Mechanical driving type concrete internal structure three-dimensional imaging and defect precise positioning equipment

Abstract

This invention provides a mechanically driven 3D imaging and precision defect location device for internal concrete structures, relating to the field of civil engineering technology. It includes a device housing and several drive wheels. A camera is rotatably connected to one end of the housing, and an antenna is movably connected to the end of the housing furthest from the camera. Several ultrasonic sensors are fixedly connected to the inner wall of the housing. The drive wheels are rotatably connected to the perimeter of the housing via several shafts. The invention also includes a mechanical drive and scanning module, used to achieve precise, stable, and automated multidimensional movement of ultrasonic sensors on the surface or in space of the concrete structure via the drive wheels. This invention reduces the probability of inaccurate determination of the coordinates, shape, and extension trend of defects in three-dimensional space, leading to incomplete defect assessment and difficulties in subsequent repair work. It also reduces reliance on experience-based manual interpretation, lowering the probability of subjective misjudgment and omissions.

Description

Technical Field

[0001] This invention relates to the field of civil engineering technology, and in particular to a mechanically driven three-dimensional imaging and defect precision positioning device for the internal structure of concrete. Background Technology

[0002] Mechanically driven 3D imaging and defect localization equipment for concrete internal structures is a high-tech device used to detect the internal condition of concrete structures. This equipment typically combines technologies from multiple fields such as mechanical engineering, electronic engineering, computer vision, and image processing. It can perform non-destructive testing of concrete structures, achieve 3D imaging of the internal structure, and accurately locate internal defects.

[0003] Currently, concrete structures are widely used in critical infrastructure projects such as bridges, tunnels, high-rise buildings, and water conservancy projects due to their high strength and durability. However, during the pouring, curing, and long-term service of concrete, internal defects can easily develop due to factors such as material performance deterioration, load application, and environmental erosion. If these defects are not detected and addressed in a timely manner, they may gradually expand and endanger structural safety, even leading to major safety accidents such as collapses. Therefore, accurate detection and defect location of the internal structure of concrete is a core element in ensuring the safe operation of infrastructure and extending its service life. With the development of non-destructive testing technology, mechanically driven concrete internal structure testing equipment has gradually emerged. This type of equipment uses a mechanical drive mechanism... Moving the detection probe across the concrete surface enables automated scanning over a larger area, collecting more detection data. However, early equipment had significant shortcomings in 3D reconstruction and imaging: Firstly, its 3D reconstruction was mostly based on simple stitching algorithms, superimposing multiple 2D cross-sectional images to form a 3D model. This method ignored the spatial correlation between cross-sections, resulting in low resolution and poor continuity of the reconstructed 3D images, making it difficult to accurately reflect the complex internal structure of the concrete. Secondly, in terms of defect localization, the lack of effective 3D imaging algorithms made it impossible to accurately determine the coordinates, shape, and extension trend of defects in 3D space, leading to an incomplete assessment of defects and causing difficulties for subsequent repair work.

[0004] Therefore, it is necessary to provide new mechanically driven 3D imaging and defect precision localization equipment for the internal structure of concrete to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a mechanically driven 3D imaging and defect precision positioning device for the internal structure of concrete.

[0006] The mechanically driven 3D imaging and defect precision positioning device for internal concrete structures provided by this invention includes a device shell and several device drive wheels. A camera is rotatably connected to one end of the device shell, and an antenna is movably connected to the end of the device shell away from the camera. Several ultrasonic sensors are fixedly connected to the inner wall of the device shell. The several drive wheels are rotatably connected to the four sides of the device shell through several rotating shafts. The device also includes a mechanical drive and scanning module, which is used to realize the precise, stable, and automated multidimensional movement of the ultrasonic sensors on the surface or in space of the concrete structure through the several device drive wheels, and to complete a systematic and full-coverage scan of the target area. The sensing and detection module is used to transmit signals into the concrete and receive reflected / transmitted signals to obtain internal structural information. The data acquisition and signal processing module is used to acquire raw signals from the sensor in real time and perform preprocessing. At the same time, the Radon inverse transform is used to convert the preprocessed signals into tomographic image data. The 3D reconstruction and imaging module is used to address the problems of fault matching deviation and structural continuity breakage that easily occur in the stitching of 2D slice data by adopting a multimodal fusion tomographic imaging algorithm. By introducing an adaptive weight iterative optimization mechanism, it performs high-precision spatial registration and dynamic fusion of discrete 2D scan data to reconstruct a continuous 3D stereo image. At the same time, it innovatively integrates a multi-dimensional morphological parameter intelligent extraction algorithm to automatically quantify and analyze the volume, surface area, extension direction, and connectivity characteristic parameters of defects. The defect identification and location analysis module is used to automatically identify and label internal defects, providing location, size and type information, while introducing AI algorithms for defect classification and intelligent identification. The control system and human-machine interface module are used to realize equipment operation control, parameter setting and result display; Power supply and auxiliary modules are used to ensure stable system operation.

