A method and system for on-line measurement of the ovality of a continuous casting round billet by laser

Abstract

This invention discloses a laser online measurement method and system for the ellipticity of continuously cast round billets, belonging to the field of online detection and machine vision measurement in metallurgical processes. To solve the problem of contour distortion caused by high-temperature radiation, steam, oxide scale, and billet vibration, and the difficulty in distinguishing between runout and true elliptical deformation, this invention achieves the technical effect of reducing false out-of-tolerance and missed reports and outputting reliable continuous out-of-tolerance sections by synchronous multi-view differential contour reconstruction, continuous mass weight propagation, dual-domain state decomposition of acquisition time and material coordinates, polar temperature registration, and robust elliptic fitting at reference temperature.

Description

Technical Field

[0001] This invention relates to the field of online inspection and machine vision measurement in metallurgical processes, and particularly to a method and system for online laser measurement of the ellipticity of continuously cast round billets. Background Technology

[0002] Continuously cast round billets typically run continuously along roller conveyors at high temperatures. To obtain cross-sectional dimension information before the billet enters subsequent rolling processes, existing online diameter measurement solutions usually employ laser profile measurement or machine vision to acquire the surface profile of the billet and calculate the diameter or cross-sectional shape parameters accordingly.

[0003] However, the intense heat radiation at the continuous casting site can easily saturate the laser stripes. Secondary cooling steam and randomly detached oxide scale can cause stripe breakage, contour edge drift, or local occlusion. Furthermore, billet vibration and roller table runout can cause common-mode displacement in the measured contour. Relying solely on unidirectional diameter or contour processing without continuous reliability constraints makes it difficult to distinguish between true elliptical deformation, thermal expansion, roller table runout, and measurement noise. It is also difficult to reliably suppress the output for insufficient cross-sections, which may lead to false out-of-tolerance or missed reports.

[0004] Therefore, there is a need for a laser-based online measurement method and system for the ellipticity of continuously cast round billets that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose an online laser measurement method and system for the ellipticity of continuously cast round billets. Addressing the issues of contour distortion caused by high-temperature radiation, steam, oxide scale, and billet vibration, and the difficulty in separating true elliptical deformation, roller runout, thermal expansion, and measurement noise in unidirectional diameter measurement, this invention proposes a method for simultaneously acquiring paired frames of multi-view laser switches, circumferential temperature, and encoder data. It retains the same quality weight formed by imaging quality and cross-view geometric consistency at the same point, performs robust solution calculations, and determines validity. The acquisition time and material coordinates are used to index the center displacement state and shape state respectively. Combined with polar angle temperature registration, point-by-point reference temperature conversion, and positive definite constraint robust ellipse fitting techniques, this invention achieves the technical effects of separating common-mode runout from the true shape, limiting the influence of abnormal contours, and only alarming reliable continuous over-limit sections.

[0006] This invention provides a method for online laser measurement of the ellipticity of continuously cast round billets, comprising: S1, simultaneously acquiring multi-view laser-closed background frames, laser-open contour frames, circumferential temperature data, and roll speed encoder data of the same nominal cross-section of the continuously cast round billet; S2, performing differential processing on the background frames and contour frames and reconstructing contour points in a common coordinate system, and determining the physical quality weights bound to the contour points based on the imaging quality of the differential fringes; S3, correcting the physical quality weights based on the geometric consistency of contour arc segments from different viewpoints, and ensuring that the corrected quality weights are applied throughout the contour point selection, robust solution, and cross-sectional validity determination; S4, establishing material coordinates based on the roll speed encoder data, and using... The system collects data on the displacement state of the center indexed by the time and the elliptical shape state indexed by the material coordinates. Based on the weighted robust residuals of the multi-view contour points, it jointly solves for the two states to obtain the cross-sectional center and the thermal elliptical shape after separating the common mode runout. S5. Based on the cross-sectional center, it determines the polar angle of the contour points, registers the circumferential temperature data to the contour points according to the polar angle, and converts the contour points to the reference temperature according to the linear expansion function. S6. It performs robust elliptical fitting with mass weights on the contour points converted to the reference temperature, determines the measurement state based on the sufficiency of the point set and the fitting uncertainty, and outputs an alarm segment when the measurement state is reliable, the ellipticity exceeds the upper limit of the allowable ellipticity and is continuous along the material coordinates.

[0007] Optionally, S1 includes: setting at least four sets of narrow-strip line laser camera probes on the outer periphery of the continuous casting round billet roller table, and establishing the transformation relationship between the probe coordinate system and the common coordinate system through common calibration and pointing to the same nominal cross section;

[0008] A unified trigger signal causes each narrowband line laser camera probe to sequentially obtain the laser-off background frame and the laser-on contour frame within a cross-sectional acquisition cycle, and the circumferential temperature sampling time and the roller speed encoder count are associated with the cross-sectional acquisition cycle.

[0009] The material coordinate increment is determined based on the difference in counts between adjacent encoders and the unit pulse displacement of the roller speed encoder, and the material coordinate increment is associated with the cross-section acquisition cycle to form a synchronous cross-section data group.

[0010] Optionally, S2 includes: estimating the gain difference and offset difference between the background frame and the contour frame within the corresponding pixel area, and subtracting the paired frames after gain and offset matching;

[0011] Calculate the gray-level centroid or curve peak position along a direction perpendicular to the local direction of the differential fringes to obtain the sub-pixel center;

[0012] The sub-pixel center triangle is reconstructed to the common coordinate system using the common calibration parameters of the probes of each narrowband line laser camera;

[0013] Mass components with values ​​from zero to one are generated based on the saturation pixel ratio of the differential stripes, the maximum break length, the deviation of the stripe width from the calibration width, and the differential signal-to-noise ratio. The physical mass weight is determined based on the product or weighted geometric mean of the mass components.

[0014] Optionally, S3 includes: generating a geometric mass component with a value of zero to one based on at least two of the following: the chord length difference of the opposing view profile arc segment, the distance of the common support line of the adjacent view profile arc segment, and the initial ellipse fitting residual.

[0015] The geometric mass component is multiplied by the physical mass weight determined according to each mass component to obtain the mass weight with a value of zero to one;

[0016] Contour points whose mass weights are greater than the retention threshold determined by the mass weight distribution of non-round blank calibration data and qualified round blank calibration data are retained, and the mass weights corresponding to the same contour points are continued to be used in subsequent robust distance loss terms and cross-sectional validity determinations.

[0017] Furthermore, when the polar angle range continuously traversed by contour points whose quality weight is not greater than the point retention threshold in a view reaches the failure angle threshold determined by the nominal coverage angle of the view, the contour arc segment of the view in the current section is deactivated and the effective angle coverage and the number of opposing effective viewpoints are recalculated.

[0018] When the recalculated effective angular coverage reaches the angular coverage threshold and the number of opposing effective viewing angles reaches the viewing angle number threshold, continue to execute steps S4 to S6. When the recalculated effective angular coverage does not reach the angular coverage threshold or the number of opposing effective viewing angles does not reach the viewing angle number threshold, the measurement status of the current section is determined to be invalid.

[0019] Optionally, S4 includes: forming a time state sequence by assembling the center displacement states of each cross section within the same sliding window according to the acquisition time, and forming a shape state sequence by assembling the major semi-axis state, minor semi-axis state, and principal axis direction state of each cross section according to material coordinates;

[0020] Construct a joint objective function that includes a robust distance term from the contour points with the mass weights to the corresponding candidate ellipses, a temporal smoothing constraint term for the temporal state sequence, and a material coordinate smoothing constraint term for the shape state sequence;

[0021] Solve the joint objective function, and include the displacements of the contour points at each viewpoint that act together with the acquisition time into the central displacement state. Include the major semi-axis, minor semi-axis, and principal axis directions of the ellipse that change with the material coordinates into the shape state to obtain the center of the cross section and the shape parameters of the hot ellipse.

[0022] Optionally, S5 includes: transforming the installation azimuth angle of each circumferential temperature sensor to a polar coordinate system with the center of the cross section as the origin, and interpolating between valid circumferential temperature samples located on both sides of the contour point according to the polar angle of the contour point to obtain the temperature of the contour point, and determining a temperature confidence level with a value of zero to one according to the validity mark of the circumferential temperature sample used for interpolation, the angular distance between each circumferential temperature sample and the polar angle of the contour point, and the allowable error of the temperature sensor.

[0023] When there are fewer than two valid circumferential temperature samples for the profile point temperature interpolation, or when the azimuth interval between the valid circumferential temperature samples on both sides of the profile point exceeds the temperature interpolation angle threshold determined by the circumferential temperature sensor installation interval, the temperature confidence of the profile point is set to zero and the profile point is not included in the reference temperature quality weighted point set.

[0024] Based on the steel grade identification, the effective linear expansion coefficient with temperature variation within the range from the profile point temperature to the reference temperature is queried, and the radial coordinate of the profile point relative to the center of the cross section is converted point by point.

