. 2) deployed on the foundation structure of the offshore wind turbine (for example, F shown in FIG. 1) collects vibration signals in a predetermined time period, and a rotation speed sensor (for example, T shown in FIG. 2) collects blade rotation speed values at a plurality of predetermined time points in the predetermined time period. Figure 1 Figure 1 Figure 1 Figure 1 ​​​the B) rotational speed value as shown in the cloud server S) to generate a classification result indicating whether the performance of the foundation structure of the offshore wind turbine meets the predetermined requirements. Figure 1

[0076] After introducing the basic principles of the present application, the various non-limiting embodiments of the present application will be specifically introduced below with reference to the accompanying drawings.

[0077] Exemplary system

[0078] Figure 2 The block diagram of the offshore wind turbine structure state monitoring system based on artificial intelligence according to the embodiments of the present application is illustrated. As shown in the figure, Figure 2 The offshore wind turbine structure state monitoring system 200 based on artificial intelligence according to the embodiments of the present application includes: a data acquisition module 210 for acquiring vibration signals of a predetermined time period collected by vibration sensors deployed on the foundation structure of the offshore wind turbine, and blade rotational speed values at a plurality of predetermined time points within the predetermined time period; a rotational speed data encoding module 220 for obtaining rotational speed feature vectors by passing the blade rotational speed values at a plurality of predetermined time points within the predetermined time period through a time series encoder containing a one-dimensional convolution layer; a waveform generation module 230 for obtaining generated vibration signals by passing the rotational speed feature vectors through a waveform generator; a generated waveform encoding module 240 for obtaining a generated vibration feature matrix by passing the waveform diagram of the generated vibration signals through a first convolutional neural network as a feature extractor; a detected vibration waveform encoding module 250 for obtaining a detected vibration feature matrix by passing the waveform diagram of the vibration signals of the predetermined time period through a second convolutional neural network as a feature extractor; a difference module 260 for calculating a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix; a feature correction module 270 for correcting the feature values at each position in the difference feature matrix based on the mean and variance of the feature value set of the difference feature matrix to obtain a corrected difference feature matrix; and a state monitoring result generation module 280 for obtaining a classification result by passing the corrected difference feature matrix through a classifier, wherein the classification result is used to indicate whether the performance of the foundation structure of the offshore wind turbine meets the predetermined requirements.

[0079] ​Specifically, in the embodiments of the present application, the data acquisition module 210 and the rotating speed data encoding module 220 are configured to acquire vibration signals collected by vibration sensors deployed on the foundation structure of the offshore wind turbine within a predetermined time period and blade rotating speed values at a plurality of predetermined time points within the predetermined time period, and encode the blade rotating speed values at the plurality of predetermined time points within the predetermined time period through a time series encoder comprising a one-dimensional convolution layer to obtain a rotating speed feature vector. As described above, it can be understood that, since the traditional modal identification method is based on the assumption that the external excitation has or approximately has the characteristic of random white noise, and the periodic load is mixed into the wind turbine due to the rotation of the blade, the assumption of random white noise cannot be met, so when the natural frequency is close to the harmonic, especially when the harmonic component accounts for a large proportion in the response, the method cannot distinguish between true and false modalities, resulting in low identification accuracy. Therefore, in the technical solution of the present application, it is expected to generate vibration signals excited by the blade operating characteristics through the implicit correlation features of the blade rotating speed values at a plurality of predetermined time points, and to compare the difference between the vibration signals actually detected within the predetermined time period, so as to extract the excitation vibration signals caused by the rotation of the blade from the vibration signals of the foundation structure of the offshore wind turbine, so that the difference feature matrix can more intuitively reflect the vibration mode of the foundation structure of the offshore wind turbine, and then the performance state monitoring of the foundation structure of the offshore wind turbine is performed based on the vibration mode of the foundation structure of the offshore wind turbine.

