Pinwheel tooth profile error detection system based on adaptive calibration of laser point cloud

Abstract

The present application belongs to the field of mechanical engineering, and particularly relates to a pin wheel tooth profile error detection system based on laser point cloud adaptive calibration, aiming to solve the problem of inaccurate tooth profile error detection caused by mechanical deformation, point cloud distortion, coordinate registration misalignment and environmental disturbance in the traditional measurement method. The system integrates laser scanning array, six-degree-of-freedom fine adjustment platform, environmental sensing unit and point cloud processing engine, realizes real-time updating of calibration parameters and coordinate system offset correction through static internal parameter self-calibration and dynamic external parameter compensation double mechanism; combined with multi-view point cloud adaptive weighted fusion and improved Hausdorff distance evaluation algorithm, a closed-loop detection process is constructed. It can effectively improve the tooth surface point cloud integrity and measurement stability, support rapid model change of multiple models, realize full-automatic and high-repetitive online detection in non-constant temperature workshop environment, and significantly improve the quality control efficiency and process consistency of pin wheel transmission parts.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, specifically relating to a needle wheel tooth profile error detection system based on laser point cloud adaptive calibration. Background Technology

[0002] In the field of precision machinery manufacturing and testing, gears, as core transmission components, directly determine the operational stability, noise level, and service life of mechanical equipment through their machining accuracy. With the development of high-end equipment manufacturing, the requirements for gear geometric accuracy are becoming increasingly stringent, especially for pinwheels with complex tooth profiles. The measurement accuracy of their tooth profile error has become a key factor restricting product performance improvement. Traditional contact measurement methods rely on coordinate measuring machines or gear measuring centers, acquiring tooth surface data point-by-point through probe scanning. While possessing a certain level of accuracy, these methods suffer from inherent drawbacks such as low measurement efficiency, easy damage to the tooth surface, and difficulty in covering the entire tooth profile.

[0003] Among these, non-contact optical measurement techniques have gradually become a research hotspot, particularly laser triangulation and structured light scanning. These methods, due to their high sampling rate, non-destructive nature, and ability to acquire full-field data, are widely used in the 3D reconstruction of complex curved surfaces. These methods emit laser beams or grating patterns onto the surface being measured, using a camera to capture the deformation information of the reflected light, and then calculate the 3D point cloud data of the object's surface. Theoretically, this technology can achieve rapid digitization of the tooth surface of a pinwheel, providing a data foundation for tooth profile error analysis.

[0004] While existing technologies have achieved a leap from contact to non-contact methods, they still face multiple technical obstacles in practical applications for pinwheel tooth profile inspection: the matching relationship between the laser incident angle and the tooth surface normal is significantly affected by the pinwheel's geometric characteristics, resulting in widespread occlusion and missing data in the tooth root and tip regions of the point cloud data; the lack of a high-precision initial registration benchmark during multi-view point cloud stitching makes it difficult to control cumulative errors; more importantly, pinwheels inevitably experience installation eccentricity and tilting during clamping, which, without real-time attitude compensation, directly introduces systematic measurement deviations; furthermore, the surface reflectivity of different batches of pinwheel materials varies considerably, making it difficult for fixed-parameter laser acquisition systems to adapt to the signal-to-noise ratio requirements under various working conditions, leading to unstable point cloud quality. These problems collectively result in current non-contact inspection systems failing to meet the demands of batch, high-reliability inspection in terms of repeatability, global registration accuracy, and environmental adaptability, especially when assessing micron-level tooth profile errors, where the reliability of measurement results significantly decreases. Therefore, there is an urgent need to build a needle wheel tooth profile error detection system that can achieve adaptive calibration of laser point clouds, so as to overcome the combined errors caused by installation deviation, viewing angle obstruction and environmental disturbance, and improve detection accuracy and engineering applicability. Summary of the Invention

[0005] The purpose of this invention is to provide a needle wheel tooth profile error detection system based on laser point cloud adaptive calibration. This system addresses the technical contradictions in the high-precision manufacturing and assembly of needle wheel transmission components, stemming from problems such as mechanical deformation introduced by traditional contact measurement methods, limited measurement paths, and point cloud data distortion, coordinate registration misalignment, and sensitivity to environmental disturbances commonly found in non-contact optical inspection. These problems lead to inaccurate tooth profile error quantification, poor repeatability, and low automation. As high-end equipment develops towards precision, lightweight, and high dynamic performance, the needle wheel, as a key component of core transmission parts such as harmonic reducers, directly determines transmission accuracy, backlash characteristics, and fatigue life through its tooth profile geometric consistency. Existing technologies rely on fixed calibration processes under manual intervention, which struggle to handle systematic drift caused by multiple batches of workpieces with varying specifications. Furthermore, under complex lighting, slight vibrations, or temperature variations, the three-dimensional point cloud acquired by laser scanning is prone to localized defects, noise accumulation, and spatial misalignment. This results in extraction deviations of the tooth tip circle, pitch curve, and tooth root transition surface exceeding process tolerances, failing to meet the requirements for online full inspection.

