An automatic mapping system for on-site layout of occupational health equipment

Abstract

The present application provides a kind of professional health field equipment layout automatic mapping system, including mobile bearing terminal and laser radar scanning unit on it, and artificial intelligence processing module and automatic mapping module including point cloud preprocessing unit and identification classification unit, and identification classification unit includes shared feature extraction subunit and parallel equipment type classification branch and occupational hazard attribute classification branch;The clustering subunit of point cloud preprocessing unit and the identification classification unit constitute closed loop: the clustering subunit is with first merging threshold value to point cloud clustering and obtains initial target cluster, and then the reference size corresponding to the label output by equipment type classification branch is used as second merging threshold value, and the distance less than the threshold value is merged and returned for re-judgment;Automatic mapping module is mapped and annotated occupational disease hazard factor legend according to the two types of labels after re-judgment;Through mobile bearing terminal carrying laser radar scanning, cooperate with artificial intelligence processing and automatic mapping, reduce the time cost of field investigation.

Description

Technical Field

[0001] This invention relates to the field of health technology, and in particular to an automatic mapping system for the layout of occupational health field equipment. Background Technology

[0002] In occupational health field surveys, it is necessary to accurately map the distribution and layout of various production equipment, pipelines, operating tables and other facilities in the enterprise's production workshop and work area, and draw standard equipment layout diagrams to provide basic drawing data for occupational hazard factor detection, risk assessment and health evaluation.

[0003] Currently, the mainstream surveying and mapping methods in the industry are divided into two categories: 1. Manual on-site surveying: Workers carry tools such as tape measures, rangefinders, and paper and pens to the work site to manually measure and record the location, dimensions, spacing between equipment, and area boundaries of each piece of equipment. After returning to the office, they manually reproduce and draw the equipment layout diagram using CAD and drafting software. The entire process relies on manual operation, starting with on-site data measurement and then manual drawing.

[0004] 2. Traditional ordinary camera surveying: In some scenarios, ordinary cameras are used to capture on-site images. The layout is drawn by combining two-dimensional images with manual annotation. It can only obtain planar visual information and cannot accurately collect three-dimensional spatial dimensions and the three-dimensional position of equipment.

[0005] However, manual point-to-point measurement and hand-drawn maps are cumbersome. In large factories and multi-workshop settings, a single on-site survey and map production often takes several days, which cannot meet the needs of rapid on-site mapping and batch surveys in occupational health. Furthermore, manual handheld measurement tools have low accuracy. Ordinary cameras can only capture two-dimensional images and cannot reproduce the three-dimensional shape, spatial height, and three-dimensional spacing of equipment. The layout maps drawn are mostly pure planar diagrams, which are difficult to support in-depth analysis of occupational hazard diffusion and protective facility layout in three-dimensional space. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An automatic mapping system for occupational health field equipment layout includes a mobile carrier terminal and a lidar scanning unit on it, as well as an artificial intelligence processing module and an automatic drawing module containing a point cloud preprocessing unit and an identification and classification unit. The identification and classification unit includes a shared feature extraction subunit and parallel equipment type classification branches and occupational hazard attribute classification branches. The clustering subunit of the point cloud preprocessing unit and the identification and classification unit form a closed loop: the clustering subunit clusters the point cloud to obtain an initial target cluster using a first merging threshold, and then uses the reference size corresponding to the label output by the equipment type classification branch as a second merging threshold, merging those with a spacing smaller than the threshold and sending them back for re-judgment; the automatic drawing module generates a map based on the two types of labels after re-judgment and labels occupational hazard factor legends.

[0007] Furthermore, the shared feature extraction subunit outputs a point-by-point local feature set and a global feature vector to the input target cluster; the equipment type classification branch takes the global feature vector as input, and outputs the probability distribution of each equipment category through a fully connected layer and a normalized exponential function, taking the one with the highest probability as the category label of the target cluster; the occupational hazard attribute classification branch takes the feature obtained by concatenating the point-by-point local feature set with the global feature vector as input, and outputs a judgment value of 0 to 1 for each type of occupational hazard factor after passing through a shared multilayer perceptron and a pooling layer, and each judgment value is independently compared with a preset threshold to form a multi-label hazard attribute label.

[0008] Furthermore, the shared feature extraction subunit samples a preset number of representative points from the target cluster using the farthest point sampling method, supplementing the insufficient number of points with repeated sampling; each sampling point forms a four-dimensional input with three-dimensional coordinates and laser reflection intensity, which is then processed by a first set of shared weight multilayer perceptrons to obtain the point-by-point local feature set, and then processed by a second set of shared weight multilayer perceptrons and max pooling to obtain the global feature vector with permutation invariance.

[0009] Furthermore, the artificial intelligence processing module also includes a device voxel model library, which stores the reference length, reference width, reference height, basic geometric topology, and main axis direction under standard posture according to device category. In the first stage, the clustering subunit performs Euclidean distance clustering using a first merging threshold and calculates the center coordinates, length, width, height, and main axis direction of the minimum bounding box of each initial target cluster. In the second stage, the reference size corresponding to the category label of the initial target cluster is retrieved from the device voxel model library, and the product of the reference size and a preset scaling factor is taken as the second merging threshold. Two initial target clusters whose bounding box center distance is less than the second merging threshold and whose bounding box main axis direction angle is less than a preset angle are determined to be the same target and merged. After merging, the center coordinates, length, width, height, and main axis direction of the bounding box are recalculated and sent back to the identification and classification unit for re-judgment.

[0010] Furthermore, for slender target clusters with bounding box aspect ratios greater than a preset multiple, the second merging threshold of the clustering sub-unit along its bounding box principal axis is taken as an integer multiple of the product of the reference size and the preset scaling factor, while the second merging threshold perpendicular to the principal axis remains unchanged; for non-linear continuous target clusters, the first principal component direction obtained from the principal component analysis of its point cloud is taken as the principal axis direction, and the aspect ratio is calculated as the ratio of the projected length in that direction to the equivalent diameter perpendicular to that direction.

