An Intelligent Task Scheduling and Management Method Based on Visual Inspection Robot Localization

Abstract

This invention discloses an intelligent task scheduling and management method based on vision-based inspection robot localization. This invention relates to the field of task scheduling and management technology, including digital archiving of inspection asset ledgers and binding with positioning targets, inspection task scheduling and path planning, precise absolute positioning execution of the robot, inspection data collection, coordinate labeling and reporting, intelligent identification and handling of abnormal events, and work order management. The advantages of this invention are: by using no fewer than four total stations to calibrate positioning targets for redundant absolute positioning, centimeter-level three-dimensional absolute coordinate positioning accuracy is achieved, fundamentally eliminating the cumulative error of relative positioning methods; the CSG semantic occlusion model can dynamically quantify the degree of scene occlusion; and adaptively scheduling multiple sensors such as binocular vision, LiDAR, inertial measurement unit, RTK-GPS, and ultra-wideband positioning module to work collaboratively, ensuring stable and reliable positioning in complex environments such as mild, moderate, and extreme occlusion.

Description

Technical Field

[0001] This invention relates to the field of task scheduling and management technology, specifically to an intelligent task scheduling and management method based on the localization of a vision-based inspection robot. Background Technology

[0002] With the rapid development of industrial automation, smart city operation and maintenance, and intelligent energy facilities, inspection operations, as a core link in ensuring equipment safety and stable facility operation, are gradually transforming from manual and semi-automatic inspections to autonomous and intelligent inspection modes. Visual inspection robots, with visual sensors as the core positioning and perception unit, integrate functions such as image acquisition, environmental recognition, and status detection. They can replace manual labor to complete routine inspection operations in scenarios such as substations, integrated pipe corridors, factory workshops, and tunnels, becoming an important hardware carrier for intelligent operation and maintenance systems. Against this backdrop, relying on visual positioning technology to achieve autonomous operation of inspection robots, coupled with an efficient task scheduling and management mechanism, has become an important application direction for data processing technology in the fields of administration and operation and maintenance management. Common intelligent task scheduling and management methods often rely on relative positioning, which can easily lead to continuous cumulative errors and make it difficult to guarantee long-term positioning accuracy. Single sensors are prone to failure in occluded environments, and positioning stability is poor in complex scenarios. The spherical positioning equation has a high degree of nonlinearity, making direct solution difficult and limiting accuracy. Furthermore, it lacks a reliable residual verification mechanism, making it difficult to achieve centimeter-level coordinate accuracy. It also cannot dynamically quantify the degree of occlusion, and multi-sensor scheduling is rigid. Under extreme occlusion conditions, it is prone to problems such as positioning interruption, drift, and insufficient reliability. To address these issues, we propose an intelligent task scheduling and management method based on vision inspection robot positioning. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent task scheduling and management method based on the positioning of visual inspection robots.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent task scheduling and management method based on visual inspection robot localization, comprising the following steps: Digital archiving and target binding of inspection asset ledgers: Obtain asset ledger information of industrial sites, establish a digital inspection point map including equipment number, equipment type, rated inspection parameters and spatial location, and pre-deploy no less than four circular positioning targets with absolute coordinates calibrated by a total station in the work site. Link and bind the three-dimensional absolute coordinates of each positioning target with the corresponding inspection point in the asset ledger to form a two-way index database of physical coordinates and asset information. Inspection task scheduling and path planning: Based on the preset inspection cycle strategy, task priority rules and the real-time position of each inspection robot, the task scheduling management engine dynamically generates an inspection task queue, uses the shortest path optimization algorithm to plan the inspection path, and issues task instructions containing target location identifiers, inspection item parameters and arrival time limits to the corresponding inspection robots. Robot Precise Absolute Positioning Execution: After the inspection robot moves to the target area and stops according to the task instructions, it performs adaptive binocular vision absolute positioning based on redundancy back calculation to obtain the three-dimensional absolute coordinates of the optical center of the robot's laser rangefinder in the total station coordinate system. The absolute positioning includes the following steps: CSG semantic occlusion model construction and adaptive sensor scheduling, depth feature map generation with joint compensation of binocular time difference and illumination, global target search and distance adaptive closed-loop target shooting, target physical eccentricity extraction and depth triangulation compensation, and linearized redundant multi-point coordinate inverse calculation based on difference order reduction. Inspection data collection, coordinate labeling and reporting: After the robot completes its positioning, it starts environmental data collection under precise three-dimensional absolute coordinates. The collected data is associated with the current three-dimensional absolute coordinates, UTC timestamp, task number and robot number and encapsulated into a structured inspection data packet, which is then uploaded to the inspection management information system via wireless network. Intelligent identification and handling work order management for abnormal events: After receiving the inspection data packet, the inspection management information system compares and analyzes it with the historical benchmark data and early warning threshold of the corresponding equipment to identify abnormal events. For the identified abnormal events, it automatically generates a handling work order containing precise location coordinates, abnormality type, severity and on-site image evidence. The work order is pushed to the corresponding handling entity according to the responsibility attribution rules, and the handling status of the work order is tracked throughout the process until the work order is closed.