[0006] The technical solution of this invention is to construct a closed-loop detection system integrating dynamic sensing, real-time compensation, and intelligent reconstruction capabilities. This system consists of a laser scanning array module, a six-degree-of-freedom fine-tuning platform, an environmental parameter sensing unit, a point cloud processing engine, and an error assessment core. Each module achieves millisecond-level data interaction via a high-speed synchronous bus. The laser scanning array module employs a multi-view, cross-arranged line structured light sensor, with four groups evenly distributed around the circumference of the needle wheel under test. Each group includes a laser projector with an adjustable emission angle and a high-frame-rate industrial camera, used to acquire raw point cloud clusters covering the entire tooth surface within the rotation sampling period. The six-degree-of-freedom fine-tuning platform carries a standard reference target sphere group, which consists of three 10mm diameter ceramic spheres rigidly connected in a non-collinear manner. The platform performs sub-micron-level translation and rotation according to a preset excitation sequence to generate a spatial motion trajectory for iterative optimization of calibration parameters. The environmental parameter sensing unit integrates a temperature gradient sensor, a triaxial accelerometer, and an illuminance detector, distributed at key locations on the inner wall of the scanning cavity, to monitor in real-time changes in the physical field affecting optical propagation and structural stability. The point cloud processing engine, acting as the system's central hub, receives raw point cloud streams from the laser scanning array module and auxiliary data from the environmental parameter sensing unit. It first performs point cloud denoising and completion operations based on spatiotemporal correlation: outliers are identified using the initial ICP registration results between adjacent frames, and the missing regions are geometrically filled with workpiece pose predicted by Kalman filtering. Furthermore, an adaptive weighting mechanism is introduced to normalize and weight-fuse the point cloud density from different viewpoints, forming a complete tooth surface mesh model in a unified coordinate system. This mesh model, smoothed by a subdivision surface algorithm, is input to the feature extraction submodule. The actual tooth tip center coordinates are determined by fitting a least-squares circle, and 256 equal-angle cross-sectional lines are generated radially. The distance sequence between each cross-sectional line and the nearest point on the theoretical tooth profile template is extracted as the original deviation dataset.

[0007] Furthermore, the system establishes a dual calibration mechanism to ensure the long-term stability of the measurement benchmark. The first layer is static intrinsic parameter self-calibration: before each detection task starts, the six-degree-of-freedom fine-tuning platform drives the standard reference target sphere group to enter the center of the laser scanning field of view, collecting its point cloud data in multiple postures; the point cloud processing engine runs a closed-loop solution algorithm to invert the internal parameters of each line structured light sensor, including focal length, principal point offset, and lens distortion coefficient, and updates them to the imaging model. The second layer is dynamic extrinsic parameter compensation: during the needle wheel rotation scanning process, when the environmental parameter sensing unit detects a temperature change rate exceeding 0.5℃ / min or a vibration acceleration peak greater than 0.2g, it triggers the extrinsic parameter re-estimation process; at this time, the system pauses the main measurement process, calls the cylindrical surface template in the built-in standard geometric primitive library, performs local fitting on the currently acquired part of the point cloud, calculates the rigid body transformation matrix of the current coordinate system relative to the initial calibration coordinate system, and injects this matrix into the subsequent point cloud registration process in real time, thereby realizing dynamic correction of systematic offsets caused by external disturbances. The error assessment core receives the calibrated tooth surface mesh model and compares it point-by-point with the theoretical CAD model stored in the database. It calculates the maximum deviation using an improved Hausdorff distance algorithm and generates an error report conforming to national standards, based on indicators such as total profile deviation Fa and tooth inclination deviation fHβ as defined in GB / T 10095.1-2008. The core also includes a built-in statistical process control module that performs trend analysis on the inspection data of 10 consecutive batches of similar pinwheels. When the range of a certain error term exceeds the control limit three times consecutively, a maintenance warning signal is automatically pushed to the equipment management system.

[0008] Preferably, the line structured light sensor in the laser scanning array module has a dynamic depth-of-field adjustment function, with an object distance adjustment range of 150mm to 250mm and an adjustment step of 0.1mm. The controller automatically matches the optimal working distance according to the pinwheel module size, ensuring that the laser stripe still has a width resolution of not less than 8 pixels at the bottom of the tooth groove. Preferably, the adaptive weight mechanism in the point cloud processing engine dynamically allocates fusion weights based on the number of effective points per unit area under each viewpoint and the local normal consistency index. The normal consistency index is obtained by calculating the direction variance of the first eigenvector obtained from the principal component analysis of the neighboring point cloud, and the weight value is positively correlated with the point cloud quality. Preferably, the motion controller of the six-degree-of-freedom fine-tuning platform adopts a PID feedforward composite control strategy, and its position feedback comes from a laser interferometer integrated into the platform base, with a resolution of 0.05μm, ensuring accurate reproduction of the calibration trajectory. Preferably, the sampling frequency of the illuminance detector in the environmental parameter sensing unit is set to 100Hz. When sudden light source interference is detected, causing the image signal-to-noise ratio to drop by more than 15dB, the system automatically activates the multi-frame averaging mode, continuously acquiring 5 frames of point cloud data and performing median filtering fusion to suppress instantaneous optical noise. Preferably, the feature extraction submodule adopts an adaptive sampling density strategy when generating equiangular cross-section lines. That is, in areas with large curvature at the tooth tip arc and tooth root transition, the sampling density is increased to 5 sampling points per degree, while in straight involute segments, it is sparsed to 1 sampling point per degree, thereby improving the overall processing efficiency while ensuring the accuracy of key areas. Preferably, the improved Hausdorff distance algorithm in the error evaluation core introduces a direction-sensitive factor to distinguish between overcut and undercut tooth profiles, and records the two types of deviations separately, providing a basis for directional correction for subsequent grinding compensation. Preferably, the system supports rapid changeover detection for multiple types of pinwheels. By pre-storing calibration parameter groups and theoretical model index tables of different specifications, operators only need to input the product code to complete all configuration switching, and the entire preparation time does not exceed 30 seconds.

