A line section detection system based on non-contact scanning and a detection method thereof

Abstract

The present application relates to the technical field of line section detection, in particular to a line section detection system based on non-contact scanning and a detection method thereof, which comprises a profile feature intelligent extraction module, a defect multi-dimensional quantitative evaluation module, a quality intelligent decision module and an information transmission between algorithms; the profile feature intelligent extraction module converts the original laser scanning point cloud into profile feature data through front-end processing and outputs the profile feature set; the defect multi-dimensional quantitative evaluation module detects, classifies and quantifies defects on the profile feature set and outputs the defect evaluation result; based on the profile feature set and the defect evaluation result, the quality of the section is checked in size, the degree of defect is evaluated, the good product is determined, and the quality decision is made by grading and classification. Through the combination of machine vision, deep learning and automation technology, the automation and intelligent production of line section size measurement and surface defect detection are realized.

Description

Technical Field

[0001] This invention relates to the field of linear profile inspection technology, specifically to a linear profile inspection system and method based on non-contact scanning. Background Technology

[0002] With the rapid development of the profile processing industry, higher demands are placed on the measurement accuracy and efficiency of linear profiles. Traditional manual measurement methods suffer from subjective errors and low efficiency, and can no longer meet the needs of modern production. Linear profiles have a wide range of dimensions and are highly variable, requiring high precision. Linear profiles, especially strip timber, have numerous defect types with significant inconsistencies; many uncommon defect types are not recorded. Even within the same defect type, there are significant differences in color, texture, and shape. Traditional manual measurement methods suffer from subjective errors and low efficiency, or rely on visual sampling, which is incomplete and cannot guarantee the objectivity and consistency of test results. This leads to the entry of substandard products into the market, and the lack of statistical analysis and quality traceability functions makes them unsuitable for modern refined management and standardized product production. Summary of the Invention

[0003] The purpose of this invention is to provide a non-contact scanning-based linear profile inspection system and method to solve the problems mentioned in the background. This invention combines machine vision, deep learning and automation technologies to achieve automation and intelligence in linear profile size measurement and surface defect detection, thereby reducing labor costs and increasing production capacity.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a non-contact scanning-based line profile inspection system, comprising a contour feature intelligent extraction module (CFIEC), a defect multidimensional quantitative evaluation module (DMDE), and a quality intelligent decision module (IQD), and information transmission between the algorithms; The contour feature intelligent extraction module processes the original laser scan point cloud through front-end processing. ) is converted into contour feature data, and a contour feature set is output ( ); The multidimensional quantitative assessment module for defects detects, classifies, and quantifies defects based on the contour feature set, and outputs the defect assessment results. ); Based on the profile feature set and defect assessment results, quality decisions are made by performing dimensional inspections, defect assessments, good product determination, and grading of profiles. ).

[0005] As a further embodiment of the present invention, in the detection method of the non-contact scanning-based line profile detection system, the contour feature intelligent extraction module (CFIEC) uses a contour feature intelligent extraction algorithm to retrieve and calculate a contour feature set from the database. The Defect Multidimensional Quantization Evaluation (DMDE) module uses the Defect Multidimensional Quantization Evaluation algorithm to evaluate the transmitted contour feature set ( )Analyze and output defect assessment results ( The Intelligent Quality Decision (IQD) module uses an intelligent quality decision algorithm to re-evaluate the defect assessment results and feed them back in real time to the intelligent feature extraction module and the multi-dimensional quantitative assessment module for defect assessment, forming a closed loop for adaptive adjustment.

[0006] As a further embodiment of the present invention, the intelligent contour feature extraction algorithm includes a unified model for tilt correction and feature extraction: ; in: ; In the formula, The number of point clouds in a single frame represents the total number of 3D points acquired in a single frame of laser scanning, and is a positive integer determined by the sensor scanning accuracy. This represents the original single-point cloud data, indicating the first point in a single frame of the point cloud. The three-dimensional coordinates of each point, components Indicates the direction of the conveyor belt, Refers to the direction perpendicular to the conveyor belt, In the direction of height, , , All are in mm; The corrected feature point cloud represents the set of valid 3D point clouds after tilt correction and invalid point filtering. The extracted contour feature parameter set represents the set of core profile features extracted from the corrected point cloud; Configure parameters for the algorithm; a fixed set of parameters configured for all sub-algorithms of the CFIEC module. This is the CFIEC module algorithm mapping function, representing the overall algorithm logic function for intelligent extraction of contour features; Step 1: Obtain the tilt-corrected point cloud ( ): ; in, The tilt correction operator is a mathematical operator for tilt correction of the original point cloud. Step 2: Calculate the adaptive reference height ( ): ,in, The operator for calculating the reference height represents the operator for calculating the reference height based on the corrected point cloud and the statistical region; Step 3: Valid point determination value ( ): ; Step 4: Calculate the width ( ): ; in, The unfiltered contour width of a single frame. The coordinates of a single point perpendicular to the conveyor belt; Step 5: .

[0007] As a further embodiment of the present invention, in step two, the reference height The result calculated using the weighted regression method is: ; in: The middle region of the outline represents the selected effective area in the middle of the point cloud (usually the middle 100 points). Weights are based on height distribution; The height of a single point after tilt correction is represented as follows. The Middle The z-coordinates of each point after tilt correction; The average height of the middle region represents All points inside The arithmetic mean; The standard deviation of the height in the middle region represents... All points inside Standard deviation; In step three, the valid point determination value is the valid point mask. Through adaptive threshold determination, the first Valid judgment value of each point : ; Adaptive height threshold Determined by local statistical characteristics: ; in, This represents the standard deviation of the local height difference. This is the amplification factor for the standard deviation. The minimum adaptive threshold, The maximum adaptive threshold; This is the invalid point determination value for the SDK, representing the height value of an invalid point defined by the sensor SDK. In step five, the final contour width Obtained through Kalman filtering: ; in, for Width after filtering at any time for Width after filtering at any time for Moment-time Kalman gain, for Width is measured at all times. for Time-based prediction error covariance For process noise covariance, To measure the noise covariance.

