A method and device for monitoring icing state of a wind turbine blade

Abstract

The application discloses a kind of wind generating set blade icing state monitoring method and device, first, the electric signal of piezoelectric film sensor attached on each blade under different time, different weather condition is collected, and impedance curve is drawn;Then the time domain feature, frequency domain feature and time-frequency domain feature are obtained by feature extraction to electric signal, the fusion feature is obtained after the time domain feature, frequency domain feature and time-frequency domain feature are processed and spatially fused;Finally, the fusion feature is input into blade state monitoring neural network, and blade icing state monitoring is carried out;The application has the advantages of strong applicability, high detection accuracy, can accurately identify blade icing state, meet the safety production monitoring demand of wind power generation enterprise, provide reference for generator set blade deicing work.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of blade icing state monitoring, and relates to a wind turbine generator set blade icing state monitoring method and device, in particular to a wind turbine generator set blade icing state monitoring method and device based on a piezoelectric film sensor. BACKGROUND

[0002] Wind power generation technology plays an important role in energy saving and emission reduction and air pollution prevention. In order to obtain higher wind speed, wind turbines are usually installed on the sea and remote high-altitude land areas, which can cause the blades to be prone to icing. After the wind turbine blades are iced, the blade airfoil will change, causing the aerodynamic efficiency to decrease, reducing power output, and in severe cases, the imbalance of ice load on the blade can cause the wind turbine to vibrate, affecting the service life of the equipment. In addition, part of the ice will be thrown out due to wind speed and rotational centrifugal force, which will cause great safety hazards to the unit and on-site workers. Therefore, timely monitoring of the icing condition of the wind turbine blade is of great significance to ensure the safe operation of the unit and maximize the reduction of economic losses.

[0003] The current blade icing monitoring technology can be divided into two categories: direct method and indirect method. The commonly used direct method includes grating detection method and image recognition method, and the indirect method usually uses machine learning to predict the icing thickness. The image recognition method obtains the icing thickness information by comparing the blade images before and after icing. This method can obtain relatively accurate icing thickness, but cannot realize real-time dynamic monitoring. The grating detection method measures the strain ratio of the upper and lower surfaces of the blade, and calculates the icing thickness according to the strain ratio. It has the characteristics of high sensitivity, fast response, high resolution and long distance coverage, but the cost is too high. The data source of the icing prediction model based on machine learning is generally directly taken from the prediction target, which is suitable for specified models of wind turbine generators and specific regions, and has limited scope of application. Therefore, it is necessary to propose a blade icing state real-time monitoring method with wide application range. SUMMARY

[0004] In view of the deficiencies and improvement needs of the prior art, the present application provides a wind turbine generator set blade icing state monitoring device and method based on a piezoelectric film sensor. The purpose is to generate signals by piezoelectric film, and to classify the signals by deep residual convolutional neural network, so as to realize real-time monitoring of the blade icing state and thickness, and to provide a reference for the safety evaluation of wind power generation and the deicing work of the generator set blade.

[0005] The technical scheme adopted by the method of the present application is as follows: a wind turbine generator set blade icing state monitoring method, comprising the following steps:

[0006] Step 1: Collect the electric signals of the piezoelectric film sensor attached to each blade under different time and weather conditions, and draw the impedance curve;

[0007] Step 2: Extract features from the electrical signal to obtain time-domain features, frequency-domain features, and time-frequency-domain features. After processing the time-domain features, frequency-domain features, and time-frequency-domain features, perform spatial fusion to obtain fused features.

[0008] Step 3: Input the fused features into the leaf state monitoring neural network to monitor the leaf icing status;

[0009] The leaf state monitoring neural network consists of seven layers. The first layer consists of convolutional layers and pooling layers. The fused feature Concat1 is input into the convolutional layer to obtain convolutional feature C4. The convolutional feature C4 is then pooled to obtain pooled feature P4. The second layer consists of three residual convolutional modules. The third layer consists of a downsampling convolutional module and three residual convolutional modules. The fourth layer consists of a downsampling convolutional module and five residual convolutional modules. The fifth layer consists of a downsampling convolutional module, two residual convolutional modules, and a pooling layer. The sixth layer is a fully connected layer. The seventh layer is a Softmax classifier.

