Intelligent detection and modeling method for composite material damage based on ultrasonic point cloud and deep learning

Abstract

The present application relates to the technical field of composite material damage detection, and specifically discloses a composite material damage intelligent detection and modeling method based on ultrasonic point cloud and deep learning, comprising: S1, data acquisition, S2, ultrasonic data processing, S3, simulation model construction, S4, PVT-RCNN model processing, S5, RICP processing and S6, damage model generation; the present application designs an ultrasonic point cloud data acquisition platform and a data processing method, constructs an adequate composite material laminate ultrasonic point cloud data set with obvious characteristics, designs a PVT-RCNN model improved by Swin Transformer based on PV-RCNN, improves the feature extraction capability of the model on point cloud, intelligently acquires three-dimensional spatial information of damage, designs an RICP registration algorithm and a damage model generation method, directly generates a finite element model with damage according to the damage point cloud, and material performance can be obtained through calculation.

Description

Technical Field

[0001] This invention belongs to the field of composite material damage detection technology, specifically involving a method for intelligent detection and modeling of composite material damage based on ultrasonic point cloud and deep learning. Background Technology

[0002] Advanced fiber-reinforced composite materials (FRPs) are increasingly used in high-performance structural applications, such as main structural components of aircraft and wind turbine blades, due to their high specific stiffness and strength, corrosion resistance, and fatigue resistance. During use, FRPs are subjected to alternating loads and impacts from foreign objects, which can easily cause damage such as debonding and separation, affecting the material's performance and lifespan. Therefore, regular inspection and monitoring of FRPs are necessary to identify, assess, and address damage before it develops to the point where the structural strength falls below acceptable limits.

[0003] Several common non-destructive testing (NDT) methods exist for inspecting composite materials, including visual inspection, ultrasonic testing, infrared thermography, terahertz scanning, and X-ray computed tomography (CT). Ultrasonic testing is the most widely used NDT technique for composite materials due to its advantages such as high resolution, real-time dynamic imaging, high safety, and convenience. Common ultrasonic testing results include A-scan, B-scan, and C-scan, which can visually distinguish damaged areas from normal areas, but cannot directly obtain three-dimensional spatial information of the damage. Therefore, much research has been conducted on ultrasonic three-dimensional imaging in recent years. Benjamin et al. used a high-resolution ultrasonic C-scan system to detect low-velocity impact damage in composite materials and presented the internal 3D contour of the damage using an unbiased layer-by-layer technique. McKee et al. used a two-dimensional phased array probe to experimentally image artificial damage within a hyperbolic specimen under immersion conditions, successfully obtaining a three-dimensional image of the specimen's interior with a root mean square error of only 0.04 mm. Ohara et al. used a Phased-Array system based on a piezoelectric and laser ultrasonic system (PLUS) to create high-resolution 3D images of internal damage. Wang et al. constructed a phased array ultrasonic scanning system to detect low-velocity impact damage in thin composite plates, achieving three-dimensional dimensional quantification of the damage. Jiang et al. used in-situ computed tomography and digital volume correlation techniques to detect composite materials and construct three-dimensional models. Three-dimensional imaging methods mainly include point clouds and voxels; ultrasonic three-dimensional imaging based on the principle of ultrasonic wave reflection is a point cloud imaging method. Ultrasonic point clouds contain not only material surface information but also internal material information, allowing direct acquisition of the three-dimensional spatial information of the damage.

[0004] The performance evaluation methods for damaged composite materials include: obtaining empirical parameters through experiments, which is a commonly used method in industry; or conducting numerical simulations using finite element modeling. Finite element analysis of composite material damage has been extensively studied. Zhou et al. used a progressive damage model to study the mechanical response and damage development of laminates under low-velocity impact. Higuchi et al. designed a microscale simulation scheme to evaluate the in-situ damage and strength characteristics of laminates, effectively predicting transverse crack propagation. The 3D finite element-progressive damage method (FE-PDM) proposed by Liu et al. can accurately predict the complex mechanical behavior and damage evolution of composite materials. However, constructing composite material damage models is complex. Therefore, it is of great significance to rapidly generate composite material damage models using ultrasonic testing results and conduct finite element analysis experiments.