[0008] As a further aspect of the present invention, the contour quality score Comprehensive evaluation of extraction quality: ; in, To score the contour quality, As the weight for quality score, The standard deviation of the width sequence. The mean of the width sequence. For effective point density, Surface noise level; Output a unified set of output features: ; in, As the reference height, This is the width after filtering. , , This represents the surface noise level.

[0009] As a further embodiment of the present invention, the defect multidimensional quantitative evaluation algorithm includes a unified defect detection and quantification model as follows: ; in: This is the mapping function for the DMDE module algorithm. For profile type parameters, Configure parameters for defect detection. For the set of detected defects, For individual defect data, This represents the total number of defects. These are statistical characteristics of defects; For the detected defects , its first The multidimensional feature vector of each defect is: ; in: : No. The depth of each defect represents the maximum depth difference between the defects. : No. The width of the defect indicates its maximum size in the direction perpendicular to the conveyor belt. : No. The length of a defect indicates its maximum size along the conveyor belt direction. : No. The crack width of a defect represents the actual crack width of the cracked defect (only applicable to cracked defects). : No. The area of ​​a defect, representing the two-dimensional projected area of ​​the defect ( ), : No. The location characteristics of each defect, the ratio of defect length to width ( ), : No. The location characteristics of the defect, and the distance of the defect from the edge of the profile; Unified processing for defect feature normalization: All defect features are normalized to the [0,1] interval. ; No. The normalized feature vector of each defect, The maximum defect depth threshold. The maximum defect width threshold. The maximum defect length threshold. The maximum crack width threshold. This is a truncation function.

[0010] As a further embodiment of the present invention, the defect multidimensional quantitative evaluation module quantifies defects, tracks defects, and assesses overall surface defects; The quantitative defects are scored for severity, and the severity score is calculated as follows: ) Calculated through an ensemble learning model: ; in: For the first The severity score of each defect. For ensemble learning models, To fuse weights in the model, For gradient boosting decision tree models, For linear models, It is the characteristic transformation function; The defect tracking performs cross-frame defect tracking using a data association algorithm: ; The association cost function considers spatial distance, feature similarity, and motion consistency: ; in, for The first moment One defect, for The first moment One defect, for The set of defects at any given moment For defect-related costs, For the two defects to be matched, For associated cost weights, The coordinates of the center position of the defect are: For the Euclidean norm, This is the normalized feature vector of the defect. The velocity vector of the defect; The overall surface defects are assessed by evaluating the overall surface defect degree of a single line profile. )calculate: ; in, The overall defect level of a single profile surface. The maximum defect severity, This is the defect area weighting coefficient. This represents the total area of ​​the defects. This refers to the surface area of ​​the profile inspection surface.

[0011] As a further aspect of the present invention, the defect multidimensional quantification evaluation algorithm outputs a standardized set of defect descriptions: ; in, The location of the defect center. For defect trajectory ID, The confidence level for defect detection.

[0012] As a further embodiment of the present invention, the quality intelligent decision-making algorithm includes inputting various surface features and defect data, sequentially performing dimensional conformity checks, defect degree calculation and aggregation, critical defect detection, good product determination, quality grade determination, and non-conformity cause analysis, and finally outputting quality grade, decision basis, and improvement suggestions to form a complete quality decision-making system. Make the linear profile have There are 1 detection surface, and the contour features of each surface are: The defect assessment result is , The unified mathematical model for quality decision-making is: ; in: This is the mapping function for the IQD module algorithm. For the first The contour feature set of each detection surface. No. A set of defects on each inspection surface. This represents the total number of inspection surfaces on the profile. For the quality grade of the profile, For decision-making basis and detailed information, These are quality standard parameters; The multi-criteria decision-making process for dimensional conformity inspection evaluates dimensional conformity using a multi-criteria decision function. ; Specifically: ; in, This is the result of the dimensional conformity assessment. For dimensional compliance decision function, The total length of the profile. For the set of dimensional tolerances, For the target width value, For width tolerance, For the target length value, For length tolerance; The aggregation model for calculating and aggregating the overall defect degree is as follows: Overall defect rate of the entire profile The calculation is performed by aggregating the surface defect levels: ; Weighted maximum aggregation: ; Weight This reflects the importance of different surfaces (e.g., the top surface usually has a higher weight); in, The overall defect level of the entire profile. For the defect degree aggregation operator, For the first Surface defect degree of each inspection surface For the detection surface weight; The fatal flaw detection is performed using Boolean logic, and the fatal flaw detection is implemented through Boolean logic expressions: ; in, This is the result of the fatal defect determination. The threshold for fatal crack width; The good product determination is achieved through a decision function for good product determination, including: ; The specific decision-making rules are as follows: ; in, For the good product determination decision function, To add contour features, The defect threshold for superior grade products. The defect threshold for qualified products. It has no fatal defects; The quality level determination is based on a fuzzy logic model for quality grading. For boundary cases, fuzzy logic is used for quality grading. ; The quality grade is determined by the maximum value of the membership degree of each grade; among which, For quality level membership function, For fuzzy classification threshold; The analysis of the causes of nonconforming products involves a multi-dimensional analysis of the causes of nonconformity. The vector of causes of nonconforming products is as follows: ; Each causal component is calculated using the rule system: ; ; ; in, This is the vector of reasons for non-compliance. The reason is that the dimensions are not up to standard. The reason for the defect rate exceeding the standard is... Due to the fatal defect, For other reasons, This is an indicator function.

