A blade condition detection method and apparatus

Abstract

The application provides a kind of blade state detection method and device, it is related to wind power generator operation and maintenance field, to real-time, accurately monitor and evaluate the state of blade.The method comprises: obtaining real-time measurement data, a real-time measurement data includes the data of each dimension sampled at each sampling time point in a sampling time range;The real-time measurement data is input into the mapping network to obtain the data representation of the real-time measurement data, wherein the mapping network is trained using health data as training samples, and the health data refers to measurement data collected when the blade is in a healthy state;Obtain the data representation of the health data, which is obtained using the mapping network;According to the similarity between the data representation of the real-time measurement data and the data representation of the health data, determine the state of the blade.

Description

Technical Field

[0001] This application relates to the field of wind turbine operation and maintenance, and in particular to a method and device for blade condition detection. Background Technology

[0002] Wind power generation is one of the most technologically mature and promising renewable energy utilization methods. Vigorously developing wind power generation is of great significance for alleviating the energy crisis and solving environmental problems. Wind turbines are crucial equipment in wind power generation, comprising blades, towers, transmission systems, power generation systems, pitch systems, and yaw systems. Among all components of a wind turbine, the blades can be considered vital pilot components. They directly contact the external wind, using their aerodynamic shape to convert wind energy into mechanical energy, which then drives the transmission and power generation systems, ultimately converting mechanical energy into electrical energy. Therefore, the service condition of the blades has a decisive impact on the overall operation of the wind turbine.

[0003] Due to their large size and complex structure, wind turbine blades have a high failure rate, and the prolonged downtime caused by blade failures can result in significant power generation losses. Severe blade failures can even lead to blade breakage, causing injuries and fatalities, posing a major threat to production safety. Currently, wind turbine blade failure prevention still largely relies on regular manual inspections. However, this method is labor-intensive, involves hazardous working environments, and cannot provide real-time status monitoring during blade operation, exhibiting a significant time lag. Therefore, a method is needed to monitor blade status in real time, detect early signs of blade failure, avoid catastrophic consequences from severe failures, and ensure the efficiency and safety of wind power generation.

[0004] Most wind turbines are currently equipped with Supervisory Control and Data Acquisition (SCADA) systems, which use various sensors such as temperature, pressure, and vibration to monitor the turbine's operating status and remotely implement control functions such as pitch and yaw. Therefore, SCADA measurement data provides a foundation for monitoring the condition of wind turbine blades. Data-driven pattern recognition methods can be used to identify blade health conditions. Data-driven methods based on a large amount of historical measurement data reduce reliance on expert knowledge related to fault mechanisms, which is beneficial for wind power engineers to master and apply. However, the long-term operation of wind turbines with blade failures poses safety risks. Therefore, SCADA measurement data under fault conditions is scarce compared to healthy measurement data. This lack of fault measurement data makes it difficult for data-driven models to learn information related to fault states during training, and it cannot guarantee the separability of measurement data under different health states, thus affecting the accuracy of blade health condition identification. Therefore, current technologies do not consider the interference of the scarcity of blade fault measurement data, making it impossible to monitor and assess the health condition of wind turbine blades in real time and accurately based on SCADA measurement data. Summary of the Invention

[0005] In view of the above problems, embodiments of the present invention provide a blade state detection method and apparatus to overcome or at least partially solve the above problems.

[0006] A first aspect of the present invention provides a blade state detection method, comprising:

[0007] Acquire real-time measurement data, which includes data from each dimension sampled at each sampling time point within a sampling time range;

[0008] The real-time measurement data is input into a mapping network to obtain a data representation of the real-time measurement data. The mapping network is trained using healthy data as training samples. The healthy data refers to the measurement data collected when the leaf is in a healthy state.

[0009] The data representation of the health data is obtained using the mapping network;

[0010] The state of the leaf is determined based on the similarity between the data representation of the real-time measurement data and the data representation of the health data.

[0011] Optionally, determining the state of the leaf based on the similarity between the data representation of the real-time measurement data and the data representation of the health data includes:

[0012] Calculate the mean vector and covariance matrix of the data representation of the health data;

[0013] Based on the data representation of the real-time measurement data, the mean vector, and the covariance matrix, calculate the first distance between the data representation of the real-time measurement data and the data representation of the health data;

[0014] The state of the blade is determined based on the first distance.

