Deep Learning-Based 3D Reconstruction Defect Detection System and Method Based on Laser Ultrasound

Abstract

This invention provides a deep learning-based 3D reconstruction defect detection system and method based on laser ultrasound. The system includes a motion control module, a laser scanning module, a data acquisition module, a signal processing module, and a result display module. It generates laser ultrasound signals on the surface of aerospace composite materials through step-by-step laser scanning, and uses a 2D motion platform to carry a laser vibrometer for signal acquisition. Finally, it automatically reconstructs the internal structure of the part using a LANDT (Laser-Oriented Derivative) network based on generative adversarial networks, outputting a stitched 3D image of the part. This allows for automated detection of surface and internal defects in composite materials, and real-time 3D imaging of the detection results, enabling intuitive observation of internal defects. This achieves the goal of effectively improving the accuracy of surface and internal defect detection in composite materials while significantly reducing production costs.

Description

Technical Field

[0001] This invention relates to the field of laser ultrasonic nondestructive testing technology, and in particular to a deep learning-based three-dimensional reconstruction composite defect detection system and method based on laser ultrasonics. Background Technology

[0002] With the rapid development of the aerospace manufacturing industry, fiber-reinforced composite materials have been widely used in aerospace vehicles in recent years. This is considered an effective way to reduce the weight of aircraft and lower direct maintenance costs. Carbon fiber composites are a perfect combination of carbonized fibers and plastics, and have unique advantages compared to metals: they are strong, very rigid, and especially lightweight. The application of carbon fiber composites in aircraft can reduce their weight by up to 30%, a feat difficult to achieve with other technologies. Furthermore, this material combines the flexibility and weavability of fibers, allowing it to serve both as a structural load-bearing material and a functional material. Therefore, developing larger and lighter carbon fiber composite blades is a trend, and researching more testing methods for carbon fiber composite blades is also of great significance.

[0003] Laser ultrasonic testing technology is a non-destructive testing technique that uses lasers to excite and detect ultrasound. Compared with traditional piezoelectric ultrasonic technology, laser ultrasonic testing technology has advantages such as non-contact, broadband, and point emission and reception. Therefore, it is applied in material characterization, defect detection, processing monitoring, and the detection or monitoring of workpieces with complex morphologies or equipment in special environments such as high temperature, high pressure, corrosion, and radiation.

[0004] Laser ultrasonic testing technology, as an interdisciplinary technology, combines the advantages of ultrasonics and laser technology. It utilizes the strong penetrating power of traditional ultrasound and the non-contact nature of optical testing, achieving non-destructive testing of the test sample by exciting and receiving ultrasound waves with a laser.

[0005] Compared to other non-destructive testing (NDT) technologies, laser ultrasonic testing offers advantages such as low cost, ease of operation, and high precision, and has great potential for future development in the field of NDT. However, existing laser ultrasonic excitation and testing methods generally use LabVIEW for programming and control. While this can accurately produce test results, the programming method is not flexible enough and cannot visually display the test results inside the material, increasing the difficulty of analyzing internal defects. Industrial CT technology can visually visualize internal defects in materials, but it is limited by the high cost of the equipment and cannot inspect large or thick parts.

[0006] Chinese Patent Publication No. CN101435784B discloses a CT inspection device and method for turbine blades. This method uses X-rays to irradiate the blade and adjusts the blade's tilt angle to detect cracks in various directions. However, this method uses an expensive X-ray source, the device is bulky and cannot perform in-situ inspections, and requires adjusting the blade angle to prevent missing cracks perpendicular to the X-rays, resulting in slow imaging speed and low inspection efficiency.