[0013] As a further embodiment of the present invention, the quality intelligent decision-making algorithm outputs a quality assessment report including quality level, decision basis, and improvement suggestions: ; in, For quality assessment report, In terms of size status, For details of the fatal defect, Suggestions for improvement, For decision confidence level.

[0014] Compared with the prior art, the beneficial effects of the present invention are: the present invention includes a contour feature intelligent extraction module (CFIEC), a defect multidimensional quantitative evaluation module (DMDE), and a quality intelligent decision module (IQD), and the inter-algorithm information transmission between them; The contour feature intelligent extraction module processes the original laser scan point cloud through front-end processing. This algorithm transforms the original laser-scanned point cloud into stable and accurate contour feature data. It integrates multiple steps, including tilt correction, baseline calculation, effective point filtering, and width calculation, and uses an adaptive mechanism to ensure reliable contour features are obtained under various profile surface conditions. It outputs a contour feature set (…). ); The intelligent contour feature extraction algorithm adapts to different profile surfaces through dynamic threshold adjustment, exhibiting strong adaptability and robust regression to resist abnormal point interference. It is robust, supports high-speed production lines through flow processing, and its parameters have configurable performance that can be adjusted according to profile type and process.

[0015] (3) The multidimensional quantitative assessment module for defects detects defects, classifies defects, and quantifies defects on the contour feature set, and outputs the defect assessment results. As the core detection module of the system, it is responsible for detecting, classifying, and quantifying various defects from contour features. This algorithm integrates surface defect detection, crack detection, defect tracking, and machine learning weight optimization, achieving a comprehensive and accurate quantitative assessment of defects.

[0016] (4) Defect multidimensional quantitative evaluation algorithm, which uniformly processes multiple types of defects such as surface defects and cracking defects under the same framework, dynamically adjusts the feature weights of each defect based on machine learning, ensures the integrity of defect parameters through cross-frame tracking, and calculates the interpretability score by combining model prediction and physical parameters to calculate the severity of defects.

[0017] (5) Based on the profile feature set and defect assessment results, perform dimensional inspection, defect assessment, good product determination, and grade classification to make quality decisions regarding profile quality. The system's final decision-making module, based on contour features and defect assessment results, achieves comprehensive evaluation and intelligent decision-making regarding material quality. This algorithm integrates dimensional inspection, defect assessment, good product determination, and grading, forming a complete quality decision-making system.

[0018] (6) Quality intelligent decision-making algorithm, multi-level decision-making: progressive multi-level decision-making from dimensional inspection to defect assessment, fuzzy boundary processing of quality level through fuzzy logic processing, while providing detailed reasons for non-conformity and improvement suggestions for interpretable decision-making, and adaptively adjusting decision thresholds according to profile type and customer requirements.

[0019] The three algorithms have clear division of labor and standardized interfaces. Each algorithm has a complete mathematical description and rigorous formula expression. After integration and optimization, the overall detection accuracy is improved by 25-30%. It supports the rapid integration of new algorithm modules, has strong scalability, and the modular design facilitates algorithm updates and debugging. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0021] A non-contact scanning-based line profile inspection system includes a contour feature intelligent extraction module (CFIEC), a defect multidimensional quantitative evaluation module (DMDE), and a quality intelligent decision module (IQD), and information is transmitted between the algorithms; The contour feature intelligent extraction module processes the original laser scan point cloud through front-end processing. ) is converted into contour feature data, and a contour feature set is output ( ); The multidimensional quantitative assessment module for defects detects, classifies, and quantifies defects based on the contour feature set, and outputs the defect assessment results. ); Based on the profile feature set and defect assessment results, quality decisions are made by performing dimensional inspections, defect assessments, good product determination, and grading of profiles. ).

[0022] The detection method of the non-contact scanning-based line profile inspection system, wherein the contour feature intelligent extraction module (CFIEC) uses a contour feature intelligent extraction algorithm to retrieve and calculate the contour feature set from the database. The Defect Multidimensional Quantization Evaluation (DMDE) module uses the Defect Multidimensional Quantization Evaluation algorithm to evaluate the transmitted contour feature set ( )Analyze and output defect assessment results ( The Intelligent Quality Decision (IQD) module uses an intelligent quality decision algorithm to re-evaluate the defect assessment results and feed them back in real time to the intelligent feature extraction module and the multi-dimensional quantitative assessment module for defect assessment, forming a closed loop for adaptive adjustment.

[0023] Example 2 The intelligent contour feature extraction algorithm includes a unified model for tilt correction and feature extraction: ; in: ; In the formula, The number of point clouds in a single frame represents the total number of 3D points acquired in a single frame of laser scanning, and is a positive integer determined by the sensor scanning accuracy. This represents the original single-point cloud data, indicating the first point in a single frame of the point cloud. The three-dimensional coordinates of each point, components Indicates the direction of the conveyor belt, Refers to the direction perpendicular to the conveyor belt, In the direction of height, , , All are in mm; The corrected feature point cloud represents the set of valid 3D point clouds after tilt correction and invalid point filtering. The extracted contour feature parameter set represents the set of core profile features extracted from the corrected point cloud; Configure parameters for the algorithm; a fixed set of parameters configured for all sub-algorithms of the CFIEC module. This is a composite function for the CFIEC module algorithm mapping function, including sub-steps such as tilt correction, reference height calculation, valid point determination, and width filtering.

[0024] Example 3 Feature extraction is performed, and the feature extraction process includes the following steps: Step 1: Obtain the tilt-corrected point cloud ( ): ; in, The tilt correction operator is a mathematical operator for tilt correction of the original point cloud. Step 2: Calculate the adaptive reference height ( ): ,in, The reference height calculation operator represents the operator for calculating the reference height based on the corrected point cloud and the statistical region; it includes sub-logic for region selection, weight calculation, and weighted summation.

