Nuclear power station shear wall vibrating robot and intelligent control method

Abstract

The invention relates to the technical field of vibrating robots, and discloses a nuclear power station shear wall vibrating robot and a control method, the nuclear power station shear wall vibrating robot comprises a main frame, a folding arm mechanism, a cable connected with a plug-in vibrator and a cable winding mechanism; the folding arm mechanism comprises a first knuckle arm assembly and a second knuckle arm assembly which are connected through a connecting component, a connecting frame is arranged at one end of the second knuckle arm assembly, and a guide wheel, a laser range finder and a binocular camera are rotationally connected to the connecting frame; walking wheels are arranged on the main frame, and a 360-degree camera is arranged on the outer side of the main frame and used for shooting images around the pouring template platform; a control system is arranged in the main frame and used for receiving data of the laser range finder, the binocular camera and the 360-degree camera and instructions of an upper computer and controlling the extending length of the folding arm mechanism and operation of the cable winding mechanism. The problems that operation is inconvenient, the vibration quality cannot be accurately controlled, and vibration operation control is inaccurate in the shear wall vibration operation can be solved.

Description

Technical Field

[0001] This invention relates to the field of vibration robot technology, specifically to a vibration robot for shear walls in nuclear power plants and its intelligent control method. Background Technology

[0002] Currently, during concrete pouring construction, in order to ensure uniform concrete distribution and improve and enhance the structural strength of the concrete after solidification, it is necessary to vibrate the concrete while pouring. This is especially important in some special application scenarios, such as shear walls in nuclear power plants.

[0003] Vibration operations have a significant impact on the quality of concrete components. The concrete vibration process must meet technical requirements such as "vertical insertion," "quick insertion," "slow withdrawal," and "no contact with the ground." However, judging whether concrete vibration is qualified often relies on the experience of construction personnel, which cannot accurately determine whether vibration has been completed and is highly subjective. Currently, the main vibration methods are manual vibration and mechanical vibration. Manual vibration has drawbacks such as slow vibration progress, low construction efficiency, and poor vibration control precision. Moreover, some construction areas are not suitable for manual operation. Mechanical vibration, by introducing vibration equipment, improves work efficiency compared to manual vibration, but it suffers from inconvenience in operation, inaccurate control of vibration quality, and imprecise vibration operation control.

[0004] Currently there are such Figure 1 The vertical wall shown has a complex structure with a detachable composite material formwork on the outside, a relatively dense steel mesh inside, and vertical steel bars protruding from the formwork at the top. The shear wall is poured in different sequences and positions, with a maximum single pouring depth of about 6m. Pouring defects are prone to occur in areas with small thickness and large depth. The exposed vertical steel bars at the top of the formwork occupy most of the space, making it difficult to construct manually using a handheld vibrator. Summary of the Invention

[0005] Based on the above description, the present invention provides a nuclear power plant shear wall vibration robot and control method to solve the problems of inconvenient operation, inaccurate control of vibration quality, and imprecise vibration operation control in shear wall vibration operations.

[0006] On the one hand, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: a nuclear power plant shear wall vibration robot, including an insert vibrator; and also including a main frame, a folding arm mechanism, and a cable and rope winding mechanism connected to the insert vibrator; the folding arm mechanism includes a first arm assembly and a second arm assembly rotatably connected to the main frame, and the first arm assembly and the second arm assembly are connected by a connecting member;

[0007] The second arm assembly has a connecting frame at the end away from the main frame. The connecting frame is rotatably connected to a guide wheel for guiding the cable, a laser rangefinder with the test surface always facing downwards, and a binocular camera with the shooting lens always facing downwards. The main frame is equipped with walking wheels, and a 360° camera is installed on the outside of the main frame to capture images of the area around the casting template platform.

[0008] The main frame is equipped with a control system, which receives data from the laser rangefinder, binocular camera, and 360° camera, as well as instructions from the host computer, and controls the extension length of the folding arm mechanism and the operation of the cable winding mechanism.

[0009] The above technical solution integrates a folding arm mechanism, a cable winding mechanism, and multiple sensors such as a 360° camera, a binocular camera, and a laser rangefinder on the main frame, and is uniformly scheduled by the control system. This allows the immersion vibrator to achieve rapid positioning and stable insertion in the narrow and highly interference-prone environment of the shear wall formwork platform in a nuclear power plant. The guide wheel standardizes the cable routing, and the winding mechanism enables dynamic management of tension and allowance, avoiding tangling, dragging, and accidental contact. The multi-sensor fusion perception significantly improves the accuracy of vibration point identification and posture maintenance, reduces manual intervention and safety risks, shortens process changeover time, improves the consistency of vibration coverage and concrete density, and ensures the repeatability and traceability of construction quality.

[0010] Based on the above technical solution, the present invention can be further improved as follows.

[0011] Furthermore, the first arm assembly includes a first arm section, a first connecting rod, and a connecting plate. The two ends of the first arm section and the first connecting rod are rotatably connected to the connecting plate and the connecting member, respectively, and the four rotation points form a parallelogram.

[0012] The second boom assembly includes a second boom section and a second connecting rod. The two ends of the second boom section and the second connecting rod are rotatably connected to the connecting frame and the connecting member, respectively, and the four rotation points form a parallelogram.

[0013] Both the first and second boom sections have meshing synchronous gears fixedly connected to their respective ends that are close to each other.

[0014] Through the above technical solution, both the first and second arm assemblies adopt a parallelogram connecting rod composed of four rotation points, so that the posture of the end effector remains basically unchanged when the arm span changes. Synchronous gears are set on the close side of the two arm sections to ensure that the two mechanisms are linked in a consistent manner, with small transmission gaps and low cumulative errors. The mechanical redundancy is low and the kinematic solution is simple, which makes it easy for the control system to achieve accurate end positioning based on angle measurement, i.e., the position of the connecting frame. The improved structural stiffness and trajectory smoothness make the immersion vibrator swing less and have a lower risk of collision when passing through dense steel bars. The insertion path is more controllable, thereby improving the vibration landing rate and positioning repeatability, and reducing the impact of repeated insertion on the uniformity of concrete.

[0015] Furthermore, the main frame is equipped with a folding arm drive motor that drives the first arm section to rotate, and an angle sensor that measures the rotation angle of the first arm section is installed inside the main frame. The angle sensor sends an electrical signal to the control system.

[0016] Through the above technical solution, the articulated arm drive motor provides stable and controllable joint torque, and the angle sensor provides real-time feedback on the angle of the first arm section, forming a closed-loop control of position and attitude. The control system can implement a feedforward + feedback composite control strategy to suppress overshoot and oscillation, improve dynamic response and steady-state accuracy, and ensure that the deviation between the end-effector pose and the target point is minimized during the arm extension process. This configuration enables the robot to complete alignment with a small safety margin even in confined spaces such as template boundaries and areas with dense rebar, reducing human correction and waiting time. At the same time, it facilitates calibration with binocular vision coordinates, establishing a consistent mapping of "pixel-angle-space", improving the repeatability and work efficiency of the whole machine.

[0017] Furthermore, the main frame is provided with walking wheels, and the casting template platform is provided with a limiting guide rail to restrict the vertical displacement of the walking wheels. An extension frame is fixedly connected to the limiting guide rail, and the extension frame is provided with a clamping component that can be detachably engaged with the vertical reinforcing bars.

[0018] Through the above technical solution, the vertical displacement of the traveling wheels on the template platform is constrained by the positioning guide rail, which can effectively suppress the jumping caused by platform unevenness and vibration. The extension frame and detachable clamping components form a rigid reference point and attachment channel that can be quickly established with the vertical reinforcing bars. This combination improves the overall walking stability and anti-tipping ability of the machine, improves the stability of the base posture during the extension and insertion of the folding arm, and thus improves the end alignment accuracy. The detachable clamping components are easy to adapt to different reinforcing bar spacing and diameter, reduce temporary reinforcement and positioning fixtures, and reduce setup time. The dual constraint of guide rail and clamping also helps to achieve rhythmic execution of repetitive path operation and multi-point operation.

[0019] Furthermore, the casting template platform is equipped with a limiting guide rail to restrict the vertical displacement of the traveling wheels, and a strong magnet that can be attracted to the casting template platform is connected to the limiting guide rail.