[0011] Furthermore, when the category label obtained by the identification and classification unit is inconsistent with that before the re-judgment, the clustering subunit redetermines the second merging threshold based on the reference size corresponding to the new category label and merges again. If the number of iterations reaches the preset upper limit and the category labels are still inconsistent, the target cluster is marked as pending manual confirmation. When the maximum probability value output by the equipment type classification branch is lower than the first confidence threshold, or when all the judgment values ​​of the occupational hazard attribute classification branch are higher than the second confidence threshold and the differences between them are all less than the preset difference, the target cluster is also marked as pending manual confirmation. The marked target cluster is highlighted in the drawing generated by the automatic drawing module with an outline line type that is different from other targets, and a list of items to be confirmed is listed at the boundary of the drawing.

[0012] Furthermore, the point cloud preprocessing unit also includes a filtering subunit and a ground segmentation subunit that are sequentially connected to the clustering subunit. The filtering subunit first downsamples the global point cloud using a voxel grid, then calculates the average distance of all points in its neighborhood for each point, and identifies and removes points whose average distance exceeds a preset multiple of the global average distance. The ground segmentation subunit fits the maximum horizontal plane in the filtered point cloud using a random sampling consistency method, identifies points belonging to the plane and its neighborhood as ground surfaces, and removes the remaining points as target candidate points.

[0013] Furthermore, the mobile carrier terminal is also equipped with an inertial attitude sensor and a positioning module. The positioning module uses the angular velocity and linear acceleration output by the inertial attitude sensor to perform a short-term recursive calculation of the pose change between two adjacent scan frames, and then corrects the recursive value with the registration result of the point cloud of the adjacent frames. The artificial intelligence processing module also includes a real-time registration unit connected to the point cloud preprocessing unit. The real-time registration unit maintains a sliding window containing the scan data of the most recent few frames and the corresponding pose estimation values. The newly entered frame point cloud is transformed to the site coordinate system according to the pose estimation value of the positioning module. After voxel downsampling, corner points and planar points are extracted. The normal distribution transformation algorithm is used to register with the existing frames in the window and calculate the registration residual. The residual is used as the observation value to jointly optimize the pose of all frames in the window through factor graph optimization, and then the optimized pose is reprojected and incorporated into the global point cloud.

[0014] Furthermore, the automatic drawing module includes a 3D model reconstruction unit, a 2D projection mapping unit, a hazard factor identification and labeling unit, a mapping specification storage unit, and a format conversion and output unit. The 3D model reconstruction unit retrieves the corresponding parametric 3D volume from the device voxel model library based on the target cluster's category label. It independently sets the dimensions of the volume's length, width, and height on the three orthogonal axes using the measured bounding box, rotates its principal axis to the measured principal axis direction, and then translates it to the measured center coordinates, ensuring the reconstructed volume's outer contour fits the measured bounding box. The 2D projection mapping unit takes the dimensions of each reconstructed volume on the horizontal plane... The orthographic projection outline is drawn as a base map according to a selected scale, and the equipment spacing is marked by the nearest point distance between the outer contours of any two target bounding boxes; the hazard factor identification and labeling unit requests the corresponding legend from the mapping specification storage unit according to the hazard attribute label and places it on one side of the corresponding equipment outline; the mapping specification storage unit pre-stores occupational hazard factor legend set, occupational disease prevention facility legend set, and layer division, line type, text height and frame format parameters; the format conversion and output unit exports the base map by layer into a graphic exchange format or image format file, and exports the reconstructed spatial model into a three-dimensional model format file.

[0015] Furthermore, the mobile carrier terminal includes a wheeled chassis, a walking control unit disposed within the wheeled chassis, a lifting column vertically fixed to the upper surface of the wheeled chassis, a pitch adjustment gimbal disposed at the top of the lifting column, and a radar mounting base fixed to the rotating part of the pitch adjustment gimbal. The laser radar scanning unit is fixed to the upper surface of the radar mounting base. An inertial attitude sensor is installed in the inner cavity of the radar mounting base, and its mounting base surface and the mounting base surface of the laser radar scanning unit are the upper and lower surfaces of the same component. The outer cover of the laser radar scanning unit is provided with a protective cover with an open lower end. The lower edge of the protective cover is sealed to the edge of the radar mounting base. A light-transmitting window is opened around the laser transceiver surface of the protective cover. Several purge nozzles are evenly distributed circumferentially on the outer side of the light-transmitting window. The air inlet of each purge nozzle is connected to an air source fixed in the inner cavity of the wheeled chassis through an air pipe arranged along the inner cavity of the lifting column, and continuously sprays airflow onto the window surface of the light-transmitting window to form an adhering air curtain.

[0016] Compared with the prior art, the beneficial effects of the present invention include: By using a mobile terminal equipped with lidar for automatic walking scanning, combined with real-time point cloud registration, identification and classification, and automatic mapping by the back-end artificial intelligence processing module, the process that would have taken several days for traditional manual point-by-point measurement and manual mapping can be completed immediately or within a few hours after on-site scanning, thus reducing the time cost of occupational health field investigations. Attached Figure Description

[0017] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein: Figure 1 This is a structural diagram of a mobile bearer terminal.

[0018] Figure 2 for Figure 1 Enlarged view of section A.

[0019] Figure 3 This is a side view of the mobile carrier terminal.

[0020] Figure 4 for Figure 3 Enlarged view of section B in the middle.

[0021] Figure 5 This is a diagram of the front-end data acquisition layer architecture.

[0022] Figure 6 Human-processing layer architecture diagram.

[0023] Figure 7 This is a diagram of the data transmission module architecture.

[0024] Figure 8 It is the mobile control architecture layer for mobile bearer terminals.

