Small-scale sensing array composite material impact positioning method based on deep learning

Abstract

A small-scale sensing array composite material impact positioning method based on deep learning comprises the following steps: firstly, dividing a grid on an impacted structure, installing one or two sensors on the impacted structure, and obtaining an impact signal of a monitoring point; then taking the whole impact signal as a reference target, extracting real part and imaginary part signals after Fourier transform as features, establishing a data set, normalizing the features of the real part and the imaginary part, making a grey-scale map, and dividing the data set into a training set, a verification set and a test set; secondly, constructing a neural network model, loading a pre-training weight of the OfficientNetV2, starting training by using a preheating learning rate, and optimizing model training by adopting a cosine annealing learning rate scheduling strategy in a subsequent stage of training; and finally, inputting the grey-scale map of the impact signal test set into the trained neural network model to obtain the position of a region where an impact source of the impact signal is located. While the positioning precision and the searching speed are ensured, the number of sensors is reduced, and the method can adapt to the complexity of a composite material medium.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and in particular to a small-scale sensor array composite material 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. Therefore, they have been widely used in the wind power industry and 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 their integrity and reliability may be affected by sudden events such as wind and sand, hail and flying animals. These unexpected impacts may cause hidden delamination, cracks and other minor damages inside the composite material, thereby reducing the bearing capacity of the structure. If effective monitoring and repair are not carried out in time, these damages may further expand and cause significant losses. Therefore, impact monitoring is particularly important in the structural health monitoring of wind turbine blades.

[0003] At present, the sensors commonly used in impact location 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 arrival time difference of the impact response signal, and determined the impact point position through an equivalent right-angle triangle sensor array. In 2024, Zhong Zhaozhen et al. used OFDR optical fiber sensors to obtain the strain response information of composite laminate structures, and on this basis proposed a method for impact load location based on the principle of nonlinear weighting of strain amplitude. These methods usually require multiple sensing nodes, which increases the burden of cables, acquisition channels, storage requirements and processing units, thereby increasing the overall weight and cost of the structural health monitoring system; in addition, the characteristics of composite materials make the propagation and attenuation laws of acoustic emission signals significantly different from those of single metal materials, which limits the application of traditional impact location 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 obtain satisfactory positioning accuracy with fewer sensors. They are also applicable to simple flat 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 problem that errors are difficult to eliminate and the positioning accuracy is insufficient, and the actual engineering application effect is poor. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a small-scale sensor array composite material impact positioning method based on deep learning, which selects feature signals and converts them into images for deep learning, uses a lighter EfficientNetV2 network structure as the backbone network, extracts multi-scale features to classify and locate grayscale images, and effectively reduces the number of sensor channels.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

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

[0008] Step 1: According to the positioning accuracy requirements, grids are divided on the impacted structure, and the divided grids are numbered 1 to N;

[0009] Step 2: Install M piezoelectric acceleration sensors on the impacted structure to obtain the impact signal generated when the object falls, where M=1 or 2, and arrange the sensor positions;

[0010] Step 3, obtaining the impact signal f(t) of multiple monitoring points in the target area, where the impact signal f(t) includes a direct wave component and a boundary reflection component;

[0011] Step 4: According to the principle of multipath propagation, the impulse signal f(t) as a whole is used as a reference target, and the real and imaginary signals after Fourier transformation are extracted as features;

[0012] Step 5: Create a data set, normalize the real and imaginary features and make them into grayscale images, and divide the data set into training set, validation set, and test set;

[0013] Step 6: Build a neural network model, load the pre-trained weights of EfficientNetV2, and start training using the warm-up learning rate. In the subsequent stages of training, use the cosine annealing learning rate scheduling strategy to optimize model training. The cosine annealing learning rate scheduling strategy maintains a high learning rate at the beginning of training and then gradually reduces the learning rate. In the training iteration, the loss is calculated and the model weights are updated. After each training cycle, the validation set is used to evaluate the model performance, and the loss and accuracy indicators on the validation set are monitored.

[0014] Step 7, input the grayscale image of the impact signal test set into the trained neural network model to obtain the location of the impact source of the impact signal;

[0015] Step 8: Display the positioning results.

[0016] In the step 2, if M=1, the sensors are arranged at the center point of the target area; if M=2, the sensors are arranged symmetrically around the center point of the target area.