[0005] In response, the inventors proposed an intelligent detection and modeling method for composite material damage based on ultrasonic point clouds and deep learning to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent detection and modeling method for composite material damage based on ultrasonic point cloud and deep learning, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for intelligent detection and modeling of composite material damage based on ultrasonic point cloud and deep learning includes the following steps:

[0009] S1. Data acquisition: Set up an experimental platform, use a robotic arm to control the probe to scan the material along a specified path, and acquire ultrasonic data, including A-scan signal, B-scan image, C-scan image and ultrasonic point cloud.

[0010] S2. Ultrasonic data processing: The ultrasonic signal and the coordinates of the robotic arm end effector are fused to obtain the original ultrasonic point cloud data. Attenuation compensation, body mass subsampling, half-wave height method and data enhancement processing are performed to obtain the required ultrasonic data features.

[0011] S3. Construct a simulation model. Based on the characteristics of the composite material specimen, construct a non-destructive material model. After the node point cloud is meshed according to the non-destructive material model, extract the coordinates corresponding to each node in the inp file to form the final point cloud.

[0012] S4, PVT-RCNN model processing, including using the SwinTransformer feature extraction network to extract voxel features, the region proposal network to generate preliminary 3D bounding boxes, obtaining keypoints mapped to voxel feature maps of different scales in the backbone network, concatenating these keypoint features and feeding them into the keypoint prediction module to obtain fused keypoint features, and feeding the keypoint features and the preliminary predicted 3D bounding boxes together into the RoI-grid Pooling Module to obtain the final predicted bounding boxes and confidence scores;

[0013] S5 and RICP processing are used to carry out point cloud registration. The damage corresponding to the damage point cloud identified by the PVT-RCNN model is generated into the finite element model and point cloud registration is performed.

[0014] S6. Generate a damage model. After completing the coordinate transformation and damage extraction of the ultrasonic point cloud, the information of the damage point cloud is reflected in the model.

[0015] Preferably, the attenuation expression for the plane wave in the attenuation compensation in step S2 is:

[0016] P x =P0e -αx

[0017] Where P0 is the initial sound pressure of the wave source, P x Let α be the sound pressure at a distance x from the wave source, e be the natural logarithm, α be the medium attenuation coefficient in dB / mm, and x be the distance from the wave source.

[0018] Based on the ultrasonic attenuation formula, the compensation formula is as follows:

[0019] I n =I0e -αx

[0020] In the formula, I0 is the initial signal strength; I n denoted as the compensated signal strength; x represents the distance from the acquired signal to the upper surface of the material.

[0021] Preferably, in step S2, the centroid downsampling of the voxel data divides the point cloud data space into a series of small three-dimensional grids, which are called voxels. The centroid of all points within each voxel is calculated, and all points within that voxel are replaced by the centroid. In this way, the original point cloud data is replaced by a set of evenly distributed and relatively few points, thus achieving downsampling. The centroid calculation expression is:

[0022]

[0023] By comparing the two types of data, it can be seen that physical fitness downsampling effectively reduces the amount of data while preserving the characteristics of the data.

[0024] Preferably, the half-wave height method, also known as the 6dB-drop method, processes the damage point cloud. After the network identifies the damage point cloud, it finds the largest signal intensity and sets half of that signal intensity as a threshold. Points with less than this threshold are removed, thereby achieving further processing of the damage point cloud and obtaining more accurate damage information.

[0025] Preferably, in step S2, data augmentation processing is performed on 40 composite materials with different pre-embedded damages. In order to obtain more experimental data, 160 sets of ultrasonic point cloud data are obtained by selecting different ultrasonic gains and front and back acquisition methods. Random flipping, random rotation, random scaling and noise addition are used to finally obtain 800 sets of data, of which 600 sets are for training and 200 sets are for testing. "CloudCompare" is used to label the ultrasonic point cloud of the damaged area.

[0026] Preferably, the Voxel feature extraction using the Swin Transformer feature extraction network in step S4 is as follows: the ultrasonic point cloud is voxelized, then input into Patch Partition processing to obtain multiple patches of the same size, then input into Linear Enbedding and two Swin Transformer Blocks to obtain the first layer feature map, and then Patch Merging and Swin Transformer Blocks are performed multiple times to obtain multi-scale feature maps with resolutions of 1 / 2, 1 / 4 and 1 / 8 of the original input 2D voxel feature map.

[0027] Preferably, the Swing Transformer block consists of a shift-window based MSA module, followed by two MLP layers, with a GELU nonlinearity in between. A LayerNorm (LN) layer is applied before each MSA module and each MLP, and a residual connection is applied after each module.