[0025] For each circumferential temperature sample, the predicted temperature is obtained by interpolating the two adjacent and valid circumferential temperature samples according to the installation azimuth angle. When there are fewer than two valid circumferential temperature samples for prediction, or when the difference between the measured temperature of the circumferential temperature sample and the predicted temperature exceeds the temperature difference threshold determined by the allowable error of the temperature sensor and the statistical value of the temperature gradient of the qualified round billet, the polar angle interval covered by the circumferential temperature sample is marked as a temperature anomaly interval, and the temperature confidence of the contour points in the temperature anomaly interval is limited to no greater than the upper limit of the temperature anomaly weight determined by the calibration data containing temperature occlusion interference.

[0026] The reference temperature mass weight is obtained by multiplying the temperature confidence level by the mass weight.

[0027] Optionally, S6 includes: representing the candidate ellipse using a parameterization method that guarantees the positive definiteness of the quadratic matrix, and solving for the candidate ellipse by weighting and summing the robust distances from the contour points to the candidate ellipse according to the reference temperature quality weight;

[0028] The ellipticity is determined based on the ratio of the difference between the major axis and the minor axis to the minor axis obtained from the solution, and the covariance of the fitting parameters is determined based on the local curvature of the weighted residuals.

[0029] When the effective angular coverage reaches the angular coverage threshold, the number of effective viewing angles reaches the viewing angle threshold, and the spectral radius of the covariance of the fitting parameters does not exceed the covariance threshold, the measurement state is determined to be reliable.

[0030] When one of the three conditions is not met but the other two are met, the measurement status is determined to be pending verification.

[0031] The measurement status is determined to be invalid if at least two conditions are not met.

[0032] The corner coverage threshold, the number of viewing angles threshold, and the covariance threshold are determined by the jointly calibrated qualified round blank verification data.

[0033] Furthermore, the upper limit of the allowable ellipticity is set according to the round billet acceptance standard corresponding to the target steel grade and specifications;

[0034] The sections with a reliable measurement status and an ellipticity exceeding the upper limit of the allowable ellipticity are sorted according to material coordinates. The sections to be reviewed, invalid sections, and sections with a reliable measurement status but an ellipticity not exceeding the upper limit of the allowable ellipticity are all taken as the breakpoints of the candidate over-limit sections, so that the reliable over-limit sections on both sides of the breakpoint are respectively assigned to different candidate over-limit sections.

[0035] Within each candidate over-limit section, when the material coordinate interval between adjacent over-limit sections is not greater than the connection interval determined by the sampling material coordinate spacing and the material coordinate length between the first and last over-limit sections reaches the alarm length determined by the minimum allowable defect length at the rolling inlet, the material coordinate interval between the first and last over-limit sections is determined as the alarm section.

[0036] Write the sections to be reviewed into the review queue, and write the invalid sections and the validity conditions they do not meet into the invalid record.

[0037] On the other hand, the present invention also provides a laser online measurement system for the ellipticity of continuously cast round billets, comprising:

[0038] The synchronous acquisition module is used to synchronously acquire laser-off background frames, laser-on contour frames, circumferential temperature data, and roller speed encoder data from multiple perspectives to form a synchronous cross-sectional data set. The contour reconstruction module is used to perform differential processing on paired frames, extract sub-pixel centers, reconstruct contour points in a common coordinate system, generate each mass component based on the differential fringe imaging quality, and determine the physical mass weights bound to the contour points based on each mass component. The mass propagation module is used to correct the physical mass weights based on the geometric consistency between contour arc segments from different perspectives, and use the obtained mass weights for contour point retention, robust distance weighting, and cross-sectional validity determination. The dual-domain state solution module is used to index the center displacement state by acquisition time, the major semi-axis state, the minor semi-axis state, and the principal axis direction state by material coordinates, and jointly solve the center displacement state and each shape state. The temperature registration compensation module is used to register the circumferential temperature data to the contour points by polar angle based on the obtained cross-sectional center and convert it to the reference temperature. The ellipse fitting and alarm module is used to perform robust ellipse fitting with positive definite constraints, determine the measurement state, and form alarm segments for continuous over-limit along the material coordinates.

[0039] The beneficial effects of this invention are:

[0040] 1. By synchronously acquiring data through paired multi-view laser switches and reconstructing differential fringes into sub-pixel images, the influence of thermal radiation background on contour extraction can be reduced, and multiple contour arcs in a common coordinate system can be used to supplement the insufficient information in unidirectional diameter measurement.

[0041] 2. By combining the physical quality weights formed by imaging quality with cross-view geometric consistency correction, and continuously using the resulting quality weights for point retention, robust distance weighting, and cross-section validity determination, the influence of saturation, fracture, vapor obstruction, and oxide scale abnormal contours on ellipse results can be limited.

[0042] 3. By jointly solving the time domain center displacement state and the material coordinate domain shape state, as well as the polar angle temperature registration based on the cross-section center, point-by-point reference temperature conversion and point set sufficiency gating, it is possible to separate the common mode runout, the true elliptical shape and the influence of thermal expansion, and form an alarm basis for reliable and continuously exceeding material limits. Attached Figure Description

[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0044] Figure 1 This is a flowchart of the laser online measurement method for the ellipticity of continuously cast round billets according to the present invention.

[0045] Figure 2 This is a flowchart of the joint solution of the S4 dual-domain states in this invention. Detailed Implementation

[0046] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0047] refer to Figures 1-2A method for online laser measurement of the ellipticity of continuously cast round billets includes: S1, simultaneously acquiring multi-view laser-closed background frames, laser-open contour frames, circumferential temperature data, and roll speed encoder data for the same nominal cross-section of the continuously cast round billet; S2, performing differential processing on the background frames and contour frames and reconstructing contour points in a common coordinate system, and determining physical quality weights bound to the contour points based on the imaging quality of the differential fringes; S3, correcting the physical quality weights based on the geometric consistency of contour arc segments from different viewpoints, and ensuring that the corrected quality weights are applied throughout the contour point selection, robust solution, and cross-sectional validity determination; S4, establishing material coordinates based on the roll speed encoder data to collect... The time-indexed center displacement state and the material coordinate-indexed elliptical shape state are used to solve the two states jointly based on the weighted robust residuals of the multi-view contour points, resulting in the cross-sectional center and the thermal elliptical shape after separating the common mode runout; S5, the polar angle of the contour point is determined according to the cross-sectional center, the circumferential temperature data is registered to the contour point according to the polar angle, and the contour point is converted to the reference temperature according to the linear expansion function; S6, the contour points converted to the reference temperature are subjected to robust elliptical fitting with mass weights, the measurement state is determined according to the sufficiency of the point set and the fitting uncertainty, and an alarm segment is output when the measurement state is reliable, the ellipticity exceeds the upper limit of the allowable ellipticity and is continuous along the material coordinates.

[0048] In this specific embodiment, S1 includes:

[0049] In this specific embodiment, a measurement section without direct water mist scouring is selected between the end of the secondary cooling section of the continuous casting machine and the entrance of the straightening section. Four sets of narrow-band line laser camera probes are arranged circumferentially around the outer periphery of the roller conveyor. Each set of probes includes a 660nm wavelength line laser, a 10nm bandwidth narrow-band filter, and a global shutter camera. The nominal coverage arcs of the four sets of probes overlap each other and all point to the same nominal cross section. The acquisition controller writes the probe number, installation azimuth angle, camera intrinsic parameters, laser plane parameters, and exposure parameters for each set of probes.

[0050] During common calibration, a heat-resistant calibration component with circumferential feature points and axial engravings is placed on the measurement section. The rigid body transformation from each probe coordinate system to the common coordinate system with the roller centerline as the vertical axis is calculated sequentially. The root mean square distance of the three-dimensional feature points with the same name in the overlapping area of ​​adjacent probes is no greater than 0.15mm as the calibration pass condition. The calibration results are stored in the common calibration parameter table with the probe number and calibration version as the joint key. The table records the rotation matrix, translation vector, laser plane coefficient, validity date, and cumulative running time of recalibration trigger.

[0051] The unified trigger outputs a trigger sequence every 10ms for a cross-sectional acquisition cycle. In the first half of the cycle, the four-line lasers are turned off and synchronously exposed to obtain the laser-off background frame. In the second half, the line lasers are turned on and synchronously exposed with the same exposure time and analog gain to obtain the laser-on contour frame. The interval between the two exposures is set to 2ms. The acquisition controller records the image using the cross-sectional group number, probe number, frame type, hardware timestamp, exposure time, and gain. Only paired frames with the same group number and a maximum hardware timestamp difference of no more than 50μs are assigned to the same cross-sectional acquisition cycle.