[0015] Optionally, it also includes:

[0016] Acquire a second distance between the data representations of multiple historical measurement data and the data representations of the health data;

[0017] Based on multiple second distances, an exponentially weighted moving average control chart is established;

[0018] The warning distance is determined based on the exponentially weighted moving average control chart.

[0019] Determining the state of the blade based on the first distance includes:

[0020] If the first distance is not greater than the warning distance, the blade is determined to be in a healthy state;

[0021] If the first distance is greater than the warning distance, the blade is determined to be in an unhealthy state.

[0022] Optionally, the mapping network is trained through the following steps:

[0023] Obtain multiple sets of the aforementioned health data;

[0024] Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point.

[0025] Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data;

[0026] The mapping network to be trained is obtained by performing self-supervised training based on the data representation of each of the augmented data.

[0027] Optionally, acquiring multiple health data includes:

[0028] When the blade is in a healthy state, for each dimension, data sampled at each sampling time point within multiple sampling time ranges are acquired;

[0029] Filter out invalid data from the data to obtain multiple remaining data for each dimension within multiple sampling time ranges;

[0030] Eliminate the influence of the data volume of multiple remaining data in each dimension within multiple sampling time ranges to obtain multiple target data in each dimension within multiple sampling time ranges;

[0031] Based on each sampling time point within each sampling time range, multiple target data points for each dimension within that sampling time range are stacked to obtain multiple health data points.

[0032] Optionally, the step of performing self-supervised training on the mapping network to be trained based on the data representation of each of the augmented data to obtain the mapping network includes:

[0033] The data representation of each augmented data is input into the projection network to obtain the projected data representation of each augmented data.

[0034] The projected data representation of each augmented data is input into the prediction network to obtain the prediction result of whether each two augmented data are the same health data.

[0035] A loss function is established based on the multiple prediction results and the projected data representation of each of the augmented data;

[0036] Based on the loss function, the mapping network to be trained is trained to obtain the mapping network.

[0037] Optionally, establishing a loss function based on the multiple prediction results and the projected data representation of each augmented data includes:

[0038] Based on the multiple prediction results and the projected data representation of each of the augmented data, the negative cosine similarity is calculated;

[0039] The loss function is established based on the negative cosine similarity.

[0040] Optionally, the two random data augmentation processes include, but are not limited to, any two of the following: amplitude axis flipping, variable axis flipping, truncation, adding noise, cropping adjustment, time wrapping, and smoothing.

[0041] A second aspect of the present invention provides a blade condition detection device, comprising:

[0042] The data acquisition module is used to acquire real-time measurement data. A real-time measurement data includes data of each dimension sampled at each sampling time point within a sampling time range.

[0043] The first data representation module is used to input the real-time measurement data into the mapping network to obtain the data representation of the real-time measurement data. The mapping network is trained using healthy data as training samples. The healthy data refers to the measurement data collected when the leaf is in a healthy state.

[0044] The second data representation module is used to obtain a data representation of the health data, which is obtained using the mapping network.

[0045] The state determination module is used to determine the state of the blade based on the similarity between the data representation of the real-time measurement data and the data representation of the health data.

[0046] Optionally, the mapping network is trained through the following steps:

[0047] Obtain multiple sets of the aforementioned health data;

[0048] Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point.

[0049] Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data;

[0050] The mapping network to be trained is obtained by performing self-supervised training based on the data representation of each of the augmented data.

[0051] The embodiments of the present invention have the following advantages:

[0052] In this embodiment, real-time measurement data can be acquired. A single real-time measurement data set includes data from each dimension sampled at various sampling time points within a sampling time range. This real-time measurement data is input into a mapping network to obtain its data representation. The mapping network is trained using healthy data as training samples, where healthy data refers to measurement data collected when the blade is in a healthy state. The data representation of the healthy data is obtained using the mapping network. Based on the similarity between the data representation of the real-time measurement data and the data representation of the healthy data, the state of the blade is determined. Thus, the mapping network only needs to output the data representation of the data without directly determining the blade's state. Therefore, the mapping network can be trained using only healthy data, avoiding the drawback of the model struggling to learn information related to the fault state during training due to the scarcity of fault measurement data. By comparing the similarity between the data representation of the real-time measurement data generated through the same mapping network and the data representation of the healthy data, the similarity between the current state and the healthy state of the blade can be determined, thereby determining the blade's state. Furthermore, without the need for regular manual inspections, real-time measurement data can be directly acquired, and a data representation of the real-time measurement data can be quickly generated through a mapping network. Therefore, the condition of the blades can be quickly determined, thereby enabling real-time and accurate monitoring and evaluation of the blade condition. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a flowchart of the steps of a blade condition detection method in an embodiment of the present invention;

