Edge computing fault detection method for offshore wind turbines during off-grid period

Abstract

The present invention discloses an edge computing fault detection method for an offshore wind turbine during the off-grid period, which belongs to the field of wind power generation technology. The method includes constructing a finite element simulation model and a wind load calculation model of the turbine and obtaining strain data of each point on the supporting structure, selecting the strain data of the first k points with the largest discrimination index and constructing a first fault data set; constructing a computational fluid dynamics model and obtaining the blade vortex-induced vibration amplitude and the strain data of each blade root, and constructing a second fault data set; arranging strain sensors at the positions of the first k points and the root of each blade, and constructing two measured data sets, and then obtaining a spatial transformation matrix; training a machine learning model based on the two fault data sets, the two measured data sets and the spatial transformation matrix, obtaining two detection models and performing fault detection on offshore wind turbines during the off-grid period. The present invention can realize fault diagnosis of offshore wind turbines during off-grid operation with limited measurement points and limited communication.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and in particular relates to an edge computing fault detection method for offshore wind turbines during off-grid periods. Background Art

[0002] my country's offshore wind power generation is steadily expanding, with deep-sea areas emerging as a new frontier for offshore wind power development. Offshore wind turbine installation typically has a continuous operation window of less than three months per year. Due to more severe wind and sea conditions, the transportation and installation of deep-sea wind turbine equipment is challenging. A large-scale deep-sea wind power project typically requires turbine installation in batches across multiple windows, with commissioning and grid connection completed over many years. Consequently, the first batch of installed turbines will be off-grid for at least two years, creating a blind spot in the status of various deep-sea wind turbines during this period.

[0003] The operating patterns and fault characteristics of offshore wind turbines during off-grid operation differ significantly from those during grid-connected operation. During conventional off-grid operation, the wind turbine's drive system is locked, generating no power and accepting no external energy. During this time, the turbine's Supervisory Control and Data Acquisition (SCADA) system is inoperative and disconnected from the control center. When a tropical cyclone passes, the wind farm switches to typhoon mitigation mode, generating power from backup diesel generators to assist with active nacelle yaw control and mitigate the threat of the typhoon. During this time, the station's centralized control center monitors the turbine's real-time status via 4G networks or communication satellites. Compared to grid-connected turbines, off-grid turbines suffer from significant deficiencies in comprehensive and timely monitoring.

[0004] Wind turbine failures during off-grid operation are primarily due to blade vortex excitation and yaw misalignment. Blade vortex excitation, caused by fluid-structure interaction between wind and blades, induces significant blade vibration and increases the risk of blade fatigue. Yaw misalignment, on the other hand, increases bending and torsional loads on the turbine support structure during the passage of a tropical cyclone, potentially leading to serious failures such as tower bending. Due to the offline nature of SCADA systems and limited transmission channels, the centralized control center struggles to accurately and real-timely detect wind turbine failures during off-grid periods.

[0005] Existing technologies lack research on algorithms, solutions, and systems for off-grid wind turbine fault diagnosis. To ensure the safe and stable operation of deep-sea wind turbines during an off-grid period of at least two years, it is necessary to consider the potential failure modes of the turbines and design targeted condition monitoring solutions and fault detection methods to guide the control, regulation, and operation and maintenance of offshore wind farms during the off-grid period. Summary of the Invention

[0006] In order to solve the problems in the prior art, the present invention provides an edge computing fault detection method for offshore wind turbines during the off-grid period.

[0007] The technical solutions of the present invention are as follows:

[0008] The present invention discloses an edge computing fault detection method for an offshore wind turbine during an off-grid period, comprising: constructing a finite element simulation model of a support structure, a wind load calculation model of the turbine, and a computational fluid dynamics model of blade loads; obtaining multiple sets of historical wind field parameter data and yaw error angles, wherein the historical wind field parameter data include wind speed, wind direction, turbulence intensity, and wind shear; inputting the historical wind field parameter data and yaw error angle into the turbine wind load calculation model to obtain multiple input load spectra of the support structure under yaw misalignment, and then obtaining strain data of each point on the support structure through the finite element simulation model; selecting the strain data of the first k points with the largest discrimination index and constructing a first fault data set; inputting the wind speed, wind direction, and turbulence intensity into the wind load calculation model to obtain multiple input load spectra of the support structure under yaw misalignment, and then obtaining the strain data of each point on the support structure through the finite element simulation model; selecting the strain data of the first k points with the largest discrimination index and constructing a first fault data set; and ... A computational fluid dynamics model is input to obtain the blade vortex-induced vibration amplitude and the strain data of each blade root, and a second fault data set is constructed; strain sensors are arranged at the first k point positions and at the root of each blade, and a first measured data set and a second measured data set are constructed based on the actual strain data collected by the strain sensors; a spatial transformation matrix is ​​then obtained through a migration component analysis method; a machine learning model is trained based on the two fault data sets, the two measured data sets and the spatial transformation matrix to obtain a first detection model and a second detection model; finally, the first detection model is used to detect the yaw misalignment fault of the offshore wind turbine during the off-grid period, and the second detection model is used to detect the blade vortex-induced fault of the offshore wind turbine during the off-grid period.

