A 3D profile scanning thickness measurement system for a conveyor belt

Abstract

The application relates to the technical field of 3D thickness measurement, and discloses a 3D profile scanning thickness measurement system for a conveying belt, which comprises the following steps: tracking the longitudinal displacement of the belt body by using a speed measurement unit, synchronously triggering a double-side scanning unit to construct a double-side point cloud field; analyzing the point cloud field to extract the span of a supporting roller and generate the phase of the supporting roller; flattening the profile along the middle line of a section to establish a width coordinate, extracting a local normal direction to obtain an initial thickness field; constructing a periodic disturbance model by using the phase of the supporting roller, separating reversible periodic deformation variables, and obtaining a real thickness field; combining the transverse bending state of the belt body to perform weight distribution on the real thickness field, generating a comprehensive thickness value and a phase thickness map; and performing self-adaptive recursive updating based on a measurement residual error. The application effectively eliminates the interference of dynamic deformation and transverse deviation of a supporting roller on thickness measurement, and realizes stable and reliable online loss evaluation.

Description

Technical Field

[0001] This invention relates to the field of 3D thickness measurement technology, and more specifically, to a 3D contour scanning thickness measurement system for conveyor belts. Background Technology

[0002] In bulk material transportation, conveyor belts serve as continuous, closed-loop, flexible composite carriers. Under continuous operation, the conveyor belt exhibits a complex dynamic spatial morphology due to the combined effects of idler support, tensioning devices, and material loads. Existing three-dimensional profile thickness measurement solutions typically fix a scanner on a frame, acquiring spatial point clouds of the belt surface and directly calculating the vertical height difference to assess thickness loss.

[0003] However, in actual transportation conditions, the conveyor belt experiences periodic sagging between adjacent idlers, lateral bending when passing through trough idler sets, and viscoelastic indentation when the rubber layer contacts the idlers. These factors cause a large number of reversible geometric variables to be superimposed on the apparent profile acquired by the fixed scanner, resulting in localized true thickness loss being masked by dynamic deformation. Furthermore, off-center loading and idler skew often cause lateral movement and edge warping of the belt during operation, making it difficult to stably map the thickness map acquired under fixed coordinates to the same material location within the belt itself. This causes wear data at the vane-roller junction or edge areas to drift when longitudinally unfolded. These multiple dynamic interferences are coupled together, making it difficult for conventional direct height difference comparison methods to output accurate and repeatable loss assessment data in complex industrial environments. Summary of the Invention

[0004] This invention provides a 3D contour scanning thickness measurement system for conveyor belts, which solves the technical problems mentioned in the background art.

[0005] This invention provides a 3D contour scanning thickness measurement system for conveyor belts, applied to online thickness measurement operations including an upper 3D contour scanning unit and a lower 3D contour scanning unit arranged on a frame, and a speed measurement unit, comprising:

[0006] The longitudinal displacement of the conveyor belt is tracked using the data from the speed measurement unit, and the upper three-dimensional contour scanning unit and the lower three-dimensional contour scanning unit are simultaneously triggered to acquire the upper contour point cloud and the lower contour point cloud, so as to construct a dual-sided point cloud field that eliminates the interference of time sampling misalignment.

[0007] The lateral curvature caused by the support structure in the double-sided point cloud field is analyzed to extract the idler span and generate the idler phase characterizing the periodic sag of the idler support.

[0008] Extract the centerline of the current measurement section corresponding to the dual-sided point cloud field, flatten the contour along the centerline of the section, and establish the width coordinates for correcting lateral deviation drift.

[0009] The initial normal thickness is extracted along the centerline of the cross section in the local normal direction to construct an initial thickness field that reflects the true spatial thickness properties of the belt.

[0010] A periodic disturbance model is constructed based on the idler roller phase to separate and remove the periodic deformation caused by the idler roller support from the initial thickness field, and to restore the true thickness field.

[0011] The average bending curvature of the belt in the lateral direction is combined to perform local structural weight allocation on the true thickness field to generate a thickness feature set that reflects the structural wear tendency of the edge and the bottom of the groove; wherein, the thickness feature set includes a comprehensive thickness value and a phase thickness map;

[0012] The historical reference thickness field of the previous measurement cycle is obtained. Based on the measurement residual of the real thickness field relative to the historical reference thickness field, adaptive recursion is performed. While filtering transient environmental noise, the thickness measurement output set and reference thickness field of the current cycle are updated and generated.

[0013] The beneficial effects of this invention include: by integrating the longitudinal displacement and span rhythm of the conveyor belt with the centerline flattening state, the periodic reversible deformation caused by the idler roller support is effectively separated, and the thickness measurement position drift error caused by lateral wandering is corrected; after obtaining the true thickness field, specific weight allocation is performed in combination with the lateral bending state of the belt, and transient environmental noise is filtered through adaptive recursive updates, which significantly improves the accuracy and repeatability of online thickness measurement under transportation conditions, and provides highly reliable data support for belt wear monitoring and predictive maintenance decisions. Attached Figure Description

[0014] Figure 1 This is a flowchart of the operation of a 3D contour scanning thickness measurement system for conveyor belts according to the present invention. Detailed Implementation

[0015] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0016] The following combination Figure 1The implementation process of a 3D contour scanning thickness measurement system for conveyor belts is described. Figure 1 From top to bottom, the process corresponds to the following steps: construction of the dual-sided point cloud field, generation of the idler roller phase, establishment of the width coordinate, construction of the initial thickness field, restoration of the true thickness field, generation of the thickness feature set, and recursive updating of the thickness measurement output set. The flowcharts are connected in the order of data input, data processing, and data output. The output of the previous flowchart serves as the input of the next, ensuring that longitudinal displacement, dual-sided point cloud field, idler roller phase, width coordinate, initial thickness field, true thickness field, thickness feature set, thickness measurement output set, and reference thickness field form a continuous data link.

[0017] The upper and lower 3D contour scanning units are installed on the upper and lower frames of the conveyor belt, respectively. Their scanning planes are coplanar and perpendicular to the running direction of the conveyor belt. The speed measurement unit is installed on the driven roller of the conveyor belt. If the driven roller cannot be installed, an unloaded idler is selected first, and a loaded idler is selected last. When installed on a loaded idler, an elastic deformation correction coefficient of 1.03 is introduced. The GPIO output of the system control unit is connected to the trigger input of the upper and lower 3D contour scanning units, and the two units are synchronously triggered to acquire the upper and lower contour point clouds at the same physical cross-section using TTL level hardware.

[0018] The upper and lower 3D contour scanning units have a scanning frequency of 1000Hz, a lateral resolution of 0.5mm, a measurement range of 200mm, and a measurement accuracy of 0.1mm. Their coplanarity error does not exceed 0.1mm / m, and their perpendicularity error does not exceed 0.2 degrees. The velocity measurement unit has a sampling frequency of 1000Hz and a measurement accuracy of 0.01m / s. These parameters are used to define the preferred implementation method, allowing for equivalent configurations while satisfying the same data processing logic.

[0019] The equipment coordinate system is defined with the center of the lens of the upper 3D contour scanning unit as the origin, the x-axis along the belt running direction, the y-axis along the horizontal direction, and the z-axis vertically upward. The system obtains the transformation matrix between the equipment coordinate system and the original point cloud coordinate system through standard gauge blocks calibration. The transformation matrix includes translation and rotation matrices. After the standard gauge blocks are placed at the center of the conveyor belt, the point clouds of the gauge blocks from the upper and lower 3D contour scanning units are collected respectively, the coordinates of the four corner points are extracted, and the coordinate transformation matrix is ​​obtained by least squares fitting.

[0020] Example 1:

[0021] Figure 1 The first flowchart corresponds to S201 to S203. This part uses a velocity measurement unit to track displacement and synchronously triggers dual-sided scanning, so that the upper and lower contour point clouds are continuously expanded based on the running mileage of the belt material, forming a dual-sided point cloud field that eliminates the interference of time sampling misalignment.

