Offshore wind turbine blade hoisting intelligent control system and method

Abstract

The present application relates to the field of intelligent lifting of offshore wind turbine blades, and specifically discloses an intelligent control system and method for lifting offshore wind turbine blades, which adopts a convolutional neural network model based on deep learning as a feature extractor, and dynamically adjusts the tension of two wind ropes in real time through the wind force change characteristics at various times and the center of gravity change characteristics of the wind turbine blades. In this process, the control force of the wind rope of the wind turbine is intelligently adjusted by considering the wind force and the control of the winch as a whole and taking the change of the center of gravity data as a reference, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation.

Description

Technical Field

[0001] The present invention relates to the field of intelligent hoisting of offshore wind turbine blades, and more particularly to an intelligent control system and method for hoisting offshore wind turbine blades. Background Art

[0002] The development and utilization of wind energy resources is a crucial, environmentally friendly energy development strategy for adjusting my country's power structure and has garnered widespread attention in recent years. To increase wind power production, wind farms are often located in areas with high wind speeds. However, due to their complex geometry, low stiffness, and low mass, blades can only be installed at wind speeds below 12 m / s. This requires expensive lifting equipment (installation vessels and cranes) and increases crew waiting time, leading to increased installation costs.

[0003] Therefore, during the construction of offshore wind farms, how to reduce the wind speed restrictions imposed by blade installation and improve installation efficiency is a common concern for researchers and engineers. Therefore, an intelligent control solution for offshore wind turbine blade installation is desired to ensure accurate and safe blade installation while improving installation efficiency. Summary of the Invention

[0004] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an intelligent control system and method for offshore wind turbine blade hoisting, which uses a convolutional neural network model based on deep learning as a feature extractor, and dynamically adjusts the tension of two cables in real time through the wind force change characteristics at different times and the center of gravity change characteristics of the wind turbine blades. In this process, the wind force and the control of the winch are regarded as a whole, and the change of the center of gravity data is used as a reference to intelligently adjust the control force of the cable of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation.

[0005] According to one aspect of the present application, an intelligent control system for offshore wind turbine blade installation is provided, comprising:

[0006] a hoisting data acquisition module, configured to acquire blade hoisting monitoring video within a predetermined time period, wind force data at a plurality of predetermined time points within the predetermined time period, and tension values of the first cable wind rope and the second cable wind rope at a plurality of predetermined time points within the predetermined time period;

[0007] a cable guy rope data structuring module, configured to respectively arrange the tension values of the first cable guy rope and the tension values of the second cable guy rope at a plurality of predetermined time points within the predetermined time period into a first tension force input vector and a second tension force input vector, and calculate the product between the transposed vector of the first tension force input vector and the second tension force input vector to obtain a tension force correlation input matrix;

[0008] a cable wind rope data encoding module, configured to pass the tension force correlation input matrix through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain a tension force correlation feature matrix;

[0009] A key frame extraction module, configured to extract a plurality of key frames from the blade hoisting monitoring video;

[0010] A center of gravity data extraction module is used to extract the center of gravity data of the hoisted blade from each key frame of the plurality of key frames to obtain a center of gravity data input vector;

[0011] A first multi-scale neighborhood feature extraction module, configured to pass the center of gravity data input vector through the multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector;

[0012] a second multi-scale neighborhood feature extraction module, configured to pass the wind data at a plurality of predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector;

[0013] a correction module, configured to correct the eigenvalues of each position in the tension-related characteristic matrix based on a transposed matrix of the tension-related characteristic matrix to obtain a corrected tension-related characteristic matrix;

[0014] a response control module, configured to multiply the corrected tension force correlation characteristic matrix by the wind force time series characteristic vector to obtain a response control characteristic vector;

[0015] a responsiveness estimation module, configured to calculate a responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector to obtain a classification feature matrix; and

[0016] The hoisting control result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the control force of the first cable guy rope should be increased or decreased, and whether the control force of the second cable guy rope should be increased or decreased at the current time point.

[0017] In the above-mentioned offshore wind turbine blade hoisting intelligent control system, the cable wind rope data encoding module includes: a shallow feature matrix extraction unit, used to extract the shallow feature matrix from the Mth layer of the first convolutional neural network, where M is an even number; a deep feature matrix extraction unit, used to extract the deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number and N is greater than 2 times of M; and a feature matrix fusion unit, used to fuse the shallow feature matrix and the deep feature matrix to generate the tension correlation feature matrix.

[0018] In the above-mentioned intelligent control system for offshore wind turbine blade installation, the key frame extraction module is further used to extract the multiple key frames from the blade installation monitoring video at a predetermined sampling frequency.

[0019] In the above-mentioned offshore wind turbine blade hoisting intelligent control system, the first multi-scale neighborhood feature extraction module includes: a first-scale center of gravity encoding unit, used to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the center of gravity data input vector with a first one-dimensional convolution kernel to obtain a first-scale center of gravity feature vector; a second-scale center of gravity encoding unit, used to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the center of gravity data input vector with a second one-dimensional convolution kernel to obtain a second-scale center of gravity feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and a center of gravity scale fusion unit, used to cascade the first-scale center of gravity feature vector and the second-scale center of gravity feature vector to obtain the center of gravity change feature vector.

[0020] In the above-mentioned intelligent control system for offshore wind turbine blade hoisting, the second multi-scale neighborhood feature extraction module includes: a wind input vector construction unit, which is used to arrange the wind data of multiple predetermined time points within the predetermined time period into an input vector to obtain a wind time series input vector; a first-scale wind encoding unit, which is used to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a first one-dimensional convolution kernel to obtain a first-scale wind time series feature vector; a second-scale wind encoding unit, which is used to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a second one-dimensional convolution kernel to obtain a second-scale wind time series feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and a wind scale fusion unit, which is used to cascade the first-scale wind time series feature vector and the second-scale wind time series feature vector to obtain the wind time series feature vector.

[0021] In the above-mentioned offshore wind turbine blade hoisting intelligent control system, the correction module is further used to: based on the transposed matrix of the tension force correlation characteristic matrix, correct the eigenvalues of each position in the tension force correlation characteristic matrix using the following formula to obtain the corrected tension force correlation characteristic matrix;

[0022] Wherein, the formula is:

[0023]

[0024] Where M represents the tension-related feature matrix, Cov1() and Cov2() are both single convolutional layers, Used to map two-dimensional position coordinates to one-dimensional values, P M Represents the (x, y) coordinate matrix of matrix M, and ⊙ represents the point multiplication by position.

[0025] In the above-mentioned offshore wind turbine blade hoisting intelligent control system, the responsiveness estimation module is further used to: calculate the responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector using the following formula to obtain the classification feature matrix;

[0026] Wherein, the formula is:

[0027] V1=M*V2

[0028] Wherein, V1 represents the response control feature vector, V2 represents the center of gravity change feature vector, and M represents the classification feature matrix.