[0080] That is, specifically, in the technical solution of the present application, first, the vibration signals collected by the vibration sensors deployed on the foundation structure of the offshore wind turbine within a predetermined time period and the blade rotating speed values at a plurality of predetermined time points within the predetermined time period are acquired. It can be understood that, when the performance state of the foundation structure of the offshore wind turbine is good, the vibration signals generated by the normal operation of the offshore wind turbine have certain implicit rules, and when the state is abnormal, the vibration signals of the foundation structure of the offshore wind turbine will change, so the vibration signals of the foundation structure of the offshore wind turbine are used as data input to represent the state performance characteristics of the foundation structure of the offshore wind turbine.

[0081] Then, it can be understood that, considering that the blade rotating speed values have dynamic correlation rules in the time series dimension, in the technical solution of the present application, in order to fully extract such dynamic implicit correlation features, the blade rotating speed values at a plurality of predetermined time points within the predetermined time period are encoded through a time series encoder comprising a one-dimensional convolution layer to extract the change features of the blade rotating speed values in the time series dimension, thereby obtaining a rotating speed feature vector. In one example of the present application, the time series encoder is composed of fully connected layers and one-dimensional convolution layers arranged alternately, which extracts the correlation of the blade rotating speed values in the time series dimension through one-dimensional convolution encoding and extracts the high-dimensional implicit features of the blade rotating speed values through fully connected encoding.

[0082] More specifically, in the embodiments of the present application, the rotating speed data encoding module comprises: an input vector construction unit, configured to arrange the blade rotating speed values at a plurality of predetermined time points in a predetermined time period into an input vector in time dimension; a fully connected encoding unit, configured to use the fully connected layer of the time series encoder to perform fully connected encoding on the input vector according to the following formula to extract high-dimensional hidden features of the feature values at each position in the input vector, wherein the formula is: wherein X is the input vector, Y is the output vector, W is the weight matrix, B is the bias vector, represents matrix multiplication; a one-dimensional convolutional encoding unit, configured to use the one-dimensional convolutional layer of the time series encoder to perform one-dimensional convolutional encoding on the input vector according to the following formula to extract high-dimensional hidden correlation features between the feature values at each position in the input vector, wherein the formula is:

[0083]

[0084] wherein a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix operated with the convolution kernel function, w is the size of the convolution kernel, and X represents the input vector.

[0085] Specifically, in the embodiments of the present application, the waveform generation module 230 is configured to pass the rotating speed feature vector through a waveform generator to obtain a generated vibration signal. That is, in the technical solution of the present application, in order to generate a vibration signal excited by the working characteristics of the blade through the hidden correlation features of the blade rotating speed values at a plurality of predetermined time points, the rotating speed feature vector is further passed through a waveform generator to generate the generated vibration signal. In particular, in one specific example, here, the waveform generator is a generative adversarial network, and the waveform generator comprises a generator model and a discriminator neural network model. The generator model of the waveform generator can use the dynamic features of the blade rotating speed values in the time series dimension to generate a waveform generated image representing the working characteristics of the blade; and the discriminator neural network model of the waveform generator can make the feature information extracted from the waveform generated image and the reference waveform image similar by "cheating" the discriminator, so that the generated vibration signal can represent a vibration signal excited by the working characteristics of the blade.

[0086] More specifically, in the embodiments of the present application, the training process of the waveform generator comprises: obtaining training data, the training data comprising training blade speed values at a plurality of predetermined time points in a predetermined time period and a reference waveform image of the predetermined time period; passing the training blade speed values at the plurality of predetermined time points in the predetermined time period in the training data through the time series encoder comprising a one-dimensional convolution layer to obtain a training speed feature vector; passing the training speed feature vector through the generator model of the waveform generator to obtain a training waveform generated image; passing the training waveform generated image through the discriminator neural network model of the waveform generator to obtain a first feature map; passing the reference waveform image through the discriminator neural network model of the waveform generator to obtain a second feature map; determining whether the values at a predetermined position in the first feature map and the second feature map are the same; in response to the values at the predetermined position in the first feature map and the second feature map being the same, calculating a negative value of a base-2 logarithm value of the value at the predetermined position as a first value; in response to the values at the predetermined position in the first feature map and the second feature map being different, calculating a base-2 logarithm value of the value at the predetermined position as a second value; calculating a sum of an average value of the first values at the positions with the same values and an average value of the second values at the positions with different values as the discriminator loss function value; and training the waveform generator with the discriminator loss function value and through the back propagation of gradient descent.