[0007] Chinese Patent Publication No. CN111537444A discloses a laser ultrasonic nondestructive testing method and system with virtual repetition frequency control. It utilizes a laser ultrasonic system to scan the structure under test, acquiring guided wave field data excited at a low repetition frequency; reconstructs guided wave field data excited at a virtual repetition frequency based on the low repetition frequency excited wave field data; calculates damage images for different repetition frequencies based on the virtual repetition frequency excited wave field data; determines the laser excitation repetition frequency with optimal detection effect based on the damage images; and performs nondestructive testing based on the laser excitation repetition frequency with optimal detection effect. This testing method can achieve high detection accuracy. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the above-mentioned background technology, realize the location and three-dimensional reconstruction of internal defects in materials, and perform internal three-dimensional reconstruction of the inspected parts by combining laser ultrasound and deep learning, so as to intuitively display the material surface and internal defects. This invention is used to solve the technical problems of expensive industrial CT, inability to inspect large or thick parts, and inflexible programming of laser ultrasound inspection in industrial inspection.

[0009] To achieve the above objectives, this invention first provides a low-cost, high-performance laser ultrasound-based deep learning-based 3D reconstruction defect detection system, which includes:

[0010] XY two-dimensional motion platform, Q-switched Nd:YAG laser, XZ two-dimensional motion platform, Vibroflex laser vibrometer, two-dimensional deflection mirror, aerospace composite material parts under test, industrial computer, monitor.

[0011] The XY two-dimensional motion platform provides displacement in the XY direction, the Q-switched Nd:YAG laser generates continuous laser pulses, the XZ two-dimensional motion platform provides two-dimensional motion in the XZ direction, the Vibroflex laser vibrometer detects the vibration of the parts, the two-dimensional deflection mirror realizes continuous deflection scanning of the laser, and the industrial control computer includes a control module and a signal processing module.

[0012] Preferably, the industrial control computer is the core of the entire system's control and data processing, and its software includes the aforementioned control module and signal processing module. The control module communicates via serial port with the aforementioned XY two-dimensional motion platform, Q-switched Nd:YAG laser, XZ two-dimensional motion platform, Vibroflex laser vibrometer, and two-dimensional deflection mirror to achieve continuous and efficient laser scanning and to acquire data on part vibration. The signal processing module processes the images obtained from the vibrometer, progressively realizing signal waveform plotting, part scanning point thickness measurement, part outline plotting, part scanning point defect depth location, and layered plotting of the internal image of the part scanning points, ultimately completing the three-dimensional reconstruction of the part under test.

[0013] Secondly, this invention also provides a laser ultrasound-based deep learning-based three-dimensional reconstruction defect detection method, which, based on the aforementioned laser ultrasound-based deep learning-based three-dimensional reconstruction defect detection system, includes the following steps:

[0014] S101: Initializes hardware parameters via industrial control computer, including parameters for the pulsed laser, 2D galvanometer, laser vibrometer, and 2D motion platform. Sets the scanning parameters for the laser and 2D galvanometer, and matches the displacement parameters of the 2D motion platform equipped with the laser vibrometer to them.

[0015] S102: The laser emits continuous pulsed laser light, which directly hits the two-dimensional galvanometer. The laser light is continuously deflected by the synchronous scanning of the two-dimensional galvanometer, completing the continuous scanning of the current position on the aerospace composite material, thereby generating a laser ultrasonic signal on the surface of the composite material to be tested.

[0016] S103: By controlling the XZ-axis two-dimensional motion platform, the laser vibrometer moves synchronously with the laser scanning point on the part under test, always maintaining a vertical position to collect laser ultrasonic signals, thereby obtaining continuous laser scanning signals, and saving these signal data to the industrial control computer data processing module.

[0017] S104: The laser ultrasonic signal collected by the vibration meter is preprocessed, the ultrasonic signal is analyzed and the waveform diagrams in the time domain and frequency domain are obtained, and the thickness of the material at the current scanning point is obtained according to the pulse time interval, so that the three-dimensional contour map of the material at the scanning position can be obtained.

[0018] S105: After obtaining the laser ultrasonic waveform, it is imported into the innovatively proposed LANDT neural network for training. The initial training parameters, learning rate and batch size, are set. After 300 training rounds, the defects on the surface and inside of the part under test can be automatically analyzed, and the defects can be reconstructed layer by layer using position and depth information to generate the three-dimensional contour map in the above steps.