[0028] Layer l employs a conventional window partitioning scheme, which is used to divide the voxel feature map to generate local windows. Focusing self-attention within the local window can effectively improve computational efficiency and avoid interference from irrelevant distant pixels.

[0029] In Layer l+1, the window partitions are shifted to create a new window. This new window is created by moving the window position by the same length along the X and Y axes. The self-attention calculation in the new window crosses the boundaries of the previous window in the previous layer, thus providing a connection between them.

[0030] Preferably, the Swin Transformer Block is constructed by replacing the standard multi-head self-attention MSA module in the Transformer block with a shift-window based module, while keeping other layers unchanged. The calculation process of the Swin Transformer Block twice is expressed as follows:

[0031]

[0032] in, and z l These represent the output characteristics of the SW-MSA module and the MLP module of the l-th block, respectively.

[0033] Preferably, in step S5, the RICP processing based on the Rodrigues rotation matrix prevents registration from getting trapped in local optima and effectively replaces the coarse registration step. The formula for the Rodrigues rotation matrix is:

[0034]

[0035] By multiplying the ultrasonic point cloud by a uniformly decreasing rotation matrix, the point cloud is rotated around the Z-axis by an angle that decreases uniformly from 90° to 0° with each iteration. Corresponding points are searched in the rotated ultrasonic point cloud and the node point cloud. The rotation matrix R and translation matrix t are calculated, and the coordinates of the ultrasonic point cloud are updated based on R and t. The mean square error E(R,t) between the ultrasonic point cloud and the node point cloud is calculated. Finally, the error value is compared with the set iteration threshold. If it is greater than or equal to the iteration threshold, the iteration continues; if it is less than or equal to the iteration threshold, the iteration ends, and the point cloud registration is completed.

[0036] Preferably, the experimental platform in step S1 includes a robotic arm, a water tank, an ultrasonic probe, and a probe holder.

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

[0038] (1) This invention designs an ultrasonic point cloud data acquisition platform and data processing method, constructs a sufficient ultrasonic point cloud dataset of composite laminates with obvious characteristics, designs a PVT-RCNN model based on PV-RCNN and improved by Swin Transformer, enhances the model's feature extraction capability of point cloud, intelligently acquires the three-dimensional spatial information of damage, designs RICP registration algorithm and damage model generation method, directly generates a damaged finite element model based on the damage point cloud, and obtains material properties through calculation.

[0039] (2) In this invention, the robotic arm controls the ultrasonic probe to scan the composite material to obtain the original data. After data processing, the ultrasonic point cloud data with obvious features is obtained. At the same time, a non-damaged composite material model is constructed according to the characteristics of the composite material, and its node coordinates are extracted to obtain the node point cloud. Then, the PVT-RCNN intelligently detects the damage point cloud in the ultrasonic point cloud, and the spatial location and size of the damage are obtained by the 6dB-drop method. The RICP algorithm is used to register the ultrasonic point cloud and the node point cloud to obtain the transformation matrix (R,t). By fusing the two results, a damage point cloud that matches the composite material model can be obtained. Finally, the damage model generation module generates the damage represented by the damage point cloud into the composite material model to obtain a composite material model with damage. Finite element analysis of the model can realize the performance evaluation of the tested composite material. Attached Figure Description

[0040] Figure 1 This is a graph of ultrasound data obtained in this invention;

[0041] Figure 2 This is a flowchart of the damage model generation process based on damage point cloud according to the present invention;

[0042] Figure 3 This is a diagram illustrating the effect of the body mass index sampling method of the present invention.

[0043] Figure 4 This is a schematic diagram of the effect of the through filter of the present invention;

[0044] Figure 5 This is an enhanced version of a set of ultrasonic point cloud data from the present invention.

[0045] Figure 6 This is a schematic diagram of the MSA module based on a shift window according to the present invention;

[0046] Figure 7 This is a diagram showing the strip-shaped damage detection results of the present invention;

[0047] Figure 8 This is a diagram showing the square damage detection results of the present invention;

[0048] Figure 9 This is a diagram showing the results of circular damage detection according to the present invention;

[0049] Figure 10 This is a diagram showing the experimental results of registration between ultrasonic point clouds and nodal point clouds according to the present invention.