[0087] Specifically, in the embodiments of the present application, the generated waveform encoding module 240 and the detected vibration waveform encoding module 250 are configured to pass the waveform graph of the generated vibration signal through a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix, and pass the waveform graph of the vibration signal in the predetermined time period through a second convolutional neural network as a feature extractor to obtain a detected vibration feature matrix. That is, in the technical solution of the present application, the waveform graph of the generated vibration signal and the waveform graph of the vibration signal in the predetermined time period are further extracted through the convolutional neural network model as a feature extractor, respectively, to extract the local high-dimensional implicit feature distribution information of the waveform graph of the generated vibration signal and the waveform graph of the vibration signal, respectively, so as to obtain the generated vibration feature matrix and the detected vibration feature matrix.

[0088] More specifically, in the embodiments of the present application, the waveform encoding module is further configured to: each layer of the first convolutional neural network respectively performs, in forward propagation of the layer, convolution processing on input data to obtain a convolution feature map, mean pooling based on a local channel dimension on the convolution feature map to obtain a pooling feature map, and non-linear activation on the pooling feature map to obtain an activated feature map; and an output of a last layer of the first convolutional neural network is the generated vibration feature matrix, and an input of a first layer of the first convolutional neural network is a waveform graph of the generated vibration signal.

[0089] More specifically, in the embodiments of the present application, the vibration waveform encoding module is further configured to: each layer of the second convolutional neural network respectively performs, in forward propagation of the layer, convolution processing on input data to obtain a convolution feature map, mean pooling based on a local channel dimension on the convolution feature map to obtain a pooling feature map, and non-linear activation on the pooling feature map to obtain an activated feature map; and an output of a last layer of the second convolutional neural network is the detected vibration feature matrix, and an input of a first layer of the second convolutional neural network is a waveform graph of the vibration signal in the predetermined time period.

[0090] Specifically, in the embodiments of the present application, the difference module 260 is configured to calculate a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix. That is, in the technical solution of the present application, further, a difference between the generated vibration feature matrix and the detected vibration feature matrix in a high-dimensional feature space is calculated to obtain a difference feature matrix, that is, an excitation vibration signal caused by rotation of a blade is extracted from the vibration signal of the offshore wind turbine foundation structure, so that the difference feature matrix can more intuitively reflect a vibration mode of the offshore wind turbine foundation structure, thereby improving accuracy of subsequent classification.

[0091] Specifically, in the embodiments of the present application, the feature correction module 270 is configured to correct the feature values of each position in the difference feature matrix based on the mean and variance of the feature value set of the difference feature matrix to obtain a corrected difference feature matrix. It should be understood that when the rotation speed feature vector is input into the waveform generator to obtain the generated vibration signal and the feature extraction is performed to obtain the generated vibration feature matrix, the generated vibration signal obtained by the waveform generator may deviate from the inherent distribution of the source data, so that the difference feature matrix of the detection vibration feature matrix and the generated vibration feature matrix may contain abnormal instances (instance) containing abnormal vibration information features, which affects the classification effect of the difference feature matrix. Therefore, in the technical solution of the present application, the difference feature matrix is preferably optimized. Here, the above optimization takes the feature set as an adaptive instance (adaptive instance), uses the intrinsic prior information of the statistical features to dynamically generate information normalization of the single feature value, and takes the normalized length information of the feature set as a bias to describe the invariance in the set distribution domain. In this way, the feature optimization that shields the disturbance distribution of the abnormal instances as much as possible is realized, the classification effect of the difference feature matrix is improved, and the classification accuracy is further improved.

