Real-time code spraying and position retesting method and system of fixed testing robot in signal construction

Abstract

The invention provides a real-time code spraying and position retesting method and system for a positioning and testing robot in signal construction, and relates to the technical field of robot positioning and testing, and the method comprises the steps: obtaining the three-dimensional point cloud data of a construction area, and generating a grid map; identifying a positioning mark point and planning an optimal motion track; controlling the fixed measurement robot to move and keeping the code spraying device horizontal; when a preset position is reached, the height of the nozzle is adjusted and code spraying is performed; collecting a code spraying image and extracting contour features; comparing the actual sprayed code with a preset template, and calculating a deviation; optimizing the motion parameter when the deviation exceeds a threshold; and uploading the deviation data and generating a quality report. The method improves the accuracy and automation degree of signal construction.

Description

Technical Field

[0001] The present invention relates to the field of robot measurement technology, and in particular to a real-time coding and position re-measurement method and system for a measurement robot in signal construction. Background Art

[0002] As the scale of railway construction continues to expand, the fixed measurement marks in signal engineering construction are an important benchmark to ensure the track alignment and signal equipment installation accuracy. The coding operation of the fixed measurement marks directly affects the positioning accuracy and construction quality of the subsequent construction links. At present, railway engineering construction has put forward higher requirements for the position accuracy and construction efficiency of fixed measurement marks, and automated and intelligent fixed measurement coding technology has gradually become an industry development trend.

[0003] The existing fixed-measurement marking coding mainly adopts manual handheld inkjet printers or simple automatic coding equipment. These methods have many shortcomings: the accuracy of manual coding depends on the experience of the operator, and it is difficult to maintain a stable coding posture and uniform spraying effect; the existing automatic coding equipment lacks intelligent navigation and posture adjustment functions, and cannot adapt to complex construction environments; at the same time, the lack of real-time quality inspection and data management means makes it difficult to ensure the construction quality, and increases the cost of rework and maintenance in the later stage.

[0004] To sum up, there is an urgent need to develop an intelligent surveying and marking system that integrates robot technology, computer vision and artificial intelligence to realize automatic surveying and path planning of construction areas, precise coding position control, real-time posture adjustment and online detection of coding quality, improve the construction efficiency and quality of surveying and marking, and enhance the automation level and quality management capabilities of railway engineering construction. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for real-time coding and position re-measurement of a measurement robot in signal construction, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for real-time coding and position re-measurement of a fixed-measurement robot in signal construction is provided, comprising:

[0008] The construction area is scanned by a surveying robot to obtain three-dimensional point cloud data of the construction area, and the construction area is rasterized based on the three-dimensional point cloud data to generate a raster map. The image information of the construction area is collected by the surveying robot, and the positioning mark points in the construction area are identified by an image recognition algorithm. The spatial coordinates of the positioning mark points are mapped to the raster map, the distribution of the positioning mark points in the raster map is determined, and the optimal motion trajectory of the surveying robot is planned by using a time-series reinforcement iterative neural network algorithm.

[0009] The measurement robot is controlled to move along the optimal motion trajectory, and the measurement posture data is collected in real time through the inertial measurement unit of the measurement robot. Based on the measurement posture data, the inkjet device of the measurement robot is controlled to maintain a horizontal state. When the measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted. The image information of the preset inkjet position is collected through the visual positioning system, and the inkjet device is controlled to perform the inkjet operation in combination with the preset inkjet template.

[0010] After the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

[0011] In an optional embodiment,

[0012] Scanning the construction area with a fixed-measurement robot to obtain three-dimensional point cloud data of the construction area, and rasterizing the construction area based on the three-dimensional point cloud data to generate a raster map, including:

[0013] Scanning the construction area with a fixed measurement robot to obtain three-dimensional point cloud data of the construction area, calculating the statistical mean and standard deviation of the three-dimensional point cloud data, determining a valid data range, removing abnormal data outside the valid data range, obtaining filtered three-dimensional point cloud data, dividing the filtered three-dimensional point cloud data into voxels according to a preset voxel size, calculating the centroid coordinates of the three-dimensional points contained in each voxel, and using the centroid coordinates as downsampling points of the corresponding voxels to obtain downsampled three-dimensional point cloud data;

[0014] According to a preset radius, select neighboring 3D points within a corresponding radius range of each 3D point in the downsampled 3D point cloud data, construct a covariance matrix based on the neighboring 3D points, calculate the normal vector of each 3D point through the covariance matrix, determine the vector difference between the normal vectors, use the median of the vector difference as a Gaussian kernel parameter, construct a Gaussian kernel function based on the Gaussian kernel parameter, and calculate the significance score of each 3D point;

[0015] The significance scores are sorted in descending order, and the top thirty percent of the three-dimensional points in the significance score sorting are selected as candidate feature points. The spacing between the candidate feature points is determined, and the candidate feature points whose spacing is less than a preset spacing threshold are eliminated to obtain a feature point set; the point feature histogram descriptor of each feature point in the feature point set is calculated, and feature matching is performed to obtain an initial registration result;

[0016] Calculate the distance value and the normal angle value between the corresponding feature points in the initial registration result, use the median value of the distance value as the distance parameter, and the median value of the normal angle value as the angle parameter; construct a distance weight item based on the distance parameter, and construct an angle weight item based on the angle parameter, multiply the distance weight item by the angle weight item to obtain an adaptive weight, and use the adaptive weight to perform iterative optimization to obtain an accurate registration result;

[0017] The construction area is divided into grids of a preset size, the distance between each grid and the three-dimensional point cloud data in the precise registration result is calculated, the observation likelihood of the Gaussian distribution is calculated, the observation likelihood is multiplied by the prior probability of the corresponding grid to obtain the probability product, the reciprocal of the cumulative sum of the probability products is calculated as a normalization factor, the probability product is normalized using the normalization factor to obtain the state probability of each grid, each grid is marked as occupied, idle or unknown based on the state probability, and a grid map of the construction area is generated.

[0018] In an optional embodiment, the positioning mark points in the construction area are identified by an image recognition algorithm, and the spatial coordinates of the positioning mark points are mapped to a grid map. The distribution of the positioning mark points in the grid map is determined, including:

[0019] The image collected by the measurement robot is input into the backbone feature network, and the backbone feature network extracts image features through the residual module. In the residual module, the spatial dimension and channel dimension of the feature map are weightedly reconstructed by using the spatial attention mechanism and the channel attention mechanism, respectively, to obtain the first-layer feature map; the first-layer feature map is input into the feature pyramid structure, and different scale features are fused through upsampling and downsampling operations to obtain a multi-level feature map;

[0020] The multi-level feature map is input into the detection head network, and the detection head network outputs the category score and position coordinates of the positioning mark point. Based on the category score and the position coordinates, the two-dimensional pixel coordinates of the positioning mark point in the image are determined; based on the internal parameter matrix of the binocular camera, the image is distorted and corrected, and the parallax information of the left and right images is calculated. In combination with the external parameter matrix of the binocular camera, the two-dimensional pixel coordinates are converted into depth information through the triangulation principle, and the two-dimensional pixel coordinates are combined with the depth information to obtain the three-dimensional space coordinates of the positioning mark point;

[0021] The corresponding feature points in the coordinate system corresponding to the camera coordinate system and the grid map are extracted, and feature point matching pairs are constructed. The rotation matrix and the translation vector are calculated, and the rotation matrix and the translation vector are optimized using a nonlinear optimization algorithm to obtain a coordinate transformation matrix. The three-dimensional space coordinates are mapped to the coordinate system of the grid map through the coordinate transformation matrix to obtain the spatial distribution of the positioning mark points.

[0022] In an optional embodiment,

[0023] Using the time-series enhanced iterative neural network algorithm, the optimal motion trajectory of the robot is planned and measured, including:

[0024] Constructing a time-series reinforcement iterative neural network, combining the real-time position information of the measuring robot, the spatial distribution of the positioning mark points and the local information of the grid map into a state vector, and combining the discretized linear velocity and angular velocity into an action vector, wherein the time-series reinforcement iterative neural network includes a main action network and a target value network;

[0025] The main action network selects an action vector based on the current state vector, executes the action vector to obtain the next state vector, and the target value network estimates the value based on the next state vector; calculates the path length value, motion smoothness value and marker coverage value of the action vector, and combines them to form a reward value;

[0026] The current state vector, action vector, reward value and next state vector are used to construct state transfer data, the time difference error of the state transfer data is calculated, priorities are assigned to the state transfer data based on the time difference error, and the state transfer data of different priorities are stored in an experience replay pool; state transfer data are sampled from the experience replay pool according to priority to train the main action network, the parameters of the main action network are updated to the target value network according to a preset number of steps, iterations are repeated until convergence, and the optimal motion trajectory for the measurement robot to visit the positioning mark point is output.

[0027] In an optional embodiment,

[0028] Controlling the measurement robot to move along the optimal motion trajectory, collecting measurement posture data in real time through the inertial measurement unit of the measurement robot, controlling the inkjet device of the measurement robot to maintain a horizontal state based on the measurement posture data, measuring and automatically adjusting the vertical distance between the nozzle of the inkjet device and the ground when the measurement robot moves to a preset inkjet position, collecting image information of the preset inkjet position through a visual positioning system, and controlling the inkjet device to perform inkjet operation in combination with a preset inkjet template, including:

[0029] The inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by an extended Kalman filter algorithm, and the roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion;

[0030] Based on the attitude quaternion, a kinematic model of the inkjet printer is established, the target angle of the first axis and the target angle of the second axis of the two-degree-of-freedom gimbal are calculated, the stepper motors of the first axis and the second axis are controlled to rotate respectively by a PID controller, and the inkjet printer is set in a horizontal state;

[0031] When the preset coding position is reached, the vertical distance between the nozzle and the ground is measured by a laser distance sensor, and the vertical distance is compared with the preset vertical distance boundary to obtain the distance deviation value and the distance deviation change rate, establish a fuzzy control rule, input the fuzzy control rule into the fuzzy PID controller, adaptively adjust the PID parameters, control the speed of the DC servo motor of the lifting mechanism, and adjust the nozzle height;

[0032] Collect an image of a preset inkjet printing position, perform grayscale and edge detection processing on the image to obtain contour features, perform feature point matching between a preset inkjet printing template and the contour features, determine the inkjet printing position coordinates and the spraying direction angle based on the matching results, generate a spraying instruction sequence, the spraying instruction sequence includes a spraying instruction sequence of nozzle position information, spraying pressure information and spraying time information, and send the spraying instruction sequence to the inkjet printing controller via a field bus;

[0033] During the inkjet printing process, the spraying image is collected in real time, and the detection parameters of the spraying lines in the spraying image are analyzed, the detection parameters including width value, continuity value and clarity value, and the analysis results are generated. When any one of the detection parameters is lower than the corresponding preset threshold value, the spraying pressure information and the spraying time information are adaptively adjusted based on the analysis result until each detection parameter reaches the corresponding preset threshold value, and the inkjet printing operation is completed.