[0029] In the above-mentioned offshore wind turbine blade hoisting intelligent control system, the hoisting control result generation module is further used for: the classifier processes the classification 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)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0030] According to another aspect of the present application, an intelligent control method for hoisting offshore wind turbine blades includes:

[0031] Obtaining blade hoisting monitoring video within a predetermined time period, wind force data at multiple predetermined time points within the predetermined time period, and tension values of the first cable wind rope and the second cable wind rope at multiple predetermined time points within the predetermined time period;

[0032] After arranging the tension values of the first guy rope and the tension values of the second guy rope at a plurality of predetermined time points within the predetermined time period as a first tension input vector and a second tension input vector, respectively, calculating the product between the transposed vector of the first tension input vector and the second tension input vector to obtain a tension correlation input matrix;

[0033] Passing the tension-related input matrix through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix;

[0034] Extracting multiple key frames from the blade hoisting monitoring video;

[0035] Extracting the center of gravity data of the hoisted blade from each of the plurality of key frames to obtain a center of gravity data input vector;

[0036] Passing the center of gravity data input vector through a multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector;

[0037] Passing the wind data at a plurality of predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector;

[0038] Based on the transposed matrix of the tension-related characteristic matrix, correcting the eigenvalues of each position in the tension-related characteristic matrix to obtain a corrected tension-related characteristic matrix;

[0039] Multiplying the corrected tension force correlation characteristic matrix by the wind force time series characteristic vector to obtain a response control characteristic vector;

[0040] calculating a responsiveness estimate of the response control eigenvector relative to the center of gravity change eigenvector to obtain a classification feature matrix; and

[0041] The classification feature matrix is passed through a classifier to obtain a classification result, which is used to indicate whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point.

[0042] In the above-mentioned intelligent control method for offshore wind turbine blade hoisting, the tension-related input matrix is passed through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix, including: extracting a shallow feature matrix from the Mth layer of the first convolutional neural network, where M is an even number; extracting a deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number and N is greater than 2 times of M; and fusing the shallow feature matrix and the deep feature matrix to generate the tension-related feature matrix.

[0043] In the above-mentioned intelligent control method for offshore wind turbine blade installation, extracting multiple key frames from the blade installation monitoring video includes: extracting the multiple key frames from the blade installation monitoring video at a predetermined sampling frequency.

[0044] In the above-mentioned intelligent control method for offshore wind turbine blade hoisting, the center of gravity data input vector is passed through a multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector, including: using the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the center of gravity data input vector with a first one-dimensional convolution kernel to obtain a first-scale center of gravity feature vector; using the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the center of gravity data input vector with a second one-dimensional convolution kernel to obtain a second-scale center of gravity feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and, cascading the first-scale center of gravity feature vector and the second-scale center of gravity feature vector to obtain the center of gravity change feature vector.

[0045] In the above-mentioned intelligent control method for offshore wind turbine blade hoisting, the wind data at multiple predetermined time points within the predetermined time period are passed through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector, including: arranging the wind data at multiple predetermined time points within the predetermined time period as an input vector to obtain a wind time series input vector; using the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a first one-dimensional convolution kernel to obtain a first-scale wind time series feature vector; using the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a second one-dimensional convolution kernel to obtain a second-scale wind time series feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and, cascading the first-scale wind time series feature vector and the second-scale wind time series feature vector to obtain the wind time series feature vector.

[0046] In the above-mentioned offshore wind turbine blade hoisting intelligent control method, based on the transpose matrix of the tension force correlation characteristic matrix, correcting the eigenvalues of each position in the tension force correlation characteristic matrix to obtain a corrected tension force correlation characteristic matrix, including: based on the transpose matrix of the tension force correlation characteristic matrix, correcting the eigenvalues of each position in the tension force correlation characteristic matrix using the following formula to obtain the corrected tension force correlation characteristic matrix;

[0047] Wherein, the formula is:

[0048]

[0049] Where M represents the tension-related feature matrix, Cov1() and Cov2() are both single convolutional layers, Used to map two-dimensional position coordinates to one-dimensional values, P M Represents the (x, y) coordinate matrix of matrix M, and ⊙ represents the point multiplication by position.

[0050] In the above-mentioned offshore wind turbine blade hoisting intelligent control method, calculating the responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector to obtain a classification feature matrix includes: calculating the responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector using the following formula to obtain the classification feature matrix;

[0051] Wherein, the formula is:

[0052] V1=M*V2

[0053] Wherein, V1 represents the response control feature vector, V2 represents the center of gravity change feature vector, and M represents the classification feature matrix.

[0054] In the above-mentioned offshore wind turbine blade hoisting intelligent control method, the classification feature matrix is passed through a classifier to obtain a classification result, including: the classifier processes the classification 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)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0055] Compared with the existing technology, the intelligent control system and method for offshore wind turbine blade hoisting provided by the present application adopt a convolutional neural network model based on deep learning as a feature extractor, and dynamically adjust the tension of the two wind ropes in real time through the wind force change characteristics at different times and the center of gravity change characteristics of the wind turbine blades. In this process, the control of the wind force and the winch is regarded as a whole, and the change of the center of gravity data is used as a reference to intelligently adjust the control force of the wind rope of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0057] Figure 1 This is a diagram of an application scenario of an intelligent control system for offshore wind turbine blade hoisting according to an embodiment of the present application.

[0058] Figure 2 This is a block diagram of an intelligent control system for offshore wind turbine blade installation according to an embodiment of the present application.

[0059] Figure 3 This is a block diagram of a first multi-scale neighborhood feature extraction module of an intelligent control system for offshore wind turbine blade installation according to an embodiment of the present application.

[0060] Figure 4 This is a flow chart of an intelligent control method for offshore wind turbine blade installation according to an embodiment of the present application.

[0061] Figure 5 Schematic diagram of the architecture of the intelligent control method for offshore wind turbine blade installation according to an embodiment of the present application. DETAILED DESCRIPTION

[0062] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0063] Scenario Overview

[0064] As previously mentioned, the development and utilization of wind energy resources is a crucial, environmentally friendly energy development strategy for adjusting my country's power structure and has garnered widespread attention in recent years. To increase wind power production, wind farms are often located in areas with high wind speeds. However, due to their complex geometry, low stiffness, and low mass, blades can only be installed at wind speeds below 12 m / s. This requires expensive lifting equipment (installation vessels and cranes) and increases crew waiting time, leading to increased installation costs.

[0065] Therefore, during the construction of offshore wind farms, how to reduce the wind speed restrictions imposed by blade installation and improve installation efficiency is a common concern for researchers and engineers. Therefore, an intelligent control solution for offshore wind turbine blade installation is desired to ensure accurate and safe blade installation while improving installation efficiency.

[0066] In the existing blade installation method, such as Figure 1 As shown, the blades are connected to the slings through a clamp and lifted by a winch. The clamping point of the clamp is arranged symmetrically with the center of gravity of the blade. The two guy ropes controlled by the winch reduce the blade swing caused by wind load by adjusting their control force in real time, ensuring that the blades are lifted according to the predetermined route and posture.

[0067] Based on this, the inventors of the present application found that the intelligent control of the two cables for wind rope control by the winch is a key factor in reducing the wind speed restriction of blade installation and improving the accuracy and effectiveness of the blade installation. Therefore, in the technical solution of the present application, it is expected to dynamically adjust the tension of the two cables in real time through the wind force change characteristics at different times and the center of gravity change characteristics of the wind turbine blades. In this process, it is chosen to regard the wind force and the control of the winch as a whole, and use the change of the center of gravity data as a reference to intelligently control the control force of the cable of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the blade installation.

[0068] Specifically, in the technical solution of the present application, first, an anemometer is used to obtain wind data at multiple predetermined time points within a predetermined time period, and tension meters are respectively deployed on the first and second guy ropes to obtain the tension values of the first and second guy ropes at multiple predetermined time points within the predetermined time period. Furthermore, considering that when the tension of the two guy ropes is dynamically adjusted in real time based on the wind force variation characteristics at various times, it is also necessary to monitor the installation effect of the wind turbine blades in real time and monitor the position of the center of gravity to use the changes in the center of gravity data as a reference, it is also necessary to use a camera to obtain a monitoring video of the blade hoisting within the predetermined time period.