[0025] In step two, the reference height The result calculated using the weighted regression method is: ; in: The middle region of the outline represents the selected effective area in the middle of the point cloud (usually the middle 100 points); the patent specifies that the selected effective area in the middle of the point cloud is usually the middle 100 points. A subset of.

[0026] The weights are based on the height distribution; they take values ​​(0,1], with greater weights the closer the height is to the mean. The height of a single point after tilt correction is represented as follows. The Middle The z-coordinates of each point after tilt correction; The average height of the middle region represents All points inside The arithmetic mean; is the arithmetic mean of the arithmetic mean; it is used to calculate Statistical benchmark; The standard deviation of the height in the middle region represents... All points inside The standard deviation reflects the degree of dispersion of the height in the intermediate region and is used for weight decay adjustment.

[0027] Step 3: Valid point determination value ( ): ; In step three, the valid point determination value is the valid point mask. Through adaptive threshold determination, the first Valid judgment value of each point : ; Adaptive height threshold Determined by local statistical characteristics: ; Among them, adaptive height threshold Calculated dynamically from the statistical characteristics of local point clouds, not a fixed value; The "local area" is configured by the sensor's scanning accuracy, and is usually 5-10 points around a single point.

[0028] This represents the standard deviation of local height difference, reflecting the degree of fluctuation in height within a local area.

[0029] This is the standard deviation amplification factor, an empirical value used in the algorithm, typically between 2 and 3, which can be adjusted to suit different profile types. The minimum adaptive threshold is fixed at 0.5mm to avoid misjudgment caused by an excessively small threshold.

[0030] The maximum adaptive threshold is fixed at 10mm in the algorithm to avoid missed detections due to an excessively large threshold. This is the invalid point determination value for the SDK, representing the invalid point height value defined by the sensor SDK. In the algorithm, it is fixed at 1.0E9, which is an invalid value defined by the sensor hardware. Step 4: Calculate the width ( ): ; in, is the unfiltered contour width of a single frame, and is the original input value for width filtering; For the coordinates of a single point perpendicular to the conveyor belt, only take... The valid point coordinates.

[0031] Step 5: .

[0032] In step five, the final contour width Obtained through Kalman filtering: ; in, for Filtered width at time 1, 1 The frame's contour width after Kalman filtering. The core component is the final width measurement.

[0033] for The filtered width at any given time, and the filtered width of the previous frame, are the predicted values ​​of the Kalman filter and serve as historical reference values ​​for filtering.

[0034] for Kalman gain at time 1, 2 The frame's filter gain coefficient, with a value of (0,1), determines the degree of influence of the current measurement value on the filtering result.

[0035] for Width measured at time, the first Unfiltered profile width of the frame That is, the raw measurement value of a single frame.

[0036] for The prediction error covariance at any given time, and the error covariance of the filtered prediction value in the previous frame, reflect the reliability of the prediction value and are updated iteratively.

[0037] The process noise covariance is represented by the noise covariance caused by conveyor belt vibration, which is an empirical value ranging from 0.01 to 0.1. It is adapted to the speed of the conveyor belt.

[0038] To measure the noise covariance, the measurement noise covariance introduced by the sensor scanning is determined by the sensor's accuracy and is typically taken as 0.001 to 0.01. .

[0039] The contour quality score Comprehensive evaluation of extraction quality: ; in, The contour quality score comprehensively evaluates the quality of contour feature extraction, with values ​​ranging from (0,1). Values ​​closer to 1 indicate higher extraction quality. Quantity The quality score weights are used as weighting coefficients for each evaluation dimension, satisfying the following conditions: upper surface width stability It has the highest weight (usually 0.4).

[0040] The standard deviation of the width sequence, the width after filtering of multiple consecutive frames. The standard deviation reflects the stability of width measurement.

[0041] The width is the average of the width sequence, and the width after filtering of multiple consecutive frames. The arithmetic mean of the values ​​is used as the baseline for width statistics.

[0042] The effective point density takes values ​​in the range (0,1). Number of valid points / Total number of point clouds in a single frame .

[0043] Surface noise level; The core component reflects the smoothness of the profile surface; the lower the noise level, the smaller the value.

[0044] Output a unified set of output features: ; in, As the reference height, This is the width after filtering. , , These are the slope / angle parameters of the profile, respectively, determined by the tilt correction operator. The calculations reflect the degree of tilt in the placement of the profiles, providing tilt compensation for subsequent algorithms. for This represents the surface noise level.

[0045] Example 4 The multidimensional quantitative evaluation algorithm for defects includes a unified model for defect detection and quantification: ; in: This is the DMDE module algorithm mapping function, the overall algorithm logic function for multidimensional quantitative evaluation of defects, and a composite function that includes sub-steps such as defect detection, feature extraction, quantization, and tracking.

[0046] For profile type parameters, adapt to the defect detection parameters of different profile types, and configure according to material and process requirements, such as defect threshold and crack judgment criteria.

[0047] The DMDE module has a fixed set of configuration parameters for defect detection, including minimum defect size, number of tracking frames, and model fusion weights.

[0048] For the set of detected defects, For a single defect, the data includes all information such as the defect's location, size, type, and severity. The total number of defects, a non-negative integer. Indicates no defects; The statistical characteristics of defects include the number of defects, type distribution, size statistics, and average severity, providing a statistical basis for quality decisions.

[0049] For the detected defects , its first The multidimensional feature vector of each defect is: ; in: : No. The depth of each defect represents the maximum depth difference between the defects, and the pit is... , the protrusion is Take the maximum value, in mm.

[0050] : No. The width of a defect represents the maximum size of the defect in the direction perpendicular to the conveyor belt. It is the maximum difference in the defect area in the y-direction, and the unit is mm.

[0051] : No. The length of a defect represents the maximum size of the defect along the conveyor belt direction. It is calculated by the number of defect tracking frames × conveyor belt speed × acquisition interval, and the unit is mm.

