Detection method for intelligent steel reinforcement framework mesh

Abstract

The invention discloses a reinforcement cage intelligent mesh detection method, which comprises the following steps of: obtaining a two-dimensional image of a reinforcement mesh through an industrial camera, and synchronously obtaining three-dimensional point cloud data by using LiDAR; establishing an internal and external parameter mapping relation between the industrial camera and the LiDAR based on a Zhang's calibration method, and fusing two-dimensional image features and three-dimensional point cloud features into a multi-modal feature vector; an improved FS-DETR model is adopted to carry out cross point detection, and the model dynamically adjusts an NMS threshold alpha t through reinforcement learning; the detected steel bar spacing deviation delta is input into a welding parameter adjusting module, and the mapping relation between the spacing deviation delta and the welding current I and between the spacing deviation delta and the welding pressure P is established; the detection model is compressed through a model quantification and pruning algorithm, and is deployed in an edge computing device to realize real-time reasoning. According to the invention, the industrial camera and the LiDAR synchronously collect data and fuse the three-dimensional features, so that high-precision detection in a complex scene is realized, and the problems of shielding, uneven illumination and the like are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of quality inspection of steel bar mesh sheets, and particularly to a detection method for steel bar framework intelligent mesh sheets. Background Technique

[0002] As a core component of high-speed rail precast box girders, the dimensional accuracy of steel bar mesh sheets directly affects structural safety. Traditional manual inspection has low efficiency. It takes 3 hours to inspect a 32m mesh sheet, and relying on off-line inspection results in high rework costs. Automated inspection technologies based on computer vision have been gradually popularized, such as the YOLO algorithm, LiDAR, FS-DETR model, etc. However, the YOLO algorithm has a missed detection rate of 8% under low contrast, and the intersection point positioning error > 2mm; the existing LiDAR is only used for auxiliary ranging and does not participate in feature fusion, resulting in a three-dimensional coordinate conversion error > 1.5mm; the FS-DETR model takes 180 minutes for block detection of a 32m mesh sheet, which cannot meet the production line beat requirements.

[0003] In addition, the following problems are also faced:

[0004] Limitations of single-modal detection: Relying only on two-dimensional vision or laser scanning, it cannot cope with complex lighting scenarios, such as workshop shadows and occlusion scenarios, such as dense steel bar overlaps, and the detection error reaches 2.15 - 5.36mm.

[0005] Poor model adaptability: The detection accuracy of the fixed-parameter model drops sharply in the dense intersection point scenario, and the AP value is only 0.85.

[0006] High cost of off-line inspection: Inspection after welding results in a rework rate of 5%, and the annual loss of a single beam yard exceeds 150,000 yuan, and production parameters cannot be corrected in real time. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a detection method for steel bar framework intelligent mesh sheets.

[0008] To solve the above technical problems, the present invention provides the following technical solutions:

[0009] The present invention provides a detection method for steel bar framework intelligent mesh sheets, including the following steps:

[0010] Step 1: Multi-modal data acquisition: Obtain the two-dimensional image of the steel bar mesh sheet through an industrial camera, and simultaneously obtain the three-dimensional point cloud data by using LiDAR, and the two are triggered synchronously through hardware;

[0011] Step 2: Multi-modal feature fusion: Based on the Zhang's calibration method, establish the internal and external parameter mapping relationship between the industrial camera and LiDAR, and fuse the two-dimensional image features and three-dimensional point cloud features into a multi-modal feature vector;

[0012] Step 3: Dynamic model detection: An improved FS-DETR model is used for intersection detection, and the model dynamically adjusts the NMS threshold α through reinforcement learning t , and the formula is:

[0013] α t = α0 + γ·ΔL t ;

[0014] where α t is the NMS threshold for the t-th iteration, α0 is the initial threshold, γ is the learning rate, and ΔL t is the current loss change;

[0015] Step 4: Real-time quality feedback: The detected steel bar spacing deviation δ is input into the welding parameter adjustment module, and the mapping relationship between the spacing deviation δ and the welding current I and welding pressure P is established. The formula is:

[0016]

[0017] where I new is the adjusted welding current, I old is the original current, d max is the maximum allowable spacing deviation, P new is the adjusted welding pressure, P old is the original pressure, d max = 10mm, and the adjustment response time ≤ 100ms;

[0018] Step 5: Lightweight edge computing: The detection model is compressed to ≤ 10MB through model quantization and pruning algorithms and deployed on the edge computing device to achieve real-time inference.