[0007] Preferably, the step of reconstructing a three-dimensional stereoscopic image with a continuous structure from two-dimensional scan data using a multimodal fusion tomography algorithm includes the following steps: S1. Multi-source data preprocessing and standardization: grayscale normalization and noise suppression are performed on the collected two-dimensional slice data, and the coordinate system of all slices is unified through spatial coordinate calibration to eliminate the initial position deviation caused by sensor displacement error. S2. Feature point extraction and multimodal registration: In view of the feature blurring of ultrasound images, an improved SIFT algorithm is used to extract the edge contour of the slice and the stable feature points of the defect area. At the same time, phase consistency features are introduced to enhance the recognition of weak edges. Meanwhile, based on the multimodal fusion strategy, coarse and fine registration of the feature point sets of adjacent slices are performed, and the spatial transformation matrix between slices is calculated. S3. Adaptive weighted iterative optimization fusion: First, a dynamic weight function is constructed. Fusion weights are assigned based on the similarity of overlapping areas of adjacent slices and the confidence of feature point matching. The pixel values ​​of overlapping areas are weighted and averaged, and an iterative optimization mechanism is introduced. S4. Three-dimensional voxel mesh construction: The registered and fused two-dimensional slice data is mapped to the three-dimensional voxel space, and an adaptive voxel resolution strategy is adopted. S5. Hole filling and boundary smoothing: Holes in the voxel mesh are identified by a 3D region growing algorithm. Interpolation filling is performed based on the gray-level gradient trend of the surrounding effective voxels. At the same time, bilateral filtering is applied to the surface boundary of the 3D model to eliminate jagged noise while preserving defect details. S6. Three-dimensional structure verification and accuracy correction: A reference marker verification mechanism is introduced, which compares the actual coordinates of the markers in the reconstructed model with the theoretical coordinates.

[0008] Preferably, the AI ​​algorithm for defect classification and intelligent identification includes the following steps: S10. Defect data acquisition and standardization: First, collect various types of data generated by the ultrasonic testing system and then standardize the data. S20. Defect sample labeling and dataset partitioning: First, the data is labeled and then divided into training set, validation set and test set in a 6:2:2 ratio. Stratified sampling is used to ensure that the distribution of defect type and size is consistent in each subset. S30. Data augmentation and noise suppression: First, the signal and image in the data are augmented, and then the noise in the signal and image is suppressed. S40. Defect feature extraction: Using the 3D-Unet network, the spatial volume, connectivity, and extension direction of defects are captured through 3D convolutional kernels. S50 AI model training and parameter optimization uses the ResNet18 architecture in deep convolutional neural networks as the base network, while tracking the accuracy and F1 score of the training set / validation set in real time. When the accuracy of the validation set drops for 3 consecutive rounds, an early stop mechanism is triggered. S60. Model evaluation and error correction: First, calculate the overall accuracy, recall rate and mIoU of the model, and count the misclassified samples. At the same time, add a feature attention module for this type of defect to the model. S70, model deployment and online inference: First, the TensorRT tool is used to quantize the model to INT8 precision and prune redundant network layers. At the same time, after the system collects new data, it is automatically input into the data acquisition and signal processing module, and the results are visualized on the human-machine interface.