[0010] Preferably, in step 1, the piezoelectric thin film sensor consists of two parts: a piezoelectric thin film and a signal conditioning circuit. The conditioning circuit is composed of passive components, including capacitors, resistors, and inductors, and is disposed below the piezoelectric thin film in an integrated circuit substrate.

[0011] Preferably, the upper and lower electrodes of the piezoelectric film have an irregular pentagonal structure.

[0012] Preferably, in step 1, the electrical signal acquisition circuit adopts a combination of oscillation circuit and bulk acoustic wave reflection coefficient measurement.

[0013] As a preferred embodiment, step 2 includes the following sub-steps:

[0014] Step 2.1: Process the electrical signal using the variational mode decomposition algorithm to obtain a reconstructed signal that is only related to the intrinsic modes;

[0015] Step 2.2: Use continuous wavelet transform to plot the time-frequency spectrum of the reconstructed signal to obtain feature samples containing time-domain features, frequency-domain features, and time-frequency-domain features;

[0016] Step 2.3: Perform convolution operations on the time-domain features, frequency-domain features, and time-frequency-domain features with the convolution kernel to obtain convolution features C1, C2, and C3;

[0017] Step 2.4: Pool the features C1, C2, and C3 to obtain pooled features P1, P2, and P3;

[0018] Step 2.5: Concatenate and merge pooled features P1, P2, and P3 into feature Concat1.

[0019] Preferably, in step 3, when the blade is found to be covered by ice, the static capacitance change of the piezoelectric thin film sensor is calculated by the series resonant spectrum and parallel resonant frequency in the impedance curve, and the thickness of the ice covering the blade is determined.

[0020] As a preferred embodiment, the relationship between the parallel resonant frequency and the series resonant frequency in the piezoelectric film capacitor and impedance curve is as follows:

[0021] ;

[0022] in, It is a dynamic capacitor, which is related to the input voltage; and The series resonant frequency and the parallel resonant frequency are taken from the impedance curve. The frequency corresponding to the highest point of the impedance curve is the series resonant frequency, and the frequency corresponding to the lowest point of the impedance curve is the parallel resonant frequency.

[0023] Preferably, in step 3, the Softmax cross-entropy loss function is:

[0024] ;

[0025] ;

[0026] in, Indicates the sample label value. Indicates the output category of the neural network. For the output vector, for The Middle The value of an output or category. This indicates the category that needs to be calculated. Training will stop when the recognition accuracy is greater than 90%.

[0027] Preferably, the blade state monitoring neural network mentioned in step 3 is a trained network; the training process includes the following steps:

[0028] (1) Collect electrical signals from several piezoelectric thin film sensors attached to each blade at different times and under different weather conditions, and plot impedance curves;

[0029] (2) Extract features from the electrical signal to obtain time-domain features, frequency-domain features and time-frequency-domain features, process the time-domain features, frequency-domain features and time-frequency-domain features and then perform spatial fusion to obtain fused features;

[0030] Step 3: Input the fused features into the blade state monitoring neural network for network training; obtain the classification results through the Softmax classifier and backpropagate the error to obtain the trained blade state monitoring neural network.

[0031] The technical solution adopted by the device of the present invention is: a wind turbine blade icing status monitoring device, comprising:

[0032] One or more processors;

[0033] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the wind turbine blade icing status monitoring method.

[0034] The beneficial effects of this invention are as follows:

[0035] 1. Combining a piezoelectric thin-film sensor with a deep residual convolutional neural network (blade condition monitoring neural network) enables real-time monitoring of the material adhering to blades. In the case of icing, it can calculate the ice thickness and issue an alarm when the ice thickness exceeds a set threshold, providing a reference for blade de-icing operations.

[0036] 2. Noise reduction of thin-film bulk acoustic sensor signals is achieved using variational mode decomposition to prevent overfitting of the neural network. Parallel multi-spatial features considering time-domain, frequency-domain, and time-frequency-domain features are established as inputs. A deep residual convolutional neural network (blade state monitoring neural network) is employed. By increasing the depth of the neural network and utilizing skip links in the residual convolution module, the vanishing gradient problem is mitigated, improving model accuracy.