[0052] The circumferential temperature data is collected by eight narrow-band infrared temperature sensors arranged around the measurement section. Each sample records the sensor number, installation azimuth angle, measured temperature, allowable error, validity mark and sampling timestamp. The roller speed encoder latches the cumulative pulse value when the same unified trigger signal arrives with a 1MHz counter. The acquisition controller writes the temperature sample closest to the center of the section acquisition cycle and with a time difference of no more than 5ms, as well as the encoder count latched in the same cycle, into the synchronous buffer corresponding to the image record.

[0053] Material coordinate increments according to calculate, Indicates the first The increment of material coordinates for each cross-section acquisition cycle relative to the previous cycle is expressed in mm. This parameter represents the unit pulse displacement of the roller speed encoder. After being calibrated using a standard displacement scale, it is entered into the encoder parameter table, and the unit is mm / pulse. This parameter is only used for linear displacement conversion and does not involve torque or force value conversion. and They represent the first The cumulative pulse count latched in the previous cycle and the current cycle, with units of pulses and taken as a non-negative integer allowed by the encoder's counting bit width. This represents a periodic index that increments by the collection time from the beginning, traversing all collection periods of the current casting. The index itself is a dimensionless integer and no unit is applicable.

[0054] The data acquisition controller performs direction and jump checks on the material coordinate increment. When the pulse count rewinds, compensation is made according to the encoder count bit width. The system saves the cumulative pulse count, material coordinates, timestamp, and abnormal start group number of the last valid cycle. When the material coordinate increment is opposite to the given direction of the roller speed or its absolute value exceeds three times the median of the increments of the fifty consecutive qualified cycles at the same roller speed, the current cycle is marked as an encoder abnormality and the formation of a solvable section is stopped. When subsequent cycles return to normal, the total number of pulses across the abnormal segment is obtained by subtracting the last valid cumulative pulse count from the cumulative pulse count of the recovery cycle. After rewind compensation, the total displacement across the abnormal segment is calculated and expressed in the direction of roller speed and abnormality. The average roller speed integral value within the continuous time and the upper limit of three times the normal sampling interval are checked for consistency. If the check passes, the total displacement across the abnormal section is added to the last valid material coordinate to obtain the material coordinate of the restored section. At the same time, the abnormal span is marked as a material gap and the center smoothing and shape smoothing items on both sides of the gap in step S4 are cut off. If the check fails, the material coordinate is re-anchored with the current reading of the production line entrance length counter. The section from the start of the abnormality to the recovery period is written into the unified format invalid record and the check queue, and a breakpoint event is sent to the alarm section generator. If no abnormality is triggered, the previous valid material coordinate is directly added to the material coordinate increment of the current cycle to obtain the current material coordinate.

[0055] Before the equipment is put into operation, the synchronization threshold is calibrated by running a unified trigger continuously for 10,000 cycles of no-load recording. The 99.9th percentile of the hardware timestamp range of the four cameras plus 10 microseconds is taken as the camera synchronization threshold. Half of the upper limit of the sampling interval of the temperature sensor is taken as the temperature correlation threshold. The average unit pulse displacement obtained by the standard displacement platform running back and forth at three roller speeds is used as the encoder conversion parameter. The threshold table records the field name, unit, number of calibration samples, statistical percentile, applicable roller speed, version and update time. If any threshold version does not match the common calibration version, the synchronization validity flag is set to false.

[0056] The final synchronized cross-section data set uses the cross-section group number as the primary key and includes file offsets for four background frames and contour frames, eight circumferential temperature samples, encoder counts, material coordinates, common calibration version, steel grade identification, billet specifications, and synchronized validity flags. The data set is written by appending only to the ring buffer and passing the buffer address to step S2. When any probe pairing frame is missing, the data set is still retained, but the corresponding probe is marked as missing a frame, so that subsequent steps are determined by the number of valid viewing angles instead of replacing it with a blank image.

[0057] In this specific embodiment, S2 includes:

[0058] The contour reconstruction module first reads the synchronization validity flag of the synchronization cross-section data group. When the synchronization validity flag is false, steps S2 to S6 are blocked, the paired frames are not read or processed, the current cross-section measurement status is determined to be invalid, the timestamp range, temperature correlation time difference, threshold version, common calibration version, original synchronization cross-section data group index and failure reason are written, and a breakpoint event is sent to the alarm section generator. When the synchronization validity flag is true, the contour reconstruction module reads the paired frames of each valid probe. In the common valid pixel area, the gray median and interquartile range of unsaturated background pixels and contour pixels are statistically analyzed according to a 32×32 pixel grid. The slope and intercept of the robust linear regression of the gray values ​​of the background frame to the gray values ​​of the non-striped area of ​​the contour frame are used as the gain difference and the bias difference, respectively. When the effective regression samples are less than 60% of the number of pixels per grid, the gain and bias of the most recently calibrated probe are used and the parameter backoff flag is written. After the effective regression samples of five consecutive cycles recover to more than 60%, the backoff flag is cleared and online estimation is restored.

[0059] The gain bias parameter table uses probe number, exposure time and analog gain level as joint keys to store robust regression slope, intercept, median grayscale residual, effective grayscale range, calibration version and update time. The parameters in the table are jointly fitted by the laser-free blank background sequence and the uniform reflection calibration plate sequence. The online estimated value is only written into the current cycle record when the slope is between 0.8 and 1.2 and the absolute value of the intercept does not exceed 20 grayscale levels. When it exceeds the range, the aforementioned parameter rollback is triggered.

[0060] Paired frame differential values ​​according to calculate, Indicates the first Each probe at pixel coordinates The difference in grayscale at a given location, expressed in grayscale levels. This indicates the grayscale value of the laser-activated contour frame at the same pixel, expressed in grayscale levels. This indicates the grayscale value of the background frame at the same pixel when the laser is turned off, expressed in grayscale levels. Indicates the first Dimensionless gain matching coefficients for each probe-paired frame This indicates the corresponding bias matching value, expressed in gray levels. Represent the head index and traverse the set , Represents an integer column index from zero to the image width minus one. Represents an integer row index from zero to the image height minus one. , and All are dimensionless integer indices and units are not applicable;

[0061] After differential processing, firstly, using three times the width of the calibrated stripe as the structural element, connect the discontinuities shorter than the length along the stripe extension direction. Then, determine the stripe tangent by the main direction of the local gray-level gradient and take its perpendicular direction as the normal. On each scanning normal, select a continuous pixel segment whose differential gray-level exceeds the median of the grid noise plus six times the absolute median difference. When the continuous segment is empty, the scanning position is recorded as the break position. When there is more than one continuous segment, retain the pixel segment with the largest integrated gray-level.

[0062] Sub-pixel center according to calculate, Indicates the first The first probe Subpixel center coordinates on the scan normal, in pixels. The first line on the normal line The center coordinates of candidate pixels, in pixels. This represents the differential gray level of the candidate pixel, expressed in gray levels. This represents the differential threshold determined by the corresponding grid noise statistics, expressed in gray levels. Indicates in The truncation operation takes the larger value between zero and zero, and the unit varies. , The pixel half-width represents the number of dimensionless pixels, and in this specific embodiment, it is taken as four. Indicates traversal Dimensionless candidate pixel indices containing all integers. Represents a dimensionless integer index from one to the current number of probe scan normals. Using the probe index set The units in the above indexes are not applicable as they are only used for positioning;

[0063] The sub-pixel center is back-projected using the camera intrinsic parameters to the line of sight in a common coordinate system, and its intersection with the laser plane obtained from common calibration is determined. The intersection point is calculated according to... Sure, This represents a contour point in a common coordinate system, with coordinate units in mm. Indicates the first The position vectors of the optical centers of each camera in a common coordinate system, with units of mm. Indicates the first The dimensionless unit normal vector of a laser plane This indicates the plane offset of the laser plane, and the unit is mm. This represents the dimensionless unit line-of-sight vector in the common coordinate system determined by the sub-pixel centers. This indicates a vector transpose operation where units are not applicable. and Use the dimensionless index and traversal range defined in the previous section;

[0064] When the absolute value of the normal inlay product of the line of sight and the laser plane is less than the lower limit of the intersection angle in the common calibration parameter table, the intersection point is deleted. Otherwise, for each contour point, the saturation pixel ratio, continuous break length, relative deviation of the half-height fringe width relative to the calibration width, and differential signal-to-noise ratio within its seven-pixel normal window are calculated. The saturation ratio is linearly attenuated from zero to twenty percent, the break length is linearly attenuated from zero to three times the calibration fringe width, the absolute value of the relative deviation of the width is linearly attenuated from zero to 0.5, and the differential signal-to-noise ratio is linearly increased from six to twenty. The results are then truncated to zero to one to obtain the saturation mass component, break mass component, width mass component, and signal-to-noise ratio mass component.

[0065] Physical quality weights according to Sure, Indicates the relationship with the first The first probe The physical quality weights bound to each contour point are dimensionless numbers ranging from zero to one. A dimensionless number representing the saturated mass component, ranging from zero to one. A dimensionless number representing the fracture mass component, ranging from zero to one. A dimensionless number representing the width mass component, ranging from zero to one. denoted as a dimensionless number between zero and one, representing the differential signal-to-noise ratio mass component. The physical mass weight is zero when any mass component is zero.

