Virtual maintenance data processing system for maintenance scene optimization

Abstract

The invention relates to the technical field of data processing, and particularly discloses a virtual maintenance data processing system for maintenance scene optimization. Comprising an optical contact probe, a scanning measurement system, an image recognition and defect analysis module, an image auxiliary measurement module, a part visual tracking module, an augmented reality assembly guide module, a data fusion and state modeling module and a man-machine interaction and task excitation module. By fusing optical contact measurement, image recognition, point cloud modeling and augmented reality guidance, intelligent and high-precision execution of defect detection, size measurement, assembly deviation correction and state modeling in the disassembly and reassembly process of the gearbox is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and more specifically, to a virtual maintenance data processing system for optimizing maintenance scenarios. Background Art

[0002] The charging pump (high-pressure charging pump) is a key safety system, and its gearbox undertakes the task of high-load and high-precision power transmission. Due to the complex long-term operating environment, the gearbox is prone to gear meshing abnormalities (such as eccentric wear, pitting, and spalling), bearing damage, unqualified matching clearance, and reinstallation errors. During the maintenance process, the disassembly and restoration of the gearbox include typical high-difficulty processes such as precise disassembly and marking, multi-dimensional measurement and recording (matching surface, tooth pitch, meshing angle), status assessment and damage judgment, parts repair / replacement and reinstallation, reinstallation accuracy confirmation and function verification. During the disassembly process, there are many parts, complex structures, confusing position markings, and a high risk of reinstallation errors. Traditional measuring tools are inefficient for high-precision clearance, tooth shape, and meshing angle measurement, and the measurement methods are complex. The identification of tooth surface defects is highly subjective, and it is difficult to achieve standardized maintenance based on experience. The complex assembly structure of the gearbox maintenance will lead to difficulties in identifying reinstallation posture errors, low accuracy in identifying structural defects, scattered measurement data that is difficult to integrate and model, and manual task execution is prone to errors and lacks an incentive feedback mechanism, as well as a lack of intelligent adaptation and learning capabilities. Summary of the invention

[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a virtual maintenance data processing system for optimizing maintenance scenarios, which realizes the precision, intelligence and visualization of the maintenance process through virtual maintenance data processing with the coordinated operation of multi-source fusion, augmented reality guidance, state modeling and task incentives.

[0004] To achieve the above object, the present invention provides the following technical solutions: A virtual maintenance data processing system for maintenance scenario optimization includes an optical contact probe, a scanning measurement system, an image recognition and defect analysis module, an image-assisted measurement module, a part visual tracking module, an augmented reality assembly guidance module, a data fusion and state modeling module, and a human-computer interaction and task incentive module. The optical contact probe performs contact measurement of nanometer-level size and displacement data of key parts of gear meshing surfaces, shaft shoulders, matching holes, and gaskets in a gearbox. The scanning measurement system obtains three-dimensional point cloud data of key areas of the gearbox. The image recognition and defect analysis module collects image data of target areas of the gearbox, detects defects such as tooth surface wear, cracks, pitting, and spalling through a recognition algorithm based on deep learning, and jointly judges the posture deviation and poor contact of the assembly state. The image-assisted measurement module performs geometric analysis through calibrated benchmarks and structural feature points in the figure, and forms a multi-source comparison with the point cloud model and contact measurement results. The part visual tracking module performs geometric analysis through calibrated benchmarks and structural feature points in the figure, and forms a multi-source comparison with the point cloud model and contact measurement results. The image and unique identification of parts are recorded during the process, the whole process tracking and assembly sequence are managed through image registration, and the state model is combined to detect missing parts and verify assembly during the reinstallation stage. The augmented reality assembly guidance module spatially aligns the standard 3D assembly model with the current on-site image, identifies assembly deviations in real time and generates a posture error heat map, combines image recognition and point cloud data to jointly drive assembly correction guidance, and provides posture adjustment prompts with multi-source information enhancement. The data fusion and state modeling module fuses the optical probe measurement results, image recognition output and point cloud morphology data to build a digital twin model of the gearbox assembly status, and generates personalized virtual maintenance suggestions through trend analysis, structural degradation judgment and residual life assessment. When the human-computer interaction and task incentive module identifies potential risk areas and key operation nodes, it actively generates measurement and assembly tasks, and guides operators to complete the gearbox maintenance actions in combination with task scoring, behavior confidence assessment and achievement incentive strategies.

[0005] As a further technical solution of the present invention, the image recognition and defect analysis module includes a recognition unit based on a deep convolutional neural network. The network structure of the recognition unit includes a feature extraction layer, a region proposal network and a classification regression branch. By introducing an attention channel, weighted learning is performed on the spatial area of ​​the gear surface image to extract crack edge and texture degradation features in the presence of local occlusion and oil interference.

[0006] As a further technical solution of the present invention, the image-assisted measurement module includes a two-dimensional image correction unit and a size conversion module, and is provided with image and point cloud error comparison control logic. The two-dimensional image correction unit is used to perform geometric distortion correction and angle projection correction on the collected original image, solve the image measurement error problem caused by lens distortion, shooting angle tilt or perspective distortion, correct fisheye distortion, barrel or pincushion deformation, and convert the tilted image into a standard front view through geometric transformation to ensure the geometric authenticity of the calibration reference and structure edge in the image, laying the foundation for subsequent precise size extraction. The size conversion module converts the pixel unit in the image into the actual physical size unit (such as mm), and performs structural feature The size calculation is based on the calibrated reference objects in the drawing (such as rulers, correction blocks) or preset scale factors, calculates the pixel-millimeter conversion coefficient, extracts the geometric distances between key structural edges, hole positions, center points and other features, and outputs two-dimensional measurement data for comparison with point cloud measurement results or probe measurement results. It supports dimensional anomaly warnings and preliminary judgments on assembly deviations. When the difference between the image measurement result and the point cloud measurement result is greater than 0.5mm, the image and point cloud error comparison control logic performs automatic image recalibration and size mapping coefficient adjustment operations until the measurement error is lower than the set tolerance range of 0.5mm. The scanning measurement system is a structured light scanner with a point pitch resolution of 0.05mm and a scanning range of 300mm. 300mm. During the acquisition process, point cloud data of five viewing angles are obtained through several angles. The point cloud data are used to generate a three-dimensional model of the gearbox surface after uniform posture alignment.