[0092] More specifically, in the embodiments of the present application, the feature correction module is further configured to correct the feature values of each position in the difference feature matrix based on the mean and variance of the feature value set of the difference feature matrix to obtain the corrected difference feature matrix according to the following formula:

[0093] The formula is:

[0094]

[0095] wherein m i，j represents the feature values of each position in the difference feature matrix, W and H are the width and height of the difference feature matrix respectively, μ and σ are the mean and variance of the feature value set of the difference feature matrix, and α is a weight hyperparameter.

[0096] Specifically, in the embodiments of the present application, the state monitoring result generation module 280 is configured to pass the corrected differential feature matrix through a classifier to obtain a classification result, which is used to indicate whether the performance of the offshore wind turbine foundation structure meets the predetermined requirements. That is, in the technical solution of the present application, the corrected differential feature matrix is further passed through a classifier to obtain a classification result used to indicate whether the performance of the offshore wind turbine foundation structure meets the predetermined requirements. Accordingly, in one specific example, the classifier processes the differential feature matrix to generate a classification result according to the following formula: softmax{(W n , B n ):…:(W1, B1)|Project(F)}, where Project(F) represents projecting the differential feature matrix into a vector, W1 to W n are weight matrices of each layer of fully connected layers, and B1 to B n are bias matrices of each layer of fully connected layers.

[0097] In summary, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence according to the embodiments of the present application is illustrated, which adopts an artificial intelligence-based monitoring method, uses a convolutional neural network model as a feature extractor, generates high-dimensional implicit features of the vibration signal excited by the blade operating characteristics through the implicit correlation features of the blade speed values at multiple predetermined time points, and differentially compares the high-dimensional implicit features of the vibration signal actually detected in the predetermined time period, so as to extract the excitation vibration signal caused by the blade rotation from the vibration signal of the offshore wind turbine foundation structure, so that the differential feature matrix can more directly reflect the vibration mode of the offshore wind turbine foundation structure, and then the performance state monitoring of the offshore wind turbine foundation structure is performed based on the vibration mode of the offshore wind turbine foundation structure.

[0098] As described above, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence according to the embodiments of the present application can be implemented in various terminal devices, such as a server of an offshore wind turbine structure state monitoring algorithm based on artificial intelligence. In one example, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence according to the embodiments of the present application can be integrated into a terminal device as a software module and / or a hardware module. For example, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence can also be one of the many hardware modules of the terminal device.

[0099] Alternatively, in another example, the offshore wind turbine structure state monitoring system 200 based on artificial intelligence and the terminal device can also be separate devices, and the offshore wind turbine structure state monitoring system 200 based on artificial intelligence can be connected to the terminal device through a wired and / or wireless network, and transmit interactive information in a predetermined data format.

[0100] Exemplary method

[0101] Figure 3 A flowchart of the offshore wind turbine structure state monitoring method based on artificial intelligence is shown. As shown in Figure 3 The offshore wind turbine structure state monitoring method based on artificial intelligence according to the embodiments of the present application includes the following steps: S110, obtaining vibration signals collected by a vibration sensor deployed on the foundation structure of an offshore wind turbine within a predetermined time period, and blade rotation speed values at a plurality of predetermined time points within the predetermined time period; S120, passing the blade rotation speed values at the plurality of predetermined time points within the predetermined time period through a time series encoder containing a one-dimensional convolution layer to obtain a rotation speed feature vector; S130, passing the rotation speed feature vector through a waveform generator to obtain a generated vibration signal; S140, passing a waveform graph of the generated vibration signal through a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix; S150, passing a waveform graph of the vibration signal within the predetermined time period through a second convolutional neural network as a feature extractor to obtain a detected vibration feature matrix; S160, calculating a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix; S170, based on the mean and variance of the feature value set of the difference feature matrix, correcting the feature values at each position in the difference feature matrix to obtain a corrected difference feature matrix; and S180, passing the corrected difference feature matrix through a classifier to obtain a classification result, which is used to indicate whether the performance of the foundation structure of the offshore wind turbine meets the predetermined requirements.