[0022] S201: The instantaneous speed of the conveyor belt is obtained through the speed measurement unit, and the longitudinal displacement corresponding to the current sampling moment is calculated based on the kinematic transformation relationship between speed and displacement to characterize the actual running mileage of the belt material.

[0023] Under normal conditions, the system calculates the longitudinal displacement using the trapezoidal integral method on the velocity sampling data. The calculation form of the longitudinal displacement is:

[0024]

[0025] in, For longitudinal displacement, As the initial reference point, The scan start time, At the current sampling time, For the conveyor belt in time variable The instantaneous velocity. The initial reference point is automatically calibrated through the joint feature of the belt body, which is used to form a physical positioning reference for longitudinal displacement.

[0026] The identification steps for the joint features are as follows: Calculate the longitudinal thickness gradient along a sliding window with 20 longitudinal sampling points and a step size of 1 sampling point. When the maximum longitudinal thickness gradient within the window exceeds 5 mm / m and the thickness values ​​of 20 consecutive sampling points deviate from the nominal thickness by more than 20%, it is identified as a joint area, and the center longitudinal displacement of the joint area is taken as the initial reference point. During the initial run before obtaining the true thickness field, the longitudinal thickness gradient is calculated longitudinally from the height difference between the upper and lower contour point clouds at the same horizontal coordinate. After completing the initial normal thickness calculation, the longitudinal thickness gradient is calculated longitudinally from the initial normal thickness. After restoring the true thickness field, the longitudinal thickness gradient is calculated longitudinally from the true thickness field. The system re-extracts joint features every 100 cycles to correct reference point drift.

[0027] The system detects belt slippage in real time. When the instantaneous speed is lower than 70% of the average speed over the past second and the duration exceeds 100ms, slippage is identified. In this case, a visual feature tracking method based on the belt surface is used to supplement displacement measurement. An industrial camera is mounted above the conveyor belt, with a frame rate of 30fps and a resolution of 1280x720. The ORB algorithm extracts texture feature points from the belt surface, calculates the average displacement of feature points in adjacent frames, and combines this with camera calibration parameters to convert it into the actual displacement of the belt. This displacement is then weighted and fused with the encoder displacement, with a fusion weight of 0.3 for the encoder and 0.7 for the vision.

[0028] S202: At the current sampling time, trigger the upper 3D contour scanning unit and the lower 3D contour scanning unit to acquire the upper contour point cloud and the lower contour point cloud at the same physical cross section, respectively.

[0029] The system uses TTL level hardware synchronous triggering, with the trigger time difference between the upper and lower 3D contour scanning units not exceeding 10µs. The trigger signal is active high, with a high-level duration of 10µs. Each acquired point cloud point contains five attributes: x-coordinate, y-coordinate, z-coordinate, reflection intensity, and timestamp, with the data type being a 32-bit floating-point number.

[0030] After the raw point cloud was acquired, outliers were first removed using statistical filtering with 50 neighboring points and a standard deviation factor of 3. Then, radius filtering was used to further filter noise with a search radius of 5mm and a minimum neighboring point count of 10. Points with a reflection intensity below 100 were considered low-quality and removed, while points with a reflection intensity above 2000 were considered specular reflection points and replaced with the median of their five neighboring points. If the height values ​​of three or more consecutive point cloud points exceeded the measurement range, they were considered occluded; if the occlusion length was less than 10mm, linear interpolation was used for completion; if it was greater than 10mm, it was marked as an invalid region.

[0031] There is a trigger time difference between the upper and lower scan units. At that time, if the belt speed is Then, the x-coordinate of the lower contour point cloud is corrected for time synchronization:

[0032]

[0033] in, The x-coordinates of the lower contour point cloud before correction. The corrected x-coordinates of the lower contour point cloud. This represents the trigger time difference between the upper 3D contour scanning unit and the lower 3D contour scanning unit.

[0034] S203: The longitudinal displacement is used as the entity positioning label and spatially bound to the corresponding upper and lower contour point clouds to generate a double-sided point cloud field that is continuously expanded based on the running mileage of the belt material, so as to eliminate the longitudinal spatial distortion interference introduced by belt speed fluctuation.

[0035] For a single-frame point cloud, if the bandwidth changes by more than 5% within a single frame scan time, the system uses linear interpolation to assign an independent displacement value to each point based on the difference between the timestamp of each point and the frame start time, thus avoiding intra-frame spatial distortion. The bilateral point cloud field is organized as follows:

[0036]

[0037] in, For a two-sided point cloud field, Add cloud points to the upper outline. Point clouds for the lower contour. For longitudinal displacement, The horizontal axis is... For height coordinates, Reflection intensity, For timestamps. After point cloud and displacement are bound, they are stored in binary format, with each point written in the order of x-coordinate, y-coordinate, z-coordinate, reflection intensity, timestamp, and displacement value.

[0038] Example 2:

[0039] Figure 1 The second flowchart corresponds to S301 to S304. This part analyzes the lateral curvature caused by the support structure in the bilateral point cloud field, extracts the idler span, and generates the idler phase characterizing the periodic sag of the idler support.

[0040] S301: Fuse the height features of the upper and lower contour point clouds at the same lateral coordinates to fit and generate the centerline height that represents the overall undulation state of the current measurement section.

[0041] The system interpolates the upper and lower contour point clouds onto a uniform horizontal grid with a spacing of 0.5 mm, and uses the nearest neighbor interpolation method to obtain height features at the same horizontal coordinates. Effective bandwidth is identified using a combination of gradient and density methods; locations with a horizontal gradient value exceeding 5 mm / pixel are considered edge points, and areas with point cloud density below 30% of the average density are considered outside-edge regions. If the height difference between five consecutive edge points exceeds 10 mm, it is determined to be a stuck point, which is then removed before refitting the edge. The centerline height is used to characterize the centerline of the current measurement section, and its calculation form is as follows:

[0042]

[0043] in, The height of the centerline. and These represent the heights of the upper and lower contour point clouds at the same horizontal coordinate, respectively.

[0044] S302: Extract the curvature variation characteristics of the centerline height within the effective bandwidth of the conveyor belt and perform transverse spatial aggregation to evaluate the transverse curvature, which characterizes the degree of transverse bending deformation at each longitudinal displacement.

[0045] Before calculating curvature, the centerline height is first Gaussian smoothed with a Gaussian kernel value of 1.5 times the horizontal grid spacing and a smoothing window size of 5x5. Then, the curvature variation characteristics are calculated using the central difference method, with boundary points using either forward or backward differencing. To avoid positive and negative curvatures canceling each other out during horizontal spatial convergence, the horizontal curvature is expressed as an integral of the absolute value of curvature.

[0046]

[0047]

[0048] in, Features of curvature variation For lateral curvature, and These represent the left and right boundaries of the current measurement section, respectively. In the discrete case, the system uses the trapezoidal integral method to calculate the lateral curvature.

[0049] S303: Perform longitudinal morphological similarity matching on the transverse curvature along the conveyor belt running direction, and determine the cycle length corresponding to the peak value of the matching correlation as the idler span value.

[0050] When the idler span is unknown, the initial calculation interval is 10m. If no effective autocorrelation peak is detected within this interval, it is successively extended to 20m, 30m, and a maximum of 50m. The hysteresis domain can be calculated from 0.5m to 5m, covering the common idler span ranges in industrial settings. The autocorrelation function is:

[0051]

[0052]

[0053] in, It is the autocorrelation function. This is the longitudinal displacement hysteresis. This represents the average lateral curvature. and To calculate the start and end positions of the interval, This represents the idler roller span value. and These represent the lower and upper limits of the computable hysteresis domain, respectively. After the autocorrelation coefficient is calculated, a triple criterion is used to screen peak values: the peak height is greater than 2.5 times the mean of the autocorrelation coefficient; the relative deviation between adjacent peaks is less than 10%; and the peak spacing is stable for more than three consecutive periods. When there are multiple idler groups with different spans, a db4 wavelet is used for three-level decomposition to separate the periodic components of different frequencies.

