Mobile bridge multi-sensor 3D scanning system

Abstract

This application discloses a mobile bridge-type multi-sensor 3D scanning system, belonging to the field of 3D scanning technology. It includes: a measurement platform for placing the object to be measured; a three-coordinate bridge measurement structure for omnidirectional measurement of the object; a dynamic encoding transmission module for dynamically adjusting the signal transmission direction or frequency; a multimodal sensing module for receiving light reflected from the object and capturing data from different types of sensors to achieve full-dimensional data acquisition of the object; and a data fusion processor with a built-in hypergraph neural network model to achieve feature fusion of multi-source heterogeneous data received from different types of sensors. This invention achieves high-precision, omnidirectional, and intelligent measurement of the object.

Description

Technical Field

[0001] This application relates to a mobile bridge-type multi-sensor 3D scanning system, belonging to the field of 3D scanning technology. Background Technology

[0002] With the increasing demands for precision and comprehensiveness in acquiring three-dimensional morphology and material information of objects in various fields such as industrial manufacturing, quality inspection, and digital preservation of cultural relics, 3D scanning technology has received widespread attention and application. In existing technologies, traditional 3D scanning systems mostly employ single sensors or simple multi-sensor combinations, resulting in insufficient flexibility in the measurement structure. This makes it difficult to achieve comprehensive and accurate measurement of the object under complex spatial postures. Data collected by multiple sensors is often processed independently, making it difficult to fully explore the potential correlations between different types of sensor data. Consequently, the capture of the object's morphology and material characteristics is not comprehensive or in-depth enough, hindering the ability to quickly and accurately provide risk decision-making results and failing to meet the practical application requirements for high-precision, comprehensive, and intelligent measurement of objects. Summary of the Invention

[0003] According to one aspect of this application, a mobile bridge-type multi-sensor 3D scanning system is provided, which enables high-precision, omnidirectional, and intelligent measurement of the object being measured.

[0004] A mobile bridge-type multi-sensor 3D scanning system includes:

[0005] A measuring platform, used to place the object being measured;

[0006] A three-coordinate bridge measurement structure is installed on a measurement platform for adjusting the measurement orientation. A detection head is installed on the three-coordinate bridge measurement structure, and the detection head integrates a dynamic encoding transmission module and a multi-modal sensing module to realize omnidirectional measurement of the object under test.

[0007] A dynamic encoding transmission module, integrated into the detection head, is configured to dynamically adjust the signal transmission direction or frequency; the polarization encoding unit of the dynamic encoding transmission module generates annular polarized light containing topological phase;

[0008] A multimodal sensing module, integrated into the detection head, includes a visual layer and a tactile layer. The visual layer captures a dynamic image sequence of the surface of the object under test through a multi-camera array. The tactile layer generates a non-contact force distribution matrix through a pressure sensor array and uses light reflected from the object under test and data received by different types of sensors to capture data of the object under test in all dimensions based on the non-contact force distribution matrix.

[0009] The data fusion processor incorporates a hypergraph neural network model to achieve feature fusion of multi-source heterogeneous data received from different types of sensors. Based on the data of the object under test, the hypergraph neural network model constructs a multi-view view of the object's shape and material. Nodes represent features from different types of sensors, and hyperedges represent cross-modal correlations, forming a multi-view hypergraph. Simultaneously, it identifies a subset of key features related to the object's shape and material characteristics, and uses this subset to reconstruct the multi-view hypergraph structure, thereby updating the hypergraph neural network. Based on a preset risk assessment threshold for the object's shape and material characteristics, the updated hypergraph neural network determines the risk decision result for the object.

[0010] Furthermore, the coordinate measuring bridge structure includes:

[0011] An X-axis motion mechanism is disposed on the upper surface of the measuring platform. The object to be measured is located above the X-axis motion mechanism, and the X-axis motion mechanism guides the object to be measured to move along the X-axis direction of the measuring platform.

[0012] A cable tray is horizontally mounted above the measuring platform and fixed on both sides of the measuring platform. A Z-axis motion mechanism is provided on the cable tray. The Z-axis motion mechanism moves horizontally along the cable tray to realize the Y-axis movement of the detection head. The detection head is mounted on the Z-axis motion mechanism to realize the Z-axis movement of the detection head.

[0013] The X-axis motion mechanism is located below the bridge frame.

[0014] Furthermore, the dynamic coding transmission module includes:

[0015] Wavelength adaptive unit automatically selects the illumination wavelength based on the surface material of the object being measured;

[0016] Polarization coding unit modulates the topological phase distribution of ring-polarized light;

[0017] The safety protection unit monitors optical power in real time and triggers an emergency shutdown.

[0018] The frequency range of the signal is dynamically limited by the encoding sequence, and invalid encoding sequences are automatically filtered out.

[0019] Furthermore, the multimodal sensing module is implemented through a sensor array;

[0020] The sensor array includes a visual layer and a tactile layer;

[0021] The visual layer includes:

[0022] The central camera defines the first search area with the image center as the origin, and its radius adaptively expands with the target's movement speed.

[0023] Multiple ring cameras are arranged circumferentially around the central camera, and the fan-shaped imaging area of ​​each camera covers the blind area of ​​the first search area;

[0024] The visual layer uses an edge prediction algorithm to estimate the position of a high-speed moving target and aligns it with the data from the tactile layer in time and space.

[0025] The tactile layer includes:

[0026] A capacitive sensing array, deployed in the central region, maps the surface morphology of an object through changes in capacitance.

[0027] An ultrasonic sensing unit is deployed in the peripheral area to improve resolution through echo superposition.

[0028] The data from the capacitive sensing array and the ultrasonic sensing unit are fused using a hypergraph neural network.

[0029] Furthermore, the data fusion processor includes:

[0030] The data calibration module establishes the mapping relationship between sensor feature vectors and hypergraph nodes;

[0031] The dynamic reconstruction module filters key feature subsets and reconstructs the hypergraph structure based on gradient risk assessment.

[0032] The risk assessment module outputs risk probabilities through a fully connected layer and a Softmax function;

[0033] The set of hyperedges in the hypergraph neural network is adaptively adjusted through k-means clustering.

[0034] Furthermore, the hypergraph neural network performs the following operations:

[0035] Construct a node feature matrix and a hyperedge correlation matrix, where the nodes represent different sensor feature dimensions;

[0036] Node features are updated iteratively through hypergraph convolutional layers, fusing local and global relationships;

[0037] A continuous learning framework is used to dynamically optimize the hyperedge weights and redundant nodes are removed based on feature similarity.

[0038] Furthermore, it also includes cross-control logic:

[0039] The dynamic coding emission module adjusts the polarization coding parameters in real time based on the mechanical distribution data of the tactile layer.

[0040] The three-coordinate bridge measurement structure dynamically plans the obstacle avoidance scanning path based on the output of the risk assessment module.

[0041] When the security protection unit detects an abnormal optical power, it synchronously triggers the self-test protocol of the data fusion processor.