[0019] S106: The XY two-dimensional motion platform moves to the next position and repeats the above steps S102-S105. After the parts in each laser scanning area are reconstructed in three dimensions, the images of the reconstructed parts in each scanning area are stitched together according to the displacement of the XY two-dimensional motion platform to obtain the final overall three-dimensional reconstruction image of the parts.

[0020] S107: Outputs the overall 3D reconstructed part image to the display, thereby completing the automated detection of surface and internal defects of the part and allowing for intuitive observation of the location and size of defects.

[0021] Furthermore, in step S105, this embodiment innovatively proposes a LANDT neural network, which is based on a generative adversarial network. Its model parameters and dataset are obtained through the following steps:

[0022] Step 1: Perform laser ultrasonic testing on as many aerospace composite materials containing known defects as possible. Manually annotate the obtained time-domain and frequency-domain waveforms, indicate the defect locations and record their image positions, and generate the txt file required by the LANDT neural network from the label information.

[0023] Step 2: Manually generate 3D images of the inspected parts, mark the defect locations, and correlate them with the waveform and label information in Step 1 to obtain the training set for the Generative Adversarial Network (LANDT).

[0024] Step 3: Set the training parameters: learning rate, batch size, and number of training epochs. Input the waveform, 3D image, and labels generated in the previous two steps into the LANDT network for training. Observe the loss curve during training to determine whether the LANDT network has converged. If it has converged, stop training; if it has not converged, continue training. Preferably, the learning rate is set to 0.01, the batch size is set to 16, and the number of training epochs is set to 300 to obtain the trained model.

[0025] Step 4: In actual testing, the vibration meter data after laser scanning is preprocessed to generate a waveform diagram, which is then imported into the training result model from Step 3 for forward input. This yields the final 3D reconstruction result, completing the detection and visualization of part defects. From the above technical solution, it can be seen that the embodiments of this application have the following advantages:

[0026] This application provides a laser ultrasound-based deep learning-based 3D reconstruction defect detection system and method, which creatively combines the advantages of laser ultrasound detection technology and industrial CT technology. It performs regional laser ultrasound scanning on aerospace composite materials, controls the laser to always follow the scanning point in two-dimensional motion via a two-dimensional motion platform, and collects the laser ultrasound signals excited by the laser at the corresponding positions. Then, it constructs a LANDT network based on generative adversarial networks to perform regional 3D reconstruction of the part's interior, and finally stitches the 3D reconstructed images of each region into a complete 3D image of the part. Compared with other industrial CT and laser ultrasound systems, this system is lower in cost, simpler in structure, and has powerful and flexible software functions, making it easy to deploy and apply. It achieves the goal of real-time automated detection and 3D visualization of the interior of large and thick parts, improving the accuracy and effectiveness of the detection, and enabling inspectors to perform product analysis more conveniently and intuitively in the future. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A schematic diagram of the installation location of a laser ultrasound-based deep learning-based three-dimensional reconstruction defect detection device according to an embodiment of the present invention is shown.

[0029] Figure 2 A schematic diagram of the overall process of a deep learning-based 3D reconstruction defect detection method based on laser ultrasound is shown.

[0030] Figure 3 A schematic diagram illustrating the working principle of a deep learning-based 3D reconstruction defect detection system based on laser ultrasound is shown.

[0031] Figure 4 A schematic diagram of a typical defective aerospace composite material part is shown.

[0032] Figure 1 In the diagram: 101 is the XY two-dimensional motion platform, 102 is the Q-switched Nd:YAG laser, 103 is the XZ two-dimensional motion platform, 104 is the Vibroflex laser vibrometer, 105 is the two-dimensional deflection mirror, 106 is the aerospace composite material part to be tested, 107 is the industrial control computer, and 108 is the display. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] For easier understanding, please refer to Figure 1 This invention provides a laser-ultrasound-based deep learning-based three-dimensional reconstruction system for detecting defects in aerospace composite materials. The system includes an XY two-dimensional motion platform 101 providing XY-axis displacement, a Q-switched Nd:YAG laser 102 generating continuous laser pulses, an XZ two-dimensional motion platform 103 providing XZ-axis two-dimensional motion, a Vibroflex laser vibrometer 104 detecting part vibration, a two-dimensional deflection mirror 105 realizing continuous deflection scanning of the laser, an aerospace composite material part to be tested 106, an industrial control computer 107 including a control module and a signal processing module, and a display 108.