[0102] Figure 4 A schematic diagram of the offshore wind turbine structure state monitoring method based on artificial intelligence according to the embodiments of the present application is shown. As shown in Figure 4 In the network architecture of the offshore wind turbine structure state monitoring method based on artificial intelligence, first, the obtained blade rotation speed values at a plurality of predetermined time points within the predetermined time period (for example, P1 as shown in Figure 4 ) are passed through a time series encoder containing a one-dimensional convolution layer (for example, E as shown in Figure 4 ) to obtain a rotation speed feature vector (for example, VF1 as shown in Figure 4 ); then, the rotation speed feature vector is passed through a waveform generator (for example, WG as shown in Figure 4 ) to obtain a generated vibration signal (for example, Vg as shown inFigure 4 P2) shown in FIG. 1; then, the waveform graph of the generated vibration signal is passed through a first convolutional neural network (e.g., CNN1) shown in FIG. 1 as a feature extractor to obtain a generated vibration feature matrix (e.g., MF1) shown in FIG. 1. Figure 4 Figure 4 Figure 4 Q) shown in FIG. 1 is passed through a second convolutional neural network (e.g., CNN2) shown in FIG. 1 as a feature extractor to obtain a detected vibration feature matrix (e.g., MF2) shown in FIG. 1. Figure 4 Figure 4 Figure 4 Figure 4 Figure 4

[0103] More specifically, in steps S110 and S120, a vibration signal of a predetermined time period collected by a vibration sensor deployed on the foundation structure of the offshore wind turbine is obtained, and a blade rotating speed value at a plurality of predetermined time points in the predetermined time period is obtained, and the blade rotating speed value at the plurality of predetermined time points in the predetermined time period is passed through a time series encoder containing a one-dimensional convolutional layer to obtain a rotating speed feature vector. It can be understood that since the conventional modal identification method is based on the assumption that the external excitation has or is approximately random white noise characteristics, and for the wind turbine, due to the periodic load mixed by the blade rotation, it often cannot meet the assumption requirement of random white noise, so for the natural frequency close to the harmonic, especially when the harmonic component accounts for a large proportion in the response, these methods cannot distinguish between true and false modalities, resulting in low identification accuracy. Therefore, in the technical solution of the present application, it is expected to generate a vibration signal excited by the blade working characteristics through the implicit correlation features of the blade rotating speed values at a plurality of predetermined time points, and compare it with the actually detected vibration signal in the predetermined time period, so as to extract the excitation vibration signal caused by the blade rotation from the vibration signal of the offshore wind turbine foundation structure, so that the difference feature matrix can more intuitively reflect the vibration mode of the offshore wind turbine foundation structure, and then the performance state monitoring of the offshore wind turbine foundation structure is performed based on the vibration mode of the offshore wind turbine foundation structure. ​​​​​​​

[0104] That is, specifically, in the technical solution of the present application, first, the vibration signals of a predetermined time period and the blade speed values of a plurality of predetermined time points in the predetermined time period are collected by the vibration sensor deployed on the foundation structure of the offshore wind turbine. It should be understood that, since the vibration signals generated by the normal operation of the offshore wind turbine have certain implicit regularities when the performance state of the foundation structure of the offshore wind turbine is good, and the vibration signals of the foundation structure of the offshore wind turbine will change when the state is abnormal, therefore, using the vibration signals of the foundation structure of the offshore wind turbine as data input can represent the state performance characteristics of the foundation structure of the offshore wind turbine.