[0055] Figure 2 This is a structural diagram of the mapping network, projection network, and prediction network in an embodiment of the present invention;

[0056] Figure 3 This is a flowchart illustrating the blade state detection method in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the structure of a blade condition detection device according to an embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] Reference Figure 1 As shown, a flowchart illustrating the steps of a blade state detection method according to an embodiment of the present invention is presented, as follows: Figure 1 As shown, the blade condition detection method may specifically include the following steps:

[0060] Step S11: Obtain real-time measurement data. Real-time measurement data includes data of each dimension sampled at each sampling time point within a sampling time range.

[0061] Real-time measurement data is data collected by the SCADA system and can be stored. A single sampling channel of the SCADA system samples data in one dimension. This dimension includes at least one or more of the following: wind speed, wind direction, ambient temperature, generator speed, power output, lubricating oil temperature, etc. A real-time measurement dataset includes data from each dimension sampled at various sampling time points within the most recent sampling time range. Optionally, the data from each dimension sampled at various sampling time points within a sampling time range can be stacked to obtain a two-dimensional matrix, which represents a real-time measurement dataset. One of the row and column dimensions of this two-dimensional matrix can represent the data sampled at each sampling time point, and the other can represent the data in each dimension.

[0062] For example, a sampling time range is half an hour, with each sampling time point spaced 10 minutes apart. That is, sampling will be performed at the 10th, 20th, and 30th minutes of the sampling time range. At each sampling time point, each sampling channel of the SCADA system will perform one sampling. If there are i channels, 3×i data points can be collected within a sampling time range, and these 3×i data points together constitute a real-time measurement. The blade state usually changes gradually, and the change in the blade state within a sampling time range is negligible. Therefore, although a real-time measurement is actually the data from the end of the previous sampling time range (the beginning of the current sampling time range) to the end of the current sampling time range over the past 30 minutes, it can still be considered real-time data. Optionally, the size of the sampling time range and the interval between sampling time points can be set according to actual needs.

[0063] Because the data volumes differ across dimensions, normalization can be applied to each dimension to eliminate the influence of data volume. Normalization can be performed using the following formula:

[0064]

[0065] Where, x iThis represents the data in the i-th dimension; i = 1, 2, ..., I; I represents the total number of dimensions; the superscript "0" represents the original data before normalization, μ represents the mean, and σ represents the standard deviation.

[0066] Step S12: Input the real-time measurement data into the mapping network to obtain the data representation of the real-time measurement data, wherein the mapping network is trained using healthy data as training samples, and the healthy data refers to the measurement data collected when the leaf is in a healthy state.

[0067] Health data also includes data from various dimensions sampled at different time points within a sampling time range. Health data can be collected through a SCADA system. A mapping network trained using health data as training samples can accurately output data representations of each measurement based on the input measurement data. Inputting real-time measurement data into the pre-trained mapping network allows the network to directly output a data representation of that real-time measurement data. Optionally, the data representation can be a vector.

[0068] Step S13: Obtain the data representation of the health data, which is obtained using the mapping network.

[0069] The health data representation can be obtained by inputting the health data into a trained mapping network. This representation can then be stored. When it's necessary to determine the blade's condition based on real-time measurement data, the stored health data representation can be directly retrieved.

[0070] Step S14: Determine the state of the leaf based on the similarity between the data representation of the real-time measurement data and the data representation of the health data.

[0071] The similarity between the real-time measurement data representation and the health data representation can be determined by calculating the distance between them. A higher similarity between the real-time measurement data representation and the health data representation indicates a higher similarity between the blade's state and its health state; therefore, the blade's state can be determined based on this similarity. The blade can be a wind turbine blade.