[0054] S304: Position the physical location corresponding to the peak of the lateral curvature as the support reference point, and generate the idler phase representing the idler support sag state based on the periodic proportion of the longitudinal displacement relative to the support reference point in the idler span value.

[0055] Before extracting the peak value, a 5-point moving average filter is applied to the lateral curvature sequence to eliminate spurious peak values ​​caused by random noise. When multiple peak values ​​exist, the peak value corresponding to the peak spacing closest to the theoretical idler span is selected as the support reference point. Every 10 idler spans, the system re-extracts the support reference point, calculates the deviation between the current phase and the theoretical phase, and performs a global phase correction if the deviation exceeds a set threshold.

[0056]

[0057]

[0058] in, As the support reference point, For the idler roller phase, To be a function that takes the fractional part, the range of values ​​for the idler roller phase is: .when When it approaches 0, the corresponding longitudinal displacement approaches the support reference point; when When the value approaches 0.5, the longitudinal displacement is close to the mid-span position between adjacent support reference points. When the idler roller is skewed, the system extracts the lateral coordinates of the peak lateral curvature at 10 consecutive idler roller support positions, calculates the linear fitting slope of the peak lateral coordinates, and performs phase correction accordingly.

[0059] Example 3:

[0060] Figure 1 The third flowchart corresponds to S401 to S403. This part extracts the centerline of the current measurement section, flattens the profile along the centerline, and establishes width coordinates for correcting lateral drift.

[0061] S401: Extract local slope features based on centerline height, accumulate curve path lengths inward along the effective data boundary of the current measurement section, and calculate the arc length coordinates of the centerline of the section after eliminating the groove bending deformation.

[0062] The local slope is calculated using the central difference method with a difference step size of 0.5 mm. For the left boundary point, forward difference is used, and for the right boundary point, backward difference is used. For cases where the boundary point cloud is missing, a linear extrapolation method is used, with the extrapolation length not exceeding 5% of the total arc length; any extrapolation exceeding this is marked as an invalid section. The arc length coordinates are:

[0063]

[0064] in, Let the coordinates be arc length. This represents the left boundary of the current measurement section. For integration, the curve path length is accumulated sequentially from the left boundary along the horizontal grid points in the discrete case.

[0065] S402: Determine the total arc length parameter after the centerline of the cross section is fully unfolded, and proportionally map the arc length coordinates to the two end boundaries to convert them into standardized width coordinates within the specified reference interval.

[0066] If the current total arc length parameter deviates by more than 5% from the average total arc length parameter of the past 100 sections, the edge is re-identified and the total arc length parameter is recalculated. The standardized width coordinates use the specified reference interval by default. .

[0067]

[0068]

[0069] in, This is the total arc length parameter. The right boundary For width coordinates, when performing inverse solving of width coordinates, a mapping table between arc length coordinates and original horizontal coordinates is established. For a given width coordinate, the corresponding arc length coordinate is first calculated, and then the corresponding original horizontal coordinate is obtained through linear interpolation.

[0070]

[0071] in, These are the arc length coordinates obtained by inverse solving of the width coordinates. This inverse solution relation is used for subsequent resampling and lateral spatial variability transformation.

[0072] S403: The discrete points of the upper and lower contour point clouds are resampled using the width coordinates to convert the bilateral height data in the fixed equipment coordinate system to the width coordinate system composed of longitudinal displacement and width coordinates, so as to correct the thickness measurement position drift error caused by lateral snake deviation.

[0073] The grid resolution is set to 0.5mm vertically and 0.5mm horizontally, not less than half the original resolution of the scanning unit. For each target grid point, the height value is calculated using bilinear interpolation.

[0074]

[0075] in, and For adjacent vertical grid points, and For adjacent width coordinate grid points, , , and The height values ​​are for four adjacent points. Boundary extrapolation is used within 10% of the edge width, while zero-fill extrapolation is used for the internal area. Normal correction occurs when the deviation is less than 5% of the bandwidth; deviation between 5% and 10% is marked as slight deviation; and deviation greater than 10% triggers an alarm and marks the data at that location as low reliability.

[0076] The upper height data in the fixed equipment coordinate system is denoted as The lower height data in the fixed equipment coordinate system is denoted as After inverse kinematics in width coordinates and mesh resampling, the bilateral height data transformed to the width coordinate system are as follows:

[0077]

[0078]

[0079] in, and These are the bilateral height data after resampling and transformation to the width coordinate system. This represents the original horizontal coordinates obtained by inverse solving of the width coordinates. In this expression... and Used only to distinguish input data in a fixed equipment coordinate system. and This serves as the input data for subsequent spatial surface reconstruction.

[0080] Example 4:

[0081] Figure 1 The fourth flowchart corresponds to S501 to S504. This part extracts the initial normal thickness along the local normal direction along the centerline of the cross section, and constructs an initial thickness field that reflects the true spatial thickness properties of the strip.

[0082] S501: In the width coordinate system, spatial surface reconstruction is performed on the resampled bilateral height data to generate upper and lower surface surfaces that reflect longitudinal displacement and width changes, and the spatial mean layer of the two is extracted as the mid-surface surface.

[0083] The spatial surface reconstruction employs a bicubic B-spline interpolation method, with the B-spline order being 3. The number of control points is calculated as (vertical grid points / 4) multiplied by (horizontal grid points / 4), and natural boundary conditions are used. After reconstruction, the mean square error between the reconstructed surface and the original point cloud is calculated. If the error exceeds 0.3 mm, the number of control points is increased, and reconstruction is repeated. For missing point cloud regions, radial basis function interpolation is used to complete the area if it is less than 100 mm², while areas greater than 100 mm² are marked as invalid regions.

[0084] The upper, lower, and middle surfaces are all spatial surfaces, each containing longitudinal physical coordinates, transverse physical coordinates obtained by inverse solving the width coordinates, and height coordinates, respectively. The upper height component after bicubic B-spline interpolation is denoted as... The lower height component after bicubic B-spline interpolation is denoted as... Then the upper surface, lower surface, and middle surface are:

[0085]

[0086]

[0087]

[0088] in, The upper surface is curved. The lower surface is curved. It is a mid-surface curved surface. and The height component is obtained after bicubic B-spline interpolation. These are the original transverse coordinates obtained by inverse solving of the width coordinates. The radial basis function interpolation uses a Gaussian kernel function with a shape parameter of 0.1.

[0089] S502: Extract orthogonal tangential features along the longitudinal and transverse directions of the mid-surface surface, and determine the mid-surface normal vector representing the local real spatial attitude of the band body based on the orthogonal normal constraint.

[0090] Tangential features are calculated using the central difference method, with a longitudinal difference step size of 0.5 mm and a transverse difference step size of 0.5 mm. Since the mid-surface is a spatial surface, differentiating along the longitudinal and width coordinates yields spatial tangent vectors. The cross product of these two spatial tangent vectors is used to determine the mid-surface normal vector. After calculating the mid-surface normal vector, if the z-component is negative, the direction is reversed to ensure the mid-surface normal vector points upwards on the belt. Subsequently, the mid-surface normal vector is averaged and smoothed using a 3x3 window, and then re-normalized.

[0091]

[0092] in, Let be the normal vector of the mid-plane. Let be the spatial tangent vector of the mid-surface surface along the longitudinal displacement direction. Let be the spatial tangent vector of the mid-surface surface along the width coordinate direction. For vector cross product, Calculate the vector magnitude.

[0093] S503: Extract the spatial difference vector between the upper and lower surface surfaces, and project this spatial difference vector into the orientation direction of the mid-surface normal vector to separate the initial normal thickness perpendicular to the local surface.