[0020] The above technical solution involves configuring a strong magnet on the limiting guide rail that can be attracted to the template platform. This provides adhesion and anti-lifting capability without the need for additional mechanical clamping, suppressing vertical micro-displacement caused by pouring vibration and personnel passage. The magnetic attraction solution is quick to deploy, requires minimal platform modification, and is suitable for steel or steel-containing platform environments. Its passive nature enhances system reliability and safety redundancy. In conjunction with the geometric constraints of the guide rail, the magnetic attraction can reduce visual jitter and ranging fluctuations caused by micro-vibrations of the machine body, improve the micro-motion stability and insertion repeatability of the folding arm end, and reduce the need for additional counterweights, thus balancing portability and positioning accuracy.

[0021] Secondly, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: a method for determining the compaction density of concrete based on video images, set in the above-mentioned compaction determination module, including the following steps:

[0022] S1 Video Acquisition and Preprocessing: Based on the video data captured in real time by the binocular camera (24) of the concrete vibration process, an original video image sequence containing the entire vibration process is obtained. The video image sequence is divided into several time segments, each time segment containing a preset number of consecutive frames. The region of interest is extracted for each frame image, only the concrete area is retained, and grayscale, normalization and optional image alignment and de-shaking processing are performed in sequence to obtain a preprocessed image sequence.

[0023] S2 Manual Feature Extraction: Based on the preprocessed image sequence, time-domain dynamic features reflecting vibration intensity changes are extracted from the time dimension, and spatial surface features reflecting the state of the concrete surface are extracted from the spatial dimension. The time-domain dynamic features and spatial surface features are subjected to time-series statistical compression to form a fixed-dimensional manual feature vector.

[0024] S3 Deep Feature Extraction and Temporal Aggregation: The preprocessed image sequence is input into the frame-level convolutional feature extraction subnet of a pre-trained deep neural network in chronological order. Convolutional encoding is performed on each frame to obtain the frame-level deep feature sequence. Then, the frame-level deep feature sequence is input into the temporal aggregation subnet for temporal modeling to obtain temporal deep features that characterize the entire vibration process.

[0025] S4 Feature Fusion and Compaction Classification: The manual feature vector and the temporal depth feature are fused to obtain a fused feature vector. The fused feature vector is then input into a classifier to output the discrete category of concrete compaction and / or the determination result of whether the compaction requirement is met.

[0026] S5 Vibration Condition Decision: Based on the judgment result and the preset decision rules, output vibration end prompt and / or continue vibration prompt during real-time vibration process. The preset decision rules include at least: outputting vibration end prompt when the confidence of the "meets compaction requirements" category exceeds the threshold and remains stable within a preset duration; outputting under-vibration warning when the vibration time exceeds the recommended time in the specification and the confidence of the "under-vibration" category is still high.

[0027] The deep neural network is obtained through supervised training using historical vibration video samples labeled with concrete density levels.

[0028] Through the above technical solution, a binocular camera captures real-time video of the entire vibration process. In the preprocessing stage, it performs region-of-interest extraction, grayscale conversion, normalization, and alignment stabilization to suppress background interference and camera shake on the judgment results, improving feature stability and cross-site adaptability. Based on temporal dynamic features and spatial surface features such as surface texture, it characterizes the vibration energy decay process and the apparent density of concrete. Simultaneously, it utilizes a convolutional neural network and a temporal aggregation subnet to extract deep temporal features of the entire vibration process, fusing them at the feature level to achieve high-precision classification of concrete vibration density and sensitive identification of under-vibration conditions. Combining continuous judgment results and preset decision rules, it outputs vibration end prompts and under-vibration warnings, achieving real-time, quantitative, and intelligent control of the vibration process. This reduces reliance on human experience, minimizes the risk of under-vibration or over-vibration, and improves the durability of concrete structures and the controllability of engineering quality.

[0029] Furthermore, the manual feature extraction in S2 specifically includes:

[0030] S201 Temporal Domain Dynamic Feature Extraction: Constructing a Difference Map Using the Difference Between Adjacent Frames: ;

[0031] S202 and calculates the inter-frame differential energy within the concrete region: ;

[0032] S203 yields a time-varying differential energy sequence. Statistical features, including the maximum differential energy, the mean and variance of the differential energy in the stable phase, and the decay time required for the differential energy to decay from the peak to a set proportional threshold, are extracted from the differential energy sequence. In step S204, a dense optical flow algorithm is used to calculate the optical flow vector between adjacent frames. The velocity amplitude is obtained as follows: The frame-level mean and variance of optical flow amplitude are calculated within the concrete region. The concrete region is then divided into multiple sub-regions. The mean optical flow amplitude of each sub-region is statistically analyzed, and the overall motion intensity, motion non-uniformity, and their time-varying statistical characteristics are extracted.

[0033] S205 Spatial Surface Feature Extraction: Several later-stage frame images are selected and averaged during the later stages of vibration or when vibration tends to stabilize to obtain a stable-stage surface image. Edge detection is performed on the stable-stage surface image, and the number of edge pixels per unit area is calculated as the edge density feature. Gray-scale quantization is performed on the stable-stage surface image, constructing gray-level co-occurrence matrices in multiple directions, and calculating texture statistics such as contrast, energy, homogeneity, and correlation. Local binary pattern encoding is performed on the stable-stage surface image, and a local binary pattern histogram is calculated as a texture description feature. Porous or void areas on the surface are detected through threshold segmentation and connected component analysis, and the number of pores, average area, maximum area, and the ratio of pore area to concrete area are calculated. Furthermore, the mean gray level, standard deviation gray level, and proportion of bright pixels are calculated for the stable-stage surface image to characterize the degree of slurry uplift and surface bleeding. The temporal dynamic features and spatial surface features are statistically compressed and spliced ​​in the temporal dimension to form the fixed-dimensional hand-crafted feature vector.

[0034] By jointly extracting the aforementioned temporal dynamic features and spatial surface features, this invention can finely characterize the vibration process and the apparent state of the concrete after vibration from multiple dimensions, such as vibration energy attenuation, motion intensity and its spatial uniformity, surface texture roughness, pore distribution, and bleed bright spots. The differential energy sequence and optical flow statistics provide quantifiable indicators for the changes in vibration intensity and whether "sufficient vibration has been achieved." The edge density, GLCM, LBP, and pore parameters of the surface image in the stable stage sensitively reflect the differences in compaction and the concentrated areas of defects. The fixed-dimensional manual feature vector formed after time statistical compression is rich in information and has a clear physical meaning, which can effectively enhance the ability of the subsequent classifier to distinguish between under-vibration, basically compacted, and compaction categories, and improve the stability and reliability of compaction determination under complex working conditions.

[0035] Furthermore, the deep neural network and feature fusion structure in steps S3 and S4 are specifically as follows:

[0036] The frame-level convolutional feature extraction subnetwork includes a two-dimensional convolutional neural network with a residual network as the backbone. It performs convolution, pooling, and residual operations on each preprocessed frame of the input image and outputs a fixed-dimensional sequence of frame-level feature vectors.

[0037] The temporal aggregation subnet is a bidirectional long short-term memory network or its stacked structure. It receives the frame-level feature vector sequence, models and fuses the forward and backward hidden states along the time dimension, and outputs temporal deep features that characterize the entire vibration process.

[0038] The feature fusion includes: performing dimensionality reduction and normalization processing on the handcrafted feature vector through a fully connected layer and nonlinear activation to obtain a compressed handcrafted feature vector; and concatenating the compressed handcrafted feature vector with the temporal deep features in the feature dimension to obtain a fused feature vector.

[0039] The classifier includes at least one fully connected layer, a nonlinear activation layer, and an output layer. The output layer uses the number of neurons corresponding to the number of concrete compaction density categories and uses a soft maximum function to output the probability distribution of each category. The concrete compaction density category is determined according to the principle of maximum probability, and the category is mapped to a judgment result of "meets compaction requirements" or "does not meet compaction requirements" in combination with a preset threshold.