[0019] As a preferred technical solution of the present invention, the multi-modal feature fusion in the above step 2 includes:

[0020] Perform non-maximum suppression on the two-dimensional image to generate a heat map of the intersection position;

[0021] Perform voxel filtering and plane segmentation on the LiDAR point cloud to extract the steel bar contour point cloud;

[0022] Align the heat map coordinates and the point cloud coordinates through a spatial transformation matrix. The formula is:

[0023] P cloud = R·P image + T;

[0024] where R is the rotation matrix, T is the translation vector, P cloud is the image coordinate, and P image is the point cloud coordinate;

[0025] Establish the conversion relationship between the three-dimensional coordinates of the intersection point and the actual size, and the conversion formula is:

[0026]

[0027] Where d real is the actual distance, d pixel is the pixel distance, f is the camera focal length, and s is the pixel size.

[0028] As a preferred technical solution of the present invention, the improved FS-DETR model in step three includes:

[0029] Introduce a deformable attention module in the Transformer encoder to dynamically adjust the number of sampling points for each query vector;

[0030] Optimize the intersection point detection accuracy through a contrastive learning loss function, and the formula is:

[0031]

[0032] Where f i and f j are the positive and negative sample feature vectors, and τ is the temperature parameter.

[0033] As a preferred technical solution of the present invention, the reinforcement learning strategy for dynamically adjusting the NMS threshold α t is:

[0034] Use the PPO algorithm to optimize the threshold, and the reward function R is defined as:

[0035] R = AP current -AP previous + 0.1·ΔT;

[0036] Where AP is the mean average precision, ΔT is the change in inference time, and the convergence threshold of the reward function R is 0.02.

[0037] As a preferred technical solution of the present invention, the welding parameter adjustment module in step four includes:

[0038] When it is continuously detected that δ > 5mm three times, trigger the welding current I to increase by 5%;

[0039] When δ > 10mm, trigger the device to stop and alarm, and at the same time generate a rework guidance document.

[0040] As a preferred technical solution of the present invention, the edge computing device in step five includes:

[0041] A processor based on the ARM architecture, integrated with a TensorRT acceleration engine;

[0042] The floating - point precision of the model is compressed from FP32 to INT8 using the mixed - precision quantization algorithm, and the formula is:

[0043]

[0044] Redundant convolutional kernels are deleted through the channel pruning algorithm, and the pruning rate is ≥ 80%.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] The present invention synchronously collects data through an industrial camera and LiDAR and fuses three - dimensional features to achieve high - precision detection in complex scenarios, effectively solving problems such as occlusion and uneven illumination.

[0047] The present invention introduces a deformable attention module and a contrastive learning mechanism to improve the feature extraction ability of dense intersection points, and combines reinforcement learning to dynamically adjust detection parameters, enhancing the adaptability of the model to different mesh densities and illumination conditions.

[0048] The detection results of the present invention are mapped to welding parameters in real - time, and the production parameters are dynamically corrected through a proportional adjustment algorithm, reducing the rework cost and improving the qualification rate.

[0049] The model quantization and pruning technology of the present invention compresses the volume of the detection model, and combines with an edge computing device with an ARM architecture to achieve portable real - time inference, reducing the hardware cost and deployment complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0051] Figure 1 is the flowchart of the detection method of the present invention;

[0052] Figure 2 is the schematic diagram of the connection of the hardware system of the present invention;

[0053] Figure 3 is the flowchart of the training of the improved FS - DETR model of the present invention;

[0054] Figure 4 is the flowchart of the compression of the lightweight edge - computing model of the present invention;

[0055] Figure 5 is the flowchart of Zhang's calibration and coordinate alignment of the present invention;

[0056] Figure 6 is the flowchart of real - time closed - loop control of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0058] Embodiment

[0059] As Figure 1-6 shown, the present invention provides a method for detecting a steel bar skeleton intelligent manufacturing mesh, including the following steps:

[0060] Step 1: Multimodal data acquisition: Obtain a two-dimensional image of the steel bar mesh through an industrial camera with a frame rate of 60fps, and simultaneously obtain three-dimensional point cloud data using LiDAR, with a point cloud density ≥ 100 points / cm 2 , and the two are triggered synchronously by hardware. The LiDAR is 16 lines, and the scanning range is 0 - 20m;