[0009] Compared with the prior art, the present invention has the following beneficial effects:

[0010] The present invention is used for fault detection of offshore wind turbines during the off-grid period. It proposes a full-process solution covering data construction, model training, and model deployment for two types of faults that may occur in the turbine: yaw misalignment and blade vortex excitation.

[0011] Based on the analysis results of a finite element simulation model of the support structure, this paper optimizes the layout of strain sensors with the goal of maximizing the discrimination of strain data under different yaw error angles. By calculating the KL divergence between strain data under different simulation conditions, the deployed strain sensors are ensured to more effectively capture the fault characteristics caused by yaw misalignment. This process achieves an optimized layout of strain sensors, maximizing the accuracy and effectiveness of fault monitoring within the limited number of sensors.

[0012] This paper constructs a dataset covering various fault conditions through simulation, providing a sufficient sample base for subsequent fault warning models in the absence of field operating data. Transfer learning methods are used to map and align source domain simulation data and target domain field data into a feature space, minimizing data distribution differences between the source and target domains and ensuring that the model trained with simulation data is capable of performing field fault warning tasks.

[0013] This invention addresses the limited signal transmission issues of traditional monitoring methods during off-grid conditions. An edge-end signal acquisition card records strain data from the turbine and acquires wind farm parameter data from the station via a local area network. This data is then processed and fault diagnosis performed on-site by a single-chip microcomputer, ensuring real-time data transmission and processing while reducing reliance on the control center. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a schematic flow chart of the method involved in the present invention;

[0015] Figure 2 This is a schematic diagram of the system structure involved in the present invention;

[0016] Figure 3 is a flow chart of a method for optimizing the placement of strain sensors on a support structure of the present invention;

[0017] Figure 4 is a flow chart of a method for constructing a measured data set of the present invention;

[0018] Figure 5 It is a flow chart of a method for obtaining representations of a fault data set and a measured data set in a feature space according to the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described and illustrated below in conjunction with specific embodiments. The embodiments are merely illustrative of the present disclosure and do not limit its scope. The technical features of the various embodiments of the present invention may be combined accordingly, provided that there is no conflict between them.

[0020] The technical problem to be solved by the present invention is: fault detection of offshore wind turbines during the off-grid period. In order to solve the problem of limited transmission of status monitoring signals during the off-grid period, an edge computing fault detection method for offshore wind turbines during the off-grid period is proposed. The present invention obtains strain data of offshore wind turbines during the off-grid period through a limited number of optimally arranged strain sensors and constructs a measured data set, namely, a target domain data set; and implements fault simulation under various working conditions to obtain strain data at the location of the strain sensor and construct a fault data set, namely, a source domain data set; maps the source domain data set and the target domain data set to a new feature space, and trains a machine learning model in the new feature space using the source domain data set and the target domain data set; finally, the trained model is deployed in the edge terminal device of the offshore wind turbine during the off-grid period, and the edge terminal device performs real-time data acquisition and fault diagnosis on site, and sends fault information to the centralized control center.

[0021] like Figure 1 As shown, in order to solve the above technical problems, the present invention adopts the following technical solutions:

[0022] Step 1: Simulation model construction

[0023] Based on the design parameters of offshore wind turbines during the off-grid period, the following three simulation models are constructed:

[0024] The first model is a finite element simulation model of the offshore wind turbine's support structure. This model was constructed using finite element software. The input data used to construct the finite element simulation model included the dimensions and materials of the offshore wind turbine tower and the underlying support structure. The finite element simulation model aims to capture the support structure's response to external loads, specifically the strain data at each point on the support structure.

[0025] The second model is a wind load calculation model for the turbine. This model was constructed using the open-source aerodynamic simulation code OpenFAST. The wind load calculation model takes as input information the offshore wind turbine's blade profile, wind field parameters, and basic structural data for the offshore wind turbine's nacelle (including the mass and relative position of the nacelle's main shaft and gearbox). This wind load calculation model aims to derive the equivalent wind load acting on the supporting structure, i.e., to obtain multiple input load spectra for the supporting structure under yaw misalignment.