[0042] The beneficial effects that this application can produce include:

[0043] The mobile bridge-type multi-sensor 3D scanning system provided in this application features a three-coordinate bridge measurement structure, enabling flexible adjustment of the measurement orientation. Combined with a detection head integrating a dynamic encoding transmission module and a multi-modal sensing module, it can dynamically optimize signal transmission and achieve synchronous acquisition of multi-modal data, completing omnidirectional, high-precision measurement of the object under test. The hypergraph neural network model built into the data fusion processor can deeply mine cross-modal correlations of multi-source heterogeneous data, accurately fuse features from different sensors, construct multi-view perspectives, and effectively capture rich information about the shape and material of the object under test. By identifying key feature subsets to reconstruct the hypergraph structure and update the model, data processing efficiency and accuracy are significantly improved. Finally, based on risk assessment thresholds, it quickly outputs risk decision results, providing an efficient, intelligent, and accurate 3D measurement solution for industrial inspection and other scenarios. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a mobile bridge-type multi-sensor 3D scanning system according to one embodiment of this application;

[0045] Figure 2 This is a schematic diagram of a three-coordinate bridge measurement structure in one embodiment of this application;

[0046] List of components and reference numerals: 1-Measuring platform; 2-Detection head; 3-X-axis motion mechanism; 4-Bridge; 5-Z-axis motion mechanism. Detailed Implementation

[0047] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.

[0048] See Figure 1-2 A mobile bridge-type multi-sensor 3D scanning system, characterized in that it includes:

[0049] Measurement platform 1 is used to place the object to be measured;

[0050] A three-coordinate bridge measurement structure is installed on a measurement platform 1 for adjusting the measurement orientation. A detection head 2 is installed on the three-coordinate bridge measurement structure. The detection head 2 integrates a dynamic encoding transmission module and a multi-modal sensing module to realize omnidirectional measurement of the object under test.

[0051] A dynamic encoding transmission module, integrated into the detection head, is configured to dynamically adjust the signal transmission direction or frequency; the polarization encoding unit of the dynamic encoding transmission module generates annular polarized light containing topological phase;

[0052] A multimodal sensing module, integrated into the detection head, includes a visual layer and a tactile layer. The visual layer captures a dynamic image sequence of the surface of the object under test through a multi-camera array. The tactile layer generates a non-contact force distribution matrix through a pressure sensor array and uses light reflected from the object under test and data received by different types of sensors to capture data of the object under test in all dimensions based on the non-contact force distribution matrix.

[0053] The data fusion processor incorporates a hypergraph neural network model to achieve feature fusion of multi-source heterogeneous data received from different types of sensors. Based on the data of the object under test, the hypergraph neural network model constructs a multi-view view of the object's shape and material. Nodes represent features from different types of sensors, and hyperedges represent cross-modal correlations, forming a multi-view hypergraph. Simultaneously, it identifies a subset of key features related to the object's shape and material characteristics, and uses this subset to reconstruct the multi-view hypergraph structure, thereby updating the hypergraph neural network. Based on a preset risk assessment threshold for the object's shape and material characteristics, the updated hypergraph neural network determines the risk decision result for the object.

[0054] Specifically, the measurement platform 1 is the basic support structure of the entire scanning system. Its main function is to place the object under test, providing a stable placement position for subsequent measurement work and ensuring that the object remains relatively stationary or moves in a predetermined manner during the scanning process to accurately acquire its data. The coordinate measuring machine (CMM) bridge structure is installed on the measurement platform 1. As the core mechanical structure of the measurement system, its key function is to adjust the measurement orientation. Through precise movement and positioning in three coordinate directions (usually the X, Y, and Z axes), the position and orientation of the detection head 2 can be flexibly changed, thereby enabling measurements of different parts and angles of the object under test, ensuring comprehensive coverage of the object's surface. The detection head 2, installed on the CMM bridge structure, is a key component for data acquisition. The detection head 2 integrates a dynamic encoding transmission module and a multimodal sensing module. These two modules work together to complete the measurement task of the object under test.

[0055] The dynamic encoding emission module dynamically adjusts the signal emission direction or frequency. During 3D scanning, the dynamic encoding emission module can change the signal characteristics in real time according to the shape, material, and measurement requirements of the object being measured. For example, for objects with complex shapes, the emission direction may need to be adjusted to better illuminate all surfaces of the object; for objects of different materials, the emission frequency may need to be changed to obtain more accurate reflected signals, thus providing a higher quality signal source for subsequent data acquisition. The multimodal sensing module is responsible for receiving the light reflected from the object being measured. This light carries various information about the surface of the object, such as shape, texture, and color. Through data received from different types of sensors, the multimodal sensing module can achieve full-dimensional capture of the object's data. Different types of sensors may have different working principles and measurement characteristics; for example, laser sensors can accurately measure the distance and shape of an object, while optical sensors can acquire color and texture information. The data collected by these different types of sensors complement each other, jointly constructing complete information about the object being measured.

[0056] The data fusion processor, with its built-in hypergraph neural network model, is the core of the entire system's data processing. Its main function is to achieve feature fusion of multi-source heterogeneous data received from different types of sensors. During 3D scanning, the data acquired by various sensors has different formats, structures, and semantic information, constituting multi-source heterogeneous data. The hypergraph neural network model can effectively process this complex data and organically fuse it. Based on the measured object data, the hypergraph neural network model constructs a multi-view view of the object's shape and material. In this hypergraph structure, nodes represent features from different types of sensors, and hyperedges represent cross-modal correlations. This representation method can intuitively show the interrelationships between different sensor features and describe the measured object from multiple perspectives. The model identifies key feature subsets related to the object's shape and material characteristics. These key feature subsets are crucial for accurately describing the object's properties. Using the identified key feature subsets, the multi-view hypergraph structure is reconstructed, and the hypergraph neural network is updated. In this way, the model can continuously optimize itself, improving its ability to extract and analyze the features of the measured object. Based on preset risk assessment thresholds for the morphology and material characteristics of the object under test, an updated hypergraph neural network is used to determine the risk decision result for the object under test. For example, in industrial inspection, factors such as the degree of morphological defects and material abnormalities of the object under test, combined with preset risk assessment thresholds, can be used to determine whether the object under test meets quality standards, whether there are potential risks, and to provide corresponding decision results, such as qualified, unqualified, or requiring further testing.

[0057] The coordinate measuring bridge structure includes:

[0058] X-axis motion mechanism 3 is disposed on the upper surface of the measuring platform 1. The object to be measured is located above the X-axis motion mechanism 3. The X-axis motion mechanism 3 guides the object to be measured to move along the X-axis direction of the measuring platform 1.

[0059] A cable tray 4 is horizontally mounted above the measuring platform 1 and fixed on both sides of the measuring platform 1. A Z-axis motion mechanism 5 is provided on the cable tray 4. The Z-axis motion mechanism 5 moves horizontally along the cable tray 4 to realize the Y-axis movement of the detection head 2. The detection head 2 is mounted on the Z-axis motion mechanism 5 to realize the Z-axis movement of the detection head 2.

[0060] The X-axis motion mechanism 3 is located below the bridge frame 4.