[0009] Preferably, the mechanical drive and scanning module includes drive execution devices, motion guidance devices, precision control devices, and auxiliary adaptation devices. The drive execution devices are responsible for converting electrical energy into mechanical motion, which is the power source for the sensor to achieve multi-dimensional movement. The motion guidance devices are used to limit the sensor's motion path, reduce vibration and offset, and ensure stable coupling between the sensor and the concrete surface. The precision control devices are used to correct motion deviations through real-time position detection and closed-loop feedback. The auxiliary adaptation devices are used to provide adaptability for different detection environments, while ensuring the safety of equipment and personnel.

[0010] Preferably, the sensing and detection module includes a millimeter-wave sensor, an ultrasonic sensor, a grating ranging module, an optical camera, and a ranging sensor. The millimeter-wave sensor transmits echo signals to a data acquisition and signal processing module to image the internal structure of the concrete, enabling non-destructive testing of internal defects and reinforcing bars. The ultrasonic sensor maps and creates a real-time 3D visualization image of the internal structure of the concrete. The grating ranging module provides the location information of the detection device and, in conjunction with the millimeter-wave sensor module, measures the distance between reinforcing bars. The optical camera identifies target points attached to the corners of the detection surface, thereby determining the precise spatial location of the sub-test area. The ranging sensor monitors the distance between the ultrasonic probe of the 3D ultrasonic imager and the concrete detection surface in real time, ensuring good contact between the probe and the detection surface.

[0011] Preferably, the acquired sensor raw signals include raw signals from ultrasonic sensors, raw signals from millimeter-wave / radar sensors, raw signals from positioning and attitude sensors, and raw signals from environmental and auxiliary sensors.

[0012] Compared with related technologies, the mechanically driven three-dimensional imaging and defect precision positioning equipment for the internal structure of concrete provided by this invention has the following advantages: This invention employs a multimodal fusion tomography algorithm to reconstruct a continuous three-dimensional image from two-dimensional scan data. It performs grayscale normalization and noise suppression on multi-source acquired two-dimensional slice data, and unifies the coordinate system of all slices through spatial coordinate calibration. Addressing the issues of blurred features and weak edges in ultrasound and other modal images, an improved SIFT algorithm is used to extract stable feature points of slice edge contours and defect regions. Phase consistency features are introduced to enhance the identification of weak edges. A multimodal fusion strategy is combined to achieve coarse and fine registration between adjacent slices. A dynamic weighting function is constructed to adaptively allocate fusion weights based on the similarity of overlapping areas and feature point matching confidence of adjacent slices. A weighted average is applied to pixels in overlapping areas, and an iterative optimization mechanism is introduced. Then, the registration... The fused 2D data is mapped to a 3D voxel space. An adaptive voxel resolution strategy is adopted to improve resolution in key areas. A 3D region growing algorithm is used to identify void regions in the voxel mesh and interpolation is performed based on the gray-level gradient trend of the surrounding effective voxels. Bilateral filtering is used to smooth the surface boundary of the 3D model. Finally, a reference marker verification mechanism is introduced. By comparing the actual coordinates of the markers in the reconstructed model with the theoretical coordinates, the reconstruction error is quantified and feedback is provided for correction. This device reduces the probability of inaccurate determination of the coordinates, shape and extension trend of defects in 3D space, which leads to incomplete defect assessment and difficulties in subsequent repair work. It also reduces reliance on experience-based manual interpretation and lowers the probability of subjective misjudgment and omission. Attached Figure Description

[0013] Figure 1 A schematic diagram of the mechanically driven 3D imaging and defect precision positioning device for the internal structure of concrete provided by the present invention; Figure 2 This is a structural schematic diagram from another perspective provided by the present invention; Figure 3 This is a schematic diagram of the cross-sectional structure provided by the present invention; Figure 4 A structural block diagram of the mechanically driven 3D imaging and defect precision location device for internal concrete structures provided by the present invention; Figure 5 The flowchart provided by the present invention describes the process of reconstructing a three-dimensional stereoscopic image with a continuous structure from two-dimensional scan data using a multimodal fusion tomography algorithm. Figure 6 This is a flowchart illustrating the process of defect classification and intelligent recognition using the AI ​​algorithm provided in this invention.