[0061] Step 2: Multimodal feature fusion: Establish the internal and external parameter mapping relationship between the industrial camera and LiDAR based on the Zhang's calibration method, and fuse the two-dimensional image features and three-dimensional point cloud features into a multimodal feature vector;

[0062] Step 3: Dynamic model detection: Use an improved FS-DETR model for intersection detection, and the model dynamically adjusts the NMS threshold α through reinforcement learning t , and the formula is:

[0063] α t = α0 + γ·ΔL t ,

[0064] where α t is the NMS threshold for the t-th iteration, α0 is the initial threshold, γ is the learning rate, γ = 0.1, and ΔL t is the current loss change amount, and the detection AP ≥ 0.98 in complex lighting with an illumination intensity ≤ 500lux and a dense intersection scene with a spacing ≤ 50mm;

[0065] Step 4: Real-time quality feedback: Input the detected steel bar spacing deviation δ into the welding parameter adjustment module, and establish the mapping relationship between the spacing deviation δ and the welding current I and welding pressure P. The formula is:

[0066]

[0067] where I new is the adjusted welding current, I old is the original current, d max is the maximum allowable spacing deviation, P new is the adjusted welding pressure, P old is the original pressure, d max= 10mm, adjust the response time ≤ 100ms;

[0068] Step Five: Lightweight edge computing: Compress the detection model to ≤ 10MB through model quantization and pruning algorithms, deploy it on the edge computing device to achieve real-time inference, and the inference time for a single image ≤ 20ms.

[0069] Furthermore, the multi-modal feature fusion in Step Two includes:

[0070] Perform non-maximum suppression on the two-dimensional image, with the NMS threshold dynamic range of 0.3 - 0.7, and generate a heat map of the intersection point positions;

[0071] Perform voxel filtering and plane segmentation on the LiDAR point cloud to extract the rebar contour point cloud;

[0072] Align the heat map coordinates with the point cloud coordinates through a spatial transformation matrix, with the conversion error ≤ 0.8mm, and the formula is:

[0073] P cloud = R · P image + T;

[0074] where, R is the rotation matrix, T is the translation vector, P cloud is the image coordinate, P image is the point cloud coordinate;

[0075] And establish the conversion relationship between the three-dimensional coordinates of the intersection point and the actual size, and the formula is:

[0076]

[0077] where, d real is the actual distance, d pixel is the pixel distance, f is the camera focal length, s is the pixel size, the camera focal length f = 12mm, the pixel size s = 3.45μm, and the conversion error ≤ 0.5mm.

[0078] Furthermore, the improved FS-DETR model in Step Three includes:

[0079] Introduce a deformable attention module in the Transformer encoder to dynamically adjust the number of sampling points for each query vector, with the range of 8 - 16 points;

[0080] Optimize the intersection point detection accuracy through a contrastive learning loss function, and the formula is:

[0081]

[0082] where, f i and f j are the positive and negative sample feature vectors, τ is the temperature parameter, τ = 0.07, and the positive and negative sample feature vectors fi and f j has a cosine similarity ≥ 0.85.

[0083] Furthermore, in step three, the NMS threshold α is dynamically adjusted t using the following reinforcement learning strategy:

[0084] The PPO algorithm is used to optimize the threshold, and the reward function R is defined as:

[0085] R = AP current - AP previous + 0.1·ΔT;

[0086] where AP is the mean average precision, ΔT is the change in inference time, and the convergence threshold of the reward function R is 0.02.

[0087] Furthermore, in step four, the welding parameter adjustment module includes:

[0088] When δ > 5mm is detected continuously for 3 times, the welding current I is increased by 5%;

[0089] When δ > 10mm, the device is triggered to stop and alarm, and at the same time, a rework guidance file is generated.

[0090] Furthermore, in step five, the edge computing device includes:

[0091] A processor based on the ARM architecture, model number RK3588, integrated with the TensorRT acceleration engine, model number TensorRT8.5;

[0092] The model floating-point precision is compressed from FP32 to INT8 using the mixed-precision quantization algorithm, and the formula is:

[0093]

[0094] Redundant convolution kernels are deleted through the channel pruning algorithm, and the pruning rate ≥ 80%.