[0061] Specifically, the X-axis motion mechanism 3 is positioned on the upper surface of the measuring platform 1, below the cable tray 4. This design ensures both the compactness of the overall structure and facilitates coordinated movement with the measuring platform 1 and the cable tray 4. The object to be measured is placed above the X-axis motion mechanism 3. When the X-axis motion mechanism 3 is activated, it directly moves the object to be measured along the X-axis of the measuring platform 1. This design allows for precise positioning and movement of the object to be measured along the X-axis, providing a foundation for subsequent measurements at different locations. By guiding the object to be measured along the X-axis, scanning measurements of different parts of the object in the X-axis direction can be achieved, thereby obtaining complete morphological and feature information of the object in the X-axis direction. The cable tray 4 is horizontally mounted above the measuring platform 1 and fixed to both sides of the measuring platform 1. This fixing method ensures the stability and reliability of the cable tray 4, enabling it to withstand the weight of the Z-axis motion mechanism 5 and the detection head 2, as well as various forces generated during movement. The cable tray 4, as a crucial supporting component of the entire measurement structure, not only provides the mounting foundation for the Z-axis motion mechanism 5 but also, through its horizontally mounted structure, enables the Z-axis motion mechanism 5 to move horizontally on the cable tray 4, thereby indirectly achieving the movement of the detection head 2 in the Y-axis direction. The Z-axis motion mechanism 5, mounted on the cable tray 4, can move horizontally along the cable tray 4 in cooperation with it. The detection head 2 is mounted on the Z-axis motion mechanism 5, which can not only move horizontally on the cable tray 4 (achieving the Y-axis movement of the detection head 2) but also drive the detection head 2 to move vertically in the Z-axis direction. This dual-axis motion capability allows the detection head 2 to be flexibly positioned and moved in three-dimensional space. The horizontal movement of the Z-axis motion mechanism 5 on the cable tray 4 achieves the positioning and scanning of the detection head 2 in the Y-axis direction; the vertical movement of the Z-axis motion mechanism 5 itself achieves the positioning and scanning of the detection head 2 in the Z-axis direction. Combined with the X-axis motion mechanism 3 driving the object under test to move in the X-axis direction, the three work together to enable the detection head 2 to perform all-round, multi-angle scanning and measurement of the object under test, thereby obtaining complete and accurate three-dimensional data of the object under test.

[0062] During the measurement process, the X-axis motion mechanism 3, the Z-axis motion mechanism 5 on the bridge 4, and the detection head 2 cooperate to form a complete three-coordinate measurement system. The X-axis motion mechanism 3 is responsible for the movement of the object being measured in the X-axis direction, and the Z-axis motion mechanism 5 drives the detection head 2 to move in the Y-axis and Z-axis directions. Through the coordinated movement in the three coordinate directions, the detection head 2 can accurately reach each measurement point of the object being measured, realizing omnidirectional and high-precision measurement of the object being measured.

[0063] It is worth noting that the motion drive of the X, Y, and Z axes in the three-coordinate bridge measurement structure can be achieved using existing drive structures.

[0064] The dynamic coding transmission module includes:

[0065] Wavelength adaptive unit automatically selects the illumination wavelength based on the surface material of the object being measured;

[0066] Polarization coding unit modulates the topological phase distribution of ring-polarized light;

[0067] The safety protection unit monitors optical power in real time and triggers an emergency shutdown.

[0068] The frequency range of the signal is dynamically limited by the encoding sequence, and invalid encoding sequences are automatically filtered out.

[0069] The modulation of the topological phase includes:

[0070] ;

[0071] In the formula, l represents the topological charge number, l∈{1,2,3,4}; f represents the equivalent lens focal length; The wavelength is represented by r; the radial distance is represented by r. Indicates azimuth;

[0072] The safety protection unit monitors the light power in real time and triggers an emergency shutdown when it exceeds a threshold.

[0073] The frequency range of the transmitted signal is (A, B), where A and B are both positive integers, and A is the lower limit of the range and B is the upper limit of the range. If the frequency corresponding to the coded sequence exceeds the range limit, the coded sequence corresponding to the transmitted signal exceeding the range limit is set as invalid code, and the coded sequence corresponding to the transmitted signal within the range limit is set as valid code.

[0074] When the effective coding period is 0, it indicates a single transmission. When it is not 0, the coding period should be greater than the sum of the interval times of all frequencies corresponding to the coding sequence, and the difference between the coding period and the sum of the interval times of all frequencies should be greater than F.

[0075] Specifically, the wavelength adaptive unit has the ability to automatically select the appropriate illumination wavelength based on the object being measured. During 3D scanning, objects of different materials and properties exhibit varying characteristics in reflecting and absorbing light of different wavelengths. For example, some materials have higher reflectivity for specific wavelengths, while others may absorb certain wavelengths. The wavelength adaptive unit, through built-in sensors or algorithms, analyzes relevant information about the object being measured (such as color and material type) and automatically selects the most suitable illumination wavelength. This adaptive selection significantly improves the interaction between the signal and the object, enhancing the intensity and quality of the reflected signal, thus providing a more reliable foundation for subsequent measurement and data processing. By optimizing the illumination wavelength, the adaptability and measurement accuracy of the scanning system to objects of various materials can be effectively improved, reducing measurement errors caused by wavelength mismatch and ensuring accurate and clear data of the object being measured in different measurement scenarios. The polarization encoding unit is responsible for generating ring-polarized light containing topological phase. Topological phase is a special phase characteristic that can carry additional information. By introducing topological phase into ring-polarized light, unique encoding characteristics can be assigned to the emitted signal. In the mentioned topological phase modulation formula, the topological charge number *l* is a key parameter, belonging to the set {1, 2, 3, 4}. Different topological charge numbers correspond to different topological phase distributions, providing multiple possibilities for signal encoding. The equivalent lens focal length *f* also affects the modulation effect of the topological phase. By adjusting the values ​​of *l* and *f*, the characteristics of the topological phase in ring-polarized light can be precisely controlled, enabling the encoding of the transmitted signal. This encoding method increases the complexity and information content of the signal, helping to improve the system's anti-interference capability and measurement accuracy in complex environments. Utilizing topological phase encoding provides a novel and effective signal encoding method for 3D scanning systems, enhancing the uniqueness and identifiability of the signal. Especially in complex measurement environments with noise interference or multipath reflections, it helps to accurately extract information about the measured object, improving system reliability and measurement accuracy. The primary responsibility of the safety protection unit is to monitor the light power in real time. During the operation of the dynamic encoding transmission module, the stability and safety of the light power are crucial. Excessive light power can damage the object being measured (e.g., causing thermal or chemical effects on sensitive materials) and the optical components of the measurement system itself, affecting its normal operation and lifespan. The safety protection unit monitors the light power in real time. If the power exceeds a preset threshold, it immediately triggers an emergency shutdown mechanism to cut off light emission, thus protecting both the object being measured and the measurement system. This safety protection unit provides reliable safety for the entire scanning system, ensuring it operates within a safe range under various measurement conditions. It prevents equipment damage or measurement accidents caused by abnormal light power, improving system stability and reliability.

[0076] The frequency range of the transmitted signal is limited to (A, B), where A and B are both positive integers, A representing the lower limit of the range and B representing the upper limit. This frequency range setting ensures that the transmitted signal operates within a specific frequency band to meet the system's measurement requirements and electromagnetic compatibility requirements. For example, it avoids frequency interference with other wireless devices or electronic systems while ensuring the stability and accuracy of the signal during transmission and processing. When the frequency corresponding to the coded sequence exceeds the range limit, the coded sequence corresponding to the transmitted signal exceeding the range limit is set as invalid; while the coded sequence corresponding to the transmitted signal within the range limit is set as valid. This judgment rule clearly defines the validity of the transmitted signal encoding, ensuring that only coded sequences that meet the frequency range requirements can be recognized and processed by the system, avoiding interference from invalid signals to the measurement results. For valid encoding, there are specific requirements for the encoding period. When the encoding period is 0, it indicates a single transmission, meaning that the signal is transmitted once and then not repeated. When the encoding period is non-zero, the encoding period should be greater than the sum of the interval times of all frequencies corresponding to the coded sequence, and the difference between the encoding period and the sum of the interval times of all frequencies should be greater than F. This rule is set to ensure that the coded signals have a reasonable time interval and rhythm during transmission, avoid signal overlap or confusion due to improper frequency interval scheduling, and ensure that each coded signal can be accurately identified and processed, thereby guaranteeing the accuracy and reliability of the measurement data.