[0072] The technical solution adopted in this application embodiment only requires the mapping network to output the data representation of the data, without directly determining the state of the blade. Therefore, the mapping network can be trained using only health data, avoiding the disadvantage that the scarcity of fault measurement data makes it difficult for the model to learn information related to the fault state during training. By comparing the similarity between the data representation of real-time measurement data generated through the same mapping network and the data representation of health data, the similarity between the current state and the health state of the blade can be determined, thereby determining the state of the blade. In addition, there is no need for regular manual inspections; real-time measurement data can be directly acquired, and the data representation of the real-time measurement data can be quickly generated through the mapping network. Therefore, the state of the blade can be quickly determined, thereby achieving real-time and accurate monitoring and evaluation of the blade's state.

[0073] Optionally, the mean vector and covariance matrix of the health data representation can be calculated. Based on the data representation of the real-time measurement data, and the mean vector and covariance matrix of the health data representation, a first distance between the data representations of the real-time measurement data and the health data can be calculated. This first distance can be a Mahalanobis distance. The similarity between the data representations of the real-time measurement data and the health data can be determined based on this first distance, thereby determining the state of the blade.

[0074] The first distance between the data representation of real-time measurement data and the data representation of health data can be determined using the following formula:

[0075]

[0076] Where S is the first distance, y r The data representation characterizes real-time measurement data, where T is the total number of sampling time points, and μ h The mean vector, Σ, representing the data representation of health data. h The covariance matrix representing the data representation of health data.

[0077] Thus, the mean vector of the health data representation can synthesize the characteristics of each data point, and the covariance matrix of the health data representation can characterize the correlation of data in each dimension. Therefore, the first distance calculated based on the mean vector and covariance matrix of the health data representation can accurately measure the similarity between the data representation of real-time measurement data and the data representation of health data, thereby accurately determining the state of the blade.

[0078] Optionally, based on the above technical solution, when determining whether a leaf is healthy according to a first distance, the size of the first distance and the warning distance can be compared. If the first distance is not greater than the warning distance, the leaf is determined to be in a healthy state; if the first distance is greater than the warning distance, the leaf is determined to be in an unhealthy state.

[0079] The warning distance can be a control chart based on an exponentially weighted moving average established from multiple second distances, where the second distance is the distance between the data representation of historical measurement data and the data representation of health data. The method for obtaining the second distance between the data representation of historical measurement data and the data representation of health data can refer to the method for obtaining the first distance, and will not be elaborated here.

[0080] An exponentially weighted moving average control chart can be created using the following formula:

[0081] EWMA(t)=γS(t)+(1-γ)EWMA(t-1)

[0082] Where t = 1, 2, ...; S(t) represents the second distance between the data representation of the t-th historical measurement data and the data representation of the health data; EWMA(t) represents the statistic of the exponentially weighted moving average control chart, and EWMA(0) represents the average value of S(t); 0 < γ ≤ 1 is the memory parameter that controls the weights of short-term and long-term memory.

[0083] The upper limit of the warning distance can be determined by the following formula:

[0084] UCL=EWMA(0)+3σ EWMA

[0085] Wherein, UCL represents the upper control limit of the warning distance, σ EWMA This represents the standard deviation of the variance in the EWMA statistic.

[0086] Thus, the exponentially weighted moving average control chart incorporates information from all historical measurement data, ensuring that smaller or larger values ​​do not significantly affect the chart during calculation. By varying the weights used and the amount of historical measurement data for control limits, a control chart can be constructed that can detect shifts of almost any size during the process, thus enabling the detection of even small shifts and accurate assessment of the blade's condition.

[0087] Alternatively, based on the above technical solution, the mapping network can be trained through the following steps:

[0088] Step S21: Obtain multiple health data points.

[0089] Multiple health data can be obtained through the following steps: when the leaf is in a healthy state, for each dimension, acquire data sampled at each sampling time point within multiple sampling time ranges; filter invalid data in the data to obtain multiple remaining data for each dimension within multiple sampling time ranges; eliminate the influence of the data volume of the multiple remaining data for each dimension within multiple sampling time ranges to obtain multiple target data for each dimension within multiple sampling time ranges; stack the multiple target data for each dimension within each sampling time range according to each sampling time point within each sampling time range to obtain multiple health data.

[0090] Invalid data refers to data generated when the wind turbine is operating at its limited power or outside the cut-in / cut-out wind speed range.