[0034] In an optional embodiment,

[0035] The inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by the extended Kalman filter algorithm. The roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion, including:

[0036] The three-axis acceleration sensor of the inertial measurement unit is used to collect the acceleration components of gravity acceleration in three orthogonal axes, and the three-axis gyroscope is used to collect the angular velocity components of the robot rotating around the three orthogonal axes, and the acceleration components and the angular velocity components are sampled and filtered through the signal acquisition module;

[0037] A system state equation is established based on the angular velocity component, the attitude quaternion is set as the state vector, the angular velocity is set as the system input, and a nonlinear state transfer model is constructed; an observation equation is established based on the acceleration component, the projection of the gravity vector in the body coordinate system is set as the observation quantity, and a nonlinear observation model is constructed according to the relationship between the attitude quaternion and the gravity vector;

[0038] Performing a first-order Taylor expansion on the nonlinear state transfer model to obtain a state Jacobian matrix, and performing a first-order Taylor expansion on the nonlinear observation model to obtain an observation Jacobian matrix;

[0039] The extended Kalman filter algorithm is used to estimate the attitude, including: using the state Jacobian matrix and the nonlinear state transfer model to perform state prediction to obtain a predicted attitude quaternion and a predicted state covariance; using the observation Jacobian matrix, the nonlinear observation model and the predicted state covariance to calculate the Kalman gain, multiplying the Kalman gain with the observation residual to obtain a state correction, adding the state correction to the predicted attitude quaternion to obtain an estimated attitude quaternion, and updating the state covariance at the same time;

[0040] The estimated quaternion is normalized, and each component is divided by the corresponding modulus to obtain a unit attitude quaternion; according to the four components of the unit attitude quaternion, three rotation angles describing the robot attitude are calculated through a conversion formula from the unit attitude quaternion to the Euler angle, including a roll angle around the X-axis, a pitch angle around the Y-axis, and a yaw angle around the Z-axis.

[0041] In an optional embodiment,

[0042] After the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report including:

[0043] The spectral image of the coding area is collected by a hyperspectral camera, a grid pattern is projected by a structured light projector and a structured light image is collected, and the spectral image and the structured light image are input into a synchronous trigger circuit for timing alignment to obtain standard dual-modal image data;

[0044] The standard bimodal image data is input into a spectral-spatial joint feature extraction network, the spectral attention module of the spectral-spatial joint feature extraction network processes the spectral image to obtain spectral features, the spatial pyramid pooling layer processes the structured light image to obtain spatial features, and the spectral features and the spatial features are fused through a cross-modal feature fusion module to obtain fused features of the coding area;

[0045] Perform edge detection on the fused features to obtain the edge point set of the coding pattern, input the edge point set into the dynamic graph convolution network, extract the local structure information through the adaptive graph convolution layer, and use the graph attention mechanism to weight the local structure information to obtain the contour features of the coding pattern;

[0046] The contour feature is matched with the preset coding template, and the initial posture deviation is obtained by performing a first-stage registration through a graph matching network. Based on the initial posture deviation, a second-stage registration is performed through an iterative closest point algorithm to obtain the actual posture deviation of the coding pattern;

[0047] The actual posture deviation is input into a hierarchical reinforcement learning network, a high-level strategy network of the hierarchical reinforcement learning network generates a parameter optimization strategy based on the actual posture deviation, and a low-level execution network adjusts the motion parameters of the measurement robot based on the parameter optimization strategy;

[0048] The motion parameters are input into a hybrid parameter optimizer, the hybrid parameter optimizer uses Bayesian optimization to determine a parameter search space, and an evolutionary algorithm is used to iteratively optimize the motion parameters in the parameter search space to obtain optimized motion parameters;

[0049] Based on a pre-constructed uncertainty assessment model, the optimized motion parameters are input into the uncertainty assessment model, and the measurement uncertainty introduced by the measurement process, the model uncertainty introduced by the model prediction and the parameter uncertainty introduced by the parameter optimization are calculated respectively. The measurement uncertainty, the model uncertainty and the parameter uncertainty are input into the evidence theory model to obtain the credibility evaluation value of the parameter optimization result.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provided is a real-time coding and position re-measurement system for a fixed-measurement robot in signal construction, including:

[0052] The first unit is used to scan the construction area by a surveying robot, obtain three-dimensional point cloud data of the construction area, perform rasterization processing on the construction area based on the three-dimensional point cloud data, generate a raster map, collect image information of the construction area by the surveying robot, identify positioning mark points in the construction area by an image recognition algorithm, map the spatial coordinates of the positioning mark points to the raster map, determine the distribution of the positioning mark points in the raster map, and use a time-series reinforcement iterative neural network algorithm to plan the optimal motion trajectory of the surveying robot;

[0053] The second unit is used to control the movement of the fixed measurement robot along the optimal motion trajectory, collect fixed measurement posture data in real time through the inertial measurement unit of the fixed measurement robot, and control the inkjet device of the fixed measurement robot to maintain a horizontal state based on the fixed measurement posture data. When the fixed measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted, and the image information of the preset inkjet position is collected through the visual positioning system. In combination with the preset inkjet template, the inkjet device is controlled to perform the inkjet operation;

[0054] The third unit is used to collect high-definition images of the coding area after the coding operation is completed, use image processing algorithms to extract contour features of the coding pattern, compare the contour features with the preset coding template, and calculate the actual deviation of the coding position. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

[0055] According to a third aspect of the embodiments of the present invention,

[0056] An electronic device is provided, comprising:

[0057] processor;

[0058] a memory for storing processor-executable instructions;

[0059] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0060] A fourth aspect of the embodiments of the present invention is:

[0061] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0062] In an embodiment of the present invention, by scanning the construction area to generate a grid map, using an image recognition algorithm to identify positioning markers, and combining the time-series enhanced iterative neural network algorithm to plan the optimal motion trajectory, the working efficiency of the measurement robot can be greatly improved. At the same time, the measurement posture data is collected in real time and the inkjet device is automatically adjusted to ensure the accuracy of the inkjet operation; after the inkjet operation is completed, the contour features of the inkjet pattern are extracted by the image processing algorithm and compared with the preset template to calculate the actual deviation. When the deviation exceeds the preset threshold, the system will adaptively optimize the motion parameters of the measurement robot, thereby continuously improving the construction quality. This real-time monitoring and self-optimization mechanism greatly enhances the control ability of construction quality; the actual deviation data is uploaded to the construction management system in real time, and a construction quality report is generated, providing accurate and timely data support for construction management. This intelligent data collection and analysis method helps construction management personnel to quickly understand the construction quality status and make timely decisions, thereby improving the overall intelligent level of construction management. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of a method for real-time coding and position re-measurement of a measurement robot in signal construction according to an embodiment of the present invention;

[0064] Figure 2It is a structural schematic diagram of the real-time coding and position re-measurement system of the measurement robot in signal construction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0067] Figure 1 Schematic diagram of the flow of the real-time coding and position re-measurement method of the measurement robot in the signal construction of the embodiment of the present invention, as shown in FIG. Figure 1 As shown, the method includes:

[0068] S101. Scan the construction area with a surveying robot to obtain three-dimensional point cloud data of the construction area, perform rasterization processing on the construction area based on the three-dimensional point cloud data, generate a raster map, collect image information of the construction area with the surveying robot, identify positioning mark points in the construction area with an image recognition algorithm, map the spatial coordinates of the positioning mark points to the raster map, determine the distribution of the positioning mark points in the raster map, and use a time-series reinforcement iterative neural network algorithm to plan the optimal motion trajectory of the surveying robot;

[0069] S102. Control the measurement robot to move along the optimal motion trajectory, collect measurement posture data in real time through the inertial measurement unit of the measurement robot, control the inkjet device of the measurement robot to maintain a horizontal state based on the measurement posture data, and when the measurement robot moves to a preset inkjet position, measure and automatically adjust the vertical distance between the nozzle of the inkjet device and the ground, collect image information of the preset inkjet position through the visual positioning system, and control the inkjet device to perform inkjet operation in combination with a preset inkjet template;

[0070] S103. After the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

[0071] First, the survey robot uses the 3D laser scanner to scan the construction area in all directions to obtain high-precision 3D point cloud data. The voxel filtering algorithm is used to downsample and denoise the point cloud data to improve data quality. Then, the octree algorithm is used to divide the processed point cloud data into grid cells of uniform size, and the size of each grid cell is set to 0.1mx 0.1m. The point cloud in each grid cell is statistically analyzed to extract the grid's elevation, flatness and other feature information to generate a 2D grid map.

[0072] Next, the measurement robot uses the high-definition camera on board to collect image information of the construction area. The YOLOv5 target detection algorithm is used to process the image and identify the positioning markers in the construction area. The positioning markers are usually black and white checkerboard patterns with a size of 20cm x 20cm. The YOLOv5 algorithm can quickly and accurately detect the position and outline of these markers. After obtaining the image coordinates of the positioning markers, they are converted into three-dimensional space coordinates using the internal and external parameters of the camera. Then, these three-dimensional coordinates are mapped to the previously generated grid map to obtain the precise position distribution of the positioning markers in the grid map.