[0077] The multimodal sensing module is implemented through a sensor array;

[0078] The sensor array includes a visual layer and a tactile layer;

[0079] The visual layer includes:

[0080] The central camera defines the first search area with the image center as the origin, and its radius adaptively expands with the target's movement speed.

[0081] Multiple ring cameras are arranged circumferentially around the central camera, and the fan-shaped imaging area of ​​each camera covers the blind area of ​​the first search area;

[0082] The visual layer uses an edge prediction algorithm to estimate the position of a high-speed moving target and aligns it with the data from the tactile layer in time and space.

[0083] The tactile layer includes:

[0084] A capacitive sensing array, deployed in the central region, maps the surface morphology of an object through changes in capacitance.

[0085] An ultrasonic sensing unit is deployed in the peripheral area to improve resolution through echo superposition.

[0086] The data from the capacitive sensing array and the ultrasonic sensing unit are fused using a hypergraph neural network.

[0087] Specifically, this multimodal sensing module utilizes a sensor array to acquire multidimensional information. The sensor array employs a hierarchical structure, including a visual layer and a tactile layer. This hierarchical design integrates sensors of different sensing types, fully leveraging the advantages of each sensor to comprehensively acquire information about the object being measured from various angles, providing a rich data source for subsequent data fusion and analysis.

[0088] The vision layer consists of a central camera and multiple ring cameras arranged circumferentially around it. The central camera, as the core of the vision system, typically possesses high resolution and imaging quality, enabling it to acquire clear and accurate image information of the central area of ​​the object being measured. The multiple ring cameras arranged around the central camera expand the visual coverage, achieving omnidirectional, blind-spot-free image capture of the object. Through the collaborative work of the central and ring cameras, the vision layer captures dynamic image sequences of the object's surface. During 3D scanning, the object may be stationary or in motion; the dynamic image sequences record changes in the object's surface features at different times, including shape, texture, and color. This image data provides crucial visual evidence for subsequent 3D reconstruction, feature extraction, and data analysis, helping to more accurately reconstruct the true shape of the object.

[0089] The tactile layer employs a pressure sensor array, composed of multiple pressure sensors arranged according to a specific pattern. These pressure sensors can detect the magnitude and distribution of applied pressure. By integrating multiple pressure sensors into an array, pressure information can be collected over a large area. This pressure sensor array is used to generate a non-contact force distribution matrix. In the 3D scanning scene, although named the "tactile layer," it actually acquires information through non-contact methods (such as sensing the interaction forces or pressure distribution between the object and the sensor based on optical or electromagnetic principles). This non-contact force distribution matrix reflects the distribution of forces acting on the surface of the object at different locations, such as local deformation of the surface and the influence of surface roughness on forces. By analyzing the non-contact force distribution matrix, information on the surface mechanical properties of the object, such as hardness and elasticity, can be obtained. This information complements the image information acquired by the visual layer, providing strong support for a comprehensive understanding of the object's characteristics.

[0090] It's worth noting that the visual and tactile layers acquire information about the object being measured from different perceptual dimensions. Visual information intuitively displays the object's appearance features, while tactile information reflects its mechanical properties. The combination of these two layers overcomes the limitations of a single sensor in information acquisition, providing a more comprehensive and accurate description of the object. The fusion of multimodal information allows for mutual verification and complementarity, reducing measurement bias caused by errors from a single sensor. For example, when visual information is affected by factors such as lighting or occlusion, tactile information can provide auxiliary judgment; conversely, tactile information may have some ambiguity during acquisition, while visual information can provide more precise spatial positioning and feature recognition. The design of the multimodal sensing module enables the 3D scanning system to adapt to a wider range of application scenarios. In industrial inspection, it can not only accurately measure the size and shape of objects but also assess their surface quality and mechanical properties; in the medical field, it can be used for 3D modeling and mechanical property analysis of human tissues or organs; and in robot grasping and manipulation tasks, it helps robots better perceive the characteristics of objects and environmental information, achieving more intelligent and precise operations.

[0091] The images captured by the central camera are labeled P1, and the images captured by the ring camera are labeled P2, P3, ..., P... n Where n is the number of ring cameras, which is a positive integer;

[0092] A first search region is set for the marked image P1. A circular region with radius R1 is defined with the image center as the origin, where R1 = image width × 0.4. When the target speed of the object being measured is greater than V1, R1 is automatically expanded to R2 = R1 × 1.5, and the edge prediction algorithm is enabled to estimate the target position. The first search region covers the object being measured.

[0093] For labeled images P2, P3, ..., P n A second search region is set up, and based on the camera mounting angle θi, the angle range [θi] of each sector within the second search region is defined. i -Δθ,θ i +Δθ], Δθ=30°, the sector area is the imaging area of ​​each ring camera, wherein the second search area covers the blind zone of the first search area.

[0094] Specifically, the images captured by the central camera are labeled P1, and the images captured by the ring cameras are labeled P2, P3, ..., P... n(n is the number of ring cameras and is a positive integer). This marking method facilitates the differentiation and management of images acquired by different cameras, providing clear identification for subsequent image processing and target search operations. A circular region with radius R1 is defined as the first search region, centered at the image center, where R1 = image width × 0.4. In most cases, the main features or target of the object being measured may be concentrated near the central region of the image. Setting the initial search region as a circular region with a radius of 0.4 times the image width and centered at the image center can effectively narrow the search range, improve search efficiency, and reduce unnecessary computation. When the target's movement speed > V1, R1 automatically expands to R2 = R1 × 1.5, and the edge prediction algorithm is enabled to estimate the target position. This is because when the target moves quickly, it may have moved out of the initially set small search region in the next frame. By increasing the search region radius by 1.5 times, the probability of finding a fast-moving target is increased. Meanwhile, enabling edge prediction algorithms can estimate the target's possible position in the current frame based on information such as its position, velocity, and direction of motion in the previous frame, further improving the accuracy and efficiency of the search and ensuring effective tracking and capture of the target even when it is moving rapidly. The first search area needs to cover the object being measured; this is a fundamental prerequisite for accurately acquiring information about the object. If the search area cannot cover the object, it may lead to target loss or incomplete information acquisition, affecting subsequent measurement and analysis results.