[0040] In three aspects, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: A smart control method for a nuclear power plant shear wall vibration robot, using the above-mentioned nuclear power plant shear wall vibration robot, includes the following specific steps: the image analysis module in the control system identifies the position of the vertical reinforcing bars based on the images captured by the 360° camera; the vibration robot moves to a suitable position, the folding arm mechanism unfolds, and at the same time, the image analysis module obtains the vibration point based on the images captured by the binocular camera using deep learning; the immersion vibrator is aligned with the vibration point, the cable winding mechanism releases the line, and the immersion vibrator is inserted into the concrete; combined with the distance to the concrete surface obtained by the laser rangefinder, the immersion vibrator is placed to the required depth; the vibration quality evaluation module analyzes the concrete during the vibration process based on the images captured by the binocular camera until the vibration is qualified.

[0041] Through the above technical solutions, the control method connects "environmental recognition—path positioning—end alignment and insertion—depth control—quality evaluation" into a closed-loop process; a 360° camera provides panoramic recognition of the rebar position, reducing manual point-to-point identification; the robot positions itself according to the plan and unfolds its folding arm, with binocular vision and deep learning determining the vibration point, improving the alignment success rate; cable deployment and retraction enable low-resistance insertion, and laser ranging provides real-time distance measurements to the surface, ensuring accurate insertion depth and avoiding under-vibration and over-vibration; the quality evaluation module dynamically determines the compaction based on image features, automatically stopping when it meets the requirements, significantly improving construction consistency and reducing reliance on manual labor and rework rates.

[0042] Furthermore, the image analysis module performs distortion correction and image stitching on the images captured by the 360° camera to generate a circumferential panoramic view of the environment, and identifies the distribution area of ​​vertical reinforcing bars and the template boundary based on the panoramic view;

[0043] The control system calculates the robot's walking path based on the identified vertical steel bar positions and the main frame's walking path planning module, and then automatically moves the robot to the target area along the limit guide rail via the walking wheel drive motor.

[0044] During the deployment of the articulated arm mechanism, the control system performs closed-loop control based on the real-time signal from the angle sensor and the distance information measured by the laser rangefinder, ensuring that the end of the insert vibrator remains vertical and aligned with the target vibration point.

[0045] Through the above technical solutions, the distortion of 360° camera images is corrected and a panoramic view is stitched together to establish a stable dimensional relationship with the platform and template. Based on the panoramic recognition of the vertical rebar array and template boundary, reliable information such as row, column, and spacing is obtained, improving the accuracy of coarse positioning and path planning, reducing deviations caused by false detections, and providing prior knowledge for stereo vision-based fine alignment. Combined with the rebar recognition results, a smooth collision avoidance path is planned under the constraint of the guide rail and automatically tracked by the walking wheels to bypass dense rebar areas, reduce ineffective turning stops, and improve work efficiency. The guide rail constraint also improves trajectory repeatability, reduces energy consumption, and reduces manual intervention. During the folding arm unfolding stage, dual closed-loop control using joint angle and laser ranging is used to simultaneously constrain the posture and end distance from the interface when approaching the vibration point, making the end approximately vertical and precisely aligned, reducing "hitting rebar - retreating - re-inserting", and triggering speed limits and soft stops when approaching the template or rebar to avoid collisions and grout splashing, achieving faster alignment, more stable insertion, and more controllable processes.

[0046] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0047] 1. Multi-source sensor fusion and rigid-flexible synergy: 360° camera, binocular vision, laser ranging and angle sensing closed loop, combined with parallelogram folding arm, synchronous gear and cable winding, to achieve fast and accurate alignment and vertical insertion; the guide rail / magnetic / clamping three-in-one stable base improves the passability and work efficiency in environments with dense steel bars and limited space, and significantly reduces manual dependence and safety risks;

[0048] 2. End-to-end quality closed loop: Panoramic distortion correction and stitching complete environmental modeling, binocular vision + deep learning acquire vibration points, laser ranging controls depth, and the quality evaluation module determines density based on texture / bubble / brightness gradient, forming an adaptive process of "identification - positioning - insertion - evaluation - pass and stop"; effectively avoiding under-vibration / over-vibration, improving coverage uniformity, repeatability and first-pass yield, and reducing rework;

[0049] 3. Reliability and Maintainability: Guide wheels with standardized routing and a winding mechanism that dynamically manages tension; finger cylinders with scraping claws that promptly remove slurry, reducing entanglement and sensor contamination, and lowering the frequency of downtime maintenance; modular walking and attachment solutions are compatible with various template platforms, reducing on-site modification costs; operation data can be recorded and reported, supporting quality traceability and process optimization, extending the overall machine lifespan, and improving the economic efficiency throughout the entire life cycle. Attached Figure Description

[0050] Figure 1 This is a structural schematic diagram of a shear wall formwork;

[0051] Figure 2 This is a schematic diagram of the overall structure of the vibrating robot according to Embodiment 1 of the present invention;

[0052] Figure 3 This is a schematic diagram of the internal structure of the vibrating robot according to Embodiment 1 of the present invention;

[0053] Figure 4 This is a schematic diagram of the folding arm mechanism of the vibrating robot according to Embodiment 1 of the present invention;

[0054] Figure 5 This is a schematic diagram of the connecting frame of the vibrating robot according to Embodiment 1 of the present invention;

[0055] Figure 6 This is a schematic diagram of the mud-scraping gripper of the vibrating robot in Embodiment 1 of the present invention;

[0056] Figure 7 This is a schematic diagram of the vibration robot and the casting template platform in Embodiment 1 of the present invention.

[0057] Figure 8 This is a schematic diagram of the cooperation between the vibrating robot and the casting template platform in Embodiment 2 of the present invention;

[0058] Figure 9 This is a schematic diagram of the structure of a vibrating unit composed of multiple vibrating robots according to Embodiment 3 of the present invention;

[0059] Figure 10 This is a schematic diagram of the concrete vibration compaction determination method based on video images in Embodiment 4 of the present invention.

[0060] Figure 11 This is a schematic diagram of the video preprocessing and ROI extraction process in Embodiment 4 of the present invention;

[0061] Figure 12 This is a schematic diagram of the feature extraction and fusion network structure in Embodiment 4 of the present invention;

[0062] Figure 13 This is a schematic diagram of the deep neural network training process in Embodiment 4 of the present invention;

[0063] Figure 14 This is a schematic diagram of the concrete vibration surface image feature recognition model of Embodiment 4 of the present invention;

[0064] Figure 15 This is a schematic diagram of concrete surface images under different vibration times in Embodiment 4 of the present invention;

[0065] Figure 16 This is a flowchart illustrating the vibration robot control method of Embodiment 5 of the present invention.

[0066] Reference numerals: 1. Immersion vibrator; 2. Main frame; 21. Folding arm mechanism; 22. Cable; 23. Guide wheel; 24. Binocular camera; 25. Pouring template platform; 26. Limiting guide rail; 27. Traveling wheel; 28. Clamping assembly; 29. ​​Strong magnet; 3. First arm assembly; 31. First arm section; 32. First connecting rod; 33. Connecting plate; 34. Folding arm drive motor; 36. Synchronous gear; 30. Connecting component; 4. Second arm assembly; 41. Second arm section; 42. Second connecting rod; 43. Connecting frame; 44. Finger cylinder; 45. Scraper gripper; 46. Scraper groove; 5. Cable winding mechanism; 6. Control system; 61. Laser rangefinder; 7. Template; 71. Vertical reinforcement; 8. Fixing frame; 81. Horizontal slide rail; 82. Moving wheel; Detailed Implementation

[0067] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0069] Example 1: Reference Figures 1-3 A nuclear power plant shear wall vibration robot includes an immersion vibrator 1, a main frame 2, a folding arm mechanism 21, a cable 22 connected to the immersion vibrator 1 and a cable winding mechanism 5, and a control system 6.

[0070] The main frame 2 is rotatably connected to the lower part of the main frame 2. The main frame 2 is fixed with a reducer and a motor that drive the rotation of the walking wheel 27. The reducer and the walking wheel 27 are driven by a chain and a sprocket. The motor is a servo motor. The servo motor communicates with the PLC controller. The PLC controller is electrically connected to the control system 6.

