CN110161119A - Wind electricity blade defect identification method - Google Patents

Abstract

The invention discloses a method for identifying defects of wind power blades, which is to perform ultrasonic detection on wind power blades, perform wavelet packet transformation on the obtained ultrasonic detection signals according to the frequency band feature window, and input the obtained energy spectrum coefficients into the BP neural network as feature vectors, The neural network outputs the corresponding defect types to realize the automatic identification of different defects of wind turbine blades. The defect recognition method provided by the invention is effective and feasible, and makes automatic recognition of wind power blade defects possible, with an average recognition rate as high as 90%.

Description

Wind turbine blade defect identification method

technical field

The invention belongs to the technical field of intelligent detection, and in particular relates to a defect identification method of a wind power blade.

Background technique

As the core component of wind turbines, wind power blades have the characteristics of complex structure, various processes, and special materials, which inevitably lead to defects in the production process, such as inclusions, lack of glue, wrinkles, etc. Different types of defects have great impact on blade stiffness and strength. Therefore, in order to ensure the service life of the blade, it is very necessary to identify the type of blade defect. Ultrasonic testing has been widely used in the field of composite materials such as wind turbine blades due to its strong penetrating ability and convenient use. However, it is mainly used to detect the existence and location of bonding defects in the trailing edge of wind turbine blades. The identification of different defect modes in the web and girder mainly relies on manual experience, and comprehensive judgment is made based on the processing technology, material, structure and inspection data of the inspected workpiece; this method has a large workload and low efficiency, and the inspection level is subject to inspection The reason is that for wind power blade materials, due to the sound attenuation, scattering and reflection effects of composite materials, the blade ultrasonic signals are complex, and it is difficult to extract hidden defect information from a large number of complex ultrasonic signals. It is not possible to accurately assess the condition of defects in the blade.

Contents of the invention

In order to solve the above technical problems, the present invention provides a defect identification method for wind power blades based on frequency band characteristic windows, which is used to automatically identify defects inside the web or girder of wind power blades.

The wind power blade defect identification method provided by the present invention is specifically:

Step a) Ultrasonic detection is carried out on the sample containing defects, and the db4 function with the greatest correlation with the signal waveform is selected as the wavelet base for wavelet packet decomposition through the time domain diagram of the ultrasonic detection signal;

Step b) performing spectral transformation on the ultrasonic detection signal obtained in step a, combining spectral analysis and wavelet packet decomposition to find out the characteristic frequency band window, thereby determining the optimal number of layers m of wavelet packet transformation;

Step c) performing m-layer db4 wavelet packet decomposition on the ultrasonic detection signal obtained in step a, and obtaining decomposed nodes and energy spectral coefficients thereof through wavelet packet decomposition;

Step d) constructing a BP neural network, using the energy spectrum coefficient obtained in step c as an input vector, and the corresponding defect type as an output result, and training and verifying the neural network;

Step e) Ultrasonic detection is performed on the wind power blade to be tested, and the ultrasonic detection signal is input to the neural network constructed in step d to automatically identify the type of defects inside the wind power blade.

Preferably, the specific operation of step (a) is: perform ultrasonic detection on the wind turbine blade sample containing defects to obtain corresponding defect ultrasonic detection signals, and select the most relevant db4 as Wavelet basis for wavelet packet decomposition.

Preferably, the specific operation of step (b) is: perform spectral transformation on the defect ultrasonic detection signal provided in step a, analyze the frequency characteristics of the obtained spectral signal, perform k-level wavelet packet decomposition on the frequency characteristic of the spectral signal, and find out the characteristic frequency band Window (characteristic frequency band window is the data window corresponding to different characteristic frequency ranges in the data analysis software, the present invention adopts matlab to carry out data analysis), determines the optimal wavelet packet decomposition layer number m, thereby ensures each node after the wavelet packet m layer decomposition It can effectively separate the feature information in the spectrum signal; where, k is a natural number, k≠0, m∈k.

Preferably, the specific operation of step (c) is: perform m-level wavelet packet decomposition on the defect ultrasonic detection signal provided by step a with db4 wavelet, and extract the node energy spectrum coefficient corresponding to its node; select the node as the abscissa, and select the node corresponding to The energy spectrum coefficient is the ordinate, and the energy spectrum coefficient map is obtained; according to the energy of different nodes of different defects in the energy spectrum coefficient map, the types of defects are distinguished, and the correctness of the optimal decomposition layer number m is verified.