[0007] As a further technical solution of the present invention, the augmented reality assembly guidance module includes an image feature point matching submodule and a point cloud registration submodule. The image feature point matching submodule (used to extract and match key feature points between the real-time collected assembly site image and the standard assembly model image, and realize the preliminary positioning and posture estimation of the assembly parts in the two-dimensional image) and the point cloud registration submodule (used to spatially align the three-dimensional point cloud data obtained by structured light scanning or depth camera with the standard three-dimensional assembly model, and accurately calculate the position and posture of the current part in the three-dimensional space) are connected to a heat map generation unit. The outputs of the image feature point matching submodule and the point cloud registration submodule are combined to calculate the three-dimensional rigid body transformation matrix of the assembly part relative to the standard assembly model through least squares registration, and when the rotation amount of the three-dimensional rigid body transformation matrix is ​​greater than 3° and the translation amount is greater than 2 mm, the heat map generation unit is triggered to superimpose a red warning block in the AR display.

[0008] As a further technical solution of the present invention, in the augmented reality assembly guidance module, the process of detecting the position and posture changes of the assembly parts in the three-dimensional space when the operator performs the reassembly operation includes: Step 1: Image acquisition and initial feature extraction: Collect the image of the reassembled component, apply the SHIF algorithm to extract the feature points of the reassembled component, and call the standard assembly model to establish the target area to be registered; Step 2: Feature matching registration based on center point correction: perform registration correction on the center point calculation of the bolt on the gear in the detection target. If the overall component registration deviation exceeds the preset threshold within the set range, but the target center point alignment error is less than the set threshold, it is judged as surface misleading registration, and the transformation matrix is ​​recalculated using the center point as the reference; Step 3: Edge contour detection and straight line fitting: extract the edge contour of the target component, use the Hough straight line fitting method to determine its boundary direction, and obtain the assembly rotation angle error by comparing the angle between the boundary direction and the standard direction; Step 4: Point cloud-assisted spatial posture detection: Based on the point cloud data obtained by the scanning measurement system, it automatically switches to 3D point cloud posture analysis after image registration, maps the area of ​​interest to the point cloud space, intercepts the point cloud of the target area, performs RANSAC plane fitting on the target assembly surface, analyzes the spatial distance, parallelism, and angle between the top and bottom surfaces, and determines whether there is any misalignment or incomplete insertion; Step 5, Error analysis and type judgment: Comprehensive image and point cloud information, classify and judge assembly errors, calculate translation error through center point offset or point cloud center point difference, calculate rotation error through boundary line direction or fitted plane normal vector, calculate vertical error through point cloud surface distance to judge whether it is pressed tightly, and identify incomplete insertion through edge contour difference map area greater than the set threshold or incomplete fitting area; Step 6: Abnormal filtering and self-learning: For areas with abnormal center point distribution and abnormal contour disturbance, perform an evaluation of the assembly error recognition optimization mechanism based on dynamic confidence weighting, and adaptively optimize the recognition algorithm.

[0009] As a further technical solution of the present invention, in the augmented reality assembly guidance module, an assembly error recognition optimization mechanism based on dynamic confidence weighting generates a confidence score for each assembly error recognition result. The confidence score is obtained by weighted summation of the number of matching points, boundary integrity, point cloud surface closure rate, center point offset and image clarity. If the confidence score is lower than the set threshold, the control pauses the task push and prompts the operator to re-collect the image. The assembly error recognition optimization mechanism based on dynamic confidence weighting includes a regression learning model based on operation behavior correction feedback, which records and marks the error correction between the operator's actual correction behavior and the recognition guidance result during the assembly process, and dynamically adjusts the parameter weights of image and point cloud recognition through a lightweight regression learning model to perform self-learning, self-optimization and confidence adaptive prompts for the assembly error recognition algorithm.

[0010] As a further technical solution of the present invention, the human-computer interaction and task incentive module includes a behavior data backtracking and strategy optimization mechanism, which is used to record the operator's execution path, assembly adjustment behavior, error level change and confidence fluctuation process in each task, generate a behavior log based on time series and performance labels, and adaptively optimize and adjust the recommended path, recognition parameter configuration and incentive level weight of subsequent tasks. The human-computer interaction and task incentive module uses a task scoring formula to calculate the task score, and the task scoring formula is:

[0011] Where: Scoring tasks is used to cumulatively update the operator's incentive level based on this score, and to control their task unlocking permissions and challenge prompt push. Scoring task benchmarks, which are pre-set by administrators and engineers for each type of task. is the measurement error rate, which is equal to the mean absolute error rate of the optical touch probe, ,in is the number of measurement points involved in the task, The index of the measurement point for this task, , The tasks are The measured and expected values ​​of the measurement points, To identify the confidence coefficient, the Softmax layer classification probability output from the deep learning network of the image recognition module, ,in is the normalized value of the matching points and the total points, To identify the mean of the pair residuals, is the normalized value of the boundary closure rate, is the normalized value of the center point offset ratio, is the normalized value for measuring image clarity, , , , , is the weight factor, , automatically lock the task submission button and pop up a prompt box "Please collect again".