[0069] Then, it should be understood that since the first and second guy ropes need to cooperate with each other to install the wind turbine blades, there are special correlation features between the first and second guy ropes. If we want to deeply explore this special implicit correlation feature to subsequently adjust the control force of the guy ropes, we can further arrange the tension values of the first guy rope and the second guy rope at multiple predetermined time points within the predetermined time period into a first tension input vector and a second tension input vector, respectively, and calculate the product between the transposed vector of the first tension input vector and the second tension input vector to integrate the correlation information of the tensions of the two guy ropes, thereby obtaining a tension correlation input matrix. Then, the tension correlation input matrix is processed by a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to extract high-dimensional implicit correlation features of the tensions of the two guy ropes, thereby obtaining a tension correlation feature matrix. It should be understood that, in particular, here, the adjacent convolutional layers of the first convolutional neural network use mutually transposed convolution kernels, so that during training, the network parameters can be updated and the network parameter structure suitable for the specific data structure can be searched simultaneously, thereby improving the adequacy of the associated feature mining and improving the accuracy of classification.

[0070] For the blade hoisting monitoring video within the predetermined time period, since the data information of the blade hoisting monitoring video is relatively complex and contains a lot of interfering and useless information, in order to improve the accuracy of the wind turbine blade installation in the monitoring video, the multiple key frames are further extracted from the blade hoisting monitoring video at a predetermined sampling frequency.

[0071] Then, in order to regard the control of wind force and winch as a whole, the change of key data is used as a reference to intelligently control the control force adjustment of the winch's cable rope, and then the center of gravity data of the hoisted blade is extracted from each key frame of the multiple key frames to obtain the center of gravity data input vector.

[0072] It's understandable that convolutional neural networks were originally developed for image processing, but their concept of local feature extraction can also be applied to time series data analysis. For example, a time series convolution architecture with a convolution kernel size of 3 applies a sliding window along the time dimension to the input time series data, outputting a weighted sum of the data within each time series segment. Each convolution unit stacks multiple convolution kernels to produce multidimensional features. Large convolution kernels extract features from large-scale time series neighborhoods, where the influence of each value within the neighborhood is smaller, thereby reducing input data fluctuations and mitigating the impact of noise on the output features. However, large convolution kernels reduce differences in numerical changes, which can lead to oversmoothing and loss of discriminative power in the output features. In contrast, small convolution kernels better preserve information in the input data but are also more susceptible to noise. Therefore, considering the characteristics of convolutions of different scales, convolution units of different sizes are combined to extract features at different time series scales. Feature fusion is then achieved through feature concatenation, resulting in multi-scale neighborhood features.

[0073] Specifically, in the technical solution of the present application, the convolution layer of the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales is further used to perform one-dimensional convolution encoding on the center of gravity data input vector respectively, and then the obtained feature vectors corresponding to the two one-dimensional convolution kernels of different scales are cascaded to obtain the center of gravity change feature vector. In particular, in this way, the multi-scale neighborhood association of the center of gravity change characteristics of the wind turbine blade in the time series dimension is extracted, that is, the dynamic change of the center of gravity of the wind turbine blade contains implicit feature information, so that the output features contain both the smoothed features and the original input features, avoiding information loss and improving the accuracy of subsequent classification.

[0074] Similarly, for the wind data at multiple predetermined time points within the predetermined time period, after being arranged into wind input vectors according to the time dimension, the wind input vectors are also passed through the multi-scale neighborhood feature extraction module to obtain wind time series feature vectors. In this way, the dynamic change characteristics of the wind data in the time series dimension can be extracted, and the output features can include both smoothed features and original input features, thus avoiding information loss and improving the accuracy of subsequent classification.

[0075] Here, the center of gravity change feature vector obtained by passing the time series of the center of gravity data through the multi-scale neighborhood feature extraction module can express the correlation characteristics across large time series. However, although the first convolutional neural network using mutually transposed convolution kernels in adjacent layers can focus on the specific structure of the data, it is still unable to focus on the tension correlation characteristics at large scales due to the size limitation of the convolution kernel.

[0076] Therefore, preferably, the tension-related feature matrix, for example, denoted as M, is further subjected to local semantic reasoning based on position proposals, specifically:

[0077]

[0078] Where M represents the tension-related feature matrix, Cov1() and Cov2() are both single convolutional layers, Used to map two-dimensional position coordinates to one-dimensional values, P M Represents the (x, y) coordinate matrix of matrix M, and ⊙ represents the point multiplication by position.

[0079] In other words, this location proposal-based local semantic reasoning can use location information as a proposal to derive local scene semantics through the local perception field of the convolutional layer. This fully integrates the captured local semantics while further deriving larger-scale semantics, thereby achieving the migration of semantic information from small scale to large scale and the inference and prediction of large-scale semantic information. In this way, the large-scale associative semantic expression capability of the response control feature vector can be improved, thereby improving the classification performance of the classification feature matrix and improving classification accuracy.

[0080] Furthermore, the corrected tension-related characteristic matrix is multiplied by the wind time-series characteristic vector to map the wind time-series characteristic vector into the high-dimensional space of the corrected tension-related characteristic matrix, thereby obtaining a response control characteristic vector. It should be understood that this allows the wind and winch control to be considered as a whole to obtain responsive control characteristics for tension changes caused by wind changes, thereby enabling subsequent responsive control of center of gravity changes to ensure effective installation of the wind turbine blades.

[0081] It should be understood that, since the responsiveness control can be viewed as a responsiveness characteristic of the center of gravity change in a high-dimensional feature space, in order to adjust the control force of the guy rope using the change characteristics of the center of gravity data as a reference, a responsiveness estimate of the responsiveness control feature vector relative to the center of gravity change feature vector is further calculated to fuse the feature information of both, thereby obtaining a classification feature matrix. Furthermore, the classification feature matrix is then classified by a classifier to obtain a classification result indicating whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point. Furthermore, the control force of the first guy rope and the control force of the second guy rope are regulated based on the classification results.

[0082] Based on this, the present application proposes an intelligent control system for hoisting blades of offshore wind turbines, which includes: a hoisting data acquisition module for acquiring blade hoisting monitoring video within a predetermined time period, wind data at multiple predetermined time points within the predetermined time period, and the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period; a cable wind rope data structuring module for arranging the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period into a first tension input vector and a second tension input vector, respectively. After the tension input vector is input, the product between the transposed vector of the first tension input vector and the second tension input vector is calculated to obtain a tension-related input matrix; a cable data encoding module is used to pass the tension-related input matrix through the first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix; a key frame extraction module is used to extract multiple key frames from the blade hoisting monitoring video; a center of gravity data extraction module is used to extract the center of gravity data of the hoisted blade from each key frame of the multiple key frames to obtain the center of gravity a data input vector; a first multi-scale neighborhood feature extraction module for passing the center of gravity data input vector through the multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector; a second multi-scale neighborhood feature extraction module for passing the wind data at multiple predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector; a correction module for correcting the eigenvalues of each position in the tension force association feature matrix based on the transpose matrix of the tension force association feature matrix to obtain a corrected tension force association feature matrix; a response control module for multiplying the corrected tension force association feature matrix with the wind time series feature vector to obtain a response control feature vector; a responsiveness estimation module for calculating a responsiveness estimate of the response control feature vector relative to the center of gravity change feature vector to obtain a classification feature matrix; and a hoisting control result generation module for passing the classification feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the control force of the first cable wind rope at the current time point should be increased or decreased, and whether the control force of the second cable wind rope should be increased or decreased.