Preferably, the concrete operation of step (d) is: construct BP neural network, wherein, the input node of BP neural network is set to the number of energy spectrum coefficients after m layer wavelet packet decomposition, and output node is set to the kind number of defect, Through the training and verification of the neural network, the mean square error of the output layer of the neural network reaches the minimum value.

Preferably, the specific operation of step (e) is: conduct ultrasonic detection on the wind power blade to be tested, and input the energy spectrum coefficients obtained after decomposing the ultrasonic detection signal into the m-layer db4 wavelet packet into the neural network constructed in step d, and automatically identify Defect types of wind turbine blades.

It is worth mentioning that when determining the optimal number of decomposition layers of wavelet packets, the frequency band aliasing phenomenon will appear in the defect ultrasonic detection signal under the general wavelet packet analysis, and the defect characteristic information cannot be distinguished. The coefficient cannot be used as a basis for distinguishing between different defect types. But only when the m-th layer wavelet packet decomposition is performed based on the frequency band feature window, the different defect signal eigenvalues have differences (the characteristic frequency band separation of the defect ultrasonic signal). That is, in this decomposition method, we have found the characteristic parameters that can distinguish different defect types.

Compared with the prior art, the present invention proposes a defect identification method for wind power blades based on the frequency band characteristic window, which is to conduct ultrasonic detection on the wind power blade, perform wavelet packet transformation on the ultrasonic signal according to the frequency band characteristic window, and convert the obtained energy spectrum coefficient As a defect frequency feature, it is input to the constructed BP neural network, so as to realize the intelligent identification of different defects inside the wind turbine blade. By combining spectrum analysis and wavelet packet decomposition, the frequency band feature window is found, so as to realize the separation of defect frequency features. The spectrum energy features after wavelet packet decomposition are extracted, and the feature vector of defect signal is constructed as the input vector, which is input to the trained BP neural network, and the BP neural network outputs the corresponding defect type to realize the automatic identification of wind turbine blade defects. The defect recognition method provided by the invention is effective and feasible, avoids subjective errors of manual recognition, makes automatic recognition of wind power blade defects possible, and has an average automatic recognition rate of up to 90%.

Description of drawings

Figure 1 is a prefabricated drawing of a defect in a wind turbine blade;

Fig. 2a is a time-domain diagram of the ultrasonic detection signal of the non-defective part of the wind turbine blade obtained in step a of the embodiment of the present invention;

Fig. 2b is a time-domain diagram of the ultrasonic detection signal of the inclusion defect part of the wind turbine blade obtained in step a of the embodiment of the present invention;

Fig. 2c is a time-domain diagram of the ultrasonic detection signal of the wrinkle defect part of the wind turbine blade obtained in step a of the embodiment of the present invention;

Fig. 2d is a time-domain diagram of the ultrasonic detection signal of the wind power blade lacking glue defect obtained in step a of the embodiment of the present invention;

Fig. 3 a is the frequency spectrum of the wind power blade lack of glue defect obtained in step b of the embodiment of the present invention;

Fig. 3b is a non-destructive spectrum diagram of the wind power blade obtained in step b of the embodiment of the present invention;

Fig. 4a is an energy spectrum diagram of the ultrasonic detection signal of the non-defective part of the wind turbine blade obtained in step c of the embodiment of the present invention;

Fig. 4b is an energy spectrum diagram of the ultrasonic detection signal of the inclusion defect part of the wind turbine blade obtained in step c of the embodiment of the present invention;

Fig. 4c is an energy spectrum diagram of the ultrasonic detection signal of the wrinkle defect part of the wind turbine blade obtained in step c of the embodiment of the present invention;

Fig. 4d is an energy spectrum diagram of the ultrasonic detection signal of the glue-lack defect part of the wind power blade obtained in step c of the embodiment of the present invention;

Fig. 5 is a diagram of the training results of the BP neural network provided by the embodiment of the present invention.

Detailed ways

The present invention will be further specifically described below in conjunction with the accompanying drawings and embodiments.