[0025] Numbering on the map: 1. Mobile carrier terminal; 11. Wheeled chassis; 111. Drive wheel; 112. Caster wheel; 12. Walking control unit; 13. Lifting column; 14. Radar mounting base; 15. Pitch adjustment gimbal; 16. Obstacle avoidance sensor; 17. Remote control receiver unit; 18. Power module; 19. Air source; 2. LiDAR scanning unit; 21. Protective cover; 22. Light transmission window; 23. Blowing nozzle; 3. Inertial attitude sensor; 4. Positioning module; 5. Data transmission module; 51. Wireless communication unit; 52. Wired data interface; 6. Local storage module 7. Artificial Intelligence Processing Module; 71. Real-time Registration Unit; 72. Point Cloud Preprocessing Unit; 721. Filtering Subunit; 722. Ground Segmentation Subunit; 723. Clustering Subunit; 73. Identification and Classification Unit; 731. Shared Feature Extraction Subunit; 732. Equipment Type Classification Branch; 733. Occupational Hazard Attribute Classification Branch; 8. Automatic Drawing Module; 81. 3D Model Reconstruction Unit; 82. 2D Projection Drawing Unit; 83. Hazard Factor Identification and Labeling Unit; 84. Drawing Standard Storage Unit; 85. Format Conversion and Output Unit. Detailed Implementation

[0026] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0027] The occupational health field equipment layout automatic mapping system of this embodiment consists of a data acquisition end that arrives at the work site with the personnel and enters the site, and a back-end processing end located outside the hazardous area. The two are connected by a data transmission module 5. The data acquisition end is responsible for the raw acquisition and local caching of field point cloud and pose data, while the back-end processing end is responsible for data processing and mapping. During the scanning process, the personnel are always outside the hazardous area and can remotely control the data acquisition end via a remote control or a preset path. The system specifically includes the following structure.

[0028] The bottom of the mobile carrier terminal 1 is a wheeled chassis 11, which is a load-bearing box that can hold counterweights. A pair of drive wheels 111 are symmetrically arranged on the rear side of the chassis bottom, and omnidirectional wheels 112 are arranged on the front side, forming a differential steering walking mechanism. The drive wheels 111 are connected to the walking control unit 12 located within the wheeled chassis 11 via a reduction mechanism. The walking control unit 12 includes a drive motor and a driver; adjusting the speed and direction of the two drive wheels 111 respectively enables straight-line movement, turning, and stationary rotation, which can be achieved using conventional geared motors and differential control methods. A power module 18, preferably a rechargeable lithium battery pack, is fixed inside the box, supplying power to components requiring electricity, such as the walking control unit 12, the lidar scanning unit 2, the inertial attitude sensor 3, the positioning module 4, the data transmission module 5, the local storage module 6, and the remote control receiver unit 17, via power lines. The data acquisition terminal does not require an external power supply during on-site operation, avoiding the entanglement of dragging cables in equipment-intensive workshops. The power module 18 can also serve as a counterweight. The installation position must ensure that the overall center of gravity of the acquisition end falls within the area enclosed by the drive wheel and the caster wheel to avoid the equipment tipping over due to center of gravity issues.

[0029] Several (e.g., six) obstacle avoidance sensors 16 are embedded in the front and side walls of the wheeled chassis 11. In this embodiment, ultrasonic ranging devices are used. The signal output terminals of the obstacle avoidance sensors 16 are connected to the walking control unit 12. A remote control receiving unit 17 is provided on the top of the wheeled chassis 11. The signal output terminal of the remote control receiving unit 17 is also connected to the walking control unit 12 to receive walking commands issued by workers located outside the hazardous area via a handheld remote control. When using the preset path mode, the path point sequence is sent from the background to the walking control unit 12 via the data transmission module 5, which drives the drive wheels 111 in sequence to complete the full-area passage. The control logic for manual remote control and path tracking can adopt the conventional implementation method in the field of industrial automated guided vehicles.

[0030] A lifting column 13 (preferably a retractable cylinder) is vertically fixed to the upper surface of the self-wheeled chassis 11. A pitch adjustment gimbal 15 is located at the top of the lifting column 13. The fixed part of the pitch adjustment gimbal 15 is fixed to the top of the lifting column 13, and the rotating part is fixed to the bottom surface of the radar mounting base 14. The rotation axis is horizontal. In this embodiment, the pitch adjustment gimbal 15 is a stepper motor-driven worm gear type gimbal with a power-off self-locking function. The radar mounting base 14 tilts and rotates around a horizontal axis, allowing the scanning surface of the lidar scanning unit 2 to both scan upwards to supplement high-level pipelines, cable trays, and top air supply and exhaust ducts, and to downwards to densify sampling of pits, trenches, and equipment foundations.

[0031] Each of the above modules can be modularly assembled by purchasing, and the assembly / debugging of these modules is a mature technology, so this application will not elaborate on it.

[0032] It is known that for sites with narrow passages, large ground elevation differences, and where wheeled chassis are difficult to pass, the mobile bearing terminal 1 can be replaced by a portable terminal that is held by a person and equipped with a backpack power supply. In this case, the power supply is replaced by a backpack battery pack, and the zero-bias static calibration time of the inertial attitude sensor 3 is extended accordingly to suppress low-frequency jitter when the person holds it and enters the zero-bias average value.

[0033] It is known that for the upper space of high-level pipe corridors, roof trusses and large-span factory buildings, the mobile bearing terminal 1 can be replaced by a drone frame. The lidar scanning unit 2 and the inertial attitude sensor 3 can be installed on the mounting base under the frame via rubber shock absorbers. The other connection relationships and data processing flow remain unchanged.

[0034] The lidar scanning unit 2 is fixed to the upper surface of the lidar mounting base 14. This embodiment uses a mechanically rotating lidar. The contour features of machine tool columns, pipe supports, dust collector housings, etc., can all be covered by two or more scanning lines. The number of lines and the measurement range are selected according to the span, floor height, and equipment density of the workshop being measured, with the spacing of the scanning lines at the farthest equipment ensuring that the contour of the equipment is not missed.

[0035] The inertial attitude sensor 3 is installed inside the radar mounting base 14. Its mounting surface and the mounting surface of the lidar scanning unit 2 are the upper and lower surfaces of the same component, forming a rigid connection. This design ensures that the translation and rotation amounts between the measurement coordinate system of the inertial attitude sensor 3 and the scanning coordinate system of the lidar scanning unit 2 remain fixed after assembly and factory calibration, and do not change during movement due to the extension / retraction of the lifting column 13 or the pitch state of the pitch adjustment gimbal 15. The inertial attitude sensor 3 is equipped with a three-axis gyroscope and a three-axis accelerometer.

[0036] The positioning module 4 is located within the wheeled chassis 11. Its input terminals are connected to the lidar scanning unit 2 and the inertial attitude sensor 3, respectively, to calculate the position and heading of the mobile carrier terminal 1 in the site coordinate system in real time. The core logic of the positioning module 4 is as follows: first, it uses the angular velocity and linear acceleration output by the inertial attitude sensor 3 to perform a short-term recursive calculation of the carrier pose change between two adjacent scan frames; then, it corrects this recursive value using the registration result of the current frame point cloud and the previous frame point cloud, thus suppressing long-term drift of the inertial integral.