[0050] Figure 11 This is a damage model diagram generated based on circular damage point cloud according to the present invention;

[0051] Figure 12 This is a damage model diagram generated based on square damage point cloud according to the present invention;

[0052] Figure 13 This is a schematic diagram of the overall experimental framework of the present invention;

[0053] Figure 14 This is a schematic diagram of the experimental platform of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1:

[0056] Please see Figures 1 to 14 As shown, the intelligent detection and modeling method for composite material damage based on ultrasonic point cloud and deep learning includes the following steps:

[0057] S1. Data acquisition and experimental platform setup, such as... Figure 14 As shown, the experimental platform includes a robotic arm, a water tank, an ultrasonic probe, and a probe holder. The robotic arm controls the probe to scan the material along a designated path, acquiring ultrasonic data including A-scan signals, B-scan images, C-scan images, and ultrasonic point clouds. The robotic arm is a JAKA Zu 7 collaborative robot, the ultrasonic A-scan device is a UFD BOX, and the ultrasonic probe is a 5MHz straight probe. The experimental material is a pre-fabricated internally damaged carbon fiber laminate, measuring 150mm × 100mm × 5mm, with a sound velocity of 2900m / s and a sampling range of 7mm. The dashed line in the figure represents the probe's movement path planning; the robotic arm controls the probe to scan the material along an S-shaped path. The data sampling range is: X = 170mm, Y = 120mm. Figure 1 The ultrasonic data obtained in the experiments of this invention includes A-scan signals, B-scan images, C-scan images, and ultrasonic point clouds, where T represents the signal from the upper surface of the specimen, B represents the signal from the lower surface of the specimen, and D represents the damage signal. As can be seen from the figures, A, B, and C scans can all identify damage, but they cannot obtain the spatial information of the damage. The ultrasonic point cloud contains information such as the spatial location and size of the damage, based on which a finite element model of the composite material including the damage can be established.

[0058] S2. Ultrasonic data processing: The ultrasonic signal and the coordinates of the robotic arm end are fused to obtain the original ultrasonic point cloud data. This data contains 4 sets of data (x, y, z, intensity), where x and y are provided by the robotic arm, z is the depth corresponding to the ultrasonic signal, and intensity is the intensity of the ultrasonic signal. The required ultrasonic data features are obtained by performing attenuation compensation, body mass subsampling, pass-through filtering, and data enhancement processing.

[0059] S3. Construct a simulation model. Based on the characteristics of the composite material specimens, a simulation model was constructed as follows: Figure 13 The non-destructive material model in the overall experimental framework diagram measures 150mm x 100mm x 5mm and consists of 40 layers with zero-thickness cohesive elements between them. The layer layup sequence is: [45 / 0 / -45 / 0 / 90 / 0 / 45 / 0 / -45 / 0 / 45 / 0 / -45 / 0 / 90 / 0 / 45 / 0 / -45 / 0]s. To balance computational efficiency and accuracy, the mesh density is set to 1mm x 1mm for the central 70mm square region, surrounded by 1mm x 2mm and 2mm x 2mm meshes. Figure 13 The point cloud of nodes in the overall experimental framework diagram is formed by extracting the coordinates of each node from the inp file after meshing the non-destructive material model.

[0060] S4, PVT-RCNN model processing, including using the Swin Transformer feature extraction network to extract voxel features, the region proposal network to generate preliminary 3D bounding boxes, obtaining keypoints mapped to voxel feature maps at different scales of the backbone network, concatenating these keypoint features and feeding them into the keypoint prediction module to obtain fused keypoint features, and feeding the keypoint features and the preliminary predicted 3D bounding boxes together into the RoI-grid Pooling Module to obtain the final predicted bounding boxes and confidence scores;

[0061] S5 and RICP processing are used to carry out point cloud registration. The damage corresponding to the damage point cloud identified by the PVT-RCNN model is generated into the finite element model and point cloud registration is performed.

[0062] S6. Generate a damage model. After completing the coordinate transformation and damage extraction of the ultrasound point cloud, the damage point cloud information is reflected in the model. The region of the element is calculated using the coordinates of the element nodes in the inp file. A loop is used to check whether all elements in the intermediate layer contain damage sites from the scan results. If one or more damage sites are contained, the element will be included in a set and written to a new inp file, and then deleted in AbaqusCAE. Because only layered damage is considered, only the cohesive layer elements of the model are checked and deleted.