Figure 1 shows the prefabricated defect-containing wind turbine blade sample (glass fiber cloth is glass fiber cloth) in this embodiment. The preset defect types are wrinkles, lack of glue and inclusions. The specific defect parameters are shown in Table 1. The wind turbine blades used are made of glass fiber cloth and epoxy resin, and are formed by vacuum injection technology. The manufacturing process is to lay a layer of glass fiber and then a layer of epoxy resin. There are 37 layers in total. The first layer And the 37th layer is glass fiber cloth, and its molding thickness is 1.2mm per layer, 44.4mm in total. Three types of defects including inclusions, lack of glue and wrinkles were prefabricated in the sample, that is, the wrinkles were formed by pre-embedding cylindrical FRP between the 17th and 18th layers of glass fiber cloth, and between the 19th and 20th layers of glass fiber cloth. The upper blade simulates the inclusion defect and the polytetrafluoroethylene simulates the lack of glue defect. The present invention adopts matlab software to carry out data analysis.

Table 1 Experimental sample parameter list

The defect identification method provided by the present invention is used to detect the defects of the wind turbine blade samples prepared in this embodiment as follows:

Step (1): Ultrasonic detection is performed on the sample by an ultrasonic detector, and the db4 function with the greatest correlation with the signal waveform is selected as the wavelet basis for wavelet packet decomposition through the time-domain diagram of the ultrasonic detection signal.

Specifically, the sampling frequency of the instrument is 10 MHz, and the probe adopts an Olympus R101 probe with a frequency of 0.5 MHz. The defect signals of the prepared samples were detected by the water coupling method, and 60 sets of ultrasonic detection signals were collected for each defect. Fourier transform is performed on the obtained ultrasonic detection signal, and the time-domain signal is shown in Figure 2. The ordinate is the amplitude, and the abscissa is the depth of the A-type reflected wave. As shown in Figure 2a, the bottom echo appears at the position of 44.4mm, which corresponds to the thickness of the prefabricated sample. Figure 2b, Figure 2c, and Figure 2d have defect echoes at 21.4mm, 23.8mm, and 25.0mm, respectively, at the defect prefabricated position, in addition to the bottom surface echo at the 44.4mm position, which is very similar to the position of the inventor's prefabricated defect. From the waveform diagram shown in Figure 2, it can be seen whether the sample has defects and the depth of the defects, but the corresponding defect type cannot be identified. According to the characteristics of the wavelet basis function and combined with the time domain waveform of the ultrasonic signal, in this embodiment, the db4 function most similar to the detection signal is selected as the wavelet basis for the wavelet packet decomposition.

Step (2): Perform spectral transformation on the ultrasonic detection signal obtained in step 1, combine the spectrum and wavelet packet decomposition, and determine the frequency band feature window, so as to determine the optimal decomposition layer number m of wavelet packet decomposition.

Specifically, in this embodiment, the wavelet packet decomposition of 1 to 5 layers is performed on the ultrasonic detection signal obtained in step 1, and the frequency band characteristic window obtained by analyzing each layer is selected to select the frequency band characteristic window in which the difference in energy distribution of the ultrasonic signal can be seen. Layer is the optimal number of decomposition layers of wavelet packet decomposition. In this embodiment, the frequency spectrum analysis results of samples lacking glue and non-destructive samples are taken as examples for illustration. It can be seen from Figure 3 that the main frequencies of both the lack of glue and the non-destructive signals are concentrated around 0.5MHZ, which is consistent with the transmitting frequency of the probe being 0.5MHZ. The lossless samples are troughs at 0.4MHZ and 0.5MHZ, and peaks at 0.35MHZ, 0.45MHZ and 0.55MHZ. These two troughs indicate the difference between the two spectrograms, so these two The eigenvalues of the signal can be found by separating the troughs. In this step, after the ultrasonic signal spectrum is decomposed by 4 layers of wavelet packets, 16 nodes are obtained, and the frequency bandwidth of each node is 0.3125MHZ, then the frequency of the second node is 0.3125-0.625MHZ, it can be clearly seen that the second Each node contains 0.4MHZ and 0.5MHZ, without separating the two, it can be seen that the decomposition scale is too small; after decomposing the signal spectrum with 5 layers of wavelet packets, 32 wavelet packet decomposition frequency bands will be obtained, each with a bandwidth of 0.15625MHZ After calculation, we can know that 0.4MHZ is at the third node, that is, 0.3125-0.46875MHZ, and 0.5MHZ is at the fourth node, that is, 0.46875-0.625

MHZ. It can be seen that the frequency band feature window can be found by decomposing the 5-layer wavelet packet, thereby separating the feature points of the two kinds of ultrasonic signals and reflecting the difference in energy distribution of the two kinds of ultrasonic signals. Therefore, the optimal decomposition layer number of the wavelet packet in this embodiment for 5.