[0012] As a further technical solution of the present invention, in the data fusion and state modeling module, the digital twin model takes the three-dimensional CAD model as the structural basis, and sets a parameter set that can update the status for each key component. The parameter set includes wear level, participating thickness, cumulative loading time and deviation history. The model update cycle is refreshed immediately after each measurement and identification task, and is used to determine whether to enter the maintenance window. The wear level includes 0 to 5 levels, the unit of participating thickness is mm, and the unit of cumulative loading time is h.

[0013] As a further technical solution of the present invention, the parts visual tracking module uses image texture hash coding to compare with the preset parts sequence dictionary, and verifies in real time during the reinstallation process whether the image ID of the current component matches its preset position. If misalignment, sequence jump, and direction flipping occur, the maintenance task is controlled to be paused and a prompt box of "abnormal assembly sequence" is popped up.

[0014] The technical effects of the virtual maintenance data processing system for optimizing maintenance scenarios of the present invention are as follows: The present invention effectively improves the intelligence level of assembly precision control, structural defect detection, maintenance data modeling and operation task management through the deep integration of multi-source data collection, intelligent identification, three-dimensional alignment, error judgment and guidance incentives throughout the whole process of gearbox disassembly and reinstallation. In the process of assembly guidance, the image recognition and point cloud posture recognition algorithms are combined to dynamically obtain the spatial posture changes of components, and accurately generate hot zone prompts and correct them in real time by identifying translation errors, rotation errors and misfit states to ensure high-precision assembly. At the same time, the deep learning algorithm is used to achieve high-robustness recognition of tooth surface cracks, wear and pitting defects, enhance the structural health assessment capability of the maintenance process, and achieve quantitative feedback and strategy optimization of operation quality through scoring mechanism and behavioral data retrospective analysis, guide maintenance personnel to actively complete high-quality operations, effectively avoid human problems such as wrong installation and missing installation, and realize the transformation of gearbox maintenance from experience-driven to data-driven, which significantly improves maintenance efficiency, safety and intelligence level. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a marked schematic diagram of the optical contact probe of the present invention measuring the inner diameter of a cylinder; Figure 2 This is an explosion demonstration interface diagram of the gearbox disassembly and overhaul of the present invention; Figure 3 This is a diagram of the disassembly interface of the gear box base during the disassembly and overhaul of the gear box of the present invention; Figure 4 This is a diagram of the disassembly interface of the upper cover of the gear box during the disassembly and overhaul of the gear box of the present invention; Figure 5 The interface diagram of the gearbox disassembly and overhaul steps of the present invention is played continuously; Figure 6 It is the comparison interface between the measurement model of the present invention and the standard model; Figure 7 This is the gearbox component defect identification interface of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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.

[0017] Example 1

[0018] The invention proposes a virtual maintenance data processing system for optimizing maintenance scenarios, including an optical contact probe, a scanning measurement system, an image recognition and defect analysis module, an image-assisted measurement module, a part visual tracking module, an augmented reality assembly guidance module, a data fusion and state modeling module, a human-computer interaction and task incentive module. The optical contact probe performs contact measurement of nanometer-level dimensions and displacement data of key parts of the gear meshing surface, shaft shoulder, matching hole, and gasket in the gearbox (such as Figure 1 As shown in the figure, the scanning measurement system obtains the 3D point cloud data of the key area of ​​the gearbox, and the image recognition and defect analysis module collects the image data of the target area of ​​the gearbox. The recognition algorithm based on deep learning is used to detect the defects of tooth surface wear, cracks, pitting and spalling (such as Figure 7 The image-assisted measurement module performs geometric analysis through the calibration benchmark and structural feature points in the figure, and forms a multi-source comparison with the point cloud model and contact measurement results. The part visual tracking module records the image and unique identification of the parts during the gearbox disassembly process, manages the whole process tracking and assembly sequence through image registration, and performs missing parts detection and assembly verification in combination with the state model during the reassembly stage. The augmented reality assembly guidance module spatially aligns the standard 3D assembly model with the current on-site image, identifies the assembly deviation in real time and generates a posture error heat map, combines image recognition and point cloud data to jointly drive the assembly correction guidance, and performs multi-source information enhanced posture adjustment prompts. The data fusion and state modeling module fuses the optical probe measurement results, image recognition output and point cloud morphology data to construct a digital twin model of the gearbox assembly state, and generates personalized virtual maintenance suggestions through trend analysis, structural degradation judgment and residual life assessment. When the human-computer interaction and task incentive module identifies potential risk areas and key operation nodes, it actively generates measurement and assembly tasks, and guides the operator to complete the gearbox maintenance actions in combination with task scoring, behavior confidence assessment and achievement incentive strategies.

[0019] It should be noted that the image recognition and defect analysis module includes a recognition unit based on a deep convolutional neural network. The network structure of the recognition unit includes a feature extraction layer, a region proposal network and a classification regression branch. By introducing an attention channel, weighted learning is performed on the spatial area of ​​the gear surface image to extract crack edge and texture degradation features in the presence of local occlusion and oil interference.