[0091] The impact of data volume refers to the fact that the value ranges of data in different dimensions may differ. For example, the value range of data in one dimension may be 0 to 10, while the value range of data in another dimension may be 0 to 1000. To eliminate the impact of data volume, the data in each dimension can be normalized to ensure that the value range of each dimension is 0 to 1, and that the processed data in each dimension conforms to a normal distribution, thus avoiding the influence of different data volumes on the representation of health data.

[0092] After eliminating the impact of data volume, the target data can be obtained. Stacking multiple target data points for each dimension within a sampling time range according to their sampling time points yields health data. Health data can be a two-dimensional matrix, where one of the row and column dimensions represents the data sampled at each sampling time point, and the other represents the data for each dimension.

[0093] Step S22: Perform two random data augmentation processes on each of the health data to obtain two different augmented data corresponding to each of the health data.

[0094] After obtaining the health data, two different data augmentation processes are performed on each health data point. The data augmentation process for each health data point is random, resulting in two augmented data points for each health data point. Because the two augmented data points have undergone different data augmentation processes, they are different.

[0095] Optionally, the data augmentation processes performed can include amplitude axis flipping, variable axis flipping, truncation, adding noise, cropping adjustments, time wrapping, and smoothing. The two data augmentation processes to be performed on each health dataset can be determined by generating two random integers from 1 to 7, each corresponding to one of the data augmentation processes.

[0096] Among them, amplitude axis flipping involves flipping the health data from left to right. For data x, the augmented data with amplitude axis flipping is... It can be represented as: Where A1 = [a m,n ] I×I Representing element a m,n The transformation matrix, where:

[0097]

[0098] Variable axis flipping is achieved by changing the labels of elements in data x, augmenting the data representation as follows: Where A2∈R I×I This represents a diagonal matrix with diagonal elements equal to -1.

[0099] The purpose of truncation is to randomly set the elements in data x to zero, with the number of truncated elements accounting for 0 < η³ < 1. Augmented data can be represented as... Where A3∈R I×I This represents a diagonal matrix, where the diagonal elements of int(Iη3) are 0, and the diagonal elements of [I-int(Iη3)] are 1. int() represents an integer function.

[0100] Adding noise involves adding Gaussian white noise to the data x. Augmented data can be represented as Where σ a =σ x η4 represents the standard deviation of the added noise, σ x Let denot x represent the standard deviation of the data, and η4>0 represent the relative noise intensity.

[0101] Pruning adjustment refers to randomly removing the portion of data x that contains elements with values ​​0 < η5 < 1, and then resampling the pruned data using linear interpolation to enhance the data. It has the same length as the original data x.

[0102] Temporal wrapping is a special resampling method where elements in data x are randomly discarded or interpolated. Furthermore, to ensure the length of the augmented sample remains constant, the number of discarded data points and the number of interpolated data points are the same. Specifically, it can be described as generating a decision vector a6∈R from elements -1, 0, 1. I The proportions of elements 0 and 1 in the total number of elements are both 0 < η6 < 0.5. Augmented data. In the decision vector a6, the element corresponding to element -1 in the data x is discarded, the element corresponding to element 0 in a6 is retained, and linear interpolation is performed on the element corresponding to element 1 in a6. The augmented data can be represented as: and

[0103] Smoothing involves filtering out high-frequency components from data x, and then convolving data x with a sliding window a7 of length η7 to obtain enhanced data, represented as... In this case, augmented data The transient portion of the data is suppressed, while the general form of the data x is preserved.

[0104] Step S23: Input the two different augmented data corresponding to each health data into the mapping network to be trained to obtain the data representation of each augmented data.

[0105] Build three convolutional neural networks (CNNs) in the PyTorch platform (an open-source machine learning library). Each CNN contains a convolutional layer, a batch normalization layer, and a non-linear activation function. One of the CNNs is the mapping network to be trained, and the other two CNNs can be the projection network and the prediction network, respectively.

[0106] The mapping network to be trained can obtain a data representation of each input data point.

[0107] Step S24: Perform self-supervised training on the mapping network to be trained based on the data representation of each augmented data to obtain the mapping network.