[0052] : No. The crack width of a defect indicates the actual crack width of the cracked defect (only for cracked defects), while non-cracked defects... It is calculated from the position difference of the normal points before and after the cracked section, and the unit is mm.

[0053] : No. The area of ​​a defect, representing the two-dimensional projected area of ​​the defect ( ), which is an approximate area.

[0054] : No. The location characteristics of each defect, the ratio of defect length to width ( ( ), reflecting the morphological characteristics of defects.

[0055] : No. The location characteristics of a defect, the distance of the defect from the edge of the profile; reflecting the position of the defect on the profile surface, and the basis for determining edge defects / center defects.

[0056] Unified processing for defect feature normalization: All defect features are normalized to the [0,1] interval. ; No. The normalized feature vector of each defect, with all components taking values ​​[0,1], eliminates the influence of dimensions and serves as the input for the defect severity score.

[0057] The maximum defect depth threshold is determined by... For example, pine wood is typically 5mm thick.

[0058] The maximum defect width threshold is determined by... Configuration, such as the usual 10mm.

[0059] The maximum defect length threshold is determined by... Configuration, such as the usual 50mm.

[0060] The maximum crack width threshold is determined by... Configuration, such as the usual 8mm, This is a fatal flaw.

[0061] This is a truncation function that ensures the normalization result is 1 when the feature value exceeds the threshold, reflecting that the defect exceeds the standard.

[0062] The multidimensional quantitative assessment module for defects quantifies defects, tracks defects, and assesses overall surface defects. The quantitative defects are scored for severity, and the severity score is calculated as follows: ) Calculated through an ensemble learning model: ; in: For the first The severity score of each defect is a quantitative value of the severity of a single defect, with a value of [0,1]. The closer it is to 1, the more severe the defect.

[0063] This is an ensemble learning model for defect severity scoring, which is a fusion of GBDT and a linear model, taking into account both nonlinear fitting and interpretability.

[0064] This is the model fusion weight, the weight in ensemble learning, with a value of (0,1), usually 0.7, which can be adjusted according to the model performance.

[0065] This involves using a gradient boosting decision tree model and a nonlinear feature fitting model to capture the nonlinear relationships between defective features and improve scoring accuracy.

[0066] It is a linear model, a linear feature fitting model, to ensure the interpretability of the scoring results.

[0067] The feature transformation function performs transformation operations on the normalized feature vector, such as standardization, normalization, and feature crossing, to adapt the input to the linear model.

[0068] The defect tracking performs cross-frame defect tracking using a data association algorithm: ; The association cost function considers spatial distance, feature similarity, and motion consistency: ; in, for The first moment The first defect, the The first frame detected Each defect is a historical defect data point for defect tracking.

[0069] for The first moment One defect, No. Frame and The drawbacks of matching are due to association costs. Minimize the result to achieve cross-frame defect tracking.

[0070] for The set of defects at time 1, the first All defects detected in the frame constitute a candidate set for defect matching.

[0071] This represents the cost of associating defects, the cost of matching two defects. It takes a value greater than or equal to 0, and the smaller the value, the higher the degree of matching between the two defects.

[0072] Let be the two defects to be matched, and any two defects to be matched (such as historical defects and current defects) be the objects of the associated cost calculation.

[0073] As the association cost weight, the weighting coefficients for each matching dimension satisfy... Location features It has the highest weight (usually 0.5).

[0074] The coordinates of the center position of the defect are given, the two-dimensional center positions (x, y) of the two defects are given, and the geometric center coordinates of the defect region are given.

[0075] The distance is the Euclidean norm, which reflects the similarity between two features; the smaller the distance, the higher the similarity.

[0076] This is the normalized feature vector of the defect, reflecting the similarity of the defect's morphological features.

[0077] The velocity vector of the defect is calculated from the change in the defect position and the acquisition interval, reflecting the consistency of the motion, and is expressed in mm / s. The overall surface defects are assessed by evaluating the overall surface defect degree of a single line profile. )calculate: ; in, The overall defect degree of a single profile surface is a quantitative value of the defect severity of a single inspected surface of the profile, with a value of [0,1], which is the core input for quality decision-making.

[0078] For the maximum defect severity, among all defects on this surface The maximum value of , which takes the range [0,1], reflects the most severe degree of a single defect.

[0079] This is the defect area weighting coefficient, which is the amplification factor for the proportion of the total defect area.

[0080] The total defect area represents the sum of the areas of all defects on the surface. It is an empirical value, usually taken as 0.8 to 1.0, balancing the influence of individual defects and the total defect area. .

[0081] The surface area of ​​the profile inspection surface is the two-dimensional projected area of ​​a single inspection surface, calculated from the length and width of the profile, in units of... .

[0082] The defect multidimensional quantification evaluation algorithm outputs a standardized set of defect descriptions: ; in, This indicates the center location of the defect, reflecting its specific location on the profile surface.

[0083] The defect trajectory ID is a positive integer. The trajectory ID of the same defect is the same in different frames, enabling full lifecycle tracking of defects.

[0084] This represents the defect detection confidence level, with a value of [0,1]. The closer it is to 1, the more reliable the defect detection result is. It is output by the DMDE algorithm.

[0085] Example 5 The quality intelligent decision-making algorithm includes inputting various surface features and defect data, and sequentially performing dimensional conformity checks, defect degree calculation and aggregation, critical defect detection, good product determination, quality registration determination, non-conformity cause analysis, and finally outputting quality grade, decision basis and improvement suggestions to form a complete quality decision-making system. Make the linear profile have There are 1 detection surface, and the contour features of each surface are: The defect assessment result is , The unified mathematical model for quality decision-making is: ; in: This is the algorithm mapping function for the IQD module, the overall algorithm logic function for intelligent quality decision-making, and a composite function containing sub-steps such as dimensional conformity inspection, defect aggregation, critical defect detection, and good product determination.

