Method for impact localization of small scale sensor array composites based on deep learning

Abstract

The application discloses a small-scale sensing array composite material impact positioning method based on deep learning, which comprises the following steps: first, dividing a grid on a structure to be impacted, installing 1 or 2 sensors on the structure to be impacted, and acquiring impact signals of monitoring points; then, taking the impact signals as a whole as a reference target, extracting real part and imaginary part signals after Fourier transform as features, establishing a data set, normalizing the real part and imaginary part features and making gray-scale images, and dividing the data set into a training set, a verification set and a test set; then, constructing a neural network model, loading pre-training weights of EfficientNetV2, starting training by using a preheating learning rate, and adopting a cosine annealing learning rate scheduling strategy to optimize model training in a subsequent training stage; finally, inputting the gray-scale images of the impact signal test set into the trained neural network model to obtain a region position of an impact source of the impact signal; the application can reduce the number of sensors while ensuring positioning accuracy and searching speed, and can adapt to the complexity of a composite material medium.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structural health monitoring, and particularly relates to a small-scale sensor array composite impact positioning method based on deep learning. TECHNICAL BACKGROUND

[0002] Compared with traditional materials such as aluminum and steel, glass fiber reinforced epoxy resin composites have the advantages of light weight, corrosion resistance and excellent fatigue performance, and have been widely used in the wind power industry, which can effectively improve the performance and service life of wind turbine blades. However, the external environment of wind turbine blades is complex and changeable, and its integrity and reliability may be affected by sudden events such as wind sand, hail and flying animals. These unexpected impacts may cause hidden delamination, cracks and other micro-damage inside the composite material, thereby reducing the load-carrying capacity of the structure. If not timely monitored and repaired, these damages may further expand, causing significant losses. Therefore, impact monitoring is particularly important in the structural health monitoring of wind turbine blades.

[0003] Currently, the sensors commonly used for impact positioning research at home and abroad mainly involve piezoelectric sensors and grating optical fiber sensors. In 2016, Sai et al. used a method based on Morlet wavelet basis function to extract the wave arrival time difference of the impact response signal, and determined the impact point position through an equivalent right triangle sensor array. In 2024, Zhong et al. used OFDR optical fiber sensors to obtain the strain response information of the composite laminate structure, and on this basis proposed an impact load positioning method based on the nonlinear weighting principle of strain amplitude. These methods usually require multiple sensor nodes, which increases the burden of cables, acquisition channels, storage requirements and processing units, and further increases the overall weight and cost of the structural health monitoring system; in addition, the characteristics of composite materials make the propagation and attenuation law of acoustic emission signals significantly different from single metal materials, limiting the application of traditional impact positioning methods in composite materials.

[0004] In recent years, with the development of artificial intelligence and computer technology, data-driven methods have the ability to extract features without prior knowledge, and can achieve satisfactory positioning accuracy with fewer sensors, and are suitable for simple plate structures and complex structures; this method directly inputs the original impact signal into the designed network and directly outputs the positioning result. However, the current impact positioning method has the problems of difficult to eliminate errors and insufficient positioning accuracy, and the engineering practical effect is poor. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a small-scale sensor array composite impact positioning method based on deep learning, which converts feature signals into images for deep learning, uses an EfficientNetV2 network structure as the backbone network to extract multi-scale features for classification and positioning of gray images, and effectively reduces the number of sensor channels.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is:

[0007] A small-scale sensor array composite impact positioning method based on deep learning, comprising the following steps:

[0008] Step one, according to the positioning accuracy requirement, divide the grid on the impacted structure, and number the divided grid 1~N;

[0009] Step two, install M piezoelectric acceleration sensors on the impacted structure to obtain the impact signal generated when the object falls, wherein M=1 or 2, and layout the sensor position;

[0010] Step three, obtain the impact signal f(t) of multiple monitoring points in the target area, the impact signal f(t) including direct wave component and boundary reflection component;

[0011] Step four, according to the principle of multi-path propagation, the impact signal f(t) is taken as a reference target, and the real part and imaginary part signals after Fourier transform are extracted as features;

[0012] Step five, establish a data set, normalize the real part and imaginary part features and make them into gray images, and divide the data set into training set, validation set and test set;