[0095] Based on camera installation angle θ i Define the angle range of each sector in the second search region as [θ]. i −Δθ,θ i +Δθ], where Δθ = 30°. Since the ring cameras are arranged circumferentially around the central camera, each ring camera has its specific installation angle θ. i Centered on this installation angle, a fan-shaped area is defined extending 30° to the left and right as the search area. This setup fully considers the imaging angle and layout characteristics of the ring camera, rationally dividing the search area and ensuring that each ring camera is responsible for searching for targets within its specific field of view. The second search area needs to cover the blind spots of the first search area. Since the first search area of ​​the central camera is circular, some locations at the edges or corners of the image may not be effectively covered, forming blind spots in the first search area. By rationally arranging the ring cameras and setting their search areas to fan-shaped, the fan-shaped search areas of multiple ring cameras cooperate to fill the blind spots of the central camera's circular search area, achieving omnidirectional, blind-spot-free search and monitoring of the object under test, ensuring that complete information about the object can be obtained.

[0096] It is worth noting that by setting reasonable search areas for different cameras and dynamically adjusting the search areas according to the target's motion state, combined with edge prediction algorithms, the search efficiency and accuracy of the measured object can be significantly improved. In complex 3D scanning scenarios, quickly and accurately locating and tracking the measured object is crucial to ensuring measurement quality, and this search area setting method helps achieve this goal. Considering factors such as rapid target movement and blind spots in the search area, the system's adaptability and robustness under different operating conditions are enhanced through dynamically expanding the search area and the complementary setting of multiple camera search areas. Even when the target's motion state changes or some areas are difficult to search, the system can still reliably acquire information about the measured object, ensuring the stability and reliability of the 3D scanning system. Simultaneously, accurately searching for and acquiring information about the measured object is the foundation of multimodal data fusion. Through reasonable search area settings, it is ensured that the image data acquired by different cameras in the visual layer can completely and accurately reflect the characteristics of the measured object, providing high-quality visual information for subsequent fusion analysis with tactile layer data, and contributing to a more comprehensive and in-depth understanding and measurement of the measured object.

[0097] The pressure sensing array is arranged in a grid pattern, with a capacitor array deployed in the central area and ultrasonic units deployed in the outer area;

[0098] The capacitor array establishes a mapping relationship between capacitance changes and the surface topography of the object:

[0099] ;

[0100] in, This is the change in capacitance, reflecting the degree of electric field distortion caused by the approach of an object; is the vacuum permittivity, which characterizes the ability of an electric field to propagate in a vacuum; ρ is the relative permittivity, describing the dielectric properties of a material relative to vacuum. Let be the distance distribution function from the object surface to the sensor, which varies with position (x,y); Let be the integral area element, representing the differential region covered by the sensor array;

[0101] The ultrasonic unit is used to improve resolution:

[0102] ;

[0103] in, The image intensity represents the superposition result of the echo signals at position (x, y); The number of sensors represents the total number of ultrasonic units involved in imaging. For the echo signal of the i-th sensor, The speed of sound is used to enhance imaging accuracy through superposition; t is the echo time, representing the time delay of the sound wave from transmission to reception; (x i , y i Let be the coordinates of the i-th sensor, defining its spatial position in the array.

[0104] Specifically, the pressure sensor array employs a grid-interlaced layout, which increases sensor density and uniformity. Compared to a regular grid layout, the interlaced layout reduces interference between sensors, improves the accuracy of pressure distribution sensing, and allows the sensor array to more finely capture pressure changes applied to the object surface, providing richer data for subsequent pressure distribution analysis and 3D reconstruction. A capacitor array is deployed in the central region of the array. Capacitive sensors are highly sensitive to changes in the electric field caused by the approach of an object. When an object approaches the sensor surface, it alters the electric field distribution around the sensor, causing a change in capacitance. The central region is typically the most concentrated and critical area for pressure sensing by the sensor array; deploying a capacitor array fully utilizes its sensitivity to changes in the electric field to accurately capture the surface morphology information of the object in this region. Ultrasonic units are deployed in the outer region. Ultrasonic sensors acquire information by emitting and receiving ultrasonic waves, possessing good penetration and spatial resolution. Building upon the preliminary morphological information obtained from the capacitor array in the central region, the ultrasonic units in the outer region can further extend the measurement range and supplement measurements in areas that the capacitor array struggles to accurately detect, while simultaneously leveraging the characteristics of ultrasound to improve the overall system resolution. In the formula, ΔC represents the change in capacitance, which directly reflects the degree of electric field distortion caused by the approach of an object. When the surface of an object approaches the sensor, it changes the electric field distribution between the sensor plates, causing a change in capacitance. The magnitude of ΔC is closely related to the degree of this electric field distortion. ν is the vacuum permittivity, a physical constant that characterizes the ability of an electric field to propagate in a vacuum and is one of the fundamental parameters for calculating changes in capacitance. ρ is the relative permittivity, describing the dielectric properties of a material relative to vacuum. Different materials have different relative permittivity, which affects the distribution of the electric field between the object and the sensor, thus influencing the capacitance change. d(x,y) is the distance distribution function from the object surface to the sensor, varying with position (x,y). It reflects the distance differences between different positions on the object surface and the sensor, and is one of the key factors affecting the capacitance change. dA is the integral area element, representing the differential region covered by the sensor array. By integrating over the entire sensor array coverage area, the total capacitance change caused by changes in the object surface topography can be calculated. By establishing this mapping relationship, the capacitance change data measured by the capacitance array can be converted into topographic information of the object surface. In practical applications, by measuring the capacitance change at different positions and combining it with known parameters and formulas, the height or topographic features of the object surface at various positions can be deduced, providing important data for 3D reconstruction. Meanwhile, I(x,y) in the formula represents the imaging intensity, indicating the superposition result of the echo signals at position (x,y). N represents the number of sensors, i.e., the total number of ultrasonic units involved in imaging. More sensors mean more echo signal information can be acquired. i Let be the echo signal from the i-th sensor. Each sensor emits ultrasonic waves and receives echo signals reflected from the object's surface. These echo signals contain information such as the object's surface position and reflection characteristics. Let v be the speed of sound. In a given medium, the speed of sound is a relatively stable value. By measuring the echo time t (the time delay from emission to reception of the sound wave) and combining it with the speed of sound, the distance between the object's surface and the sensor can be calculated. (x) i , y i Let be the coordinate position of the i-th sensor, defining its spatial position in the array. By superimposing and spatially correcting the echo signals from multiple sensors, the position of the object's surface in space can be accurately determined. The ultrasonic unit emits ultrasonic waves and receives echo signals. The echo signals received by sensors at different positions exhibit time delay differences, reflecting the distance between different positions on the object's surface and the sensors. By superimposing the echo signals from multiple sensors, the intensity of the imaging signal can be enhanced, noise interference reduced, and thus the accuracy and resolution of the imaging improved. Simultaneously, utilizing the spatial position information of the sensors, the coordinates of the object's surface in space can be accurately determined, further refining the perception of the object's surface morphology and achieving more precise three-dimensional imaging of the object's surface.

[0105] It is worth noting that the capacitive array and ultrasonic unit complement each other functionally. The capacitive array is sensitive to near-field morphological changes on an object's surface, quickly capturing minute surface undulations; while the ultrasonic unit has better penetration and spatial resolution, enabling the measurement of object surface information at greater distances and compensating for the limitations of the capacitive array in measurement range and depth perception. The combination of the two provides more comprehensive and accurate information about the object's surface morphology. By establishing a mapping relationship between capacitance changes and object surface morphology, and utilizing the resolution enhancement mechanism of the ultrasonic unit, this pressure sensing array can more accurately acquire three-dimensional information about the object's surface. In fields such as 3D scanning, industrial inspection, and medical imaging, high-precision surface morphology measurement is crucial for product quality control, disease diagnosis, and scientific research. This sensor array design, combining a capacitive array and an ultrasonic unit, allows it to adapt to more complex measurement scenarios. For example, when measuring objects with diverse surface materials, irregular shapes, or obstructions, the capacitive array and ultrasonic unit can each leverage their strengths to jointly achieve accurate surface measurement, improving the system's adaptability and reliability.