[0071] refer to Figures 1-4A 360° camera (not shown in the figure) is fixedly connected to the outside of the main frame 2. The 360° camera is used to capture images around the casting template platform 25 and send the captured data to the control system 6. The image analysis module in the control system 6 generates a panoramic view of the environment around the casting template platform 25 based on the image data and identifies the position of the vertical reinforcing bars. The control system 6 plans the robot's walking path based on the position of the vertical reinforcing bars and the position of the folding arm mechanism 21, visualizes the walking path and converts it into PLC control signals, and sends them to the PLC controller. The PLC controller controls the servo motor to run according to the control signals of the walking path, so that the main frame 2 moves to the appropriate position.

[0072] The articulated boom mechanism 21 includes a first boom assembly 3 and a second boom assembly 4 rotatably connected to the main frame 2. The first boom assembly 3 and the second boom assembly 4 are connected by a connecting member 30. A connecting frame 43 is provided at the end of the second boom assembly 4 away from the main frame 2. The first boom assembly 3 includes a first boom rod 31, a first connecting rod 32 and a connecting plate 33. The two ends of the first boom rod 31 and the first connecting rod 32 are rotatably connected to the connecting plate 33 and the connecting member 30, respectively, and the four rotation points form a parallelogram.

[0073] The second arm assembly 4 includes a second arm 41 and a second connecting rod 42. The two ends of the second arm 41 and the second connecting rod 42 are rotatably connected to the connecting frame 43 and the connecting member 30, respectively, and the four rotation points form a parallelogram. The ends of the first arm 31 and the second arm 41 that are close to each other are fixedly connected to a synchronous gear 36 that meshes with each other. The synchronous gear 36 enables the first arm 31 and the second arm 41 to rotate synchronously.

[0074] refer to Figures 1-4 The main frame 2 is equipped with a folding arm drive motor 34 that drives the first arm 31 to rotate. An angle sensor that measures the rotation angle of the first arm 31 is installed inside the main frame 2. The angle sensor sends an electrical signal to the control system 6. The folding arm drive motor 34 is composed of a servo motor and a reducer. The control system 6 plans the unfolding angle of the first arm assembly 3 and the second arm assembly 4 according to the thickness of the shear wall, and controls the folding arm drive motor 34 to work. The angle sensor feeds back the rotation angle of the first arm 31 to the control system 6 to realize a feedback closed loop, so as to ensure that the first arm 31 and the second arm 41 can be unfolded to the predetermined angle.

[0075] refer to Figures 3-5The connecting frame 43 is rotatably connected to a guide wheel 23 for guiding the cable 22, a laser rangefinder 61, and a binocular camera 24. The laser rangefinder 61 is fixedly installed at the bottom of the connecting frame 43, with its test surface always facing downwards. The binocular camera 24 is fixedly installed on the side wall of the connecting frame 43, with its camera lens always facing downwards. One end of the cable 22 extends downwards under the guidance of the guide wheel 23. The laser rangefinder 61 sends the measured signal to the control system 6. The control system 6 calculates the required downward distance of the immersion vibrator 1 based on the position of the concrete surface, and then controls the cable winding mechanism 5 to operate, unwinding the cable 22 so that the immersion vibrator 1 is lowered to the predetermined depth. The binocular camera 24 sends the captured image data to the control system 6. The image analysis module in the control system 6 obtains the position of the steel mesh based on the deep learning algorithm, and then identifies the position of the vibration point. Based on the position of the vibration point, the control system 6 calculates the position of the main frame 2, the angle at which the first boom 31 and the second boom 41 need to be extended, and sends it to the PLC controller. The PLC controller controls the corresponding servo motor to work.

[0076] refer to Figure 6 A finger cylinder 44 is bolted to the connecting frame 43. The solenoid valve controlling the operation of the finger cylinder 44 is controlled by a PLC controller. Each of the two grippers of the finger cylinder 44 is connected to a mud scraper 45. Each of the two mud scraper 45 has a mud scraping groove 46 that can fit against the side wall of the cable. The cable passes through the through hole formed by the two mud scraper grooves 46. When the cable winding mechanism 5 unwinds the cable 22, the finger cylinder 44 drives the mud scraper 45 to open, allowing the cable 22 to pass through quickly. When the cable winding mechanism 5 winds up the cable 22, the finger cylinder 44 drives the mud scraper 45 to close, and the mud scraper groove 46 fits against the outer wall of the cable 22. As the cable 22 moves, the mud adhering to the cable 22 can be scraped off. When the immersion vibrator 1 is vibrating, the mud scraper 45 on the finger cylinder 44 is in the closed state, and the mud scraper 45 fixes the cable 22 to the connecting frame 43, making the vibrator more stable during operation.

[0077] refer to Figure 7 A limiting guide rail 26 is fixedly installed on the casting formwork platform 25 to limit the vertical displacement of the walking wheels 27. An extension frame is fixedly connected to the limiting guide rail 26, and a clamping component 28 that can be detachably matched with the vertical reinforcing bars is installed on the extension frame.

[0078] Example 2: Reference Figure 8A nuclear power plant shear wall vibration robot differs from Embodiment 1 in that a strong magnet 29, which can be attracted to the casting template platform 25, is connected to the limiting guide rail 26. The strong magnet 29 can be a permanent magnet, electromagnet, or electro-permanent magnet. The strong magnet 29 can be detachably fixed to the limiting guide rail 26, or the limiting guide rail 26 can be made of a magnetic metal, directly magnetically connecting the limiting guide rail 26 to the casting template platform 25.

[0079] Example 3: Reference Figures 1-9 A vibration robot for shear walls in a nuclear power plant, in embodiment 1, consists of an insert vibrator 1, a folding arm mechanism 21, a cable winding mechanism 5, and a control system 6, which constitute a vibration unit; multiple vibration units are fixedly mounted on a fixed frame 8, and a moving wheel 82 is provided below the fixed frame 8. The moving wheel 82 slides and cooperates with the limiting guide rail 26, and the moving wheel 82 is driven to move by a combination of a servo motor and a reducer.

[0080] A transverse slide rail 81 is fixedly connected to the fixed frame 8. The length direction of the transverse slide rail 81 is set along the length direction of the fixed frame 8, that is, along the length direction of the shear wall. A rack is fixedly installed on the fixed frame 8, parallel to the transverse slide rail 81. A drive gear that meshes with the rack is rotatably connected to the main frame 2. A servo motor and reducer that drive the drive gear are fixedly installed on the main frame 2. The servo motor is electrically connected to the PLC controller. The limiting guide rail 26 is fixed to the casting template platform 25 by the clamping assembly 28 or the strong magnet 29. The fixed frame 8 slides along the length direction of the limiting guide rail 26 to the predetermined position. Under the control of the control system 6, the main frame 2 moves to a more precise position, the folding arm mechanism 21 unfolds, and the immersion vibrator 1 is lowered to perform vibration operation.

[0081] Example 4: This example provides a method for determining the compaction density of concrete based on video images, specifically including the following steps S1 to S6: S1 Video acquisition and preprocessing: Based on the video data captured in real time by the binocular camera 24 of the concrete vibration process, an original video image sequence containing the entire vibration process is obtained. The video image sequence is divided into several time segments, each time segment containing a preset number of consecutive frames. The region of interest is extracted from each frame image, retaining only the concrete area, and grayscale conversion, normalization, and optional image alignment and de-jitter processing are performed sequentially to obtain a preprocessed image sequence.

[0082] S2 Manual Feature Extraction: Based on the preprocessed image sequence, time-domain dynamic features reflecting vibration intensity changes are extracted from the time dimension, and spatial surface features reflecting the state of the concrete surface are extracted from the spatial dimension. The time-domain dynamic features and spatial surface features are subjected to time-series statistical compression to form a fixed-dimensional manual feature vector.

[0083] S3 Deep Feature Extraction and Temporal Aggregation: The preprocessed image sequence is input into the frame-level convolutional feature extraction subnet of a pre-trained deep neural network in chronological order. Convolutional encoding is performed on each frame to obtain the frame-level deep feature sequence. Then, the frame-level deep feature sequence is input into the temporal aggregation subnet for temporal modeling to obtain temporal deep features that characterize the entire vibration process.