[0094] The validity of thickness values ​​is verified using a set range for the nominal thickness; thickness values ​​outside this range are marked as abnormal and replaced with the average of three neighboring points. Since both the upper and lower surface surfaces are spatial surfaces, their difference is used to obtain a spatial difference vector. The initial normal thickness is:

[0095]

[0096] in, The initial normal thickness, This is the dot product in spatial projection calculations. This process is used to avoid geometric measurement errors that occur when the vertical height difference is used to replace the true normal thickness.

[0097] S504: Assemble the initial normal thickness with the corresponding width coordinates and idler phase into an attribute matrix to form an initial thickness field, so as to eliminate the geometric measurement error caused by the vertical height difference deviating from the true normal thickness.

[0098] The three-dimensional attribute matrix of the initial thickness field has the following dimensions: vertical grid number, horizontal grid number, and 3. The three attributes are, in order, the initial normal thickness, the idler phase, and the width coordinate. The data type is a 32-bit floating-point number. The rules for handling missing data are as follows: when a single grid point is missing, neighborhood interpolation is used to complete it; when three or fewer consecutive grid points are missing, linear interpolation is used to complete it; and when more than three consecutive grid points are missing, it is marked as an invalid region.

[0099]

[0100] in, This represents the initial thickness field. The initial thickness field is stored in row-major order, with the storage order being vertical grid points, horizontal grid points, and attribute dimensions. Invalid regions are marked with a value of -9999.0 and are automatically skipped in subsequent calculations.

[0101] Example 5:

[0102] Figure 1 The fifth flowchart corresponds to S601 to S604. This part constructs a periodic disturbance model based on the idler roller phase, separates and removes the periodic deformation caused by the idler roller support from the initial thickness field, and restores the true thickness field.

[0103] S601: Decouples the initial normal thickness in the initial thickness field into three independent physical components: the true thickness field, the phase disturbance field representing the periodic deformation of the idler roller, and the occasional measurement noise.

[0104] The system expresses the initial normal thickness as the sum of the true thickness field, the phase disturbance field, and the occasional measurement noise. The true thickness field represents the non-idler phase synchronization component that varies with the position of the belt material; the phase disturbance field represents the periodic deformation related to the phase synchronization of the idler and its harmonics; and the occasional measurement noise represents the unstable residual component that recurs with longitudinal displacement or idler phase.

[0105]

[0106] in, For the true thickness field, For phase perturbation field, To mitigate occasional measurement noise, after estimating the phase perturbation field and reconstructing the true thickness field, the system uses cross-correlation coefficients to verify the independence of the true thickness field and the phase perturbation field; a cross-correlation coefficient with an absolute value less than 0.1 is considered independent. If the absolute value of the cross-correlation coefficient is greater than 0.1, an observation matrix is ​​constructed using the initial normal thickness sequence, the estimated phase perturbation field sequence, and the residual sequence at multiple width coordinates, and independent component analysis is used for secondary decoupling. The independent component analysis uses the FastICA algorithm, with an upper limit of 1000 iterations and a convergence threshold of [value missing]. The nonlinear function is the tanh function. The variance of sporadic measurement noise is estimated by db4 wavelet 3-level decomposition, extracting the third-level detail components and calculating their variance.

[0107] S602: Using the complexity of the balancing disturbance model and the convergence of the residuals as constraints, adaptive optimization is used to determine the optimal order for characterizing the deformation features of the idler support.

[0108] The system normalizes the residual by dividing it by the nominal thickness to obtain the normalized residual, and then calculates the Akaike Information Criterion. The order search range is 1 to 10. When the change in the Akaike Information Criterion value for three consecutive orders is less than 10, the system calculates the normalized residual. If the search is terminated prematurely, then for any order... Thickness estimation components The disturbance component is obtained by subtracting the disturbance component of the corresponding order from the initial normal thickness and then smoothing it longitudinally for a length greater than one idler roller span; the disturbance component The thickness is obtained by subtracting the estimated thickness component from the initial normal thickness and then fitting it using a periodic basis function. Both are iteratively updated at the same order until the change in the normalized residual is less than [the required value]. Or the number of iterations reaches 20.

[0109]

[0110] in, For optimal order, For order variables, For the thickness estimation components at the corresponding order, These are the perturbation components at the corresponding order. The total number of discrete grid points. For the first Coordinates of discrete grid points This represents the nominal thickness of the belt. The model parameter dimensions are... Each order corresponds to one cosine coefficient and one sine coefficient, and includes one DC component.

[0111] S603: At the optimal order, the initial normal thickness at different width coordinates is expanded in the time-frequency domain using periodic basis functions to extract the phase disturbance field that is synchronous with the phase of the idler roller.

[0112] For each width coordinate, the QR decomposition method is used to solve the least squares problem to obtain the Fourier coefficients, avoiding the numerical instability problem of the normal equation method. The matrix is ​​designed... Each line is composed of , and Composed of, the Fourier coefficient vector is , and The Fourier coefficients are then smoothed using a three-point moving average along the width coordinate direction, with smoothing coefficients of 0.25, 0.5, and 0.25. The phase perturbation field is:

[0113]

[0114] in, The DC component, and For the first Fourier coefficients. The objective function for least squares fitting is:

[0115]

[0116] in, This represents the number of vertical sampling points. This represents the thickness estimation components at the optimal order. The QR decomposition method will use the design matrix... Decompose into orthogonal matrices and upper triangular matrix Then, solve for the Fourier coefficient vector.

[0117] S604: Removes the phase disturbance field from the initial normal thickness, eliminates the periodic deformation caused by the sag of the idler roller and rubber indentation, and restores the true thickness field that reflects the physical wear state.

[0118] During stripping, the portion of the phase disturbance field that varies with the idler roller phase is the primary target for stripping, and the DC component is used as the fitting benchmark in the Fourier coefficient calculation process. This approach avoids interpreting the stable bias of the true thickness field as a periodic deformation. The periodic deformation in the phase disturbance field is:

[0119]

[0120] in, Let be the periodic deformation variable in the phase disturbance field that varies with the phase of the idler roller. The true thickness field is:

[0121]

[0122] in, The true thickness field is represented. After stripping, autocorrelation analysis is performed on the true thickness field. If a peak with the same frequency as the idler roller exists and its height is greater than 1.5 times the mean, a secondary phase perturbation stripping is performed. The termination condition for secondary phase perturbation stripping is that the residual autocorrelation peak height is less than 1.5 times the mean autocorrelation coefficient, or the number of stripping operations reaches 3. During offline verification, a cross-section is sampled every 10m along the longitudinal direction of the belt, and 5 points are sampled at each cross-section. An ultrasonic thickness gauge is used for verification, and each measurement point is measured 3 times and the average value is taken.

[0123] Example 6:

[0124] Figure 1 The sixth flowchart corresponds to S701 to S704. This part combines the average bending curvature of the belt in the lateral direction to perform local structural weighting on the true thickness field, so as to generate a thickness feature set that reflects the structural wear tendency of the edge and the bottom of the groove; the thickness feature set includes a comprehensive thickness value and a phase thickness map.

[0125] S701: Extract the spatial variation rate of the true thickness field in the two dimensions of longitudinal displacement and width coordinate, and fuse them to generate a thickness gradient that characterizes the local wear gradient of the belt.

[0126] The spatial variability was calculated using the central difference method, with a longitudinal difference step size of 0.5 mm and a transverse difference step size of 0.5 mm. Since the width coordinate satisfies... The lateral variability is converted to a physical scale consistent with the longitudinal variability through the total arc length parameter.