[0013] Step six, build a neural network model, load the pre-trained weights of EfficientNetV2, and use the preheating learning rate to start training; in the subsequent stage of training, a cosine annealing learning rate scheduling strategy is adopted to optimize the model training, which maintains a high learning rate in the early stage of training, and then gradually reduces the learning rate, in the training iteration, the loss is calculated and the model weight is updated; after each training period, the model performance is evaluated using the validation set, and the loss and accuracy indicators on the validation set are monitored;

[0014] Step seven, input the gray image of the impact signal test set into the trained neural network model to obtain the position of the impact source of the impact signal;

[0015] Step eight, display the positioning result.

[0016] In the step two, if M=1, the sensor is arranged at the center point of the target area; if M=2, the sensor is symmetrically arranged with the center point of the target area.

[0017] The impact signal f(t) obtained by each piezoelectric acceleration sensor in step four is subjected to Fourier transform according to formula (1),

[0018]

[0019] Wherein F(ω) is the frequency spectrum function after Fourier transform, the independent variable is ω, which is the frequency; e -jωt is a complex signal; f(t) is an impact signal;

[0020] After Fourier transform, the real part and the imaginary part are extracted and sliced, and the first K non-zero data are taken.

[0021] In step five, the data of the M piezoelectric acceleration sensors are spliced to obtain 2×M×K data points, and normalization is performed according to formula (2) to make the output between 0 and 1, so that the dimensional expression becomes dimensionless, facilitating the making of a gray scale diagram;

[0022]

[0023] Wherein X i is the amplitude of the current point; X min is the minimum amplitude; X max is the maximum amplitude;

[0024] According to formula (3), the output interval of the one-dimensional signal is converted into a matrix of 0-255, and the matrix is rearranged to obtain a gray scale diagram;

[0025] Z i :Y i ×255(3)

[0026] Wherein Z i is the gray scale diagram matrix value; Y i is the normalized value.

[0027] In step five, the ratio of the training set, the verification set and the test set is 7:2:1, and before inputting the neural network model, the gray scale diagram resolution is adjusted to 128×128.

[0028] In step six, the preheating learning rate is first trained for a number of epochs using a small learning rate at the beginning of training, and then modified to a preset learning rate for training. The learning rate calculation formula is as follows:

[0029]

[0030] Wherein base_lr is the initial learning rate; step is the current iteration step; warm_epoch is the set warm-up epoch number; iter_one_epoch is the step number required for training one epoch;

[0031] The cosine annealing learning rate is accelerated decay and slow decay, and the specific formula is as follows:

[0032]

[0033] Where i is the current round index; η max is the maximum value of the learning rate, η min is the minimum value, T cur is the current round, T i is half a period.

[0034] The method is suitable for single metal medium and composite material medium.

[0035] The device used in the method comprises a processor, a memory and a computer program stored in the memory, and the processor executes the computer program to realize the small-scale sensing array composite material impact positioning method based on deep learning when working.

[0036] The storage medium used in the method stores computer instructions, and when the computer instructions are executed by the computer, the computer executes the small-scale sensing array composite material impact positioning method based on deep learning.

[0037] Compared with the prior art, the beneficial effects of the present application are:

[0038] Since the present application selects appropriate feature signals to convert into images for deep learning, uses the lightweight EfficientNetV2 network structure as the backbone network, extracts multi-scale features to classify and locate the gray-scale images, has the advantages of reducing the number of required sensors while ensuring positioning accuracy and search speed, and can adapt to the complexity of the composite material structure. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is the front view of the wind power blade medium of the embodiment of the present application.

[0040] Figure 2 is the side view of the wind power blade medium of the embodiment of the present application.

[0041] Figure 3 is the method flowchart of the embodiment of the present application.

[0042] Figure 4 is the grid area distribution layout schematic diagram of the embodiment of the present application.

[0043] Figure 5 is the impact signal diagram received by the sensor of the embodiment of the present application.

[0044] Figure 6This is a schematic diagram of the propagation of stress waves along different paths according to an embodiment of the present invention.

[0045] Figure 7 This is a spliced ​​image of the real and imaginary parts of the Fourier transform signal after an embodiment of the present invention.