[0005] As a further aspect of the present invention, it also includes multi-inspection robot collaborative scheduling management: the task scheduling management engine receives the three-dimensional absolute coordinates reported by each inspection robot in real time and dynamically displays the position of each robot in the digital inspection point map. When a robot fails to locate, the robot closest to the target point is automatically selected from the idle robot queue to take over the task, and the task queue is updated. When a robot task is detected to have timed out, the robot closest to the target location is automatically selected from the idle robot queue to take over the task, and the task queue is updated to ensure the continuity of inspection coverage.

[0006] As a further aspect of the present invention, it also includes: Inspection performance statistical analysis and management decision support: Summarize inspection execution data according to preset statistical cycles, calculate and output key performance indicators such as inspection coverage rate, task on-time completion rate, anomaly detection rate in each area, average positioning accuracy and robot effective operation time. High-risk inspection areas are identified based on the frequency and location distribution of historical anomalies. The inspection frequency of these areas is automatically increased, and a facility maintenance priority recommendation report is generated for management personnel, providing data support for preventive maintenance decisions.

[0007] As a further aspect of the present invention: the CSG semantic occlusion model construction and adaptive sensor scheduling specifically involve: The occlusion objects in the working environment are simplified into basic geometric shapes according to semantic categories, and a CSG semantic occlusion model is constructed. The boundary of the occlusion area is calculated by geometric projection and the occlusion level is determined by combining the motion state information of the occlusion objects. The sensor combination is adaptively selected according to the occlusion level, and the fusion weight of each sensor is assigned by the CSG semantic occlusion model. The occlusion level is divided into three levels: when the proportion of the occlusion area to the total area of ​​the observation field of view is less than the first preset threshold, it is judged as mild occlusion, and the binocular vision is the main method, supplemented by the lidar data of the occluded area. When the ratio is not less than the first preset threshold and less than the second preset threshold, it is determined to be a moderate occlusion, and LiDAR, inertial measurement unit and RTK-GPS are used for collaborative sensing. When the ratio is not less than the second preset threshold, it is determined to be extreme occlusion. The ultra-wideband positioning module and the inertial measurement unit are used as a backup, and the CSG semantic occlusion model provides the historical frame environment geometric constraints. The occlusion level determination results are synchronously reported to the task scheduling management engine. When the number of times the robot continuously determines that there is extreme occlusion at a certain inspection point reaches the preset consecutive number threshold, the task scheduling management engine automatically generates a site environment rectification suggestion work order and pushes it to the facility management personnel.

[0008] As a further aspect of the present invention: the depth feature map generation for the combined binocular time difference and illumination compensation specifically involves: Hardware-level clock synchronization of the binocular camera is performed by using an FPGA chip to send trigger pulses to the binocular camera synchronously through LVDS differential signal lines, keeping the acquisition time difference within the range of 100 picoseconds. ORB feature points are extracted from the left and right images and bidirectional matching is performed. The residual time deviation value is inferred from the displacement difference corresponding to the matched feature points. Based on the dynamic and static attributes of the image region, a multi-channel joint compensation process is performed: adaptive exposure compensation and parallax correction are performed sequentially for static regions, and dynamic interpolation correction and directional edge feature extraction are performed sequentially for dynamic regions. A three-way fusion calibration is performed using inertial measurement unit data, parallax correction, and illumination compensation coefficient to generate a depth feature map; In the aforementioned three-way fusion calibration, the weighting coefficient for motion offset is 0.4, the weighting coefficient for disparity correction is 0.4, and the weighting coefficient for illumination compensation is 0.2.