[0073] Based on the grid map and the distribution information of the positioning markers, the time-series reinforcement iterative neural network (STRIN) algorithm is used to plan the optimal motion trajectory of the measurement robot. The STRIN algorithm combines the advantages of reinforcement learning and recurrent neural networks and can effectively handle time-series decision-making problems. First, the grid map and the position of the positioning markers are used as state inputs, and the action space is defined as the movement direction and speed of the robot. Then, the reward function is designed, taking into account factors such as path length, energy consumption, and obstacle avoidance. Through multiple iterative training, the STRIN algorithm can learn the optimal strategy and generate a smooth and efficient motion trajectory.

[0074] The robot is controlled to move along the planned optimal trajectory. The inertial measurement unit (IMU) carried by the robot collects the robot's attitude data in real time, including three-axis acceleration, angular velocity, and magnetic field strength. The extended Kalman filter algorithm is used to fuse the IMU data and odometer data to obtain high-precision attitude estimation results. Based on the attitude estimation results, the PID control algorithm is used to adjust the attitude of the robot chassis to ensure that the inkjet printer always remains horizontal and improve the inkjet printer accuracy.

[0075] When the measuring robot moves to the preset coding position, the laser distance sensor on the robot is activated to measure the vertical distance between the coding device nozzle and the ground. The target distance is set to 30cm, and the stepper motor is used to control the lifting and lowering of the coding device to achieve precise adjustment of the nozzle height. At the same time, the robot's visual positioning system is started, which includes two orthogonally installed industrial cameras to collect stereo images of the coding area. The precise three-dimensional coordinates of the coding position are calculated using SIFT feature matching and triangulation principles.

[0076] Combined with the pre-designed inkjet template, the inkjet device is controlled to perform inkjet operation. The inkjet template contains information such as the inkjet content, font, size, etc. The vector graphics library is used to convert the template into an inkjet instruction sequence to accurately control the switch and movement of the nozzle. The inkjet device uses piezoelectric inkjet technology to achieve an injection accuracy of 0.1mm. By adjusting the ink drop size and injection frequency, the inkjet pattern is ensured to be clear and uniform.

[0077] After the coding operation is completed, the measurement robot uses a high-resolution camera to collect images of the coding area with an image resolution of no less than 4K. The adaptive threshold segmentation algorithm is used to extract the outline of the coding pattern, and then the morphological operation is used to refine and smooth the outline. The processed contour features are aligned with the preset coding template, and the iterative closest point (ICP) algorithm is used to calculate the deviation between the actual coding position and the ideal position. The deviation threshold is set to 2mm. When the actual deviation exceeds the threshold, the adaptive optimization process is triggered.

[0078] Adaptive optimization uses a method based on deep reinforcement learning. A multi-layer perceptron network is constructed, with inputs including current motion parameters, environmental characteristics and deviation data, and outputs optimized motion parameters. Through interaction with the environment and guidance of the reward function, the network can gradually learn better motion strategies and automatically adjust the robot's speed, acceleration and other parameters to adapt to different ground conditions and external interference.

[0079] Finally, the actual deviation data and optimized parameters are uploaded to the construction management system in real time. Based on these data, the system generates a detailed construction quality report, including the distribution map of the coding position, deviation statistical analysis, optimization effect evaluation, etc. Construction management personnel can use the report to timely understand the construction quality and make corresponding adjustments and improvements.

[0080] In this embodiment, the whole process from area scanning, trajectory planning to coding operation is automated through the fixed measurement robot, which greatly reduces the need for manual operation. At the same time, the use of high-precision sensors and advanced algorithms ensures the accuracy of the coding position and significantly improves the construction quality; through continuous position re-measurement and deviation analysis, the system can timely detect and correct construction errors. The introduction of the deep learning model enables the robot to automatically adjust the motion parameters according to the actual situation, improving the adaptability and robustness of the system; all construction data are uploaded and analyzed in real time to generate a comprehensive quality report. This not only facilitates construction supervision and acceptance, but also provides important data support for subsequent maintenance and optimization, which is conducive to long-term improvement of construction efficiency and quality.

[0081] In an optional implementation, scanning the construction area by a fixed measurement robot to obtain three-dimensional point cloud data of the construction area, rasterizing the construction area based on the three-dimensional point cloud data, and generating a raster map includes:

[0082] Scanning the construction area with a fixed measurement robot to obtain three-dimensional point cloud data of the construction area, calculating the statistical mean and standard deviation of the three-dimensional point cloud data, determining a valid data range, removing abnormal data outside the valid data range, obtaining filtered three-dimensional point cloud data, dividing the filtered three-dimensional point cloud data into voxels according to a preset voxel size, calculating the centroid coordinates of the three-dimensional points contained in each voxel, and using the centroid coordinates as downsampling points of the corresponding voxels to obtain downsampled three-dimensional point cloud data;

[0083] According to a preset radius, select neighboring 3D points within a corresponding radius range of each 3D point in the downsampled 3D point cloud data, construct a covariance matrix based on the neighboring 3D points, calculate the normal vector of each 3D point through the covariance matrix, determine the vector difference between the normal vectors, use the median of the vector difference as a Gaussian kernel parameter, construct a Gaussian kernel function based on the Gaussian kernel parameter, and calculate the significance score of each 3D point;

[0084] The significance scores are sorted in descending order, and the top thirty percent of the three-dimensional points in the significance score sorting are selected as candidate feature points. The spacing between the candidate feature points is determined, and the candidate feature points whose spacing is less than a preset spacing threshold are eliminated to obtain a feature point set; the point feature histogram descriptor of each feature point in the feature point set is calculated, and feature matching is performed to obtain an initial registration result;

[0085] Calculate the distance value and the normal angle value between the corresponding feature points in the initial registration result, use the median value of the distance value as the distance parameter, and the median value of the normal angle value as the angle parameter; construct a distance weight item based on the distance parameter, and construct an angle weight item based on the angle parameter, multiply the distance weight item by the angle weight item to obtain an adaptive weight, and use the adaptive weight to perform iterative optimization to obtain an accurate registration result;

[0086] The construction area is divided into grids of a preset size, the distance between each grid and the three-dimensional point cloud data in the precise registration result is calculated, the observation likelihood of the Gaussian distribution is calculated, the observation likelihood is multiplied by the prior probability of the corresponding grid to obtain the probability product, the reciprocal of the cumulative sum of the probability products is calculated as a normalization factor, the probability product is normalized using the normalization factor to obtain the state probability of each grid, each grid is marked as occupied, idle or unknown based on the state probability, and a grid map of the construction area is generated.

[0087] Specifically, a fixed measurement robot is used to scan the construction area to obtain the three-dimensional point cloud data of the construction area. In order to improve the data quality, the original point cloud data needs to be preprocessed. The statistical mean and standard deviation of the three-dimensional point cloud data are calculated, and the effective data range is set to the mean plus or minus 3 times the standard deviation. Abnormal points outside this range are eliminated to obtain the filtered three-dimensional point cloud data.

[0088] Next, the filtered point cloud data is downsampled. The three-dimensional space is divided into cubic voxels with a side length of 10 cm. For each voxel, the centroid coordinates of all the three-dimensional points inside it are calculated and used as the downsampling point of the voxel. In this way, the amount of point cloud data can be greatly reduced while retaining the geometric features of the original point cloud.

[0089] Then, feature extraction is performed on the downsampled point cloud. The search radius is set to 30 cm, and the neighboring points within this radius are searched for each 3D point. A 3x3 covariance matrix is ​​constructed based on these neighboring points, and the normal vector of each point is obtained by eigenvalue decomposition. The angle between the normal vectors of adjacent points is calculated, and the value is taken as the Gaussian kernel parameter σ. This parameter is used to construct the Gaussian kernel function: exp(-d 2 / (2σ 2 )), where d is the distance between two points. For each point, the sum of its Gaussian kernel function value and that of its neighboring points is calculated to obtain the significance score of the point.

[0090] The significance scores of all points are sorted in descending order, and the top 30% of the points are selected as candidate feature points. In order to make the feature points evenly distributed, the minimum spacing threshold is set to 5cm, and the candidate points with spacing less than the threshold are eliminated, and finally the feature point set is obtained.

[0091] For each point in the feature point set, its point feature histogram (PFH) descriptor is calculated. The PFH descriptor characterizes local features by counting the geometric relationship between point pairs in the neighborhood of the feature point. After the calculation is completed, the KD tree is used to match the feature points to obtain the initial point cloud registration result.

[0092] In order to further optimize the registration results, the adaptive weighted iterative closest point (ICP) algorithm is introduced. First, the Euclidean distance and normal angle between the corresponding feature point pairs in the initial registration results are calculated. The median of the distance is taken as the distance parameter d0, and the median of the normal angle is taken as the angle parameter θ0. Construct the distance weight term wd = exp(-d 2 / d0 2 ) and the angle weight term wθ=exp(-θ 2 / θ0 2 ). Multiplying the two weight terms together yields the adaptive weight w = wd*wθ. In the ICP iteration process, using the adaptive weight to weight the error can effectively suppress the influence of abnormal matching pairs and improve the registration accuracy.

[0093] Finally, a grid map is generated based on the accurately registered point cloud data. The construction area is divided into square grids with a side length of 20 cm. For each grid, the distance d from its center point to the nearest point cloud point is calculated. Assuming that the measurement error follows a Gaussian distribution with a mean of 0 and a standard deviation of σ, the observation likelihood can be expressed as p(z|x)=exp(-d 2 / (2σ 2 )). Multiply the observation likelihood by the grid’s prior probability p(x) to obtain the posterior probability p(x|z)∝p(z|x)*p(x). Normalize the posterior probabilities of all grids to obtain the state probability of each grid.

[0094] The grid is marked into three states according to the state probability: occupied, idle, or unknown. Specifically, when the state probability is greater than 0.7, it is marked as occupied, when it is less than 0.3, it is marked as idle, and when it is between 0.3 and 0.7, it is marked as unknown. In this way, a probabilistic grid map of the construction area is finally generated.

[0095] In this embodiment, through preprocessing steps such as filtering and downsampling, outliers and redundant data are effectively removed, thereby improving the point cloud data quality and processing efficiency; the introduction of the ICP algorithm with adaptive weights can effectively suppress the influence of abnormal matching and significantly improve the accuracy and robustness of point cloud registration; the probabilistic grid mapping method is adopted, which takes into account measurement uncertainty and can more accurately characterize the environmental state, providing a reliable basis for subsequent path planning and navigation decisions.