[0017] In step 4, the impact signal f(t) obtained by each piezoelectric acceleration sensor is Fourier transformed according to formula (1):

[0018]

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

[0020] After Fourier transform, extract the real and imaginary parts and slice them, taking the first K non-zero data of each.

[0021] In the 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) so that the output is between 0 and 1, and the dimensional expression is converted into a dimensionless expression to facilitate the production of a grayscale image;

[0022]

[0023] Where X i Current point amplitude; X min is the minimum amplitude; X max is the maximum value;

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

[0025] Z i :Y i ×255(3)

[0026] Where Z i is the grayscale image matrix value; Y i is the normalized value.

[0027] In the step 5, the ratio of the training set, validation set and test set is 7:2:1, and the grayscale image resolution is adjusted to 128×128 before inputting into the neural network model.

[0028] In step 6, the warm-up learning rate is to first use a small learning rate to train for several epochs at the beginning of training, and then change it to the preset learning rate for training. The learning rate calculation formula is as follows:

[0029]

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

[0031] The cosine annealing learning rate first decays rapidly and then decays slowly. The specific formula is as follows:

[0032]

[0033] 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 For half a cycle.

[0034] The method is applicable to single metal media and composite material media.

[0035] The device used in the method includes: a processor, a memory, and a computer program stored in the memory. When the processor is working, it executes the computer program to implement the small-scale sensor array composite material impact positioning method based on deep learning.

[0036] The method adopts a storage medium, which stores computer instructions. When the computer instructions are executed by a computer, the computer executes the small-scale sensor array composite material impact positioning method based on deep learning.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

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

[0039] Figure 1 It is a front view of the wind turbine blade medium according to an embodiment of the present invention.

[0040] Figure 2 It is a side view of the medium of a wind turbine blade according to an embodiment of the present invention.

[0041] Figure 3 It is a flow chart of a method according to an embodiment of the present invention.

[0042] Figure 4 It is a schematic diagram of the grid area distribution layout of an embodiment of the present invention.

[0043] Figure 5 4 is a diagram of an impact signal received by a sensor according to an embodiment of the present invention.

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

[0045] Figure 7 It is a spliced ​​image of the real and imaginary signal after Fourier transformation in an embodiment of the present invention.

[0046] Figure 8 It is a schematic diagram of a grayscale image according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below in conjunction with the embodiments and drawings. The embodiments use wind turbine blades, which are important structural components widely installed on wind turbines. During the operation cycle of the wind turbine, the wind turbine blades will experience various external loads that may affect the health of their structures. Dynamic impact loads are one of the most common load types when the wind turbine is running, which may cause damage to the wind turbine blades and endanger the normal operation of the wind turbine. Wind turbine blades have certain spatial dimensions, but vibration sensors can often only be deployed in limited locations, and it is impossible to measure at each location. Therefore, using a limited number of vibration sensors to accurately locate the impact has become a major difficulty in dynamic impact monitoring of wind turbine blades.

[0048] Example Experimental conditions:

[0049] The impact source position monitoring system experimental platform is mainly composed of wind turbine blades, piezoelectric acceleration sensors, oscilloscopes and computers;

[0050] The test objects include wind turbine blades, steel balls and their supports; Figure 1 , Figure 2 The size of the wind turbine blade is 100cm×100cm×5cm. In order to minimize the impact of environmental noise, buffer isolation is added under the four edges of the wind turbine blade. Each buffer isolation consists of a supporting steel plate and a rubber plate. Starting from the bottom layer, they are supporting steel plate and rubber plate respectively, with a total thickness of about 3cm; among them, the size of the supporting steel plate is 20cm×20cm×1.2cm, and the size of the rubber plate is 20cm×20cm×2cm; the weight of the steel ball used in the experiment is 176g; two piezoelectric accelerometers are arranged on the wind turbine blade, and the piezoelectric accelerometers are diagonally distributed. The sampling frequency of each piezoelectric accelerometer is 2MHz.