Step (3): Perform 5-layer wavelet packet decomposition on the defect ultrasonic detection signal provided in step 1 with db4 wavelet to obtain 2 ⁵ =32 nodes and their corresponding energy spectrum coefficients, each node corresponds to a frequency segment of the defect signal, The characteristics of the frequency bands of different defects are different, and the corresponding energy spectrum coefficients are also different. Extract the energy spectrum coefficients of these 32 nodes, and make a wavelet packet energy spectrum coefficient histogram (two-dimensional histogram), as shown in Figure 4. It can be seen from Figure 4 that most of the energy is concentrated within the first 10 nodes, and the energy of the next 22 nodes is almost 0, and the signal energy is mainly concentrated in the 3rd, 4th and 7th, 8th nodes. It can be clearly seen from the spectrum energy histogram that different defect types have different distributions of wavelet energy spectral coefficients, which proves the correctness of the inventor's selection of 5-layer wavelet packet decomposition and using energy spectral coefficients as characteristic parameters to distinguish defect types.

The wavelet packet transform is used to extract the energy spectrum coefficients of the ultrasonic detection signals of the three defects of glue, inclusion and wrinkle, and 60 sets of each defect are tested for a total of 180 sets of data, of which 40 sets of data are selected from the 60 sets of data for each defect to form a training sample set , and the remaining 20 sets of data constitute the test sample set. Due to the large amount of data, this embodiment selects representative data in the training set and displays them in a table, as shown in Table 2. The test set selects 9 representative sets of data (3-10 nodes) for display. This type of data is the digital feature that can distinguish the defect type, as shown in Table 3.

Table 2 training sample set

Table 3 Test sample set

Step (4): The energy spectrum is a quantifiable characteristic parameter, so the wavelet energy spectrum coefficient calculated after the wavelet packet decomposition is used as the input vector of the neural network to construct the BP neural network. When constructing the BP neural network, considering the five-layer wavelet packet decomposition of the defect ultrasonic detection signal, 32 spectral energy feature values after wavelet packet decomposition are extracted, and the corresponding neural network input nodes are set to 32. Three kinds of defects need to be automatically identified, and the output nodes of the network are set to 3, among which (100) represents inclusion defects, (010) represents fold defects, and (001) represents glue-lack defects. A general neural network has only one hidden layer. If the BP neural network contains one hidden layer, it can realize the hyperplane division of less sample space, but in this embodiment, the number of samples is large, so two hidden layers are selected, which can reduce the network scale and improve the neural network. training accuracy. Now the number of hidden layer nodes is generally selected by the following empirical formula:

Among them, n is the number of nodes in the hidden layer, n _i is the number of nodes in the input layer, n ₀ is the number of nodes in the output layer, and a is a constant from 1 to 10.

In this embodiment, based on empirical formulas, it is found through trial and error that when the number of nodes in the hidden layer is 8, the error of the prediction result of the BP neural network is the smallest, and the structure of the neural network is relatively stable. Finally, it is determined that the number of nodes in the first hidden layer is set to 8, and the number of nodes in the second hidden layer is set to 4. Since the input and output data are all in the range of [0,1] after normalization, in order to reduce the error, both hidden layers use the S-type Sigmoid function. The network is trained using an algorithm with a variable learning rate, and only the initial learning rate needs to be determined. In order to ensure the stability of the system, the learning rate value is generally selected between 0.01-1. In this embodiment, the initial learning rate is selected as 0.8, the set target error is 0.001, and the maximum number of training times is 5000.

The data (Table 2) extracted from the three types of defects after wavelet packet decomposition is input to the BP neural network as a training set. The training results are shown in Figure 5, where the broken line is the training error and the dotted line is the target error (0.001). During the 9-step round-trip training process in the BP neural network, the broken line shows a steady downward trend and finally drops below the target error value to achieve network convergence. . Use the test sample set (Table 3) to input the trained network for testing, and the recognition results are shown in Table 4. It can be seen from Table 4 that the recognition effect of the three kinds of defects is better, and the average recognition accuracy rate reaches more than 90%.