[0105] Then, it should be understood that, considering that the blade speed values have dynamic correlation in the time sequence dimension, therefore, in the technical solution of the present application, in order to fully extract the dynamic implicit correlation characteristics, the blade speed values of a plurality of predetermined time points in the predetermined time period are encoded by a time sequence encoder containing a one-dimensional convolution layer to extract the change characteristics of the blade speed in the time sequence dimension, thereby obtaining a speed feature vector. In one example of the present application, the time sequence encoder is composed of fully connected layers and one-dimensional convolution layers arranged alternately, which extracts the correlation of the blade speed values in the time sequence dimension by one-dimensional convolution encoding and extracts the high-dimensional implicit features of the blade speed values by fully connected encoding.

[0106] More specifically, in step S130, the speed feature vector is input into a waveform generator to obtain a generated vibration signal. That is, in the technical solution of the present application, in order to generate a vibration signal excited by the working characteristics of the blade by the implicit correlation characteristics of the blade speed values of a plurality of predetermined time points, further, the speed feature vector is input into the waveform generator to generate the generated vibration signal. In particular, in one specific example, here, the waveform generator is a generative adversarial network, and the waveform generator includes a generator model and a discriminator neural network model, the generator model of the waveform generator can utilize the dynamic characteristics of the blade speed values in the time sequence dimension to generate a waveform generated image representing the working characteristics of the blade; and the discriminator neural network model of the waveform generator can make the feature information extracted from the waveform generated image and the reference waveform image similar by the method of "cheating" the discriminator, so that the generated vibration signal can represent the vibration signal excited by the working characteristics of the blade.

[0107] More specifically, in step S140 and step S150, the waveform diagram of the generated vibration signal is passed through a first convolutional neural network as a feature extractor to obtain a generated vibration feature matrix, and the waveform diagram of the vibration signal of the predetermined time period is passed through a second convolutional neural network as a feature extractor to obtain a detected vibration feature matrix. That is, in the technical solution of the present application, the waveform diagram of the generated vibration signal and the waveform diagram of the vibration signal of the predetermined time period are further subjected to feature extraction by the convolutional neural network model as a feature extractor, respectively, to extract the local high-dimensional implicit feature distribution information of the waveform diagram of the generated vibration signal and the waveform diagram of the vibration signal, respectively, so as to obtain the generated vibration feature matrix and the detected vibration feature matrix.

[0108] More specifically, in step S160, a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix is calculated. That is, in the technical solution of the present application, the difference between the generated vibration feature matrix and the detected vibration feature matrix in the high-dimensional feature space is further calculated to obtain a difference feature matrix, that is, the excitation vibration signal caused by the rotation of the blade is extracted from the vibration signal of the offshore wind turbine foundation structure, so that the difference feature matrix can more intuitively reflect the vibration mode of the offshore wind turbine foundation structure, thereby improving the accuracy of subsequent classification.

[0109] More specifically, in step S170, based on the mean and variance of the feature value set of the difference feature matrix, the feature values at each position in the difference feature matrix are corrected to obtain a corrected difference feature matrix. It should be understood that when the rotation speed feature vector is passed through the waveform generator to obtain the generated vibration signal and the feature extraction obtains the generated vibration feature matrix, the generated vibration signal obtained by the waveform generator may deviate from the inherent distribution of the source data, so that the difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix may contain abnormal vibration information features, affecting the classification effect of the difference feature matrix. Therefore, in the technical solution of the present application, the difference feature matrix is preferably optimized. Here, the above optimization takes the feature set as an adaptive instance, uses the essential intrinsic prior information of its statistical characteristics to dynamically generate information normalization of individual feature values, and takes the normalized length information of the feature set as a bias as an invariance description in the set distribution domain. In this way, the feature optimization that shields the disturbance distribution of abnormal instances as much as possible is realized, the classification effect of the difference feature matrix is improved, and the accuracy of classification is further improved.