[0095] Specifically, the detailed steps of a method for detecting steel bar skeleton manufacturing mesh sheets:

[0096] I. Multi-modal data collection

[0097] 1.1. Hardware layout

[0098] Industrial cameras: Two MV-CS016-10GC industrial cameras are selected. Their resolution is 1440×1080, the frame rate reaches 60fps, and the focal length is 12mm. These two cameras are installed on both sides above the detection area of the steel mesh at appropriate angles and heights to ensure that the camera's field of view can completely cover the super-large steel mesh of 32m×10m, and the relative positions between the cameras should be fixed to facilitate subsequent multi-modal data fusion processing.

[0099] LiDAR: A 16-line VLP-16 LiDAR is adopted, with a scanning frequency of 10Hz and a ranging error ≤2mm. The LiDAR is installed above the center of the detection area at a height of 5m, which can ensure that its scanning range can cover the entire steel mesh and obtain relatively uniform and dense point cloud data at the same time.

[0100] Synchronous triggering device: Use hardware I / O signals to achieve synchronous triggering of the industrial camera and LiDAR. Connect it to the camera and LiDAR through a dedicated synchronous controller to ensure that both start data acquisition at the same moment, and the time deviation is controlled within ≤50μs.

[0101] 1.2. Acquisition parameter settings

[0102] Working distance: Set the working distances of both the industrial camera and LiDAR to 5m. This distance can not only ensure that the camera obtains clear two-dimensional images but also enable the LiDAR to obtain point cloud data with sufficient density to cover the entire 32m-long mesh.

[0103] Point cloud density: To more accurately extract features such as steel bar contours later, by adjusting the scanning parameters and filtering settings of the LiDAR, the point cloud density is made to reach 100 points / cm 2 , and at the same time, voxel filtering is adopted, and the voxel size is set to 5mm to remove noise points and reduce the data volume.

[0104] II. Multi-modal feature fusion (as Figure 5 shown)

[0105] 2.1. Zhang's calibration process

[0106] Preparation work: Make a checkerboard target with a specification of 10×10 corner points, and the side length of each corner point is 25mm. Place the target at different positions and angles within the detection area of the steel mesh to ensure that both the industrial camera and LiDAR can clearly capture the target.

[0107] Data acquisition: Take images of the checkerboard target at different positions and angles through the industrial camera, and at the same time use the LiDAR to obtain the corresponding point cloud data. The number of collected data is not less than 20 groups to ensure the accuracy of subsequent calibration.

[0108] Parameter calculation: Using Zhang's calibration algorithm, the collected image and point cloud data are processed to calculate the rotation matrix R and translation vector T between the industrial camera and the LiDAR. During the calculation process, the parameters are continuously iteratively optimized to control the calibration error within ≤0.5mm.

[0109] 2.2. Feature extraction

[0110] Visual feature extraction: Use the YOLOv5 algorithm to process the two-dimensional images collected by the industrial camera and extract the heat map of the intersection points. During this process, the non-maximum suppression (NMS) method is used to remove redundant detection boxes, and the threshold of NMS is set to a dynamic range of 0.3 - 0.7 to adapt to different complex image scenes.

[0111] Point cloud feature extraction: The point cloud data collected by the LiDAR is processed using the DBSCAN clustering algorithm. Set the clustering parameters ε = 10mm and MinPts = 5. Through this algorithm, the point cloud data is clustered into different regions, thereby extracting the contour point cloud of the steel bars.

[0112] 2.3. Coordinate alignment and transformation

[0113] Coordinate alignment: According to the previously calculated rotation matrix R and translation vector T, use the spatial transformation formula P cloud = R·P image + T to align the coordinates of the visual features and point cloud features. During the alignment process, the parameters are continuously adjusted to control the alignment error within ≤0.8mm.

[0114] Actual size conversion: In order to convert the pixel size in the image to the actual physical size, use the formula for conversion.

[0115] where d real is the pixel distance measured in the image, 12 is the focal length (mm) of the camera, and 3.45×10 -3 is the pixel size (mm). During the conversion process, through multiple measurements and calibrations, ensure that the conversion error is controlled within ≤0.5mm.

[0116] III. Dynamic model detection (as Figure 3 shown)

[0117] 3.1. Improved FS-DETR model construction

[0118] Deformable Attention Module: Introduce a deformable attention module into the Transformer encoder of the FS-DETR model. This module can dynamically adjust the number of sampling points for each query vector according to different input features, and the range of the number of sampling points is set to 8 - 16 points. In this way, the model can better adapt to the scenarios of steel bar intersections with different densities and improve the detection accuracy.