[0014] The following are the labels in the diagram: 1. Equipment casing; 2. Equipment drive wheel; 3. Camera; 4. Antenna; 5. Ultrasonic sensor. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0016] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 as well as Figure 6 ,in, Figure 1 A schematic diagram of the mechanically driven 3D imaging and defect precision positioning device for the internal structure of concrete provided by the present invention; Figure 2 This is a structural schematic diagram from another perspective provided by the present invention; Figure 3 This is a schematic diagram of the cross-sectional structure provided by the present invention; Figure 4 A structural block diagram of the mechanically driven 3D imaging and defect precision location device for internal concrete structures provided by the present invention; Figure 5 The flowchart provided by the present invention describes the process of reconstructing a three-dimensional stereoscopic image with a continuous structure from two-dimensional scan data using a multimodal fusion tomography algorithm. Figure 6 This is a flowchart illustrating the process of defect classification and intelligent recognition using the AI ​​algorithm provided in this invention.

[0017] In the specific implementation process, such as Figures 1 to 6 As shown, the device includes a housing 1 and several drive wheels 2. A camera 3 is rotatably connected to one end of the housing 1, and an antenna 4 is movably connected to the end of the housing 1 away from the camera 3. Several ultrasonic sensors 5 are fixedly connected to the inner wall of the housing 1. The drive wheels 2 are rotatably connected to the four sides of the housing 1 through several shafts. The device also includes a mechanical drive and scanning module, which is used to realize the precise, stable, and automated multidimensional movement of the ultrasonic sensors on the concrete structure surface or in space through the drive wheels 2, and to complete a systematic and full-coverage scan of the target area. It should be noted that the mechanical drive and scanning module includes drive execution devices, motion guidance devices, precision control devices, and auxiliary adaptation devices. Drive execution devices are responsible for converting electrical energy into mechanical motion, which is the power source for the sensor to achieve multi-dimensional movement. Motion guidance devices are used to limit the sensor's movement path, reduce vibration and offset, and ensure stable coupling between the sensor and the concrete surface. Precision control devices are used to correct motion deviations through real-time position detection and closed-loop feedback. Auxiliary adaptation devices are used to provide adaptability for different detection environments, while ensuring the safety of equipment and personnel. The module's working logic consists of a preparation phase, a motion phase, a precision correction phase, and an end phase. In the preparation phase, the module is placed on the surface of the concrete component, and the "grid scanning" mode is selected through the human-machine interface. The parameters are set, and the drive wheel adsorption device is started. Motion Phase: The main motion controller issues commands, and the active drive wheel moves along a preset path—the X-axis drive wheel moves at a speed of 10mm / s, pausing every 5mm to allow the sensor to collect signals; after the X-axis single stroke is completed, the Y-axis drive wheel differentially steers, moving 5mm to enter the next stroke, with the driven guide wheel assisting in adjusting the direction; Precision correction stage: The wheel speed encoder provides real-time feedback on the driving wheel's movement distance. If the X-axis is detected to have moved 4.9mm, the main controller instructs the X-axis driving wheel to move an additional 0.1mm. When the laser positioning sensor detects a surface tilt of 1.5°, the wheel assembly lifting mechanism adjusts the height of the corresponding driving wheel to maintain the sensor's horizontal position. End phase: After the scan is completed, the drive wheel automatically returns to the initial position, the wheel brakes are engaged and fixed, the module generates a scan report, and waits for the next detection command; The sensing and detection module is used to transmit signals into the concrete and receive reflected / transmitted signals to obtain internal structural information. It should be noted that the sensing and detection module includes a millimeter-wave sensor, an ultrasonic sensor 5, a grating ranging module, an optical camera, and a ranging sensor. The millimeter-wave sensor transmits the echo signal to the data acquisition and signal processing module to image the internal structure of the concrete, enabling non-destructive testing of internal defects and reinforcing bars. The ultrasonic sensor 5 is used to map and create a real-time 3D visualization image of the internal structure of the concrete. The grating ranging module provides the position information of the detection device and, in conjunction with the millimeter-wave sensor module, measures the distance between reinforcing bars. The optical camera is used to identify target points pasted on the corners of the detection surface, thereby determining the precise spatial position of the sub-test area. The ranging sensor is used to monitor the distance between the ultrasonic probe of the 3D ultrasonic imager and the concrete detection surface in real time, ensuring good contact between the probe and the detection surface. Furthermore, in the mechanically driven 3D imaging and defect location equipment for the internal structure of concrete, the invisible internal structure of concrete is transformed into quantifiable electrical signal data through the process of "actively transmitting detection signals - receiving feedback signals - extracting structural features", providing the original basis for subsequent 3D imaging and defect analysis. The data acquisition and signal processing module is used to acquire raw signals from the sensor in real time and perform preprocessing. At the same time, the Radon inverse transform is used to convert the preprocessed signals into tomographic image data. It should be noted that the raw sensor signals collected include the raw signals from the ultrasonic sensor 5, the raw signals from the millimeter-wave / radar sensor, the raw signals from the positioning and attitude sensors, and the raw signals from the environmental and auxiliary sensors. The data acquisition and signal processing module is a key component for achieving three-dimensional imaging and precise defect localization of the internal structure of concrete. It is primarily responsible for acquiring, preprocessing, and converting the raw signals from the sensors into tomographic image data. Specifically, this includes: Step 1: Data Acquisition. The data acquisition module converts the analog signals output by the sensor into digital signals through an analog-to-digital converter so that the computer can recognize and process them. Then, the data is processed by a chip that typically integrates a high-speed ADC. Step 2: Signal preprocessing. This involves removing noise from the signal, normalizing the signal, adjusting the signal amplitude to a specific range, and extracting parameters from the original signal that reflect its essential characteristics. The 3D reconstruction and imaging module is used to address the problems of fault matching deviation and structural continuity breakage that easily occur in the stitching of 2D slice data by adopting a multimodal fusion tomographic imaging algorithm. By introducing an adaptive weight iterative optimization mechanism, it performs high-precision spatial registration and dynamic fusion of discrete 2D scan data to reconstruct a continuous 3D stereo image. At the same time, it innovatively integrates a multi-dimensional morphological parameter intelligent extraction algorithm to automatically quantify and analyze the volume, surface area, extension direction, and connectivity characteristic parameters of defects. It should be noted that the process of reconstructing a continuous three-dimensional image from two-dimensional scan data using a multimodal fusion tomography algorithm includes the following steps: S1. Multi-source data preprocessing and standardization: grayscale normalization and noise suppression are performed on the collected two-dimensional slice data, and the coordinate system of all slices is unified through spatial coordinate calibration to eliminate the initial position deviation caused by sensor displacement error. It should be noted that the core of grayscale normalization is to map the grayscale values ​​of all slices to a uniform numerical range, eliminating differences in units and intensity. Furthermore, linear normalization employs a "min-max normalization" algorithm to normalize the original grayscale values. Map to target interval Its formula is:

[0018] In the formula, , These are the maximum and minimum grayscale values ​​for a single slice, respectively. S2. Feature point extraction and multimodal registration: In view of the feature blurring of ultrasound images, an improved SIFT algorithm is used to extract the edge contour of the slice and the stable feature points of the defect area. At the same time, phase consistency features are introduced to enhance the recognition of weak edges. Meanwhile, based on the multimodal fusion strategy, coarse and fine registration of the feature point sets of adjacent slices are performed, and the spatial transformation matrix between slices is calculated. It should be noted that feature points of a single ultrasonic modality may cause registration deviations due to "defect occlusion or signal loss". By introducing "ultrasonic + electromagnetic" multimodal data, high-precision spatial registration between slices can be achieved through "coarse registration for rapid alignment and fine registration to correct deviations". Coarse registration: Fast alignment based on multimodal feature points The goal is to quickly reduce the initial positional deviation between adjacent slices, laying the foundation for precise registration. This is achieved using "multimodal feature point set matching + rigid transformation". 1. Feature point set construction: Feature points are extracted from the ultrasound and electromagnetic images of adjacent slices to form a "multimodal feature point set"; 2. Feature point matching: "kd-tree nearest neighbor search" is used to quickly match the feature point sets of adjacent slices. For each feature point in slice A, the two candidate points with the smallest Euclidean distance are searched in slice B. If the ratio of the smallest distance to the second smallest distance is <0.6, it is determined to be a matching point pair. The matching success rate must be ≥60%. 3. Calculation of the rigid transformation matrix: Based on matching point pairs, the rigid transformation matrix between slices is calculated using the "least squares method," as shown in the following formula:

[0019] The matrix is ​​applied to transform slice B to achieve coarse registration, at which point the positional deviation between adjacent slices is reduced to within 0.2mm; Fine-tuning: Nonlinear optimization based on mutual information entropy Even a deviation of 0.1-0.2 mm after coarse registration may still lead to the breakage of the three-dimensional structure, which needs to be further corrected by fine registration. The core is to "optimize the spatial transformation matrix with the goal of information consistency of multimodal data". 1. Objective function: Maximize the multimodal mutual information entropy. Mutual information entropy measures the degree of information correlation between two modal data. The more accurate the registration, the higher the overlap of information in corresponding regions of ultrasound slices and electromagnetic slices, and the greater the mutual information entropy. The multimodal mutual information entropy is defined as:

[0020] In the formula, For ultrasound sections, Electromagnetic slices, For information entropy, the goal of fine registration is to find the optimal transformation matrix. ,make maximum; 2. Transformation model, affine transformation Compared to the rigid transformation of coarse registration, the affine transformation adds a scaling factor. With shear factor This can correct the "slice stretching / compression" caused by sensor angle deviation. The transformation matrix is ​​as follows:

[0021] 3. Optimize the algorithm: gradient descent method Using mutual information entropy as the objective function, the parameters of the affine transformation matrix are iteratively optimized using the gradient descent method. ; S3. Adaptive weighted iterative optimization fusion: First, a dynamic weight function is constructed. Fusion weights are assigned based on the similarity of overlapping areas of adjacent slices and the confidence of feature point matching. The pixel values ​​of overlapping areas are weighted and averaged, and an iterative optimization mechanism is introduced. S4. Three-dimensional voxel mesh construction: The registered and fused two-dimensional slice data is mapped to the three-dimensional voxel space, and an adaptive voxel resolution strategy is adopted. S5. Hole filling and boundary smoothing: Holes in the voxel mesh are identified by a 3D region growing algorithm. Interpolation filling is performed based on the gray-level gradient trend of the surrounding effective voxels. At the same time, bilateral filtering is applied to the surface boundary of the 3D model to eliminate jagged noise while preserving defect details. S6. Three-dimensional structure verification and accuracy correction: A reference marker verification mechanism is introduced, which compares the actual coordinates of the markers in the reconstructed model with the theoretical coordinates. The defect identification and location analysis module is used to automatically identify and label internal defects, providing location, size and type information, while introducing AI algorithms for defect classification and intelligent identification. It should be noted that the AI ​​algorithm for defect classification and intelligent identification includes the following steps: S10. Defect data acquisition and standardization: First, collect various types of data generated by the ultrasonic testing system and then standardize the data. S20. Defect sample labeling and dataset partitioning: First, the data is labeled and then divided into training set, validation set and test set in a 6:2:2 ratio. Stratified sampling is used to ensure that the distribution of defect type and size is consistent in each subset. S30. Data augmentation and noise suppression: First, the signal and image in the data are augmented, and then the noise in the signal and image is suppressed. S40. Defect feature extraction: Using the 3D-Unet network, the spatial volume, connectivity, and extension direction of defects are captured through 3D convolutional kernels. It should be noted that 3D-Unet is a 3D semantic segmentation network extended from 2D-Unet, particularly suitable for processing 3D volumetric data of concrete structures. It can accurately segment internal defects while preserving spatial topology. Its core value lies in solving problems such as "large differences in defect scale, blurred boundaries, and easy omission of small defects" in 3D data through an "encode-decode" architecture and skip connections, providing an end-to-end solution for automatic defect identification in 3D concrete reconstruction. S50 AI model training and parameter optimization uses the ResNet18 architecture in deep convolutional neural networks as the base network, while tracking the accuracy and F1 score of the training set / validation set in real time. When the accuracy of the validation set drops for 3 consecutive rounds, an early stop mechanism is triggered. It should be noted that the ResNet18 architecture is a lightweight model in the ResNet family. With an 18-layer network depth, it achieves a balance between accuracy and efficiency in tasks such as image classification and feature extraction. Its core innovation is residual connection, which effectively solves the gradient vanishing / exploding problem in the training of deep neural networks. S60. Model evaluation and error correction: First, calculate the overall accuracy, recall rate and mIoU of the model, and count the misclassified samples. At the same time, add a feature attention module for this type of defect to the model. S70, Model Deployment and Online Inference: First, the TensorRT tool is used to quantize the model to INT8 precision and redundant network layers are pruned. At the same time, after the system collects new data, it is automatically input into the data acquisition and signal processing module, and the results are visualized on the human-machine interface. The control system and human-machine interface module are used to realize equipment operation control, parameter setting and result display; It should be noted that the human-computer interaction module is designed with "ease of use, intuitiveness, and engineering" as its design principles. It combines "hardware interface + software function" to meet the needs of different users. Its core components include an industrial touch screen, host computer software, and a data communication and storage module. Power supply and auxiliary modules are used to ensure stable system operation; It should be noted that the power module needs to provide continuous and stable power to various types of loads; while the auxiliary module solves the pain points of operation in complex construction site scenarios through functions such as environmental adaptation, equipment protection, and status monitoring, ensuring that the entire set of equipment can maintain high precision and high reliability in harsh environments.