[0009] As a further aspect of the present invention: the global target search and distance-adaptive closed-loop target acquisition specifically comprises: Start the short focal length fixed-focus camera to scan and lock onto the circular positioning target, and extract the geometric center pixel coordinates and pixel diameter of the target surface; Based on the comparison results of the pixel diameter and the lens switching pixel threshold, a shooting mode is selected between the near-field shooting mode and the far-field shooting mode. The corresponding parallax baseline parameters and adaptive scale are retrieved to perform parallax compensation calculation. The gimbal deflection is input into the closed-loop controller to drive the gimbal motor. When the laser spot falls into the effective contour of the target surface, the gimbal motor is locked. The short focal length fixed-focus camera is an 8mm fixed-focus industrial camera, and the long focal length fixed-focus camera is a 75mm fixed-focus industrial camera. Both are rigidly connected to the laser rangefinder on a two-degree-of-freedom gimbal with pitch and yaw degrees of freedom, and the optical axes of the three satisfy the spatial parallel constraint. The closed-loop controller is an incremental PID controller that drives the gimbal motor at high frequency via an industrial communication bus.

[0010] As a further aspect of the present invention: the targeted physical eccentricity extraction, depth triangulation compensation, and linearized redundant multi-point coordinate inverse calculation based on difference reduction are specifically as follows: Extract the pixel coordinates of the laser spot landing point in the stable image frame, use an adaptive scale to convert the pixel deviation into the total physical eccentricity, read the original straight-line distance value returned by the laser rangefinder, and use the Pythagorean theorem to calculate the compensated true distance from the original straight-line distance value and the total physical eccentricity. For at least four known absolute coordinates of the positioning target, the shooting and ranging steps are performed respectively to obtain multiple sets of target point coordinates and the actual distance value after compensation. One set of target point data is selected as the reference. The spatial spherical equations of the remaining sets are algebraically subtracted from the reference spherical equations to eliminate the quadratic terms. The nonlinear equation system is transformed into an overdetermined linear equation system. The robot's three-dimensional absolute coordinates are obtained by solving the least squares method. Substitute the three-dimensional absolute coordinates back into the original spatial spherical equations of each group to calculate the reverse average residual; When the reverse average residual is lower than the preset accuracy tolerance, the coordinates are output as the positioning result and reported to the task scheduling management engine. When the reverse average residual is not lower than the preset accuracy tolerance, target data groups whose residual contribution exceeds the outlier threshold are removed and the difference reduction and least squares solution are re-executed. When the number of solution failures exceeds the preset solution threshold, the task scheduling management engine automatically marks the inspection task as a location anomaly and schedules a backup robot to take over.

[0011] As a further aspect of the present invention: the inspection data collection, coordinate labeling, and reporting also include: Each inspection data packet is attached with a positioning accuracy confidence label, which is calculated by combining the inverse average residual value, the number of effective targets involved in the calculation, and the current occlusion level. When identifying abnormal events, the inspection management information system sets a manual review flag for low-confidence data packets based on confidence level labels to ensure the reliability of the data on which management decisions are based; the inspection points corresponding to the low-confidence data packets are automatically added to the next round of priority re-inspection queue.

[0012] As a further aspect of the present invention: the task scheduling management engine also includes a target health status monitoring function. The probability of each positioning target being successfully identified in each positioning operation is statistically analyzed. When the identification success rate of a positioning target is lower than the preset health threshold, a positioning target inspection and maintenance work order is automatically generated and pushed to the facility operation and maintenance personnel. During target maintenance, the task scheduling management engine automatically blocks the target and re-verifies whether the number of remaining targets is no less than four. If the number of remaining targets is less than four, the inspection tasks that depend on that area are suspended and a management warning is issued.

[0013] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: 1. This invention employs at least four total stations to calibrate positioning targets for redundant absolute positioning. It linearizes the nonlinear spherical positioning equation by subtraction and order reduction, and combines least squares solution with a reverse average residual verification mechanism to achieve centimeter-level three-dimensional absolute coordinate positioning accuracy. This fundamentally eliminates the cumulative error of relative positioning methods. The CSG semantic occlusion model can dynamically quantify the degree of scene occlusion and adaptively schedule the collaborative work of multiple sensors such as binocular vision, lidar, inertial measurement unit, RTK-GPS, and ultra-wideband positioning module to ensure stable and reliable positioning in complex environments such as mild, moderate, and extreme occlusion. 2. This invention eliminates the systematic ranging error introduced by the physical eccentricity of the laser spot relative to the target center through targeted physical eccentricity extraction and depth triangulation compensation, thereby further improving the positioning accuracy. The FPGA hardware-level clock synchronization suppresses the binocular frame time difference to the order of 100 picoseconds. Combined with ORB feature point residual deviation back-inference and channel dynamic and static joint compensation, the accuracy of the depth feature map is significantly improved. 3. This invention organically integrates the robot positioning and execution layer, task scheduling layer, data quality management layer, abnormal work order management layer, and performance statistics and decision-making layer to construct a complete closed-loop intelligent inspection management system, which significantly improves the digitalization level of industrial facility management. Through mechanisms such as confidence labeling, target health monitoring, multi-machine collaborative takeover, and backup robot scheduling, the continuity and reliability of inspection tasks are guaranteed. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention. Detailed Implementation

[0015] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] This invention provides an intelligent task scheduling and management method based on the positioning of a vision-based inspection robot.