[0083] Figure 1 The diagram shows an application scenario of the offshore wind turbine blade hoisting intelligent control system according to an embodiment of the present application. Figure 1 As shown, in this application scenario, first, an anemometer (e.g., Figure 1 A) as shown in the figure obtains wind data at multiple predetermined time points within a predetermined time period, and wind data are respectively deployed on the first cable wind rope (for example, Figure 1 R1 shown in FIG) and a second guy line (e.g. Figure 1 A tensiometer (e.g., R2) is provided (e.g., Figure 1 The tension value of the first guy rope and the tension value of the second guy rope at a plurality of predetermined time points within the predetermined time period are obtained by using a camera (e.g., T1 and T2 shown in FIG. Figure 1 C) Obtain the leaves within the predetermined time period (for example, Figure 1 Then, the wind force data obtained at multiple predetermined time points within the predetermined time period, the tension values of the first and second wind ropes, and the blade hoisting monitoring video within the predetermined time period are input into a server deployed with an offshore wind turbine blade hoisting intelligent control algorithm (for example, Figure 1 The cloud server S shown in the figure) is capable of processing the wind data at multiple predetermined time points within the predetermined time period, the tension value of the first cable wind rope and the tension value of the second cable wind rope, and the blade hoisting monitoring video within the predetermined time period using an intelligent control algorithm for offshore wind turbine blade hoisting, so as to generate a classification result indicating whether the control force of the first cable wind rope should be increased or decreased, and whether the control force of the second cable wind rope should be increased or decreased at the current time point.

[0084] In this application scenario, the fan blades are clamped by a fixture (e.g. Figure 1 F) and slings (e.g. Figure 1 G) and is connected by a winch (e.g. Figure 1 The blade is lifted by a winch (e.g., W) with the clamping point symmetrical to the center of gravity of the blade; Figure 1 The two guy ropes (H) shown in the figure reduce the blade swing caused by wind load by adjusting their control force in real time, ensuring that the blades are lifted according to the predetermined route and posture.

[0085] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0086] Exemplary Systems

[0087] Figure 2 The figure shows a block diagram of an intelligent control system for offshore wind turbine blade installation according to an embodiment of the present application. Figure 2As shown, according to an embodiment of the present application, the offshore wind turbine blade hoisting intelligent control system 200 includes: a hoisting data acquisition module 210, which is used to obtain a blade hoisting monitoring video within a predetermined time period, wind data at multiple predetermined time points within the predetermined time period, and the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period; a cable wind rope data structuring module 220, which is used to arrange the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period into a first tension input vector and a second tension input vector, respectively. After the force input vector is input, the product between the transposed vector of the first tensioning force input vector and the second tensioning force input vector is calculated to obtain a tensioning force associated input matrix; the cable rope data encoding module 230 is used to pass the tensioning force associated input matrix through the first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tensioning force associated feature matrix; the key frame extraction module 240 is used to extract multiple key frames from the blade hoisting monitoring video; the center of gravity data extraction module 250 is used to extract the center of gravity data of the hoisted blade from each key frame of the multiple key frames to obtain the center of gravity data According to the input vector; a first multi-scale neighborhood feature extraction module 260 is used to pass the center of gravity data input vector through the multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector; a second multi-scale neighborhood feature extraction module 270 is used to pass the wind data of multiple predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector; a correction module 280 is used to correct the eigenvalues of each position in the tension correlation feature matrix based on the transpose matrix of the tension correlation feature matrix to obtain a corrected tension correlation feature matrix; A control module 290 is used to multiply the corrected tensioning force associated characteristic matrix with the wind time series characteristic vector to obtain a response control characteristic vector; a responsiveness estimation module 300 is used to calculate a responsiveness estimate of the response control characteristic vector relative to the center of gravity change characteristic vector to obtain a classification characteristic matrix; and a hoisting control result generation module 310 is used to pass the classification characteristic matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the control force of the first guy rope at the current time point should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased.

[0088] Specifically, in an embodiment of the present application, the hoisting data acquisition module 210 is configured to acquire blade hoisting monitoring video within a predetermined time period, wind force data at multiple predetermined time points within the predetermined time period, and the tension values of the first and second wind ropes at multiple predetermined time points within the predetermined time period. As previously mentioned, since intelligent control of the two wind ropes used to control the winch is a key factor in reducing the wind speed restrictions on blade installation and improving the accuracy and effectiveness of blade installation, the technical solution of the present application is intended to dynamically adjust the tension of the two wind ropes in real time based on the wind force variation characteristics at various times and the center of gravity variation characteristics of the wind turbine blades. In this process, the wind force and winch control are considered as a whole, and the changes in the center of gravity data are used as a reference to intelligently control the control force of the wind ropes of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation.

[0089] Specifically, in the technical solution of the present application, first, an anemometer is used to obtain wind data at multiple predetermined time points within a predetermined time period, and tension meters are respectively deployed on the first and second guy ropes to obtain the tension values of the first and second guy ropes at multiple predetermined time points within the predetermined time period. Furthermore, considering that in order to dynamically adjust the tension of the two guy ropes in real time based on the wind force variation characteristics at various times, it is also necessary to monitor the installation effect of the wind turbine blades in real time and to monitor the position of the center of gravity to use the changes in the center of gravity data as a reference, it is also necessary to use a camera to obtain a monitoring video of the blade hoisting within the predetermined time period.

[0090] Specifically, in an embodiment of the present application, the cable guy rope data structuring module 220 and the cable guy rope data encoding module 230 are used to arrange the tension values of the first cable guy rope and the tension values of the second cable guy rope at multiple predetermined time points within the predetermined time period as a first tension input vector and a second tension input vector, respectively, calculate the product between the transposed vector of the first tension input vector and the second tension input vector to obtain a tension correlation input matrix, and pass the tension correlation input matrix through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension correlation feature matrix. It should be understood that since the first guy rope and the second guy rope need to cooperate with each other to install the wind turbine blades, there are special correlation characteristics between the first guy rope and the second guy rope. If one wants to deeply explore this special implicit correlation characteristic to carry out subsequent control force regulation of the guy rope, in the technical solution of the present application, the tension values of the first guy rope and the tension values of the second guy rope at multiple predetermined time points within the predetermined time period are further arranged as the first tension input vector and the second tension input vector, respectively, and the product between the transposed vector of the first tension input vector and the second tension input vector is calculated to integrate the correlation information of the tensions of the two guy ropes, thereby obtaining a tension correlation input matrix.

[0091] The tension force correlation input matrix is then processed through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to extract high-dimensional implicit correlation features of the tension forces of the two guy ropes, thereby obtaining a tension force correlation feature matrix. It should be understood that, in particular, the use of mutually transposed convolution kernels in adjacent convolution layers of the first convolutional neural network allows for simultaneous updating of network parameters and search for a network parameter structure suitable for a specific data structure during training, thereby improving the adequacy of correlation feature mining and enhancing classification accuracy.

[0092] More specifically, in an embodiment of the present application, the cable wind rope data encoding module includes: a shallow feature matrix extraction unit for extracting a shallow feature matrix from the Mth layer of the first convolutional neural network, where M is an even number; a deep feature matrix extraction unit for extracting a deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number and N is greater than 2 times of M; and a feature matrix fusion unit for fusing the shallow feature matrix and the deep feature matrix to generate the tension-related feature matrix.

[0093] Specifically, in this embodiment of the present application, the keyframe extraction module 240 and the center of gravity data extraction module 250 are configured to extract multiple keyframes from the blade installation monitoring video and extract the center of gravity data of the installed blade from each of the multiple keyframes to obtain a center of gravity data input vector. Specifically, in the technical solution of the present application, for the blade installation monitoring video within the predetermined time period, since the data information in the blade installation monitoring video is relatively complex and contains a lot of interfering and useless information, in order to improve the accuracy of the wind turbine blade installation in the monitoring video, the multiple keyframes are further extracted from the blade installation monitoring video at a predetermined sampling frequency. Then, in order to integrate wind power and winch control, changes in key data are used as a reference to intelligently control the control force of the winch's cable. The center of gravity data of the installed blade is then extracted from each of the multiple keyframes to obtain a center of gravity data input vector.