[0127]

[0128]

[0129]

[0130] in, and For spatial variability, The thickness gradient is the ratio of the thickness gradient to the nominal thickness. The warning threshold for the relative thickness gradient is 0.2. 0 to 0.1 indicates normal wear, 0.1 to 0.2 indicates mild abnormal wear, 0.2 to 0.3 indicates moderate abnormal wear, and greater than 0.3 indicates severe abnormal wear.

[0131] S702: Perform spatial smoothing operation on the mid-surface curved surface along the specified longitudinal measurement range to extract the average groove line of the mid-surface that reflects the transverse morphology of the solid support structure of the conveyor belt.

[0132] Spatial smoothing employs a Hanning window moving average filter, with the window size equal to the number of longitudinal sampling points corresponding to one idler span. The length of the longitudinal measurement range must include at least one complete idler span; when the idler span is unknown, the extracted idler span value is used.

[0133]

[0134] in, The average groove shape of the mid-surface. For the height component of the mid-surface surface, For the specified longitudinal measurement range, This is the Hanning window coefficient. During the validation of the groove line validity, the second derivative of the average groove line on the mid-surface is calculated, and the number of peaks is detected. If the number of peaks is not 3, the longitudinal measurement range is expanded and the calculation is repeated.

[0135] S703: Extract the second derivative of the average groove line of the mid-surface with respect to the width coordinate as the average bending curvature, and construct the width weight accordingly to increase the weight of structural failure assessment in bending areas such as the junction of the vane rollers.

[0136] The average bending curvature is converted to a physical scale corresponding to the width coordinate using the total arc length parameter. The average total arc length parameter within the specified longitudinal measurement range is:

[0137]

[0138] in, This is the average total arc length parameter. This is the total arc length parameter. For the specified longitudinal measurement range, Here is the Hanning window coefficient. The average curvature is:

[0139]

[0140] in, The average bending curvature. To avoid numerical scale inconsistencies when directly incorporating the average bending curvature into weight calculations, the system first constructs a normalized average bending curvature:

[0141]

[0142] in, To normalize the average bending curvature, For the width coordinate integral variable, To prevent positive numbers with a denominator of zero, the width weight is:

[0143]

[0144] in, For width weights. During normalization validation, the sum of all width weights is calculated; if the sum deviates from 1 by more than 1... Then, normalization is performed again. Areas with a width weight greater than twice the mean are considered as bending areas at the junction of the airfoil rollers.

[0145] S704: The thickness gradient is spatially weighted and aggregated using the width weight to generate a comprehensive thickness value in a single-point dimension; and the real thickness field is redistributed in two-dimensional space according to the phase and width coordinates of the idler roller to generate a phase thickness map that characterizes the periodic stress coupling wear law.

[0146] The overall thickness value is obtained by spatially weighting and aggregating the thickness gradient using width weights. Therefore, the overall thickness value reflects the weighted thickness change level of the belt at the corresponding longitudinal displacement. The warning threshold for the overall thickness value is 0.3 mm / m; exceeding this value triggers a level-two maintenance warning. In the calculation of the phase thickness map, the idler roller phase is divided into 36 equal sub-boxes, and the two-dimensional spatial redistribution result is formed by averaging the sub-boxes.

[0147]

[0148]

[0149]

[0150] in, For the overall thickness value, For the first The set of longitudinal sampling points within each bin The number of sampling points within the set. For phase thickness map, The range is from 0 to 35. When When the value is 0, the system uses the phase thickness map of adjacent sub-boxes for linear interpolation. If both adjacent sub-boxes are empty, the sub-box is marked as an invalid region.

[0151] Example 7:

[0152] Figure 1The seventh and eighth flowcharts correspond to S801 to S804. This part obtains the historical reference thickness field of the previous measurement cycle, performs adaptive recursion based on the measurement residual of the real thickness field relative to the historical reference thickness field, and updates and generates the thickness measurement output set and reference thickness field of the current cycle while filtering transient environmental noise.

[0153] S801: Compare the actual thickness field of the current period with the historical reference thickness field of the previous measurement period to extract the measurement residual that reflects the characteristics of the new thickness deviation.

[0154] The conveyor belt running cycle is automatically identified by the joint characteristics. The longitudinal displacement difference between two adjacent joints is the belt length, which is the length of the running cycle. When the system is in the first measurement cycle, the measurement data of the first loop is processed by a 5-point median filter to remove the influence of joint areas and outliers, and this data is used as the initial historical reference thickness field. The initial thickness fluctuation parameter is set to 0.1 mm^2. Areas 500 mm before and after the joint area are marked as invalid areas and are not included in the calculation of the initial reference thickness field.

[0155]

[0156] in, For the first Measurement residuals for each measurement cycle For the current period's true thickness field, This is the historical reference thickness field for the previous measurement cycle.

[0157] S802: Comprehensively evaluate the relationship between the thickness fluctuation parameters of historical periods and the measurement residuals of the current period, and adjust the feedback gain to generate update coefficients to resist transient abnormal interference.

[0158] The upper and lower limits of the update coefficient are constrained to 0.01 to 0.5 to avoid abrupt changes in the reference thickness field or long-term inability to update due to a single measurement anomaly. The thickness fluctuation parameter is recursively derived based on the square of the measurement residual. Abnormal residuals are judged by comparing the absolute value of the measurement residual with three times the level after taking the square root of the thickness fluctuation parameter. When the abnormal residual condition is met, the update coefficient is set to the lower limit of 0.01.

[0159]

[0160] in, To measure the residual, This refers to the thickness fluctuation parameter over historical periods. The update coefficient is:

[0161]

[0162] in, The update coefficient is set at a lower limit of 0.01 to ensure that the reference thickness field has a minimum update capability, while the upper limit of 0.5 is used to limit the magnitude of a single update.

[0163] S803: The measurement residual is attenuated and adjusted using the update coefficient, and the compensation is superimposed on the historical reference thickness field to complete the smoothing and denoising recursion of the reference thickness field for the current period; the thickness fluctuation parameters are corrected by jointly using the update coefficient and the measurement residual, and the thickness error field characterizing the stability of the current thickness value is iteratively output.

[0164] When the thickness measurement error field exceeds 0.5mm, the update coefficient for that area is increased by 20% for priority correction; the increased update coefficient is still subject to the upper limit constraint. After the conveyor belt is replaced, the joint is repaired, or the system hardware is recalibrated, the reference thickness field and thickness fluctuation parameters are manually reset.

[0165]

[0166]

[0167]

[0168] in, This serves as the reference thickness field for the current period. The thickness fluctuation parameter for the current cycle. This represents the thickness measurement error field. In low-reliability regions, a reliability flag is added to the thickness measurement output set; 0 indicates high reliability, and 1 indicates low reliability.

[0169] S804: Packages and integrates the smoothed and denoised multidimensional attribute parameters to generate a thickness measurement output set.

[0170] Data transmission supports ModbusTCP and Profinet protocols. The JSON-formatted output includes timestamps, strip length, idler span, true thickness field array, phase thickness map array, comprehensive thickness value array, and thickness measurement error field array. Complete thickness measurement data for the past 30 days is stored locally, while historical periodic thickness measurement output sets are stored in the cloud, with a data retention period of one year. The real-time output frequency is 10Hz, outputting the comprehensive thickness value and thickness measurement error of the current section; the periodic output frequency is once per week, outputting a complete thickness measurement output set.

[0171] The belt length during the conveyor belt's operating cycle is:

[0172]

[0173] in, For the length of the belt, and This represents the longitudinal displacement corresponding to two adjacent joint features. The thickness measurement output set is:

[0174]

[0175] in, For thickness measurement output set, For timestamps, For the length of the belt, This serves as the reference thickness field for the current period. For the current period's true thickness field, For phase thickness map, For the overall thickness value, For the thickness measurement error field, This refers to the idler span value. The thickness measurement output set is used to provide the host system with the reference thickness field, actual thickness field, phase thickness map, comprehensive thickness value, thickness measurement error field, belt length, timestamp, and idler span value for the current cycle.