[0106] The data fusion processor includes:

[0107] The data calibration module is used to calibrate the sensor and establish the mapping relationship between the sensor feature vector and the hypergraph structure.

[0108] A sensor node set is defined for different sensors, wherein the sensor node set V = {v1,...,v...} n} Corresponds to n types of sensor features;

[0109] If the system contains n types of sensor features, then:

[0110] Node set:

[0111] ;

[0112] In the formula, The eigenvector represents the feature vector of the i-th type of sensor; n represents the total number of sensor types (n≥1); d i Represents the feature dimension of the i-th type of sensor; express 3D real space;

[0113] Hyperedge set:

[0114] ;

[0115] In the formula, Indicates the j-th superedge; Represents the set of positive integers; j≥1;

[0116] The data preprocessing module is used to achieve spatiotemporal alignment and benchmark unification of multi-sensor data;

[0117] S1. Unify the data space through timestamp synchronization and coordinate transformation matrix:

[0118] ;

[0119] In the formula, Indicates alignment time; This indicates the search for a time transformation T that minimizes the difference between the measurements from the two sensors; ∑ is the summation symbol, representing the accumulation of differences over all time points; and These represent the measurement values ​​of sensor i and sensor j at time t, respectively. This represents the result of time transformation T of the measurement value of sensor j at time t; This represents the square of the L2 norm, used to quantify the difference between the measurement value of sensor i and the measurement value of sensor j after time transformation T;

[0120] S2. Extract visual and tactile features;

[0121] S3. Output the standardized feature vector set;

[0122] Hypergraph neural networks are used to construct multi-view hypergraphs to achieve cross-modal feature fusion.

[0123] The dynamic reconstruction module is used to identify key feature subsets related to the shape and material characteristics of the object under test and reconstruct the multi-view hypergraph structure.

[0124] The risk assessment module is used to generate the final risk decision.

[0125] Specifically, the core task of the data calibration module is to calibrate multimodal sensors and establish a mapping relationship between sensor feature vectors and the hypergraph structure. This is the foundation for subsequent efficient data fusion and analysis. Through this mapping, raw data from different types of sensors can be transformed into a form that can be processed and utilized within the hypergraph structure, facilitating the discovery of intrinsic relationships between data. A sensor node set V={v1,...,v...} is defined for different sensors. n}, each node v i This corresponds to a type of sensor feature. In the node set, the feature vector of the i-th type of sensor is represented as... Where n is the total number of sensor types (n≥1), d i This represents the feature dimension of the i-th type of sensor. express A 3D real-valued space clearly describes the feature information carried by each sensor node, providing a data foundation for subsequent construction of the hypergraph structure. By mapping different sensors to nodes in the hypergraph and using sensor features as node attributes, the relationships between different sensors and the information they provide can be intuitively represented. This node- and feature vector-based representation method facilitates unified processing and analysis of multi-sensor data within the hypergraph structure. The hyperedge set is used to describe the complex relationships between sensor nodes. Each hyperedge e... j (j≥1) can connect multiple nodes. This connection method differs from the limitation of traditional graph edges that can only connect two nodes, and can more flexibly represent the interaction and dependency between multi-sensor data. For example, in some application scenarios, the data from multiple sensors may be simultaneously affected by a specific feature of the measured object. Hyperedges can connect these related sensor nodes to form an associated subgraph, which facilitates subsequent feature fusion and data analysis.

[0126] The data preprocessing module achieves spatiotemporal alignment and benchmark unification of multi-sensor data, providing high-quality data for subsequent feature extraction and fusion. Due to differences in the operating principles, sampling frequencies, and installation locations of different sensors, the data they acquire may exhibit inconsistencies in time and space, thus requiring preprocessing. The goal is to find a time transformation T that minimizes the difference between the measurements from two sensors. During processing... and Let represent the measurement values ​​of sensor i and sensor j at time t, respectively. This represents the result of time transformation T after applying a time transformation T to the measurement value of sensor j at time t. The square of the L2 norm is used to quantify the difference between two measurements. By summing the differences at all time points and finding the minimum value, the optimal time transformation T can be found, achieving temporal alignment between the two sensors. Besides time alignment, it is also necessary to unify the data spaces of different sensors. This is usually achieved through a coordinate transformation matrix, converting the coordinate systems of different sensors to a unified coordinate system, making their acquired data spatially comparable. Spatiotemporal alignment and benchmark unification are key steps in multi-sensor data fusion, avoiding analytical errors caused by data inconsistencies. After completing spatiotemporal alignment and benchmark unification, feature extraction is performed on the data from the visual layer (central camera and ring camera) and the tactile layer (pressure sensor array). Visual features may include information such as the shape, texture, and color of objects, while tactile features may include mechanical properties such as pressure distribution and surface hardness. Feature extraction utilizes existing techniques for feature stitching and fusion, which is feasible. By extracting these features, the raw sensor data can be transformed into more representative and analyzable feature vectors. The extracted feature vectors are then standardized to have the same scale and distribution characteristics. The standardized feature vector set can be used as input for subsequent modules such as hypergraph neural networks, facilitating feature fusion and analysis.

[0127] Hypergraph neural networks are used to construct multi-view hypergraphs to achieve cross-modal feature fusion. In multimodal sensor systems, sensor data from different modalities provide different perspectives about the measured object. Hypergraph neural networks integrate feature information from different modalities by constructing a hypergraph structure and leverage the powerful learning capabilities of neural networks to uncover potential correlations and complex patterns between data, achieving effective cross-modal feature fusion. Hypergraph neural networks perform information propagation and feature updates on the constructed hypergraph structure. Each node interacts with other nodes through its connected hyperedges, which can be seen as channels for information transmission between nodes. During information propagation, the neural network updates and fuses the node features based on the node's features and the weights of the hyperedges. Through multiple rounds of information propagation and feature updates, hypergraph neural networks can learn the intrinsic connections between features of different modalities, generating more discriminative fused features, providing strong support for subsequent tasks such as risk assessment.