[0066] The quality component thresholds are established by four types of calibration sequences: no round billet, qualified round billet, water mist interference, and oxide scale interference. The calibration program statistically analyzes the empirical distribution of saturation ratio, fracture length, width deviation, and differential signal-to-noise ratio for each type of sequence. The endpoints of each linear mapping are determined by the 5th percentile of qualified round billet and the 95th percentile of interference data. The quality parameter table saves the endpoints, units, sample batches, calibration versions, and recalibration trigger conditions according to the probe number and exposure level. When the online index falls outside the endpoint, only the zero to one truncation is performed without extrapolation.

[0067] Step S2 writes each valid intersection point, along with the probe number, scan normal index, common coordinates, acquisition time, material coordinates, four mass components, physical mass weights, and calibration version, into the weighted contour point record. It then forms a multi-view weighted contour point set according to the section group number and passes it to step S3. For probes with missing frames, insufficient intersection angles, or continuous breaks exceeding 40% of the probe's nominal coverage arc length, it writes the reason for the view anomaly but does not falsify the contour points.

[0068] In this specific embodiment, S3 includes:

[0069] Step S3 reads the multi-view weighted contour point set. First, sort the points in each view according to the scan normal index and divide the contour arc segment with the condition that the distance between adjacent points does not exceed three times the distance between calibration points. Then, form opposing view pairs with probes whose azimuth angles differ by 180 degrees within the allowable error range in the common calibration parameter table, and form adjacent view pairs with probes whose nominal coverage arc segments overlap. The arc segment record saves the endpoints, chord length, common support line, effective polar angle interval, and source probe.

[0070] The dimensionless chord length inconsistency is obtained by dividing the chord length difference of opposing viewpoints by the mean of the chord lengths of the two arc segments and truncating to zero or one. The dimensionless support line inconsistency is also obtained by dividing the distance of the common support line between adjacent viewpoints in the overlapping area by the 95th percentile of that distance in the common calibration verification data and truncating to zero or one. Simultaneously, an initial ellipse fitting without geometric correction is performed on contour points with a physical quality weight greater than 0.2. The initial fitting requires retaining at least twenty points, an effective polar angle coverage of at least 0.5, a normalized design matrix condition number no greater than 10^8, and ensuring that the relative change of the target is less than 10^-6 within twenty iterations. The fitting must satisfy all conditions and converge. When the orthogonal distance from each point to the initial ellipse is divided by the nominal radius of the blank and truncated to zero or one, the dimensionless initial residual is obtained. If any condition is not met or the solution fails, the initial ellipse parameters are not output, and the initial residual of all contour points participating in the initial fitting is uniformly set to one. The current section is still screened based on the inconsistency of chord length, the inconsistency of support line, and the physical quality weight, and the failure reason, number of points, effective angle coverage, number of conditions, and number of iterations are saved in the geometric quality record. If less than ten points are retained after rollback, the current section is directly written into the unified format invalid record and a breakpoint event is sent to the alarm section generator, and step S4 is not entered again.

[0071] Geometric mass components according to generate, Indicates the first The first probe The geometric mass components of each contour point are dimensionless numbers ranging from zero to one. This indicates the inconsistency in chord length between opposite arc segments to which the point belongs. This indicates the inconsistency in the support lines of the adjacent arc segments to which the point belongs. This represents the initial elliptic residual at that point. It represents a natural exponential function, and when a certain opposite or overlapping arc segment is missing, the corresponding inconsistent quantity is set to one instead of being replaced by zero;

[0072] The final quality weight of the contour points is based on Sure, This represents the dimensionless mass weight, ranging from zero to one, output in step S3 and maintained in subsequent steps, that is bound to the same contour point. This represents the physical mass weight obtained in step S2. The geometric quality components obtained in this step are multiplied so that differential imaging defects and cross-view geometric conflicts can only be reduced, but not improved, in terms of the reliability of contour points.

[0073] The retention threshold is established by jointly calibrated data of blankless and qualified blanks. The calibration program performs the same difference, reconstruction and geometric correction on the two types of data as in the online process, forming weighted empirical distributions respectively. The intersection of the 99th percentile of the blankless distribution and the 5th percentile of the qualified blank distribution with equal error rates is taken as the threshold. When the two distributions have no intersection, the average of the two percentiles is taken. In this specific embodiment, the calibration yields 0.28. The threshold, along with the calibration version, applicable steel grade and specification, sample batch and update time, is written into the quality threshold table.

[0074] The candidate set for point selection consists of contour points where all common coordinates have been reconstructed and intersection angles are valid in the current section. The selection predicate is that the final quality weight is strictly greater than the point retention threshold. Points that satisfy the predicate are written into the solvable point set in ascending order of probe number and polar angle. Points that do not satisfy the predicate are written into the deletion list. When the thresholds are equal, they are not retained. When there are multiple overlapping viewpoints at the same polar angle, all of them are retained and continue to participate in robust solution according to their respective quality weights without pre-selecting one.

[0075] The quality threshold table and geometric consistency parameters are constructed from the same batch of jointly calibrated data. Records the weight distribution of blanks without round billets, the weight distribution of qualified round billets, the chord length reference, the support line distance reference, the initial residual reference, the applicable specifications and versions. The entire table is updated when the equipment rolls are changed, any probe is reinstalled, or the joint calibration residual of the overlap area exceeds 0.2 mm. The old version is only used for traceability and must not be mixed with the new version of the contour points.

[0076] Only contour points with a quality weight greater than the point retention threshold are retained in the solveable point set. Deleted points still retain their quality weight and deletion reason in the original record. When calculating the robust distance loss in subsequent steps, the same quality weight is directly used as the multiplier. When calculating the effective angular coverage and the number of effective viewpoints, only points that pass the point retention threshold are counted. It is forbidden to reset the quality weight of retained points to one before subsequent fitting.

[0077] When points within a certain viewpoint that are lower than or equal to the point retention threshold form a continuous polar angle interval, step S3 records the ratio of the start and end angles, span, and nominal coverage angle of the interval, and passes the list of continuous low-quality intervals for each viewpoint along with the solveable point set to steps S4 and S6, and the subsequent steps perform viewpoint deactivation and validity determination according to the failure angle threshold.

[0078] The quality propagation record output in step S3 uses the cross section group number and contour point identifier as a joint key. It includes physical quality weight, three geometric inconsistencies, geometric quality components, final quality weight, point retention status, low quality range of viewpoint, and quality threshold version. This record, along with the solvable point set, is written into the status window cache for continuous reading in step S4 joint solution, step S5 reference temperature weight calculation, and step S6 final fitting.

[0079] In this specific embodiment, S4 includes:

[0080] Step S4 defines the reading of the production line inlet length counter at the first effective section in this casting as the initial material coordinate. The unit is mm; for the first Each effective section according to Update material coordinates. and These represent the material coordinates of the current and previous effective cross sections, respectively, in mm. This represents the current material coordinate increment calculated in step S1 and verified for anomalies, in mm. This represents a dimensionless integer index that iterates through all effective sections of the current casting from the beginning, and the unit is not applicable; when the anomaly described in step S1 occurs and re-anchoring is performed, a new index is formed based on the inlet length counter reading. All smoothing items are cut off at the anchor point, and the window records and retains three indices: acquisition time, material coordinates, and section group number.

[0081] When the status window is running normally, it reads the current section and the fifteen synchronized sections before and after it. The current section is finalized after obtaining the next fifteen valid sections, with a waiting limit of thirty normal acquisition cycles. The first fifteen sections at the start end are formed by using all the obtained preceding sections and fifteen subsequent sections to form a shortened window. After the stop signal arrives, the last fifteen sections are immediately finished with all the remaining subsequent sections. The minimum number of solvable sections in the shortened window is sixteen, and the same pairwise scale normalization and smoothing weights are used. Points cannot be filled across the gap on both sides of the synchronization gap. If there are insufficient records on either side, a single-sided shortened window is used. If the total number of windows is still less than sixteen or the waiting timeout occurs, the current section is directly written into a unified format invalid record and a breakpoint event is sent to the alarm section generator. The boundary state of the shortened window is used as the initial value of the overlapping solution of the previous complete window. If there is no overlapping solution, the initial value is fitted by the weighted algebraic ellipse of the current section.

[0082] For each viewpoint, the continuous low-quality polar angle span recorded in step S3 is divided by the nominal coverage angle of the viewpoint. When the ratio reaches the failure angle threshold, the contour arc segment of the current section of the viewpoint is deactivated. The failure angle threshold is determined by the nominal coverage angle of the viewpoint and the common calibration verification data, and is set to 40% in this specific embodiment. After deactivation, the effective angle coverage is recalculated according to the union of the polar angle intervals of the retained points, and the number of effective viewpoints in opposite directions is calculated according to the rule that there are effective arc segments within the allowable error range of 180 degrees difference in azimuth angle.