[0176] Through the above process Figure 1 A continuous data link is formed between each process frame: the speed measurement unit provides longitudinal displacement, the upper three-dimensional contour scanning unit and the lower three-dimensional contour scanning unit provide double-sided point cloud fields, the double-sided point cloud fields are processed by the idler phase, width coordinates and initial normal thickness to obtain the initial thickness field, and then the periodic deformation caused by the idler support is stripped by the periodic disturbance model. Finally, the average bending curvature and measurement residual are combined to generate the thickness measurement output set and the reference thickness field.

[0177] Example 8:

[0178] The station coordinate system uses the ground of the charging station as the reference plane, selects the ground corner point at the entrance of the charging station as the origin, takes the horizontal direction parallel to the long side of the main building of the charging station as the positive x-axis, the horizontal direction perpendicular to the x-axis as the positive y-axis, and the direction perpendicular to the ground upward as the positive z-axis. The panoramic coordinate system is the ground plane projection coordinate system of the station coordinate system, retaining only the x-axis and y-axis components. The panoramic coordinates correspond one-to-one with the ground plane coordinates of the station coordinate system, without the need for additional conversion.

[0179] The local coordinate system takes the lower left corner of the physical boundary of the corresponding workstation unit as its origin, the direction parallel to the long side of the parking space as the positive x-axis, and the direction parallel to the short side of the parking space as the positive y-axis. The transformation between the local coordinate system and the station coordinate system is achieved through a preset translation and rotation matrix, and the corresponding transformation formula is as follows:

[0180]

[0181] in, Let be the rotation matrix of the local coordinate system relative to the station coordinate system. It is a translation vector. Coordinates in the local coordinate system The coordinates are in the station coordinate system; the ground coordinates output by the plane mapping table are two-dimensional ground plane coordinates in the station coordinate system, and the panoramic coordinates output by the panoramic mapping table are two-dimensional coordinates in the panoramic coordinate system. The two have the same dimension and are applicable to target physical positioning and panoramic image stitching, respectively.

[0182] When generating the planar mapping table and the panoramic mapping table, at least four non-collinear structural anchor points are selected as calibration references for each workstation unit. Based on the pinhole camera model and the homography matrix solution method, the mapping relationship from image pixels to physical coordinates is calculated, and the homography matrix... The solution formula is:

[0183]

[0184] in, The image pixel coordinates are in homogeneous form. The physical coordinates are in homogeneous form. After obtaining the homography matrix by solving through at least 4 sets of anchor point correspondences, the corresponding mapping table is generated and stored by traversing all pixel positions of the image.

[0185] The reverse mapping is implemented using a lookup table interpolation method. For a given panoramic coordinate, the corresponding image pixel grid position is first found through the reverse index of the mapping table, and then the sub-pixel precision image pixel coordinates are calculated through bilinear interpolation. The reverse addressing formula of the panoramic mapping table is as follows:

[0186]

[0187] in, A pre-generated reverse mapping lookup table stores the correspondence between panoramic coordinates and image pixel coordinates.

[0188] The structural anchor points are selected from points in the fixed structure that have obvious corner features and unique spatial locations, including the top corner of the anti-collision post, the bottom corner of the charging terminal, the corner endpoint of the parking space line, and the edge inflection point of the pile side gun seat. All of the above points meet the requirements of repeatability and fixed position.

[0189] The automatic extraction of structural anchor points adopts the Harris corner detection algorithm. First, corner points are detected in the corrected image to obtain candidate corner points. Then, they are matched with the preset physical coordinates of fixed structures. The candidate corner points that are successfully matched are the valid structural anchor points. For anchor points that fail to be detected automatically, they are supplemented by manual annotation. When annotating, the pixel center position of the corner point is selected as the pixel coordinate of the anchor point. The physical coordinates in the corresponding field coordinate system are obtained by field measurement. The pixel coordinates and physical coordinates of each anchor point are stored in a one-to-one correspondence.

[0190] Example 9:

[0191] The preset coverage ratio for the camera and workstation unit is 0.6. The system determines when the camera covers the corresponding workstation unit and completes the binding process.

[0192] Preset time threshold for time slices A value of 0.033 corresponds to the single-frame time interval at a capture frame rate of 30fps, ensuring that all video frames within the same time slice are in the same capture cycle. Time slices are generated using a fixed-step sliding window method, with the sliding step size consistent with the video capture frame rate; that is, a new time slice is generated for each new video frame captured. The duration of each time slice is related to a preset time threshold. Maintain consistency.

[0193] The time reference for the time slice is the frame timestamp of the main camera; for other cameras, the timestamp difference from the reference timestamp is no greater than [value missing]. The most recent video frame is included in the current time slice; when a certain camera loses a frame, the previous valid frame closest to the reference timestamp of that camera is selected as the replacement frame and included in the time slice, and the frame is marked as a frame loss compensation frame. If the number of consecutive frame losses exceeds 3 frames, the camera is marked as unavailable in the corresponding time slice and will not participate in the subsequent fusion calculation.

[0194] Lens distortion correction uses the Brownian distortion model, which calculates the remapping coordinates of each pixel based on the camera's intrinsic parameters and distortion parameters. The resampling process uses a bilinear interpolation algorithm to ensure that the edges of the corrected image are smooth and free of jagged edges, and that the geometry of the fixed structure after correction is consistent with the physical shape in the field coordinate system.

[0195] When cropping a local image, the physical boundary vertices of the workstation unit are first converted into pixel coordinates in the image using a planar mapping table to obtain the polygonal projection area of ​​the physical boundary in the image. Then, the corrected image is cropped along the polygonal boundary. The pixels within the polygonal area are the local images of the corresponding workstation unit. The cropped local images retain the pixel coordinate system of the original image and correspond one-to-one with the physical boundary of the workstation unit.

[0196] Reference luminance normalized The image is obtained through offline camera calibration. During calibration, the camera is pointed at a standard grayscale chart to capture images. The average brightness of the grayscale chart area is taken as the reference brightness of the camera. When calculating the average brightness, pixels with grayscale values ​​less than 10 and greater than 245 in the image that are too dark or overexposed are removed. Only the average value of pixels in the effective grayscale range is calculated.

[0197] Example 10:

[0198] The background template of the workstation unit is generated by multi-frame fusion. Ten consecutive frames of idle workstation images without dynamic targets are selected, and the median value of the corresponding pixels in each frame is taken as the pixel value of the background template. The generated background template is aligned with the local coordinate system, and each workstation unit has an independent background template.

[0199] The background template uses a periodic update strategy. Every 30 minutes, a frame of workstation image without dynamic targets is selected to perform a weighted update on the background template. The update formula is as follows:

[0200]

[0201] in, We set 0.9 as the background weight coefficient. These are the pixel values ​​of the original background template. The image pixel values ​​where there are currently no moving targets. These are the updated background template pixel values.

[0202] Background difference overlap texture determination preset value Set the value to 15, which corresponds to the pixel difference threshold of an 8-bit grayscale image. When the absolute value of the difference between pixel positions is not greater than 15, it is determined to be an overlapping texture and the erasure is performed.

[0203] The connection component determination of dynamic structures adopts the 8-connectivity rule, that is, the top, bottom, left, right and four diagonal directions of a pixel are all considered connected. The minimum area threshold of dynamic structures is 30 pixels. Connected components with an area smaller than this threshold are judged as noise and are removed.

[0204] After dynamic structure extraction, morphological post-processing is performed. First, a 3x3 rectangular structuring element is used to perform an erosion operation to remove noise points. Then, the same structuring element is used to perform an expansion operation to fill the holes inside the connected domain, finally obtaining the complete dynamic structure region.

[0205] In the calculation of occlusion ratio, the range of occluded pixels It includes three types of pixels: first, fixed structure pixels covered by dynamic structures; second, dark area pixels at the edge of the field of view without effective image pixels; and third, blurry area pixels where structural anchor points cannot be identified. These three types of pixels together constitute the occlusion pixel range, which is used to calculate the occlusion ratio within the corresponding window.