[0037] After the system starts up, positioning module 4 automatically completes the following steps in sequence: S001: The initial parking position of the mobile terminal 1 is taken as the origin O, the orientation of the terminal when it starts is taken as the positive X-axis, the direction perpendicular to the ground and upward is taken as the positive Z-axis, and the positive Y-axis is determined by the right-hand rule. This coordinate system is fixed once established.

[0038] S002: The terminal remains stationary for 2 to 5 seconds. The positioning module 4 averages the angular velocity and linear acceleration measurements of the inertial attitude sensor 3 during this period, using these values ​​as the zero bias records for the gyroscope and accelerometer, and subtracts them from subsequent recursions. During the stationary period, the lidar scanning unit 2 simultaneously acquires the first frame of point cloud data as a reference frame.

[0039] S003: The pitch and roll angles are determined by the gravity direction observations of the accelerometer during the stationary period, the heading angle is determined by the scan start direction set by the user, and the position coordinates are set to zero.

[0040] S004: Within the first 5 meters of travel after the terminal starts, the scale factor and installation angle error of the inertial attitude sensor 3 are calibrated online using the registration results between the wheel odometer and the scanning frame.

[0041] After the above steps are completed, the positioning module 4 enters the formal synchronous positioning and mapping state. When the same workshop needs to be scanned multiple times or along multiple paths, the starting point coordinates and starting heading of each scan are written into the record file header of the positioning module 4.

[0042] Considering the common occurrence of dust and oil mist in occupational health environments, the outer cover of the lidar scanning unit 2 is equipped with a protective cover 21, which is a cylindrical shell with an open lower end, and its lower edge is sealed to the edge of the lidar mounting base 14. A light-transmitting window 22 is provided around the perimeter of the cover corresponding to the laser transceiver surface, made of an optical material with a transmittance greater than 92% for the 905 nm wavelength band. Several purge nozzles 23 are evenly distributed circumferentially on the outer side of the light-transmitting window 22 and along the upper edge of the cover. The air inlet of each nozzle is connected to an air source 19 fixed in the inner cavity of the wheeled chassis 11 via an air pipe arranged along the outside of the lifting column 13. The air source 19 is a small oil-free air compressor. During scanning, the air source 19 continuously sprays airflow through the purge nozzles 23 onto the window surface of the light-transmitting window 22, forming an adhering air curtain on the outer side of the window surface, causing settled dust and oil mist droplets to be blown away before adhering. It is known that a filter with a certain pore size is installed at the inlet of the compressor.

[0043] The local storage module 6 is located inside the wheeled chassis 11. Its input end is connected to the data output end of the lidar scanning unit 2, the inertial attitude sensor 3, and the positioning module 4. It stores the original scan frames, timestamps, inertial attitude data, and positioning data together on the disk. Its output end is connected to the data transmission module 5. This embodiment preferably uses industrial-grade solid-state storage.

[0044] The data transmission module 5 includes a wireless communication unit 51 and a wired data interface 52. The wireless communication unit 51 is used to transmit data back to the artificial intelligence processing module 7 in real time when network conditions are available on site; the wired data interface 52 is used in closed workshops, underground factories, and other places where there is no network coverage or wireless transmission is not allowed, and exports the data in the local storage module 6 to the backend in one go after the operation is completed.

[0045] The artificial intelligence processing module 7 is deployed on a back-end server or industrial control computer. The artificial intelligence processing module 7 consists of a real-time registration unit 71, a point cloud preprocessing unit 72, and a recognition and classification unit 73 connected in sequence, as well as other necessary electrical components.

[0046] The real-time registration unit 71 maintains a sliding window containing recent frames of scan data and corresponding pose estimates. Upon receiving a new frame of data, the real-time registration unit 71 transforms the frame's point cloud from the scan coordinate system to the site coordinate system based on the pose estimate provided by the positioning module 4; it downsamples the transformed point cloud using a voxel grid with a side length of 0.05 meters to extract corner points and planar points; it registers the downsampled current frame point cloud with existing frame point clouds within the window using a normal distribution transformation algorithm, calculating the registration residual; it uses this residual as an observation value and performs joint optimization of the poses of all frames within the window using a factor graph optimization method; it updates the coordinate transformation parameters of each frame with the optimized poses and reprojects them; finally, it merges the current frame point cloud into the global point cloud set. Feature extraction and registration calculations are accelerated in parallel by a graphics processing accelerator card; in this embodiment, the single-frame processing latency is 62 to 94 milliseconds.

[0047] When scanning multiple times or along multiple paths, the real-time registration unit 71 first reads the initial pose information in the header of each scan record file, and performs an initial rigid body transformation on each point cloud according to its initial pose. Then, it extracts feature points in the estimated overlapping area, removes mismatched point pairs using a random sampling consistency method, and calculates the coordinate transformation matrix between each scan based on the remaining matching point pairs. Finally, it unifies the coordinates to the site coordinate system with the origin of the first scan as the reference.

[0048] The filtering subunit 721 first downsamples the global point cloud using a voxel grid, and then uses statistical filtering to remove outliers. The statistical filtering is implemented by calculating the average distance of all points in the neighborhood of each point. Points whose average distance exceeds 2.5 times the standard deviation of the global average distance are identified as outliers and removed. These outliers mainly originate from the scattering echoes of laser light from suspended dust and water mist in the air, as well as the trails created by temporarily moving personnel and forklifts on site.

[0049] Ground segmentation subunit 722 fits the maximum horizontal plane in the filtered point cloud using a random sampling consensus method. Points belonging to this plane and its neighborhood are identified as ground level and stripped away, while the remaining points are retained as target candidate points. Most of the equipment is located on the ground level via foundations, and stripping away the ground level cuts off the point cloud pathways connecting the equipment through the ground level.

[0050] The clustering subunit 723 performs two-stage clustering on the target candidate points and forms a closed loop with the identification and classification unit 73.