[0108] The time-varying operating conditions and variable operating environment of wind turbines can interfere with measurement data, making the measurement data under the same health condition also variable. This makes it difficult for the measurement data to reflect the health condition of the wind turbine blades, thus compromising the accuracy of the wind turbine blade condition monitoring results.

[0109] This application embodiment trains a mapping network using augmented data, which enhances the network's robustness. Therefore, the mapping network can overcome interference from operating conditions and the environment on the measurement data, thereby generating accurate data representations. Consequently, the blade state determined based on the data representation obtained from the mapping network has high accuracy.

[0110] Optionally, based on the above technical solution, in order to perform self-supervised training of the mapping network, projection networks and prediction networks can be used to assist in the training of the mapping network.

[0111] Figure 2 The diagram shows the structural composition of the mapping network, projection network, and prediction network, which are connected in series.

[0112] After obtaining the data representation of the augmented data output by the mapping network to be trained, the data representation of the augmented data is input into the projection network to obtain the projected data representation of each augmented data. The projected data representation of each augmented data is then input into the prediction network, which predicts whether any two augmented data are augmented data of the same healthy data, thus obtaining the prediction result.

[0113] Based on the prediction results and the projected data representation of each augmented data, a loss function can be established. The mapping network to be trained is then trained based on this loss function. Simultaneously, the projection network and the prediction network are also trained.

[0114] Optionally, negative cosine similarity can be calculated based on multiple prediction results and the projected data representation of each augmented data; a loss function can then be established based on the negative cosine similarity. The loss function L can be calculated using the following formula:

[0115]

[0116] in, For negative cosine similarity, p1 and p2 are the projected data of the two augmented data, q1 and q2 are the prediction results of the prediction network output; τ() indicates that the gradient stops backpropagating during training.

[0117] Negative cosine similarity It can be calculated using the following formula:

[0118]

[0119] Where p represents the projected data representation of the augmented data, and q represents the prediction result.

[0120] Based on the loss function, mini-batch gradient descent is used to iteratively train the neural network to conduct self-supervised learning. Health data is used to enhance predictions between samples, extracting information related to the health status of wind turbine blades. Training is stopped based on pre-set conditions, either after a preset number of training iterations or when training converges.

[0121] Optionally, the data representation of the augmented data obtained by the mapping network can be normalized using the 2-norm; the projection network projects the data representation of the augmented data that has undergone the 2-norm normalization process to obtain the projected data representation of the augmented data; the projected data representation of the augmented data is then normalized using the 2-norm; the projected data representation of the augmented data that has undergone the 2-norm normalization process is input into the prediction network, and the prediction network determines whether the two augmented data are augmented data of the same healthy data by calculating the cosine similarity between the data representations of the two input augmented data.

[0122] In this way, the mapping network can output accurate data representations, the projection network can isolate the mapping network and the prediction network to prevent coupling between them, and the prediction network can generate prediction results to complete self-supervised training.

[0123] Figure 3 A flowchart illustrating the blade state detection method is shown. The training of the mapping network can be performed offline, while blade state detection can be performed online in real time.

[0124] In the training phase, this application employs a self-supervised learning approach to extract effective representations from SCADA measurement data to reflect the health status of wind turbine blades. Only measurement data from healthy wind turbines are needed to establish a reference model for condition monitoring, without requiring measurement data under blade failure conditions. Therefore, it can solve the problem of scarce measurement data under wind turbine blade failure conditions, which is beneficial for carrying out data-driven condition monitoring under conditions of limited measurement data, thereby providing important reference information for the formulation of wind turbine operation and maintenance plans.

[0125] This application embodiment predicts measurement data from healthy wind turbine blades by executing a self-supervised agent task, which narrows the distance between healthy measurement data samples in the feature space. This helps to enhance the separability of healthy measurement data and fault measurement data in the learned data representation, suppresses the interference of operating condition changes and environmental noise on the measurement data, improves the accuracy of wind turbine blade condition monitoring results, and avoids catastrophic accidents caused by serious blade failures.

[0126] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0127] Figure 4 This is a schematic diagram of the structure of a blade condition detection device according to an embodiment of the present invention, as shown below. Figure 4 As shown, a blade state detection device includes a data acquisition module, a first data representation module, a second data representation module, and a state determination module, wherein:

[0128] The data acquisition module is used to acquire real-time measurement data. A real-time measurement data includes data of each dimension sampled at each sampling time point within a sampling time range.