[0128] The main task of the dynamic reconstruction module is to identify a subset of key features related to the shape and material characteristics of the object under test, and to reconstruct a multi-view hypergraph structure based on these key features. In practical applications, different features of the object under test may have different importance for different analysis tasks. Through the dynamic reconstruction module, the most critical feature subset for the current task can be selected based on specific application requirements and the characteristics of the object under test, and the hypergraph structure can be adjusted to focus more on these key features, improving the efficiency and accuracy of data analysis. This module can employ feature selection algorithms, such as statistical methods and machine learning methods, to evaluate the importance of each feature to shape and material characteristics. Based on the evaluation results, features with higher importance are selected to form a key feature subset. Then, the node and hyperedge connections in the hypergraph are adjusted based on the key feature subset to reconstruct the multi-view hypergraph structure. For example, if certain tactile features are found to be crucial for material recognition, while certain visual features are more important for shape measurement, then the connections between the nodes corresponding to these key features can be strengthened to construct a hypergraph structure that better meets the needs of the current task. Risk assessment module overview: The risk assessment module generates the final risk decision based on the data processed and analyzed by the previous modules. In many applications, such as industrial inspection and medical diagnosis, it is necessary to conduct risk assessments on the state of the object being tested and make corresponding decisions based on the assessment results. The risk assessment module comprehensively utilizes information fused from multiple sensors, combined with pre-set risk assessment models and rules, to quantitatively assess the risk level of the object being tested and output the final risk decision result. Assessment basis: This module may consider the morphological characteristics of the object being tested (such as the presence of defects, deformation, etc.), material characteristics (such as hardness, elasticity, etc.), and other relevant information. Based on the correlation between these characteristics and risk, it calculates risk indicators. For example, in industrial product quality inspection, if serious defects (morphological characteristics) are detected on the product surface and the material hardness does not meet requirements (material characteristics), the risk assessment module may determine that the product has a high quality risk and output corresponding risk decisions, such as rejecting the product or recommending further testing.

[0129] The implementation of the hypergraph neural network includes:

[0130] S4. Hypergraph construction: Each sensor feature vector is used as an independent node, and hyperedges are built based on physical constraints.

[0131] Node feature matrix: X∈R {n×d} ;

[0132] Hyperedge incidence matrix: H∈{0,1} {n×m} H[i,j]=1 indicates that node v i Belongs to hyperedge e j ;

[0133] In the formula, d is the unified feature dimension; n represents the total number of nodes; and m represents the total number of hyperedges.

[0134] K-means clustering is used to discover implicit associations, and the dimensions of hyperedges are adaptively adjusted.

[0135] S5, Hypergraph Convolution:

[0136] ;

[0137] in, Indicates the first The node feature matrix of the layer; H represents the hyperedge association matrix; Represents a diagonal matrix of node degrees; Denotes the hyperedge degree diagonal matrix; W denotes the hyperedge weight matrix; Indicates the first The trainable parameter matrix of the layer; Represents a nonlinear activation function;

[0138] S6, Multi-level fusion:

[0139] Multiple hypergraph convolutional layers are stacked to gradually fuse local and global features.

[0140] Specifically, a Hypergraph Neural Network (HGNN) is a graph neural network capable of handling complex high-order relationships, particularly suitable for feature fusion of multimodal sensor data. In this model, each sensor feature vector is treated as an independent node, and the node feature matrix is ​​represented as X∈R. {n×d} Where n is the total number of nodes and d is the uniform feature dimension, this definition ensures that each sensor node carries its unique feature information, providing a foundation for subsequent hypergraph construction and feature fusion. The hyperedge association matrix describes the relationship between nodes and hyperedges. In this way, hyperedges can connect multiple nodes, thus representing complex relationships between multi-sensor data. The k-means clustering algorithm is used to cluster node features, discovering potential implicit associations. The clustering results can be used to construct hyperedges, so that each hyperedge corresponds to a cluster, thereby connecting nodes with similar features. The dimension of the hyperedge is dynamically adjusted according to the clustering results and actual needs. For example, if a cluster has a large number of nodes or significant feature differences, the dimension of the hyperedge can be appropriately increased to more precisely describe the relationships between nodes. Hypergraph convolution uses the hyperedge association matrix H and the node degree diagonal matrix... Hyperdiagonal matrix The weight matrix W is used to propagate and update node features. In the formula, Let H represent the node feature matrix of layer l. H is the hyperedge correlation matrix, used to map node features to the hyperedge space. W is the hyperedge weight matrix, used to adjust the importance of different hyperedges. Map the features of the hyperspace back to the node space. It is the trainable parameter matrix of the l-th layer, used to learn the transformation of node features. It is a non-linear activation function used to introduce non-linearity and enhance the expressive power of the model. (Node degree diagonal matrix) This describes the number of hyperedges connected to each node, used to normalize node features. Hyperedge degree diagonal matrix. This describes the number of nodes connected by each hyperedge, used to normalize hyperedge features. In practice, H is typically normalized to balance the influence of nodes and hyperedges. By stacking multiple hypergraph convolutional layers, local and global features are gradually fused. Each convolutional operation fuses the node features of the current layer with those of its neighboring nodes, capturing more complex feature patterns. During feature fusion, in shallow networks, hypergraph convolution primarily focuses on the local feature interactions between nodes and directly connected hyperedges. In deep networks, as the number of layers increases, node features are propagated and fused multiple times through hyperedges, thus capturing global feature patterns. Multi-level fusion effectively combines feature information from different levels, preserving local details while capturing global structural information, thereby improving the model's performance in complex tasks.

[0141] Therefore, hypergraph neural networks, by constructing hypergraph structures, implementing hypergraph convolution operations, and fusing multi-level features, can effectively handle the complex relationships in multimodal sensor data. Their unique hyperedge representation and adaptive adjustment mechanism enable the model to flexibly capture the interactions and dependencies between different sensor data.

[0142] The implementation of the dynamic reconfiguration module includes:

[0143] S7. Key Feature Recognition:

[0144] Based on the gradient risk assessment feature region, the node importance score S is output. i ∈[0,1];

[0145] Specifically, gradient risk assessment technology identifies feature regions that significantly impact risk assessment and generates an importance score Si ∈ [0,1] for each node. The influence of each feature on the final risk assessment result is evaluated by calculating the gradient of the model output relative to the input features. The magnitude of the gradient reflects the importance of the feature; a larger gradient indicates a more significant impact on the output. Based on the gradient information, an importance score Si is calculated for each node (i.e., each sensor feature vector). The importance score is typically obtained by normalizing the gradient value, ranging from 0 to 1, where 1 indicates the node is crucial to risk assessment, and 0 indicates the node has almost no impact.

[0146] S8, Hypergraph Reconstruction:

[0147] Delete S i <τ nodes or superedges, where τ is a preset threshold;

[0148] Dynamically adjust the coverage of the superedge based on feature similarity;

[0149] A continuous learning framework is used to incrementally update the hypergraph parameters.

[0150] Specifically, based on the key feature identification results, the hypergraph structure is adjusted by deleting unimportant nodes or hyperedges and dynamically adjusting the coverage of hyperedges to improve the model's efficiency and accuracy. In the operation, an importance threshold τ is set; nodes with an importance score Si < τ, or their associated hyperedges, are removed from the hypergraph. This step effectively reduces computation and removes the influence of noisy data on the model. For the retained nodes and hyperedges, the connection range of the hyperedges is dynamically adjusted based on their feature similarity. Feature similarity can be measured by calculating the cosine similarity or Euclidean distance between node feature vectors; nodes with high similarity can have their connected hyperedges increased to better capture the correlation between features. A continuous learning framework is adopted to incrementally update the hypergraph parameters. Continuous learning allows the model to retain previously learned knowledge while continuously receiving new data, avoiding catastrophic forgetting. Through incremental updates, the model can adapt to changes in the environment and the arrival of new data, maintaining the stability and accuracy of its performance.

[0151] The implementation of the risk assessment module includes:

[0152] S9. Input the updated hypergraph node features X';

[0153] S10, a fully connected layer and Softmax, outputs the risk probability p∈[0,1];

[0154] S11. If p > θ, it is judged as high risk, where θ is determined by the ROC curve.