[0051] The first stage involves Euclidean distance clustering using a first merging threshold. Spatially adjacent points are grouped into the same initial target cluster, while points with intervals exceeding the threshold are assigned to different clusters. The center coordinates, length, width, height, and principal axis direction of the minimum bounding box for each cluster are calculated. The first merging threshold is set according to the density of equipment distribution on-site: in this embodiment, 0.30 meters is used in a machining workshop with a span of 24 meters and a floor height of 9 meters; 0.22 meters is used in the reactor area of ​​a fine chemical workshop with even smaller equipment spacing; and 0.45 meters is used in sparse areas such as finished product storage yards. The purpose of this stage is to provide the identification and classification unit 73 with classification input that is complete in size and has clear boundaries. Merging of occluded and fragmented pieces has not yet been considered.

[0052] Phase 1: Clustering subunit 723 receives the category label output by device type classification branch 732 for each initial target cluster, retrieves the reference size of that category from the device voxel model library, and takes the product of the reference size and a preset scaling factor as the second merging threshold for that target cluster. In this embodiment, the scaling factor is 0.30. Two initial target clusters whose bounding box center distance is less than the second merging threshold and whose bounding box main axis angle is less than 15° are determined to be the same target and merged. After merging, the center coordinates, length, width, height dimensions, and main axis direction of the bounding box are recalculated and sent back to the identification and classification unit 73 for re-judgment.

[0053] The parameters of some categories in the device voxel model library and the resulting second merging threshold are shown in Table 1.

[0054] Table 1: Parameters of some categories in the device voxel model library and the second merging threshold For slender targets with a bounding box aspect ratio greater than 3, such as pipes, belt conveyors, and cable trays, occlusion splitting mostly occurs along the length direction. Therefore, the second merging threshold along the principal axis is taken as three times the value in Table 1, while the value perpendicular to the principal axis remains the same. For example, the threshold along the principal axis for the belt conveyor in item 7 is 0.72 meters. For non-linear continuous targets (such as curved pipes), the direction of the first principal component in the principal component analysis of the target's point cloud is taken as the principal axis direction, and the aspect ratio is calculated as the ratio of the projected length in that direction to the equivalent diameter perpendicular to that direction.

[0055] If the category label is inconsistent with the label before the re-judgment, clustering subunit 723 redetermines the second merging threshold based on the baseline size corresponding to the new label and merges again, with an iteration limit of 3 times. Target clusters whose labels are still inconsistent after reaching the iteration limit are marked as awaiting manual confirmation. In this embodiment, out of a total of 150 devices in the two sites, 7 target clusters triggered the second iteration, and 1 cluster reached the iteration limit.

[0056] The identification and classification unit 73 consists of a shared feature extraction subunit 731 and parallel equipment type classification branch 732 and occupational hazard attribute classification branch 733.

[0057] The shared feature extraction subunit 731 samples 1024 representative points from the input target cluster using the farthest point sampling method. Small target clusters with fewer than 1024 points, such as junction boxes, eyewash stations, and branch pipes of local exhaust hoods, are supplemented by repeated sampling. Each sampling point, with its three-dimensional coordinates and reflection intensity, constitutes a four-dimensional input. This input is processed by the first set of shared weight multilayer perceptrons to obtain a point-by-point local feature set, and then by the second set of shared weight multilayer perceptrons and max pooling to obtain a global feature vector. The network layer parameters are shown in Table 2.

[0058] Table 2: Layer parameters of shared feature extraction subunit and dual-branch system The equipment type classification branch 732 is based on common equipment in occupational health sites. This embodiment has a total of 34 categories, including metal cutting machine tools (lathes, milling machines, grinding machines, boring machines, drilling machines), forging and pressing equipment (presses, punch presses, shearing machines, air hammers), grinding wheels and polishing equipment, welding stations, reaction vessels, storage tanks (vertical and horizontal), centrifuges, drying equipment, belt conveyors, bucket elevators, crushers, ball mills, dust collectors, fume hoods, local exhaust hoods, air compressors, water pump units, power distribution cabinets, industrial robots, operating tables, material stacks, pipes, cable trays, walls, structural columns, passageways, safety exits, etc.

[0059] The label set of occupational hazard attribute classification branch 733 is constructed according to the needs of occupational health professions. This embodiment has a total of 7 categories: generating dust, generating noise, releasing chemical toxins, generating high temperature, generating vibration, generating ionizing radiation, and generating non-ionizing radiation. This branch adopts multi-label output, and the judgment threshold of each output neuron is 0.5, in order to adapt to the actual situation where the same equipment generates several hazard factors at the same time.

[0060] The two branches have different input features. Branch 732, for equipment type classification, uses a global feature vector as input, representing the overall geometric shape of the target, which aligns with the requirements for equipment category discrimination. Branch 733, for occupational hazard attribute classification, uses a concatenated global feature vector and point-by-point local feature set as input. The point-by-point local features preserve the geometric structure and reflection intensity of the neighborhood of each sampling point. Openings, gaps, and edges of the cover on the target surface constitute geometric discontinuities, which can be used to distinguish between closed and open states of equipment; the pipe morphology at the connection between the exhaust hood and the equipment can be used to determine the installation status of protective facilities; the reflection intensity value can characterize the degree of dust accumulation and liquid residue on the equipment surface. The determination of hazard attributes is based on these local quantities, without going through the intermediate step of equipment category.

[0061] The two branches share the same core features and are derived from the same forward operation. The geometric category and hygiene attribute are given for the same target cluster from the beginning, and the two are bound together by the index of the target cluster in the point cloud preprocessing unit 72. Subsequent model reconstruction and map annotation no longer require manual establishment of a correspondence between the two sets of data.

[0062] In a machining workshop with a span of 24 meters, a floor height of 9 meters, a building area of ​​1860 square meters, and 87 pieces of equipment, and a fine chemical workshop with a span of 18 meters, a floor height of 12 meters, a building area of ​​1240 square meters, and 63 pieces of equipment (all equipment was in normal production and in place when the point cloud was collected, and temporary obstructions had been cleared), the same batch of point clouds was subjected to single-stage fixed threshold clustering and the two-stage closed-loop clustering of the present invention, respectively. The results are shown in Table 3.

[0063] Table 3: Comparison of segmentation results between two-stage closed-loop clustering and single-stage fixed-threshold clustering Note: Target recall rate refers to the ratio of the number of targets that correspond to the on-site list and whose enclosed box size deviation is within ±10% to the actual number of equipment units; the average deviation of enclosed box size is the arithmetic mean of the relative deviations of the length, width, and height of each target enclosed box relative to the actual measured values ​​of the measuring tape, and a negative value indicates that it is too small.