[0037] 3. This invention has the advantages of strong applicability and high detection accuracy. It can accurately identify the icing status of blades, meet the safety production monitoring needs of wind power generation companies, and provide a reference for the de-icing work of generator blades. Attached Figure Description

[0038] The technical solutions described herein are further illustrated below using examples and specific implementation methods. Additionally, accompanying drawings are used in the description of the technical solutions. Those skilled in the art can, without any creative effort, obtain other drawings and the intent of the present invention based on these drawings.

[0039] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of a wind turbine generator in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of a piezoelectric thin film sensor in an embodiment of the present invention;

[0042] Figure 4 This is a schematic diagram of the piezoelectric thin film sensor conditioning circuit in an embodiment of the present invention;

[0043] Figure 5 This is a schematic diagram illustrating the relationship between the piezoelectric thin film mass load, capacitance, and frequency in an embodiment of the present invention;

[0044] Figure 6 This is an architecture diagram of a deep residual convolutional neural network (blade state monitoring neural network) in an embodiment of the present invention. Detailed Implementation

[0045] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0046] Please see Figure 1 This embodiment provides a method for monitoring the icing status of wind turbine blades, which includes the following steps:

[0047] Step 1: Collect electrical signals from the piezoelectric thin film sensor attached to each blade at different times and under different weather conditions, and plot the impedance curve;

[0048] In another implementation, such as Figure 2 As shown, piezoelectric thin-film sensors (1, 2, 3) are attached to the middle of each blade. Under different weather conditions (air, ice, water, ice-water mixture), the magnitude and distribution of the mass load on the piezoelectric thin film vary, causing changes in the capacitance of the piezoelectric thin film and emitting an electrical signal. Among them, the piezoelectric thin-film sensors (such as...) Figure 3 (As shown) It consists of two parts: a piezoelectric film and a signal conditioning circuit. The upper and lower electrodes 4 of the piezoelectric film have an irregular pentagonal structure and are disposed on the piezoelectric material 5; the conditioning circuit consists of passive components, including capacitors, resistors, and inductors (such as...). Figure 4 As shown), it is disposed below the piezoelectric thin film in the integrated circuit substrate 6;

[0049] In another implementation, a signal acquisition circuit is provided to receive the electrical signal emitted by the sensor and convert the change in piezoelectric film capacitance into a change in the output frequency of an oscillator (e.g., Figure 5 As shown in the figure, the impedance curves of the piezoelectric thin film sensor were collected under different times and weather conditions; the signal acquisition circuit adopts a combination of oscillation circuit and bulk acoustic wave reflection coefficient measurement.

[0050] Step 2: Extract features from the electrical signal to obtain time-domain features, frequency-domain features, and time-frequency-domain features. After processing the time-domain features, frequency-domain features, and time-frequency-domain features, perform spatial fusion to obtain fused features.

[0051] In another implementation, step 2 includes the following sub-steps:

[0052] Step 2.1: Process the electrical signal using the variational mode decomposition algorithm to obtain a reconstructed signal that is only related to the intrinsic modes;

[0053] Step 2.2: Use continuous wavelet transform to plot the time-frequency spectrum of the reconstructed signal to obtain feature samples containing time-domain features, frequency-domain features, and time-frequency-domain features;

[0054] Step 2.3: Perform convolution operations on the time-domain features, frequency-domain features, and time-frequency-domain features with the convolution kernel to obtain convolution features C1, C2, and C3;

[0055] Step 2.4: Pool the features C1, C2, and C3 to obtain pooled features P1, P2, and P3;

[0056] Step 2.5: Concatenate and merge pooled features P1, P2, and P3 into feature Concat1.