[0009] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0010] This solution fundamentally solves the measurement drift problem caused by external environmental fluctuations in traditional optical inspection systems by constructing a laser point cloud adaptive calibration closed loop. The dynamic extrinsic parameter compensation mechanism can correct coordinate system offsets caused by thermal expansion and contraction or mechanical vibration in real time without stopping the system, controlling the long-term dimensional measurement stability within ±1.5μm, significantly better than the ±5μm level of uncompensated systems. The dual calibration system, combining static intrinsic parameter updates and dynamic extrinsic parameter adjustments, enables the inspection equipment to have continuous autonomous capabilities, significantly reducing the frequency of manual calibration from once a day to once a month, lowering maintenance costs and improving production line integration. The multi-view adaptive fusion strategy effectively overcomes the defects of single-view occlusion, such as in the grooves at the root of the pinwheel tooth. The system achieves over 98% point cloud coverage in complex structural areas, a 40 percentage point improvement compared to single-view scanning, ensuring full tooth surface integrity detection. Its closed-loop control capability in response to environmental disturbances allows for stable operation in ordinary workshop environments, eliminating the need for additional temperature- and humidity-controlled cleanrooms and saving on infrastructure investment. The entire detection process is automated, with a time of less than 90 seconds from material loading and positioning to report output, meeting the demands of modern intelligent manufacturing for high-frequency, high-volume online quality inspection. The directionally sensitive deviation data output by the error assessment module can be directly connected to the CNC grinding machine control system, forming an integrated "detection-feedback-correction" process chain. This drives the evolution of pinwheel manufacturing towards an adaptive machining mode, comprehensively improving the quality consistency and production efficiency of high-end transmission components. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall technical solution architecture of the needle wheel tooth profile error detection system based on laser point cloud adaptive calibration proposed in this invention;

[0012] Figure 2 This is a schematic diagram of the principle framework of the adaptive fusion algorithm in this invention. Detailed Implementation

[0013] Please refer to Figure 1 and Figure 2This invention provides a needle wheel tooth profile error detection system based on laser point cloud adaptive calibration, aiming to solve the technical bottlenecks in needle wheel tooth profile geometry measurement in high-precision manufacturing scenarios, such as poor measurement repeatability, low automation level, and weak process adaptability caused by mechanical contact deformation, optical path obstruction, environmental disturbance sensitivity, and coordinate reference instability. The system constructs a closed-loop detection architecture integrating dynamic sensing, real-time compensation, and intelligent reconstruction functions to achieve high-completeness and high-stability non-contact scanning of the three-dimensional morphology of the entire needle wheel tooth surface, maintaining sub-micron-level spatial positioning consistency under multi-physics interference conditions. The system consists of five main components: a laser scanning array module, a six-degree-of-freedom fine-tuning platform, an environmental parameter sensing unit, a point cloud processing engine, and an error assessment core. Each module relies on a high-speed synchronous bus to complete millisecond-level data interaction and control command coordination, forming a complete technical link from raw data acquisition to final error report output.

[0014] The overall system operation begins with the triggering of the detection task start signal. The system first executes a static intrinsic parameter self-calibration process, using a six-degree-of-freedom fine-tuning platform to drive the standard reference target sphere assembly into the center of the field of view, acquiring point cloud observation sequences under multiple attitudes, and inverting and updating the internal imaging parameters of each line structured light sensor. Subsequently, the main detection phase begins, where the test pinwheel is clamped on a rotary table and rotates continuously in preset angular steps. Simultaneously, the laser scanning array module acquires the original point cloud flow covering the entire circumferential tooth surface, while the environmental parameter sensing unit monitors in real time key physical parameters affecting the optical propagation path and mechanical structure stability, such as temperature gradient, vibration acceleration, and light intensity. The point cloud processing engine receives the aforementioned multi-source heterogeneous data streams and sequentially performs spatiotemporal denoising, adaptive fusion, dynamic extrinsic parameter compensation, and feature extraction operations to generate a complete tooth surface mesh model in a unified coordinate system. After smoothing, this model is input into the error assessment core and compared point-by-point with the theoretical CAD model to calculate tooth profile deviation indices that meet national standards. Combined with statistical process control logic, it generates quality trend warning information. The entire inspection cycle is completed automatically without manual intervention, with a single full-process time of less than 90 seconds, meeting the high-frequency online quality inspection requirements of intelligent manufacturing production lines.