[0063] As shown above, by designing an ultrasonic point cloud data acquisition platform and data processing method, a sufficient ultrasonic point cloud dataset of composite laminates with obvious characteristics was constructed. A PVT-RCNN model based on PV-RCNN and improved by SwinTransformer was designed to enhance the model's feature extraction capability of point clouds and intelligently acquire the three-dimensional spatial information of damage. An RICP registration algorithm and damage model generation method were designed to directly generate a damaged finite element model based on the damage point cloud. The material properties can be obtained through calculation.

[0064] Example 2:

[0065] Please see Figures 1 to 14 As shown, the attenuation expression for a plane wave in attenuation compensation is:

[0066] P x =P0e -αx

[0067] Where P0 is the initial sound pressure of the wave source, P x Let α be the sound pressure at a distance x from the wave source, e be the natural logarithm, α be the medium attenuation coefficient in dB / mm, and x be the distance from the wave source.

[0068] Based on the ultrasonic attenuation formula, the compensation formula is as follows:

[0069] I n =I0e -αx

[0070] In the formula, I0 is the initial signal strength; I n The signal strength after compensation is represented by x; x is the distance from the acquired signal to the upper surface of the material.

[0071] The compensation formula for this material, derived from the ultrasonic attenuation formula, is: I n =I0e -0.002x

[0072] like Figure 2 The images show the effect of attenuation compensation. The top image is the original ultrasound A-scan signal, and the bottom image is the compensated ultrasound A-scan signal. The results show that attenuation compensation can effectively correct the energy loss of the signal during transmission, ensuring the accuracy and reliability of the signal.

[0073] Cenozoic centroid downsampling divides the point cloud data space into a series of small 3D grids called voxels. The centroid of all points within each voxel is calculated, and these centroids replace all points within that voxel. In this way, the original point cloud data is replaced by a set of evenly distributed, relatively few points, achieving downsampling. The centroid calculation expression is:

[0074]

[0075] like Figure 3 The image shows the effect of voxel downsampling. In this invention, a voxel is a cube with a side length of 0.1 mm. By comparing the two types of data, it can be seen that voxel downsampling effectively reduces the amount of data, and the features of the data can also be preserved.

[0076] The 6dB-drop method uses a pass-through filter with a minimum strength threshold of 1000 to remove interference signals, such as... Figure 4 This is an illustration of the effect of a pass-through filter. The top image shows the point cloud data before filtering, which is square and lacks obvious features. The bottom image shows the point cloud data after filtering, which can directly distinguish the upper surface, lower surface, and damage of the material. The filter then processes the damage point cloud. After the network identifies the damage point cloud, it finds the largest signal intensity and sets half of that signal intensity as a threshold. Points with intensity below this threshold are removed, allowing for further processing of the damage point cloud and obtaining more accurate damage information.

[0077] Data augmentation processing was performed on 40 composite materials with different pre-embedded damages. To obtain more experimental data, 160 sets of ultrasonic point cloud data were obtained by selecting different ultrasonic gains and acquiring data from both the front and back sides. Random flipping, random rotation, random scaling, and noise addition were used to ultimately obtain 800 sets of data, with 600 sets for training and 200 sets for testing. "CloudCompare" was used to annotate the ultrasonic point clouds of the damaged areas, such as... Figure 5 This is an enhanced image of a set of ultrasonic point cloud data.

[0078] The Voxel feature extraction using the Swing Transformer network is as follows: the ultrasonic point cloud is voxelized, then input into Patch Partition to obtain multiple patches of the same size, then input into LinearEnbedding and two Swing Transformer Blocks to obtain the first layer feature map, and then PatchMerging and Swing Transformer Blocks are performed multiple times to obtain multi-scale feature maps with resolutions of 1 / 2, 1 / 4 and 1 / 8 of the original input 2D voxel feature map.

[0079] The Swin Transformer block consists of shift-window based MSA modules, followed by two MLP layers, with a GELU nonlinearity in between. A LayerNorm (LN) layer is applied before each MSA module and each MLP, and a residual connection is applied after each module.

[0080] like Figure 6The schematic diagram of the MSA module based on shift window is shown. Layer l adopts a conventional window partitioning scheme, which is used to divide the voxel feature map to generate local windows. Focusing self-attention within the local window can effectively improve computational efficiency and avoid interference from irrelevant distant pixels.

[0081] In Layer l+1, the window partitions are shifted to create a new window. This new window is created by moving the window position by the same length along the X and Y axes. The self-attention calculation in the new window crosses the boundaries of the previous window in the previous layer, thus providing a connection between them.