[0094] Specifically, in this embodiment of the present application, the first multi-scale neighborhood feature extraction module 260 is configured to pass the centroid data input vector through the multi-scale neighborhood feature extraction module to obtain a centroid change feature vector. It should be understood that convolutional neural networks were originally applied to image processing, but their concept of local feature extraction can also be applied to time series data analysis. For example, in a time series convolution structure with a convolution kernel size of 3, the convolution kernel moves along the time dimension in the form of a sliding window for time series data input and outputs a weighted sum of the data within each time series segment. Each convolution unit stacks multiple convolution kernels to output multidimensional features. Large convolution kernels extract features from a large-scale time series neighborhood, where the impact of each value within the neighborhood is smaller, thereby reducing fluctuations in the input data and mitigating the impact of noise on the output features. However, large-scale convolution kernels reduce differences in numerical changes, which can easily lead to oversmoothing and cause the output features to lose their discriminative power. In contrast, small-scale convolution kernels can better preserve information in the input data but are also more susceptible to interference from noise. Therefore, considering the characteristics of convolutions at different scales, we combine convolution units of different sizes to extract features at different temporal scales. Feature fusion is then performed using feature concatenation to obtain multi-scale neighborhood features.

[0095] Specifically, in the technical solution of the present application, the convolution layer of the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales is further used to perform one-dimensional convolution encoding on the center of gravity data input vector respectively, and then the obtained feature vectors corresponding to the two one-dimensional convolution kernels of different scales are cascaded to obtain the center of gravity change feature vector. In particular, in this way, the multi-scale neighborhood association of the center of gravity change characteristics of the wind turbine blade in the time series dimension is extracted, that is, the dynamic change of the center of gravity of the wind turbine blade contains implicit feature information, so that the output features contain both the smoothed features and the original input features, avoiding information loss and improving the accuracy of subsequent classification.

[0096] More specifically, in an embodiment of the present application, the first multi-scale neighborhood feature extraction module includes: a first-scale centroid encoding unit, used to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector with a first one-dimensional convolution kernel to obtain a first-scale centroid feature vector; a second-scale centroid encoding unit, used to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector with a second one-dimensional convolution kernel to obtain a second-scale centroid feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and a centroid-scale fusion unit, used to cascade the first-scale centroid feature vector and the second-scale centroid feature vector to obtain the centroid change feature vector.

[0097] Figure 3 FIG1 is a block diagram of a first multi-scale neighborhood feature extraction module in an offshore wind turbine blade hoisting intelligent control system according to an embodiment of the present application. Figure 3 As shown, the first multi-scale neighborhood feature extraction module 260 includes: a first-scale centroid encoding unit 261, used to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector with a first one-dimensional convolution kernel to obtain a first-scale centroid feature vector; a second-scale centroid encoding unit 262, used to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector with a second one-dimensional convolution kernel to obtain a second-scale centroid feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and a centroid-scale fusion unit 263, used to cascade the first-scale centroid feature vector and the second-scale centroid feature vector to obtain the centroid change feature vector.

[0098] Specifically, in an embodiment of the present application, the second multi-scale neighborhood feature extraction module 270 is used to pass the wind data at multiple predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector. That is, in the technical solution of the present application, similarly, for the wind data at multiple predetermined time points within the predetermined time period, after arranging them into wind input vectors according to the time dimension, the wind input vectors are also passed through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector. In this way, the dynamic change characteristics of the wind data in the time series dimension can be extracted, and the output features can also include both the smoothed features and the original input features, thereby avoiding information loss and improving the accuracy of subsequent classification.

[0099] More specifically, in an embodiment of the present application, the second multi-scale neighborhood feature extraction module includes: a wind input vector construction unit, used to arrange the wind data of multiple predetermined time points within the predetermined time period as an input vector to obtain a wind time series input vector; a first-scale wind encoding unit, used to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a first one-dimensional convolution kernel to obtain a first-scale wind time series feature vector; a second-scale wind encoding unit, used to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a second one-dimensional convolution kernel to obtain a second-scale wind time series feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and a wind scale fusion unit, used to cascade the first-scale wind time series feature vector and the second-scale wind time series feature vector to obtain the wind time series feature vector.

[0100] Specifically, in an embodiment of the present application, the correction module 280 is used to correct the eigenvalues of each position in the tension-related feature matrix based on the transposed matrix of the tension-related feature matrix to obtain a corrected tension-related feature matrix. It should be understood that in the technical solution of the present application, the center of gravity change feature vector obtained by passing the time series of the center of gravity data through the multi-scale neighborhood feature extraction module can express the correlation features across a large time series. However, although the first convolutional neural network using mutually transposed convolution kernels in the adjacent layers can focus on the specific structure of the data, it is still unable to focus on the tension-related features at a large scale due to the size limitation of the convolution kernel. Therefore, preferably, in the technical solution of the present application, the tension-related feature matrix is further subjected to local semantic reasoning based on, for example, position proposals.

[0101] That is, the location proposal-based local semantic reasoning can use location information as a proposal to derive and infer local scene semantics through the local perception field of the convolutional layer. This fully integrates the captured local semantics while further deriving larger-scale semantics, thereby achieving the migration of semantic information from small scale to large scale and the inference and prediction of large-scale semantic information. In this way, the large-scale associative semantic expression capability of the response control feature vector can be improved, thereby improving the classification performance of the classification feature matrix and improving classification accuracy.

[0102] More specifically, in an embodiment of the present application, the correction module is further configured to: based on the transposed matrix of the tension-related characteristic matrix, correct the eigenvalues of each position in the tension-related characteristic matrix using the following formula to obtain the corrected tension-related characteristic matrix;

[0103] Wherein, the formula is:

[0104]

[0105] Where M represents the tension-related feature matrix, Cov1() and Cov2() are both single convolutional layers, Used to map two-dimensional position coordinates to one-dimensional values, P M Represents the (x, y) coordinate matrix of matrix M, and ⊙ represents the point multiplication by position.

[0106] Specifically, in the embodiment of the present application, the response control module 290 is configured to multiply the corrected tension-related characteristic matrix by the wind time-series characteristic vector to obtain a response control characteristic vector. That is, in the technical solution of the present application, the tension-related characteristic matrix is further multiplied by the wind time-series characteristic vector to map the wind time-series characteristic vector to the high-dimensional space of the corrected tension-related characteristic matrix, thereby obtaining a response control characteristic vector. It should be understood that in this way, the control of wind power and the winch can be considered as a whole to obtain the responsive control characteristics of the tension changes caused by the wind changes, and then perform responsive control of the subsequent center of gravity changes to ensure the installation effect of the wind turbine blades.

[0107] Specifically, in this embodiment of the present application, the responsiveness estimation module 300 and the hoisting control result generation module 310 are configured to calculate the responsiveness estimate of the responsiveness control feature vector relative to the center of gravity change feature vector to obtain a classification feature matrix, and then pass the classification feature matrix through a classifier to obtain a classification result. The classification result indicates whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point. It should be understood that since the responsiveness control can be viewed as the responsiveness feature of the center of gravity change in a high-dimensional feature space, in the technical solution of the present application, in order to adjust the control force of the guy ropes based on the change features of the center of gravity data as a reference, the responsiveness estimate of the responsiveness control feature vector relative to the center of gravity change feature vector is further calculated to fuse the feature information of the two, thereby obtaining a classification feature matrix. Furthermore, the classification feature matrix is then passed through a classifier for classification to obtain a classification result indicating whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point. Then, the control force of the first guy rope and the control force of the second guy rope are regulated based on the classification result. In a specific example, the classifier processes the classification 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)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

[0108] More specifically, in an embodiment of the present application, the responsiveness estimation module is further configured to: calculate a responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector using the following formula to obtain the classification feature matrix;

[0109] Wherein, the formula is:

[0110] V1=M*V2

[0111] Wherein, V1 represents the response control feature vector, V2 represents the center of gravity change feature vector, and M represents the classification feature matrix.