[0015] The laser scanning array module, serving as the system's front-end sensing unit, undertakes the task of high-density sampling of the three-dimensional topography of the needle wheel surface. The module employs four sets of line structured light sensors evenly distributed along the circumference of the workpiece. Each sensor set consists of an adjustable-focus laser projector and a high-frame-rate industrial camera, with their optical axes arranged at a fixed angle to form a triangulation system. The laser projector emits a 650nm red line structured beam, projecting it onto the needle wheel surface to form reflective stripes. The industrial camera captures the stripe images at a sampling frequency of 2000 frames per second, calculating the depth information of the corresponding spatial points through pixel displacement. To adapt to needle wheel products with different module specifications, the laser projector features dynamic depth-of-field adjustment, with an object distance adjustment range set from 150mm to 250mm and an adjustment step accuracy of 0.1mm. The controller automatically matches the optimal working distance by calling a pre-stored optical configuration parameter table based on the current product code, ensuring a clear stripe response of at least 8 pixels wide even at the deepest point of the tooth groove, thus guaranteeing effective reconstruction of the bottom curved surface. The emission angle of each sensor group can be electrically adjusted within a range of ±15°, and the incident angle is dynamically optimized based on the pinwheel tooth height and tooth flank slope to avoid fringe blurring or breakage caused by grazing. All four sensor groups are connected to the point cloud processing engine via a gigabit Ethernet interface to achieve synchronous uploading of raw point cloud data. The timestamp alignment error is less than 0.5ms, ensuring the temporal consistency of data across different viewing angles.

[0016] The six-degree-of-freedom fine-tuning platform, as the core calibration actuator of the system, is used to support and precisely control the standard reference target ball assembly. The platform adopts a parallel Stewart structure, with six servo-driven linear actuators supporting the upper motion platform. It can independently apply sub-micron-level displacement and micro-radian-level rotation in the X, Y, and Z translational directions and rotational directions around the three axes. The standard reference target ball assembly consists of three 10mm diameter high-purity ceramic spheres rigidly connected by a low-expansion-coefficient alloy bracket. The center positions of the spheres are verified by precision grinding and coordinate measuring machines, ensuring they are non-collinearly distributed with a minimum spacing of 30mm, forming a geometric unit with a unique spatial configuration. Before each detection task is initiated, the platform executes a series of composite motions according to a preset excitation trajectory, including sinusoidal oscillation along the X-axis (amplitude ±20μm, frequency 1Hz), stepped rotation around the Z-axis (5° step, 72 steps), and Z-axis step lift (10μm step, 5 levels), generating a calibration point cloud sequence covering the entire attitude space. The trajectory design fully excites the motion modes of each degree of freedom, enhancing the observability of parameter identification and providing high-quality input data for subsequent closed-loop calculations. The platform's motion controller employs a PID feedforward composite control strategy, with the position feedback signal sourced from a dual-beam laser interferometer system integrated into the base. This system achieves a resolution of 0.05 μm and a closed-loop control bandwidth exceeding 200 Hz, ensuring that the deviation between the actual motion trajectory and the theoretical trajectory is less than 0.1 μm, meeting the stringent requirements of high-precision calibration for motion reproducibility.

[0017] The environmental parameter sensing unit is deployed in key thermodynamic and mechanically sensitive areas on the inner wall of the scanning cavity to capture external disturbances affecting the accuracy of optical measurements in real time. The unit includes three types of sensors: a temperature gradient sensor with six nodes located on the upper and lower covers, front and rear side walls, and the middle sections of the left and right columns, with a sampling frequency of 10Hz, a measurement range of 15℃ to 40℃, and a resolution of 0.01℃, used to monitor refractive index changes caused by temperature differences within the cavity and non-uniform thermal expansion of the metal structure; a triaxial accelerometer mounted at the bottom of the main optical support, with a range of ±2g, a sensitivity of 100mV / g, and a sampling frequency of 1kHz, used to detect minute vibrations transmitted from the ground or mechanical shocks generated by the equipment's own operation; and an illuminance detector using a broadband response photodiode, positioned near the viewing window, with a sampling frequency of 100Hz and a dynamic range covering 10lux to 10000lux, used to identify sudden light source interference such as personnel movement or lighting changes that may affect the camera's signal-to-noise ratio. All sensor data is aggregated to the point cloud processing engine via the CAN bus, and precisely aligned with the point cloud frames in the time domain to form an environmental context label bound to each frame of scan data, which serves as the triggering condition and correction basis for subsequent dynamic compensation algorithms.