[0082] The Swin Transformer Block is constructed by replacing the standard multi-head self-attention MSA module in the Transformer Block with a shift-window based module, while keeping other layers unchanged. The calculation process of the Swin Transformer Block twice is expressed as follows:

[0083]

[0084] in, and z l These represent the output characteristics of the SW-MSA module and the MLP module of the l-th block, respectively.

[0085] In step S5, the RICP processing based on the Rodrigues rotation matrix prevents registration from getting trapped in local optima and effectively replaces the coarse registration step. The formula for the Rodrigues rotation matrix is:

[0086]

[0087] By multiplying the ultrasonic point cloud by a uniformly decreasing rotation matrix, the point cloud is rotated around the Z-axis by an angle that decreases uniformly from 90° to 0° with each iteration. Corresponding points are searched in the rotated ultrasonic point cloud and the node point cloud. The rotation matrix R and translation matrix t are calculated, and the coordinates of the ultrasonic point cloud are updated based on R and t. The mean square error E(R,t) between the ultrasonic point cloud and the node point cloud is calculated. Finally, the error value is compared with the set iteration threshold. If it is greater than or equal to the iteration threshold, the iteration continues; if it is less than or equal to the iteration threshold, the iteration ends, and the point cloud registration is completed.

[0088] As shown above, the robotic arm controls the ultrasonic probe to scan the composite material and acquire raw data. After data processing, characteristic ultrasonic point cloud data is obtained. Simultaneously, a damage-free composite material model is constructed based on the composite material characteristics, and its node coordinates are extracted to obtain the node point cloud. Then, PVT-RCNN intelligently detects the damage point cloud in the ultrasonic point cloud, and the spatial location and size of the damage are obtained using the 6dB-drop method. The RICP algorithm is used to register the ultrasonic point cloud and the node point cloud, obtaining the transformation matrix (R,t). Fusing the two results yields a damage point cloud that matches the composite material model. Finally, the damage model generation module generates the damage represented by the damage point cloud into the composite material model, obtaining a damaged composite material model. Finite element analysis of this model enables the performance evaluation of the tested composite material.

[0089] Example 3:

[0090] Please see Figures 1 to 14 As shown, the hardware platform used in this invention is an Intel(R) Core(R) 10700k CPU, one NVIDIA RTX 3090 graphics card, Windows 10 operating system, Python development language, using PyTorch version 1.7, CUDA version 11.1, and ABAQUS 2019 simulation software.

[0091] 1) Results of ultrasonic point cloud intelligent detection experiment:

[0092] The training process for the ultrasonic point cloud intelligent detection experiment used the Adam optimizer for parameter updates, with an initial learning rate of 0.01, a decay factor of 0.8, a training epoch of 80, and a batch size of 2. Since the detection object is only damage point clouds, this invention uses 3D AP (representing the accuracy of 3D detection boxes) as the evaluation metric, with an IoU threshold of 0.7.

[0093] Figure 7-9 This is a diagram showing the results of ultrasonic point cloud damage detection. Figure 7 The results of strip-shaped damage detection. Figure 8 The results of the square damage detection are as follows. Figure 9 This is the result of circular damage detection. The first row shows the point cloud detection results, and the second row shows the damage point cloud obtained after processing with the 6dB drop method. In the figure, x, y, z represent the center coordinates of the detection box, and l, w, h represent the length, width, and height of the detection box, in mm.

[0094] Depend on Figure 7The PVT-RCNN model accurately detected the strip-shaped damage point cloud in the ultrasound point cloud. The center point coordinates of the detection box were (64.50, 81.51, 2.58), and the size was (9.00, 55.22, 1.21). After processing with the 6dB drop method, the center point coordinates of the damage point cloud became (64.50, 81.63, 2.44), and the size became (7.00, 53.81, 0.85). The length decreased by 2, the width decreased by 1.41, and the height decreased by 0.36.

[0095] Figure 8 The center point coordinates of the square damage are (67.50, 82.21, 2.50), and the dimensions are (33.00, 66.62, 0.82).

[0096] Figure 9 The center point coordinates of the circular damage are (67.50, 84.67, 2.53), and the dimensions are (61.00, 63.99, 0.76).

[0097] In summary, the method of this invention can accurately identify damage point clouds in ultrasonic point clouds and obtain three-dimensional spatial information of the damage point clouds. The 6dB drop method can effectively remove interference signals and obtain more accurate damage point cloud information.