[0206] Example 11:

[0207] The continuous edge extraction adopts the Canny edge detection algorithm. First, the grayscale image of the dynamic structure region is smoothed by Gaussian filtering. Then, the gradient magnitude and direction of the image are calculated. The thinned edge is obtained by non-maximum suppression. Finally, the continuous edge pixels are obtained by double thresholding.

[0208] The strip region recognition uses a strip detection algorithm to perform directional filtering on dynamic structural regions and extract strip-shaped pixel regions with a width of 3 to 15 pixels and a length of more than 30 pixels as strip regions.

[0209] The central skeleton extraction uses the Zhang-Suen parallel thinning algorithm to perform thinning processing on continuous edges and strip regions respectively, resulting in skeleton lines with a single pixel width. Then, the skeleton lines are broken and short branches are removed to finally obtain continuous candidate line segments. Each candidate line segment records the coordinates and direction information of its two endpoints.

[0210] The image localization of the pile-side gun seat adopts the template matching method. A standard image of the pile-side gun seat is pre-acquired as a matching template. Normalized cross-correlation template matching is performed in the pile body area of ​​the local image. The area with the highest matching degree and greater than 0.8 is the image position of the pile-side gun seat. The center pixel of this area is taken as the positioning point of the pile-side gun seat.

[0211] The vehicle-side area is divided based on the physical boundary of the workstation unit. The boundary away from the pile-side gun seat is the vehicle entrance side. The corresponding one-third area away from the pile-side gun seat in the local image is defined as the vehicle side, which is used to determine whether the candidate line segment extends into the vehicle area.

[0212] Set of fixed structural extension directions The results were obtained by fitting straight lines to the edges of the fixed structure within the workstation unit, using the Hough linear transform to extract straight line segments from the fixed structure, calculating the direction angle of each straight line, and categorizing them into a set. The direction angle is calculated with the positive x-axis of the image as the reference, and counterclockwise rotation is positive.

[0213] Angle normalization function Its function is to normalize the angle difference to The interval is calculated using the following formula:

[0214]

[0215] in, This is a rounding function that ensures that angle differences are compared within a uniform range.

[0216] The minimum directional difference between the candidate line segment and the extension direction of the fixed structure is preset to be... ,when The candidate line segment is determined to be not parallel to the extension direction of the fixed structure.

[0217] The method for independently determining the pixel position of the fixed end is as follows: first, the overall area of ​​the positioning pile side gun seat is matched by template, and then edge detection is performed at the outlet position of the gun seat area. The midpoint of the lower edge of the outlet is extracted as the pixel coordinate of the fixed end. This positioning process does not depend on the position information of the candidate line segment.

[0218] The method for independently determining the pixel position of the vehicle-side charging terminal is as follows: First, the charging gun head feature is detected in the dynamic vehicle-side region. The charging gun head region is identified by combining the directional gradient histogram feature with the support vector machine classifier. The center point of the contact position between the charging gun head region and the vehicle body is taken as the pixel coordinate of the vehicle-side charging terminal. This localization process also does not depend on the position information of the candidate line segments.

[0219] After independently determining the two endpoints, they are matched with the candidate line segments. The pixel distance from the two endpoints to the candidate line segments is calculated. The direction deviation is combined to determine whether the candidate line segment is the center line of the gun line. After successful matching, the position of the candidate line segment closest to the two endpoints is taken as the fixed end of the center line of the gun line and the charging end on the vehicle side, respectively.

[0220] Example 12:

[0221] Image features To obtain the gradient magnitude feature of a grayscale image, the gradients in the x and y directions are calculated using the Sobel operator, and the gradient magnitude is taken as the image feature value. The calculation formula is as follows:

[0222]

[0223] in, The gradient value in the x-direction. The gradient value in the y-direction is used for both structural repeatability calculation and panoramic feature aggregation. The image features used for these two methods are the same gradient magnitude features.

[0224] In image blur calculation, pixel position is used. Centered pixel window Using a 5x5 square pixel neighborhood, normalized parameters Set the value to 256 to adjust the image blur level. The output range falls within Interval.

[0225] In structural repeatability calculation, displacement set Take integer displacement combinations of ±3 pixels in both the horizontal and vertical directions. The correlation calculation uses a normalized cross-correlation algorithm, and the corresponding calculation formula is as follows:

[0226]

[0227] in, This represents the mean of the features of the original image within the window. This represents the mean of the image features after displacement within the window.

[0228] Normalized parameters in endpoint fit calculation Set to 50, which corresponds to the normalized baseline for pixel distance; the angle normalization parameter in the direction deviation calculation. Pick This is used to control the decay rate of directional deviation.

[0229] Preset positive coefficients in the calculation of fusion intermediate values Take 1.2, Take 0.8, The value of 0.6 corresponds to the positive gain coefficient of the topological eigenvalue and the negative penalty coefficients of structural redundancy and image blur, respectively.

[0230] Camera coverage set The criteria for determination are: panoramic coordinates Cameras that fall within the effective field of view of a camera through reverse mapping, and whose image blur at that location is less than 0.7 and occlusion ratio is less than 0.5, are included in the coverage set. .

[0231] Panoramic pixel fusion performs weighted fusion on the three RGB channels of a color image, with each channel independently calculating the fusion weight and weighted pixel value, and finally synthesizing the color panoramic pixels.

[0232] Example 13:

[0233] Preset positive weights in splicing traversal cost calculation Take 3.0, Take 1.5, A value of 0.5 is used to assign cost weights to crossing gun lines, dynamic structures, and fixed structures, respectively, ensuring that crossing gun lines has the highest priority penalty; the preset value for determining crossing cost. A value of 0.8 is used to filter splicing paths that meet the requirement of low traversal cost.

[0234] The width of the buffer zone at the physical boundary is 0.3 times the width of the overlapping area of ​​adjacent workstations, and the splicing path is only selected within this buffer zone; the preset positive number in the observation point offset calculation. The value is set to 2, which is used to control the offset distance between the observation point and the target coordinates.

[0235] The logic for the combination of splicing seams and fusion weights is as follows: the non-overlapping areas of adjacent workstations are directly filled with image pixels from a single camera, while the overlapping areas are weighted and fused using fusion weights to determine the splicing seam as the path with the lowest crossing cost within the overlapping area. The non-overlapping areas on both sides of the splicing seam are respectively assigned to the corresponding workstation camera, and the overlapping zone where the splicing seam is located is smoothly transitioned using fusion weights.

[0236] The seam serves only as a boundary reference between overlapping and non-overlapping areas. All panoramic pixels within the overlapping area are generated through a weighted aggregation method with fusion weights, eliminating hard switching boundaries and ensuring a natural transition at the stitching point without obvious stitching marks.

[0237] The dynamic panoramic image adopts a planar unfolding form, using the ground plane of the site coordinate system as the unfolding reference, and unfolds the image of each workstation unit according to its physical spatial position to the panoramic coordinate system; the mapping between the physical boundary of the workstation and the panoramic pixel coordinates adopts a fixed scale, with the scale set to 100 panoramic pixels per meter. The pixel position of the workstation unit in the panoramic coordinate system is calculated based on the site coordinates of the physical boundary of the workstation unit, and the display area of ​​each workstation unit in the panoramic image is determined.

[0238] The arrangement logic of multiple workstations in the panoramic view is completely consistent with the actual spatial layout of the site. Adjacent workstation units are arranged in sequence according to their physical location, and the workstations maintain the same spacing distance as in reality, ensuring that the spatial layout of the panoramic view completely corresponds to the charging station site.