[0064] When the threshold is 0.20 meters, the average deviation of the bounding box size is -18.7%, reflecting the smaller size due to fragments not being combined; when the threshold is 0.50 meters, the deviation turns to +12.4%, reflecting the larger size due to adhesion between adjacent devices. Under the two-stage closed loop, this deviation decreases to -3.2% and -4.1%.

[0065] 412 target clusters that had been segmented and manually verified were taken from three sites. Occupational hazard factors were estimated by looking up tables by equipment category and directly identified by dual branches. The results were compared with the on-site occupational health survey results. The results are shown in Table 4 (the remaining 301 samples, which cover common equipment such as pumps, fans, compressors, and distribution cabinets, are not listed).

[0066] Table 4: Comparison of methods for determining occupational hazard factors Note: The false negative rate refers to the ratio of actual hazardous factors that were not detected to the total number of actual hazardous factors in the sample; the false positive rate refers to the ratio of hazardous factors that were detected but did not actually exist to the total number of detected hazardous factors in the sample.

[0067] For the four types of samples—closed reactors, covered grinding machines, fully enclosed conveyors, and forging equipment with soundproof covers—the equipment categories were the opposite of the actual hazard conditions, and the misclassification rate using the lookup table method was 100% for all four types. The misclassification rate for direct identification of dual-branch systems on the same batch of samples was 0-8.3%. Across all 412 samples, the misclassification rate decreased from 16.3% to 5.6%, and the false negative rate decreased from 11.4% to 4.1%.

[0068] The deep learning model relied upon by the classification unit 73 was trained using a combination of field-measured point clouds and manual annotation. Point cloud data was collected from 12 occupational health sites of different industries and scales, and after segmentation by the point cloud preprocessing unit 72, 13,800 target cluster samples were obtained. Each target cluster was simultaneously labeled with equipment category tags (single choice) and occupational hazard factor tags (multiple choice) by annotators trained in occupational health. The annotation results were cross-validated by at least two annotators before being stored in the database. Samples whose categories and hazards do not correspond, such as closed reactors, covered grinding machines, enclosed conveyors, and forging equipment with soundproof covers, accounted for no less than 8% of the training set to ensure that the hazard attribute branch learns local criteria rather than category shortcuts.

[0069] During training, the training, validation, and test sets are divided in an 8:1:1 ratio. The Adam optimizer is used with an initial learning rate of 0.001, which decays to 0.5 times every 5 training epochs, with a maximum of 200 epochs. Training is terminated early when the validation set loss stops decreasing after 10 consecutive epochs. The cross-entropy loss of the equipment type classification branch 732 and the binary classification cross-entropy loss of the occupational hazard attribute classification branch 733 are weighted and summed in a 1:1 ratio and then backpropagated. The gradients are backpropagated layer by layer through the shared feature extraction subunit 731 to update all network parameters. When a new equipment type or a new combination of hazard factors is added, incremental training can be performed after supplementing the corresponding samples to expand the category set without changing the connection structure of the system.

[0070] The system incorporates a confidence assessment mechanism. When the maximum probability value of the equipment type classification branch 732 is below 0.68, this threshold is determined based on the inflection point of the accuracy-coverage curve on the validation set, and the target category is marked as requiring manual confirmation. When all output values ​​of the occupational hazard attribute classification branch 733 are below 0.5, it is marked as undetected; when all output values ​​are above 0.5 and the difference between them is less than 0.05, it is marked as requiring manual confirmation. Targets marked as requiring manual confirmation are highlighted in the drawings generated by the automatic drawing module 8 with a red dashed outline and a question mark. Reviewers click on the highlighted target through the system's drawing review interface, modify its equipment category or hazard factor label in the attribute editing window, and after confirmation, the system only updates the corresponding graphic elements and attribute table entries on the drawing, without re-calling the identification and classification unit 73.

[0071] The data fields for each target cluster output by the identification and classification unit 73 are shown in Table 5.

[0072] Table 5: Target cluster data fields output by the identification classification unit.

[0073] The data in Table 5 is transmitted to the automatic plotting module 8 in JSON format.

[0074] The 3D model reconstruction unit 81 retrieves the corresponding parametric 3D volume from the device voxel model library based on the category label of each target cluster. The model library is stored locally on the system as a lightweight relational database, storing the reference length L0, reference width W0, reference height H0, basic geometric topology, and principal axis direction definitions under standard attitude according to category, and its content is consistent with Table 1. After obtaining the category label and measured bounding box dimensions (L1, W1, H1) of a target, the 3D model reconstruction unit 81 uses the center of the measured bounding box as the placement point, and independently sets the dimensions on the three orthogonal axes of the volume in the model library with L1, W1, and H1 respectively, with scaling factors of L1 / L0, W1 / W0, and H1 / H0. After rotating the principal axis to the measured principal axis direction, it translates to the measured center coordinates so that the outer contour of the reconstructed volume fits the measured bounding box. For devices with proportional constraints (slender shaft devices that must maintain the length-to-diameter ratio and standard model devices), the dimensions are set independently only in the unconstrained direction, while the constrained direction maintains the reference proportion. Irregular continuous targets such as pipes and walls are directly reconstructed from the convex hull or contour patch of their point cloud without matching with the model library.

[0075] The two-dimensional projection drawing unit 82 extracts the orthographic projection outline of each piece of equipment on the horizontal plane and draws it as a planar graphic according to a selected scale. In this embodiment, the 1860 square meter machining workshop is drawn at a scale of 1:100. The workshop boundary lines and passageway boundaries are drawn according to the walls and passageway targets to form the base map of the equipment layout. For any two pieces of equipment, the distance between the closest points of their respective enclosing box outer contours is taken as the equipment spacing label; the spacing between equipment and the wall or passageway boundary is taken as the shortest distance from the outer contour of the equipment enclosing box to the boundary line. This method of value determination is consistent with the definition of the clear distance between operating positions and the clear width of passageways in the occupational health evaluation report.