[0084] S4 Feature Fusion and Compaction Classification: The manual feature vector and the temporal depth feature are fused to obtain a fused feature vector. The fused feature vector is then input into a classifier to output the discrete category of concrete compaction and / or the determination result of whether the compaction requirement is met.

[0085] S5 Vibration Condition Decision: Based on the judgment result and the preset decision rules, output vibration end prompt and / or continue vibration prompt during real-time vibration process. The preset decision rules include at least: outputting vibration end prompt when the confidence of the "meets compaction requirements" category exceeds the threshold and remains stable within a preset duration; outputting under-vibration warning when the vibration time exceeds the recommended time in the specification and the confidence of the "under-vibration" category is still high.

[0086] The deep neural network is obtained through supervised training using historical vibration video samples labeled with concrete density levels.

[0087] As a preferred option, S1 video acquisition and preprocessing:

[0088] Video acquisition: Images of the concrete are captured using a binocular camera, with the camera's field of view covering the area to be vibrated; the resolution is preferably no less than 1920×1080; the frame rate is preferably no less than 20fps to ensure temporal resolution of the vibration process. At the start of the vibration operation, a raw video sequence containing the entire vibration process is acquired. .

[0089] Region of Interest (ROI) extraction: In implementation, the concrete area mask can be obtained in several ways: During the preparation stage, operators can manually outline the concrete area on a representative image through a graphical interface, and the system records the polygonal area as a static ROI; For scenes with clear template boundaries and obvious color contrast, a semi-automatic segmentation algorithm based on color thresholding, morphological processing, and edge detection can be used to determine the outline of the concrete area; When the sample size is large and the scene changes frequently, a simple semantic segmentation network (such as U-Net or a small Deeplab) can be pre-trained to automatically segment the concrete area, further improving adaptability.

[0090] The concrete area mask is obtained through manual calibration or simple segmentation. The non-concrete areas are set to zero, resulting in:

[0091] ;

[0092] Grayscale conversion and normalization; to reduce the impact of lighting differences between different construction environments on features, this embodiment can optionally perform histogram equalization or adaptive histogram equalization (CLAHE) on the image after grayscale conversion to improve local contrast. For normalization, subtracting the mean and dividing by the variance can also be used. ;

[0093] in and Statistics can be performed globally or within the current frame. The specific normalization method used can be adjusted based on data distribution and model accuracy.

[0094] Additional explanation regarding image stabilization and alignment: For cameras mounted on equipment subject to significant vibration, to prevent the overall camera motion from being incorrectly interpreted as concrete motion, this embodiment preferably uses an inter-frame alignment method based on feature point matching. This involves extracting feature points from the template region using feature operators such as ORB / SIFT / SURF, estimating the affine or homography transformation matrix between frames using RANSAC, and mapping the current frame to the reference frame coordinate system to achieve image alignment. Performing inter-frame differencing and optical flow calculations after alignment significantly reduces spurious motion.

[0095] As a preferred embodiment, S2 manual feature extraction: In this embodiment, manual features are extracted from both the time and spatial dimensions, and the entire vibration process is subjected to time statistical compression to construct a fixed-dimensional manual feature vector. .

[0096] S201 Temporal Domain Dynamic Feature Extraction: Inter-frame Differential Energy, constructing inter-frame difference maps for adjacent two frames:

[0097] ;

[0098] Calculate the differential energy within the concrete region: ;in, This represents the number of pixels within the ROI.

[0099] This yields the time series. Further, the following statistical features were extracted: maximum difference energy. The mean value during the later stage of vibration (such as the last 1 second or the preset duration). with standard deviation The differential energy decays from the peak value to... (like Time required , is used to characterize the speed at which "stability is reached".

[0100] Optical flow field characteristics; dense optical flow algorithms (such as Farneback method) are used to calculate the optical flow vectors between adjacent frames: And calculate the velocity amplitude: ;

[0101] Within the ROI: Frame-level average velocity: ;

[0102] Frame-level speed variance: ;

[0103] In addition, the ROI is divided into several sub-blocks (e.g., a 4×4 grid), and the average velocity of each sub-block is calculated to obtain the sub-block velocity vector. The variance of each sub-block is used to characterize the uniformity of vibration.

[0104] The above optical flow sequence is statistically compressed over time to extract: the entire process Maximum value, average value, standard deviation; later stage of vibration and Mean; the variance of the average velocity of the sub-blocks over the entire process (and later stages), used to quantify the uniformity of spatial vibration.

[0105] S202 Spatial Surface Features: Select a period of time in the later stage of vibration (e.g., several frames in the last 1 second), and take the average pixel value of these frames to obtain the surface image of the stable stage. ,exist Extract surface state features.

[0106] Edge density, for Perform edge detection (such as using the Canny operator) to obtain the edge map.

[0107] Calculate the proportion of edge pixels within the ROI: Used to characterize the degree of surface roughness.

[0108] GLCM texture features, The grayscale is quantized into several levels (e.g., 8 or 16 levels), and the grayscale co-occurrence matrix (GLCM) is calculated in multiple directions (0°, 45°, 90°, 135°), thereby obtaining: contrast, energy, homogeneity, and correlation. The average of the results in different directions can be taken as the global texture feature.

[0109] LBP texture features, for Local binary pattern (LBP) encoding (such as uniform LBP with P=8 and R=1) is performed to obtain LBP images, and the LBP pattern histogram is statistically analyzed as a texture descriptor to achieve quantization of fine surface textures.

[0110] Pore / cavity characteristics, for Potential pore regions are segmented using an adaptive or fixed threshold, and the area of ​​each pore region is obtained using connected component analysis: number of pores (num_pores); average pore area (avg_area); maximum pore area (max_area); and the ratio of total pore area to ROI area (porosity_ratio).

[0111] Gray-level and brightness distribution characteristics, statistically analyzed within the ROI: mean gray-level value Gray standard deviation High-brightness pixel ratio (the proportion of pixels with grayscale greater than a set threshold) is used to characterize the degree of bright spots caused by slurry floating and seepage.

[0112] The aforementioned time-domain dynamic features and spatial surface features are statistically compressed and concatenated along the time dimension to obtain a fixed-dimensional handcrafted feature vector. .

[0113] As a preferred option, S3 deep feature extraction and temporal aggregation

[0114] In this embodiment, a deep neural network is introduced to perform end-to-end modeling of the preprocessed image sequence.

[0115] Frame-level convolutional feature extraction; for each sample, T frames are randomly selected at equal intervals or within the vibration time period from the original vibration video to form a subsequence. Each frame is scaled to a uniform size (e.g., 224×224) and used as network input.

[0116] A two-dimensional convolutional neural network (such as ResNet-18) with a residual network as its backbone is used as the frame-level feature extraction subnetwork. Multiple layers of convolution, pooling, and residual operations are performed on each frame of the image. Finally, the penultimate layer outputs a fixed-dimensional (e.g., 512-dimensional) frame-level feature vector, forming a frame-level feature sequence. ;

[0117] Temporal aggregation: The frame-level feature sequence is input into a bidirectional long short-term memory network (BiLSTM) or its stacked structure. BiLSTM models both the forward and backward directions of the time series simultaneously, outputting the hidden state at each time step. Preferably, the forward and backward hidden states of the last frame are concatenated, or the hidden states of all time steps are averaged and pooled to obtain the temporal depth feature vector. It is used to characterize the temporal evolution of the entire vibration process.

[0118] As a preferred approach, S4 feature fusion and density classification are used.

[0119] Handcrafted feature branches: For handcrafted feature vectors Compressed handcrafted feature vectors are obtained by using one or more fully connected layers with nonlinear activations (such as ReLU) and configurable batch normalization (BatchNorm) for scaling and normalization. .

[0120] Feature fusion: combining temporal deep feature vectors Compressed handmade feature vectors Cascade along the feature dimension:

[0121] ; serves as the input to the downstream classifier.

[0122] Density classification: fusing feature vectors The input consists of a classification network containing at least one fully connected layer and non-linear activations. The number of neurons in the final output layer corresponds to the number of density categories (e.g., 3 categories: under-vibration, basically dense, and meets density requirements). The probability distribution of each category is obtained through the softmax function, and the density category prediction result is given according to the principle of maximizing probability. Furthermore, the category can be mapped to a binary judgment result of "meets density requirements" or "does not meet density requirements" according to engineering needs.