[0096] In an optional implementation, the positioning mark points in the construction area are identified by an image recognition algorithm, and the spatial coordinates of the positioning mark points are mapped to a grid map. The distribution of the positioning mark points in the grid map is determined, including:

[0097] The image collected by the measurement robot is input into the backbone feature network, and the backbone feature network extracts image features through the residual module. In the residual module, the spatial dimension and channel dimension of the feature map are weightedly reconstructed by using the spatial attention mechanism and the channel attention mechanism, respectively, to obtain the first-layer feature map; the first-layer feature map is input into the feature pyramid structure, and different scale features are fused through upsampling and downsampling operations to obtain a multi-level feature map;

[0098] The multi-level feature map is input into the detection head network, and the detection head network outputs the category score and position coordinates of the positioning mark point. Based on the category score and the position coordinates, the two-dimensional pixel coordinates of the positioning mark point in the image are determined; based on the internal parameter matrix of the binocular camera, the image is distorted and corrected, and the parallax information of the left and right images is calculated. In combination with the external parameter matrix of the binocular camera, the two-dimensional pixel coordinates are converted into depth information through the triangulation principle, and the two-dimensional pixel coordinates are combined with the depth information to obtain the three-dimensional space coordinates of the positioning mark point;

[0099] The corresponding feature points in the coordinate system corresponding to the camera coordinate system and the grid map are extracted, and feature point matching pairs are constructed. The rotation matrix and the translation vector are calculated, and the rotation matrix and the translation vector are optimized using a nonlinear optimization algorithm to obtain a coordinate transformation matrix. The three-dimensional space coordinates are mapped to the coordinate system of the grid map through the coordinate transformation matrix to obtain the spatial distribution of the positioning mark points.

[0100] First, the images collected by the measurement robot are input into the backbone feature network. The backbone feature network adopts the ResNet-50 structure and extracts image features through residual modules. In each residual module, the spatial attention mechanism and channel attention mechanism are introduced to reconstruct the feature map in a weighted manner. The spatial attention mechanism generates a spatial weight map through convolution operations and weights each spatial position of the feature map. The channel attention mechanism uses global average pooling and fully connected layers to generate a channel weight vector and weight each channel of the feature map. After feature extraction and attention weighting by multiple residual modules, the first layer of feature maps is obtained.

[0101] Next, the first layer feature map is input into the feature pyramid structure. The feature pyramid structure contains 5 scale layers, corresponding to 1 / 4, 1 / 8, 1 / 16, 1 / 32 and 1 / 64 of the original image. Features of different scales are fused through upsampling and downsampling operations. Specifically, the smaller scale feature map is upsampled by 2 times and element-wise fused with the larger scale feature map; the larger scale feature map is downsampled by 2 times and element-wise fused with the smaller scale feature map. This multi-scale feature fusion can effectively extract semantic information of different scales and obtain multi-level feature maps.

[0102] Then, the multi-level feature map is input into the detection head network. The detection head network contains a classification branch and a regression branch. The classification branch outputs the category score of each candidate box through the convolution layer and the Sigmoid activation function, and the regression branch outputs the position coordinate offset of each candidate box through the convolution layer. According to the preset confidence threshold (such as 0.5) and non-maximum suppression threshold (such as 0.3), the final detection result is screened to determine the two-dimensional pixel coordinates of the positioning mark point in the image. For example, the two-dimensional pixel coordinates of a certain positioning mark point detected are (320, 240).

[0103] Next, the image is distorted based on the internal parameter matrix of the binocular camera. The internal parameter matrix contains parameters such as focal length, principal point coordinates, and distortion coefficients. Using these parameters, the distortion is corrected through a polynomial model to obtain the corrected left and right images. Then, the disparity information of the left and right images is calculated. A stereo matching algorithm based on local block matching is used to calculate the disparity value for each pixel of the left and right images. Combined with the external parameter matrix of the binocular camera (including the rotation matrix and translation vector between the cameras), the two-dimensional pixel coordinates are converted into depth information through the principle of triangulation. For example, the depth value of a certain positioning marker is 5 meters. The two-dimensional pixel coordinates are combined with the depth information to obtain the three-dimensional spatial coordinates of the positioning marker, such as (1.6m, 1.2m, 5m).

[0104] Finally, extract the corresponding feature points in the coordinate system corresponding to the camera coordinate system and the grid map, and construct feature point matching pairs. Select at least 3 pairs of non-collinear feature point matching pairs, such as the points (0, 0, 0), (1, 0, 0), (0, 1, 0) in the camera coordinate system and the points (10, 10, 0), (11, 10, 0), (10, 11, 0) in the grid map coordinate system. Based on these matching pairs, the SVD decomposition method is used to calculate the initial rotation matrix and translation vector. Then, the Levenberg-Marquardt algorithm is used to perform nonlinear optimization on the rotation matrix and translation vector to minimize the reprojection error and obtain an accurate coordinate transformation matrix. The three-dimensional spatial coordinates of the positioning marker point are mapped to the coordinate system of the grid map through the coordinate transformation matrix to obtain the spatial distribution of the positioning marker point in the grid map. For example, the coordinates of a certain positioning marker point in the grid map are (15.6m, 12.2m).

[0105] In this embodiment, by introducing spatial attention and channel attention mechanisms, the effectiveness of feature extraction is enhanced and the detection accuracy of positioning markers is improved. Multi-scale feature fusion can effectively extract semantic information of different scales and enhance the adaptability to positioning markers of different sizes; the binocular camera and triangulation principle are used to achieve accurate acquisition of the three-dimensional coordinates of the positioning markers, which improves the accuracy of depth information compared to the monocular camera solution. The coordinate transformation method based on nonlinear optimization improves the accuracy of coordinate mapping; the overall solution realizes the full process automation processing from image acquisition to grid map mapping, improves the efficiency and accuracy of obtaining the spatial distribution of positioning markers, and provides reliable basic data for subsequent robot positioning and navigation.

[0106] In an optional implementation, the time-series enhanced iterative neural network algorithm is used to plan the optimal motion trajectory of the measurement robot, including:

[0107] Constructing a time-series reinforcement iterative neural network, combining the real-time position information of the measuring robot, the spatial distribution of the positioning mark points and the local information of the grid map into a state vector, and combining the discretized linear velocity and angular velocity into an action vector, wherein the time-series reinforcement iterative neural network includes a main action network and a target value network;

[0108] The main action network selects an action vector based on the current state vector, executes the action vector to obtain the next state vector, and the target value network estimates the value based on the next state vector; calculates the path length value, motion smoothness value and marker coverage value of the action vector, and combines them to form a reward value;

[0109] The current state vector, action vector, reward value and next state vector are used to construct state transfer data, the time difference error of the state transfer data is calculated, priorities are assigned to the state transfer data based on the time difference error, and the state transfer data of different priorities are stored in an experience replay pool; state transfer data are sampled from the experience replay pool according to priority to train the main action network, the parameters of the main action network are updated to the target value network according to a preset number of steps, iterations are repeated until convergence, and the optimal motion trajectory for the measurement robot to visit the positioning mark point is output.

[0110] Specifically, the state vector contains the following information: the current position coordinates of the robot (x, y, θ), where x and y are plane coordinates and θ is the orientation angle; the set of unvisited positioning marker coordinates {(x1, y1), (x2, y2), ..., (xn, yn)}; the local grid map centered on the robot, with a size of 20×20, and the value of each grid is 0 (idle) or 1 (obstacle). The action vector contains the linear velocity v and angular velocity ω, the value range of v is [0, 0.5] m / s, discretized into 5 values; the value range of ω is [-π / 4, π / 4] rad / s, discretized into 7 values. Therefore, the size of the action space is 5×7=35.

[0111] The main action network uses a fully connected neural network structure. The number of neurons in the input layer is the same as the dimension of the state vector. The hidden layer contains two layers, each with 128 neurons. The number of neurons in the output layer is 35, corresponding to 35 discrete actions. The target value network structure is the same as the main action network, but the output layer has only one neuron, which is used to estimate the state value. The activation function of both networks uses the ReLU function.

[0112] The main action network is based on the current state vector s t Select the action vector a t Specifically, s t Input the main action network and get the Q value estimates of 35 actions. Use the ε-greedy strategy to select actions, that is, select the action with the largest Q value with a probability of 1-ε, and randomly select actions with a probability of ε. The initial value of ε is 0.9, and it decays linearly to 0.1 during training. Execute the selected action a t , simulate the robot movement for 0.5 seconds and get the next state s t+1 .

[0113] The target value network is based on the next state vector s t+1 Estimated value V(s t+1 ). Specifically, s t+1 Input the target value network to directly get the value estimate of the state.

[0114] Next, calculate action a tThe reward value r t , consists of three parts: path length value r l , motion smoothness value r s And the coverage value r of the marking point c . The path length value r l =-d, where d is the execution action a t The distance the robot moves. The motion smoothness value r s =-|ω t -ω t-1 |, where ω t and ω t-1 are the angular velocities at the current moment and the previous moment respectively. The mark point coverage value rc = n / N, where n is the number of mark points currently visited and N is the total number of mark points. The final reward value r t =w1*r l +w2*r s +w3*r c , where w1, w2, and w3 are weight coefficients, which are 0.3, 0.3, and 0.4 respectively.

[0115] The current state s t 、Action a t , Reward t and the next state s t+1 Construct state transfer data(s t , a t , r t ,s t +1). Calculate the time difference error δ of the data t =r t +γV(s t+1 )-Q(s t , a t ), where γ is the discount factor, which is 0.99; V(s t+1 ) is the next state value estimate output by the target value network; Q(s t , a t ) is the Q value estimate of the current state-action pair output by the main action network. Based on |δ t |Assign priority p to state transition data = |δ t |+ε, where ε is a small positive number to prevent the priority from being 0. t , a t , r t ,s t+1 , p) stored in the experience replay pool.