[0098] 2) Experimental results of RICP registration algorithm:

[0099] This experiment registered ultrasonic point clouds with nodal point clouds, and used the algorithm's convergence time, root mean square error of point cloud distance (RMSE), and registration effect as evaluation indicators.

[0100] Table 2 below shows the results of different registration algorithms:

[0101]

[0102] Table 2

[0103] Figure 10 The image shows the registration results of the ultrasonic point cloud and the nodal point cloud.

[0104] The ultrasonic point cloud in Point Cloud 1 contains a circular lesion point cloud. ICP took 5.8 s with an RMSE of 0.0880 mm. The image shows that after registration, the two point clouds are 90° apart in direction, clearly misaligned. RICP took 7.0 s with an RMSE of 0.0103 mm, 1.2 s longer than ICP, but with a RMSE 0.0777 mm less. The image shows that the two point clouds are basically aligned, and the effect is significantly better than ICP.

[0105] The ultrasonic point cloud in Point Cloud 2 contains a square lesion point cloud. ICP took 5.5s with an RMSE of 0.0815mm. The image shows that after registration, the two point clouds are 90° apart in direction, clearly misaligned. RICP took 6.4s with an RMSE of 0.0136mm, 0.9s longer than ICP but with a RMSE 0.0679mm lower. The image shows that the two point clouds are basically aligned, with significantly better results than ICP. Therefore, the RICP algorithm can accurately register ultrasonic point clouds and nodal point clouds, and has application value.

[0106] 3) Results of damage model generation and evaluation experiments:

[0107] Figure 11 This is a damage model generated based on a circular damage point cloud. Figure 12 This figure shows the damage model generated based on the square damage point cloud. It reflects the Cohesive elements in the 20th and 21st plies of the model. As can be seen from the figure, the detected damage point cloud can accurately generate a finite element model with damage.

[0108] As can be seen from the above, the method in this invention can accurately identify the damage point cloud in the ultrasonic point cloud and obtain the three-dimensional spatial information of the damage point cloud. The 6dB drop method can effectively remove interference signals and obtain more accurate damage point cloud information. The RICP algorithm can accurately register the ultrasonic point cloud and the nodal point cloud. The detected damage point cloud can accurately generate a finite element model with damage. Therefore, the above method has wide application value.

[0109] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0110] The accompanying drawings of the embodiments disclosed in this invention only involve structures relevant to the embodiments disclosed in this invention. Other structures can be referred to with common designs. Unless otherwise specified, the same embodiment and different embodiments of this invention can be combined with each other.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A composite damage intelligent detection and modeling method based on ultrasonic point cloud and deep learning, characterized in that, Comprise the following steps: S1, data acquisition, build an experimental platform, use a mechanical arm to control the probe to scan the material according to the specified path, and obtain the ultrasonic data, including A-scan signal, B-scan image, C-scan image and ultrasonic point cloud; S2, ultrasonic data processing, fusion of ultrasonic signal and mechanical arm end coordinates to obtain the original ultrasonic point cloud data, attenuation compensation, voxel centroid downsampling, half-wave height method and data enhancement processing, and obtain the required ultrasonic data features; S3, construct a simulation model, construct a non-damage material model according to the characteristics of the composite material specimen, grid the node point cloud according to the non-damage material model, extract the coordinates corresponding to each node in the inp file, and finally form the point cloud; S4, PVT-RCNN model processing, including using Swin Transformer feature extraction network for voxel feature extraction, region proposal network for generating preliminary 3D bounding box, obtaining key point mapping to voxel feature map of different scales of backbone network, obtaining fused key point features by splicing these key point features and sending them into key point prediction module, and sending key point features and preliminary predicted 3D bounding box into RoI-grid Pooling Module to obtain final predicted bounding box and confidence; S5, RICP processing, carrying out point cloud registration work, generating damage corresponding to the damage point cloud identified by the PVT-RCNN model to the finite element model, and carrying out point cloud registration work; S6, generate a damage model, after completing the coordinate transformation of the ultrasonic point cloud and damage extraction, reflect the information of the damage point cloud on the model.

2. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method according to claim 1, characterized in that: The attenuation expression of the plane wave in the attenuation compensation in step S2 is: ； wherein is the initial sound pressure of the wave source, is the sound pressure at a distance x1 from the wave source, e is the natural logarithm, is the medium attenuation coefficient, in dB / mm, and x1 is the distance from the wave source; According to the ultrasonic wave attenuation formula, the compensation formula is: ； In the formula is the initial signal intensity; is the compensated signal intensity; x2 is the distance of the collected signal to the upper surface of the material.

3. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method according to claim 1, characterized in that: In step S2, the voxel centroid downsampling divides the point cloud data space into a series of small three-dimensional grids, which are called voxels. The centroid of all points inside each voxel is calculated, and all points in the voxel are replaced by the centroid. In this way, the original point cloud data is replaced by a set of uniformly distributed and fewer points, achieving downsampling. By comparing the two types of data, it can be seen that the voxel centroid downsampling effectively reduces the data volume, and the characteristics of the data can be preserved.

4. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method of claim 1, wherein: The half-wave height method, also known as the 6db-drop method, processes the damage point cloud. After the network identifies the damage point cloud, the largest signal strength is found, and the threshold is set at half of the signal strength. The points smaller than the threshold are removed to further process the damage point cloud and obtain more accurate damage information.

5. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method of claim 1, wherein: In step S2, data augmentation processing is performed on 40 different pre-embedded damage composite materials. In order to obtain more experimental data, 160 groups of ultrasonic point cloud data are obtained by selecting different ultrasonic gains and front and back collection methods. Random flipping, random rotation, random scaling and adding noise are used to finally obtain 800 groups of data, of which the training set and test set are 600 groups and 200 groups respectively. The ultrasonic point cloud in the damage area is labeled using "CloudCompare".

6. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method of claim 1, wherein: The voxel feature extraction in step S4 using the Swin Transformer feature extraction network is specifically: the ultrasound point cloud is voxelized, then input into the Patch Partition processing to obtain a plurality of patches of the same size, and then input into the Linear Enbedding and twice Swin Transformer Block to obtain a first layer feature map, and then perform a plurality of Patch Merging and Swin Transformer Block to obtain multi-scale feature maps with resolutions of 1 / 2, 1 / 4 and 1 / 8 of the original input 2D voxel feature map.

7. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method according to claim 6, characterized in that: The Swin Transformer block is composed of a shift window-based MSA module, followed by 2-layer MLP, with GELU nonlinearity in between, and a LayerNorm (LN) layer is applied before each MSA module and each MLP, and a residual connection is applied after each module; Layer l adopts a conventional window partition scheme, which is used to divide the voxel feature map to generate a local window, and focusing self-attention in the local window can effectively improve the calculation efficiency and avoid the interference caused by irrelevant distant pixels; In Layer l+1, the window partition is shifted, thereby generating a new window, which moves the window position in the X-axis and Y-axis by the same length, and the self-attention calculation in the new window crosses the boundary of the previous window in the previous layer, thereby providing a connection between them.

8. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method according to claim 7, characterized in that: The Swin Transformer Block is constructed by replacing the standard multi-head self-attention MSA module in the Transformer block with a shift window-based module, and the other layers remain unchanged. The calculation process of the two Swin Transformer Blocks is as follows: ； wherein, and denote the SW-MSA module output feature and the MLP module output feature of the l-th block, respectively.

9. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method according to claim 1, characterized in that: The RICP processing in step S5 is based on the RICP algorithm of the Rodrigues rotation matrix to prevent the registration from falling into a local optimum, and effectively replaces the coarse registration step. The formula of the Rodrigues rotation matrix is: ； Rz represents the rotation matrix around the Z-axis, which is multiplied by the ultrasound point cloud and a uniformly decreasing rotation matrix, which makes the point cloud rotate around the Z-axis by an angle, and the value uniformly decreases from 90° to 0° with the number of iterations. Corresponding points are searched in the rotated ultrasound point cloud and node point cloud, and the rotation matrix R and the translation matrix t are obtained by calculation. The ultrasound point cloud coordinates are updated according to R and t, the mean square error E(R, t) between the ultrasound point cloud and the node point cloud is calculated, and finally the error value is compared with the size of the set iteration threshold. If it is greater than or equal to the iteration threshold, it is iterated again, and if it is less than or equal to the iteration threshold, the iteration is ended, and the point cloud registration work is completed.

10. The ultrasonic point cloud and deep learning based composite damage intelligent detection and modeling method of claim 1, wherein: The experimental platform in step S1 includes a mechanical arm, a water tank, an ultrasonic probe and a probe clamp.

Images