[0018] Therefore, in order to effectively solve the above problems, this application proposes an intelligent task scheduling and management method based on the localization of a vision-based inspection robot, as shown in the attached figures of the specification. Figure 1 As shown; Example 1: Digital filing and target binding of inspection asset ledgers: First, obtain complete asset ledger data from the facility management information system of the industrial site, and enter each equipment node that needs to be inspected regularly into the digital inspection point map. Each node records the equipment number, equipment type, rated inspection parameters and its spatial position in the engineering coordinate system. In the work area, professional surveyors use a total station to calibrate the absolute coordinates of pre-installed circular positioning targets. At least four positioning targets are set up in each inspection area to obtain the three-dimensional coordinates of each target center in the total station coordinate system (i.e., the global absolute coordinate system). subscript Indicates the first One positioning target, , The above three-dimensional absolute coordinates are associated and bound with the corresponding inspection points in the asset ledger to form a two-way index database of physical coordinates and asset information, which is stored in the inspection management information system. Example 2: Inspection Task Scheduling and Path Planning: The task scheduling and management engine dynamically generates an inspection task queue based on the preset inspection cycle strategy (including daily inspection cycle, key area intensive inspection cycle and high-risk area priority weight) and the three-dimensional absolute coordinates reported by each robot in real time. It uses the shortest path optimization algorithm to plan the path of the inspection points in the task queue, generates the optimal traversal path, and then issues task instructions to the corresponding inspection robot. The task instructions include the target positioning target identifier, the inspection item parameters to be collected, and the arrival time limit. Example 3: Robot performs precise absolute positioning: Precise absolute positioning of a robot includes the following five sub-processes; Sub-process 1: Construction of CSG semantic occlusion model and adaptive sensor scheduling; After the robot arrives at the target area, the vision system first performs semantic segmentation on the scene, simplifying the identified occlusions (shelves, vehicles, people, etc.) into basic geometric shapes such as spheres, cylinders, and cuboids according to semantic categories. It then constructs a CSG semantic occlusion model using the Constructive Solid Geometry (CSG) modeling method. Combining the motion state information of the occlusions (static occlusions correspond to static geometric constraints, and moving people correspond to dynamic envelopes), it calculates the boundary of the occlusion area in the current field of view through geometric projection, calculates the proportion of the occlusion area to the total area of ​​the observed field of view, and determines the occlusion level. The occlusion level is divided into three levels: when the occlusion area ratio is less than the first preset threshold, it is judged as mild occlusion, and the sensor solution is mainly based on binocular vision, and supplemented by LiDAR data of the occluded area; when the occlusion area ratio is not less than the first preset threshold and less than the second preset threshold, it is judged as moderate occlusion, and the sensor solution switches to LiDAR, inertial measurement unit and RTK-GPS collaborative perception; when the occlusion area ratio is not less than the second preset threshold, it is judged as extreme occlusion, and the sensor solution switches to ultra-wideband positioning module and inertial measurement unit as a backup, and the CSG semantic occlusion model provides historical frame environmental geometric constraints to assist positioning. The CSG semantic occlusion model dynamically allocates the fusion weights of each sensor according to the determined occlusion level and synchronously reports the judgment results to the task scheduling management engine. When the number of times the robot continuously judges extreme occlusion at a certain inspection point reaches the preset consecutive number threshold, the task scheduling management engine automatically generates a site environment rectification suggestion work order and pushes it to the facility management personnel. Sub-process 2: Generation of depth feature maps with joint compensation of binocular time difference and illumination; To address the issues of time difference and uneven illumination between the left and right frames of a stereo camera, a hardware-level clock synchronization scheme is adopted: the FPGA chip simultaneously sends trigger pulses to the left and right stereo cameras via LVDS differential signal lines, controlling the time difference between the left and right frame acquisitions to the order of 100 picoseconds. ORB feature points are extracted from the left and right images and bidirectional matching is performed. The residual time deviation value is inferred from the pixel displacement difference of the matched feature point pairs. ; Dynamic and static analysis is performed on each region of the image, and joint compensation is applied to each channel: for static regions, adaptive exposure compensation (to eliminate brightness deviations introduced by uneven illumination) and parallax correction (based on...) are applied sequentially. Time difference compensation is performed on the disparity map, and dynamic interpolation correction is sequentially performed on the dynamic region (based on...). Interpolation correction of moving target position) and directional edge feature extraction (extracting edge direction features to enhance the robustness of depth estimation of moving region); A three-way fusion calibration is performed using inertial measurement unit data, disparity correction, and illumination compensation coefficient to generate a depth feature map. The weighted calculation formula for the three-way fusion calibration is as follows: ; in, This is the combined calibration offset after fusion (unit: pixels). Motion offset (unit: pixels) calculated from inertial measurement unit data. This is the parallax correction amount (unit: pixels). This is the image brightness offset correction amount (unit: pixels) calculated from the illumination compensation coefficient. The weighting coefficient for the motion offset. The weighting coefficients for the disparity correction amount. These are the weighting coefficients for the illumination compensation coefficient. The sum of the three weighting coefficients is 1. The weighting values ​​were determined by the researchers of this invention through extensive experimental calibration. Sub-process 3: Global target search and distance-adaptive closed-loop target acquisition; The 8mm fixed-focus short-focal-length industrial camera mounted on a two-degree-of-freedom gimbal is activated, driving the gimbal to scan within the global field of view. The target is located by locking the target through a circular target detection algorithm, and the geometric center pixel coordinates and pixel diameter of the target are extracted. Based on the comparison results between the extracted pixel diameter and the lens switching pixel threshold, a shooting mode is selected between the near-field