[0116] The experience replay pool is implemented using a priority queue data structure with a capacity of 10,000. When the pool is full, new data will replace the old data with the lowest priority. Mini-batches (size 64) are sampled from the experience replay pool according to the probability distribution of priority p to train the main action network. Specifically, the TD error of the sampled data is calculated as δ = r + γV(s') - Q(s, a), where V(s') and Q(s, a) are calculated by the target value network and the main action network, respectively. Then the loss function L = (δ - Q(s, a)) is calculated. 2 , update the main action network parameters through back-propagation.

[0117] After every 100 main action network updates, its parameters are soft updated to the target value network: θ'=τθ+(1-τ)θ', where θ and θ' are the parameters of the main action network and the target value network, respectively, and τ is the soft update coefficient, which is 0.01. Repeat the above training process until convergence or the maximum number of iterations (10,000) is reached.

[0118] After the training is completed, the trained main action network is used to plan the robot's motion trajectory. Starting from the starting position, repeat the following steps until all the marked points are visited: 1) Construct the current state vector s t ; 2) Input st into the main action network and select the action a with the largest Q value t ;3)Execute a t And update the robot position. The final trajectory is the optimal motion trajectory.

[0119] In this embodiment, a time-series reinforcement iterative neural network algorithm is adopted, which can adaptively learn the optimal decision-making strategy in a complex environment, and has stronger generalization ability and environmental adaptability than the traditional path planning algorithm; the introduction of a priority experience replay mechanism improves the learning efficiency of key experiences, accelerates the convergence speed of the network, and also improves the efficiency of the algorithm; by comprehensively considering multiple objectives such as path length, motion smoothness, and marker coverage, the obtained trajectory can achieve a good balance in multiple indicators to meet the needs of actual applications.

[0120] In an optional embodiment, the measurement robot is controlled to move along the optimal motion trajectory, the measurement posture data is collected in real time by the inertial measurement unit of the measurement robot, and the inkjet device of the measurement robot is controlled to maintain a horizontal state based on the measurement posture data. When the measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted, and the image information of the preset inkjet position is collected by the visual positioning system. In combination with the preset inkjet template, the inkjet device is controlled to perform the inkjet operation, which includes:

[0121] The inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by an extended Kalman filter algorithm, and the roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion;

[0122] Based on the attitude quaternion, a kinematic model of the inkjet printer is established, the target angle of the first axis and the target angle of the second axis of the two-degree-of-freedom gimbal are calculated, the stepper motors of the first axis and the second axis are controlled to rotate respectively by a PID controller, and the inkjet printer is set in a horizontal state;

[0123] When the preset coding position is reached, the vertical distance between the nozzle and the ground is measured by a laser distance sensor, and the vertical distance is compared with the preset vertical distance boundary to obtain the distance deviation value and the distance deviation change rate, establish a fuzzy control rule, input the fuzzy control rule into the fuzzy PID controller, adaptively adjust the PID parameters, control the speed of the DC servo motor of the lifting mechanism, and adjust the nozzle height;

[0124] Collect an image of a preset inkjet printing position, perform grayscale and edge detection processing on the image to obtain contour features, perform feature point matching between a preset inkjet printing template and the contour features, determine the inkjet printing position coordinates and the spraying direction angle based on the matching results, generate a spraying instruction sequence, the spraying instruction sequence includes a spraying instruction sequence of nozzle position information, spraying pressure information and spraying time information, and send the spraying instruction sequence to the inkjet printing controller via a field bus;

[0125] During the inkjet printing process, the spraying image is collected in real time, and the detection parameters of the spraying lines in the spraying image are analyzed, the detection parameters including width value, continuity value and clarity value, and the analysis results are generated. When any one of the detection parameters is lower than the corresponding preset threshold value, the spraying pressure information and the spraying time information are adaptively adjusted based on the analysis result until each detection parameter reaches the corresponding preset threshold value, and the inkjet printing operation is completed.

[0126] This embodiment provides a method for measuring robot coding, and the specific implementation steps are as follows:

[0127] First, the fixed-measurement robot is controlled to move along the preset optimal motion trajectory. The optimal motion trajectory can be generated by a path planning algorithm, such as the A* algorithm or the RRT algorithm, taking into account factors such as obstacle avoidance and motion stability. The fixed-measurement robot can adopt a differential drive structure to achieve movement by controlling the speed of the left and right wheel motors.

[0128] During the movement of the robot, the inertial measurement unit (IMU) on the robot collects the attitude data in real time. The IMU contains a three-axis accelerometer and a three-axis gyroscope, and the sampling frequency can be set to 100Hz. The accelerometer measures the acceleration components in three directions, and the gyroscope measures the angular velocity components in three axes.

[0129] The acceleration and angular velocity data collected by the IMU are input into the extended Kalman filter (EKF) algorithm to estimate the attitude quaternion. The EKF algorithm includes two steps: prediction and update. The prediction step predicts the state based on the system model, and the update step corrects the prediction result based on the observed data. The optimal estimated attitude quaternion is obtained through iterative calculation.

[0130] The roll angle, pitch angle and yaw angle of the robot are calculated based on the attitude quaternion. This can be achieved through the conversion formula from quaternion to Euler angle. For example, the roll angle is equal to the inverse tangent function of the four components x, y, z, and w in the quaternion.

[0131] Based on the calculated attitude quaternion, the kinematic model of the inkjet device is established. The inkjet device is installed on a two-degree-of-freedom pan-tilt, and the two rotating shafts of the pan-tilt are driven by stepper motors. Through the kinematic forward solution, the target angles of the two rotating shafts are calculated to keep the inkjet device level.

[0132] For example, when the attitude quaternion indicates that the robot is tilted forward by 5 degrees, the first axis (pitch axis) needs to be rotated backward by 5 degrees to compensate. The stepper motors of the two axes are controlled by PID controllers to rotate to the target angles. The proportional, integral, and differential parameters of the PID controller can be determined by the Ziegler-Nichols tuning method and then fine-tuned according to the actual response characteristics.

[0133] When the measuring robot moves to the preset coding position, the vertical distance between the nozzle and the ground is measured by the laser distance sensor installed on the robot. The measuring range of the laser distance sensor is 0.1-10m, and the accuracy is ±1mm. The actual distance measured is compared with the preset ideal distance to obtain the distance deviation value.

[0134] At the same time, the change rate of the distance deviation is calculated, that is, the change of the distance deviation per unit time. The distance deviation value and the change rate are used as input to establish the fuzzy control rule. The fuzzy control rule can be in the form of IF-THEN, for example: IF the deviation is positive and the change rate is positive THEN the output is positive.

[0135] The fuzzy control rules are input into the fuzzy PID controller to realize the adaptive adjustment of the PID parameters. The fuzzy PID controller looks up the rule table to obtain the adjustment amount of the PID parameters according to the fuzzy quantity of the deviation and the rate of change, thereby dynamically adjusting the PID parameters. The controller output acts on the DC servo motor of the lifting mechanism to adjust the motor speed and thus adjust the nozzle height.

[0136] The image of the preset inkjet printing position is collected through the visual positioning system. The visual system includes an industrial camera and an LED light source. The camera resolution is 1920x1080 and the frame rate is 60fps. The collected image is grayed and the RGB three-channel image is converted into a single-channel grayscale image.

[0137] Then, Canny edge detection is performed on the grayscale image to extract the contour features in the image. The Canny algorithm includes Gaussian filtering, calculating gradient amplitude and direction, non-maximum suppression, dual threshold detection and other steps. The high and low thresholds of edge detection can be set to 50 and 150.

[0138] Match the pre-stored coding template with the extracted contour features. Use the SIFT algorithm to extract feature points and descriptors, and then use the FLANN algorithm for nearest neighbor matching. Calculate the affine transformation matrix based on the matching results to determine the coding position coordinates and direction angle.

[0139] Generate a spraying instruction sequence based on the matching results, including information such as nozzle position, spraying pressure and spraying time. For example, the nozzle position can be expressed as relative coordinates (x, y, z), the spraying pressure range is 0.2-0.8MPa, and the spraying time is 0.1-2s. The instruction sequence is sent to the inkjet controller via the CAN bus for execution.

[0140] During the coding process, the spraying image is collected and analyzed in real time. Parameters such as width, continuity and clarity of the sprayed lines are extracted. The width can be measured by morphological operations, the continuity can be evaluated by detecting the number of breakpoints, and the clarity can be analyzed by edge gradient.

[0141] The detection parameters are compared with the preset thresholds, such as the width threshold of 1-3mm, the continuity threshold of 95%, and the clarity threshold of 0.8. When any parameter is lower than the threshold, the spraying pressure and time are adaptively adjusted according to the analysis results. The pressure can be adjusted within the range of ±0.1MPa, and the time can be adjusted within the range of ±0.2s. Through multiple iterations of optimization, until all parameters reach the preset threshold, high-quality inkjet printing is completed.

[0142] In this embodiment, through inertial measurement and posture estimation, combined with a two-degree-of-freedom pan-tilt platform, automatic horizontal maintenance of the coding device is achieved, thereby improving the coding accuracy and quality; a fuzzy PID control algorithm is used to adaptively adjust the nozzle height, thereby improving the robustness and adaptability of the system and being able to cope with different ground conditions; based on visual positioning and real-time image analysis, precise positioning of the coding position and dynamic optimization of the spraying parameters are achieved, thereby ensuring the consistency and reliability of the coding effect.