[0026] Model 3: A computational fluid dynamics model of offshore wind turbine blade loads. This model is constructed using computational fluid dynamics (CFD) methods. The model uses blade dimensions and materials, wind field parameters, and other information as input. The CFD model aims to capture the root load characteristics during vortex-induced vibrations (VIVs), specifically the VIV amplitudes and strain data at the root of each blade in offshore wind turbines.

[0027] Step 2: Yaw load spectrum calculation

[0028] Ideally, the plane of an offshore wind turbine's rotor faces the incoming wind, the nacelle's yaw error angle is 0 degrees, and the bending and torsional loads on the support structure are minimal. Therefore, to simulate nacelle yaw misalignment—that is, when the wind force is not perpendicular to the rotor plane—a simulation analysis was conducted using the turbine's wind load calculation model. The equivalent wind loads on the support structure were calculated at different yaw error angles, thereby obtaining the input load spectrum of the offshore wind turbine's support structure under yaw misalignment.

[0029] Specifically, multiple sets of historical wind field parameter data, including wind speed, wind direction, turbulence intensity, and wind shear, must first be obtained from an external wind station's wind tower. The angle between the wind direction and the normal to the impeller plane is then calculated to obtain multiple sets of yaw error angles. This wind field parameter data and each set of yaw error angles are then input into the turbine's wind load calculation model to obtain the input load spectrum for the support structure under yaw misalignment conditions. This means obtaining multiple input load spectra for the support structure under various yaw error angles. The input load spectrum is a curve showing how the input load (force) changes over time.

[0030] Step 3: Optimize the layout of strain sensors (i.e., install strain sensors on the bottom support structure)

[0031] like Figure 3 As shown in the figure, the goal of this stage is to optimize the layout of the strain sensors on the support structure, that is, to select the layout of the strain sensors on the support structure through simulation, so as to ensure that the strain data collected by the strain sensors have the greatest discrimination for different yaw error angles under the premise of a certain number of strain sensors.

[0032] To achieve the above objectives, the input load spectrum of the support structure under yaw misalignment obtained in step 2 is first used as the input of the finite element simulation model (i.e., the first model) of the support structure of the offshore wind turbine. The support structure is simulated and analyzed using the first model to obtain the layout position of the strain sensor on the support structure.

[0033] Specifically, the probability density function of the strain data of each point on the support structure under each input load spectrum is first calculated. The calculation formula is:

[0034] ;

[0035] in, Indicates that under the action of the jth input load spectrum, the point on the support structure Strain data The probability density function of Represents strain data The mean of Represents strain data The variance of .

[0036] Then calculate the KL divergence of the probability density function of the strain data of the same point on the support structure under each two different input load spectra. The calculation formula is:

[0037] ;

[0038] in, Indicates that under the action of the i-th input load spectrum, the point on the support structure Strain data The probability density function of .

[0039] KL divergence is used to measure the difference between two probability distributions. The larger the KL divergence value, the more obvious the difference between the two probability distributions. Therefore, in order to maximize the discrimination between different input loads, a discrimination index is constructed to define the sum of the KL divergences of all load conditions at each point on the support structure. In other words, the discrimination index of the point is obtained by summing up all the KL divergences at the same point. The formula for the discrimination index is:

[0040] ;

[0041] in, Indicates a point The discrimination index of Indicates the total number of input load spectra.

[0042] Finally, all points are sorted from highest to lowest by their discrimination index, and the strain data for the first k points are obtained. The locations of these k points are used as the locations for the strain sensors on the support structure. The number k is selected based on actual needs. In one specific embodiment of the present invention, k is 8.

[0043] Step 4: Simulation data set construction

[0044] Calculate the mean, variance, and spectral kurtosis of the strain data of the first k points obtained in step 3, and then obtain the wind speed, turbulence intensity, and yaw error angle of the wind field parameter data in step 2. Then, normalize the wind speed, turbulence intensity, yaw error angle, mean, variance, and spectral kurtosis to obtain the yaw misalignment fault data set.

[0045] The wind speed, wind direction, and turbulence intensity from the wind field parameter data obtained in step 2 are input into the blade-loaded computational fluid dynamics model to simulate and analyze the blade vortex-induced vibration (VIV) amplitude and strain data at each blade root. The mean, variance, and spectral kurtosis of the strain data at each blade root are then calculated. The wind speed, turbulence intensity, blade VIV amplitude, and the mean, variance, and spectral kurtosis of the strain data at each blade root are then normalized to obtain a blade VIV fault dataset.