[0035] It should be noted that the XY two-dimensional motion platform 101, the Q-switched Nd:YAG laser 102, and the motion software controlled by the industrial control computer 107, through XY-axis motion, drive the Q-switched Nd:YAG laser 102 and the two-dimensional deflection mirror 105 to perform X and Y-axis displacements, completing the laser scanning work of the laser ultrasonic testing system, causing the aerospace composite material part 106 to be tested to vibrate for defect detection; the Q-switched Nd:YAG laser 102 is the core of the entire testing system. The laser internally controls the continuous excitation of laser pulses by controlling the Q-switch, generating continuous laser pulses; the two-dimensional deflection mirror 105 continuously deflects the laser beam generated by the laser. The continuous laser pulses emitted by the Q-switched Nd:YAG laser 102 are directly incident on the mirror 105, and then the continuous deflection scanning by the two-dimensional deflection mirror 105 and the XY two-dimensional motion platform 101 realize the automation of laser scanning.

[0036] Furthermore, the Vibroflex laser vibrometer 104 completes the vibration data acquisition of the aerospace composite material part 106 under test. The Vibroflex laser vibrometer 104 is fixed on the XZ two-dimensional motion platform 103. As the laser scanning continues, the industrial control computer completes the overall path planning. According to the laser scanning process in the above steps, the industrial control computer can track the position of the laser scanning point obtained by the deflection of the two-dimensional deflection mirror 105 in real time. The XZ two-dimensional motion platform is also controlled by the industrial control computer for path planning. The Vibroflex laser vibrometer 104 mounted on the 103 performs continuous displacement in the XZ direction, so that the probe position of the Vibroflex laser vibrometer 104 always follows the position of the laser scanning point, thereby maintaining perpendicularity to the laser scanning point on the part 106. This allows for better reception of the vibration of the part 106, resulting in a high signal-to-noise ratio, and enabling higher accuracy in the subsequent three-dimensional reconstruction of the part.

[0037] Furthermore, the industrial control computer 107 is the core of the entire system's control and data processing, and its software includes the aforementioned control module and signal processing module. The control module communicates via serial port with the XY two-dimensional motion platform 101, the Q-switched Nd:YAG laser 102, the XZ two-dimensional motion platform 103, the Vibroflex laser vibrometer 104, and the two-dimensional deflection mirror 105 to achieve continuous and efficient laser scanning and to acquire data on part vibration. The signal processing module processes the images obtained from the vibrometer. First, it reads the signal data acquired by the laser and vibrometer using C++ and Python, calls a noise reduction function for noise reduction, and then uses a time-frequency domain analysis algorithm to plot the two-dimensional waveform of the denoised signal data in the time-frequency domain, resulting in a visualized waveform. Two-dimensional images are generated. Secondly, by analyzing the laser-ultrasonic echo acquisition time of the vibration meter signal, the laser pulse energy is gradually increased. Since the transmission speed of laser-ultrasonic waves in the part is a fixed value, the thickness of the part at the scanning point can be measured. The thickness value of each point of the part is combined with the laser scanning path. By calling the 3D reconstruction function in the OpenCV library, a 3D drawing result of the part's outline can be obtained. Finally, by matching the obtained laser-ultrasonic shape image with a self-drawn 3D defect sample image through labels, and importing them into the LANDT neural network for training, the final 3D reconstructed image is obtained. The specific implementation process of the LANDT neural network will be shown in the following embodiments. Compared with existing modeling methods, this method can accurately locate the defect depth and draw the layered image inside the scanning point, ultimately completing the 3D reconstruction of the part under test.