[0083] The candidate set for view selection consists of all views with at least ten reconstructed points in the current section. First, invalid views are deleted according to the predicate of continuous low-quality polar angle span. Then, the remaining view list is formed by sorting the starting angle of the effective polar angle interval. If multiple views have the same starting angle, they are sorted in descending order according to the sum of the quality weights of the retained points. The selection results include retained views, deactivated views, reasons for deletion, and coverage before and after recalculation. These are used as input for joint solution instead of supplementing points from the original view set.

[0084] If, after recalculation, the effective angular coverage is lower than the angular coverage threshold used in step S6 or the number of effective viewing angles in the opposite direction is lower than the number of viewing angles threshold, step S4 will not start the joint solution. Instead, the current section measurement status will be directly determined as invalid. An invalid record will be written in the unified format of step S6, recording the coverage, the number of effective viewing angles in the opposite direction, the disabled viewing angle, the reason for deletion, and the original synchronization data index. A breakpoint event will be sent to the alarm section generator. Steps S5 and S6 will no longer process this section. If both conditions are met, the remaining viewing angles will continue to be processed. The disabled viewing angle status will only apply to the current section and will be saved together with the original quality propagation record. The next section will be re-determined based on its continuous low-quality interval.

[0085] Within the sliding window, the center displacement state is indexed as a two-dimensional vector based on the acquisition time, and the shape state is indexed as a material coordinate. , Indicates the first The thermal elliptical shape of each cross-section is represented by components in mm, mm, and rad, respectively. The units of the mixed-dimensional vector itself are not applicable; it is normalized component-by-component by the scaling matrix described later before entering the norm. This represents the semi-major axis of the thermal ellipse, with the unit being mm. This represents the minor semi-axis of the thermal ellipse, with the unit being mm. Indicates the direction of the principal axis and the unit is rad. This indicates a transpose operation where units are not applicable. The dimensionless integer index in the current window section set is used; the shape state is arranged in ascending order according to the material coordinates while the center displacement state is kept in the order of acquisition time, so as to separate the modeling of the roller common mode runout and the cross-sectional shape moving with the material.

[0086] The principal axis direction is restricted to a left-closed, right-open interval from zero to π radians. For each new section, the candidate principal axis direction is added to or subtracted by an integer multiple of π, and the equivalent representation closest to the principal axis direction of the previous effective section is selected, so that the difference between adjacent directions entering the shape smoothing term falls into a left-closed, right-open interval from negative half of π to positive half of π. When the relative difference between the major and minor semi-axles does not exceed the near-circular threshold, the principal axis direction smoothing weight is reset to zero and written into the direction uncertainty flag. The near-circular threshold is determined by the relative difference between the major and minor semi-axles when the direction variance begins to increase sharply in the repeated fitting data of the qualified round billet, and is written into the shape smoothing parameter table along with the sample batch and version.

[0087] Contour point normalized elliptical residuals according to calculate, Indicates the first The first section The dimensionless radial residuals of the preserved contour points relative to the candidate hot ellipse, and This indicates that the contour point is subtracted from the first... The principal axis coordinates are obtained by rotating the axis along the principal axis direction after determining the center displacement state, and the unit is mm. , and Consistent with the same type of quantity in the shape state; let Let be the set of section indices in the current full or shortened window. For the first The set of all preserved contour point indices for a given cross section, then and , and All are dimensionless integer indices used only for traversal and units are not applicable;

[0088] robust loss adopts , This represents the dimensionless Huber loss. This represents the dimensionless residual of the input. This represents the dimensionless turning point threshold determined by the 95th percentile of the residual jointly calibrated by qualified round blanks. In this specific embodiment, it is taken as 0.008. Contour points whose absolute value exceeds the turning point threshold still participate in the solution, but their influence increases linearly.

[0089] The data residuals of the joint target are calculated according to calculate, Represents the residual term of dimensionless data. and Following the previous definition, Indicates step S3 and the first The dimensionless quality weights bound to each contour point This indicates the aforementioned dimensionless robust loss. and Traverse each and Both are dimensionless integer indices and units are not applicable.

[0090] Center smoothing term according to calculate, Represents the dimensionless central smoothing term. Indicates the index of the first section in the current window. Traversal The set of window sections other than and All are dimensionless integer indices and units are not applicable. and These represent the two-dimensional center displacement states at adjacent acquisition times, with the unit of each component being mm. This represents the 2x2 central scaling matrix, as will be discussed later. This indicates L2 norm operations where units are not applicable.

[0091] Shape smoothing item according to calculate, Represents the dimensionless shape smoothness term. and The set range and dimensionless index attribute are defined as in the previous section. and This represents the mixed-dimensional shape state at coordinates of adjacent materials; the unit of the vector itself is not applicable. The term represents a 3x3 shape-scale matrix, and the 2-norm operation follows the definition from the previous section.

[0092] Joint objectives according to Build, This represents the dimensionless joint objective value to be minimized. , and These represent the aforementioned dimensionless data residual term, center smoothing term, and shape smoothing term, respectively. and These represent the dimensionless center smoothing weight and shape smoothing weight, respectively, and in this specific embodiment, they are taken as 0.2 and 0.1. This indicates a dimensionless scalar multiplication operation where units are not applicable.

[0093] The central scale matrix is ​​defined as , It is a 2x2 diagonal matrix with both diagonal elements having units of millimeters. and The shape scale matrix is ​​obtained by multiplying the allowable lateral and longitudinal velocities by the time interval between adjacent acquisitions, with units in mm; the shape scale matrix is ​​defined as follows: , It is a 3x3 diagonal matrix, where the units of its three diagonal elements are per millimeter, per millimeter, and per radian, respectively. , and The values ​​are obtained by multiplying the allowable change rate of the major semi-axis, the allowable change rate of the minor semi-axis, and the allowable change rate of the principal axis direction by the distance between the coordinates of adjacent materials, with units of mm, mm, and rad, respectively. This indicates that the diagonal matrix construction operation is not applicable and each allowable rate of change is determined by the 99th percentile of the qualified round billet sliding window verification data and the unit, specification, window length and version are recorded. When any scale is zero, the resolution of the same unit is used as the lower limit, so that the input components of the two norms are dimensionless.

[0094] The solver uses a weighted algebraic ellipse fitting of each section retention point as the initial shape value and the mean of the midpoints of two adjacent opposing view arc segments as the initial center value. It uses the iterative reweighted least squares method with a trust region to solve the joint objective function and alternately updates the center displacement state and shape state. It constrains the major semi-axis to be no less than the minor semi-axis, the minor semi-axis to be between 80% and 120% of the nominal radius, and the single principal axis direction update to be no more than 0.1 rad. The solver terminates when the relative change of the target is less than 10 to the power of negative 6 or when it reaches 30 iterations.

[0095] After the window is solved, the current cross section is moved forward by one cross section group number and the overlapping state of the previous window is retained as the initial value of the next window. If the material coordinates of the newly entered window are more than three times the normal sampling interval of the previous cross section, the shape smoothing term is cut off at that position. If the sampling time interval is more than three times the normal cycle, the center smoothing term is cut off so that the material gap or the trigger interruption will not be crossed by the smoothing constraint.

[0096] When the solution converges, the translation of the contour points of each viewpoint acting on the same cross section together with the acquisition time is written into the center displacement state, and the major semi-axis, minor semi-axis and principal axis directions that change with the material coordinates are written into the hot shape state. The solution residual, iteration number, list of disabled viewpoints and quality weight summary are associated with the cross section group number. The cross section center and hot elliptical shape of the current cross section are passed to step S5. When the solution fails to converge or there is no feasible solution due to constraints, the current cross section measurement state is directly determined as invalid, and an invalid record is written in the unified format of step S6. The solution diagnosis, original window and original synchronization data index are recorded, and a breakpoint event is sent to the alarm section generator. Steps S5 and S6 no longer process the cross section.

[0097] In this specific embodiment, S5 includes:

[0098] Step S5 takes the center of the cross section obtained in step S4 as the common polar coordinate origin, maps the installation position of each circumferential temperature sensor to the measurement plane through a common calibration rigid body transformation, and then determines the sensor azimuth angle by the direction of the installation position relative to the center of the cross section. The polar angle of the profile point is determined by subtracting the direction of the cross section center from the retained profile point. All azimuth angles are expanded in the range of zero to twice the circumferential angle and sorted in the counterclockwise direction.