[0119] Contrastive Learning Loss Function: To enhance the model's ability to distinguish intersection features, use a contrastive learning loss function, and the formula is τ = 0.07. Through this loss function, the cosine similarity between the positive and negative sample feature vectors f i and f j is ≥ 0.85, thereby improving the model's detection performance in complex scenarios.

[0120] Reinforcement Learning Strategy: Adopt the Proximal Policy Optimization (PPO) algorithm to dynamically adjust the threshold α of non-maximum suppression (NMS) t . The reward function is defined as R = AP current - AP previous + 0.1·ΔT. By continuously optimizing the reward function, the model can improve the inference speed while ensuring the detection accuracy.

[0121] 3.2. Experimental Tests and Evaluations

[0122] Light Intensity Test: Test the improved FS-DETR model in an environment with a light intensity of 200 lux. Use multiple groups of different steel bar mesh samples for detection and record the average precision (AP) of the detection. After multiple experiments, the detection AP of the model under this light intensity is 0.98, while the AP of the existing technology under the same light conditions is only 0.85.

[0123] Intersection Spacing Test: Test the model for the dense intersection scenario with an intersection spacing of 50 mm. By comparing the recall rates of different models in this scenario, it is found that the recall rate of the model of the present invention reaches 99.2%, while the recall rate of the existing technology is only 92.5%.

[0124] Inference Time Test: Test the inference time of a single image. Divide the 32m-long mesh into blocks, use the improved model to detect each block of the image, and record the total inference time. After multiple experiments, the inference time of a single image ≤ 20 ms, and the total time for the block detection of the entire 32m mesh ≤ 25 minutes, while the existing technology takes 180 minutes to detect a mesh of the same size.

[0125] IV. Real-time Quality Feedback (as Figure 6 shown)

[0126] 4.1. Welding parameter adjustment strategy parameter mapping relationship: Establish the mapping relationship between the steel bar spacing deviation δ and the welding current I and pressure P. The specific formula is:

[0127] Where I old and P old are the current welding current and pressure, and I new and P new are the adjusted welding current and pressure.

[0128] Response time control: By optimizing the communication protocol and data transmission process between the detection system and the welding control system, ensure that when the steel bar spacing deviation δ is detected, the adjustment response time of the welding parameters ≤ 100 ms.

[0129] Threshold trigger mechanism: Set two thresholds to trigger different operations. When δ > 5 mm is detected continuously for 3 times, automatically increase the welding current by 5%; when δ > 10 mm, immediately trigger the equipment to stop and send an alarm signal, and at the same time generate a detailed rework guidance document, which contains information such as the deviation position and deviation size for the workers to carry out rework.

[0130] 4.2. Experimental verification and effect evaluation

[0131] Qualified rate statistics: Conduct detection and welding parameter adjustment experiments on 300 steel bar mesh sheets, and count the number of qualified mesh sheets. After the experiment, it is found that after adopting the real-time quality feedback system of the present invention, the qualified rate of the mesh sheets reaches 99.3%, while the qualified rate of the existing technology is only 95%.

[0132] Cost saving calculation: According to the experimental data, calculate the cost saving situation of the single beam field after adopting the system of the present invention. Calculated at 12,000 yuan per mesh sheet material cost, the rework rate of the existing technology is 5%, then the annual rework cost is 150,000 yuan; while after adopting the system of the present invention, the rework rate drops to 0.7%, and the annual rework cost is only 21,000 yuan, and the single beam field saves 2.1 million yuan annually.

[0133] V. Lightweight edge computing (as Figure 4 shown)

[0134] 5.1. Hardware selection and configuration

[0135] Processor: Select the RK3588 processor, which adopts an 8-core ARM architecture and has high computing performance and low power consumption.

[0136] Model Compression: The detection model is compressed by using a method of mixed-precision quantization and channel pruning with a pruning rate of 85%. Through these methods, the model size is compressed from the original 120MB to 7.8MB, and at the same time, the TensorRT 8.5 acceleration engine is used to accelerate the model.

[0137] 5.2. Performance Index Testing

[0138] Floating-point Operations (FLOPs) Testing: A professional performance testing tool is used to test the FLOPs of the compressed model. After testing, it is found that the FLOPs of the model are reduced from the original 25 GFLOPs to 3.2 GFLOPs, greatly reducing the computational volume.