[0038] It should be noted that, in this embodiment, in order to achieve defect detection and automated 3D reconstruction of aerospace composite materials, the control module controls the laser to generate continuous pulses, and the 2D galvanometer deflects synchronously with the laser pulses, thereby enabling the laser to scan the aerospace composite material part under test line by line at the current position. The scanning path is shown in [reference needed]. Figure 3 As shown.

[0039] Furthermore, when the laser and the two-dimensional galvanometer move, the XZ two-dimensional motion platform 103 equipped with the laser vibrometer 104 also needs to move synchronously to ensure that the laser vibrometer 104 remains perpendicular to the laser scanning point, thereby extracting the data with the highest signal-to-noise ratio. For the specific positional relationship, please refer to [link to relevant documentation]. Figure 1 After the two-dimensional galvanometer is deflected at the current position, allowing the laser to complete the scan at the current position, the industrial control computer 107 controls the XY two-dimensional motion platform 101 to move to the next position, thereby performing laser scanning and data acquisition at the new position of the part to be measured, and repeating the above steps.

[0040] Furthermore, for the signal processing module, an innovative LANDT network is proposed. This network, based on a generative adversarial network (GAN), can import waveform images after training to obtain intuitive and visual 3D reconstructed images of parts. The backbone network of LANDT is a GAN, and training this network requires the creation of a dataset. In this embodiment, waveform images measured by a vibration meter, manually marked labels at defects on the waveform images, and manually constructed 3D images are used as inputs to the network. The network is trained using a GAN, thus allowing the waveform images to serve as inputs and the final 3D reconstructed image to serve as the network output.

[0041] Furthermore, when the part is large, the laser and 2D galvanometer cannot guarantee that the laser scan of the entire part can be completed from the initial position. Therefore, the part needs to be divided into a grid. By moving the XY axis of the 2D platform, the laser and 2D galvanometer can complete the scan of the entire part. The idea of ​​gridding is to divide the large part into small grids of the same size, so that each small grid can be inspected separately during defect detection. Finally, all grid parts can be stitched together to achieve the inspection of large parts.

[0042] When performing 3D reconstruction of data, each scanned part needs to be reconstructed step by step, and finally the 3D images of each part are stitched together to form a whole, thereby obtaining the 3D reconstruction image of the whole part.

[0043] For ease of understanding, this embodiment takes the testing of a 5mm thick aerospace composite laminate sample as an example. It requires the detection of surface microstructure and internal defects, and the completion of three-dimensional reconstruction of the part so that the defects can be visualized intuitively.

[0044] First, as Figure 4 The aerospace composite material shown is fixed onto the testing platform.

[0045] Next, the initial parameters of the laser and the two-dimensional galvanometer are set. After determining the initial scanning position, the initial position and parameters of the laser vibrometer are adjusted so that the laser scanner can move synchronously with the laser scanning position. After the parameter settings are completed, the laser begins continuous pulse excitation, the two-dimensional galvanometer deflects continuously, and the laser vibrometer follows the two-dimensional motion platform to perform displacement motion, so that the point measured by the vibrometer is always vertically focused on the laser scanning point, and the ultrasonic signal is acquired.

[0046] Finally, after preprocessing the acquired ultrasonic signals, waveforms in the time and frequency domains are obtained and imported into the LANDT network for three-dimensional reconstruction, ultimately resulting in a visualized three-dimensional reconstructed image of the part's interior.

[0047] The above is an embodiment of a laser ultrasonic three-dimensional reconstruction and inspection system for aerospace composite materials provided in this application. The following is an embodiment of a laser ultrasonic three-dimensional reconstruction and inspection method for aerospace composite materials provided in this application.