[0123] As a preferred option, S5 vibration condition decision-making and prompting combines the above classification results with preset decision rules to make decisions and provide prompts during real-time vibration. Specifically, this includes, but is not limited to:

[0124] Vibration completion notification rule: When the probability value of the "meets compaction requirements" category in multiple consecutive model inference results is greater than a set threshold. If the vibration intensity is 0.8 (e.g., 0.8) and the duration is not less than the preset time window (e.g., 3s), then it is determined that the concrete at the current position has reached the target density, and a vibration end prompt is issued to the operator.

[0125] Under-vibration warning rule: When the cumulative vibration time has exceeded the recommended time in the specification (e.g., the reference time determined by the layer thickness and vibrator specifications), but the probability value of the "under-vibration" or "basically dense but locally under-vibration" category in the model output is still high, and there is no continuous "meets the density requirements" judgment, an under-vibration warning will be issued to the operator or supervisor, prompting them to extend the vibration or check the concrete and vibration method.

[0126] Abnormal operating condition prompts: If obvious abnormalities appear in the features, such as a sharp drop in vibration intensity while the number of surface pores remains high, or the surface texture and bright spot features show segregation, severe water bleeding, etc., an abnormality judgment rule can be set to output a prompt of "may be over-vibration or material abnormality" for on-site verification.

[0127] Preferably, in this embodiment, the deep neural network is trained under supervision using historical vibration video samples labeled with density, based on the PyTorch model training process. This embodiment details a model training method implemented using the PyTorch framework to obtain the aforementioned deep neural network model; the specific steps are as follows:

[0128] I. Training Data Construction:

[0129] Sample collection and calibration: Multiple sets of concrete vibration videos were collected under different engineering site and laboratory conditions. Each set of videos corresponds to one or more density level labels comprehensively evaluated by core sampling, measured density, compressive strength or other non-destructive testing methods. For example: 0: obviously under-vibration; 1: basically dense but with local under-vibration; 2: meets the density requirements.

[0130] Video slicing and frame extraction divide the complete vibration video into several sample segments according to the construction process. Each sample segment covers a specific vibration operation or the entire vibration process of a certain local location.

[0131] For each sample segment, T frames of images are extracted at fixed intervals or randomly within the vibration time range to form a subsequence as model input.

[0132] Manual feature calculation and storage: For each sample segment, calculate the manual feature vector according to step S2 of Example 1. They are stored together with the corresponding video subsequences and density labels to form training sample records.

[0133] Dataset partitioning involves dividing all samples into training, validation, and test sets according to engineering batches or random principles, for example, in a 6:2:2 ratio, to avoid data leakage and overfitting.

[0134] II. Model Structure Configuration: In this embodiment, the model generally consists of a frame-level convolutional feature extraction subnetwork, a temporal aggregation subnetwork, a handcrafted feature branch, and a classifier. The structure corresponds to S3 and S4. Specific parameter configurations are as follows:

[0135] The frame-level convolutional subnet uses ResNet-18 as the backbone network, and removes the last fully connected layer to obtain a 512-dimensional output;

[0136] The temporal aggregation subnet uses a two-layer bidirectional LSTM with a hidden layer dimension of 256 and an output dimension of 512.

[0137] The handcrafted feature branch includes a fully connected layer that compresses the handcrafted features to 128 dimensions;

[0138] After feature fusion, the input is a two-layer fully connected classifier with a hidden layer dimension of 128 and an output layer dimension consistent with the number of dense classes.

[0139] The above parameters can be adjusted according to the data scale and hardware capabilities.

[0140] III. Training Hyperparameters and Loss Function

[0141] The loss function adopted is the cross-entropy loss function. For class imbalance, the class weight can be set according to the number of samples in each class to increase the penalty for misclassification of under-vibrated classes.

[0142] For the optimizer and learning rate, the AdamW optimizer is preferred, and the initial learning rate can be set to 1×10⁻. 4 ~3×10⁻ 4 The learning rate is gradually reduced by using cosine annealing or piecewise decay strategies.

[0143] Batch size and training epochs: The batch size can be set to 4-16 depending on the available GPU memory; the number of training epochs can be set to 50-100 epochs. An early stopping strategy should be adopted in conjunction with the performance of the validation set to prevent overfitting.

[0144] IV. Training Process

[0145] During training, the model parameters are updated iteratively according to the following steps:

[0146] 1. Read a batch of samples from the training set, which contains several video subsequences and corresponding handcrafted features and labels;

[0147] 2. Preprocess the video subsequences (cropping, scaling, normalization, etc.), construct tensors and input them into the frame-level convolutional subnet to obtain frame-level feature sequences;

[0148] 3. Input the frame-level feature sequence into a bidirectional LSTM to obtain temporal deep features;

[0149] 4. Feed the handcrafted feature vectors into the handcrafted feature branch to obtain compressed handcrafted features;

[0150] 5. Concatenate the temporal deep features with compressed handcrafted features, feed them into the classifier network, and output the predicted probability of each density category;

[0151] 6. Calculate the cross-entropy loss based on the predicted probabilities and the true labels, backpropagate the gradient, and update the network parameters;

[0152] 7. After completing one epoch, evaluate the model performance on the validation set, record the validation set accuracy, recall, confusion matrix and other metrics, and determine whether the convergence condition is met or whether early stopping is required.

[0153] V. Model Selection and Deployment: After training, evaluate the performance of the final model on the test set, paying particular attention to the following metrics: volume classification accuracy; recall and error rate of under-vibration classes (those that do not meet the compactness requirements); and the proportion of under-vibration classes that are misclassified as meeting the requirements.

[0154] Select the model with the best comprehensive index as the final deployment model, export the network weight file, and deploy it to an industrial computer or embedded inference device in the field to achieve real-time inference.

[0155] Example 5:

[0156] refer to Figure 10 A method for intelligent control of a nuclear power plant shear wall vibration robot, using the nuclear power plant shear wall vibration robot described in Example 1, includes the following specific steps:

[0157] The image analysis module in the S1 control system 6 identifies the position of the vertical reinforcing bars based on images captured by the 360° camera;

[0158] The S2 vibratory robot moves to the appropriate position, the folding arm mechanism 21 unfolds, and at the same time the image analysis module obtains the vibration point based on the image captured by the binocular camera 24 and deep learning.

[0159] S3 aligns the immersion vibrator 1 with the vibration point, the cable winding mechanism 5 releases the line, and the immersion vibrator 1 is inserted into the concrete.

[0160] S4 uses the distance to the concrete surface obtained by the laser rangefinder 61 to place the immersion vibrator 1 to the required depth.

[0161] The S5 vibration quality evaluation module analyzes the concrete during the vibration process based on images captured by the binocular camera 24 until the vibration is deemed satisfactory.

[0162] Furthermore, the image analysis module calibrates and corrects distortion of the original panoramic images captured by the 360° camera, and uses feature matching and stitching fusion to generate a 25-dimensional panoramic image of the casting template platform. On the panoramic image, a trained deep learning segmentation network detects the vertical reinforcing bars (including main bars and distribution bars) and the template boundary, forming a "reinforcing bar distribution area map" and a "template boundary mask". The output includes: set of reinforcing bar center lines, minimum safe spacing constraints for reinforcing bars, reachable working zones, and prohibited insertion zones.

[0163] Furthermore, the walking path planning module reads the S1 identification results and the current construction point allocation strategy (point spacing, row and column coverage) of the pouring section, and generates the shortest walking path and stopping posture from the current position to the target work area under the constraint of the limit guide rail 26; the control system 6 drives the walking wheel 27 motor to move automatically along the guide rail and stop in the target area, and completes meter-level positioning in combination with the encoder / odometer, and then ensures the safe space for the folding arm to unfold through short-range fine adjustment (±50 mm).