[0051] Reference Figure 3 , a small-scale sensor array composite material impact location method based on deep learning, comprising the following steps:

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

[0053] Step 2: Install M piezoelectric acceleration sensors on the impacted structure to obtain the impact signal generated when the object falls, where M=2; reasonably arrange the sensor positions, and arrange the piezoelectric acceleration sensors symmetrically around the center point; the number of piezoelectric acceleration sensors required is small, and two piezoelectric acceleration sensors can still complete the impact positioning, which greatly reduces the cost of structural health monitoring;

[0054] Step 3: Obtain the impact signal f(t) of multiple monitoring points in the target area, such as Figure 5 As shown, the impact signal f(t) includes a direct wave component and a boundary reflection component;

[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 location method usually only uses the direct wave component and ignores the boundary reflection component. In fact, the boundary reflection will change the propagation direction of the stress wave, enrich the signal trajectory, and contain a lot of structural state information. In order to reduce the number of sensor channels, an effective strategy is to fully explore and utilize the boundary reflection information in the response signal (i.e., multipath and multimodal information); the schematic diagram of different propagation paths is shown in Figure 6 As shown in the figure, the equivalent modes of sensors under the direct path, the primary reflection path, and the secondary reflection path are respectively shown; in addition to the direct wave, the echo path reflected by the boundary is equivalent to deploying sensors at other locations, which increases the information source of impact positioning; however, the traditional impact positioning method usually needs to fully and accurately separate these signals into the components of each path and each mode in order to perform impact positioning, which puts forward high requirements on signal processing technology, thus limiting its extensiveness in practical applications, and deep learning methods can effectively solve this problem;

[0056] Step 4: According to the principle of multipath propagation, the single sensor receiving signal is regarded as the multi-sensor receiving signal at different positions, the impulse signal f(t) as a whole is taken as the reference target, and the real and imaginary signals after Fourier transformation are extracted as features;

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

[0058]

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

[0060] After testing, it is found that using the real and imaginary features as input has better effects than using the amplitude, phase and other features;

[0061] After Fourier transform, the real part and the imaginary part are extracted and sliced, and the first K non-zero data are taken. Since the frequency distribution of the real and imaginary signals is below 20kHz, the input image size of the neural network model is 128×128, so only the first 128 data points of the signal are taken, which can effectively reduce the signal size and improve the efficiency of data processing; this embodiment takes the first K=128 non-zero data of the real and imaginary signals, such as Figure 7 As shown, the first 128 data in the figure are real part signals, and the last 128 data are imaginary part signals;

[0062] Step 5: Create a data set, normalize the real and imaginary features and make them into grayscale images, and divide the data set into training set, validation set, and test set;

[0063] In this embodiment, the data of M=1 and M=2 piezoelectric acceleration sensors are spliced ​​and normalized according to formula (2) so that the output is between 0 and 1, and the dimensionless expression is converted into a dimensionless expression to facilitate the production of grayscale images.

[0064]

[0065] Where X i Current point amplitude; X min is the minimum amplitude; X max is the maximum value;

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

[0067] Z i :Y i ×255(8) where Z i is the grayscale image matrix value; Y i is the normalized value;

[0068] In this embodiment, the ratio of the training set, validation set and test set is 7:2:1, and the grayscale image resolution is adjusted to 128×128 before input into the neural network;

[0069] Step 6: Build a neural network model, load the pre-trained weights of EfficientNetV2, and start training using the warm-up learning rate. In the subsequent stages of training, use the cosine annealing learning rate scheduling strategy to optimize model training. The cosine annealing learning rate scheduling strategy maintains a high learning rate at the beginning of training to quickly accelerate the convergence of the model, and then gradually reduces the learning rate to avoid fluctuations in the later stages of training. In the training iteration, calculate the loss and update the model weights. After each training cycle, use the validation set to evaluate the model performance and monitor the loss and accuracy indicators on the validation set.

[0070] In step 6, the warm-up learning rate is to use a small learning rate to train for several epochs at the beginning of training, and then change it to the preset learning rate for training to speed up the model convergence 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 number; warm_epoch is the set number of warm-up epochs; iter_one_epoch is the number of steps required for training iteration one epoch;

[0073] The cosine annealing learning rate first decays rapidly and then decays slowly. 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 cycle;

[0076] Step 7, input the grayscale image of the impact signal test set into the trained neural network model to obtain the location of the impact source of the impact signal;

[0077] Step eight, display the positioning results. This embodiment uses the following method to represent the positioning results: lateral difference: the lateral difference between the positioning position and the actual collision position; vertical difference: the longitudinal difference between the positioning position and the actual collision position. Table 1 is a single-sensor steel ball impact test result record table, and Table 2 is a dual-sensor steel ball impact test result record table.