[0048] For easier understanding, please refer to Figure 2 This embodiment provides a laser-ultrasonic three-dimensional reconstruction defect detection method for aerospace composite materials, which ultimately obtains a three-dimensional reconstructed image of internal defects in the part, allowing the observer to directly observe the layered defects inside the material. Based on the above embodiment, the laser-ultrasonic deep learning-based three-dimensional reconstruction defect detection system and method includes the following steps:

[0049] S101: Initializes hardware parameters via industrial control computer, including parameters for the pulsed laser, 2D galvanometer, laser vibrometer, and 2D motion platform. Sets the scanning parameters for the laser and 2D galvanometer, and matches the displacement parameters of the 2D motion platform equipped with the laser vibrometer to them.

[0050] S102: The laser emits continuous pulsed laser light, which directly hits the two-dimensional galvanometer. The laser light is continuously deflected by the synchronous scanning of the two-dimensional galvanometer, completing the continuous scanning of the current position on the aerospace composite material, thereby generating a laser ultrasonic signal on the surface of the composite material to be tested.

[0051] S103: By controlling the XZ-axis two-dimensional motion platform, the laser vibrometer moves synchronously with the laser scanning point on the part under test, always maintaining a vertical position to collect laser ultrasonic signals, thereby obtaining continuous laser scanning signals, and saving these signal data to the industrial control computer data processing module.

[0052] S104: The laser ultrasonic signal collected by the vibration meter is preprocessed, the ultrasonic signal is analyzed and the waveform diagrams in the time domain and frequency domain are obtained, and the thickness of the material at the current scanning point is obtained according to the pulse time interval, so that the three-dimensional contour map of the material at the scanning position can be obtained.

[0053] S105: After obtaining the laser ultrasonic waveform, it is imported into the innovatively proposed LANDT neural network for training. This neural network is based on a generative adversarial network and consists of a generator G and a discriminator D. The generator G consists of four fully convolutional layers, whose function is to generate simulated dataset data after training on the input dataset. The discriminator consists of four inverse convolutional layers, whose function is to determine whether the data is real data from the training dataset or data generated by the generator G. The goal is to identify as much "fake data" as possible generated by the generator. When the data generated by the generator G can no longer be recognized by the discriminator D, it proves that the network training is complete and can output the desired result. After training, the waveform is imported to obtain a visually intuitive 3D reconstructed image of the part. In this embodiment, the waveform measured by the vibration meter, the manually marked labels at the defects in the waveform, and the manually constructed 3D image are used as inputs to the network, which is then trained using a generative adversarial network (GAN). The generator G continuously generates 3D reconstructed images, while the discriminator D continuously trains to discriminate them. Training stops when the generator's true discrimination rate is sufficiently high, and the final 3D reconstructed image is used as the network's output. The initial training parameters, learning rate and batch size, are set. After 300 training iterations, the network can automatically analyze the surface and internal defects of the part under test and perform layer-by-layer 3D reconstruction using location and depth information, generating the 3D contour image from the previous steps.

[0054] S106: The XY two-dimensional motion platform moves to the next position and repeats the above steps S102-S105. After the parts in each laser scanning area are reconstructed in three dimensions, three-dimensional images of each scanning point position can be obtained. Using the image stitching function in the open-source image processing library OpenCV, the images of the three-dimensional reconstructed parts in each scanning area are stitched together according to the displacement of the XY two-dimensional motion platform, so as to obtain the final overall three-dimensional reconstruction image of the parts.

[0055] S107: Outputs the overall 3D reconstructed part image to the display, thereby completing the automated detection of surface and internal defects of the part and allowing for intuitive observation of the location and size of defects.

[0056] Furthermore, in step S105, this embodiment innovatively proposes a LANDT neural network, which is based on a generative adversarial network. Its model parameters and dataset are obtained through the following steps:

[0057] Step 1: Perform laser ultrasonic testing on as many aerospace composite materials containing known defects as possible. Manually annotate the obtained time-domain and frequency-domain waveforms, indicate the defect locations and record their image positions, and generate the txt file required by the LANDT neural network from the label information.

[0058] Step 2: Manually generate 3D images of the inspected parts, mark the defect locations, and correlate them with the waveform and label information in Step 1 to obtain the training set for the Generative Adversarial Network (LANDT).