[0046] Figure 8 This is a grayscale schematic diagram of an embodiment of the present invention. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the embodiments and accompanying drawings. The embodiments employ wind turbine blades, which are important structural components widely installed on wind turbines. During the operating cycle of a wind turbine, the blades will experience various external loads that may affect their structural health. Dynamic impact loads are one of the most common types of loads during wind turbine operation, potentially damaging the blades and jeopardizing the normal operation of the wind turbine. Wind turbine blades have a certain spatial dimension, but vibration sensors can often only be deployed in limited locations, making it impossible to measure at every location. Therefore, accurately locating impacts using a limited number of vibration sensors is a major challenge in monitoring the dynamic impact of wind turbine blades.

[0048] Example test conditions:

[0049] The experimental platform for the impact source location monitoring system mainly consists of wind turbine blades, piezoelectric accelerometers, oscilloscopes, and computers.

[0050] The test objects included wind turbine blades, steel balls, and their supports; among them, the reference... Figure 1 , Figure 2 The wind turbine blade measures 100cm × 100cm × 5cm. To minimize the impact of environmental noise, buffer barriers were added under all four edges of the blade. Each buffer barrier consists of a supporting steel plate and a rubber plate, with the supporting steel plate and rubber plate arranged from the bottom layer, and a total thickness of approximately 3cm. The supporting steel plate measures 20cm × 20cm × 1.2cm, and the rubber plate measures 20cm × 20cm × 2cm. The steel ball used in the experiment weighs 176g. Two piezoelectric accelerometers are arranged on the wind turbine blade, diagonally distributed, with each piezoelectric accelerometer having a sampling frequency of 2MHz.

[0051] Reference Figure 3 A method for impact localization of composite materials using a small-scale sensor array based on deep learning, comprising the following steps:

[0052] Step 1: Based on the positioning accuracy requirements, divide the impacted structure into grids, and number the grids from 1 to 64, such as... Figure 4As shown in the figure, the area is divided into 64 rectangular areas with equal length and width, sensor 1 is arranged at the upper left corner of the area, and sensor 2 is arranged at the lower right corner of the area;

[0053] Step two, install M piezoelectric acceleration sensors on the impacted structure to obtain impact signals generated when the object falls, wherein M = 2; reasonably arrange the sensor positions, and the piezoelectric acceleration sensors are arranged symmetrically at the center point; the required number of piezoelectric acceleration sensors is small, and 2 piezoelectric acceleration sensors can still complete impact positioning, greatly reducing the cost of structural health monitoring;

[0054] Step three, obtain the impact signals f(t) of multiple monitoring points in the target area, such as Figure 5 As shown in the figure, the impact signal f(t) includes direct wave components and boundary reflection components;

[0055] After obtaining the impact signals f(t) of multiple monitoring points in the target area, in order to reduce the difficulty of signal analysis and understanding, the traditional impact positioning method usually only uses the direct wave component and ignores the boundary reflection component. In fact, boundary reflection can change the propagation direction of stress waves, enrich the signal trajectory, and contain a lot of structure state information. In order to reduce the number of sensor channels, an effective strategy is to fully excavate and utilize the boundary reflection information (i.e. multi-path and multi-modal information) in the response signal. Different propagation path diagrams are shown in Figure 6 As shown in the figure, the figure shows the sensor equivalent mode under the direct path, the first reflection path and the second reflection path; in addition to the direct wave, the echo path after boundary reflection is equivalent to deploying sensors at other positions, which increases the information source of impact positioning. However, the traditional impact positioning method usually needs to separate the components of each path and each mode from these signals sufficiently and accurately in order to perform impact positioning, which puts high requirements on signal processing technology, thereby limiting its wide application in practical application. However, the deep learning method can effectively solve this problem.