[0076] The hazard factor labeling unit 83 reads the hazard factor multi-label vector for each target cluster and requests the corresponding legend from the cartographic specification storage unit 84. The cartographic specification storage unit 84 pre-stores a set of occupational hazard factor legends and a set of occupational disease prevention facility legends compiled according to occupational health industry cartographic specifications, as well as layer division, line type, text height, and title block format parameters. All hazard factor legends are standardized vector graphics conforming to the warning sign graphic style and color system stipulated in GBZ 158 "Occupational Disease Hazard Warning Signs in the Workplace". Layer division and legend attribution are shown in Table 6.

[0077] Table 6: Drawing Layer Division and Legend Assignment After obtaining the legend, the hazard factor labeling unit 83 places it on the upper right side of the corresponding equipment outline on the base map. If the labeling density is too high, it automatically switches to the upper left, lower right, or lower left side to avoid legend overlap. Next to the legend, the equipment's number and attribute information table are appended. Each row of the attribute information table corresponds to one piece of equipment, and the columns include equipment number, equipment name, equipment category, coordinates, length, width, and height dimensions, hazard factor category, and remarks. Targets requiring manual confirmation are represented by a red dashed outline on the map, and a list of items to be confirmed is automatically generated at the map boundary, listing the numbers and coordinates of all targets to be confirmed.

[0078] The format conversion output unit 85 exports the two-dimensional layout diagram as DXF, DWG graphics exchange format or PNG, PDF image format files, and exports the spatial model retained by the three-dimensional model reconstruction unit 81 as OBJ, STEP, glTF / GLB three-dimensional model format files for subsequent hazard diffusion simulation and protective facility location analysis. In the exported two-dimensional layout diagram, each piece of equipment, legend, and annotation text belongs to a different layer, facilitating editing or disabling of specific layers by third-party drawing software.

[0079] After arriving at the site, the operators first delineate the survey area based on the workshop's column grid and passageways. They then place the mobile support terminal 1 at the starting point of the delineated area, adjust the extension length of the lifting column 13 and the angle of the tilt adjustment platform 15 according to the height of the equipment on site, and then start the system. After system startup, sensor initialization and coordinate reference establishment are automatically performed.

[0080] After entering scanning mode, the mobile carrier terminal 1 moves along the workshop aisle according to a preset path or remote control command. The obstacle avoidance sensor 16 continuously detects obstacles in front and to the sides and sends signals to the walking control unit 12. When the measured distance is less than 0.6 meters, the walking control unit 12 decelerates, detours, or brakes. Synchronously with walking, the lidar scanning unit 2 outputs a 3D point cloud frame by frame, the inertial attitude sensor 3 and the positioning module 4 output the pose synchronously, and the air source 19 continuously blows through the light-transmitting window 22 via the blowing nozzle 23. All data is first written to the local storage module 6 and simultaneously transmitted back to the backend via the data transmission module 5. The artificial intelligence processing module 7 in the backend sequentially completes registration, filtering, ground stripping, two-stage clustering, feature extraction, and dual-branch discrimination, and the automatic drawing module 8 then completes 3D reconstruction, 2D projection, legend annotation, and format export. When real-time transmission is not available on-site, the data is first stored in the local storage module 6 and then imported into the backend for offline processing via the wired data interface 52 after the operation is completed. The processing flow is consistent with the real-time mode. Throughout the entire process, personnel remain outside the hazard area.

[0081] In workshops with dense equipment and severe obstruction, continuous walking scanning can be replaced by multi-point fixed-point scanning, that is, a stationary 360° scan is performed at several selected points, and the data from each station is stitched together by the real-time registration unit 71 according to the characteristics of the overlapping areas to form a global point cloud.

[0082] Furthermore, when the pose estimation confidence of the positioning module 4 is lower than a preset threshold, the real-time registration unit 71 pauses the coordinate transformation of new frames and reinitializes the pose estimation by feature matching between the current frame and the most recently successfully registered frame. After the registration residual converges to the normal range, subsequent frames are processed. If positioning cannot be restored for 10 consecutive frames, the system pauses walking and sends a prompt signal through the remote control receiver unit 17.

[0083] When the wireless communication unit 51 disconnects, the acquisition end continues to acquire data and writes the data completely into the local storage module 6; after communication is restored, the background artificial intelligence processing module 7 automatically requests the acquisition end to resume transmission from the point of last successful transmission, and the acquisition end starts sending subsequent data from the position of the last successful transmission.

[0084] The battery management system of the power module 18 continuously monitors the remaining power. When the power is lower than 15% of the total capacity, the walking control unit 12 records the current scanning position and performs a return operation along the original path back to the starting point, and sends a low power warning through the remote control receiving unit 17.

[0085] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An automatic mapping system for occupational health field equipment layout, comprising a mobile carrier terminal and a lidar scanning unit thereon, an artificial intelligence processing module including a point cloud preprocessing unit and a recognition and classification unit, and an automatic mapping module, characterized in that, The identification and classification unit includes a shared feature extraction subunit and parallel equipment type classification branches and occupational hazard attribute classification branches; the clustering subunit of the point cloud preprocessing unit and the identification and classification unit form a closed loop: the clustering subunit clusters the point cloud with a first merging threshold to obtain an initial target cluster, and then uses the reference size corresponding to the label output by the equipment type classification branch as a second merging threshold, merges those with a spacing smaller than the threshold and sends them back for re-judgment; the automatic drawing module generates a map based on the two types of labels after re-judgment and labels the occupational hazard factor legend.

2. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The shared feature extraction subunit outputs a point-by-point local feature set and a global feature vector to the input target cluster; the equipment type classification branch takes the global feature vector as input, and outputs the probability distribution of each equipment category through a fully connected layer and a normalized exponential function, taking the one with the highest probability as the category label of the target cluster; the occupational hazard attribute classification branch takes the feature obtained by concatenating the point-by-point local feature set with the global feature vector as input, and outputs a judgment value of 0 to 1 for each type of occupational hazard factor after passing through a shared multilayer perceptron and a pooling layer, and each judgment value is independently compared with a preset threshold to form a multi-label hazard attribute label.

3. The automatic mapping system for occupational health field equipment layout according to claim 2, characterized in that, The shared feature extraction subunit samples a preset number of representative points from the target cluster using the farthest point sampling method. If the number of points is insufficient, it is supplemented by repeated sampling. Each sampling point forms a four-dimensional input with three-dimensional coordinates and laser reflection intensity. The local feature set is obtained by the first set of shared weight multilayer perceptrons, and then obtained by the second set of shared weight multilayer perceptrons and max pooling to obtain the global feature vector with permutation invariance.

4. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The artificial intelligence processing module also includes a device voxel model library, which stores the reference length, reference width, reference height, basic geometric topology, and main axis direction under standard posture according to device category; the clustering subunit performs Euclidean distance clustering in the first stage with a first merging threshold, and calculates the center coordinates, length, width, height, and main axis direction of the minimum bounding box of each initial target cluster; In the second stage, the reference size corresponding to the category label of the initial target cluster is retrieved from the device voxel model library. The product of the reference size and the preset scaling factor is taken as the second merging threshold. Two initial target clusters whose bounding box center distance is less than the second merging threshold and whose bounding box main axis direction angle is less than the preset angle are determined to be the same target and merged. After merging, the center coordinates, length, width and height dimensions and main axis direction of the bounding box are recalculated and sent back to the identification and classification unit for re-judgment.

5. The automatic mapping system for occupational health field equipment layout according to claim 4, characterized in that, For slender target clusters whose bounding box aspect ratio is greater than a preset multiple, the second merging threshold of the clustering subunit along its bounding box main axis is taken as an integer multiple of the product of the reference size and the preset scaling factor, while the second merging threshold perpendicular to the main axis remains unchanged. For non-linear continuous target clusters, the direction of the first principal component obtained from the principal component analysis of its point cloud is the main axis direction, and the aspect ratio is calculated as the ratio of the projected length in that direction to the equivalent diameter perpendicular to that direction.

6. The automatic mapping system for occupational health field equipment layout according to claim 4, characterized in that, When the category label obtained by the identification and classification unit is inconsistent with that before the re-judgment, the clustering subunit redetermines the second merging threshold based on the reference size corresponding to the new category label and merges again. If the number of iterations reaches the preset upper limit and the category labels are still inconsistent, the target cluster is marked as pending manual confirmation. When the maximum probability value output by the equipment type classification branch is lower than the first confidence threshold, or when all the judgment values ​​of the occupational hazard attribute classification branch are higher than the second confidence threshold and the differences between them are all less than the preset difference, the target cluster is also marked as pending manual confirmation. The marked target cluster is highlighted in the drawing generated by the automatic drawing module with an outline that is different from the other targets, and a list of items to be confirmed is listed at the boundary of the drawing.

7. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The point cloud preprocessing unit further includes a filtering subunit and a ground segmentation subunit connected sequentially before the clustering subunit. The filtering subunit first downsamples the entire point cloud using a voxel grid, then calculates the average distance of all points in its neighborhood for each point, and identifies and removes points whose average distance exceeds a preset multiple of the global average distance. The ground segmentation subunit fits the maximum horizontal plane in the filtered point cloud using a random sampling consistency method, identifies points belonging to the plane and its neighborhood as ground surfaces, and peels them off. The remaining points are output as target candidate points to the clustering subunit.

8. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The mobile carrier terminal is also equipped with an inertial attitude sensor and a positioning module. The positioning module uses the angular velocity and linear acceleration output by the inertial attitude sensor to perform a short-term recursive calculation of the pose change between two adjacent scan frames, and then corrects the recursive value with the registration result of the point cloud of the adjacent frames. The artificial intelligence processing module also includes a real-time registration unit connected to the point cloud preprocessing unit. The real-time registration unit maintains a sliding window containing the scan data of the most recent few frames and the corresponding pose estimation values. The newly entered frame point cloud is transformed to the site coordinate system according to the pose estimation value of the positioning module. The corner points and plane points are extracted after voxel grid downsampling. The normal distribution transformation algorithm is used to register with the existing frames in the window and calculate the registration residual. The residual is used as the observation value to jointly optimize the pose of all frames in the window through factor graph optimization. Then, the optimized pose is reprojected and incorporated into the global point cloud.

9. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The automatic drawing module includes a 3D model reconstruction unit, a 2D projection mapping unit, a hazard factor labeling unit, a mapping specification storage unit, and a format conversion and output unit. The 3D model reconstruction unit retrieves the corresponding parametric 3D volume from the device voxel model library based on the target cluster's category label. It independently sets the length, width, and height of the measured bounding box on the three orthogonal axes of the volume, rotates its principal axis to the measured principal axis direction, and then translates it to the measured center coordinates, ensuring the reconstructed volume's outer contour fits the measured bounding box. The 2D projection mapping unit takes the orthogonal projection of each reconstructed volume on the horizontal plane. The silhouette is drawn as a base map according to a selected scale, and the equipment spacing is marked by the nearest point distance between the outer contours of any two target bounding boxes; the hazard factor identification and labeling unit requests the corresponding legend from the mapping specification storage unit according to the hazard attribute label and places it on one side of the corresponding equipment contour; the mapping specification storage unit pre-stores occupational hazard factor legend set, occupational disease prevention facility legend set, and layer division, line type, text height and frame format parameters; the format conversion and output unit exports the base map by layer into a graphic exchange format or image format file, and exports the reconstructed spatial model into a three-dimensional model format file.

10. The automatic mapping system for occupational health field equipment layout according to claim 1, characterized in that, The mobile carrier terminal includes a wheeled chassis, a walking control unit located within the wheeled chassis, a lifting column vertically fixed to the upper surface of the wheeled chassis, a pitch adjustment gimbal located at the top of the lifting column, and a radar mounting base fixed to the rotating part of the pitch adjustment gimbal. The laser radar scanning unit is fixed to the upper surface of the radar mounting base. An inertial attitude sensor is installed in the inner cavity of the radar mounting base, and its mounting base surface and the mounting base surface of the laser radar scanning unit are the upper and lower surfaces of the same component. The outer cover of the laser radar scanning unit is provided with a protective cover with an open lower end. The lower edge of the protective cover is sealed to the edge of the radar mounting base. A light-transmitting window is opened around the laser transceiver surface of the protective cover. Several purge nozzles are evenly distributed circumferentially on the outer side of the light-transmitting window. The air inlet of each purge nozzle is connected to an air source fixed in the inner cavity of the wheeled chassis via an air pipe arranged along the inner cavity of the lifting column, and continuously sprays airflow onto the window surface of the light-transmitting window to form an adhering air curtain.

Images