[0020] First, the collected gearbox images are processed with grayscale equalization, noise reduction and edge enhancement. The training data set is expanded by rotation, scaling and brightness perturbation to improve the generalization ability of the model. The oil-covered and occluded defect images are labeled with key points or bounding boxes for supervised training. The overall recognition model adopts the improved Faster R-CNN structure. ResNet-50 is used as the backbone to extract features in the feature extraction layer. The texture, edge and morphological features of the gearbox images are extracted from several set scales, and the detection of tiny cracks is supported. In the region proposal network (RPN), a sliding window is generated on the feature map to generate candidate regions, and a candidate frame is generated by combining an anchor frame with a preset size and aspect ratio. The candidate defect region is generated using foreground / background classification and regression calculation. In the classification and regression branches, each proposed region is cropped and sent to the RoI The Pooling layer unifies the scale, branch one performs defect classification (cracks, spalling, pitting), branch two regresses and predicts the precise position of the bounding box (for subsequent defect positioning and measurement), and introduces the SE (Squeeze-and-Excitation) channel attention mechanism in the feature extraction trunk. The responses of each channel are globally averaged and pooled, and the weight relationship between channels is learned. Important channels are weighted and enhanced to weaken the influence of non-critical textures. Spatial attention can also be introduced to guide the model to focus on shallow crack areas and fuzzy areas. The detection defect targets extracted and output from abnormal areas include defect classification labels, defect bounding boxes, defect area confidence (Softmax output), crack length, area and other structured data.

[0021] It should be noted that the image-assisted measurement module includes a two-dimensional image correction unit and a size conversion module, and is equipped with an image and point cloud error comparison control logic. When the difference between the image measurement result and the point cloud measurement result is greater than 0.5mm, the image and point cloud error comparison control logic performs automatic image recalibration and size mapping coefficient adjustment operations until the measurement error is lower than the set tolerance range of 0.5mm. The scanning measurement system is a structured light scanner with a point pitch resolution of 0.05mm and a scanning range of 300mm. 300mm. During the acquisition process, point cloud data of five viewing angles are obtained through several angles. The point cloud data are used to generate a three-dimensional model of the gearbox surface after uniform posture alignment.

[0022] The gearbox of the charging pump is a key equipment in a nuclear power plant. The gearbox contains multiple gear sets, bearings, shoulders, gaskets and other high-precision parts. Especially during the disassembly and reinstallation process, it is necessary to precisely measure the dimensions such as the matching aperture, axial clearance, and gasket thickness. Image measurement is interfered by factors such as structural occlusion and non-orthogonal viewing angles. Error control is particularly critical. Setting the image and point cloud error comparison control logic can dynamically correct the image measurement to ensure the reliability of key part measurement. The position of the gearbox of the charging pump is fixed and the space is limited. It is difficult for maintenance personnel to adjust the lighting or posture. Although the point cloud data is highly accurate, it is easy to produce holes in grooves and edge dead corners. Image measurement provides the advantages of texture completion and contour tracking, but it must be calibrated after point cloud calibration to trust its measurement value. The structured light scanning result is used as a reliable benchmark to enable the image measurement process to have self-calibration capabilities. Failure of the transmission device of the gearbox of the charging pump will cause serious production safety accidents. Setting 0.5mm as the error comparison trigger threshold is a rigid requirement based on the nuclear equipment maintenance technical standards, not an arbitrary parameter.

[0023] In the disassembly and maintenance scenario of the charging pump gearbox, due to the complex equipment structure, compact internal space, limited operating space, and the presence of a large number of closely fitting bearing holes, meshing surfaces, and gasket parts, it is difficult for the traditional single measurement method to take into account both measurement efficiency and accuracy requirements. In the present invention, the image-assisted measurement module realizes rapid dimension extraction through a two-dimensional image correction unit and a dimension conversion module, and sets the image and point cloud error comparison control logic. With the help of the high-resolution point cloud data obtained by the structured light scanning system as the error reference, when the difference between the image measurement result and the point cloud result exceeds 0.5mm, the image recalibration and mapping coefficient adjustment operation are automatically triggered until the error converges to the set tolerance range within the nuclear-grade maintenance requirements. This mechanism is particularly suitable for key scenarios such as nuclear power charging pump gearboxes that require high-precision measurement and reinstallation verification in limited spaces. It significantly improves the availability of image measurement and data fusion consistency, and meets the high reliability requirements of nuclear equipment for dimensional control.

[0024] It should be noted that the augmented reality assembly guidance module includes an image feature point matching submodule and a point cloud registration submodule. The image feature point matching submodule and the point cloud registration submodule are connected to a heat map generation unit. The outputs of the image feature point matching submodule and the point cloud registration submodule are combined to calculate the three-dimensional rigid body transformation matrix of the assembly relative to the standard assembly model through least squares registration. When the rotation amount of the three-dimensional rigid body transformation matrix is ​​greater than 3° and the translation amount is greater than 2 mm, the heat map generation unit is triggered to superimpose a red warning block in the AR display.

[0025] It should be specifically explained that, in the augmented reality assembly guidance module, the process of detecting the position and posture changes of the assembly parts in the three-dimensional space when the operator performs the reassembly operation includes: Step 1: Image acquisition and initial feature extraction: Collect the image of the reassembled component, apply the SHIF algorithm to extract the feature points of the reassembled component, and call the standard assembly model to establish the target area to be registered; Step 2: Feature matching registration based on center point correction: perform registration correction on the center point calculation of the bolt on the gear in the detection target. If the overall component registration deviation exceeds the preset threshold within the set range, but the target center point alignment error is less than the set threshold, it is judged as surface misleading registration, and the transformation matrix is ​​recalculated using the center point as the reference; Step 3: Edge contour detection and straight line fitting: extract the edge contour of the target component, use the Hough straight line fitting method to determine its boundary direction, and obtain the assembly rotation angle error by comparing the angle between the boundary direction and the standard direction; Step 4: Point cloud-assisted spatial posture detection: Based on the point cloud data obtained by the scanning measurement system, it automatically switches to 3D point cloud posture analysis after image registration, maps the area of ​​interest to the point cloud space, intercepts the point cloud of the target area, performs RANSAC plane fitting on the target assembly surface, analyzes the spatial distance, parallelism, and angle between the top and bottom surfaces, and determines whether there is any misalignment or incomplete insertion; Step 5, Error analysis and type judgment: Comprehensive image and point cloud information, classify and judge assembly errors, calculate translation error through center point offset or point cloud center point difference, calculate rotation error through boundary line direction or fitted plane normal vector, calculate vertical error through point cloud surface distance to judge whether it is pressed tightly, and identify incomplete insertion through edge contour difference map area greater than the set threshold or incomplete fitting area; Step 6: Abnormal filtering and self-learning: For areas with abnormal center point distribution and abnormal contour disturbance, perform an evaluation of the assembly error recognition optimization mechanism based on dynamic confidence weighting, and adaptively optimize the recognition algorithm.