[0022] The circuits and controls involved in this invention are all existing technologies and will not be described in detail here.

[0023] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A mechanically driven 3D imaging and defect precision positioning device for the internal structure of concrete, comprising a device housing (1) and a plurality of device drive wheels (2), wherein a camera (3) is rotatably connected to one end of the device housing (1), an antenna (4) is movably connected to the other end of the device housing (1) away from the camera (3), and a plurality of ultrasonic sensors (5) are fixedly connected to the inner wall of the device housing (1), and the plurality of device drive wheels (2) are rotatably connected to the periphery of the device housing (1) through a plurality of rotating shafts, characterized in that, It also includes a mechanical drive and scanning module, which is used to drive the wheels (2) through several devices to realize the precise, stable and automated multidimensional movement of the ultrasonic sensor on the surface or in space of the concrete structure, and to complete the systematic and full-coverage scanning of the target area; The sensing and detection module is used to transmit signals into the concrete and receive reflected / transmitted signals to obtain internal structural information. The data acquisition and signal processing module is used to acquire raw signals from the sensor in real time and perform preprocessing. At the same time, the Radon inverse transform is used to convert the preprocessed signals into tomographic image data. The 3D reconstruction and imaging module is used to address the problems of fault matching deviation and structural continuity breakage that easily occur in the stitching of 2D slice data by adopting a multimodal fusion tomographic imaging algorithm. By introducing an adaptive weight iterative optimization mechanism, it performs high-precision spatial registration and dynamic fusion of discrete 2D scan data to reconstruct a continuous 3D stereo image. At the same time, it innovatively integrates a multi-dimensional morphological parameter intelligent extraction algorithm to automatically quantify and analyze the volume, surface area, extension direction, and connectivity characteristic parameters of defects. The defect identification and location analysis module is used to automatically identify and label internal defects, providing location, size and type information, while introducing AI algorithms for defect classification and intelligent identification. The control system and human-machine interface module are used to realize equipment operation control, parameter setting and result display; Power supply and auxiliary modules are used to ensure stable system operation.

2. The mechanical drive concrete internal structure 3D imaging and defect precise positioning equipment according to claim 1, characterized in that, The process of reconstructing a continuous three-dimensional image from two-dimensional scan data using a multimodal fusion tomography algorithm includes the following steps: S1. Multi-source data preprocessing and standardization: grayscale normalization and noise suppression are performed on the collected two-dimensional slice data, and the coordinate system of all slices is unified through spatial coordinate calibration to eliminate the initial position deviation caused by sensor displacement error. S2. Feature point extraction and multimodal registration: In view of the feature blurring of ultrasound images, an improved SIFT algorithm is used to extract the edge contour of the slice and the stable feature points of the defect area. At the same time, phase consistency features are introduced to enhance the recognition of weak edges. Meanwhile, based on the multimodal fusion strategy, coarse and fine registration of the feature point sets of adjacent slices are performed, and the spatial transformation matrix between slices is calculated. S3. Adaptive weighted iterative optimization fusion: First, a dynamic weight function is constructed. Fusion weights are assigned based on the similarity of overlapping areas of adjacent slices and the confidence of feature point matching. The pixel values ​​of overlapping areas are weighted and averaged, and an iterative optimization mechanism is introduced. S4. Three-dimensional voxel mesh construction: The registered and fused two-dimensional slice data is mapped to the three-dimensional voxel space, and an adaptive voxel resolution strategy is adopted. S5. Hole filling and boundary smoothing: Holes in the voxel mesh are identified by a 3D region growing algorithm. Interpolation filling is performed based on the gray-level gradient trend of the surrounding effective voxels. At the same time, bilateral filtering is applied to the surface boundary of the 3D model to eliminate jagged noise while preserving defect details. S6. Three-dimensional structure verification and accuracy correction: A reference marker verification mechanism is introduced, which compares the actual coordinates of the markers in the reconstructed model with the theoretical coordinates.