[0143] In an optional embodiment, the inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by an extended Kalman filter algorithm. The roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion, including:

[0144] The three-axis acceleration sensor of the inertial measurement unit is used to collect the acceleration components of gravity acceleration in three orthogonal axes, and the three-axis gyroscope is used to collect the angular velocity components of the robot rotating around the three orthogonal axes, and the acceleration components and the angular velocity components are sampled and filtered through the signal acquisition module;

[0145] A system state equation is established based on the angular velocity component, the attitude quaternion is set as the state vector, the angular velocity is set as the system input, and a nonlinear state transfer model is constructed; an observation equation is established based on the acceleration component, the projection of the gravity vector in the body coordinate system is set as the observation quantity, and a nonlinear observation model is constructed according to the relationship between the attitude quaternion and the gravity vector;

[0146] Performing a first-order Taylor expansion on the nonlinear state transfer model to obtain a state Jacobian matrix, and performing a first-order Taylor expansion on the nonlinear observation model to obtain an observation Jacobian matrix;

[0147] The extended Kalman filter algorithm is used to estimate the attitude, including: using the state Jacobian matrix and the nonlinear state transfer model to perform state prediction to obtain a predicted attitude quaternion and a predicted state covariance; using the observation Jacobian matrix, the nonlinear observation model and the predicted state covariance to calculate the Kalman gain, multiplying the Kalman gain with the observation residual to obtain a state correction, adding the state correction to the predicted attitude quaternion to obtain an estimated attitude quaternion, and updating the state covariance at the same time;

[0148] The estimated quaternion is normalized, and each component is divided by the corresponding modulus to obtain a unit attitude quaternion; according to the four components of the unit attitude quaternion, three rotation angles describing the robot attitude are calculated through a conversion formula from the unit attitude quaternion to the Euler angle, including a roll angle around the X-axis, a pitch angle around the Y-axis, and a yaw angle around the Z-axis.

[0149] Specifically, the acceleration component and angular velocity component are collected by the inertial measurement unit of the measuring robot. The inertial measurement unit includes a three-axis acceleration sensor and a three-axis gyroscope. The three-axis acceleration sensor is used to measure the components of gravity acceleration in the three orthogonal axes of x, y, and z, with a sampling frequency of 100Hz. The three-axis gyroscope is used to measure the angular velocity component of the robot rotating around the three orthogonal axes of x, y, and z, with a sampling frequency of 200Hz. The collected acceleration and angular velocity signals are filtered through a low-pass filter with a cutoff frequency of 20Hz to eliminate high-frequency noise.

[0150] The filtered acceleration and angular velocity signals are input into the extended Kalman filter for attitude estimation. First, the system state equation is established, the attitude quaternion q = [q0, q1, q2, q3] is set as the state vector, the angular velocity ω = [ωx, ωy, ωz] is set as the system input, and the nonlinear state transfer model f(q, ω) is constructed. Then the observation equation is established, the projection of the gravity vector g in the body coordinate system is set as the observation quantity, and the nonlinear observation model h(q) is constructed based on the relationship between the attitude quaternion and the gravity vector.

[0151] The nonlinear state transfer model f(q, ω) is subjected to a first-order Taylor expansion to obtain a 4×4 state Jacobian matrix F. The nonlinear observation model h(q) is subjected to a first-order Taylor expansion to obtain a 3×4 observation Jacobian matrix H.

[0152] The state Jacobian matrix F and the nonlinear state transfer model f(q, ω) are used for state prediction to obtain the predicted attitude quaternion q_pred and the predicted state covariance P_pred. The Kalman gain K is calculated using the observation Jacobian matrix H, the nonlinear observation model h(q) and the predicted state covariance P_pred. The Kalman gain K is multiplied by the observation residual to obtain the state correction Δq, and the state correction Δq is added to the predicted attitude quaternion q_pred to obtain the estimated attitude quaternion q_est, and the state covariance P is updated at the same time.

[0153] The estimated attitude quaternion q_est is normalized, and each component is divided by the modulus of the quaternion to obtain the unit attitude quaternion q_unit = [q0, q1, q2, q3]. According to the four components of the unit attitude quaternion q_unit, through the conversion relationship from quaternion to Euler angle, the three rotation angles describing the robot attitude are calculated: the roll angle φ around the x-axis, the pitch angle θ around the y-axis, and the yaw angle ψ around the z-axis.

[0154] The specific implementation steps are as follows:

[0155] First, the acceleration and angular velocity signals are collected by the inertial measurement unit. The three-axis acceleration sensor measures the acceleration component ax = 0.1m / s 2 , ay=-0.2m / s2 , az=9.8m / s 2 The three-axis gyroscope measures the angular velocity components ωx = 0.01 rad / s, ωy = -0.02 rad / s, and ωz = 0.03 rad / s.

[0156] Then, the original signal is low-pass filtered. A second-order Butterworth low-pass filter is used with a cutoff frequency of 20 Hz. After filtering, ax_f = 0.09 m / s 2 , ay_f=-0.19m / s 2 , az_f=9.79m / s 2 , ωx_f=0.009rad / s, ωy_f=-0.018rad / s, ωz_f=0.028rad / s.

[0157] Next, a nonlinear state transfer model f(q, ω) and a nonlinear observation model h(q) are constructed. The state transfer model describes the change of the attitude quaternion q over time, and the observation model describes the relationship between the projection of the gravity vector g in the body coordinate system and the attitude quaternion q.

[0158] The nonlinear model is expanded by the first order Taylor to obtain the state Jacobian matrix F and the observation Jacobian matrix H. The F matrix reflects the influence of a small change in the state vector q on the state transfer, and the H matrix reflects the influence of a small change in the state vector q on the observed quantity.

[0159] Perform state prediction. Using the estimated attitude quaternion q at the previous moment and the measured angular velocity ω, the predicted attitude quaternion q_pred = [0.998, 0.005, -0.01, 0.015] is calculated through the nonlinear state transfer model f(q, ω). At the same time, the state Jacobian matrix F is used to update the predicted state covariance P_pred.

[0160] Calculate the observation residual. Use the nonlinear observation model h(q) to calculate the predicted observation value and compare it with the actual measured acceleration value to obtain the observation residual.

[0161] Calculate the Kalman gain K. The Kalman gain K is calculated based on the predicted state covariance P_pred, the observed Jacobian matrix H, and the observed noise covariance R.

[0162] Update state estimation. Multiply the Kalman gain K by the observation residual to get the state correction Δq = [0.001, -0.002, 0.003, -0.004], add it to the predicted attitude quaternion q_pred to get the estimated attitude quaternion q_est = [0.999, 0.003, -0.007, 0.011]. At the same time, update the state covariance P.

[0163] The estimated attitude quaternion q_est is normalized to obtain the unit attitude quaternion q_unit = [0.9990, 0.0030, -0.0070, 0.0110].

[0164] Finally, according to the unit attitude quaternion q_unit, the Euler angles are calculated: roll angle φ = 0.35°, pitch angle θ = -0.80°, yaw angle ψ = 1.26°. These three angles describe the current attitude of the robot.

[0165] In this embodiment, an extended Kalman filter algorithm is used to fuse accelerometer and gyroscope data, making full use of the complementary characteristics of the two sensors, retaining the short-term stability advantage of the gyroscope, and using the long-term stability characteristics of the accelerometer to correct the accumulated error, thereby improving the accuracy and reliability of attitude estimation; by constructing a nonlinear state space model and performing linearization processing, this method can effectively deal with nonlinear problems in attitude estimation, and has better performance than traditional linear Kalman filtering; using quaternions to represent attitude avoids the singularity problem in Euler angle representation, and can achieve stable estimation within the full attitude range, which is suitable for attitude measurement in complex motion scenarios of robots.

[0166] In an optional implementation, after the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds a preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time, and a construction quality report is generated, including:

[0167] The spectral image of the coding area is collected by a hyperspectral camera, a grid pattern is projected by a structured light projector and a structured light image is collected, and the spectral image and the structured light image are input into a synchronous trigger circuit for timing alignment to obtain standard dual-modal image data;

[0168] The standard bimodal image data is input into a spectral-spatial joint feature extraction network, the spectral attention module of the spectral-spatial joint feature extraction network processes the spectral image to obtain spectral features, the spatial pyramid pooling layer processes the structured light image to obtain spatial features, and the spectral features and the spatial features are fused through a cross-modal feature fusion module to obtain fused features of the coding area;

[0169] Perform edge detection on the fused features to obtain the edge point set of the coding pattern, input the edge point set into the dynamic graph convolution network, extract the local structure information through the adaptive graph convolution layer, and use the graph attention mechanism to weight the local structure information to obtain the contour features of the coding pattern;

[0170] The contour feature is matched with the preset coding template, and the initial posture deviation is obtained by performing a first-stage registration through a graph matching network. Based on the initial posture deviation, a second-stage registration is performed through an iterative closest point algorithm to obtain the actual posture deviation of the coding pattern;

[0171] The actual posture deviation is input into a hierarchical reinforcement learning network, a high-level strategy network of the hierarchical reinforcement learning network generates a parameter optimization strategy based on the actual posture deviation, and a low-level execution network adjusts the motion parameters of the measurement robot based on the parameter optimization strategy;

[0172] The motion parameters are input into a hybrid parameter optimizer, the hybrid parameter optimizer uses Bayesian optimization to determine a parameter search space, and an evolutionary algorithm is used to iteratively optimize the motion parameters in the parameter search space to obtain optimized motion parameters;

[0173] Based on a pre-constructed uncertainty assessment model, the optimized motion parameters are input into the uncertainty assessment model, and the measurement uncertainty introduced by the measurement process, the model uncertainty introduced by the model prediction and the parameter uncertainty introduced by the parameter optimization are calculated respectively. The measurement uncertainty, the model uncertainty and the parameter uncertainty are input into the evidence theory model to obtain the credibility evaluation value of the parameter optimization result.

[0174] After the coding operation is completed, a high-definition image of the coding area is first collected. Specifically, a high-resolution industrial camera is used to shoot the coding area to obtain an RGB color image with a resolution of no less than 4K. At the same time, a hyperspectral camera is used to image the coding area to obtain a hyperspectral data cube in the wavelength range of 400-1000nm. In addition, a grid-like stripe pattern is projected onto the coding area through a structured light projector, and a high-speed camera is used to capture the structured light image.

[0175] Next, the collected spectral image and structured light image are input into the synchronization trigger circuit for timing alignment. Specifically, a high-precision clock signal is used to control the exposure time of the hyperspectral camera and the structured light camera to ensure that the acquisition time error of the two images is less than 1ms. Then, an image registration algorithm, such as the phase correlation method, is used to align the spatial position relationship of the two images, and finally standard dual-modal image data with time and space synchronization is obtained.