shooting mode and the far-field shooting mode: when the pixel diameter is smaller than the lens switching pixel threshold, the far-field shooting mode is selected, and the 75mm fixed-focus telephoto industrial camera parameter group is switched to perform fine alignment; when the pixel diameter is not smaller than the lens switching pixel threshold, the near-field shooting mode is selected, and the 8mm fixed-focus camera parameter group is maintained. The parallax baseline parameters and adaptive scale corresponding to the selected target firing mode are retrieved to perform parallax compensation calculation. The required gimbal deflection is input into the incremental PID closed-loop controller. The controller drives the gimbal motor to adjust the pitch and yaw angles by high-frequency output control commands through the industrial communication bus. When the laser rangefinder spot falls into the effective contour of the target surface, the gimbal motor is locked to complete the target firing. The optical axes of the 8mm camera, 75mm camera and laser rangefinder satisfy the spatial parallel constraint. The three are rigidly connected to the two-degree-of-freedom gimbal. The above target firing steps are performed on no less than four positioning targets one by one. Sub-process 4: Targeted physical eccentricity extraction and depth triangulation compensation; After the gimbal locks, extract the pixel coordinates of the laser spot's landing point from the stable image frame. Simultaneously extract the pixel coordinates of the geometric center of the target surface. Calculate the horizontal pixel deviation between the two. Pixel deviation in the vertical direction Utilizing the adaptive scale corresponding to the current shooting mode Pixel deviation is converted into total physical eccentricity. The calculation formula is: ; in, The total physical eccentricity (in meters) between the laser spot landing point and the geometric center of the target surface within the target plane. The adaptive scale (unit: meters / pixel) is determined by the focal length parameter of the current firing mode and the currently estimated target distance. Horizontal pixel deviation (unit: pixels). Vertical pixel deviation (unit: pixels); Read the raw linear distance measurement value returned by the laser rangefinder (i.e., the straight-line distance from the optical center of the laser rangefinder to the point where the laser spot falls, in meters), calculate the compensated true distance using the Pythagorean theorem. (That is, the true straight-line distance from the optical center of the laser rangefinder to the geometric center of the target surface, in meters), the calculation formula is: ; in, The original straight-line distance measurement value (unit: meters) obtained by the laser rangefinder. Total physical eccentricity (unit: meters). To compensate for the physical eccentricity error and determine the true distance (in meters), when Compared to When the error is small, this compensation eliminates the systematic excessive ranging error introduced by eccentricity; Sub-process 5: Linearized redundant multi-point coordinate inverse calculation based on difference reduction; Let the three-dimensional absolute coordinates of the optical center of the robot's laser rangefinder be the unknown quantity. , for the One positioning target (the three-dimensional absolute coordinates of the target center are...) The actual distance after compensation is Establish the equations for a spatial sphere: ; in, For the first The three-dimensional absolute coordinates of the target center (unit: meters, obtained from total station calibration, are known quantities). For the laser rangefinder optical center to the first The compensated true distance of each target center (unit: meters, calculated by subprocess four, is a known quantity). The total number of positioning targets involved in the calculation (no less than four); Using the data from the first target as a baseline, the data from the second target were analyzed. ( After expanding the equations of the first sphere, subtract the expansion of the first sphere equation algebraically to eliminate the errors. , , The quadratic term yields information about The linear equation: ; in, The three-dimensional absolute coordinate components (unit: meters) of the target center (first positioning target) of the reference target point. The distance from the optical center of the laser rangefinder to the reference target point after compensation (unit: meters) is the true distance (in meters). All terms on the right side of the equation are known quantities. At that time, a total of [amount] can be obtained. A linear equation is formed, which relates to three unknowns. The overdetermined linear equations were solved using the least squares method to obtain the estimated three-dimensional absolute coordinates. ; Will Substitute the original spatial spherical equations into each set and calculate the inverse average residual. : ; in, The reverse average residual (unit: meters) reflects the overall accuracy of the positioning results. The three-dimensional absolute coordinates estimated by the least squares method (unit: meters) have the same meaning as the terms in the absolute value sign as defined in the above spherical equation; when When the output is below the preset accuracy tolerance, To locate the results and report them to the task scheduling management engine, when When the accuracy tolerance is not lower than the preset tolerance, calculate each target pair one by one. The contribution amount is used to remove target data groups whose contribution amount exceeds the outlier threshold. The difference reduction and least squares solution are re-executed. When the number of solution failures exceeds the preset solution threshold, the task scheduling management engine automatically marks the inspection task as a location anomaly and schedules a backup robot to take over. Example 4: Inspection data collection, coordinate labeling, and reporting: After completing precise positioning, the robot starts collecting environmental data from various sensors (including temperature, humidity, smoke concentration, visible light images, etc.) under the obtained precise three-dimensional absolute coordinates. The collected raw data is associated and encapsulated with the current three-dimensional absolute coordinates, UTC timestamp, task number, and robot number to generate a structured inspection data package. Meanwhile, based on the reverse average residual value of this positioning... The number of valid targets participating in this calculation and the current level of occlusion are used to comprehensively calculate the confidence level of positioning accuracy. The confidence level label is attached to the data packet. The value of the confidence level label comprehensively reflects the reliability of the positioning result and is used for subsequent data quality classification processing. The structured inspection data packet is uploaded to the inspection management information system through the wireless network. Example 5: Intelligent Identification and Handling of Abnormal Events and Work Order Management: After receiving the inspection data packet, the inspection management information system sets a manual review flag for low-confidence data packets based on the confidence level label to ensure the reliability of the data on which management decisions are based. The corresponding inspection point is automatically added to the next round of priority review queue. For high-confidence