[0129] The first data representation module is used to input the real-time measurement data into the mapping network to obtain the data representation of the real-time measurement data. The mapping network is trained using healthy data as training samples. The healthy data refers to the measurement data collected when the leaf is in a healthy state.

[0130] The second data representation module is used to obtain a data representation of the health data, which is obtained using the mapping network.

[0131] The state determination module is used to determine the state of the blade based on the similarity between the data representation of the real-time measurement data and the data representation of the health data.

[0132] Optionally, the state determination module includes:

[0133] A calculation unit is used to calculate the mean vector and covariance matrix of the data representation of the health data;

[0134] The first distance calculation unit is used to calculate a first distance between the data representation of the real-time measurement data and the data representation of the health data based on the data representation of the real-time measurement data, the mean vector, and the covariance matrix.

[0135] A state determination unit is used to determine the state of the blade based on the first distance.

[0136] Optionally, it also includes:

[0137] The second distance acquisition module is used to acquire the second distance between the data representations of multiple historical measurement data and the data representation of the health data;

[0138] The control chart building module is used to build an exponentially weighted moving average control chart based on multiple second distances;

[0139] The warning distance determination module is used to determine the warning distance based on the exponentially weighted moving average control chart.

[0140] The state determination unit includes:

[0141] The first determining subunit is used to determine that the blade is in a healthy state when the first distance is not greater than the warning distance;

[0142] The second determining subunit is used to determine that the blade is in an unhealthy state when the first distance is greater than the warning distance.

[0143] Optionally, the mapping network is trained through the following steps:

[0144] Obtain multiple sets of the aforementioned health data;

[0145] Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point.

[0146] Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data;

[0147] The mapping network to be trained is obtained by performing self-supervised training based on the data representation of each of the augmented data.

[0148] Optionally, acquiring multiple health data includes:

[0149] When the blade is in a healthy state, for each dimension, data sampled at each sampling time point within multiple sampling time ranges are acquired;

[0150] Filter out invalid data from the data to obtain multiple remaining data for each dimension within multiple sampling time ranges;

[0151] Eliminate the influence of the data volume of multiple remaining data in each dimension within multiple sampling time ranges to obtain multiple target data in each dimension within multiple sampling time ranges;

[0152] Based on each sampling time point within each sampling time range, multiple target data points for each dimension within that sampling time range are stacked to obtain multiple health data points.

[0153] Optionally, the step of performing self-supervised training on the mapping network to be trained based on the data representation of each of the augmented data to obtain the mapping network includes:

[0154] The data representation of each augmented data is input into the projection network to obtain the projected data representation of each augmented data.

[0155] The projected data representation of each augmented data is input into the prediction network to obtain the prediction result of whether each two augmented data are the same health data.

[0156] A loss function is established based on the multiple prediction results and the projected data representation of each of the augmented data;

[0157] Based on the loss function, the mapping network to be trained is trained to obtain the mapping network.

[0158] Optionally, establishing a loss function based on the multiple prediction results and the projected data representation of each augmented data includes:

[0159] Based on the multiple prediction results and the projected data representation of each of the augmented data, the negative cosine similarity is calculated;

[0160] The loss function is established based on the negative cosine similarity.

[0161] Optionally, the two random data augmentation processes include, but are not limited to, any two of the following: amplitude axis flipping, variable axis flipping, truncation, adding noise, cropping adjustment, time wrapping, and smoothing.

[0162] It should be noted that the device embodiments are similar to the method embodiments, so the description is relatively simple. For relevant details, please refer to the method embodiments.

[0163] This invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the blade state detection method disclosed in this application.

[0164] This invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements the blade state detection method disclosed in this application.

[0165] This invention also provides a computer program product, including a computer program or computer instructions, which, when executed by a processor, implement the blade state detection method disclosed in the embodiments of this application.