[0176] The standard bimodal image data is input into the pre-trained spectral-spatial joint feature extraction network. The spectral attention module of the network first processes the spectral image: the spectral features of each pixel are extracted through one-dimensional convolution, and then the features of different bands are weighted using the channel attention mechanism to highlight the key band information. At the same time, the spatial pyramid pooling layer performs multi-scale feature extraction on the structured light image: the image is pooled through pooling kernels of different sizes to obtain spatial feature maps of different scales. Finally, the cross-modal feature fusion module uses bilinear pooling to fuse the spectral features and multi-scale spatial features to obtain the feature representation of the coding area that integrates spectral and spatial information.

[0177] The fused features are edge detected to extract the contour information of the coding pattern. Specifically, the Canny operator is first used to perform edge detection on the fused feature map to obtain a preliminary edge map. Then, morphological operations are used to refine and connect the edge map to obtain a continuous edge curve. The Douglas-Peucker algorithm is then used to perform polygonal approximation on the edge curve to extract key edge points and form an edge point set. The edge point set is input into a dynamic graph convolutional network. The adaptive graph convolution layer of the network dynamically constructs a graph structure based on the spatial relationship between the points and extracts local structural features through graph convolution operations. Then, the graph attention mechanism is used to weight the features of different nodes to highlight important structural information. Finally, a feature representation that accurately describes the contour of the coding pattern is obtained.

[0178] The extracted contour features are matched with the preset inkjet template. First, a graph matching network is used for coarse registration: the contour features and template features are input into the Siamese network, the feature similarity matrix is ​​calculated, and then the Hungarian algorithm is used to solve the optimal match to obtain a preliminary pose deviation estimate. Then, based on the preliminary estimation results, the iterative closest point (ICP) algorithm is used for fine registration: the corresponding point pairs are iteratively calculated and the distance between the point pairs is minimized, the pose estimation is continuously optimized, and finally the actual pose deviation with sub-pixel accuracy is obtained.

[0179] The actual posture deviation is input into the hierarchical reinforcement learning network to optimize the motion parameters of the robot. The high-level policy network of the network first generates a parameter optimization strategy based on the posture deviation, such as adjusting the joint angle, end position, etc. Then the low-level execution network adjusts the robot's kinematic parameters, such as joint angle, linear velocity, angular velocity, etc., according to the optimization strategy. The optimization process uses a model-based reinforcement learning algorithm, such as DDPG, to continuously improve the policy network through interaction with the simulation environment, and finally obtain a robust parameter optimization solution.

[0180] The initially optimized motion parameters are input into the hybrid parameter optimizer for further optimization. First, the Bayesian optimization algorithm is used to determine the parameter search space: a Gaussian process regression model is constructed based on historical optimization data to predict the performance of different parameter combinations, and the most promising area is selected as the search space. Then, an evolutionary algorithm is used to optimize the parameters in this space: through operations such as selection, crossover, and mutation, new parameter combinations are generated and their performance is evaluated. After multiple generations of evolution, the globally optimal motion parameters are finally obtained.

[0181] Finally, the uncertainty of the optimized motion parameters is evaluated. First, the uncertainty introduced by the measurement process is evaluated based on the Monte Carlo method: by repeated measurements, the distribution characteristics of the measurement results are analyzed. Then, the uncertainty of model prediction is evaluated using the ensemble learning method: multiple models are trained and the discrete degree of their prediction results is analyzed. Finally, the uncertainty introduced by parameter optimization is evaluated using the sensitivity analysis method: the parameters are optimized by perturbing and observing the changes in system response. The three uncertainties are input into the evidence fusion model based on the Dempster-Shafer theory, and the credibility of the parameter optimization results is comprehensively calculated to provide a reliability basis for subsequent decision-making.

[0182] In this embodiment, by combining hyperspectral and structured light imaging technologies, rich spectral information is obtained and the geometric structure of the coding pattern is accurately captured, laying a solid foundation for subsequent feature extraction and matching. At the same time, advanced algorithms such as deep learning and graph networks are used to achieve accurate modeling and feature expression of the coding pattern, greatly improving the accuracy and robustness of detection; through hierarchical reinforcement learning and hybrid optimization strategies, the robot's motion parameters can be dynamically adjusted according to the actual detection results, effectively compensating for system errors and environmental disturbances. This closed-loop optimization mechanism significantly improves the adaptability and stability of the system and ensures consistent performance in long-term operation; by comprehensively considering the uncertainties introduced by measurement, model and parameter optimization, and integrating them with evidence theory, a credibility assessment of the optimization results is given. This provides a reliable basis for subsequent decision-making, helps to timely discover potential risks and take corresponding measures, and improves the reliability and safety of the entire system.

[0183] Figure 2 Schematic diagram of the structure of the real-time coding and position re-measurement system of the fixed measurement robot in the signal construction of the embodiment of the present invention, as shown in FIG. Figure 2 As shown, the system comprises:

[0184] The first unit is used to scan the construction area by a surveying robot, obtain three-dimensional point cloud data of the construction area, perform rasterization processing on the construction area based on the three-dimensional point cloud data, generate a raster map, collect image information of the construction area by the surveying robot, identify positioning mark points in the construction area by an image recognition algorithm, map the spatial coordinates of the positioning mark points to the raster map, determine the distribution of the positioning mark points in the raster map, and use a time-series reinforcement iterative neural network algorithm to plan the optimal motion trajectory of the surveying robot;

[0185] The second unit is used to control the movement of the fixed measurement robot along the optimal motion trajectory, collect fixed measurement posture data in real time through the inertial measurement unit of the fixed measurement robot, and control the inkjet device of the fixed measurement robot to maintain a horizontal state based on the fixed measurement posture data. When the fixed measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted, and the image information of the preset inkjet position is collected through the visual positioning system. In combination with the preset inkjet template, the inkjet device is controlled to perform the inkjet operation;

[0186] The third unit is used to collect high-definition images of the coding area after the coding operation is completed, use image processing algorithms to extract contour features of the coding pattern, compare the contour features with the preset coding template, and calculate the actual deviation of the coding position. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

[0187] According to a third aspect of the embodiments of the present invention,

[0188] An electronic device is provided, comprising:

[0189] processor;

[0190] a memory for storing processor-executable instructions;

[0191] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0192] A fourth aspect of the embodiments of the present invention is:

[0193] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0194] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time coding and position re-measurement of a fixed-measurement robot in signal construction, characterized in that: include: The construction area is scanned by a surveying robot to obtain three-dimensional point cloud data of the construction area, and the construction area is rasterized based on the three-dimensional point cloud data to generate a raster map. The image information of the construction area is collected by the surveying robot, and the positioning mark points in the construction area are identified by an image recognition algorithm. The spatial coordinates of the positioning mark points are mapped to the raster map, the distribution of the positioning mark points in the raster map is determined, and the optimal motion trajectory of the surveying robot is planned by using a time-series reinforcement iterative neural network algorithm. The measurement robot is controlled to move along the optimal motion trajectory, and the measurement posture data is collected in real time through the inertial measurement unit of the measurement robot. Based on the measurement posture data, the inkjet device of the measurement robot is controlled to maintain a horizontal state. When the measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted. The image information of the preset inkjet position is collected through the visual positioning system, and the inkjet device is controlled to perform the inkjet operation in combination with the preset inkjet template. After the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

2. The method according to claim 1, characterized in that Scanning the construction area with a fixed-measurement robot to obtain three-dimensional point cloud data of the construction area, and rasterizing the construction area based on the three-dimensional point cloud data to generate a raster map, including: Scanning the construction area with a fixed measurement robot to obtain three-dimensional point cloud data of the construction area, calculating the statistical mean and standard deviation of the three-dimensional point cloud data, determining a valid data range, removing abnormal data outside the valid data range, obtaining filtered three-dimensional point cloud data, dividing the filtered three-dimensional point cloud data into voxels according to a preset voxel size, calculating the centroid coordinates of the three-dimensional points contained in each voxel, and using the centroid coordinates as downsampling points of the corresponding voxels to obtain downsampled three-dimensional point cloud data; According to a preset radius, select neighboring 3D points within a corresponding radius range of each 3D point in the downsampled 3D point cloud data, construct a covariance matrix based on the neighboring 3D points, calculate the normal vector of each 3D point through the covariance matrix, determine the vector difference between the normal vectors, use the median of the vector difference as a Gaussian kernel parameter, construct a Gaussian kernel function based on the Gaussian kernel parameter, and calculate the significance score of each 3D point; The significance scores are sorted in descending order, and the top thirty percent of the three-dimensional points in the significance score sorting are selected as candidate feature points. The spacing between the candidate feature points is determined, and the candidate feature points whose spacing is less than a preset spacing threshold are eliminated to obtain a feature point set; the point feature histogram descriptor of each feature point in the feature point set is calculated, and feature matching is performed to obtain an initial registration result; Calculate the distance value and the normal angle value between the corresponding feature points in the initial registration result, use the median value of the distance value as the distance parameter, and the median value of the normal angle value as the angle parameter; construct a distance weight item based on the distance parameter, and construct an angle weight item based on the angle parameter, multiply the distance weight item by the angle weight item to obtain an adaptive weight, and use the adaptive weight to perform iterative optimization to obtain an accurate registration result; The construction area is divided into grids of a preset size, the distance between each grid and the three-dimensional point cloud data in the precise registration result is calculated, the observation likelihood of the Gaussian distribution is calculated, the observation likelihood is multiplied by the prior probability of the corresponding grid to obtain the probability product, the reciprocal of the cumulative sum of the probability products is calculated as a normalization factor, the probability product is normalized using the normalization factor to obtain the state probability of each grid, each grid is marked as occupied, idle or unknown based on the state probability, and a grid map of the construction area is generated.