data packets, the system automatically compares and analyzes them with the historical baseline data of the equipment and the preset warning threshold to identify abnormal events. For identified abnormal events, the system automatically generates a handling work order. The work order includes precise location coordinates, abnormality type, severity, and on-site image evidence. It is pushed to the corresponding handling entity according to the responsibility attribution rules, and the handling status of the work order is tracked throughout the process until the work order is closed. Example 6: Multi-machine collaborative scheduling and other management functions: The task scheduling and management engine receives the three-dimensional absolute coordinates reported by each inspection robot in real time and dynamically displays the position of each robot on the digital inspection point map. When a robot fails to locate, it automatically selects the robot closest to the target point from the idle robot queue to take over the task and updates the task queue. When a robot's task times out, it also automatically executes the same task takeover process to ensure the continuity of inspection coverage. The task scheduling and management engine continuously monitors the health status of the positioning targets and calculates the success rate of each positioning target in each positioning operation. When the success rate of a positioning target is lower than the preset health threshold, a positioning target inspection and maintenance work order is automatically generated and pushed to the facility operation and maintenance personnel. During the maintenance period, the target is blocked. After blocking, the system re-verifies whether the number of remaining targets is not less than four. If not, the inspection task dependent on the area is suspended and a management warning is issued. According to the preset statistical cycle, the system summarizes the inspection execution data, calculates and outputs key performance indicators such as inspection coverage rate, task on-time completion rate, anomaly detection rate in each area, average positioning accuracy and robot effective working time. Based on the historical anomaly frequency and location distribution, it identifies high-risk inspection areas, automatically increases the inspection frequency of the area, and generates facility maintenance priority suggestion reports for managers, providing data support for preventive maintenance decisions. While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent task scheduling management method based on visual inspection robot positioning, characterized in that: Includes the following steps: Digital archiving and target binding of inspection asset ledgers: Obtain asset ledger information of industrial sites, establish a digital inspection point map including equipment number, equipment type, rated inspection parameters and spatial location, and pre-deploy no less than four circular positioning targets with absolute coordinates calibrated by a total station in the work site. Link and bind the three-dimensional absolute coordinates of each positioning target with the corresponding inspection point in the asset ledger to form a two-way index database of physical coordinates and asset information. Inspection task scheduling and path planning: Based on the preset inspection cycle strategy, task priority rules and the real-time position of each inspection robot, the task scheduling management engine dynamically generates an inspection task queue, uses the shortest path optimization algorithm to plan the inspection path, and issues task instructions containing target location identifiers, inspection item parameters and arrival time limits to the corresponding inspection robots. Robot Precise Absolute Positioning Execution: After the inspection robot moves to the target area and stops according to the task instructions, it performs adaptive binocular vision absolute positioning based on redundancy back calculation to obtain the three-dimensional absolute coordinates of the optical center of the robot's laser rangefinder in the total station coordinate system. The absolute positioning includes the following steps: CSG semantic occlusion model construction and adaptive sensor scheduling, depth feature map generation with joint compensation of binocular time difference and illumination, global target search and distance adaptive closed-loop target shooting, target physical eccentricity extraction and depth triangulation compensation, and linearized redundant multi-point coordinate inverse calculation based on difference order reduction. Inspection data collection, coordinate labeling and reporting: After the robot completes its positioning, it starts environmental data collection under precise three-dimensional absolute coordinates. The collected data is associated with the current three-dimensional absolute coordinates, UTC timestamp, task number and robot number and encapsulated into a structured inspection data packet, which is then uploaded to the inspection management information system via wireless network. Intelligent identification and handling work order management for abnormal events: After receiving the inspection data packet, the inspection management information system compares and analyzes it with the historical benchmark data and early warning threshold of the corresponding equipment to identify abnormal events. For the identified abnormal events, it automatically generates a handling work order containing precise location coordinates, abnormality type, severity and on-site image evidence. The work order is pushed to the corresponding handling entity according to the responsibility attribution rules, and the handling status of the work order is tracked throughout the process until the work order is closed. 2.The intelligent task scheduling management method based on visual inspection robot positioning according to claim 1, characterized in that, It also includes multi-inspection robot collaborative scheduling management: the task scheduling management engine receives the three-dimensional absolute coordinates reported by each inspection robot in real time and dynamically displays the position of each robot in the digital inspection point map; When a robot fails to locate, the robot closest to the target point is automatically selected from the idle robot queue to take over the task, and the task queue is updated. When a robot task is detected to have timed out, the robot closest to the target location is automatically selected from the idle robot queue to take over the task, and the task queue is updated to ensure the continuity of inspection coverage. 3.The intelligent task scheduling management method based on visual inspection robot positioning according to claim 1, characterized in that: Also includes: Inspection performance statistical analysis and management decision support: Summarize inspection execution data according to preset statistical cycles, calculate and output key performance indicators such as inspection coverage rate, task on-time completion rate, anomaly detection rate in each area, average positioning accuracy and robot effective operation time. High-risk inspection areas are identified based on the frequency and location distribution of historical anomalies. The inspection frequency of these areas is automatically increased, and a facility maintenance priority recommendation report is generated for management personnel, providing data support for preventive maintenance decisions.

4. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The CSG semantic occlusion model construction and adaptive sensor scheduling are specifically as follows: The occlusion objects in the working environment are simplified into basic geometric shapes according to semantic categories, and a CSG semantic occlusion model is constructed. The boundary of the occlusion area is calculated by geometric projection and the occlusion level is determined by combining the motion state information of the occlusion objects. The sensor combination is adaptively selected according to the occlusion level, and the fusion weight of each sensor is assigned by the CSG semantic occlusion model. The occlusion level is divided into three levels: when the proportion of the occlusion area to the total area of ​​the observation field of view is less than the first preset threshold, it is judged as mild occlusion, and the binocular vision is the main method, supplemented by the lidar data of the occluded area. When the ratio is not less than the first preset threshold and less than the second preset threshold, it is determined to be a moderate occlusion, and LiDAR, inertial measurement unit and RTK-GPS are used for collaborative sensing. When the ratio is not less than the second preset threshold, it is determined to be extreme occlusion. The ultra-wideband positioning module and the inertial measurement unit are used as a backup, and the CSG semantic occlusion model provides the historical frame environment geometric constraints. The occlusion level determination results are synchronously reported to the task scheduling management engine. When the number of times the robot continuously determines that there is extreme occlusion at a certain inspection point reaches the preset consecutive number threshold, the task scheduling management engine automatically generates a site environment rectification suggestion work order and pushes it to the facility management personnel.

5. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The generation of the depth feature map, which is jointly compensated for by binocular time difference and illumination, is specifically as follows: Hardware-level clock synchronization of the binocular camera is performed by using an FPGA chip to send trigger pulses to the binocular camera synchronously through LVDS differential signal lines, keeping the acquisition time difference within the range of 100 picoseconds. ORB feature points are extracted from the left and right images and bidirectional matching is performed. The residual time deviation value is inferred from the displacement difference corresponding to the matched feature points. Based on the dynamic and static attributes of the image region, a multi-channel joint compensation process is performed: adaptive exposure compensation and parallax correction are performed sequentially for static regions, and dynamic interpolation correction and directional edge feature extraction are performed sequentially for dynamic regions. A three-way fusion calibration is performed using inertial measurement unit data, parallax correction, and illumination compensation coefficient to generate a depth feature map; In the aforementioned three-way fusion calibration, the weighting coefficient for motion offset is 0.4, the weighting coefficient for disparity correction is 0.4, and the weighting coefficient for illumination compensation is 0.

2.

6. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The global target search and distance-adaptive closed-loop target acquisition specifically refers to: Start the short focal length fixed-focus camera to scan and lock onto the circular positioning target, and extract the geometric center pixel coordinates and pixel diameter of the target surface; Based on the comparison results of the pixel diameter and the lens switching pixel threshold, a shooting mode is selected between the near-field shooting mode and the far-field shooting mode. The corresponding parallax baseline parameters and adaptive scale are retrieved to perform parallax compensation calculation. The gimbal deflection is input into the closed-loop controller to drive the gimbal motor. When the laser spot falls into the effective contour of the target surface, the gimbal motor is locked. The short focal length fixed-focus camera is an 8mm fixed-focus industrial camera, and the long focal length fixed-focus camera is a 75mm fixed-focus industrial camera. Both are rigidly connected to the laser rangefinder on a two-degree-of-freedom gimbal with pitch and yaw degrees of freedom, and the optical axes of the three satisfy the spatial parallel constraint. The closed-loop controller is an incremental PID controller that drives the gimbal motor at high frequency via an industrial communication bus.

7. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The targeted physical eccentricity extraction, depth triangulation compensation, and linearized redundant multi-point coordinate inverse calculation based on difference reduction are specifically as follows: Extract the pixel coordinates of the laser spot landing point in the stable image frame, use an adaptive scale to convert the pixel deviation into the total physical eccentricity, read the original straight-line distance value returned by the laser rangefinder, and use the Pythagorean theorem to calculate the compensated true distance from the original straight-line distance value and the total physical eccentricity. For at least four known absolute coordinates of the positioning target, the shooting and ranging steps are performed respectively to obtain multiple sets of target point coordinates and the actual distance value after compensation. One set of target point data is selected as the reference. The spatial spherical equations of the remaining sets are algebraically subtracted from the reference spherical equations to eliminate the quadratic terms. The nonlinear equation system is transformed into an overdetermined linear equation system. The robot's three-dimensional absolute coordinates are obtained by solving the least squares method. Substitute the three-dimensional absolute coordinates back into the original spatial spherical equations of each group to calculate the reverse average residual; When the reverse average residual is lower than the preset accuracy tolerance, the coordinates are output as the positioning result and reported to the task scheduling management engine. When the reverse average residual is not lower than the preset accuracy tolerance, target data groups whose residual contribution exceeds the outlier threshold are removed and the difference reduction and least squares solution are re-executed. When the number of solution failures exceeds the preset solution threshold, the task scheduling management engine automatically marks the inspection task as a location anomaly and schedules a backup robot to take over.

8. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The inspection data collection, coordinate labeling, and reporting also include: Each inspection data packet is attached with a positioning accuracy confidence label, which is calculated by combining the inverse average residual value, the number of effective targets involved in the calculation, and the current occlusion level. When identifying abnormal events, the inspection management information system sets a manual review flag for low-confidence data packets based on confidence level labels to ensure the reliability of the data on which management decisions are based; the inspection points corresponding to the low-confidence data packets are automatically added to the next round of priority re-inspection queue.

9. The intelligent task scheduling and management method based on vision inspection robot localization according to claim 1, characterized in that, The task scheduling and management engine also includes a target health status monitoring function: The probability of each positioning target being successfully identified in each positioning operation is statistically analyzed. When the identification success rate of a positioning target is lower than the preset health threshold, a positioning target inspection and maintenance work order is automatically generated and pushed to the facility operation and maintenance personnel. During target maintenance, the task scheduling management engine automatically blocks the target and re-verifies whether the number of remaining targets is no less than four. If the number of remaining targets is less than four, the inspection tasks that depend on that area are suspended and a management warning is issued.

Images