[0059] Step 3: Set the training parameters: learning rate, batch size, and number of training epochs. Input the waveform, 3D image, and labels generated in the previous two steps into the LANDT network for training. Observe the loss curve during training to determine whether the LANDT network has converged. If it has converged, stop training; if it has not converged, continue training. Preferably, the learning rate is set to 0.01, the batch size is set to 16, and the number of training epochs is set to 300 to obtain the trained model.

[0060] Step Four: In actual testing, the vibration meter data after laser scanning is preprocessed to generate a waveform diagram, which is then imported into the training result model from Step Three for forward input. This yields the final 3D reconstruction result, completing the detection and visualization of part defects. The technical solution of this invention has been described in detail above with reference to the accompanying drawings. This patent proposes a deep learning-based 3D reconstruction defect detection system and method based on laser ultrasound, achieving automated product inspection and 3D visualization of internal defects in parts. This facilitates the observation and analysis of internal defects, significantly reducing costs compared to industrial CT, improving the detection effect on thicker parts, effectively increasing detection accuracy, and simultaneously reducing production costs and improving product quality.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A laser-ultrasound-based deep learning-based 3D reconstruction defect detection method, comprising a laser-ultrasound-based deep learning-based 3D reconstruction defect detection system, the system comprising: XY two-dimensional motion platform (101), Q-switched Nd:YAG laser (102), XZ two-dimensional motion platform (103), Vibroflex laser vibrometer (104), two-dimensional deflection mirror (105), aerospace composite material part under test (106), industrial computer (107), display (108); The XY two-dimensional motion platform (101) provides displacement in the XY direction. The Q-switched Nd:YAG laser (102) is located on the XY two-dimensional motion platform (101) and is used to generate continuous laser pulses. The two-dimensional deflection mirror (105) realizes the continuous deflection scanning of the Q-switched Nd:YAG laser (102). The laser pulses emitted by the Q-switched Nd:YAG laser (102) hit the aerospace composite material part (106) to be tested. The XZ two-dimensional motion platform (103) provides two-dimensional motion in the XZ direction. The Vibroflex laser vibrometer (104) is located on the XZ two-dimensional motion platform (103) and is used to detect the vibration of the part. The industrial control computer (107) includes a control module and a signal processing module; The detection method is characterized by the following steps: Based on the laser ultrasonic signal collected by the detection system on the surface of the aerospace composite material part (106) to be tested, a three-dimensional contour map of the current laser scanning point is obtained according to the laser ultrasonic signal. The 3D contour map is input into the LANDT neural network to train the LANDT neural network, thereby achieving 3D reconstruction of the current laser scanning point; the LANDT neural network is implemented through the following steps: Step 1: Perform laser ultrasonic testing on the aerospace composite material with known defects, manually label the obtained time-domain and frequency-domain waveforms, indicate the defect location and record its image position, and generate the txt file required by the LANDT neural network from the label information. Step 2: Manually generate 3D images of the aerospace composite material with known defects, mark the defect locations, and match them with the waveform and label information in Step 1 to obtain the training set of the generative adversarial network LANDT neural network. Step 3: Set the training parameters, including learning rate, batch size, and number of training epochs. Input the waveform, 3D image, and labels generated in the previous two steps into the LANDT network for training. Observe the loss curve during the training process to determine whether the LANDT network has converged. If it has converged, stop training; if it has not converged, continue training. After three-dimensional reconstruction of each laser scanning point of the aerospace composite material part (106) to be tested, the three-dimensional reconstruction of the parts in each scanning area is stitched together according to the displacement of the XY two-dimensional motion platform (101) to obtain the final overall three-dimensional reconstruction image of the part. The three-dimensional reconstruction image of the aerospace composite material part (106) to be tested is output to the display (108), thereby completing the automated detection of surface and internal defects of the part and intuitively observing the location and size of the defects of the part.