[0046] Step 5: Building an edge monitoring system

[0047] like Figure 2 As shown, before the offshore wind turbine is installed, strain sensors are placed on the offshore wind turbine's support structure at locations corresponding to the k points in step 3, and strain sensors are also placed at the blade roots. An edge signal acquisition card and a single-chip microcomputer are then deployed in the offshore wind turbine's nacelle. The edge signal acquisition card receives strain data collected by the k strain sensors on the support structure and the strain data collected by the strain sensors at the blade roots. The single-chip microcomputer communicates with an external centralized control center via a communication satellite and with the wind tower at the site via a local area network.

[0048] like Figure 4 As shown, actual strain data collected by k strain sensors on the supporting structure of the offshore wind turbine during the off-grid period are obtained, and then the mean, variance, and spectral kurtosis of these k actual strain data are calculated. Then, the wind speed and turbulence intensity at the location of the offshore wind turbine are obtained. Then, the wind speed and turbulence intensity at the location of the offshore wind turbine and the mean, variance, and spectral kurtosis of the k actual strain data are normalized to obtain a target domain healthy dataset for yaw misalignment faults, that is, the first measured dataset.

[0049] The actual strain data collected by the strain sensor at the blade root of the offshore wind turbine during the off-grid period is obtained, and the mean, variance and spectral kurtosis of the actual strain data at the blade root are calculated. Then, the wind speed and turbulence intensity at the location of the offshore wind turbine and the mean, variance and spectral kurtosis of the actual strain data at the blade root are normalized to obtain the target domain healthy dataset of blade vortex-induced faults, that is, the second measured dataset.

[0050] Step 6: Feature Space Mapping

[0051] like Figure 5As shown in the figure, the source domain simulation data and the target domain health data are aligned in the new feature space through the transfer component analysis (TCA) method, where the source domain simulation data is the data in the yaw misalignment fault dataset or the blade vortex induced fault dataset; the target domain health data is the data in the target domain health dataset of the yaw misalignment fault or the blade vortex induced fault. Figure 5 The fault data set in is the yaw misalignment fault data set or the blade vortex induced fault data set, and the measured data set is the target domain health data set of the yaw misalignment fault or the target domain health data set of the blade vortex induced fault.

[0052] The maximum mean difference (MMD) is used to quantify the difference between the source domain simulation data and the target domain data in the feature space.

[0053] 6.1) The MMD calculation formula of the yaw misalignment fault dataset and the target domain healthy dataset in the feature space is:

[0054] ;

[0055] in, express and The maximum mean difference between represents the yaw misalignment fault dataset, i.e. the first fault dataset, , represents the ath sub-dataset in the yaw misalignment fault dataset, which includes a set of normalized wind speed, turbulence intensity, yaw error angle, mean, variance and spectral kurtosis; The target domain healthy dataset representing the yaw misalignment fault, i.e. the first measured dataset, , The bth sub-dataset in the target domain healthy data set representing the yaw misalignment fault, the bth sub-dataset includes a set of normalized wind speed, turbulence intensity, mean, variance and spectral kurtosis; Indicates the number of sub-datasets in the yaw misalignment fault dataset; Represents the feature mapping function, which is used to map the original data to the feature space; The number of sub-datasets in the target domain healthy dataset representing yaw misalignment faults; represents the norm in the feature space; Indicates traces; is the first kernel matrix; is the first weight matrix.

[0056] First core matrix The kernel function value of the data in the yaw misalignment fault dataset and the target domain healthy dataset containing the yaw misalignment fault is , the first kernel matrix is defined as follows:

[0057] ;

[0058] in, Represents a yaw misalignment fault dataset Yaw misalignment fault dataset The kernel function operation result between the data; Represents a yaw misalignment fault dataset Target domain healthy dataset with yaw misalignment fault The kernel function operation result between the data; Target domain healthy dataset representing yaw misalignment faults Yaw misalignment fault dataset The kernel function operation result between the data; Target domain healthy dataset representing yaw misalignment faults Target domain healthy dataset with yaw misalignment fault The kernel function operation result between the data.

[0059] is the MMD matrix representing the yaw misalignment fault dataset Target domain healthy dataset with yaw misalignment fault The relationship between the data is defined as follows:

[0060] ;

[0061] in, Represents a matrix with A rows and A columns and all elements are 1; Represents a matrix with A rows and B columns and all elements are 1; Represents a matrix with B rows and A columns and all elements are 1; Represents a matrix with B rows and B columns and all elements are 1.

[0062] In order to achieve the yaw misalignment fault dataset Target domain healthy dataset with yaw misalignment fault Alignment in the feature space requires minimizing their MMD values ​​in the feature space. Let the first feature space transformation matrix be , the optimization problem is equivalent to the following formula:

[0063] ;

[0064] in, is the first regularization term; Indicates weight, which is 0-1; is the first centralized matrix, and its expression is:

[0065] ;

[0066] in, The number of rows and columns are both The identity matrix of , whose main diagonal elements are 1 and the rest of the elements are 0; Indicates that the number of rows and columns are both A matrix whose elements are all 1.