[0110] More specifically, in step S180, the corrected differential feature matrix is passed through a classifier to obtain a classification result, which is used to indicate whether the performance of the offshore wind turbine foundation structure meets predetermined requirements. That is, in the technical solution of the present application, the corrected differential feature matrix is further passed through a classifier to obtain a classification result used to indicate whether the performance of the offshore wind turbine foundation structure meets predetermined requirements. Accordingly, in one specific example, the classifier processes the differential feature matrix in the following formula to generate a classification result, where the formula is: softmax{(W n , B n ):…:(W1, B1)|Project(F)}, where Project(F) represents projecting the differential feature matrix into a vector, W1 to W n are weight matrices of each layer of fully connected layers, and B1 to B n represent bias matrices of each layer of fully connected layers.

[0111] In summary, the offshore wind turbine structure state monitoring method based on artificial intelligence according to the embodiments of the present application is illustrated, which adopts an artificial intelligence-based monitoring method, takes a convolutional neural network model as a feature extractor, generates high-dimensional implicit features of the vibration signal of the blade working characteristics excited by the blade rotation through the implicit correlation features of the blade rotation speed values at multiple predetermined time points, and performs differential comparison with the actually detected high-dimensional implicit features of the vibration signal in the predetermined time period, so as to extract the excited vibration signal caused by the blade rotation from the vibration signal of the offshore wind turbine foundation structure, so that the differential feature matrix can more directly reflect the vibration mode of the offshore wind turbine foundation structure, and then the performance state monitoring of the offshore wind turbine foundation structure is performed based on the vibration mode of the offshore wind turbine foundation structure.

[0112] The basic principles of the present application are described above in combination with specific embodiments, but it should be pointed out that the advantages, advantages, effects and the like mentioned in the present application are only examples and not limitations, and these advantages, advantages, effects and the like cannot be considered as the must-have of each embodiment of the present application. In addition, the above-mentioned specific details are only for the purpose of example and for the purpose of understanding, and are not limited to the present application, and the above-mentioned details do not limit the present application to the must-use of the above-mentioned specific details.

[0113] 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.

[0114] 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.

[0115] 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.

[0116] 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. An offshore wind turbine structure condition monitoring system based on artificial intelligence, characterized by, The method comprises: a data acquisition module configured to acquire vibration signals collected by a vibration sensor deployed on a foundation structure of an offshore wind turbine within a predetermined time period, and blade rotation speed values at a plurality of predetermined time points within the predetermined time period; a rotation speed data encoding module configured to encode the blade rotation speed values at the plurality of predetermined time points within the predetermined time period through a time series encoder comprising a one-dimensional convolution layer to obtain a rotation speed feature vector; a waveform generation module configured to generate a generated vibration signal by inputting the rotation speed feature vector into a waveform generator; a generated waveform encoding module configured to encode a waveform image of the generated vibration signal through a first convolutional neural network serving as a feature extractor to obtain a generated vibration feature matrix; a detected vibration waveform encoding module configured to encode a waveform image of the vibration signals within the predetermined time period through a second convolutional neural network serving as a feature extractor to obtain a detected vibration feature matrix; a difference module configured to calculate a difference feature matrix between the detected vibration feature matrix and the generated vibration feature matrix; a feature correction module configured to correct feature values at each position in the difference feature matrix based on a mean and a variance of a set of feature values of the difference feature matrix to obtain a corrected difference feature matrix; and a state monitoring result generation module configured to input the corrected difference feature matrix into a classifier to obtain a classification result, wherein the classification result is used to indicate whether the performance of the foundation structure of the offshore wind turbine meets a predetermined requirement. The waveform generator is an adversarial generator. The training process of the waveform generator comprises: obtaining training data, wherein the training data comprises training blade rotation speed values at a plurality of predetermined time points within a predetermined time period and reference waveform images within the predetermined time period; encoding the training blade rotation speed values at the plurality of predetermined time points within the predetermined time period in the training data through the time series encoder comprising the one-dimensional convolution layer to obtain training rotation speed feature vectors; inputting the training rotation speed feature vectors into a generator model of the waveform generator to obtain training waveform generated images; inputting the training waveform generated images into a discriminator neural network model of the waveform generator to obtain first feature maps; inputting the reference waveform images into the discriminator neural network model of the waveform generator to obtain second feature maps; determining whether values at predetermined positions in the first feature maps and the second feature maps are the same; in response to the values at the predetermined positions in the first feature maps and the second feature maps being the same, calculating a negative value of a base-2 logarithm value of the values at the predetermined positions as a first value; in response to the values at the predetermined positions in the first feature maps and the second feature maps being different, calculating a base-2 logarithm value of the values at the predetermined positions as a second value; calculating a sum of an average of the first values at the positions with the same values and an average of the second values at the positions with different values as a discriminator loss function value; and training the waveform generator by using the discriminator loss function value and through back propagation of gradient descent. The rotation speed data encoding module comprises:

2. The offshore wind turbine structure condition monitoring system based on artificial intelligence according to claim 1, characterized in that, ​ an input vector construction unit, configured to arrange the blade speed values at a plurality of predetermined time points in a predetermined time period into an input vector in a time dimension; a fully connected encoding unit configured to use a fully connected layer of the time series encoder to fully connect encode the input vector to extract high-dimensional hidden features of feature values at each position in the input vector using a formula as follows: where X is the input vector, Y is the output vector, W is a weight matrix, B is a bias vector, denotes a matrix multiplication; a one-dimensional convolution coding unit, configured to use a one-dimensional convolution layer of the time sequence encoder to perform one-dimensional convolution coding on the input vector according to the following formula to extract high-dimensional hidden correlation features between the feature values at each position in the input vector, wherein the formula is: wherein a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix for the convolution kernel function operation, w is the size of the convolution kernel, and X represents the input vector.

3. The offshore wind turbine structure condition monitoring system based on artificial intelligence according to claim 2, characterized in that, The waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; 4. The offshore wind turbine structure condition monitoring system based on artificial intelligence according to claim 3, characterized in that, perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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.

5. The offshore wind turbine structure condition monitoring system based on artificial intelligence according to claim 4, characterized in that, The waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; where m i,j denotes the feature value at each position of the difference feature matrix, W and H are the width and height of the difference feature matrix, respectively, and μ and σ are the mean and variance of the set of feature values of the difference feature matrix, and a is a weight hyperparameter.

6. The offshore wind turbine structure condition monitoring system based on artificial intelligence according to claim 5, characterized in that, The state monitoring result generation module is further configured to: process the difference feature matrix using the classifier according to 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 difference feature matrix 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.

7. An offshore wind turbine structure state monitoring method based on artificial intelligence using the offshore wind turbine structure state monitoring system based on artificial intelligence according to claim 1, characterized by, perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension 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 waveform generation coding module is further configured to perform the following operations in the forward transmission of each layer of the first convolutional neural network: perform convolution processing on the input data to obtain a convolution feature map; perform mean pooling based on the local channel dimension on the convolution feature map to obtain a pooling feature map; The corrected differential feature matrix is input into a classifier to obtain a classification result, which indicates whether the performance of the offshore wind turbine foundation structure meets predetermined requirements.

8. The offshore wind turbine structure condition monitoring method based on artificial intelligence according to claim 7, characterized in that, The blade rotating speed values at the plurality of predetermined time points in the predetermined time period are arranged into a one-dimensional input vector according to a time dimension; The blade rotating speed values at the plurality of predetermined time points in the predetermined time period are arranged into a one-dimensional input vector according to a time dimension; use the full connection layer of the time sequence encoder to full connection encode the input vector with the following formula to extract high-dimensional hidden features of the feature values of each position in the input vector, wherein the formula is: wherein X is the input vector, Y is the output vector, W is the weight matrix, B is the bias vector, represents matrix multiplication; The one-dimensional convolution layer of the time sequence encoder is used to perform one-dimensional convolution coding on the input vector according to the following formula to extract high-dimensional hidden correlation features between feature values at different positions in the input vector, wherein the formula is as follows: wherein a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix for the convolution kernel function operation, w is the size of the convolution kernel, and X represents the input vector.

Images