[0026] For example, during a regular maintenance at a nuclear power plant, the active gear set in the gear box of the charging pump needs to be disassembled and cleaned, and restored to its standard assembly state during the reinstallation process. The comparison diagram of the measured value of the active gear assembly and the standard assembly 3D model is shown in the figure below. Figure 6 As shown in the figure, the active gear assembly is precisely matched with the spindle shoulder and the inner ring of the bearing. When reinstalling, the rotation direction error must not exceed ±3° and the translation error must not exceed ±2mm, otherwise it will cause poor meshing or early wear. Maintenance personnel use head-mounted AR glasses for assembly guidance, and the system starts the augmented reality assembly guidance module to detect and correct assembly posture errors.

[0027] The operation flow and detection process are as follows: Step 1: Image acquisition and initial feature extraction (1) The operator aligns the assembly position of the driving gear, and the system collects real-time images; (2) Use SIFT algorithm to extract key feature points such as gear profile and bolt holes; (3) Load the standard assembly model at the same time and extract the reference feature map of the corresponding parts in the model.

[0028] Step 2: Feature matching based on center point correction (1) The initial registration showed an overall deviation of 4.3 mm; (2) Further identify the alignment between the center point of the bolt hole on the driving gear and the model; (3) If the center point error is less than 1 mm, the system determines it as “surface misleading registration” and automatically recalculates the rigid body transformation matrix based on the center point.

[0029] Step 3: Edge contour and angle extraction (1) Extract the gear boundary direction through Canny edge detection + Hough line fitting; (2) Determine that the angle between this direction and the standard direction is 4.7°; (3) Determine whether the assembly rotation error exceeds the limit.

[0030] Step 4: Point cloud-assisted spatial posture detection (1) Synchronously call the structured light scanner to obtain point cloud data of the target area; (2) Map the gear area in the image to the point cloud space and perform RANSAC fitting; (3) It was identified that there was a height difference of 1.3 mm between the assembly surface and the base surface of the shell, and the angle deviation was about 1.8°, which was judged to be a loose assembly.

[0031] Step 5: Comprehensive error analysis (1) Comprehensively judge whether the rotation error exceeds the limit (>3°), the compaction is incomplete, and there are areas where the boundaries are not fitted; (2) The error type is given: "rotation direction deviation + assembly not tightened"; (3) Correction instructions are displayed: "Rotate clockwise about 3° and press down 1.5 mm."

[0032] Step 6: Abnormal area filtering and identification optimization (1) At the same time, the edge area is detected to have a reflection that causes the contour to jump, which reduces the recognition confidence; (2) Mark the area as a low-confidence image area and automatically reduce its recognition weight; (3) Call historical successful assembly data to fine-tune and optimize the current recognition weight parameters.

[0033] Final result: The operator makes adjustments according to the prompts on the heat map; the recognition posture error is lower than the set threshold (rotation angle 1.2°, translation error 0.9mm); the heat map disappears automatically, prompting "assembly qualified"; the current operation record is stored in the digital twin assembly log for subsequent tracing and learning training.

[0034] It should be noted that in the data fusion and state modeling module, the digital twin model is based on the three-dimensional CAD model as the structure, and a parameter set that can update the status is set for each key component. The parameter set includes wear level, participating thickness, cumulative loading time and deviation history. The model update cycle is refreshed immediately after each measurement and identification task, and is used to determine whether to enter the maintenance window. The wear level includes 0 to 5 levels, the unit of participating thickness is mm, and the unit of cumulative loading time is h.

[0035] There is a lack of component status evolution records before and after the disassembly and assembly of the charging pump gearbox. It relies on manual judgment and is highly subjective. It is impossible to quantify the impact of wear, assembly deviation or operating time on reliability. Maintenance decisions are mostly based on experience and lack scientific parameter support. It is difficult to achieve trend prediction and timely maintenance strategies. Image recognition, optical contact probe measurement, point cloud registration and other data are scattered, lacking structured fusion, and lacking a complete description of the life cycle of the same part. By taking the 3D CAD model as the structural basis, a digital twin structure with key components as nodes is constructed, and each node is configured with state parameters including wear level (0-5 levels), participating thickness (in millimeters), cumulative loading time (in hours), historical deviation trajectory, etc. The parameter set can be automatically refreshed after each image recognition, probe measurement or point cloud registration is completed. This mechanism significantly improves the dynamics and interpretability of assembly status modeling, enabling maintenance personnel to grasp the current status of the equipment in real time and make trend predictions, assisting in determining whether to enter the maintenance window, thereby supporting the formulation of state-based intelligent maintenance strategies and realizing the technological transition from passive maintenance to active predictive operation and maintenance.

[0036] It should be noted that the part visual tracking module uses image texture hash coding to compare with the preset part sequence dictionary, and verifies in real time during the reinstallation process whether the image ID of the current component matches its preset position. If misalignment, sequence jump, or direction flip occurs, the maintenance task is controlled to be paused and a prompt box of "abnormal assembly sequence" is popped up.