[0067] The first effective solution obtained by solving the above problem through optimization algorithm is , Then for the first effective solution Perform singular value decomposition, select the largest first several singular values ​​in the diagonal matrix and select the column vectors corresponding to each singular value from the matrix on the left side of the diagonal matrix, and select the row vectors corresponding to each singular value from the matrix on the right side of the diagonal matrix. Finally, multiply the matrix composed of all row vectors, the matrix composed of all singular values, and the matrix composed of all column vectors to obtain the first spatial transformation matrix .

[0068] Obtain the first spatial transformation matrix Then, the yaw misalignment fault dataset The data and the first spatial transformation matrix Multiply them together to obtain the representation of the yaw misalignment fault dataset in the feature space. And the target domain healthy dataset of the yaw misalignment fault The data in the first space transformation matrix Multiply them together to obtain the representation of the target domain healthy dataset of yaw misalignment fault in the feature space.

[0069] 6.2) The MMD calculation formula of the blade vortex-induced fault dataset and the target domain healthy dataset of the blade vortex-induced fault in the feature space is:

[0070] ;

[0071] in, express and The maximum mean difference between represents the blade vortex-induced fault data set, i.e., the second fault data set, , represents the cth sub-dataset in the blade vortex-induced fault dataset, which includes a set of normalized wind speed, turbulence intensity, blade vortex-induced vibration amplitude, mean, variance and spectral kurtosis; The target domain healthy data set representing the blade vortex-induced fault, i.e. the second measured data set, , The dth sub-dataset in the target domain healthy data set representing blade vortex-induced fault includes a set of normalized wind speed, turbulence intensity, mean, variance, and spectral kurtosis; Indicates the number of sub-datasets in the blade vortex-induced fault dataset; The number of sub-datasets in the target domain healthy dataset representing blade vortex-induced faults; is the second kernel matrix; is the second weight matrix.

[0072] Second core matrix Contains the kernel function values ​​of the data in the blade vortex induced failure dataset and the target domain healthy dataset of the blade vortex induced failure. , the second kernel matrix is defined as follows:

[0073] ;

[0074] in, Represents the blade vortex-induced failure dataset and blade vortex-induced failure dataset The kernel function operation result between the data; Represents the blade vortex-induced failure dataset Target domain health dataset related to blade vortex-induced failure The kernel function operation result between the data; Target domain health dataset representing blade vortex-induced failure and blade vortex-induced failure dataset The kernel function operation result between the data; Target domain health dataset representing blade vortex-induced failure Target domain health dataset related to blade vortex-induced failure The kernel function operation result between the data.

[0075] is the MMD matrix, representing the blade vortex-induced failure data set Target domain health dataset related to blade vortex-induced failure The relationship between the data is defined as follows:

[0076] ;

[0077] in, represents a matrix with C rows and C columns and all elements are 1; represents a matrix with C rows and D columns and all elements are 1; represents a matrix with D rows and C columns and all elements are 1; Represents a matrix with D rows and D columns and all elements are 1.

[0078] In order to realize the blade vortex induced failure data set Target domain health dataset related to blade vortex-induced failure Alignment in the feature space requires minimizing their MMD values ​​in the feature space. Let the second feature space transformation matrix be , the optimization problem is equivalent to the following formula:

[0079] ;

[0080] in, is the second regularization term; is the second centralized matrix, and its expression is:

[0081] ;

[0082] in, The number of rows and columns are both The identity matrix of , whose main diagonal elements are 1 and the rest of the elements are 0; Indicates that the number of rows and columns are both A matrix whose elements are all 1.

[0083] Solve the above problem through optimization algorithm and get the second effective solution , Then for the second effective solution Perform singular value decomposition, select the largest first several singular values ​​in the diagonal matrix and select the column vectors corresponding to each singular value from the matrix on the left side of the diagonal matrix, and select the row vectors corresponding to each singular value from the matrix on the right side of the diagonal matrix. Finally, multiply the matrix composed of all row vectors, the matrix composed of all singular values, and the matrix composed of all column vectors to obtain the second space transformation matrix .

[0084] Obtain the second space transformation matrix Then, the blade vortex induced failure data set The data and the second space transformation matrix Multiply them together to obtain the representation of the blade vortex-induced failure dataset in the feature space. And the target domain healthy dataset of the blade vortex-induced failure is The data in the second space transformation matrix Multiply them together to obtain the representation of the target domain healthy dataset of blade vortex-induced failure in the feature space.