[0086] For the first The contour feature set of the detection surface, the profile of the first detection surface Each detection surface The CFIEC module provides independent outputs for each detection surface.

[0087] No. The set of defects on the inspection surface, the profile's first... Each detection surface The set consists of the independent outputs of the DMDE module for each detection surface.

[0088] This refers to the total number of surfaces of the profile being inspected. It is a positive integer, determined by the inspection process, such as four surfaces: top, bottom, left, and right. .

[0089] This refers to the quality grade of the profile, the overall quality classification result of the profile, defined by the patent as superior / qualified / unqualified, which is the core output of the IQD module.

[0090] This information serves as the basis for decision-making and includes comprehensive defect rate, dimensional condition, fatal defects, reasons for nonconformity, etc., and is the core content of the quality assessment report.

[0091] These are quality standard parameters, including dimensional tolerances, defect thresholds, and quality grade classification standards, which are defined by the company's process standards.

[0092] The multi-criteria decision-making process for dimensional conformity inspection evaluates dimensional conformity using a multi-criteria decision function. ; Specifically: ; in, The result of the dimensional conformity assessment is binarized, with 1 indicating conformity and 0 indicating non-conformity.

[0093] This is the size compliance decision function, the size compliance judgment logic function, which is based on Boolean logic judgment based on multiple criteria (width + length).

[0094] The total length of the profile is calculated by the total number of acquisition frames × conveyor belt speed × acquisition interval, and the unit is mm.

[0095] This is the set of dimensional tolerances, specifically the set of tolerances for the width and length of the profile. Defined by the company's process standards, the unit is mm.

[0096] The target width value is the standard width value of the profile required by the process, such as 20mm or 22mm.

[0097] Width tolerance refers to the allowable width deviation range in the process, such as ±0.5mm, which is defined by the company's process standards.

[0098] The target length value is the standard value of the profile length required by the process, which is defined by the company's process standards, such as 1000mm or 2000mm.

[0099] The length tolerance is the length allowed by the process, such as ±5mm, as defined by the company's process standards.

[0100] The aggregation model for calculating and aggregating the overall defect degree is as follows: Overall defect rate of the entire profile The calculation is performed by aggregating the surface defect levels: ; Weighted maximum aggregation: ; Weight This reflects the importance of different surfaces (e.g., the top surface usually has a higher weight); in, The overall defect degree of the entire profile is denoted by [0,1], which is the core basis for judging good products.

[0101] The defect aggregation operator uses weighted maximum aggregation to reflect the principle that "any severe defect on any surface dominates the overall quality".

[0102] For the first The surface defect degree of each detection surface, the output value of the DMDE module for each detection surface, takes the value [0,1].

[0103] For the detection surface weight, satisfying (Maximum aggregation), the top surface has the highest weight (usually 1.0), and the secondary surfaces have a weight of 0.5~0.8.

[0104] The fatal flaw detection is performed using Boolean logic, and the fatal flaw detection is implemented through Boolean logic expressions: ; in, This is the result of a fatal defect determination; the result indicates whether a fatal defect exists. This indicates that it exists. If it does not exist, it is directly judged as a defective product; if it does exist, it is directly judged as a defective product.

[0105] This is the critical crack width threshold; the critical crack defect threshold defined by the process, and is the same as in section 2.3. Consistent, such as 8mm.

[0106] The good product determination is achieved through a decision function for good product determination, including: ; The specific decision-making rules are as follows: ; in, The decision function for determining good products is based on a three-level Boolean logic based on size, defect rate, and fatal defects.

[0107] Additional contour features, such as surface noise and tilt, are optional criteria for judgment and usually do not affect the core judgment. They are only used for fine-tuning of boundary conditions.

[0108] The defect threshold for superior products is the upper limit of the overall defect rate for superior products, with a value of [0,1], defined by the enterprise's process standards, such as 0.2.

[0109] This is the defect threshold for qualified products, representing the upper limit of the overall defect rate for qualified products. It takes a value of [0,1], defined by the company's process standards, such as 0.5, which satisfies... .

[0110] No fatal defects; fatal defect determination result is The Boolean logic NOT is a necessary condition for a product to be of superior quality or qualified quality.

[0111] The quality level determination is based on a fuzzy logic model for quality grading. For boundary cases, fuzzy logic is used for quality grading. ; The quality grade is determined by the maximum value of the membership degree of each grade; in, This is the membership function for quality levels, which combines the membership degree of defects to a certain quality level. It takes values ​​[0,1] and is used for fuzzy classification of boundary defect degrees to avoid misjudgment by hard thresholds. The threshold for fuzzy classification; the upper and lower limits of fuzzy classification, usually... , Adapt to core judgment criteria.

[0112] The analysis of the causes of nonconforming products involves a multi-dimensional analysis of the causes of nonconformity. The vector of causes of nonconformity for nonconforming products is as follows: ; Each causal component is calculated using the rule system: ; ; ; in, This is a vector of reasons for non-conformity, representing the set of reasons for non-conforming products. It is a 4×1 vector with binary components (1 = this reason causes non-conformity, 0 = no such reason). The reason for dimensional non-compliance, the judgment result of dimensional non-compliance leading to profile non-compliance, binarization. , This is an indicator function (it takes 1 if the condition is true, and 0 otherwise). The reasons for exceeding the defect limit are summarized, and the judgment result of non-conformity due to exceeding the defect limit is binarized. .

[0113] The cause of the fatal defect is the existence of a fatal defect leading to a non-conformance judgment result, which is then binarized. .

[0114] For other reasons or factors (such as sensor failure or missing point cloud data) that lead to an unqualified judgment result, binarization is performed and output by the system's anomaly detection module. .

[0115] This is an indicator function, an indicator operation for conditional judgment. It takes 1 if the condition is true and 0 otherwise. It is the core operator for binarization.