[0037] like Figures 2 to 5 As shown in the figure, the gearbox shows the step-by-step instructions for lifting the gearbox during the gearbox disassembly and overhaul process. Figures 2 to 5 There is a disassembly and maintenance step-by-step guide bar on the left side of the interface. Figure 2 A gearbox disintegration explosion demonstration was shown in Figure 3The gearbox base for gearbox hoisting is described with part numbers and precautions. The part number of the gearbox base is 56, and it contains 4 oil supply pipelines to avoid blockage. The data intelligence is displayed on the right. Figure 4 The gearbox cover is shown as part number 52, which contains the oil supply pipeline. Before reinstallation, check for foreign matter to avoid oil pipe blockage and bearing burnout. Figure 5 The disassembly and maintenance steps and animation demonstration of the gearbox are shown. The disassembly process of the gearbox of the charging pump requires that multiple gaskets, gears, bearings, etc. be numbered and saved in a fixed order. When reinstalling, it is easy for the parts to be in the wrong order or installed in the wrong direction, especially in symmetrical structures such as spiral direction and bevel angle. The manual paper record or alignment of drawings is labor-intensive. Once an error occurs, it is difficult to locate which link or component has a problem. Misinstalled bearings or gaskets can cause poor meshing and abnormal preload, affecting transmission efficiency or causing mechanical failure. The parts visual tracking module uses image texture hash coding technology to establish a unique visual identification for each disassembled component, and compares and verifies the real-time image with the preset part sequence dictionary during the reinstallation process; when it is identified that the current component position does not match its expected number, or the assembly direction is reversed or the sequence is wrong, the current maintenance task process is immediately interrupted, and a prompt box of "abnormal assembly sequence" pops up in the interactive interface. This mechanism effectively avoids low-level errors such as component misalignment and reverse assembly caused by operational negligence, greatly improving the accuracy, safety and traceability of the precision structure assembly process. It is particularly suitable for assembly scenarios of multi-level, highly sequence-dependent nuclear power key components such as charging pump gearboxes.

[0038] Example 2

[0039] The difference between Example 2 of the present invention and Example 1 is that this example introduces an augmented reality assembly guidance module in a virtual maintenance data processing system for optimizing maintenance scenarios.

[0040] It should be noted that in the augmented reality assembly guidance module, the assembly error recognition optimization mechanism based on dynamic confidence weighting generates a confidence score for each assembly error recognition result. The confidence score is obtained by weighted summation of the number of matching points, boundary integrity, point cloud surface closure rate, center point offset and image clarity. If the confidence score is lower than the set threshold, the control pauses the task push and prompts the operator to re-collect the image. The assembly error recognition optimization mechanism based on dynamic confidence weighting includes a regression learning model based on operation behavior correction feedback, which records and annotates the error correction between the operator's actual correction behavior and the recognition guidance result during the assembly process, and dynamically adjusts the parameter weights of image and point cloud recognition through a lightweight regression learning model to perform self-learning, self-optimization and confidence adaptive prompts for the assembly error recognition algorithm.

[0041] It should be specifically noted that the human-computer interaction and task incentive module includes a behavior data backtracking and strategy optimization mechanism, which is used to record the operator's execution path, assembly adjustment behavior, error level change and confidence fluctuation process in each task, generate behavior logs based on time series and performance labels, and adaptively optimize and adjust the recommended path, recognition parameter configuration and incentive level weight of subsequent tasks. The human-computer interaction and task incentive module uses the task scoring formula to calculate the task score. The task scoring formula is:

[0042] Where: Scoring tasks is used to cumulatively update the operator's incentive level based on this score, and to control their task unlocking permissions and challenge prompt push. Scoring task benchmarks, which are pre-set by administrators and engineers for each type of task. is the measurement error rate, which is equal to the mean absolute error rate of the optical touch probe, ,in is the number of measurement points involved in the task, The index of the measurement point for this task, , The tasks are The measured and expected values ​​of the measurement points, To identify the confidence coefficient, the Softmax layer classification probability output from the deep learning network of the image recognition module, ,in is the normalized value of the matching points and the total points, To identify the mean of the pair residuals, is the normalized value of the boundary closure rate, is the normalized value of the center point offset ratio, is the normalized value for measuring image clarity, , , , , is the weight factor, , automatically lock the task submission button and pop up a prompt box "Please collect again".

[0043] For example, during the reinstallation of the charging pump gearbox, a maintenance task involves the reinstallation of the active gasket assembly. The gasket fits tightly with the shaft shoulder, and the thickness change has a significant impact on the meshing clearance. Therefore, the gasket assembly must not be eccentric, uncompacted, or reversed. At the same time, the operator is required to accurately identify the assembly posture error and correct it in time. Maintenance personnel use head-mounted AR devices to assist in assembly, and the system simultaneously enables the "augmented reality assembly guidance module" and the "human-computer interaction and task incentive module." The specific implementation process is detailed as follows: Step 1: The system identifies the error and calculates the confidence score The gasket outline is identified by matching image feature points; Point cloud registration obtains the spatial position of the assembly surface; The system analysis results are as follows: The matching point ratio (FMP) is 0.91; The residual mean error (RMSE) is 0.08; The boundary closure rate (B_cl) is 0.85; The center offset (ΔC) is 0.07; The image clarity index (I) is 0.78; The weight coefficient is preset to ~ =[0.25,0.25,0.2,0.15,0.15] Calculate the confidence score: =0.25×0.91+0.25×(1-0.08)+0.2×0.85+0.15×(1-0.07)+0.15×0.78≈0.73; If the system-set threshold (0.7) is not reached, the recognition result is judged to be of insufficient confidence, the task submission is locked, and a prompt "Please re-capture the image" pops up.

[0044] Step 2: The operator fine-tunes the image and recognizes it again The operator adjusts the assembly posture according to the red hot zone map prompts; After the second image acquisition, the confidence score increased to 0.85; The identification assembly translation error is 1.5 mm and the rotation error is 2.2°, which passes the identification threshold and the task continues.