[0164] Furthermore, the articulated boom mechanism 21 unfolds within the safe zone. The control system 6 reads the joint angle and end-effector attitude angle sensor signals of the first boom section 31 and the second boom section 41 to complete the zero-position and end-effector coordinate calibration of the boom system. The image analysis module simultaneously acquires images from the binocular camera 24, generates a dense depth map of the working surface based on deep learning target detection and stereo matching, and automatically filters the candidate vibration point set of this docking position (avoiding rebars ≥1.5×maximum diameter and ≥70 mm from the inner side of the template) by combining the S1 rebar mask and the template boundary, and selects the current target vibration point according to the "coverage priority - proximity principle".

[0165] Furthermore, the articulated arm posture control module calculates the inverse kinematics of the arm system based on the spatial coordinates of the target vibration point, driving each joint to align the axis of the insert vibrator 1 with the normal direction of the target point. The control system 6 establishes a closed loop of "angle sensor + laser rangefinder": using the dual threshold range of the angle between the vibrator axis and the direction of gravity as constraints, if "deviation > upper threshold", the joint angle is automatically fine-tuned in the reverse direction; if "deviation < lower threshold", the planned trajectory is fed forward. The laser rangefinder 61 measures the distance from the end face of the vibrator to the free surface of the concrete in real time, and calculates the relative insertion depth in conjunction with the arm end posture. The cable winding mechanism 5 lays out the line according to the "fast insertion - slow insertion" speed curve, and the vibrator is inserted into the concrete to the specified depth along the alignment direction.

[0166] Furthermore, the depth control module calculates the target depth at this point based on the thickness of the poured layer and the thickness of the shear wall, and compares the depth obtained by laser ranging in real time. If it touches the template or is obstructed (depth increment stops and angle / current is abnormal), it will retreat 10–20 mm and then perform a plane micro-movement (≤80 mm) at the same stopping position before inserting again. Under normal circumstances, the closed-loop stop threshold is set at the target depth ±5 mm, and the stable vibration stage is entered.

[0167] Furthermore, the vibration quality evaluation module continuously processes the temporal features of the binocular images and the ROI of the end face region to construct a multi-index criterion:

[0168] The slope of the decrease in bubble precipitation rate is ≤ the threshold.

[0169] The continuous discharge index is consistently greater than or equal to the threshold, and the edge collapse is uniform.

[0170] Texture grayscale variance and surface micro-displacement (subpixel optical flow) reach a steady state;

[0171] The "dense consistency" score of the overlapping area of ​​adjacent points meets the standard.

[0172] When the above indicators are stably met for N consecutive frames (e.g., 2–3 s), the vibration is deemed "qualified". Otherwise, vibration is maintained or a "slight lifting-reinsertion-short vibration" reinforcement strategy is implemented until the qualification is achieved or a restricted replanning is triggered.

[0173] Furthermore, the control system 6 retracts the vibrator according to a segmented curve of "slow pull-stop-slow pull" to facilitate air release and slurry backfilling, while maintaining the vertical deviation of the axis within the allowable range; the vibration time, depth curve, and quality score at this point are recorded and reported to the work log. Then, the next target point is selected according to S2, and S3–S5 are repeated until the current stop position is covered; then S2 plans the next stop position.

[0174] As a preferred embodiment, the intelligent identification and control of the rebar mesh is as follows: To ensure that the vibrator can smoothly pass through the rebar mesh (e.g., 30 cm × 30 cm) without interfering with the rebar, this embodiment adopts an integrated solution of "panoramic prior knowledge + end-effector binocular precision positioning + attitude / depth closed loop", and the specific steps are as follows:

[0175] The 360° camera on the outside of the panoramic prior main frame 2 generates a circumferential panoramic image after distortion correction and stitching. The image analysis module completes the semantic segmentation of the steel bars and formwork based on deep learning, and outputs the macroscopic distribution of vertical / horizontal steel bars and the formwork boundary, which are used as prior constraints for the no-entry zone and candidate work zone.

[0176] The binocular camera 24, positioned at the end of the binocular precision positioning frame, acquires synchronous images under PLC I / O hard triggering. High-brightness ring digital fill light is used to suppress shadows and reflections. Dense depth is obtained through binocular correction and stereo matching. Combined with a panoramic prior rebar mask, strips and intersections of the local rebar mesh are extracted, mesh areas are segmented, and the center point and normal of each mesh are calculated. Pixel coordinates are mapped to the world system {W} through camera extrinsic parameters to obtain a set of usable mesh center coordinates. The laser rangefinder 61 does not participate in mesh recognition; it is only used for subsequent depth / collision protection closed-loop.

[0177] Target aperture selection and safety margin: The center of the target aperture is selected near the current parking position based on the principle of "nearest coverage - obstacle avoidance priority". Each candidate hole is scored based on its minimum distance from the reinforcing bar, its setback from the formwork boundary (delta), and its accessibility to the formwork frame. Holes that are obstructed or have poor accessibility are filtered out. The safety margin is set at ≥1.5 × maximum reinforcing bar diameter.

[0178] Alignment and perforation insertion of the folding arm posture control module Inverse kinematics is used to generate the desired end-effector pose so that the vibrator axis aligns with the direction of gravity. During deployment, closed-loop control is implemented by integrating joint angle sensors and laser ranging signals to correct position and perpendicularity in real time, ensuring the end-effector is aligned with the mesh normal. The cable winding mechanism 5 employs a "fast insertion-slow insertion" cable release curve, allowing the vibrator to pass through the hole center and insert to the target depth. .

[0179] If interference symptoms such as "depth increment stagnation + attitude error increase or abnormal motor current" appear during the interference detection and retry strategy, the control system 6 will immediately stop, pull up 10–20 mm, move slightly in the plane by 50–80 mm, and then retry at the nearest available hole position; if it fails continuously, the next docking column will be switched according to the strategy.

[0180] In this embodiment, instead of using a sliding rail to extend the camera travel, the visible and reachable range is expanded through "intra-column serpentine traversal + folding arm repositioning". When necessary, the exposure / gain is automatically adjusted and polarized fill light is enabled to improve the contrast of the mesh boundary in wet bright templates and dense steel reinforcement environments.

[0181] Data is sent and recorded in real time, including the center coordinates, normal and alignment posture of the selected hole position, as well as the insertion trajectory parameters, to the motion controller. The perforation results, minimum safe distance, insertion depth curve, and image evidence are written to the operation log for quality traceability and statistical optimization.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vibration robot for shear walls in nuclear power plants, comprising an immersion vibrator (1); characterized in that It also includes a main frame (2), a folding arm mechanism (21), a cable (22) connected to the immersion vibrator (1), and a cable winding mechanism (5); The folding arm mechanism (21) includes a first arm assembly (3) and a second arm assembly (4) rotatably connected to the main frame (2), and the first arm assembly (3) and the second arm assembly (4) are connected by a connecting member (30); The second arm assembly (4) is provided with a connecting frame (43) at the end away from the main frame (2). The connecting frame (43) is rotatably connected to a guide wheel (23) that guides the cable (22), a laser rangefinder (61) whose test surface is always facing down, and a binocular camera (24) whose shooting lens is always facing down. The main frame (2) is equipped with walking wheels (27), and a 360° camera is installed on the outside of the main frame (2) to capture images around the casting template platform (25); The main frame (2) is equipped with a control system (6) for receiving data from the laser rangefinder (61), binocular camera (24), 360° camera and host computer instructions, and controlling the extension length of the folding arm mechanism (21) and the operation of the cable winding mechanism (5). The control system (6) is equipped with a density determination module for judging the density of concrete. The control system (6) controls the operation of the immersion vibrator (1) according to the determination result fed back by the density determination module.

2. The nuclear power plant shear wall vibrating robot of claim 1, wherein, The first arm assembly (3) includes a first arm (31), a first connecting rod (32) and a connecting plate (33). The two ends of the first arm (31) and the first connecting rod (32) are rotatably connected to the connecting plate (33) and the connecting member (30) respectively, and the four rotation points form a parallelogram. The second arm assembly (4) includes a second arm (41) and a second connecting rod (42). The two ends of the second arm (41) and the second connecting rod (42) are rotatably connected to the connecting frame (43) and the connecting member (30) respectively, and the four rotation points form a parallelogram. Both the first boom section (31) and the second boom section (41) are fixedly connected to the ends of the boom sections that are close to each other, and are connected to the meshing synchronous gears (36).