[0112] In summary, the offshore wind turbine blade hoisting intelligent control system 200 based on the embodiment of the present application is explained, which adopts a convolutional neural network model based on deep learning as a feature extractor, and dynamically adjusts the tension of the two wind ropes in real time through the wind force change characteristics at different times and the center of gravity change characteristics of the wind turbine blades. In this process, the control of the wind force and the winch is regarded as a whole, and the change of the center of gravity data is used as a reference to intelligently adjust the control force of the wind rope of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation.

[0113] As described above, the offshore wind turbine blade installation intelligent control system 200 according to an embodiment of the present application can be implemented in various terminal devices, such as a server for an offshore wind turbine blade installation intelligent control algorithm. In one example, the offshore wind turbine blade installation intelligent control system 200 according to an embodiment 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 blade installation intelligent control system 200 can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the offshore wind turbine blade installation intelligent control system 200 can also be one of the many hardware modules of the terminal device.

[0114] Alternatively, in another example, the offshore wind turbine blade hoisting intelligent control system 200 and the terminal device may also be separate devices, and the offshore wind turbine blade hoisting intelligent control system 200 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0115] Exemplary Methods

[0116] Figure 3 The figure shows the flow chart of the intelligent control method for offshore wind turbine blade hoisting. Figure 3As shown, the intelligent control method for offshore wind turbine blade hoisting according to the embodiment of the present application includes the following steps: S110, obtaining a blade hoisting monitoring video within a predetermined time period, wind data at multiple predetermined time points within the predetermined time period, and the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period; S120, arranging the tension value of the first cable wind rope and the tension value of the second cable wind rope at multiple predetermined time points within the predetermined time period as a first tension input vector and a second tension input vector, respectively, and calculating the product between the transposed vector of the first tension input vector and the second tension input vector to obtain a tension-related input matrix; S130, passing the tension-related input matrix through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix; S140, extracting multiple key frames from the blade hoisting monitoring video; S150, extracting the center of gravity data of the hoisted blade from each key frame of the multiple key frames respectively. to obtain a center of gravity data input vector; S160, passing the center of gravity data input vector through a multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector; S170, passing the wind data of multiple predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector; S180, based on the transpose matrix of the tension association feature matrix, correcting the eigenvalues of each position in the tension association feature matrix to obtain a corrected tension association feature matrix; S190, multiplying the corrected tension association feature matrix with the wind time series feature vector to obtain a response control feature vector; S200, calculating the responsiveness estimate of the response control feature vector relative to the center of gravity change feature vector to obtain a classification feature matrix; and, S210, passing the classification feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the control force of the first cable wind rope at the current time point should be increased or decreased, and whether the control force of the second cable wind rope should be increased or decreased.

[0117] Figure 4 The figure shows a schematic diagram of the architecture of the intelligent control method for offshore wind turbine blade installation according to an embodiment of the present application. Figure 4 As shown, in the network architecture of the offshore wind turbine blade hoisting intelligent control method, first, the tension value of the first cable wind rope at multiple predetermined time points within the predetermined time period is obtained (for example, Figure 4 ) and the tension value of the second guy rope (e.g., Figure 4 P2) are arranged as the first tensioning force input vector (for example, Figure 4 V1 as shown in FIG) and a second tensioning force input vector (e.g., as shown in FIG). Figure 4After V2 is shown in FIG, the product between the transposed vector of the first tensioning force input vector and the second tensioning force input vector is calculated to obtain a tensioning force associated input matrix (for example, as shown in FIG. Figure 4 Then, the tension-related input matrix is passed through the adjacent layers using the first convolutional neural network with mutually transposed convolution kernels (e.g., Figure 4 CNN1 as shown in FIG) to obtain the tension-related feature matrix (e.g., Figure 4 Then, from the obtained blade hoisting monitoring video (for example, as shown in MF1); Figure 4 Q1) extracts multiple key frames (e.g., Figure 4 Q2 shown in FIG); then, the center of gravity data of the hoisted blade is extracted from each key frame of the plurality of key frames to obtain a center of gravity data input vector (for example, Figure 4 Then, the center of gravity data input vector is passed through a multi-scale neighborhood feature extraction module (for example, Figure 4 MS as shown in ) to obtain the center of gravity change feature vector (for example, as Figure 4 Then, the wind force data of a plurality of predetermined time points within the predetermined time period (for example, Figure 4 K) is extracted by the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector (for example, Figure 4 Then, based on the transpose matrix of the tension-related characteristic matrix, the eigenvalues of each position in the tension-related characteristic matrix are corrected to obtain a corrected tension-related characteristic matrix (for example, Figure 4 MF2 shown in FIG); Then, the corrected tension force correlation characteristic matrix is multiplied by the wind time series characteristic vector to obtain a response control characteristic vector (for example, Figure 4 Then, the responsiveness estimation of the response control feature vector relative to the center of gravity change feature vector is calculated to obtain a classification feature matrix (for example, as Figure 4 MF shown in ); and, finally, passing the classification feature matrix through a classifier (e.g., Figure 4 ) to obtain a classification result, which is used to indicate whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point.

[0118] More specifically, in step S110, blade hoisting monitoring video within a predetermined time period, wind force data at multiple predetermined time points within the predetermined time period, and the tension values of the first and second wind ropes at multiple predetermined time points within the predetermined time period are obtained. It should be understood that intelligent control of the two wind ropes controlled by the winch is a key factor in reducing the wind speed restrictions on blade installation and improving the accuracy and effectiveness of blade installation. Therefore, in the technical solution of this application, it is desired to dynamically adjust the tension of the two wind ropes in real time based on the wind force variation characteristics at various times and the center of gravity variation characteristics of the wind turbine blades. In this process, the wind force and winch control are considered as a whole, and the changes in the center of gravity data are used as a reference to intelligently control the control force of the winch's wind ropes, thereby ensuring the accuracy and safety of wind turbine blade installation while further improving installation efficiency.

[0119] Specifically, in the technical solution of the present application, first, an anemometer is used to obtain wind data at multiple predetermined time points within a predetermined time period, and tension meters are respectively deployed on the first and second guy ropes to obtain the tension values of the first and second guy ropes at multiple predetermined time points within the predetermined time period. Furthermore, considering that in order to dynamically adjust the tension of the two guy ropes in real time based on the wind force variation characteristics at various times, it is also necessary to monitor the installation effect of the wind turbine blades in real time and to monitor the position of the center of gravity to use the changes in the center of gravity data as a reference, it is also necessary to use a camera to obtain a monitoring video of the blade hoisting within the predetermined time period.

[0120] More specifically, in steps S120 and S130, after arranging the tension values of the first guy rope and the tension values of the second guy rope at multiple predetermined time points within the predetermined time period as a first tension input vector and a second tension input vector, respectively, the product between the transposed vector of the first tension input vector and the second tension input vector is calculated to obtain a tension association input matrix, and the tension association input matrix is passed through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension association feature matrix. It should be understood that since the first guy rope and the second guy rope need to cooperate with each other to install the wind turbine blades, there are special correlation characteristics between the first guy rope and the second guy rope. If one wants to deeply explore this special implicit correlation characteristic to carry out subsequent control force regulation of the guy rope, in the technical solution of the present application, the tension values of the first guy rope and the tension values of the second guy rope at multiple predetermined time points within the predetermined time period are further arranged as the first tension input vector and the second tension input vector, respectively, and the product between the transposed vector of the first tension input vector and the second tension input vector is calculated to integrate the correlation information of the tensions of the two guy ropes, thereby obtaining a tension correlation input matrix.