[0045] Step 3: Behavior backtracking and recognition weight learning The system records: Initial identification failed; The operator's actual fine-tuning direction is consistent with the hot zone guidance direction; After adjustment, the confidence of assembly recognition is improved.

[0046] Mark this behavior as "Guidance is effective + Operation correction is consistent": The regression learning module updates the weight of the "boundary closure rate B_cl" in the current task scenario from 0.2 to 0.25; Next time we encounter a similar task, we will pay more attention to contour completeness and implement dynamic learning and adaptive optimization.

[0047] Step 4: Task scoring and incentive feedback mechanism This task involves the number of measurement points =5, the difference between the actual measurement and the expected value of the optical contact probe is shown in Table 1: Table 1 Differences between actual measurement and expected values ​​of optical contact probe

[0048] Calculate the measurement error rate, which is the mean absolute error rate of the optical touch probe, , task benchmark score , final task score =100 (1-0.014) 0.85≈83.8. The system updates the operator’s points based on the score and pushes the achievement feedback “Rating of this task: excellent, high operating precision, good assembly quality”.

[0049] Through the dynamic confidence weighted recognition mechanism and behavior feedback self-learning model in the augmented reality assembly guidance module, the system can evaluate the credibility of the recognition results in real time during the assembly error recognition process to avoid misjudgment caused by poor image quality or recognition deviation; when the confidence is insufficient, the task submission is terminated in time to ensure the accuracy of the assembly data. At the same time, regression learning is carried out in combination with the operator's actual correction behavior, and the recognition parameter weights are continuously optimized to achieve adaptive updates of the assembly recognition algorithm. In conjunction with the task scoring and behavior backtracking mechanism in human-computer interaction, it not only improves the recognition accuracy and assembly success rate, but also enhances the operator's initiative and the system's intelligence, significantly improving the assembly reliability and maintenance efficiency of high-precision components such as gearboxes.

[0050] Combining Example 1 and Example 2, the present invention effectively improves the intelligence level of assembly precision control, structural defect detection, maintenance data modeling and operation task management through the deep integration of multi-source data collection, intelligent identification, three-dimensional alignment, error judgment and guidance incentives during the whole process of gearbox disassembly and reassembly. In the process of assembly guidance, image recognition and point cloud posture recognition algorithms are combined to dynamically obtain the spatial posture changes of components. By identifying translation errors, rotation errors and non-fitting states, hot zone prompts are accurately generated and corrected in real time to ensure high-precision assembly. At the same time, deep learning algorithms are used to achieve high-robustness recognition of tooth surface cracks, wear, and pitting defects, enhance the structural health assessment capability of the maintenance process, and achieve quantitative feedback and strategy optimization of operation quality through scoring mechanisms and behavioral data retrospective analysis, guide maintenance personnel to actively complete high-quality operations, effectively avoid human problems such as wrong installation and missing installation, and realize the transformation of gearbox maintenance from experience-driven to data-driven, significantly improving maintenance efficiency, safety and intelligence.

[0051] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0052] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A virtual maintenance data processing system for optimizing maintenance scenarios, comprising an optical contact probe and a scanning measurement system, characterized in that: It also includes image recognition and defect analysis module, image-assisted measurement module, part visual tracking module, augmented reality assembly guidance module, data fusion and state modeling module, human-computer interaction and task incentive module. The optical contact probe measures the data of key parts of the gearbox, and the scanning measurement system obtains the three-dimensional point cloud data of the key areas of the gearbox. The image recognition and defect analysis module collects images of the target area of ​​the gearbox. The recognition algorithm based on deep learning detects defects such as tooth surface wear, cracks, pitting and spalling, and jointly judges assembly posture deviation and poor contact. The image-assisted measurement module performs geometric analysis of the calibration benchmark and structural feature points in the image, and compares them with the point cloud model and measurement results. The part visual tracking module records the images and identifications of the disassembled parts of the gearbox, and manages the whole process tracking and assembly sequence. The augmented reality assembly guidance module aligns the standard 3D assembly model with the current on-site image space, identifies assembly deviations and posture error heat maps in real time, and combines image recognition and point cloud data to jointly drive assembly correction guidance, prompting multi-source information enhanced posture adjustment. The data fusion and state modeling module fuses measurement results, image recognition output and point cloud morphology to build a digital twin model of the gearbox assembly status. Through trend analysis, structural degradation judgment and residual life assessment, it generates virtual maintenance suggestions. When the human-computer interaction and task incentive module identifies potential risk areas and key operation nodes, it actively generates measurement and assembly tasks, and combines task scores, behavior confidence and incentive achievements to help guide the maintenance actions of the gearbox.

2. A virtual maintenance data processing system for optimizing maintenance scenarios according to claim 1, characterized in that: The image recognition and defect analysis module includes a recognition unit based on a deep convolutional neural network. The network structure of the recognition unit includes a feature extraction layer, a region proposal network and a classification regression branch. By introducing an attention channel, weighted learning is performed on the spatial area of ​​the gear surface image to extract crack edge and texture degradation features in the presence of local occlusion and oil interference.

3. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 1 is characterized in that: The image-assisted measurement module includes a two-dimensional image correction unit and a size conversion module, and is equipped with an image and point cloud error comparison control logic. When the difference between the image measurement result and the point cloud measurement result is greater than 0.5mm, the image recalibration and size mapping coefficient adjustment operation are performed automatically until the measurement error is lower than the set tolerance range of 0.5mm. The scanning measurement system is a structured light scanner with a point pitch resolution of 0.05mm and a scanning range of 300mm. 300mm. During the acquisition process, point cloud data of five viewing angles are obtained through several angles. The point cloud data are used to generate a three-dimensional model of the gearbox surface after uniform posture alignment.

4. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 1 is characterized in that: The augmented reality assembly guidance module includes an image feature point matching submodule and a point cloud registration submodule. The image feature point matching submodule and the point cloud registration submodule are connected to a heat map generation unit. The outputs of the image feature point matching submodule and the point cloud registration submodule are combined to calculate the three-dimensional rigid body transformation matrix of the assembly relative to the standard assembly model through least squares registration. When the rotation amount of the three-dimensional rigid body transformation matrix is ​​greater than 3° and the translation amount is greater than 2 mm, the heat map generation unit is triggered to superimpose a red warning block in the AR display.

5. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 4 is characterized in that: In the augmented reality assembly guidance module, the process of detecting the position and posture changes of the assembly parts in the three-dimensional space when the operator performs the reassembly operation includes: Step 1: Image acquisition and initial feature extraction: Collect the image of the reassembled component, apply the SHIF algorithm to extract the feature points of the reassembled component, and call the standard assembly model to establish the target area to be registered; Step 2: Feature matching registration based on center point correction: perform registration correction on the center point calculation of the bolt on the gear in the detection target. If the overall component registration deviation exceeds the preset threshold within the set range, but the target center point alignment error is less than the set threshold, it is judged as surface misleading registration, and the transformation matrix is ​​recalculated using the center point as the reference; Step 3: Edge contour detection and straight line fitting: extract the edge contour of the target component, use the Hough straight line fitting method to determine its boundary direction, and obtain the assembly rotation angle error by comparing the angle between the boundary direction and the standard direction; Step 4: Point cloud-assisted spatial posture detection: Based on the point cloud data obtained by the scanning measurement system, it automatically switches to 3D point cloud posture analysis after image registration, maps the area of ​​interest to the point cloud space, intercepts the point cloud of the target area, performs RANSAC plane fitting on the target assembly surface, analyzes the spatial distance, parallelism, and angle between the top and bottom surfaces, and determines whether there is any misalignment or incomplete insertion; Step 5, Error analysis and type judgment: Comprehensive image and point cloud information, classify and judge assembly errors, calculate translation error through center point offset or point cloud center point difference, calculate rotation error through boundary line direction or fitted plane normal vector, calculate vertical error through point cloud surface distance to judge whether it is pressed tightly, and identify incomplete insertion through edge contour difference map area greater than the set threshold or incomplete fitting area; Step 6: Abnormal filtering and self-learning: For areas with abnormal center point distribution and abnormal contour disturbance, perform an evaluation of the assembly error recognition optimization mechanism based on dynamic confidence weighting, and adaptively optimize the recognition algorithm.

6. A virtual maintenance data processing system for optimizing maintenance scenarios according to claim 5, characterized in that: In the augmented reality assembly guidance module, the assembly error recognition optimization mechanism based on dynamic confidence weighting generates a confidence score for each assembly error recognition result. The confidence score is obtained by weighted summation of the number of matching points, boundary integrity, point cloud surface closure rate, center point offset and image clarity. If the confidence score is lower than the set threshold, the control pauses the task push and prompts the operator to re-collect the image. The assembly error recognition optimization mechanism based on dynamic confidence weighting includes a regression learning model based on operation behavior correction feedback, which records and annotates the error correction between the operator's actual correction behavior and the recognition guidance result during the assembly process, and dynamically adjusts the parameter weights of image and point cloud recognition through a lightweight regression learning model to perform self-learning, self-optimization and confidence adaptive prompts for the assembly error recognition algorithm.

7. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 6 is characterized in that: The human-computer interaction and task incentive module includes a behavior data backtracking and strategy optimization mechanism, which is used to record the operator's execution path, assembly adjustment behavior, error level change and confidence fluctuation process in each task, generate behavior logs based on time series and performance labels, and adaptively optimize and adjust the recommended path, recognition parameter configuration and incentive level weight of subsequent tasks. The human-computer interaction and task incentive module uses the task scoring formula to calculate the task score. The task scoring formula is: ; Where: Scoring tasks is used to cumulatively update the operator's incentive level based on this score, and to control their task unlocking permissions and challenge prompt push. Scoring task benchmarks, which are pre-set by administrators and engineers for each type of task. is the measurement error rate, which is equal to the mean absolute error rate of the optical touch probe, ,in is the number of measurement points involved in the task, The index of the measurement point for this task, , The tasks are The measured and expected values ​​of the measurement points, To identify the confidence coefficient, the Softmax layer classification probability output from the deep learning network of the image recognition module, ,in is the normalized value of the matching points and the total points, To identify the mean of the pairwise residuals, is the normalized value of the boundary closure rate, is the normalized value of the center point offset ratio, is the normalized value for measuring image clarity, , , , , is the weight factor, , automatically lock the task submission button and pop up a prompt box "Please collect again".

8. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 1 is characterized in that: In the data fusion and state modeling module, the digital twin model uses the three-dimensional CAD model as its structural basis, and sets a parameter set for each key component that can update the status. The parameter set includes wear level, participating thickness, cumulative loading time and deviation history. The model update cycle is refreshed immediately after each measurement and identification task, and is used to determine whether to enter the maintenance window. The wear level includes 0 to 5 levels, the unit of participating thickness is mm, and the unit of cumulative loading time is h.

9. The virtual maintenance data processing system for optimizing maintenance scenarios according to claim 1, characterized in that: The parts visual tracking module uses image texture hash coding to compare with the preset part sequence dictionary. During the reinstallation process, it verifies in real time whether the image ID of the current component matches its preset position. If there is misalignment, sequence jump, or direction flipping, the maintenance task is controlled to be paused and a prompt box of "abnormal assembly sequence" is popped up.

Images