[0139] Memory Occupancy Testing: During the operation of the detection model, the memory occupancy of the system is monitored in real time. After testing, it is found that after adopting the lightweight edge computing solution, the memory occupancy of the system is ≤150MB, while the memory occupancy of the existing technology using the fixed industrial computer solution is as high as 1.2GB.

[0140] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A detection method for a steel bar skeleton intelligent mesh sheet, characterized in that, It includes the following steps: Step 1: Multimodal data acquisition: Obtain the 2D image of the steel bar mesh through an industrial camera, and simultaneously acquire the 3D point cloud data using LiDAR. The two are triggered synchronously by hardware; Step 2: Multimodal feature fusion: Establish the internal and external parameter mapping relationship between the industrial camera and LiDAR based on the Zhang's calibration method, and fuse the 2D image features and 3D point cloud features into a multimodal feature vector; Step 3: Dynamic model detection: An improved FS-DETR model is used for intersection detection, and the model dynamically adjusts the NMS threshold α through reinforcement learning t , and the formula is: α t = α0 + γ·ΔL t ; where α t is the NMS threshold for the t-th iteration, α0 is the initial threshold, γ is the learning rate, and ΔL t is the current loss change; Step 4: Real-time quality feedback: Input the detected steel bar spacing deviation δ into the welding parameter adjustment module, and establish the mapping relationship between the spacing deviation δ and the welding current I and welding pressure P. The formula is: Among them, I new is the adjusted welding current, and I old is the original current, d max is the maximum allowable spacing deviation, P new is the adjusted welding pressure, and P old is the original pressure, d max = 10 mm, and the adjustment response time ≤ 100 ms; Step 5: Lightweight edge computing: Compress the detection model to ≤10MB through model quantization and pruning algorithms, and deploy it on the edge computing device to achieve real-time inference.

2. The inspection method for a steel bar skeleton intelligent manufacturing mesh sheet according to claim 1, wherein The multimodal feature fusion in Step 2 includes: Perform non-maximum suppression on the 2D image to generate a heat map of the intersection point positions; Perform voxel filtering and plane segmentation on the LiDAR point cloud to extract the steel bar contour point cloud; Align the heat map coordinates and point cloud coordinates through a spatial transformation matrix. The formula is: P cloud = R·P image + T; where R is the rotation matrix, T is the translation vector, P cloud is the image coordinate, and P image is the point cloud coordinate; Establish the conversion relationship between the 3D coordinates of the intersection point and the actual size. The conversion formula is: Among them, d real is the actual distance, d pixel is the pixel distance, f is the camera focal length, and s is the pixel size.

3. The inspection method of a steel bar framework intelligent manufacturing mesh sheet according to claim 1, characterized in that, The improved FS-DETR model in Step 3 includes: Introduce a deformable attention module in the Transformer encoder to dynamically adjust the number of sampling points for each query vector; Optimize the intersection point detection accuracy through a contrastive learning loss function. The formula is: Among them, f i and f j are positive and negative sample feature vectors, and τ is a temperature parameter.

4. A method for detecting a steel bar skeleton intelligent mesh sheet according to claim 3, characterized in that, In step 3, dynamically adjust the NMS threshold α t The reinforcement learning strategy is as follows: Use the PPO algorithm to optimize the threshold, and the reward function R is defined as: R = AP current -AP previous +0.1·ΔT; Where, AP is the mean average precision, ΔT is the change in inference time, and the convergence threshold of the reward function R is 0.

02.

5. The inspection method of the steel bar framework intelligent manufacturing mesh sheet according to claim 1, characterized in that The welding parameter adjustment module in Step 4 includes: When it is detected that δ > 5mm continuously for 3 times, trigger the welding current I to increase by 5%; When δ > 10mm, trigger the device to stop and alarm, and at the same time generate a rework guidance document.

6. A method for detecting a steel bar skeleton intelligent mesh sheet according to claim 1, characterized in that, The edge computing device in Step 5 includes: A processor based on the ARM architecture, integrated with a TensorRT acceleration engine; Use a mixed-precision quantization algorithm to compress the model floating-point precision from FP32 to INT8. The formula is: Delete redundant convolution kernels through a channel pruning algorithm, and the pruning rate is ≥80%.

Images