[0121] The tension force correlation input matrix is then processed through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to extract high-dimensional implicit correlation features of the tension forces of the two guy ropes, thereby obtaining a tension force correlation feature matrix. It should be understood that, in particular, the use of mutually transposed convolution kernels in adjacent convolution layers of the first convolutional neural network allows for simultaneous updating of network parameters and search for a network parameter structure suitable for a specific data structure during training, thereby improving the adequacy of correlation feature mining and enhancing classification accuracy.

[0122] More specifically, in steps S140 and S150, multiple key frames are extracted from the blade installation monitoring video, and the center of gravity data of the installed blade is extracted from each of the multiple key frames to obtain a center of gravity data input vector. That is, in the technical solution of the present application, for the blade installation monitoring video within the predetermined time period, since the data information in the blade installation monitoring video is relatively complex and contains a lot of interfering and useless information, in order to improve the accuracy of the wind turbine blade installation in the monitoring video, the multiple key frames are further extracted from the blade installation monitoring video at a predetermined sampling frequency. Then, in order to consider the wind force and winch control as a whole, the changes in key data are used as a reference to intelligently control the control force adjustment of the winch's cable, and the center of gravity data of the installed blade is extracted from each of the multiple key frames to obtain a center of gravity data input vector.

[0123] More specifically, in step S160, the centroid data input vector is passed through a multi-scale neighborhood feature extraction module to obtain a centroid change feature vector. It should be understood that convolutional neural networks were originally applied to image processing, but their concept of local feature extraction can also be applied to time series data analysis. For example, in a time series convolution structure with a convolution kernel size of 3, the convolution kernel moves along the time dimension in the form of a sliding window for time series data input and outputs a weighted sum of the data within each time series segment. Each convolution unit stacks multiple convolution kernels to output multidimensional features. Large convolution kernels extract features from a large-scale time series neighborhood, where the impact of each value in the neighborhood is smaller, thereby reducing fluctuations in the input data and alleviating the impact of noise on the output features. However, large-scale convolution kernels reduce the differences in numerical changes, which can easily lead to oversmoothing and cause the output features to lose their discriminative power. In contrast, small-scale convolution kernels can better preserve information in the input data but are also more susceptible to interference from noise. Therefore, considering the characteristics of convolutions at different scales, we combine convolution units of different sizes to extract features at different temporal scales. Feature fusion is then performed using feature concatenation to obtain multi-scale neighborhood features.

[0124] Specifically, in the technical solution of the present application, the convolution layer of the multi-scale neighborhood feature extraction module with one-dimensional convolution kernels of different scales is further used to perform one-dimensional convolution encoding on the center of gravity data input vector respectively, and then the obtained feature vectors corresponding to the two one-dimensional convolution kernels of different scales are cascaded to obtain the center of gravity change feature vector. In particular, in this way, the multi-scale neighborhood association of the center of gravity change characteristics of the wind turbine blade in the time series dimension is extracted, that is, the dynamic change of the center of gravity of the wind turbine blade contains implicit feature information, so that the output features contain both the smoothed features and the original input features, avoiding information loss and improving the accuracy of subsequent classification.

[0125] More specifically, in step S170, the wind data at multiple predetermined time points within the predetermined time period are passed through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector. That is, in the technical solution of the present application, similarly, for the wind data at multiple predetermined time points within the predetermined time period, after arranging them into wind input vectors according to the time dimension, the wind input vectors are also passed through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector. In this way, the dynamic change characteristics of the wind data in the time series dimension can be extracted, and the output features can include both the smoothed features and the original input features, thereby avoiding information loss and improving the accuracy of subsequent classification.

[0126] More specifically, in step S180, based on the transposed matrix of the tension-related feature matrix, the eigenvalues of each position in the tension-related feature matrix are corrected to obtain a corrected tension-related feature matrix. It should be understood that in the technical solution of the present application, the center of gravity change feature vector obtained by passing the time series of the center of gravity data through the multi-scale neighborhood feature extraction module can express the correlation features across a large time series. However, although the first convolutional neural network using mutually transposed convolution kernels in the adjacent layers can focus on the specific structure of the data, it is still unable to focus on the tension-related features at a large scale due to the size limitation of the convolution kernel. Therefore, preferably, in the technical solution of the present application, the tension-related feature matrix is further subjected to local semantic reasoning based on, for example, position proposals.

[0127] That is, the location proposal-based local semantic reasoning can use location information as a proposal to derive and infer local scene semantics through the local perception field of the convolutional layer. This fully integrates the captured local semantics while further deriving larger-scale semantics, thereby achieving the migration of semantic information from small scale to large scale and the inference and prediction of large-scale semantic information. In this way, the large-scale associative semantic expression capability of the response control feature vector can be improved, thereby improving the classification performance of the classification feature matrix and improving classification accuracy.

[0128] More specifically, in step S190, the corrected tension-related characteristic matrix is multiplied by the wind time-series characteristic vector to obtain a response control characteristic vector. That is, in the technical solution of the present application, the tension-related characteristic matrix is further multiplied by the wind time-series characteristic vector to map the wind time-series characteristic vector into the high-dimensional space of the corrected tension-related characteristic matrix, thereby obtaining a response control characteristic vector. It should be understood that in this way, the control of wind power and the winch can be considered as a whole to obtain the responsive control characteristics of the tension changes caused by the wind changes, and then perform responsive control of the subsequent center of gravity changes to ensure the installation effect of the wind turbine blades.

[0129] More specifically, in steps S200 and S210, an estimated responsiveness of the response control feature vector relative to the center of gravity change feature vector is calculated to obtain a classification feature matrix. This classification feature matrix is then passed through a classifier to obtain a classification result. The classification result indicates whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased, at the current time point. It should be understood that since the responsiveness control can be viewed as a responsiveness feature of the center of gravity change in a high-dimensional feature space, in the technical solution of the present application, in order to adjust the control force of the guy ropes using the change characteristics of the center of gravity data as a reference, an estimated responsiveness of the response control feature vector relative to the center of gravity change feature vector is further calculated to fuse the feature information of these two, thereby obtaining a classification feature matrix. Furthermore, this classification feature matrix is then passed through a classifier for classification to obtain a classification result indicating whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased, at the current time point. Furthermore, the control force of the first guy rope and the control force of the second guy rope are regulated based on the classification result.

[0130] In summary, the intelligent control method for offshore wind turbine blade hoisting based on the embodiment of the present application is explained, which adopts a convolutional neural network model based on deep learning as a feature extractor, and dynamically adjusts the tension of the two wind ropes in real time through the wind force change characteristics at different times and the center of gravity change characteristics of the wind turbine blades. In this process, the control of the wind force and the winch is regarded as a whole, and the change of the center of gravity data is used as a reference to intelligently adjust the control force of the wind rope of the winch, thereby further improving the installation efficiency while ensuring the accuracy and safety of the wind turbine blade installation.

[0131] The basic principles of the present application have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in this application are merely illustrative and not restrictive, and it should not be assumed that these advantages, strengths, and effects are required of each embodiment of this application. In addition, the specific details disclosed above are merely illustrative and facilitating understanding, and are not restrictive. The above details do not limit this application to necessarily being implemented using the above specific details.

[0132] The block diagrams of the devices, devices, equipment, 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 will be appreciated by those skilled in the art, these devices, devices, equipment, and systems can be connected, arranged, or configured in any manner. Words such as "include," "comprise," "have," and the like are open-ended words, meaning "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or" and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0133] It should also be noted that in the apparatus, device, and method of the present application, each component or each step can be decomposed and / or recombined, and such decomposition and / or recombination should be regarded as equivalent solutions of the present application.