[0116] The quality intelligent decision-making algorithm outputs a quality assessment report that includes the quality level, decision basis, and improvement suggestions. ; in, The quality assessment report is the final output report of the IQD module, which includes the results, basis, and recommendations of quality judgment, providing complete information for enterprise production decisions; This refers to the dimensional status, detailing the profile dimensions, such as "Width acceptable (20.2mm), Length acceptable (1003mm)" or "Width exceeds standard (21.0mm, target 20±0.5mm)".

[0117] This section details the fatal defects, providing specific information about them. If there are no fatal defects, it will be marked "None". If there are fatal defects, it will include the location, size, and crack width of the defect.

[0118] As improvement suggestions, process improvement suggestions for the reasons for non-conformity are generated by the IQD module based on the reasons for non-conformity, such as "Width exceeds standard: Adjust conveyor belt positioning device" and "Crack defect: Optimize processing technology".

[0119] The decision confidence level represents the overall reliability of the quality level determination. It takes a value of [0,1]. The closer it is to 1, the more reliable the determination result is. It is calculated by combining the output confidence level of the CFIEC / DMDE module.

[0120] Through algorithm integration and optimization, the profile laser profilometer inspection system has formed a complete algorithm system consisting of a profile feature intelligent extraction module (CFIEC), a defect multidimensional quantitative evaluation module (DMDE), and a quality intelligent decision module (IQD). These three algorithms each have clear mathematical descriptions and unified formulas, maintaining the flexibility of modular design while ensuring the overall synergy and efficiency of the algorithms.

[0121] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-contact scanning-based linear profile inspection system, characterized in that: It includes a contour feature intelligent extraction module (CFIEC), a defect multidimensional quantitative evaluation module (DMDE), and a quality intelligent decision module (IQD), and facilitates inter-algorithm information transfer between these modules; The contour feature intelligent extraction module processes the original laser scan point cloud through front-end processing. ) is converted into contour feature data, and a contour feature set is output ( ); The multidimensional quantitative assessment module for defects detects, classifies, and quantifies defects based on the contour feature set, and outputs the defect assessment results. ); Based on the profile feature set and defect assessment results, quality decisions are made by performing dimensional inspections, defect assessments, good product determination, and grading of profiles. ).

2. A detection method for a non-contact scanning-based linear profile detection system as described in claim 1, characterized in that: The contour feature intelligent extraction module (CFIEC) uses a contour feature intelligent extraction algorithm to retrieve and calculate a contour feature set from the database. The Defect Multidimensional Quantization Evaluation (DMDE) module uses the Defect Multidimensional Quantization Evaluation algorithm to evaluate the transmitted contour feature set ( )Analyze and output defect assessment results ( The Intelligent Quality Decision (IQD) module uses an intelligent quality decision algorithm to re-evaluate the defect assessment results and feed them back in real time to the intelligent feature extraction module and the multi-dimensional quantitative assessment module for defect assessment, forming a closed loop for adaptive adjustment.

3. The detection method of the non-contact scanning-based linear profile detection system according to claim 2, characterized in that: The intelligent contour feature extraction algorithm includes a unified model for tilt correction and feature extraction: ； in: ； In the formula, The number of point clouds in a single frame represents the total number of 3D points acquired in a single frame of laser scanning, and is a positive integer determined by the sensor scanning accuracy. This represents the original single-point cloud data, indicating the first point in a single frame of the point cloud. The three-dimensional coordinates of each point, components Indicates the direction of the conveyor belt, Refers to the direction perpendicular to the conveyor belt, In the direction of height, , , All are in mm; The corrected feature point cloud represents the set of valid 3D point clouds after tilt correction and invalid point filtering. The extracted contour feature parameter set represents the set of core profile features extracted from the corrected point cloud; Configure parameters for the algorithm; a fixed set of parameters configured for all sub-algorithms of the CFIEC module. This is the CFIEC module algorithm mapping function, representing the overall algorithm logic function for intelligent extraction of contour features; The feature extraction process includes the following steps: Step 1: Obtain the tilt-corrected point cloud ( ): ； in, The tilt correction operator is a mathematical operator for tilt correction of the original point cloud. Step 2: Calculate the adaptive reference height ( ): ,in, The operator for calculating the reference height represents the operator for calculating the reference height based on the corrected point cloud and the statistical region; Step 3: Valid point determination value ( ): ； Step 4: Calculate the width ( ): ； in, The unfiltered contour width of a single frame. The coordinates of a single point perpendicular to the conveyor belt; Step 5: .

4. The non-contact scanning-based line profile inspection system according to claim 3, characterized in that: In step two, the reference height The result calculated using the weighted regression method is: ； in: The middle region of the outline represents the selected effective area in the middle of the point cloud (usually the middle 100 points). Weights are based on height distribution; The height of a single point after tilt correction is represented as follows. The Middle The z-coordinates of each point after tilt correction; The average height of the middle region represents All points inside The arithmetic mean; The standard deviation of the height in the middle region represents... All points inside Standard deviation; In step three, the valid point determination value is the valid point mask. Through adaptive threshold determination, the first Valid judgment value of each point : ； Adaptive height threshold Determined by local statistical characteristics: ； in, This represents the standard deviation of the local height difference. This is the amplification factor for the standard deviation. The minimum adaptive threshold, The maximum adaptive threshold; This is the invalid point determination value for the SDK, representing the height value of an invalid point defined by the sensor SDK. In step five, the final contour width Obtained through Kalman filtering: ； in, for Width after filtering at any time for Width after filtering at any time for Moment-time Kalman gain, for Width is measured at all times. for Time-based prediction error covariance For process noise covariance, To measure the noise covariance.