[0057] Step 3: Input the fused features into the leaf state monitoring neural network to monitor the leaf icing status;

[0058] In another implementation, please see Figure 6 The leaf state monitoring neural network consists of seven layers. The first layer consists of convolutional layers and pooling layers. The fused feature Concat1 is input into the convolutional layer to obtain convolutional feature C4. The convolutional feature C4 is then pooled to obtain pooled feature P4. The second layer consists of three residual convolutional modules. The third layer consists of a downsampling convolutional module and three residual convolutional modules. The fourth layer consists of a downsampling convolutional module and five residual convolutional modules. The fifth layer consists of a downsampling convolutional module, two residual convolutional modules, and a pooling layer. The sixth layer is a fully connected layer. The seventh layer is a Softmax classifier.

[0059] The first layer of the neural network is a convolutional layer. The fused feature Concat1 is input into the convolutional layer to obtain convolutional feature C4. Convolutional feature C4 is pooled to obtain pooled feature P4. Pooled feature P4 is then input into the second layer of the neural network, which consists of three residual convolutional modules, to obtain convolutional feature C5. Convolutional feature C5 is then input into the third layer of the neural network, which consists of a downsampling convolutional module and three residual convolutional modules, to obtain convolutional feature C6. Convolutional feature C6 is then input into the fourth layer of the neural network, which consists of a downsampling convolutional module and five residual convolutional modules, to obtain convolutional feature C7. Convolutional feature C7 is then input into the fifth layer of the neural network, which consists of a downsampling convolutional module and two residual convolutional modules, to obtain convolutional feature C8. Convolutional feature C8 is then pooled to obtain pooled feature P5. Pooled feature P5 is then connected to a fully connected layer, and finally, a Softmax classifier is used to obtain the classification result.

[0060] In another implementation, the Softmax cross-entropy loss function can be expressed as follows:

[0061] ;

[0062] ;

[0063] in, Indicates the sample label value. Indicates the output category of the neural network. For the output vector, for The Middle The value of an output or category. This indicates the category that needs to be calculated. Training will stop when the recognition accuracy is greater than 90%.

[0064] In another implementation, when the blade is detected to be covered in ice, the static capacitance change of the piezoelectric thin-film sensor is calculated using the series resonant spectrum and parallel resonant frequency in the impedance curve, and the thickness of the ice layer on the blade is determined. When the degree of icing on the blade exceeds a set threshold, an icing warning is issued.

[0065] In another embodiment, the relationship between the parallel resonant frequency and the series resonant frequency in the piezoelectric film capacitor and impedance curve is as follows:

[0066] ;

[0067] in, It is a dynamic capacitor, which is related to the input voltage; and The series resonant frequency and the parallel resonant frequency are taken from the impedance curve. The frequency corresponding to the highest point of the impedance curve is the series resonant frequency, and the frequency corresponding to the lowest point of the impedance curve is the parallel resonant frequency.

[0068] In another embodiment, the blade ice thickness is measured in the range of 0-30 mm.

[0069] In another embodiment, the blade state monitoring neural network is a trained network; the training process includes the following steps:

[0070] (1) Collect electrical signals from several piezoelectric thin film sensors attached to each blade at different times and under different weather conditions, construct a sample set, and plot impedance curves;

[0071] (2) Extract features from the electrical signal to obtain time-domain features, frequency-domain features and time-frequency-domain features, process the time-domain features, frequency-domain features and time-frequency-domain features and then perform spatial fusion to obtain fused features;

[0072] For the constructed sample set, variational mode decomposition is used for preprocessing to obtain reconstructed signals that are only related to intrinsic modes. Then, feature extraction is performed on the reconstructed signals in the time and frequency domains to obtain time-domain features, frequency-domain features, and time-frequency-domain features. After processing the time-domain features, frequency-domain features, and time-frequency-domain features, spatial fusion is performed to obtain fused features.

[0073] (3) Label the obtained fusion features and set the blade state into four categories: air, ice, water, and ice-water mixture. Input the fusion features into the blade state monitoring neural network for network training; obtain the classification results through the Softmax classifier and backpropagate the error to obtain the trained blade state monitoring neural network.

[0074] This embodiment also provides a device for monitoring the icing status of wind turbine blades, including:

[0075] One or more processors;

[0076] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the wind turbine blade icing status monitoring method.