[0134] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present application. Therefore, the present 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.

[0135] The above description has been provided for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although a number of example 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 intelligent control system for offshore wind turbine blade installation, characterized in that: include: a hoisting data acquisition module, configured to acquire blade hoisting monitoring video within a predetermined time period, wind force data at a plurality of predetermined time points within the predetermined time period, and tension values of the first cable wind rope and the second cable wind rope at a plurality of predetermined time points within the predetermined time period; a cable guy rope data structuring module, configured to respectively arrange the tension values of the first cable guy rope and the tension values of the second cable guy rope at a plurality of predetermined time points within the predetermined time period into a first tension force input vector and a second tension force input vector, and calculate the product between the transposed vector of the first tension force input vector and the second tension force input vector to obtain a tension force correlation input matrix; a cable wind rope data encoding module, configured to pass the tension force correlation input matrix through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain a tension force correlation feature matrix; A key frame extraction module, configured to extract a plurality of key frames from the blade hoisting monitoring video; A center of gravity data extraction module is used to extract the center of gravity data of the hoisted blade from each key frame of the plurality of key frames to obtain a center of gravity data input vector; A first multi-scale neighborhood feature extraction module, configured to pass the center of gravity data input vector through the multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector; a second multi-scale neighborhood feature extraction module, configured to pass the wind data at a plurality of predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector; a correction module, configured to correct the eigenvalues of each position in the tension-related characteristic matrix based on a transposed matrix of the tension-related characteristic matrix to obtain a corrected tension-related characteristic matrix; a response control module, configured to multiply the corrected tension force correlation characteristic matrix by the wind force time series characteristic vector to obtain a response control characteristic vector; a responsiveness estimation module, configured to calculate a responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector to obtain a classification feature matrix; as well as The hoisting control result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, and the classification result is used to indicate whether the control force of the first cable guy rope should be increased or decreased, and whether the control force of the second cable guy rope should be increased or decreased at the current time point.

2. The offshore wind turbine blade hoisting intelligent control system according to claim 1 is characterized in that: The cable rope data encoding module includes: a shallow feature matrix extraction unit, configured to extract a shallow feature matrix from the Mth layer of the first convolutional neural network, where M is an even number; a deep feature matrix extraction unit, configured to extract a deep feature matrix from the Nth layer of the first convolutional neural network, wherein N is an even number and is greater than 2 times M; and A feature matrix fusion unit is used to fuse the shallow feature matrix and the deep feature matrix to generate the tension-related feature matrix.

3. The offshore wind turbine blade hoisting intelligent control system according to claim 2 is characterized in that: The key frame extraction module is further configured to extract the multiple key frames from the blade hoisting monitoring video at a predetermined sampling frequency.

4. The intelligent control system for offshore wind turbine blade installation according to claim 3 is characterized in that: The first multi-scale neighborhood feature extraction module includes: a first-scale centroid encoding unit, configured to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector with a first one-dimensional convolution kernel to obtain a first-scale centroid feature vector; a second-scale centroid encoding unit, configured to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the centroid data input vector using a second one-dimensional convolution kernel to obtain a second-scale centroid feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and The center-of-gravity-scale fusion unit is configured to concatenate the first-scale center-of-gravity feature vector and the second-scale center-of-gravity feature vector to obtain the center-of-gravity change feature vector.

5. The intelligent control system for offshore wind turbine blade installation according to claim 4 is characterized in that: The second multi-scale neighborhood feature extraction module includes: a wind input vector construction unit, configured to arrange the wind data at a plurality of predetermined time points within the predetermined time period into input vectors to obtain a wind time series input vector; a first-scale wind encoding unit, configured to use the first convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector with a first one-dimensional convolution kernel to obtain a first-scale wind time series feature vector; a second-scale wind encoding unit, configured to use the second convolution layer of the multi-scale neighborhood feature extraction module to perform one-dimensional convolution encoding on the wind time series input vector using a second one-dimensional convolution kernel to obtain a second-scale wind time series feature vector, wherein the size of the first one-dimensional convolution kernel is different from the size of the second one-dimensional convolution kernel; and A wind scale fusion unit is configured to concatenate the first-scale wind time series feature vector and the second-scale wind time series feature vector to obtain the wind time series feature vector.

6. The offshore wind turbine blade hoisting intelligent control system according to claim 5, characterized in that: The correction module is further configured to: based on the transposed matrix of the tension-related characteristic matrix, correct the eigenvalues of each position in the tension-related characteristic matrix using the following formula to obtain the corrected tension-related characteristic matrix; Wherein, the formula is: Where M represents the tension-related feature matrix, Cov1() and Cov2() are both single convolutional layers, Used to map two-dimensional position coordinates to one-dimensional values, P M Represents the (x, y) coordinate matrix of matrix M, and ⊙ represents the point multiplication by position.

7. The intelligent control system for offshore wind turbine blade installation according to claim 6, characterized in that: The responsiveness estimation module is further configured to: calculate a responsiveness estimation of the response control eigenvector relative to the center of gravity change eigenvector using the following formula to obtain the classification feature matrix; Wherein, the formula is: V1=M*V2 Wherein, V1 represents the response control feature vector, V2 represents the center of gravity change feature vector, and M represents the classification feature matrix.

8. The intelligent control system for offshore wind turbine blade installation according to claim 7, characterized in that: The hoisting control result generation module is further used for: the classifier processes the classification 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)}, where Project(F) represents the projection of the classification feature matrix into a vector, W1 to W n is the weight matrix of each fully connected layer, B1 to B n Represents the bias matrix of each fully connected layer.

9. An intelligent control method for hoisting offshore wind turbine blades, characterized in that: include: Obtaining blade hoisting monitoring video within a predetermined time period, wind force data at multiple predetermined time points within the predetermined time period, and tension values of the first cable wind rope and the second cable wind rope at multiple predetermined time points within the predetermined time period; After arranging the tension values of the first guy rope and the tension values of the second guy rope at a plurality of predetermined time points within the predetermined time period as a first tension input vector and a second tension input vector, respectively, calculating the product between the transposed vector of the first tension input vector and the second tension input vector to obtain a tension correlation input matrix; Passing the tension-related input matrix through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix; Extracting multiple key frames from the blade hoisting monitoring video; Extracting the center of gravity data of the hoisted blade from each of the plurality of key frames to obtain a center of gravity data input vector; Passing the center of gravity data input vector through a multi-scale neighborhood feature extraction module to obtain a center of gravity change feature vector; Passing the wind data at a plurality of predetermined time points within the predetermined time period through the multi-scale neighborhood feature extraction module to obtain a wind time series feature vector; Based on the transposed matrix of the tension-related characteristic matrix, correcting the eigenvalues of each position in the tension-related characteristic matrix to obtain a corrected tension-related characteristic matrix; Multiplying the corrected tension force correlation characteristic matrix by the wind force time series characteristic vector to obtain a response control characteristic vector; calculating a responsiveness estimate of the response control eigenvector relative to the center of gravity change eigenvector to obtain a classification feature matrix; as well as The classification feature matrix is passed through a classifier to obtain a classification result, which is used to indicate whether the control force of the first guy rope should be increased or decreased, and whether the control force of the second guy rope should be increased or decreased at the current time point.

10. The intelligent control method for offshore wind turbine blade installation according to claim 9, characterized in that: The step of passing the tension-related input matrix through a first convolutional neural network with mutually transposed convolution kernels in adjacent layers to obtain a tension-related feature matrix includes: Extracting a shallow feature matrix from the Mth layer of the first convolutional neural network, where M is an even number; Extracting a deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number and is greater than 2 times M; and The shallow feature matrix and the deep feature matrix are fused to generate the tension-related feature matrix.

Images