[0018] The point cloud processing engine is the information hub of the system, responsible for integrating the raw point cloud stream from the laser scanning array module and the auxiliary data from the environmental parameter sensing unit, and performing a series of complex data processing operations. The engine runs on an industrial server equipped with a GPU accelerator card, using a real-time Linux distribution to ensure deterministic response for critical tasks. Its processing flow consists of multiple stages: the initial stage is spatiotemporal correlation denoising and completion. The engine first performs coarse ICP registration on two adjacent point cloud frames, judging whether there are significant attitude jumps based on the transformation matrix estimation results; if no anomalies are found, it further calculates the distance difference and normal angle between each point in the current frame and its nearest neighbor in the previous frame, marking points with distance differences greater than a set threshold (default is 5 times the local average point distance) or normal angles exceeding 30° as outliers for removal. For local point cloud missing regions caused by occlusion or insufficient reflectivity, the engine uses a Kalman filter to predict the current workpiece attitude and, combined with the geometric continuity model of the previous complete contour, performs spline interpolation filling along the tooth profile direction to restore the topological structure of the missing segments. The completed point cloud clusters then proceed to the next stage—adaptive weighted fusion.

[0019] An adaptive weighting mechanism addresses the discrepancies in density distribution and quality consistency among point clouds viewed from multiple perspectives. Due to variations in laser incident angle, surface reflection characteristics, and background noise levels at different angles, the number of effective points and local accuracy acquired from each perspective exhibit non-uniformity. The engine calculates the effective point density per unit area ρi and the local normal consistency index ηi for each subset of the point cloud from each perspective. ρi is obtained by performing a KD-tree search on points within a local neighborhood (radius r = 0.2 mm) and counting the points. ηi is obtained by performing Principal Component Analysis (PCA) on the same neighborhood point set, extracting the first principal component direction (i.e., the local normal), and then calculating the angular variance σθ of all neighboring points relative to this direction. ηi is defined as 1 / (1+σθ), where a higher consistency in the normal direction results in a higher ηi. The final fusion weight wi is determined by the product of the normalized ρi and ηi, i.e., wi = (ρi / Σρj) × (ηi / Σηj), ensuring that high-density regions with good geometric consistency dominate the fusion process. The fusion process employs Moving Least Squares (MLS) to resample and smooth the weighted point cloud under a unified target coordinate system, generating a preliminary tooth surface mesh model with uniform resolution.

[0020] The mesh model is then input into the feature extraction submodule to generate a geometric feature dataset that can be used for error assessment. The submodule first performs least-squares circle fitting on the mesh vertices, iteratively converging to find the optimal center coordinates and radius of the actual tooth tip circle, which serves as the reference origin for subsequent radial subdivision. Centered on this circle, 256 radially spaced cross-sections with equal angles are generated along the circumference, each spaced 1.40625° apart, covering the entire 360° circumference. On each cross-section, the system expands the search window radially outward, extracting the sequence of intersections with the reconstructed tooth surface mesh, and selecting the point closest to the theoretical tooth profile template as the deviation sampling point for that cross-section. To improve the evaluation accuracy of key areas, the system adopts an adaptive sampling density strategy: sampling is increased to 5 points per degree in the tooth tip arc segment (radius of curvature less than 0.5 mm) and the tooth root transition zone (curvature greater than 2000m to the power of -1), while sampling is sparsed to 1 point per degree in the intermediate involute segment. This ensures the ability to capture details in high-curvature areas while avoiding redundant computational burdens in regular areas. The spatial coordinates of all sampling points and their corresponding theoretical offsets constitute the original deviation dataset, which is used by the error assessment core.

[0021] The system establishes a dual calibration mechanism to ensure the long-term stability of the measurement benchmark throughout the entire detection cycle. The first calibration is a static intrinsic parameter self-calibration, which is forcibly executed before the start of each detection task. A six-degree-of-freedom fine-tuning platform drives the standard reference target sphere group into the common field of view center of four sets of line structured light sensors. The platform moves according to a preset excitation trajectory, and each set of sensors synchronously acquires point cloud data of the target sphere group in at least 12 different postures. The point cloud processing engine calls a closure solution algorithm, based on the known geometric relationship between the sphere center and the corresponding constraints between the observed point cloud, to inversely retrieve the internal parameters of each sensor, including focal lengths fx and fy, principal point coordinates cx and cy, and radial distortion coefficients k1 and k2 and tangential distortion coefficients p1 and p2. The solution process adopts a nonlinear least squares optimization method, and the objective function is the sum of squares of the reprojection errors between the observed point and the predicted point of the projection model:

[0022]

[0023] in, Let be the camera intrinsic parameter matrix to be solved, which includes focal length and distortion parameters; For the first The coordinates of each observation point in the image coordinate system; For projection functions, it means that given intrinsic parameters extrinsic rotation matrix Translation vector and the coordinates of the center of the sphere in space Under certain conditions, the process of mapping three-dimensional points to a two-dimensional image; This represents the total number of observation points. After optimization, the newly calculated intrinsic parameters are written into the imaging models of each sensor, replacing the old parameters and completing a full static calibration cycle. This process frees the system from dependence on factory calibration parameters and effectively eliminates systematic error sources such as lens loosening and focus drift caused by transportation, installation, or long-term use.