3. The nuclear power plant shear wall vibrating robot of claim 1, wherein, The main frame (2) is equipped with a folding arm drive motor (34) that drives the first arm (31) to rotate. An angle sensor that measures the rotation angle of the first arm (31) is installed inside the main frame (2). The angle sensor sends an electrical signal to the control system (6).

4. A nuclear power plant shear wall vibrating robot according to any one of claims 1-3, characterized in that, The main frame (2) is provided with a walking wheel (27), and the casting template platform (25) is provided with a limiting guide rail (26) to limit the vertical displacement of the walking wheel (27). An extension frame is fixedly connected to the limiting guide rail (26), and a clamping component (28) that can be detachably matched with the vertical steel bar is provided on the extension frame.

5. A nuclear power plant shear wall vibrating robot according to any one of claims 1-3, characterized in that, The casting template platform (25) is provided with a limiting guide rail (26) to limit the vertical displacement of the walking wheel (27), and a strong magnet (29) that can be attracted to the casting template platform (25) is connected to the limiting guide rail.

6. A method for determining the compaction degree of concrete based on video images, characterized in that, The density determination module described in claim 1 includes the following steps: S1 Video Acquisition and Preprocessing: Based on the video data captured in real time by the binocular camera (24) of the concrete vibration process, an original video image sequence containing the entire vibration process is obtained, and the video image sequence is divided into several time segments, each time segment containing a preset number of consecutive frames. For each frame of the image, the region of interest is extracted, only the concrete area is retained, and grayscale conversion, normalization and optional image alignment and de-jitter processing are performed in sequence to obtain a preprocessed image sequence. S2 Manual Feature Extraction: Based on the preprocessed image sequence, time-domain dynamic features reflecting vibration intensity changes are extracted from the time dimension, and spatial surface features reflecting the state of the concrete surface are extracted from the spatial dimension. The time-domain dynamic features and spatial surface features are subjected to time-series statistical compression to form a fixed-dimensional manual feature vector. S3 Deep Feature Extraction and Temporal Aggregation: The preprocessed image sequence is input into the frame-level convolutional feature extraction subnet of a pre-trained deep neural network in chronological order. Convolutional encoding is performed on each frame to obtain the frame-level deep feature sequence. Then, the frame-level deep feature sequence is input into the temporal aggregation subnet for temporal modeling to obtain temporal deep features that characterize the entire vibration process. S4 Feature Fusion and Compaction Classification: The manual feature vector and the temporal depth feature are fused to obtain a fused feature vector. The fused feature vector is then input into a classifier to output the discrete category of concrete compaction and / or the determination result of whether the compaction requirement is met. S5 Vibration Condition Decision: Based on the judgment result and the preset decision rules, output vibration end prompt and / or continue vibration prompt during real-time vibration process. The preset decision rules include at least: outputting vibration end prompt when the confidence of the "meets compaction requirements" category exceeds the threshold and remains stable within a preset duration; outputting under-vibration warning when the vibration time exceeds the recommended time in the specification and the confidence of the "under-vibration" category is still high. The deep neural network is obtained through supervised training using historical vibration video samples labeled with concrete density levels.

7. The method for determining the compaction degree of concrete based on video images according to claim 6, characterized in that, The manual feature extraction in S2 specifically includes: S201 Temporal Domain Dynamic Feature Extraction: Constructing a Difference Map Using the Difference Between Adjacent Frames: ; S202 and calculates the inter-frame differential energy within the concrete region: ; S203 yields a time-varying differential energy sequence. Statistical features are extracted from the differential energy sequence, including the maximum differential energy, the mean and variance of the differential energy in the stable phase, and the decay time required for the differential energy to decay from the peak to a set proportional threshold. S204 uses a dense optical flow algorithm to calculate the optical flow vectors between adjacent frames. The velocity amplitude is obtained as follows: ； The frame-level mean and variance of optical flow amplitude are calculated within the concrete region. The concrete region is then divided into multiple sub-regions. The mean optical flow amplitude of each sub-region is statistically analyzed, and the overall motion intensity, motion non-uniformity, and their time-varying statistical characteristics are extracted. S205 Spatial Surface Feature Extraction: Several later-stage frame images are selected and averaged during the later stages of vibration or when vibration tends to stabilize to obtain a stable-stage surface image. Edge detection is performed on the stable-stage surface image, and the number of edge pixels per unit area is calculated as the edge density feature. Gray-scale quantization is performed on the stable-stage surface image, constructing gray-level co-occurrence matrices in multiple directions, and calculating texture statistics such as contrast, energy, homogeneity, and correlation. Local binary pattern encoding is performed on the stable-stage surface image, and a local binary pattern histogram is calculated as a texture description feature. Porous or void areas on the surface are detected through threshold segmentation and connected component analysis, and the number of pores, average area, maximum area, and the ratio of pore area to concrete area are calculated. Furthermore, the mean gray level, standard deviation gray level, and proportion of bright pixels are calculated for the stable-stage surface image to characterize the degree of slurry uplift and surface bleeding. The temporal dynamic features and spatial surface features are statistically compressed and spliced ​​in the temporal dimension to form the fixed-dimensional hand-crafted feature vector.

8. The method for determining the compaction degree of concrete based on video images according to claim 6, characterized in that, The deep neural network and feature fusion structure in steps S3 and S4 are specifically as follows: The frame-level convolutional feature extraction subnetwork includes a two-dimensional convolutional neural network with a residual network as the backbone. It performs convolution, pooling, and residual operations on each preprocessed frame of the input image and outputs a fixed-dimensional sequence of frame-level feature vectors. The temporal aggregation subnet is a bidirectional long short-term memory network or its stacked structure. It receives the frame-level feature vector sequence, models and fuses the forward and backward hidden states along the time dimension, and outputs temporal deep features that characterize the entire vibration process. The feature fusion includes: performing dimensionality reduction and normalization processing on the handcrafted feature vector through a fully connected layer and nonlinear activation to obtain a compressed handcrafted feature vector; and concatenating the compressed handcrafted feature vector with the temporal deep features in the feature dimension to obtain a fused feature vector. The classifier includes at least one fully connected layer, a nonlinear activation layer, and an output layer. The output layer uses the number of neurons corresponding to the number of concrete compaction density categories and uses a soft maximum function to output the probability distribution of each category. The concrete compaction density category is determined according to the principle of maximum probability, and the category is mapped to a judgment result of "meets compaction requirements" or "does not meet compaction requirements" in combination with a preset threshold.

9. A method for intelligent control of a vibration robot for shear walls in nuclear power plants, characterized in that, The method of using the nuclear power plant shear wall vibration robot according to claim 1 includes the following specific steps: The image analysis module in the control system (6) identifies the position of the vertical steel bars based on the images captured by the 360° camera; The vibratory robot moves to the appropriate position, the folding arm mechanism (21) unfolds, and at the same time the image analysis module obtains the vibration point based on the image captured by the binocular camera (24) using deep learning; Align the immersion vibrator (1) with the vibration point, release the line using the cable winding mechanism (5), and insert the immersion vibrator (1) into the concrete; Using the distance to the concrete surface obtained by the laser rangefinder (61), the immersion vibrator (1) is placed to the required depth; The vibration quality evaluation module analyzes the concrete during the vibration process based on the images captured by the binocular camera (24) until the vibration is qualified.

10. The intelligent control method for a nuclear power plant shear wall vibration robot according to claim 9, characterized in that, The image analysis module performs distortion correction and image stitching on the images captured by the 360° camera to generate a circumferential panoramic view of the environment, and identifies the distribution area of ​​vertical steel bars and the boundary of the template based on the panoramic view. The control system (6) calculates the robot's walking path based on the identified vertical steel bar position and the walking path planning module of the main frame (2), and drives the motor along the limit guide rail (26) to automatically move to the target area through the walking wheel (27); During the unfolding of the articulated arm mechanism (21), the control system (6) performs closed-loop control based on the real-time signal from the angle sensor and the distance information measured by the laser rangefinder (61) to ensure that the end posture of the insert vibrator (1) remains vertical and aligned with the target vibration point.

Images