[0099] The temperature registration module uses all circumferential temperature samples with true validity flags for the current cross section as the candidate set, establishes a cyclic index in ascending order of unfolded azimuth angle, selects the sample with the smallest angular distance for each contour point in both clockwise and counterclockwise directions, and selects the sample with the sampling timestamp closer to the center of the cross section acquisition cycle when the angular distances are the same, and records the sample number, angular distance, time difference and whether it crosses zero degrees on both sides of the selection result as the only input for interpolation and confidence calculation;

[0100] For any contour point, only when there are nearest circumferential temperature samples with different numbers and true validity indicators on both sides of its polar angle, and the angle unfolds to satisfy the following condition: Only then, according to Calculate the temperature at the contour points; Indicates the first Temperature of each contour point, in degrees Celsius. This represents the polar angle of the unfolded point of the profile, expressed in rad. and These represent the effective temperature sample unfolding azimuth angles located on the clockwise and counterclockwise sides of the contour point, respectively, in rad. and These represent the measured temperatures of the valid samples on both sides, in degrees Celsius. This represents a dimensionless integer index that iterates through all retained contour points of the current section, and the unit is not applicable; if there are insufficient samples on both sides, the numbers are the same, or the positive angle interval is not met, this expression is blocked before performing division and is not calculated. ,Will Set it to zero and record the corresponding reason;

[0101] Temperature confidence level according to calculate, Indicates the first The temperature confidence level of each contour point is a dimensionless number between zero and one. and These represent the validity indicators of the temperature samples on both sides, and are either zero or one. and These represent the angular distances from the polar angle of the contour point to the azimuth angles of the samples on both sides, in rad. This represents the temperature interpolation angle threshold determined by the sensor installation interval, and the unit is rad. The temperature difference threshold is defined by the sum of the sensor's permissible error and the 95th percentile of the circumferential temperature gradient of a qualified round billet, and the unit is degrees Celsius.

[0102] When there are fewer than two valid circumferential temperature samples for interpolation, or when the azimuth interval between the valid samples on both sides exceeds the temperature interpolation angle threshold, the temperature confidence of the corresponding profile point is set to zero and marked as not entering the reference temperature quality weighted point set. At the same time, its geometric point, sample numbers on both sides and the reason for not entering are retained for verification. When the azimuth angle crosses zero degrees, the azimuth angle on the counterclockwise side is increased by twice the circumferential angle before the same interpolation rule is applied.

[0103] The temperature registration module determines a unique polar angle coverage interval for each circumferential temperature sensor according to the cyclical order of the installation azimuth angles: the left boundary of the interval is taken as the midpoint of the circumference of the installation azimuth angle of the sensor and the clockwise adjacent sensor, and the right boundary is taken as the midpoint of the circumference of the installation azimuth angle of the sensor and the counterclockwise adjacent sensor, and the left-closed and right-open rule is uniformly adopted; the right boundary that crosses zero degrees is doubled in circumference angle before comparison; when there is a commonly calibrated fixed field of view interval of the sensor, the intersection of the above midpoint interval and the fixed field of view interval is taken; the sensor result with lower temperature confidence is used at the intersection and overlap; the failed sensor shall not cause the coverage interval of the adjacent sensor to extend beyond its fixed field of view, and the uncovered interval formed by failure is directly marked as the temperature abnormal interval; each section saves the left and right boundaries of each interval, the zero-crossing mark, the overlap handling result and the coverage parameter version;

[0104] For each circumferential temperature sample, the temperature registration module takes the two adjacent effective temperatures and their installation azimuth angles collected in real time at the same cross section as runtime inputs, and the sensor installation order and allowable error as static inputs. It obtains the predicted temperature of the sample position according to the aforementioned deterministic angular linear interpolation function. This predicted temperature is not output by the learning model. When there are fewer than two adjacent effective samples, the unique polar angle interval of the sample is directly marked as a temperature anomaly interval. When the absolute value of the difference between the measured temperature and the predicted temperature exceeds the temperature difference threshold, it is also marked as an anomaly interval. The temperature confidence of the contour points within the interval is limited to no more than the upper limit of the temperature anomaly weight. In this specific embodiment, this upper limit is determined to be 0.35 by the quantile value of the reliable sample of the calibration data containing water mist obstruction.

[0105] Linear expansion relationship is queried by steel grade identification in the linear expansion parameter table. The table uses steel grade identification and temperature segment as keys, and records the start and end temperatures of the segment, linear expansion coefficient, coefficient unit, fitting residual, version and validity date. The coefficient is calibrated by the same batch of steel grade samples on the high temperature dilatometer at 50-degree intervals and fitted with a piecewise linear function. A valid record in the table includes target steel grade A, temperature segment 800 to 900 degrees Celsius, linear expansion coefficient of 1.4 x 10 to the power of -5 per degree Celsius, and version number. If there is no matching record for the steel grade or the temperature of the profile point exceeds the valid temperature range, the confidence level of the temperature at that point is set to zero.

[0106] The contour points are converted to a reference temperature based on the linear expansion function, and the radius of the contour points at the reference temperature is determined according to... calculate, Indicates the first The radial coordinates of each contour point relative to the center of the cross section at the reference temperature, in mm. This represents the thermal radial coordinate of the point, in mm. This indicates a reference temperature, and in this specific embodiment, it is taken as 20 degrees Celsius. This represents the temperature of the interpolated profile points, expressed in degrees Celsius. Indicates steel type At temperature The coefficient of linear expansion at a given location, expressed in degrees Celsius. This indicates that a discrete identifier is selected from the steel grade identifier set in the current valid linear expansion parameter table, and the unit is not applicable. This formula is not executed if no completely matching identifier is found. Indicates in to The temperature integral variable traversed within a closed interval, with units in degrees Celsius. Using the dimensionless index of the current section retaining the contour point set, the integration result is the dimensionless thermal expansion;

[0107] After conversion, keeping the polar angle of the profile point unchanged, the reference temperature profile point is obtained by taking the radial vector of the cross-section center plus the reference temperature radius in the direction of that polar angle, and then... Generate reference temperature mass weights. Indicates the first The mass weights of the reference temperature profile points are dimensionless numbers ranging from zero to one. This represents the quality weight associated with the same contour point in step S3. This indicates the confidence level of the temperature at that point. When the product is zero, the point is marked as not to be included in the final fit. When the product is greater than zero, the point is written into the reference temperature point set along with the weight.

[0108] Step S5 writes the coordinates of the reference temperature profile points, the polar angle of the profile points, the interpolated temperature, the temperature confidence level, the reference temperature quality weight, the version of the linear expansion table, and the temperature anomaly interval into the reference temperature point set. The point set maintains a one-to-one index relationship with the original weighted profile points using the section group number and is passed to step S6, so that the geometric quality and temperature quality are simultaneously preserved in the final ellipse fitting and validity determination.

[0109] In this specific embodiment, S6 includes:

[0110] Step S6 reads the reference temperature point set, using the candidate ellipse center, logarithmic parameters of the major and minor semi-axes, and principal axis directions as fitting variables; and takes the nominal radius from the common calibration parameter table as the reference length. and Both are in mm and are greater than zero; a dimensionless parameter is defined. and ,in and These represent the major and minor semi-axes of the candidate ellipse, respectively, with units of mm. Represents the natural logarithm operation; according to Construct a quadratic form matrix. Let f(x) represent the positive definite quadratic form matrix of the candidate ellipse, with dimensions per square millimeter. Indicates the direction from the main axis The generated two-dimensional dimensionless rotation matrix, Indicates the direction of the principal axis of the candidate ellipse, with units of rad. The dimensionless base constant of the natural exponential function is denoted by . This indicates a diagonal matrix construction operation where units are not applicable. This indicates a transpose operation and units are not applicable.

[0111] make This is the set of indices of all contour points in the current cross-section with a reference temperature mass weight greater than zero. The final robust fitting target is based on... Build, This represents the dimensionless fitting objective value to be minimized. This indicates that the result obtained in step S5 is the first... Each reference temperature mass weight, This indicates the Huber loss using the S6-specific inflection threshold. This represents the dimensionless inflection threshold, individually calibrated using the 95th percentile of the square root quadratic radial residual of a qualified round blank under the same calibration version. In this specific embodiment, it is taken as 0.004. The square root operation converts the dimensionless quadratic value into a normalized radial scale. Indicates the first The coordinates of a reference temperature profile point are given, and the unit is mm. Indicates the coordinates of the center of the candidate ellipse, in mm. Let the aforementioned positive definite quadratic form matrix be represented. Indicates traversal Dimensionless integer indexes and units are not applicable;

[0112] The solver selects initial value samples from the points where the reference temperature mass weight is greater than zero, as output in step S5. It first performs weighted algebraic ellipse fitting with positive definite constraints using this mass weight to obtain the candidate center, major semi-axis, minor semi-axis, and principal axis direction as initial values. Initial value fitting is only initiated when there are at least twenty initial value samples and their effective angular coverage reaches the angular coverage threshold described later. Otherwise, the current cross-sectional measurement state is determined to be invalid, and the number of initial value samples, effective angular coverage, reference temperature point set index, and original synchronous cross-sectional data set index are written to the solver, which then sends the data to the alarm section generator. Breakpoint events prevent subsequent fitting of the current section. The initial value principal axis direction is selected according to the equivalence relationship of the difference between the principal axis direction and the branch with the smallest difference from the principal axis direction in step S4. When the initial value is valid, the trust region iterative reweighted least squares method is used to minimize the fitting target. The major semi-axis is constrained to be no less than the minor semi-axis and both are located between 80% and 120% of the corresponding nominal radius. Convergence is determined when the relative change of the target is less than 10 to the power of negative 7 and the parameter increment L2 norm is less than 10 to the power of negative 5. If the fit fails to converge after 30 iterations or the positive definite parameter exceeds the limit, it is recorded as a fitting failure.