Table 4 BP neural network recognition results

Based on the frequency band feature window, combined with frequency spectrum analysis and wavelet packet decomposition, the invention uses BP neural network to analyze ultrasonic signals, and realizes the automatic identification of defect types of wind power blades. Specifically, combine spectrum analysis and wavelet packet analysis to find out the frequency band characteristic window, and then determine the number of wavelet decomposition layers to be 5. The five-layer wavelet packet decomposition is carried out on the defect signal of wind power blades, and the spectrum energy feature of the decomposed wavelet packet is extracted, and then the feature vector of the defect signal is constructed, and the feature vector is input into the BP neural network for training and effect inspection. Experimental results show that the defect recognition method provided by the present invention is effective and feasible, making automatic recognition of wind power blade defects possible, and the average recognition rate is as high as 90%.

Claims (6)

1. A wind turbine blade defect identification method, is characterized in that, comprises the steps: Step a) Ultrasonic detection is carried out on the sample containing defects, and the db4 function with the greatest correlation with the signal waveform is selected as the wavelet base for wavelet packet decomposition through the time domain diagram of the ultrasonic detection signal; Step b) performing spectral transformation on the ultrasonic detection signal obtained in step a, combining spectral analysis and wavelet packet decomposition to find out the characteristic frequency band window, thereby determining the optimal number of layers m of wavelet packet transformation; Step c) performing m-layer db4 wavelet packet decomposition on the ultrasonic detection signal obtained in step a, and obtaining decomposed nodes and energy spectral coefficients thereof through wavelet packet decomposition; Step d) constructing a BP neural network, using the energy spectrum coefficient obtained in step c as an input vector, and the corresponding defect type as an output result, and training and verifying the neural network; Step e) Ultrasonic detection is performed on the wind power blade to be tested, and the ultrasonic detection signal is input to the neural network constructed in step d to automatically identify the type of defects inside the wind power blade. 2. The wind turbine blade defect identification method according to claim 1, characterized in that, the specific operation of step (a) is: carry out ultrasonic detection on the wind turbine blade sample containing defects to obtain corresponding defect ultrasonic detection signals, and observe the defect ultrasonic The signal waveform in the time domain diagram of the detection signal selects db4 with the greatest correlation as the wavelet base for wavelet packet decomposition. 3. The wind turbine blade defect identification method according to claim 1, characterized in that, the specific operation of step (b) is: carry out spectrum conversion to the defect ultrasonic detection signal provided in step a, analyze the frequency characteristics of the gained spectrum signal, and The frequency characteristic of spectral signal carries out k layer wavelet packet decomposition, finds out characteristic frequency band window (the characteristic frequency band window is the data window corresponding to different characteristic frequency ranges in the data analysis software, the present invention adopts matlab to carry out data analysis), determines optimal wavelet packet Decompose the number of layers m, so as to ensure that each node after the decomposition of the wavelet packet m layer can effectively separate the characteristic information in the spectrum signal; where k is a natural number, k≠0, m∈k. 4. The wind turbine blade defect identification method according to claim 1, characterized in that, the specific operation of step (c) is: carry out m layer wavelet packet decomposition with db4 wavelet to the defect ultrasonic detection signal provided in step a, corresponding to its node Extract the node energy spectrum coefficient; select the node as the abscissa, select the corresponding energy spectrum coefficient as the ordinate, and obtain the energy spectrum coefficient map; according to the different node energies of different defects in the energy spectrum coefficient map, distinguish each defect type, and verify the most The correctness of optimal decomposition layers m. 5. The method for identifying defects of wind turbine blades according to claim 1, wherein the specific operation of step (d) is: constructing a BP neural network, wherein the input node of the BP neural network is set to m-layer wavelet packet decomposed The number of energy spectrum coefficients, the output node is set to the number of types of defects, through the training and verification of the neural network, the mean square error of the output layer of the neural network reaches the minimum value. 6. The wind power blade defect identification method according to claim 1, characterized in that, the specific operation of step (e) is: the wind power blade to be tested is subjected to ultrasonic detection, and the ultrasonic detection signal is decomposed into an m-layer db4 wavelet packet and obtained The energy spectrum coefficient is input into the neural network constructed in step d, and the defect type of the wind power blade is automatically identified by the computer.