[0155] Specifically, the risk assessment module receives updated hypergraph node features X' from the dynamic reconstruction module. These features have undergone multi-level feature fusion and dynamic reconstruction, containing the shape, material, and other relevant feature information of the object under test, comprehensively reflecting the state of the object. Before being input into the risk assessment module, the hypergraph node features X' are typically standardized or normalized to ensure consistent feature scale and avoid assessment bias caused by differences in feature scale. Through fully connected layers and the Softmax function, the updated hypergraph node features are mapped to the risk probability space, outputting the probability p that the object under test belongs to the high-risk category. The fully connected layer converts the input features X' into a high-dimensional representation, capturing the complex interactions between features. The number of neurons in the fully connected layer can be adjusted according to the specific task, typically determined experimentally. The Softmax function is applied to the output layer to convert the output of the fully connected layer into a probability distribution.

[0156] The formula for the Softmax function is: ;

[0157] in, It is the first output of the fully connected layer. The value of each neuron; This represents the corresponding risk probability; e is the natural constant; j is the index variable for traversing all categories; the probability is output through the Softmax function. Satisfy 0≤ ≤1 and , which represents the probability that the tested object belongs to different risk categories.

[0158] Based on the output risk probability and a preset threshold θ, it is determined whether the tested object belongs to a high-risk category. The threshold θ is determined using a Receiver Operating Characteristic Curve (ROC). The ROC curve helps select the optimal threshold by plotting the true positive rate (TPR) and false positive rate (FPR) at different thresholds, achieving an optimal balance between sensitivity and specificity. If the risk probability p > θ, the tested object is classified as a high-risk category; otherwise, it is classified as a low-risk category. This determination result can be directly used in subsequent decision support systems to trigger corresponding alerts or interventions.

[0159] This application also includes cross-control logic:

[0160] The dynamic coding emission module adjusts the polarization coding parameters in real time based on the mechanical distribution data of the tactile layer.

[0161] The three-coordinate bridge measurement structure dynamically plans the obstacle avoidance scanning path based on the output of the risk assessment module.

[0162] When the security protection unit detects an abnormal optical power, it synchronously triggers the self-test protocol of the data fusion processor.

[0163] Specifically, this application forms a complete perception-decision-execution chain through tactile data, coding adjustments, measurement path optimization, and safety monitoring.

[0164] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A mobile bridge-type multi-sensor 3D scanning system, characterized in that, include: Measurement platform (1), used to place the object to be measured; The three-coordinate bridge measurement structure is installed on the measurement platform (1) for adjusting the measurement orientation. The three-coordinate bridge measurement structure is also equipped with a detection head (2). The detection head (2) integrates a dynamic encoding transmission module and a multi-modal sensing module to realize the all-round measurement of the object under test. A dynamic encoding transmission module, integrated into the detection head, is configured to dynamically adjust the signal transmission direction or frequency; the polarization encoding unit of the dynamic encoding transmission module generates annular polarized light containing topological phase; A multimodal sensing module, integrated into the detection head, includes a visual layer and a tactile layer. The visual layer captures a dynamic image sequence of the surface of the object under test through a multi-camera array. The tactile layer generates a non-contact force distribution matrix through a pressure sensor array and receives light reflected from the object under test according to the non-contact force distribution matrix. Data received by different types of sensors is used to capture data of the object under test in all dimensions. The data fusion processor incorporates a hypergraph neural network model to achieve feature fusion of multi-source heterogeneous data received from different types of sensors. Based on the data of the object under test, the hypergraph neural network model constructs a multi-view view of the object's shape and material. Nodes represent features from different types of sensors, and hyperedges represent cross-modal correlations, forming a multi-view hypergraph. Simultaneously, it identifies a subset of key features related to the object's shape and material characteristics, reconstructs the multi-view hypergraph structure using these key feature subsets, and updates the hypergraph neural network. Based on a preset risk assessment threshold for the object's shape and material characteristics, the updated hypergraph neural network determines the risk decision result for the object. The dynamic coding transmission module includes: Wavelength adaptive unit automatically selects the illumination wavelength based on the surface material of the object being measured; Polarization coding unit modulates the topological phase distribution of ring-polarized light; The safety protection unit monitors optical power in real time and triggers an emergency shutdown. The frequency range of the transmitted signal is dynamically limited by the coding sequence, and invalid coding sequences are automatically filtered out. The pressure sensor array is arranged in a grid pattern, with a capacitor array deployed in the central area and ultrasonic units deployed in the outer area; The hypergraph neural network performs the following operations: Construct a node feature matrix and a hyperedge correlation matrix, where the nodes represent different sensor features; Node features are updated iteratively through hypergraph convolutional layers, fusing local and global relationships; A continuous learning framework is used to dynamically optimize the hyperedge weights and redundant nodes are removed based on feature similarity. The system also includes cross-control logic: The dynamic coding emission module adjusts the polarization coding parameters in real time based on the mechanical distribution data of the tactile layer. The three-coordinate bridge measurement structure dynamically plans the obstacle avoidance scanning path based on the output of the risk assessment module. When the security protection unit detects an abnormal optical power, it synchronously triggers the self-test protocol of the data fusion processor.

2. The mobile bridge-type multi-sensor 3D scanning system according to claim 1, characterized in that, The coordinate measuring bridge structure includes: An X-axis motion mechanism (3) is provided on the upper surface of the measuring platform (1). The object to be measured is located above the X-axis motion mechanism (3). The X-axis motion mechanism (3) guides the object to be measured to move along the X-axis direction of the measuring platform (1). A cable tray (4) is horizontally mounted above the measuring platform (1) and fixed on both sides of the measuring platform (1). A Z-axis motion mechanism (5) is provided on the cable tray (4). The Z-axis motion mechanism (5) moves horizontally along the cable tray (4) to realize the Y-axis movement of the detection head (2). The detection head (2) is mounted on the Z-axis motion mechanism (5) to realize the Z-axis movement of the detection head (2). The X-axis motion mechanism (3) is located below the bridge frame (4).

3. The mobile bridge-type multi-sensor 3D scanning system according to claim 1, characterized in that, The multimodal sensing module is implemented through a sensor array; The sensor array includes a visual layer and a tactile layer; The visual layer includes: The central camera defines the first search area with the image center as the origin, and its radius adaptively expands with the target's movement speed. Multiple ring cameras are arranged circumferentially around the central camera, and the fan-shaped imaging area of ​​each camera covers the blind area of ​​the first search area; The visual layer uses an edge prediction algorithm to estimate the position of a high-speed moving target and aligns it with the data from the tactile layer in time and space. The tactile layer includes: A capacitive sensing array, deployed in the central region, maps the surface morphology of an object through changes in capacitance. An ultrasonic sensing unit is deployed in the peripheral area to improve resolution through echo superposition. The data from the capacitive sensing array and the ultrasonic sensing unit are fused using a hypergraph neural network.

4. The mobile bridge-type multi-sensor 3D scanning system according to claim 1, characterized in that, The data fusion processor includes: The data calibration module establishes the mapping relationship between sensor feature vectors and hypergraph nodes; The dynamic reconstruction module filters key feature subsets and reconstructs the hypergraph structure based on gradient risk assessment. The risk assessment module outputs risk probabilities through a fully connected layer and a Softmax function; The set of hyperedges in the hypergraph neural network is adaptively adjusted through k-means clustering.

Images