[0085] Step 7: Fault diagnosis model training

[0086] 7.1) Concatenate the representation of the yaw misalignment fault dataset in feature space with the representation of the target domain healthy dataset in feature space. The concatenated dataset is used as a training set and input into a machine learning model for training on yaw misalignment fault detection to obtain a yaw misalignment fault detection model.

[0087] When training a model for yaw misalignment fault detection, the machine learning model inputs are the normalized mean, variance, and spectral kurtosis of strain data at k points, as well as wind speed and turbulence intensity. The yaw error angle from the yaw misalignment fault dataset serves as the label, and the final output of the machine learning model is the predicted yaw error angle. The root mean square error function is used as the loss function during training, and the machine learning model can be a convolutional neural network or support vector regression model.

[0088] 7.2) The representation of the blade vortex-induced fault dataset in the feature space and the representation of the target domain healthy dataset of the blade vortex-induced fault in the feature space are concatenated. The concatenated dataset is used as a training set and input into a machine learning model for training blade vortex-induced fault detection to obtain a blade vortex-induced fault detection model.

[0089] For training a model for blade vortex-induced fault detection, the machine learning model inputs are the normalized mean, variance, and spectral kurtosis of blade root strain data, as well as wind speed and turbulence intensity. The blade vortex-induced vibration amplitude from the blade vortex-induced fault dataset serves as the label. The final output of the machine learning model is the predicted blade vortex-induced vibration amplitude. The root mean square error (RMS) function is used as the loss function during training, and the machine learning model can be a convolutional neural network or support vector regression model.

[0090] Step 8: On-site fault diagnosis

[0091] After the yaw misalignment fault detection model and blade vortex fault detection model are trained in the centralized control center, they are transmitted via communication satellite and downloaded to the edge microcontroller.

[0092] During the off-grid operation of an offshore wind turbine, the microcontroller receives and summarizes the strain data collected by k strain sensors installed on the supporting structure and the strain data collected by the strain sensors installed at the root of the blades. At the same time, the microcontroller receives and summarizes the wind speed and turbulence intensity at the location of the offshore wind turbine input by the external station, and then uses the yaw misalignment fault detection model and the blade vortex-induced vibration fault detection model to obtain the yaw error angle and blade vortex-induced vibration amplitude respectively.

[0093] When the yaw error angle output by the yaw misalignment fault detection model exceeds a threshold, the single chip microcomputer sends the result output by the yaw misalignment fault detection model to the centralized control center via a communication satellite and issues a fault alarm.

[0094] When the blade vortex-induced vibration amplitude output by the blade vortex-induced fault detection model exceeds the threshold, the microcontroller sends the output result of the blade vortex-induced fault detection model to the centralized control center via the communication satellite and issues a fault alarm.

[0095] During the above process, the centralized control center also regularly obtains the stored data from the edge end, optimizes the yaw misalignment fault detection model and the blade vortex fault detection model based on the acquired data, and sends it back to the edge end to realize the update of the on-site model.

[0096] The present invention is used for fault detection of offshore wind turbines during the off-grid period. It proposes a full-process solution covering data construction, model training, and model deployment for two types of faults that may occur in the turbine: yaw misalignment and blade vortex excitation. Through model simulation and data migration, sufficient training samples are constructed for the fault detection model; strain, wind speed, and turbulence signals with low sampling frequency are used as input signals of the model to ensure lightweight data during the working process; the model is deployed in the edge device to collect data and diagnose faults independently, effectively reducing the cost of data transmission. The present invention can realize status monitoring and fault diagnosis of offshore wind turbines during off-grid operation with limited measurement points and restricted communications.