[0166] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0167] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0169] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0171] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0172] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0173] The blade condition detection method and apparatus provided in this application have been described in detail above. Specific examples have been used to illustrate the principle and implementation of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​this application. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for detecting the condition of a blade, characterized in that, include: Acquire real-time measurement data, which includes data from each dimension sampled at each sampling time point within a sampling time range; The real-time measurement data is input into a mapping network to obtain a data representation of the real-time measurement data. The mapping network is trained using healthy data as training samples. The healthy data refers to the measurement data collected when the leaf is in a healthy state. The data representation of the health data is obtained using the mapping network; The state of the leaf is determined based on the similarity between the data representation of the real-time measurement data and the data representation of the health data; The mapping network is trained through the following steps: Obtain multiple sets of the aforementioned health data; Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point. Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data; The data representation of each augmented data is input into the projection network to obtain the projected data representation of each augmented data. The projected data representation of each augmented data is input into the prediction network to obtain the prediction result of whether each two augmented data are the same health data. A loss function is established based on the multiple prediction results and the projected data representation of each of the augmented data; Based on the loss function, the mapping network to be trained is trained to obtain the mapping network.

2. The method according to claim 1, characterized in that, Determining the state of the leaf based on the similarity between the data representation of the real-time measurement data and the data representation of the health data includes: Calculate the mean vector and covariance matrix of the data representation of the health data; Based on the data representation of the real-time measurement data, the mean vector, and the covariance matrix, calculate the first distance between the data representation of the real-time measurement data and the data representation of the health data; The state of the blade is determined based on the first distance.

3. The method according to claim 2, characterized in that, Also includes: Acquire a second distance between the data representations of multiple historical measurement data and the data representations of the health data; Based on multiple second distances, an exponentially weighted moving average control chart is established; The warning distance is determined based on the aforementioned exponentially weighted moving average control chart; Determining the state of the blade based on the first distance includes: If the first distance is not greater than the warning distance, the blade is determined to be in a healthy state; If the first distance is greater than the warning distance, the blade is determined to be in an unhealthy state.

4. The method according to claim 1, characterized in that, The acquisition of multiple health data includes: When the blade is in a healthy state, for each dimension, data sampled at each sampling time point within multiple sampling time ranges are acquired; Filter out invalid data from the data to obtain multiple remaining data for each dimension within multiple sampling time ranges; Eliminate the influence of the data volume of multiple remaining data in each dimension within multiple sampling time ranges to obtain multiple target data in each dimension within multiple sampling time ranges; Based on each sampling time point within each sampling time range, multiple target data points for each dimension within that sampling time range are stacked to obtain multiple health data points.

5. The method according to claim 1, characterized in that, The step of establishing a loss function based on multiple prediction results and the projected data representation of each augmented data includes: Based on the multiple prediction results and the projected data representation of each of the augmented data, the negative cosine similarity is calculated; The loss function is established based on the negative cosine similarity.

6. The method according to any one of claims 4-5, characterized in that, The two random data augmentation processes include, but are not limited to, any two of the following: amplitude axis flipping, variable axis flipping, truncation, adding noise, cropping adjustment, time wrapping, and smoothing.

7. A blade condition detection device, characterized in that, include: The data acquisition module is used to acquire real-time measurement data. A real-time measurement data includes data of each dimension sampled at each sampling time point within a sampling time range. The first data representation module is used to input the real-time measurement data into the mapping network to obtain the data representation of the real-time measurement data. The mapping network is trained using healthy data as training samples. The healthy data refers to the measurement data collected when the leaf is in a healthy state. The second data representation module is used to obtain a data representation of the health data, which is obtained using the mapping network. A state determination module is used to determine the state of the leaf based on the similarity between the data representation of the real-time measurement data and the data representation of the health data; The mapping network is trained through the following steps: Obtain multiple sets of the aforementioned health data; Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point. Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data; The data representation of each augmented data is input into the projection network to obtain the projected data representation of each augmented data. The projected data representation of each augmented data is input into the prediction network to obtain the prediction result of whether each two augmented data are the same health data. A loss function is established based on the multiple prediction results and the projected data representation of each of the augmented data; Based on the loss function, the mapping network to be trained is trained to obtain the mapping network.

8. The apparatus according to claim 7, characterized in that, The mapping network is trained through the following steps: Obtain multiple sets of the aforementioned health data; Each health data point is subjected to two random data augmentation processes to obtain two different augmented data points corresponding to each health data point. Two different augmented data corresponding to each of the health data are input into the mapping network to be trained to obtain the data representation of each of the augmented data; The mapping network to be trained is obtained by performing self-supervised training based on the data representation of each of the augmented data.

Images