[0024] The second layer is dynamic extrinsic parameter compensation, used to cope with sudden environmental disturbances during the detection process. The point cloud processing engine continuously monitors the data stream of the environmental parameter sensing unit. When it detects that the temperature change rate of any temperature sensor exceeds 0.5℃ / min, or the peak acceleration of any channel of the triaxial accelerometer is greater than 0.2g, or the sudden change in light intensity recorded by the illuminance detector exceeds a preset threshold (corresponding to a decrease in image signal-to-noise ratio of more than 15dB), the extrinsic parameter re-estimation process is immediately triggered. At this time, the system pauses the main measurement process, retains the currently acquired part of the point cloud data, calls the cylindrical surface template in the built-in standard geometric primitive library, and performs local least squares fitting on the annular region located at the outer edge of the needle wheel body in this part of the point cloud to solve for the direction of the cylinder axis and the center position at the current moment. The result is compared with the standard axis stored during the initial calibration, and the rigid body transformation matrix Tcomp between the two is calculated. This matrix contains 3 translational components and 3 rotational variables. Subsequently, the system injects Tcomp into all subsequent point cloud registration steps. When aligning the new acquisition frame with the reference model, this compensation transformation is pre-stressed to offset systematic coordinate shifts caused by thermal expansion and contraction or mechanical vibration. This mechanism achieves millisecond-level response and correction to external disturbances, restoring the consistency of the measurement reference without requiring a system shutdown and restart.

[0025] The error assessment core receives the complete tooth surface mesh model after dynamic compensation and adaptive fusion, and performs the final geometric error quantification analysis. The core first globally registers the reconstructed model with the theoretical CAD model stored in the local database, and uses an improved Hausdorff distance algorithm to calculate the maximum spatial deviation between the two. Traditional Hausdorff distance only reflects the maximum deviation magnitude and cannot distinguish between overcut and undercut states, which is detrimental to subsequent machining compensation. Therefore, this system introduces a direction-sensitive factor and defines a bidirectional deviation metric:

[0026]

[0027] in, For the actual tooth surface model point set to be reconstructed, For the theoretical nominal model point set and These are the unit normal vectors at the corresponding points. This represents the vector inner product operation. The first term measures the degree to which the actual surface exceeds the theoretical surface (i.e., material overcut), and the second term measures the degree to which the actual surface is concave within the theoretical surface (i.e., material undercut). By recording the maximum values ​​of these two terms, the system can clearly identify the machining defect type of the tooth profile and provide directional correction instructions for the CNC grinding machine. Based on this, and adhering to the evaluation method defined in GB / T 10095.1-2008, the system extracts key indicators such as total tooth profile deviation Fa, tooth tilt deviation fHβ, and single tooth pitch deviation fpt from the original deviation dataset, generating a standardized error report. The report is packaged in XML format, containing numerical results, statistical charts, and a 3D deviation cloud map, and is pushed to the MES system via the OPC UA protocol.

[0028] The error assessment core also integrates a Statistical Process Control (SPC) module to monitor the quality data trends of continuous production batches. The module maintains a sliding window storing key control indicators such as Fa and fHβ for the most recent 10 batches of similar pinwheels. The moving range is calculated for each indicator. ,in For a certain key control indicator The values ​​for the next production batch; this indicator includes key control parameters such as Fa and fHβ. For the aforementioned same main control indicator, the first Record the batch measurement values ​​and update the upper limit of the control chart. Where D4 is the control chart coefficient corresponding to a sample size of n=10, with a value of 1.777. The mean range is used. When the MR value of a certain indicator exceeds the UCL three times consecutively, the judgment process shows abnormal fluctuations. The system automatically generates a maintenance warning signal, which includes the name of the abnormal indicator, the time of occurrence, the historical trend curve, and the suggested troubleshooting direction (such as "It is recommended to check the laser power stability" or "It is recommended to re-perform static calibration"). The warning is pushed to the equipment management system via RESTful API to prompt maintenance personnel to intervene.

[0029] The system supports rapid changeover testing for multiple pinwheel models. The database pre-stores product configuration packages for different combinations of module, number of teeth, and displacement coefficients. Each configuration package contains a set of exclusive parameters: theoretical CAD model file, calibration parameter set (including initial intrinsic and extrinsic parameters for each sensor), scan path planning (rotation step angle, dwell time), initial values ​​for adaptive fusion weights, error assessment template, and SPC control limits. Operators only need to input the product code on the human-machine interface, and the system automatically loads the corresponding configuration package and completes all parameter switching within 30 seconds. This design greatly improves the system's applicability in multi-variety, small-batch production modes, avoiding efficiency losses and human error risks associated with frequent manual debugging.

[0030] This embodiment constructs a closed-loop detection system with autonomous perception, real-time compensation, and intelligent decision-making capabilities through the coordinated operation of the aforementioned modules. Compared to traditional optical measurement schemes that rely on fixed calibration and manual intervention, this system achieves a technological leap from passive measurement to active control. The dynamic extrinsic parameter compensation mechanism enables the system to withstand the effects of temperature fluctuations and mechanical vibrations in ordinary workshop environments, improving the long-term dimensional measurement stability from ±5μm in the uncompensated state to within ±1.5μm, fully meeting the ±3μm assembly tolerance requirement of the harmonic reducer pinwheel. The dual calibration mechanism, combining static intrinsic parameter updates and dynamic extrinsic parameter adjustments, extends the equipment calibration cycle from once a day to once a month, significantly reducing maintenance costs. The multi-view adaptive fusion strategy overcomes the problem of single-view occlusion, achieving a point cloud coverage rate of over 98% in complex areas such as tooth root grooves, a 40 percentage point improvement compared to the single-view solution, ensuring the integrity detection of the entire tooth surface. The fully automated execution capability reduces the single-piece inspection time to within 90 seconds, supporting an inspection cycle of over 40 pieces per hour, meeting the needs of modern intelligent manufacturing for online full inspection. The direction-sensitive deviation output can be directly connected to the CNC grinding machine control system to form an integrated process closed loop of "detection-feedback-correction", which promotes the evolution of pinwheel manufacturing towards an adaptive machining mode and comprehensively improves the quality consistency and production efficiency of high-end transmission components.