3. The mechanical drive concrete internal structure 3D imaging and defect precise positioning equipment according to claim 2, characterized in that, The AI ​​algorithm performs defect classification and intelligent identification, including the following steps: S10. Defect data acquisition and standardization: First, collect various types of data generated by the ultrasonic testing system and then standardize the data. S20. Defect sample labeling and dataset partitioning: First, the data is labeled and then divided into training set, validation set and test set in a 6:2:2 ratio. Stratified sampling is used to ensure that the distribution of defect type and size is consistent in each subset. S30. Data augmentation and noise suppression: First, the signal and image in the data are augmented, and then the noise in the signal and image is suppressed. S40. Defect feature extraction: Using the 3D-Unet network, the spatial volume, connectivity, and extension direction of defects are captured through 3D convolutional kernels. S50 AI model training and parameter optimization uses the ResNet18 architecture in deep convolutional neural networks as the base network, while tracking the accuracy and F1 score of the training set / validation set in real time. When the accuracy of the validation set drops for 3 consecutive rounds, an early stop mechanism is triggered. S60. Model evaluation and error correction: First, calculate the overall accuracy, recall rate and mIoU of the model, and count the misclassified samples. At the same time, add a feature attention module for this type of defect to the model. S70, model deployment and online inference: First, the TensorRT tool is used to quantize the model to INT8 precision and prune redundant network layers. At the same time, after the system collects new data, it is automatically input into the data acquisition and signal processing module, and the results are visualized on the human-machine interface.

4. The mechanical drive concrete internal structure 3D imaging and defect precise positioning equipment according to claim 3, characterized in that, The mechanical drive and scanning module includes drive execution devices, motion guidance devices, precision control devices, and auxiliary adaptation devices. The drive execution devices are responsible for converting electrical energy into mechanical motion, which is the power source for the sensor to achieve multi-dimensional movement. The motion guidance devices are used to limit the sensor's movement path, reduce vibration and offset, and ensure stable coupling between the sensor and the concrete surface. The precision control devices are used to correct motion deviations through real-time position detection and closed-loop feedback. The auxiliary adaptation devices are used to provide adaptability for different detection environments, while ensuring the safety of equipment and personnel.

5. The mechanical drive concrete internal structure 3D imaging and defect precise positioning equipment according to claim 4, characterized in that, The sensing and detection module includes a millimeter-wave sensor, an ultrasonic sensor (5), a grating ranging module, an optical camera, and a ranging sensor. The millimeter-wave sensor transmits the echo signal to the data acquisition and signal processing module to image the internal structure of the concrete and realize non-destructive testing of internal defects and reinforcing bars. The ultrasonic sensor (5) is used to map and create a real-time 3D visualization image of the internal structure of the concrete. The grating ranging module can provide the position information of the detection device and, in conjunction with the millimeter-wave sensor module, measure the distance between reinforcing bars. The optical camera is used to identify the target points pasted on the corners of the detection surface, thereby determining the precise spatial position of the sub-test area. The ranging sensor is used to monitor the distance between the ultrasonic probe of the three-dimensional ultrasonic imager and the concrete detection surface in real time to ensure good contact between the probe and the detection surface.

6. The mechanical drive concrete internal structure 3D imaging and defect precise positioning equipment according to claim 5, characterized in that, The acquisition sensor raw signals include raw signals of ultrasonic sensors (5), raw signals of millimeter wave / radar sensors, raw signals of positioning and attitude sensors, and raw signals of environment and auxiliary sensors.

Images