[0097] The above-described embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. Persons skilled in the art will readily appreciate that variations and modifications may be made without departing from the scope of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A method for edge computing fault detection of offshore wind turbines during off-grid period, characterized in that: include: Construct a finite element simulation model of the supporting structure, a wind load calculation model for the unit, and a computational fluid dynamics model of blade loads; Obtain multiple sets of historical wind field parameter data and yaw error angles. The historical wind field parameter data include wind speed, wind direction, turbulence intensity and wind shear; Input historical wind field parameter data and yaw error angle into the wind load calculation model of the turbine unit to obtain multiple input load spectra of the support structure under yaw misalignment conditions. Then, the strain data of each point on the support structure is obtained through the finite element simulation model. The strain data of the first k points with the largest discrimination index are selected and a first fault data set is constructed; the wind speed, wind direction, and turbulence intensity are input into a computational fluid dynamics model to obtain the blade vortex-induced vibration amplitude and the strain data of each blade root, and a second fault data set is constructed; strain sensors are deployed at the positions of the first k points and at the root of each blade, and the first measured data set and the second measured data set are constructed based on the actual strain data collected by the strain sensors; Then the spatial transformation matrix is ​​obtained through the migration component analysis method; A machine learning model is trained based on two fault datasets, two measured datasets, and a spatial transformation matrix to obtain a first detection model and a second detection model. Finally, the first detection model is used to detect yaw misalignment faults in offshore wind turbines during the off-grid period, and the second detection model is used to detect blade vortex-induced faults in offshore wind turbines during the off-grid period. The step of selecting strain data of the first k points with the largest discrimination index and constructing a first fault data set includes: First, the probability density function of the strain data of each point on the support structure under each input load spectrum is calculated. Then, the KL divergence of the probability density function of the strain data of the same point on the support structure under each two different input load spectra is calculated. Then, all the KL divergences of the same point are accumulated to obtain the discrimination index of the point. The discrimination indexes of all points are sorted from large to small and the strain data of the first k points are obtained. Next, the mean, variance, and spectral kurtosis of the strain data of the k points are calculated, and then the mean, variance, spectral kurtosis, the obtained yaw error angle, and the wind speed and turbulence intensity in the historical wind field parameter data are normalized to obtain a first fault data set; The constructing of the second fault data set includes: The mean, variance, and spectral kurtosis of the strain data at the root of each blade are calculated. Then, the blade vortex-induced vibration amplitude, the mean, variance, and spectral kurtosis of the strain data at the root of each blade, and the wind speed and turbulence intensity in the historical wind field parameter data are normalized to obtain the second fault data set.

2. The edge computing fault detection method for offshore wind turbines during off-grid period according to claim 1 is characterized in that: Constructing a first measured data set includes: obtaining actual strain data collected by strain sensors at the first k point locations, calculating a mean, variance, and spectral kurtosis of the k actual strain data, obtaining a wind speed and turbulence intensity at a location where the offshore wind turbine is located, and then normalizing the wind speed and turbulence intensity at the location where the offshore wind turbine is located and the mean, variance, and spectral kurtosis of the k actual strain data to obtain the first measured data set; Constructing a second measured data set includes: obtaining actual strain data collected by a strain sensor at the blade root, calculating the mean, variance, and spectral kurtosis of the actual strain data at the blade root, and then normalizing the wind speed and turbulence intensity at the location of the offshore wind turbine and the mean, variance, and spectral kurtosis of the actual strain data at the blade root to obtain the second measured data set.

3. The edge computing fault detection method for offshore wind turbines during the off-grid period according to claim 2 is characterized in that: The spatial transformation matrix includes a first submatrix and a second submatrix, and a method for obtaining the first submatrix is ​​as follows: first, quantifying the difference between the data in the first fault data set and the data in the first measured data set in the feature space to obtain the maximum average difference between the first fault data set and the first measured data set; then, with the goal of minimizing the maximum average difference, obtaining a first valid solution; next, performing singular value decomposition on the first valid solution, selecting the first several largest singular values ​​in the diagonal matrix and selecting the column vectors corresponding to each singular value from the matrix on the left side of the diagonal matrix, and selecting the row vectors corresponding to each singular value from the matrix on the right side of the diagonal matrix; finally, multiplying the matrix composed of all row vectors, the matrix composed of all singular values, and the matrix composed of all column vectors to obtain the first submatrix; The method for obtaining the second submatrix is: first, quantify the difference between the data in the second fault data set and the data in the second measured data set in the feature space to obtain the maximum average difference between the second fault data set and the second measured data set; then, with the goal of minimizing the maximum average difference, obtain the second valid solution, and then perform singular value decomposition on the second valid solution, select the first several largest singular values ​​in the diagonal matrix and select the column vectors corresponding to each singular value from the matrix on the left side of the diagonal matrix, and select the row vectors corresponding to each singular value from the matrix on the right side of the diagonal matrix, and finally multiply the matrix composed of all row vectors, the matrix composed of all singular values, and the matrix composed of all column vectors to obtain the second submatrix.