[0031] Existing technologies generally employ non-contact measurement systems with fixed calibration parameters. These systems require a one-time calibration in a temperature-controlled cleanroom before deployment, after which the system's geometric relationships are assumed to remain constant. However, in real manufacturing environments, diurnal temperature variations can exceed 10°C, causing thermal deformation on the order of tens of micrometers in metal structures. Ground vibrations during equipment operation also transmit to the optical support, causing slight shifts in the relative positions of the camera and laser. The cumulative effect of these factors leads to significant reference drift in uncompensated measurement systems after several hours of continuous operation. This manifests as a trend of deviation in repeated measurements of the same batch of workpieces, and in severe cases, even misclassifying qualified products as defective. This solution addresses this industry pain point by introducing a dynamic external parameter compensation mechanism. It monitors environmental disturbances in real-time during the detection process and reconstructs the coordinate reference, fundamentally solving this problem. When the system detects excessive temperature change rate or vibration acceleration, it immediately pauses the main process, calls standard geometric primitives to locally fit the current point cloud, solves for the rigid body transformation matrix of the current coordinate system relative to the initial reference, and injects it into the subsequent registration process, achieving millisecond-level response and correction to external disturbances. This mechanism can restore the baseline without interrupting the detection task, ensuring the spatiotemporal consistency of the measurement results.

[0032] Furthermore, existing technologies for multi-view point cloud fusion typically employ equal-weight averaging or simple stitching strategies, failing to consider the spatial differences in point cloud quality across different viewpoints. For example, when a laser illuminates the tooth tip at a near-vertical angle, the reflected signal is strong, the point cloud is dense, and noise is low; however, when illuminating the tooth root transition region, the energy attenuation is severe due to the large incident angle, resulting in a sparse point cloud that is susceptible to stray light interference. Forcibly assigning equal weights will cause the information from high-quality areas to be dragged down by low-quality areas, leading to a decrease in overall reconstruction accuracy. The adaptive weighting mechanism proposed in this scheme dynamically allocates fusion weights based on the effective point density per unit area and the local normal consistency index, allowing regions with good geometric consistency and high sampling density to dominate the fusion process, thereby maximizing the use of high-quality observation data and improving the reconstruction fidelity of key areas. Experimental data show that after adopting this strategy, the curvature reconstruction error of the tooth root transition region decreased from 12% to 3.8%, significantly improving the detection reliability of weak points.

[0033] In the feature extraction stage, traditional methods often employ a fixed-density, equal-angle cross-section subdivision strategy, sampling at the same interval regardless of curvature, resulting in loss of detail in high-curvature areas and waste of resources in low-curvature areas. This solution introduces an adaptive sampling density strategy, dynamically adjusting the number of sampling points based on local curvature: 5 points per degree in the tooth tip arc and tooth root transition zone, and sparser at 1 point per degree in the involute segment. This strategy improves the resolution of deviation detection in critical areas by 5 times without increasing the overall computational burden, effectively capturing minute machining defects. Combined with an improved Hausdorff distance algorithm, the system can not only report the maximum deviation value but also distinguish between overcut and undercut states, providing directional guidance for subsequent grinding compensation. For example, if a region shows significant overcut, the CNC program should reduce the grinding amount at that location; conversely, if it is undercut, the depth of cut needs to be increased. This directional correction capability is significantly superior to traditional systems that only provide absolute deviations, shortening the process debugging cycle.

[0034] At the system operation and maintenance level, traditional equipment requires periodic shutdowns and calibration in the metrology laboratory, severely impacting production line continuity. This solution integrates a six-degree-of-freedom fine-tuning platform with a standard reference target assembly to achieve fully automated on-site calibration. Each time the equipment is powered on or changed, the system automatically executes a static internal parameter self-calibration process, reversing and updating the internal parameters of each sensor to eliminate imaging model drift caused by handling, vibration, or aging. Combined with a dynamic external parameter compensation mechanism, the equipment calibration cycle is extended from once daily to once monthly, improving operation and maintenance efficiency by more than 30 times. Simultaneously, the system supports rapid changeover for multiple models, enabling one-click parameter switching via pre-stored configuration packages, with preparation time controlled within 30 seconds, making it suitable for flexible manufacturing scenarios.