3. The method according to claim 2, characterized in that Identify the positioning mark points in the construction area through the image recognition algorithm, and map the spatial coordinates of the positioning mark points to the grid map. Determine the distribution of the positioning mark points in the grid map including: The image collected by the measurement robot is input into the backbone feature network, and the backbone feature network extracts image features through the residual module. In the residual module, the spatial dimension and channel dimension of the feature map are weightedly reconstructed by using the spatial attention mechanism and the channel attention mechanism, respectively, to obtain the first-layer feature map; the first-layer feature map is input into the feature pyramid structure, and different scale features are fused through upsampling and downsampling operations to obtain a multi-level feature map; The multi-level feature map is input into the detection head network, and the detection head network outputs the category score and position coordinates of the positioning mark point. Based on the category score and the position coordinates, the two-dimensional pixel coordinates of the positioning mark point in the image are determined; based on the internal parameter matrix of the binocular camera, the image is distorted and corrected, and the parallax information of the left and right images is calculated. In combination with the external parameter matrix of the binocular camera, the two-dimensional pixel coordinates are converted into depth information through the triangulation principle, and the two-dimensional pixel coordinates are combined with the depth information to obtain the three-dimensional space coordinates of the positioning mark point; The corresponding feature points in the coordinate system corresponding to the camera coordinate system and the grid map are extracted, and feature point matching pairs are constructed. The rotation matrix and the translation vector are calculated, and the rotation matrix and the translation vector are optimized using a nonlinear optimization algorithm to obtain a coordinate transformation matrix. The three-dimensional space coordinates are mapped to the coordinate system of the grid map through the coordinate transformation matrix to obtain the spatial distribution of the positioning mark points.

4. The method according to claim 3, characterized in that: Using the time-series enhanced iterative neural network algorithm, the optimal motion trajectory of the robot is planned and measured, including: Constructing a time-series reinforcement iterative neural network, combining the real-time position information of the measuring robot, the spatial distribution of the positioning mark points and the local information of the grid map into a state vector, and combining the discretized linear velocity and angular velocity into an action vector, wherein the time-series reinforcement iterative neural network includes a main action network and a target value network; The main action network selects an action vector based on the current state vector, executes the action vector to obtain the next state vector, and the target value network estimates the value based on the next state vector; calculates the path length value, motion smoothness value and marker coverage value of the action vector, and combines them to form a reward value; The current state vector, action vector, reward value and next state vector are used to construct state transfer data, the time difference error of the state transfer data is calculated, priorities are assigned to the state transfer data based on the time difference error, and the state transfer data of different priorities are stored in an experience replay pool; state transfer data are sampled from the experience replay pool according to priority to train the main action network, the parameters of the main action network are updated to the target value network according to a preset number of steps, iterations are repeated until convergence, and the optimal motion trajectory for the measurement robot to visit the positioning mark point is output.

5. The method according to claim 1, characterized in that Controlling the measurement robot to move along the optimal motion trajectory, collecting measurement posture data in real time through the inertial measurement unit of the measurement robot, controlling the inkjet device of the measurement robot to maintain a horizontal state based on the measurement posture data, measuring and automatically adjusting the vertical distance between the nozzle of the inkjet device and the ground when the measurement robot moves to a preset inkjet position, collecting image information of the preset inkjet position through a visual positioning system, and controlling the inkjet device to perform inkjet operation in combination with a preset inkjet template, including: The inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by an extended Kalman filter algorithm, and the roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion; Based on the attitude quaternion, a kinematic model of the inkjet printer is established, the target angle of the first rotation axis and the target angle of the second rotation axis of the two-degree-of-freedom gimbal are calculated, the stepper motors of the first rotation axis and the second rotation axis are controlled to rotate respectively by a PID controller, and the inkjet printer is set in a horizontal state; When the preset coding position is reached, the vertical distance between the nozzle and the ground is measured by a laser distance sensor, and the vertical distance is compared with the preset vertical distance boundary to obtain the distance deviation value and the distance deviation change rate, establish a fuzzy control rule, input the fuzzy control rule into the fuzzy PID controller, adaptively adjust the PID parameters, control the speed of the DC servo motor of the lifting mechanism, and adjust the nozzle height; Collect an image of a preset inkjet printing position, perform grayscale and edge detection processing on the image to obtain contour features, perform feature point matching between a preset inkjet printing template and the contour features, determine the inkjet printing position coordinates and the spraying direction angle based on the matching results, generate a spraying instruction sequence, the spraying instruction sequence includes a spraying instruction sequence of nozzle position information, spraying pressure information and spraying time information, and send the spraying instruction sequence to the inkjet printing controller via a field bus; During the inkjet printing process, the spraying image is collected in real time, and the detection parameters of the spraying lines in the spraying image are analyzed, the detection parameters including width value, continuity value and clarity value, and the analysis results are generated. When any one of the detection parameters is lower than the corresponding preset threshold value, the spraying pressure information and the spraying time information are adaptively adjusted based on the analysis result until each detection parameter reaches the corresponding preset threshold value, and the inkjet printing operation is completed.

6. The method according to claim 5, characterized in that The inertial measurement unit of the robot is used to collect acceleration components and angular velocity components, and the attitude quaternion is estimated by the extended Kalman filter algorithm. The roll angle, pitch angle and yaw angle are calculated according to the attitude quaternion, including: The three-axis acceleration sensor of the inertial measurement unit is used to collect the acceleration components of gravity acceleration in three orthogonal axes, and the three-axis gyroscope is used to collect the angular velocity components of the robot rotating around the three orthogonal axes, and the acceleration components and the angular velocity components are sampled and filtered through the signal acquisition module; A system state equation is established based on the angular velocity component, the attitude quaternion is set as the state vector, the angular velocity is set as the system input, and a nonlinear state transfer model is constructed; an observation equation is established based on the acceleration component, the projection of the gravity vector in the body coordinate system is set as the observation quantity, and a nonlinear observation model is constructed according to the relationship between the attitude quaternion and the gravity vector; Performing a first-order Taylor expansion on the nonlinear state transfer model to obtain a state Jacobian matrix, and performing a first-order Taylor expansion on the nonlinear observation model to obtain an observation Jacobian matrix; The extended Kalman filter algorithm is used to estimate the attitude, including: using the state Jacobian matrix and the nonlinear state transfer model to perform state prediction to obtain a predicted attitude quaternion and a predicted state covariance; using the observation Jacobian matrix, the nonlinear observation model and the predicted state covariance to calculate the Kalman gain, multiplying the Kalman gain with the observation residual to obtain a state correction, adding the state correction to the predicted attitude quaternion to obtain an estimated attitude quaternion, and updating the state covariance at the same time; The estimated quaternion is normalized, and each component is divided by the corresponding modulus to obtain a unit attitude quaternion; according to the four components of the unit attitude quaternion, three rotation angles describing the robot attitude are calculated through a conversion formula from the unit attitude quaternion to the Euler angle, including a roll angle around the X-axis, a pitch angle around the Y-axis, and a yaw angle around the Z-axis.

7. The method according to claim 1, characterized in that After the coding operation is completed, a high-definition image of the coding area is collected, and the contour features of the coding pattern are extracted using an image processing algorithm. The contour features are compared with the preset coding template, and the actual deviation of the coding position is calculated. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report including: The spectral image of the coding area is collected by a hyperspectral camera, a grid pattern is projected by a structured light projector and a structured light image is collected, and the spectral image and the structured light image are input into a synchronous trigger circuit for timing alignment to obtain standard dual-modal image data; The standard bimodal image data is input into a spectral-spatial joint feature extraction network, the spectral attention module of the spectral-spatial joint feature extraction network processes the spectral image to obtain spectral features, the spatial pyramid pooling layer processes the structured light image to obtain spatial features, and the spectral features and the spatial features are fused through a cross-modal feature fusion module to obtain fused features of the coding area; Perform edge detection on the fused features to obtain the edge point set of the coding pattern, input the edge point set into the dynamic graph convolution network, extract the local structure information through the adaptive graph convolution layer, and use the graph attention mechanism to weight the local structure information to obtain the contour features of the coding pattern; The contour feature is matched with the preset coding template, and the initial posture deviation is obtained by performing a first-stage registration through a graph matching network. Based on the initial posture deviation, a second-stage registration is performed through an iterative closest point algorithm to obtain the actual posture deviation of the coding pattern; The actual posture deviation is input into a hierarchical reinforcement learning network, a high-level strategy network of the hierarchical reinforcement learning network generates a parameter optimization strategy based on the actual posture deviation, and a low-level execution network adjusts the motion parameters of the measurement robot based on the parameter optimization strategy; The motion parameters are input into a hybrid parameter optimizer, the hybrid parameter optimizer uses Bayesian optimization to determine a parameter search space, and an evolutionary algorithm is used to iteratively optimize the motion parameters in the parameter search space to obtain optimized motion parameters; Based on a pre-constructed uncertainty assessment model, the optimized motion parameters are input into the uncertainty assessment model, and the measurement uncertainty introduced by the measurement process, the model uncertainty introduced by the model prediction and the parameter uncertainty introduced by the parameter optimization are calculated respectively. The measurement uncertainty, the model uncertainty and the parameter uncertainty are input into the evidence theory model to obtain the credibility evaluation value of the parameter optimization result.

8. A real-time coding and position re-measurement system for a measuring robot in signal construction, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to scan the construction area by a surveying robot, obtain three-dimensional point cloud data of the construction area, perform rasterization processing on the construction area based on the three-dimensional point cloud data, generate a raster map, collect image information of the construction area by the surveying robot, identify positioning mark points in the construction area by an image recognition algorithm, map the spatial coordinates of the positioning mark points to the raster map, determine the distribution of the positioning mark points in the raster map, and use a time-series reinforcement iterative neural network algorithm to plan the optimal motion trajectory of the surveying robot; The second unit is used to control the movement of the fixed measurement robot along the optimal motion trajectory, collect fixed measurement posture data in real time through the inertial measurement unit of the fixed measurement robot, and control the inkjet device of the fixed measurement robot to maintain a horizontal state based on the fixed measurement posture data. When the fixed measurement robot moves to a preset inkjet position, the vertical distance between the nozzle of the inkjet device and the ground is measured and automatically adjusted, and the image information of the preset inkjet position is collected through the visual positioning system. In combination with the preset inkjet template, the inkjet device is controlled to perform the inkjet operation; The third unit is used to collect high-definition images of the coding area after the coding operation is completed, use image processing algorithms to extract contour features of the coding pattern, compare the contour features with the preset coding template, and calculate the actual deviation of the coding position. When the actual deviation exceeds the preset threshold, a deep learning model is used based on the actual deviation to adaptively optimize the motion parameters of the measurement robot, and the actual deviation is uploaded to the construction management system in real time to generate a construction quality report.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images