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

[0079]

[0080]

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

[0082] Table 2 Dual sensor steel ball impact test result record

[0083]

[0084]

[0085] As can be seen from Table 2, the total recognition accuracy of the steel ball falling in the test set is 93.75% in all positions, and the horizontal and vertical differences of the wrong positions are not greater than 1, and there is no excessive recognition error in some areas. The recognition effect of the dual sensor is better than that of the single sensor, and it can complete the rapid recognition of the impact location.

[0086] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A small-scale sensor array composite material impact positioning method based on deep learning, characterized in that: The following steps are involved: Step 1: According to the positioning accuracy requirements, grids are divided on the impacted structure, and the divided grids are numbered 1 to N; Step 2: Install M piezoelectric acceleration sensors on the impacted structure to obtain the impact signal generated when the object falls, where M=1 or 2, and arrange the sensor positions; Step 3, obtaining the impact signal f(t) of multiple monitoring points in the target area, where the impact signal f(t) includes a direct wave component and a boundary reflection component; Step 4: According to the principle of multipath propagation, the impulse signal f(t) as a whole is used as a reference target, and the real and imaginary signals after Fourier transformation are extracted as features; Step 5: Create a data set, normalize the real and imaginary features and make them into grayscale images, and divide the data set into training set, validation set, and test set; Step 6: Build a neural network model, load the pre-trained weights of EfficientNetV2, and start training using the warm-up learning rate. In the subsequent stages of training, use the cosine annealing learning rate scheduling strategy to optimize model training. The cosine annealing learning rate scheduling strategy maintains a high learning rate at the beginning of training and then gradually reduces the learning rate. In the training iteration, the loss is calculated and the model weights are updated. After each training cycle, the validation set is used to evaluate the model performance, and the loss and accuracy indicators on the validation set are monitored. Step 7, input the grayscale image of the impact signal test set into the trained neural network model to obtain the location of the impact source of the impact signal; Step 8: Display the positioning results.

2. The positioning method according to claim 1, characterized in that: In the step 2, if M=1, the sensors are arranged at the center point of the target area; if M=2, the sensors are arranged symmetrically around the center point of the target area.

3. The positioning method according to claim 1, characterized in that: In the step 4, the impact signal f(t) obtained by each piezoelectric acceleration sensor is subjected to Fourier transform according to formula (1): Where F(ω) is the spectrum function after Fourier transformation, and the independent variable is ω, which is the frequency; e -jωt is a complex signal; f(t) is an impulse signal; After Fourier transform, extract the real and imaginary parts and slice them, taking the first K non-zero data of each.

4. The positioning method according to claim 2, characterized in that: In the 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) so that the output is between 0 and 1, and the dimensional expression is converted into a dimensionless expression to facilitate the production of a grayscale image; Where X i Current point amplitude; X min is the minimum amplitude; X max is the maximum value; According to formula (3), the output interval of the one-dimensional signal is converted into a matrix of 0 to 255, and the matrix is ​​rearranged to obtain a grayscale image; Z i :Y i ×255(13) where Z i is the grayscale image matrix value; Y i is the normalized value.

5. The positioning method according to claim 1, characterized in that: In step 5, the ratio of the training set, validation set and test set is 7:2:1, and the grayscale image resolution is adjusted to 128×128 before inputting into the neural network model.

6. The positioning method according to claim 1, characterized in that: In step 6, the warm-up learning rate is to first use a small learning rate to train for several epochs at the beginning of training, and then change it to the preset learning rate for training. The learning rate calculation formula is as follows: Where base_lr is the initial learning rate; step is the current iteration number; warm_epoch is the set number of warm-up epochs; iter_one_epoch is the number of steps required for training iteration one epoch; The cosine annealing learning rate first decays rapidly and then decays slowly. The specific formula is as follows: 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 For half a cycle.

7. The positioning method according to claim 1, characterized in that: The method is applicable to single metal media and composite material media.

8. A device used in a small-scale sensor array composite material impact positioning method based on deep learning, characterized in that: include: A processor, a memory, and a computer program stored in the memory, wherein the processor executes the computer program when working to implement a small-scale sensor array composite material impact positioning method based on deep learning as described in any one of claims 1-6.

9. A storage medium used in a small-scale sensor array composite material impact positioning method based on deep learning, characterized in that: The storage medium stores computer instructions. When the computer instructions are executed by a computer, the computer executes a small-scale sensor array composite material impact positioning method based on deep learning as described in any one of claims 1 to 6.

Images