4. The edge computing fault detection method for offshore wind turbines during off-grid period according to claim 3 is characterized in that: The calculation formula for the maximum average difference between the first fault data set and the first measured data set is: ； in, express and The maximum mean difference between represents the first fault data set, , represents the ath sub-dataset in the first fault data set, which includes a set of normalized wind speed, turbulence intensity, yaw error angle, mean, variance and spectral kurtosis; represents the first measured data set, , represents the bth sub-dataset in the first measured data set, which includes a set of normalized wind speed, turbulence intensity, mean, variance and spectral kurtosis; Indicates the number of sub-datasets in the first fault data set; represents the feature mapping function; Indicates the number of sub-datasets in the first measured data set; represents the norm in the feature space; Indicates traces; is the first kernel matrix, , Represents the first fault data set With the first fault data set The kernel function operation result between the data; Represents the first fault data set With the first measured data set The kernel function operation result between the data; Represents the first measured data set With the first fault data set The kernel function operation result between the data; Represents the first measured data set With the first measured data set The kernel function operation result between the data; is the first weight matrix, , represents a matrix with A rows and A columns and all elements are 1, represents a matrix with A rows and B columns and all elements are 1, represents a matrix with B rows and A columns and all elements are 1, represents a matrix with B rows and B columns and all elements are 1; The calculation formula for obtaining the first effective solution with the goal of minimizing the maximum average difference is: ； in, is the first regularization term, represents the weight, is the first eigenspace transformation matrix; is the first centralized matrix, , The number of rows and columns are both The identity matrix of Indicates that the number of rows and columns are both A matrix whose elements are all 1.

5. The edge computing fault detection method for offshore wind turbines during off-grid period according to claim 4 is characterized in that: The formula for the maximum average difference between the second fault data set and the second measured data set is: ； in, express and The maximum mean difference between represents the second fault data set, , represents the cth sub-dataset in the second fault data set, which includes a set of normalized wind speed, turbulence intensity, blade vortex-induced vibration amplitude, mean, variance and spectral kurtosis; represents the second measured data set, , represents the dth sub-dataset in the second measured data set, and the dth sub-dataset includes a set of normalized wind speed, turbulence intensity, mean, variance and spectral kurtosis; Indicates the number of sub-datasets in the second fault data set; Indicates the number of sub-datasets in the second measured data set; is the second kernel matrix, , Represents the second fault data set With the second fault data set The kernel function operation result between the data; Represents the second fault data set With the second measured data set The kernel function operation result between the data; Represents the second measured data set With the second fault data set The kernel function operation result between the data; Represents the second measured data set With the second measured data set The kernel function operation result between the data; is the second weight matrix, , represents a matrix with C rows and C columns and all elements are 1; represents a matrix with C rows and D columns and all elements are 1; represents a matrix with D rows and C columns and all elements are 1; represents a matrix with D rows and D columns and all elements are 1; The calculation formula for obtaining the second effective solution with the goal of minimizing the maximum average difference is: ； in, is the second regularization term; is the second eigenspace transformation matrix; is the second centralized matrix, , The number of rows and columns are both The identity matrix of Indicates that the number of rows and columns are both A matrix whose elements are all 1.

6. The edge computing fault detection method for offshore wind turbines during off-grid period according to claim 5, characterized in that: The method includes training a machine learning model based on two fault data sets, two measured data sets and a spatial transformation matrix to obtain a first detection model and a second detection model; Multiplying the first fault dataset and the first measured dataset by the first submatrix respectively to obtain a representation of the first fault dataset in the feature space and a representation of the first measured dataset in the feature space, concatenating the representation of the first fault dataset in the feature space and the representation of the first measured dataset in the feature space, and inputting the concatenated data into a machine learning model as training data to obtain a first detection model; wherein, during the training process, a root mean square error function is used as a loss function, and a yaw error angle is used as a label; The second fault data set and the second measured data set are multiplied by the second sub-matrix respectively to obtain the representation of the second fault data set in the feature space and the representation of the second measured data set in the feature space, and the representation of the second fault data set in the feature space and the representation of the second measured data set in the feature space are spliced. After splicing, they are input into the machine learning model as training data to obtain the second detection model; wherein, the root mean square error function is used as the loss function during the training process, and the blade vortex-induced vibration amplitude is used as the label.

7. The edge computing fault detection method for offshore wind turbines during off-grid period according to claim 6, characterized in that: The method of using the first detection model to detect yaw misalignment faults of offshore wind turbines during the off-grid period, and using the second detection model to detect blade vortex-induced faults of offshore wind turbines during the off-grid period, comprises: Deploying both the first detection model and the second detection model on an offshore wind turbine during an off-grid period; obtaining actual strain data collected by strain sensors at the first k points and actual strain data collected by strain sensors at the roots of each blade; and simultaneously obtaining wind speed and turbulence intensity at the location of the offshore wind turbine from an external station; The actual strain data, wind speed, and turbulence intensity collected by the strain sensors at the first k points are input into the first detection model to obtain a yaw error angle; if the yaw error angle exceeds a threshold, a yaw misalignment fault alarm is issued; otherwise, no processing is performed; The actual strain data, wind speed and turbulence intensity collected by the strain sensors at the roots of each blade are input into the second detection model to obtain the blade vortex-induced vibration amplitude; if the blade vortex-induced vibration amplitude exceeds the threshold, a blade vortex-induced fault alarm is issued; otherwise, no processing is performed.

Images