2. The method for defect detection based on laser ultrasound and deep learning-based three-dimensional reconstruction according to claim 1, characterized in that: The industrial computer (107) control module communicates with the XY two-dimensional motion platform (101), Q-switched Nd:YAG laser (102), XZ two-dimensional motion platform (103), Vibroflex laser vibrometer (104) and two-dimensional deflection mirror (105) to complete the data acquisition of the vibration of the aerospace composite material part (106) to be tested; The signal processing module of the industrial control computer (107) performs signal processing on the image obtained by the Vibroflex laser vibrometer (104) to realize signal waveform drawing, part scanning point thickness measurement, part outline drawing, part scanning point defect depth positioning, part scanning point internal image layer drawing, and finally complete the three-dimensional reconstruction of the part to be tested.

3. The deep learning-based three-dimensional reconstruction defect detection method based on laser ultrasound according to claim 1, characterized in that, The laser ultrasonic signal acquired by the detection system on the surface of the aerospace composite material part (106) to be tested, and the three-dimensional contour map of the current laser scanning point obtained based on the laser ultrasonic signal, are as follows: S101: Initialize hardware parameters via industrial computer (107); S102: The Q-switch Nd:YAG laser (102) emits continuous pulse laser light, which directly hits the two-dimensional deflection mirror (105). The laser light is continuously deflected by the synchronous scanning of the two-dimensional deflection mirror (105), and the current position is continuously scanned on the aerospace composite material part (106) to be tested, thereby generating a laser ultrasonic signal on the surface of the aerospace composite material part (106) to be tested. S103: By controlling the XZ two-dimensional motion platform (103), the Vibroflex laser vibration meter (104) moves synchronously with the laser scanning point on the part to be tested, always maintaining a vertical position to collect laser ultrasonic signals, thereby obtaining continuous laser scanning signals, and saving the continuous laser scanning signals to the data processing module of the industrial control computer (107); S104: The laser ultrasonic signal collected by the Vibroflex laser vibrometer (104) is preprocessed, the laser ultrasonic signal is analyzed and the waveform diagrams in the time domain and frequency domain are obtained, and the thickness of the current laser scanning point of the aerospace composite material part (106) under test is obtained according to the pulse time interval, thereby obtaining the three-dimensional contour diagram of the laser scanning point of the aerospace composite material part (106) under test. S105: After obtaining the laser ultrasonic shape image, it is imported into the innovatively proposed LANDT neural network for training and learning. The hyperparameters of the network training are set, and the defects on the surface and inside of the part to be tested are automatically analyzed. The part is reconstructed layer by layer using position and depth information to generate the three-dimensional contour map in the above steps. S106: The XY two-dimensional motion platform (101) moves to the next position and repeats the above steps S102-S105. After the parts in each laser scanning area are reconstructed in three dimensions, the images of the reconstructed parts in each scanning area are stitched together according to the displacement of the XY two-dimensional motion platform (101) to obtain the final overall three-dimensional reconstruction image of the parts.

4. The deep learning-based three-dimensional reconstruction defect detection method based on laser ultrasound according to claim 3, characterized in that, Specifically, S101 is: Q-switch Nd: Initializes the parameters of the YAG laser (102), two-dimensional deflection mirror (105), Vibroflex laser vibrometer (104), and two two-dimensional motion platforms (101, 103); sets the scanning parameters of the laser and two-dimensional deflection mirror (105), and matches the displacement parameters of the XZ two-dimensional motion platform (103) equipped with Vibroflex laser vibrometer (104) to it.

5. The deep learning-based three-dimensional reconstruction defect detection method based on laser ultrasound according to claim 1, characterized in that, In step three, the learning rate is set to 0.01, the batch size to 16, and the training epochs to 300. After training, a trained LANDT neural network model is obtained.

6. The method for defect detection based on laser ultrasound and deep learning-based three-dimensional reconstruction according to claim 1, characterized in that: When the part is large, the part is divided into grids, each grid is numbered, and then each grid is scanned and the image is reconstructed separately. After completion, the grids are stitched together according to the number to obtain the three-dimensional reconstruction image of the whole part.

Images