5. The non-contact scanning-based line profile inspection system according to claim 4, characterized in that: The contour quality score Comprehensive evaluation of extraction quality: ； in, To score the contour quality, As the weight for quality score, The standard deviation of the width sequence. The mean of the width sequence. For effective point density, Surface noise level; Output a unified set of output features: ； in, As the reference height, This is the width after filtering. , , This represents the surface noise level.

6. The non-contact scanning-based line profile inspection system according to claim 4, characterized in that: The multidimensional quantitative evaluation algorithm for defects includes a unified model for defect detection and quantification: ； in: This is the mapping function for the DMDE module algorithm. For profile type parameters, Configure parameters for defect detection. For the set of detected defects, For individual defect data, This represents the total number of defects. These are statistical characteristics of defects; For the detected defects , its first The multidimensional feature vector of each defect is: ； in: : No. The depth of each defect represents the maximum depth difference between the defects. : No. The width of the defect indicates its maximum size in the direction perpendicular to the conveyor belt. : No. The length of a defect indicates its maximum size along the conveyor belt direction. : No. The crack width of a defect represents the actual crack width of the cracked defect (only applicable to cracked defects). : No. The area of ​​a defect, representing the two-dimensional projected area of ​​the defect ( ), : No. The location characteristics of each defect, the ratio of defect length to width ( ), : No. The location characteristics of the defect, and the distance of the defect from the edge of the profile; Unified processing for defect feature normalization: All defect features are normalized to the [0,1] interval. ； No. The normalized feature vector of each defect, The maximum defect depth threshold. The maximum defect width threshold. The maximum defect length threshold. The maximum crack width threshold. This is a truncation function.

7. The non-contact scanning-based line profile inspection system according to claim 6, characterized in that: The multidimensional quantitative assessment module for defects quantifies defects, tracks defects, and assesses overall surface defects. The quantitative defects are scored for severity, and the severity score is calculated as follows: ) Calculated through an ensemble learning model: ； in: For the first The severity score of each defect. For ensemble learning models, To fuse weights in the model, For gradient boosting decision tree models, For linear models, It is the characteristic transformation function; The defect tracking performs cross-frame defect tracking using a data association algorithm: ； The association cost function considers spatial distance, feature similarity, and motion consistency: ； in, for The first moment One defect, for The first moment One defect, for The set of defects at any given moment For defect-related costs, For the two defects to be matched, For associated cost weights, The coordinates of the center position of the defect are: For the Euclidean norm, This is the normalized feature vector of the defect. The velocity vector of the defect; The overall surface defects are assessed by evaluating the overall surface defect degree of a single line profile. )calculate: ； in, The overall defect level of a single profile surface. The maximum defect severity, This is the defect area weighting coefficient. This represents the total area of ​​the defects. This refers to the surface area of ​​the profile inspection surface.

8. The non-contact scanning-based line profile inspection system according to claim 7, characterized in that: The defect multidimensional quantification evaluation algorithm outputs a standardized set of defect descriptions: ； in, The location of the defect center. For defect trajectory ID, The confidence level for defect detection.

9. The non-contact scanning-based line profile inspection system according to claim 7, characterized in that: The quality intelligent decision-making algorithm includes inputting various surface features and defect data, sequentially performing dimensional conformity checks, defect degree calculation and aggregation, critical defect detection, good product determination, quality grade determination, and non-conformity cause analysis, and finally outputting the quality grade, decision basis, and improvement suggestions to form a complete quality decision-making system. Make the linear profile have There are 1 detection surface, and the contour features of each surface are: The defect assessment result is , The unified mathematical model for quality decision-making is: ； in: This is the mapping function for the IQD module algorithm. For the first The contour feature set of each detection surface. No. A set of defects on each inspection surface. This represents the total number of inspection surfaces on the profile. For the quality grade of the profile, For decision-making basis and detailed information, These are quality standard parameters; The multi-criteria decision-making process for dimensional conformity inspection evaluates dimensional conformity using a multi-criteria decision function. ； Specifically: ； in, This is the result of the dimensional conformity assessment. For dimensional compliance decision function, The total length of the profile. For the set of dimensional tolerances, For the target width value, For width tolerance, For the target length value, For length tolerance; The aggregation model for calculating and aggregating the overall defect degree is as follows: Overall defect rate of the entire profile The calculation is performed by aggregating the surface defect levels: ； Weighted maximum aggregation: ； Weight This reflects the importance of different surfaces (e.g., the top surface usually has a higher weight); in, The overall defect level of the entire profile. For the defect degree aggregation operator, For the first Surface defect degree of each inspection surface For the detection surface weight; The fatal flaw detection is performed using Boolean logic, and the fatal flaw detection is implemented through Boolean logic expressions: ； in, This is the result of the fatal defect determination. The threshold for fatal crack width; The good product determination is achieved through a decision function for good product determination, including: ； The specific decision-making rules are as follows: ； in, For the good product determination decision function, To add contour features, The defect threshold for superior grade products. The defect threshold for qualified products. It has no fatal defects; The quality level determination is based on a fuzzy logic model for quality grading. For boundary cases, fuzzy logic is used for quality grading. ； The quality grade is determined by the maximum value of the membership degree of each grade; among which, For quality level membership function, For fuzzy classification threshold; The analysis of the causes of nonconforming products involves a multi-dimensional analysis of the causes of nonconformity. The vector of causes of nonconforming products is as follows: ； Each causal component is calculated using the rule system: ； ； ； in, This is the vector of reasons for non-compliance. The reason is that the dimensions are not up to standard. The reason for the defect rate exceeding the standard is... Due to the fatal defect, For other reasons, This is an indicator function.

10. The non-contact scanning-based line profile inspection system according to claim 7, characterized in that: The quality intelligent decision-making algorithm outputs a quality assessment report that includes the quality level, decision basis, and improvement suggestions. ； in, For quality assessment report, In terms of size status, For details of the fatal defect, Suggestions for improvement, For decision confidence level.

Images