[0035] In summary, the laser point cloud-based adaptive calibration pinwheel tooth profile error detection system provided in this embodiment comprehensively improves the applicability, stability, and intelligence of non-contact measurement in complex industrial environments by constructing a closed-loop technical chain of "sensing-compensation-reconstruction-evaluation." The system not only achieves sub-micron level measurement accuracy and over 98% point cloud coverage, but also possesses autonomous response capabilities to external disturbances and rapid adaptation capabilities to multiple product models, providing reliable technical support for online full inspection and adaptive processing of high-end transmission components.

Claims

1. A pin wheel tooth profile error detection system based on adaptive calibration of laser point cloud, characterized in that, include: A laser scanning array module is used to arrange multiple line structured light sensors around the circumference of the needle wheel under test to obtain a raw point cloud cluster covering the entire tooth surface; A six-degree-of-freedom fine-tuning platform is used to carry a standard reference target ball assembly and drive it to execute a preset spatial motion trajectory. An environmental parameter sensing unit is used to monitor changes in the physical field that affect the accuracy of optical measurements in real time. A point cloud processing engine is used to receive the original point cloud clusters and environmental parameter data, and to perform point cloud denoising, completion, and adaptive weighted fusion to generate a complete tooth surface mesh model. The point cloud processing engine is also used to perform a dual calibration mechanism. The core of the error assessment is used to compare the complete tooth surface mesh model with the theoretical model to generate a tooth profile error report; The dual calibration mechanism includes: static intrinsic parameter self-calibration, used to invert and update the internal parameters of each line structured light sensor based on the point cloud data of the standard reference target ball group under multiple attitudes before each detection task is started; and dynamic extrinsic parameter compensation, used to calculate and inject a rigid body transformation matrix to correct coordinate system offset in response to the trigger signal of the environmental parameter sensing unit during the needle wheel rotation scanning process, based on the local fitting results of the current partial point cloud and the built-in standard geometric primitives. When performing adaptive weighted fusion, the point cloud processing engine assigns fusion weights to point clouds from different perspectives based on the effective point density per unit area and the local normal consistency index of the point cloud subset under each perspective. The local normal consistency index is calculated by the direction variance of the first eigenvector obtained from the principal component analysis of the neighborhood point cloud. When generating the tooth profile error report, the error assessment core uses an improved Hausdorff distance algorithm that incorporates a direction-sensitive factor to calculate and record the deviation values ​​of the overcut and undercut tooth profile states, respectively. The environmental parameter sensing unit includes a temperature gradient sensor, a triaxial accelerometer, and an illuminance detector. The trigger signal for the dynamic external parameter compensation is that the temperature change rate, the peak value of vibration acceleration, or the sudden change amplitude of light intensity exceeds their respective preset thresholds. The error assessment core receives the complete tooth surface mesh model after dynamic compensation and adaptive fusion, and performs the final geometric error quantification analysis. The core first performs global registration between the reconstructed model and the theoretical CAD model stored in the local database, and uses an improved Hausdorff distance algorithm to calculate the maximum spatial deviation between the two. This system introduces a direction-sensitive factor and defines a bidirectional deviation metric. in, For the actual tooth surface model point set to be reconstructed, For the theoretical nominal model point set and These are the unit normal vectors at the corresponding points. This represents the vector dot product operation.

2. The pin wheel profile error detection system based on adaptive calibration of laser point cloud of claim 1, wherein, Each line structured light sensor in the laser scanning array module includes a laser projector with an adjustable emission angle and an industrial camera. The laser projector has a dynamic depth-of-field adjustment function, and its working distance is automatically matched by the controller according to the specifications of the needle wheel to be tested.

3. The needle wheel tooth profile error detection system based on laser point cloud adaptive calibration according to claim 2, characterized in that, When performing point cloud denoising and completion operations, the point cloud processing engine uses the initial registration results between adjacent frame point clouds to identify outliers and combines the workpiece posture prediction model to fill in the missing regions with geometric continuity.

4. The pin wheel profile error detection system based on adaptive calibration of laser point cloud of claim 3, wherein, The six-degree-of-freedom fine-tuning platform adopts a parallel structure, and its motion controller uses a composite control strategy. The position feedback comes from a high-resolution displacement sensor integrated into the platform base to ensure accurate reproduction of the calibration trajectory.

5. The pin gear tooth profile error detection system based on adaptive calibration of laser point cloud according to claim 4, characterized in that, In the feature extraction stage, the point cloud processing engine uses the actual tooth tip center obtained by fitting as a reference to generate equal-angle cross-section lines along the radial direction, and adopts an adaptive sampling density strategy to extract deviation sampling points on the cross-section lines. The sampling density in the tooth tip arc segment and the tooth root transition area is higher than that in the involute segment.

6. The pin wheel profile error detection system based on adaptive calibration of laser point cloud of claim 5, wherein, The error assessment core also integrates a statistical process control module, which is used to perform trend analysis on the detection data of similar pinwheels in consecutive batches, and automatically generate a maintenance early warning signal when the movement range exceeds the control limit multiple times in a row.

7. The pin wheel profile error detection system based on adaptive calibration of laser point cloud of claim 6, wherein, The system supports rapid changeover testing of multiple pinwheel models. By pre-storing different specifications of calibration parameter groups, theoretical models and evaluation templates, it automatically completes all configuration switching in response to product code input.

Images