[0056] Step four, according to the multi-path propagation principle, the single sensor received signal is regarded as the multi-sensor received signal at different positions, and the real part and imaginary part signals after Fourier transform of the impact signal f(t) are extracted as features;

[0057] In this embodiment, the impact signal f(t) obtained by each piezoelectric acceleration sensor is Fourier transformed according to formula (1),

[0058]

[0059] Where F(ω) is the frequency spectrum function after Fourier transform, the independent variable is ω, which is the frequency; e -jωt is a complex signal; f(t) is an impact signal;

[0060] It is found that using the real and imaginary parts as input is better than using the amplitude and phase and other features as input;

[0061] After Fourier transform, the real and imaginary parts are extracted and sliced, and the first K non-zero data are taken, since the real and imaginary part signals are distributed below 20 kHz, and the neural network model input picture size is 128x128, only the first 128 data points are taken, which can effectively reduce the signal size and improve the data processing efficiency; In this embodiment, the first K=128 non-zero data of the real and imaginary parts are taken, as shown in Figure 7 The first 128 data are real part signals, and the last 128 data are imaginary part signals;

[0062] Step five, establish a data set, normalize the real and imaginary part features and make a gray scale image, and divide the data set into a training set, a validation set and a test set;

[0063] In this embodiment, the data of M=1 and M=2 piezoelectric acceleration sensors are spliced, normalized according to formula (2), so that the output is between 0 and 1, and the dimensional expression is changed into a dimensionless expression, which is convenient for making a gray scale image;

[0064]

[0065] Where X i is the amplitude of the current point; X min is the minimum amplitude; X max is the maximum amplitude;

[0066] According to formula (3), the output interval of the one-dimensional signal is converted into a 0-255 matrix, and the matrix is rearranged to obtain a 128x128 gray scale image, as shown in Figure 8 ;

[0067] Z i : Y i x 255 (8) where Z i is the gray scale image matrix value; Y i is the normalized value;

[0068] In this embodiment, the ratio of the training set, the validation set and the test set is 7:2:1, and before inputting the neural network, the gray scale image resolution is adjusted to 128x128;

[0069] Step six, build a neural network model, load the pre-training weight of EfficientNetV2, and use the pre-warm learning rate to start training; in the subsequent stage of training, the cosine annealing learning rate scheduling strategy is used to optimize the model training, which maintains a high learning rate in the early stage of training to quickly accelerate the convergence speed of the model, and then gradually reduces the learning rate to avoid fluctuations in the later stage of training; in the training iteration, the loss is calculated and the model weight is updated; after each training period, the model performance is evaluated using the validation set, and the loss and accuracy indicators on the validation set are monitored;

[0070] The pre-warm learning rate in step six is to use a small learning rate to train for a few epochs at the beginning of training, and then modify the preset learning rate for training to accelerate the model convergence speed and improve the model effect, the learning rate calculation formula is as follows:

[0071]

[0072] Where base_lr is the initial learning rate; step is the current iteration step; warm_epoch is the set warm-up epoch number; iter_one_epoch is the step number required for training one epoch;

[0073] The cosine annealing learning rate is first accelerated and then slowly decayed, and the specific formula is as follows:

[0074]

[0075] Where i is the current round index; η max is the maximum learning rate, η min is the minimum value, T cur is the current round, T i is half a period;

[0076] Step seven, input the gray image of the impact signal test set into the trained neural network model to obtain the region position of the impact source of the impact signal;

[0077] Step eight, display the positioning result, the following method is used to represent the positioning result in this embodiment: horizontal difference: the horizontal difference between the positioning position and the actual collision position; vertical difference: the vertical difference between the positioning position and the actual collision position, Table 1 is a single sensor steel ball impact experiment result record table, and Table 2 is a double sensor steel ball impact experiment result record table.

[0078] Table 1 Single sensor steel ball impact experiment result record table

[0079]

[0080]

[0081] As can be seen from Table 1, the total correct recognition rate of the test set of steel balls falling in all positions is 89.06%, but there are some areas where the recognition error is too large. Overall, the recognition effect of the single sensor is better, and it can basically complete the rapid recognition of impact positioning.

[0082] Table 2: Double sensor steel ball impact experiment result record table

[0083]

[0084]

[0085] As can be seen from Table 2, the total correct recognition rate of the test set of steel balls falling in all positions is 93.75%, and the horizontal and vertical difference of the recognition error position is not greater than 1, and there is no case where the recognition error of some areas is too large. The recognition effect of the double sensor is better than that of the single sensor, and it can complete the rapid recognition of impact positioning.