[0077] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for monitoring the icing status of wind turbine blades, characterized in that, Includes the following steps: Step 1: Collect electrical signals from the piezoelectric thin film sensor attached to each blade at different times and under different weather conditions, and plot the impedance curve; The piezoelectric thin film sensor consists of two parts: a piezoelectric thin film and a signal conditioning circuit. The conditioning circuit is composed of passive components, including capacitors, resistors, and inductors, and is disposed below the piezoelectric thin film in an integrated circuit substrate. Step 2: Extract features from the electrical signal to obtain time-domain features, frequency-domain features, and time-frequency-domain features. After processing the time-domain features, frequency-domain features, and time-frequency-domain features, perform spatial fusion to obtain fused features. Step 3: Input the fused features into the leaf state monitoring neural network to monitor the leaf icing status; The leaf state monitoring neural network consists of seven layers. The first layer comprises a convolutional layer and a pooling layer. The fused feature Concat1 is input into the convolutional layer to obtain the convolutional feature C4. The convolutional feature C4 is then pooled to obtain the pooled feature P4. The second layer consists of three residual convolutional modules. The third layer consists of a downsampling convolutional module and three residual convolutional modules. The fourth layer consists of a downsampling convolutional module and five residual convolutional modules. The fifth layer consists of a downsampling convolutional module, two residual convolutional modules, and a pooling layer. The sixth layer is a fully connected layer. The seventh layer is a Softmax classifier. Specifically, when the blade is detected to be covered in ice, the static capacitance change of the piezoelectric thin-film sensor is calculated using the series and parallel resonant frequencies in the impedance curve, and the ice thickness on the blade is determined. The relationship between the piezoelectric thin-film capacitance and the parallel and series resonant frequencies in the impedance curve is as follows: ； in, It is a dynamic capacitor, which is related to the input voltage; and The series resonant frequency and the parallel resonant frequency are taken from the impedance curve. The frequency corresponding to the highest point of the impedance curve is the series resonant frequency, and the frequency corresponding to the lowest point of the impedance curve is the parallel resonant frequency. The Softmax cross-entropy loss function is: ； ； in, Indicates the sample label value. Indicates the output category of the neural network. For the output vector, for The Middle The training stops when the recognition accuracy is greater than 90%, and the output or category value is calculated.

2. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that: The upper and lower electrodes of the piezoelectric film have an irregular pentagonal structure.

3. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that: In step 1, the electrical signal acquisition circuit adopts a combination of oscillation circuit and bulk acoustic wave reflection coefficient measurement.

4. The method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that: Step 2 includes the following sub-steps: Step 2.1: Process the electrical signal using the variational mode decomposition algorithm to obtain a reconstructed signal that is only related to the intrinsic modes; Step 2.2: Use continuous wavelet transform to plot the time-frequency spectrum of the reconstructed signal to obtain feature samples containing time-domain features, frequency-domain features, and time-frequency-domain features; Step 2.3: Perform convolution operations on the time-domain features, frequency-domain features, and time-frequency-domain features with the convolution kernel to obtain convolution features C1, C2, and C3; Step 2.4: Pool the features C1, C2, and C3 to obtain pooled features P1, P2, and P3; Step 2.5: Concatenate and merge pooled features P1, P2, and P3 into feature Concat1.

5. The method for monitoring the icing status of wind turbine blades according to any one of claims 1-4, characterized in that: The blade state monitoring neural network mentioned in step 3 is a trained network; The training process includes the following steps: (1) Collect electrical signals from several piezoelectric thin film sensors attached to each blade at different times and under different weather conditions, and plot impedance curves; (2) Extract features from the electrical signal to obtain time-domain features, frequency-domain features and time-frequency-domain features, process the time-domain features, frequency-domain features and time-frequency-domain features and then perform spatial fusion to obtain fused features; (3) Input the fused features into the blade state monitoring neural network and train the network; obtain the classification results through the Softmax classifier and backpropagate the error to obtain the trained blade state monitoring neural network.

6. A device for monitoring the icing status of wind turbine blades, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the wind turbine blade icing status monitoring method as described in any one of claims 1 to 5.

Images