[0113] Ellipticity according to calculate, This indicates dimensionless ellipticity, and a larger value indicates a greater difference between the major and minor axes. This represents the length of the major axis of the final fitted ellipse, in mm. The length of the minor axis of the final fitted ellipse is expressed in mm. The lengths of the major axis and minor axis are twice the lengths of the corresponding semi-axis. The calculation results, along with the cross-sectional material coordinates and steel grade specifications, are written into the ellipse measurement record.

[0114] Let the original fitting parameters be at the fitted solution. , The x-coordinate of the candidate ellipse center is expressed in mm. The vertical coordinate of the candidate ellipse center is represented in mm. Indicates the semi-major axis relative to the reference length The dimensionless logarithmic parameter, Indicates the short semi-axis relative to the reference length The dimensionless logarithmic parameter, Indicates the direction of the principal axis of the candidate ellipse, with units of rad. This represents a transpose operation where units are not applicable, resulting in a mixed-dimensional vector. The unit itself is not applicable; define a parametric scale matrix. , The horizontal coordinate scale of the center is expressed in mm. The vertical coordinate scale of the center is expressed in mm. The logarithmic parameter scale representing the semi-major axis is dimensionless. The logarithmic parameter scale representing the minor semi-axis is dimensionless. Let represent the principal axis direction scale in rad, with each scale taking the standard deviation of the corresponding parameter from the common calibration verification data and using the calibration resolution of the same parameter as the lower limit; let This represents the locally linearized reference parameter vector that remains constant during the calculation of the current Jacobian and approximate Hessian matrices. Its components, order, and unit are all consistent with... Same, and take the currently converged fitted solution; increment by dimensionless parameters. Construct the weighted residual Jacobian sum approximation Hessian matrix. This represents a dimensionless local parameter increment vector; according to Determine the normalized parameter covariance. This represents the dimensionless normalized covariance matrix. This represents the dimensionless weighted residual variance. Indicates to The dimensionless approximate Hessian matrix obtained by differentiation This represents the dimensionless canonical coefficient added to it, and is taken as 10 to the power of negative 8. Denotes a dimensionless identity matrix of the same order, with the normalized covariance spectral radius taken as... The largest eigenvalue;

[0115] Effective angular coverage according to calculate, Represents the dimensionless effective angular coverage from zero to one. This represents the set of indices for all active viewpoints in the current section. Indicates traversal A dimensionless integer-view index where units are not applicable. Indicates the first Polar angle range of contour points with reference temperature mass weight greater than zero in an unused viewpoint The unit of measurement is radians representing the union of all effective polar angle intervals, expressed in rad. This represents the value of pi. The number of effective angles of view is a statistical measure of the number of angle pairs that have an effective interval within the allowable error range of 180 degrees azimuth angle difference.

[0116] The angular coverage threshold, the number of viewing angles threshold, and the normalized covariance threshold are determined by the jointly calibrated qualified round billet verification data. The verification procedure statistically analyzes the effective angular coverage lower quantile, the minimum stable value of the number of opposing viewing angles, and the normalized covariance spectrum radius upper quantile of qualified cross sections according to steel grade and specification. In this specific implementation, these are set to 0.72, 1 / 2, and 10 to the power of negative cube, respectively. The three thresholds, along with the parameter scale matrix version, calibration version, sample batch, applicable specifications, statistical quantile, and update time, are written into the validity threshold table.

[0117] When the continuous low-quality polar angle span of a certain viewpoint reaches the failure angle threshold described in step S4, the current section arc segment of that viewpoint is deactivated and the effective angle coverage and the number of opposing effective viewpoints are recalculated. After recalculation, if both items reach the corresponding threshold, the fitting and state determination are continued. If either item fails to reach the threshold, the current section measurement state is directly determined to be invalid and the unmet conditions, deactivated viewpoints, and polar angle gaps are written into the invalid record.

[0118] For a cross section that has been fitted, the measurement status is determined to be reliable when the effective angular coverage reaches the angular coverage threshold, the number of opposing effective viewing angles reaches the viewing angle threshold, and the covariance spectral radius does not exceed the covariance threshold. When only one of the three conditions is not met, it is determined to be pending verification. When at least two conditions are not met or the fitting fails, it is determined to be invalid. The cross section pending verification, along with the reference temperature point set index, is written into the verification queue. The invalid cross section, along with the unmet conditions and the original synchronous cross section data group index, is written into the invalid record.

[0119] The upper limit of ellipticity is set according to the billet acceptance standard corresponding to the target steel grade and specification. The billet acceptance standard parameter table is queried by using the steel grade identifier and billet specification in the synchronous section data group as the joint key. The table records the standard number, version, upper limit of ellipticity, minimum allowable defect length at the rolling entry, effective date and discontinuation date. Only sections with a reliable measurement status and ellipticity strictly greater than the corresponding upper limit are marked as reliable over-limit sections. Sections to be reviewed, invalid sections, and reliable but not over-limit sections are all used as candidate over-limit section breakpoints.

[0120] The round billet acceptance standard parameter table is published by the quality management system and imported into the measurement system after version signature verification. Each record includes the target steel grade, diameter specification, upper limit of 2% for ellipticity, minimum allowable defect length of 500mm, standard version, and effective date. When no completely matching record is found, the upper limit of adjacent steel grades or specifications must not be used. The current cross-section measurement status is determined as pending verification - standard missing, written into the standard missing verification queue, and a breakpoint event is sent to the alarm section generator. This status must not be used to determine whether the standard is credibly exceeded or not. When a completely matching standard is subsequently added, it is recalculated based on the raw material coordinates and the original reference temperature point set, while retaining the status before addition, the status after addition, the original query version, and the new standard version. Standard version changes only apply to cross-section records newly formed after the effective material coordinates.

[0121] The alarm section generator sorts the credible over-limit sections in ascending order of material coordinates, splits the candidate sections at the breakpoints, and requires that the material coordinate interval between adjacent over-limit sections within the same section is no greater than twice the normal sampling material coordinate interval, and that the material coordinate length between the first and last over-limit sections is no less than the minimum allowable defect length in the acceptance standard parameter table. When both conditions are met, the first and last material coordinates, maximum ellipticity, corresponding section group number list, steel grade specifications and standard version are written into the alarm section record and sent to the rolling entry quality system. If either condition is not met, only the over-limit section record is saved and no section alarm is output.

[0122] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0123] This invention uses a set of synchronous cross-section data as a common input and sequentially connects differential profile reconstruction, physical quality evaluation, cross-view geometric correction, dual-domain state joint solution, polar angle temperature registration, and robust ellipse fitting. This ensures that the influence of unreliable profiles is continuously constrained before the data enters the shape solution and alarm judgment, thereby outputting ellipse parameters and measurement status corresponding to the real round billet cross-section for on-site interference conditions.

[0124] Compared to the approach of treating differential imaging, general filtering, and temperature compensation as parallel processes, this invention uses the same quality weight across anomaly detection, loss weighting, and cross-section gating. It indexes the jumping state and shape state with different independent variables and establishes the dependency relationship between temperature samples and contour points by obtaining the center. This strengthens the constraint relationship between each processing step from a technical structure perspective, which helps to reduce false out-of-tolerance and missed detections.

Claims

1. A method for online laser measurement of the ellipticity of continuously cast round billets, characterized in that, include: S1. Simultaneously acquire multi-view laser-off background frame, laser-on contour frame, circumferential temperature data and roll speed encoder data of the same nominal cross-section of the continuously cast round billet. S2. Perform differential processing on the background frame and the contour frame and reconstruct the contour points in the common coordinate system. Determine the physical quality weights bound to the contour points based on the imaging quality of the differential fringes. S3. Correct the physical quality weights based on the geometric consistency of the contour arc segments from different perspectives, and make the corrected quality weights work through contour point selection, robust solution and section validity determination. S4. Establish material coordinates based on the roller speed encoder data, index the center displacement state with the acquisition time, and index the elliptical shape state with the material coordinates. Solve the two types of states together based on the weighted robust residuals of multi-view contour points to obtain the cross-sectional center and hot elliptical shape after separating common mode runout. S5. Determine the polar angle of the profile point based on the center of the cross section, register the circumferential temperature data to the profile point according to the polar angle, and convert the profile point to the reference temperature according to the linear expansion function. S6. Perform robust ellipse fitting with mass weights on the contour points converted to the reference temperature, determine the measurement state based on the sufficiency of the point set and the fitting uncertainty, and output the alarm segment when the measurement state is reliable, the ellipticity exceeds the upper limit of the allowable ellipticity and is continuous along the material coordinate.

2. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 1, characterized in that, S1 includes: At least four sets of narrow-band laser camera probes are set on the outer periphery of the continuous casting round billet roller table. Each narrow-band laser camera probe establishes the transformation relationship from the probe coordinate system to the common coordinate system through common calibration and points to the same nominal cross section. A unified trigger signal causes each narrowband line laser camera probe to sequentially obtain the laser-off background frame and the laser-on contour frame within a cross-sectional acquisition cycle, and the circumferential temperature sampling time and the roller speed encoder count are associated with the cross-sectional acquisition cycle. The material coordinate increment is determined based on the difference in counts between adjacent encoders and the unit pulse displacement of the roller speed encoder, and the material coordinate increment is associated with the cross-section acquisition cycle to form a synchronous cross-section data group.

3. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 2, characterized in that, S2 include: Estimate the gain difference and offset difference between the background frame and the contour frame within the corresponding pixel area, and subtract the pairs of frames after gain and offset matching; Calculate the gray-level centroid or curve peak position along a direction perpendicular to the local direction of the differential fringes to obtain the sub-pixel center; The sub-pixel center triangle is reconstructed to the common coordinate system using the common calibration parameters of the probes of each narrowband line laser camera; Mass components with values ​​from zero to one are generated based on the saturation pixel ratio of the differential stripes, the maximum break length, the deviation of the stripe width from the calibration width, and the differential signal-to-noise ratio. The physical mass weight is determined based on the product or weighted geometric mean of the mass components.

4. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 3, characterized in that, S3 include: A geometric mass component with a value of zero to one is generated based on at least two of the following: the chord length difference of the opposing view profile arc segment, the distance of the common support line of the adjacent view profile arc segment, and the initial ellipse fitting residual. The geometric mass component is multiplied by the physical mass weight determined according to each mass component to obtain the mass weight with a value of zero to one; Contour points whose mass weights are greater than the retention threshold determined by the mass weight distribution of non-circular billet calibration data and qualified circular billet calibration data are retained, and the mass weights corresponding to the same contour points are continued to be used in subsequent robust distance loss terms and cross-sectional validity determinations.

5. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 4, characterized in that, S4 include: According to the acquisition time, the center displacement state of each section within the same sliding window is formed into a time state sequence, and according to the material coordinates, the major semi-axis state, minor semi-axis state and principal axis direction state of each section are formed into a shape state sequence. Construct a joint objective function that includes a robust distance term from the contour points with the mass weights to the corresponding candidate ellipses, a temporal smoothing constraint term for the temporal state sequence, and a material coordinate smoothing constraint term for the shape state sequence; Solve the joint objective function, and include the displacements of the contour points at each viewpoint that act together with the acquisition time into the central displacement state. Include the major semi-axis, minor semi-axis, and principal axis directions of the ellipse that change with the material coordinates into the shape state to obtain the center of the cross section and the shape parameters of the hot ellipse.

6. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 5, characterized in that, S5 include: The installation orientation angle of each circumferential temperature sensor is transformed to a polar coordinate system with the center of the cross section as the origin. The temperature of the contour point is obtained by interpolation between valid circumferential temperature samples located on both sides of the contour point based on the polar angle of the contour point. The temperature confidence level with a value of zero to one is determined based on the validity mark of the circumferential temperature sample used for interpolation, the angular distance between each circumferential temperature sample and the polar angle of the contour point, and the allowable error of the temperature sensor. When there are fewer than two valid circumferential temperature samples for the profile point temperature interpolation, or when the azimuth interval between the valid circumferential temperature samples on both sides of the profile point exceeds the temperature interpolation angle threshold determined by the circumferential temperature sensor installation interval, the temperature confidence of the profile point is set to zero and the profile point is not included in the reference temperature quality weighted point set. Based on the steel grade identification, the effective linear expansion coefficient with temperature variation within the range from the profile point temperature to the reference temperature is queried, and the radial coordinate of the profile point relative to the center of the cross section is converted point by point. For each circumferential temperature sample, the predicted temperature is obtained by interpolating the two adjacent and valid circumferential temperature samples according to the installation azimuth angle. When there are fewer than two valid circumferential temperature samples for prediction, or when the difference between the measured temperature of the circumferential temperature sample and the predicted temperature exceeds the temperature difference threshold determined by the allowable error of the temperature sensor and the statistical value of the temperature gradient of the qualified round billet, the polar angle interval covered by the circumferential temperature sample is marked as a temperature anomaly interval, and the temperature confidence of the contour points in the temperature anomaly interval is limited to no greater than the upper limit of the temperature anomaly weight determined by the calibration data containing temperature occlusion interference. The reference temperature mass weight is obtained by multiplying the temperature confidence level by the mass weight.

7. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 6, characterized in that, S6 include: The candidate ellipse is represented by a parameterization method that guarantees the positive definiteness of the quadratic form matrix. The robust distances from the contour points to the candidate ellipse are weighted and summed according to the reference temperature quality weight, and the candidate ellipse is solved. The ellipticity is determined by the ratio of the difference between the major axis and the minor axis to the minor axis, and the covariance of the fitting parameters is determined by the local curvature of the weighted residuals. When the effective angular coverage reaches the angular coverage threshold, the number of effective viewing angles reaches the viewing angle threshold, and the spectral radius of the covariance of the fitting parameters does not exceed the covariance threshold, the measurement state is determined to be reliable. When one of the three conditions is not met but the other two are met, the measurement status is determined to be pending verification. The measurement status is determined to be invalid if at least two conditions are not met. The corner coverage threshold, the number of viewing angles threshold, and the covariance threshold are determined by the jointly calibrated qualified round blank verification data.

8. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 7, characterized in that, When the range of polar angles continuously traversed by contour points whose quality weight is not greater than the point retention threshold in a viewpoint reaches the failure angle threshold determined by the nominal coverage angle of that viewpoint, the contour arc segment of that viewpoint in the current section is deactivated and the effective angle coverage and the number of opposing effective viewpoints are recalculated. When the recalculated effective angular coverage reaches the angular coverage threshold and the number of opposing effective viewing angles reaches the viewing angle number threshold, continue to execute steps S4 to S6. When the recalculated effective angular coverage does not reach the angular coverage threshold or the number of opposing effective viewing angles does not reach the viewing angle number threshold, the measurement status of the current section is determined to be invalid.

9. The method for online laser measurement of the ellipticity of continuously cast round billets according to claim 7, characterized in that, The upper limit of the allowable ellipticity is set according to the round billet acceptance standard corresponding to the target steel grade and specifications. The sections with a reliable measurement status and an ellipticity exceeding the upper limit of the allowable ellipticity are sorted according to material coordinates. The sections to be reviewed, invalid sections, and sections with a reliable measurement status but an ellipticity not exceeding the upper limit of the allowable ellipticity are all taken as the breakpoints of the candidate over-limit sections, so that the reliable over-limit sections on both sides of the breakpoint are respectively assigned to different candidate over-limit sections. Within each candidate over-limit section, when the material coordinate interval between adjacent over-limit sections is not greater than the connection interval determined by the sampling material coordinate spacing and the material coordinate length between the first and last over-limit sections reaches the alarm length determined by the minimum allowable defect length at the rolling inlet, the material coordinate interval between the first and last over-limit sections is determined as the alarm section. Write the sections to be reviewed into the review queue, and write the invalid sections and the validity conditions they do not meet into the invalid record.

10. A laser online measurement system for the ellipticity of continuously cast round billets, used to perform the laser online measurement method for the ellipticity of continuously cast round billets according to any one of claims 1 to 9, characterized in that, include: The synchronous acquisition module is used to synchronously acquire laser-off background frames, laser-on contour frames, circumferential temperature data and roller speed encoder data from multiple perspectives to form a synchronous cross-sectional data set. The contour reconstruction module is used to perform differential processing on paired frames, extract sub-pixel centers, reconstruct contour points in a common coordinate system, generate each quality component based on the differential fringe imaging quality, and determine the physical quality weights bound to the contour points based on each quality component. The quality propagation module is used to correct the physical quality weights based on the geometric consistency between contour arcs from different perspectives, and to use the resulting quality weights for contour point preservation, robust distance weighting, and cross-section validity determination. The dual-domain state solving module is used to index the center displacement state by acquisition time, the major semi-axis state, the minor semi-axis state, and the principal axis direction state by material coordinates, and to jointly solve the center displacement state and each shape state. The temperature registration and compensation module is used to register the circumferential temperature data to the contour point according to the polar angle based on the obtained cross-section center and convert it to the reference temperature. The ellipse fitting and alarm module is used to perform robust ellipse fitting with positive definite constraints, determine the measurement state, and form alarm segments that continuously exceed limits along the material coordinates.

Images