[0086] The content described in the embodiments of the specification is only a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms stated in the embodiments, and the protection scope of the present application also extends to equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A method for impact localization of small scale sensor array composites based on deep learning, characterized in that, The method comprises the following steps: Step 1: According to the positioning accuracy requirement, divide the grid on the impacted structure, and number the divided grid 1-N; Step 2: Install M piezoelectric acceleration sensors on the impacted structure to obtain the impact signal generated when the object falls, wherein M = 1 or 2, and the sensor position is arranged; Step 3: Obtain the impact signal f(t) of multiple monitoring points in the target area, and the impact signal f(t) includes direct wave component and boundary reflection component; Step 4: According to the principle of multi-path propagation, the impact signal f(t) is taken as a reference target, and the real part and imaginary part signals after Fourier transform are extracted as features; Step 5: Establish a data set, normalize the real part and imaginary part features, and make a gray scale image, and divide the data set into a training set, a validation set and a test set; Step 6: Build a neural network model, load the pre-trained weights of EfficientNetV2, and use a warm-up learning rate to start training; in the subsequent stage of training, a cosine annealing learning rate scheduling strategy is used to optimize the model training, which maintains a high learning rate in the early stage of training, and then gradually reduces the learning rate, in the training iteration, the loss is calculated and the model weight is updated; after each training cycle, the model performance is evaluated using the validation set, and the loss and accuracy indicators on the validation set are monitored; Step 7: Input the gray scale image of the impact signal test set into the trained neural network model to obtain the position of the impact source of the impact signal; Step 8: Display the positioning result.

2. The positioning method of claim 1, wherein: In step 2, if M = 1, the sensor is arranged at the center point of the target area; if M = 2, the sensor is symmetrically arranged with the center point of the target area as the center.

3. The positioning method of claim 1, wherein: In step 4, the impact signal f(t) obtained by each piezoelectric acceleration sensor is Fourier transformed according to formula (1), where F(ω) is the Fourier transformed frequency spectrum function, the independent variable is ω, which is frequency; e -jωt is a complex signal; f(t) is an impact signal; After Fourier transform, the real part and imaginary part are extracted and sliced, and the first K non-zero data are taken.

4. The positioning method of claim 2, wherein: In step 5, the data of M piezoelectric acceleration sensors are spliced to obtain 2*M*K data points, which are normalized according to formula (2) to make the output between 0 and 1, and the dimensional expression is changed into a dimensionless expression, which is convenient for making a gray scale image; where X i current point amplitude; X min is the minimum amplitude; X max is the maximum amplitude; According to formula (3), the output interval of the one-dimensional signal is converted to a 0-255 matrix, and the matrix is rearranged to obtain a gray scale image; Z i :Y i ×255(13)where Z i is a grayscale matrix value; Y i is a normalized value.

5. The positioning method of claim 1, wherein: In step 5, the proportions of the training set, the validation set and the test set are 7:2:1, and the gray scale image resolution is adjusted to 128*128 before inputting into the neural network model.

6. The positioning method of claim 1, wherein: In step 6, the warm-up learning rate is calculated as follows: Wherein base_lr is the initial learning rate; step is the current iteration step; warm_epoch is the set warm-up epoch number; iter_one_epoch is the step number required for training one epoch; The cosine annealing learning rate is first accelerated and then slowly decayed, and the specific formula is as follows: Wherein i is the current round index; η max is the maximum value of the learning rate, η min is the minimum value, T cur is the current epoch, T i is half a period.

7. The positioning method of claim 1, wherein: The method is suitable for single metal medium and composite material medium.

8. A device for a small-scale sensor array composite impact location method based on deep learning, characterized in that, It comprises: The processor, the memory and the computer program stored on the memory, the processor executes the computer program to realize the method for locating impact of small-scale sensing array composite material based on deep learning according to any one of claims 1-6 when working.

9. A storage medium used in a deep learning-based small-scale sensor array composite material impact localization method, characterized in that, The storage medium stores computer instructions, when the computer instructions are executed by the computer, the computer executes the method for locating impact of small-scale sensing array composite material based on deep learning according to any one of claims 1-6.

Images