[0239] The "dynamic" attribute of the dynamic panoramic image has two aspects: First, the image content is updated in real time with each time slice. Each time a new time slice is generated, the panoramic pixels are updated once. The update frequency is consistent with the video capture frame rate, so as to realize the real-time dynamic display of the site image. Second, it supports dynamic perspective changes. The terminal can adjust the display perspective and scaling ratio of the panoramic image through rotation and zoom operations to view the detailed images of different areas of the site.

[0240] The real-time update of the dynamic panoramic image adopts an incremental update mechanism, which only updates pixels in the workstation unit area with dynamic structure, while retaining the original pixel values ​​in static areas without dynamic targets, thereby reducing the amount of computation and improving the update efficiency.

[0241] Example 14:

[0242] The 3D site model of the charging station is built based on the actual size of the site. It uses a lightweight 3D modeling method to generate 3D models of fixed structures such as the site ground, charging terminals, anti-collision posts, and parking lines. The size and spatial position of the model are completely aligned with the site coordinate system.

[0243] The texture mapping between the 2D panoramic pixels and the 3D model adopts the UV mapping method, which uses the dynamic panoramic image as a texture map and attaches it to the ground plane and fixed structure surface of the 3D model. The mapping coordinates correspond one-to-one with the station coordinate system to ensure that the texture position matches the spatial position of the 3D model.

[0244] The 3D rendering is implemented using the OpenGL graphics interface. By loading a 3D site model and dynamic panoramic texture, the 3D visualization rendering of the site scene is realized. The terminal can browse the 3D scene from different perspectives by adjusting the position and orientation of the virtual camera.

[0245] Anomaly detection employs a deep learning object detection algorithm with the YOLO network as the basic detection framework. The training dataset contains various anomaly samples in the charging station scenario, such as cables dragging on the ground, foreign objects occupying space, and facility damage. The detection input is a dynamic panoramic image of each workstation unit area.

[0246] The specific judgment rules for the three types of anomalies are as follows: The judgment rule for cable anomalies is that the vertical height of the center line of the charging gun is less than 0.1 meters, or the center line of the charging gun crosses the physical boundary of the adjacent work station unit, corresponding to the two abnormal forms of cable dragging on the ground and cable crossing the work station; the judgment rule for occupancy anomalies is that there are targets such as non-charging vehicles and debris occupying the parking space in the work station unit, and the target stays for more than 30 seconds; the judgment rule for facility anomalies is that the fixed structure such as charging terminal and anti-collision post has a morphological change of displacement, damage or missing, and the difference from the structural characteristics of the background template exceeds the set threshold.

[0247] The input carriers for anomaly recognition are the local images and dynamic structure extraction results corresponding to each workstation unit. First, abnormal targets are detected in the local images, and then mapped to the dynamic panoramic image for positioning and display.

[0248] Example 15:

[0249] For the ground projection coordinate error of suspended targets such as cable plumb points, a height correction method is used for compensation. First, the height of the cable plumb point above the ground is estimated using the tilt angle of the gun line centerline and the height information of the fixed end. Then, the height is corrected based on the projection results of the planar mapping table according to the camera's pitch angle and focal length. The correction formula is as follows:

[0250]

[0251] in, The height of the cable's perpendicular point from the ground. The camera's tilt angle. This is the horizontal offset direction vector.

[0252] The single-camera positioning method is applicable only when the target is located on the ground plane. It is only suitable for calculating the coordinates of ground targets such as those with abnormal occupancy or facilities. For suspended targets such as cables, the multi-camera aggregation method is preferred for calculating the coordinates. The multi-camera weighted aggregation method is applicable to targets covered by at least two cameras. Targets covered by a single camera are located using a single-plane mapping table. The single-camera result for selecting the perpendicular point of a cable is applicable when only a single camera can observe the complete center line of the gun line. In other scenarios with multi-camera coverage, the weighted aggregation method is used to calculate the target coordinates.

[0253] The positive direction of the y-axis in the image coordinate system is vertically downward. The y-coordinate value is the smallest at the top of the image and the largest at the bottom. Therefore, the point with the largest y-coordinate is the lowest point in the vertical direction of the image, which corresponds to the perpendicular point of the cable.

[0254] For cameras with a tilt angle, a tilt angle correction method is used to adjust the calculation of the cable perpendicular point. First, the vertical coordinates of the image are converted into the actual vertical direction based on the camera's tilt angle. Then, the lowest point in the vertical direction is taken as the cable perpendicular point. The correction formula is as follows:

[0255]

[0256] in, The vertical pixel coordinates of the image. The camera's tilt angle. Determine the installation height of the camera. These are the corrected actual vertical coordinates.

[0257] The calculation of the crossing boundary point uses the intersection of the physical projection line segment and the physical boundary. First, the pixel coordinates of the gun line centerline are converted into physical coordinates on the ground plane through a plane mapping table to obtain the gun line projection line segment in the physical space. Then, the intersection of the projection line segment and the physical boundary of the work station unit is calculated. The intersection point closest to the fixed end is selected as the crossing boundary point. The error caused by the suspension height is uniformly compensated by the height correction method of the cable plumb line.

[0258] Example 16:

[0259] The design logic for the observation orientation is to ensure that the virtual camera of the inspection terminal is aligned with the abnormal target along the camera's observation direction. The original reverse extension line logic has been modified to: the observation point is located on the line connecting the camera and the target coordinates, closer to the camera, and the observation direction is from the observation point to the target coordinates, which is consistent with the actual observation direction of the camera.

[0260] The observation orientation is represented in Euler angle form, which includes two rotational components: yaw angle and pitch angle. The positive x-axis of the panoramic coordinate system is taken as the 0-degree yaw angle reference, the positive y-axis as the 90-degree reference, and the horizontal direction as the 0-degree pitch angle reference, with upward as positive.

[0261] The jump command uses a structured data format, which includes six fields: workstation unit identifier, anomaly type, x and y components of target coordinates, and yaw and pitch components of observation orientation. It is encapsulated and transmitted in JSON format. After the terminal parses the command fields, it performs the corresponding view rotation and target centering operations.

[0262] The jumpable spatial association is triggered by a click. When the user clicks on the abnormal target icon in the dynamic panoramic image, the terminal automatically jumps to the magnified detail of the corresponding abnormal target. The target screen is a real-time video stream captured by the corresponding camera, with the center of the screen aligned with the location of the abnormal target.

[0263] The initial viewpoint of the panoramic field of view is a top-down view directly above the station, with a field of view angle of 90 degrees. The rotation reference coordinate system is the panoramic coordinate system, and the rotation operation is performed around the vertical and horizontal axes, supporting 360-degree horizontal rotation and ±90-degree vertical pitch adjustment.

[0264] The inspection terminal adopts a touch-screen tablet hardware form factor, is compatible with the Android operating system, receives dynamic panoramic image data and jump commands via wireless LAN, and renders images with a refresh rate of no less than 30 frames per second.

[0265] Example 17:

[0266] The inspection monitoring and management module, the anomaly binding unit, and the scheduling output unit are all deployed on local edge servers, adopting a centralized deployment architecture, and all data processing is completed on local servers.

[0267] The data interface between modules is implemented using internal function calls. The data flow is as follows: After the inspection monitoring and management module completes image acquisition, structural decomposition, feature extraction, weight calculation, panoramic stitching and anomaly recognition, it outputs the dynamic panoramic image and anomaly target information to the anomaly binding unit. After the anomaly binding unit completes workstation binding and jump instruction generation, it outputs the jump instruction to the scheduling output unit. The scheduling output unit then sends the data to the inspection terminal.

[0268] The index table is stored in the form of a structured data table, which includes four sub-tables: camera index table, workstation unit index table, structural anchor point index table, and abnormal target index table.

[0269] The camera index table stores camera coordinates, equipment parameters, and bound workstation identifiers. The mapping table stores path fields. The workstation unit index table stores workstation identifiers, physical boundary coordinates, a list of fixed structures, and background templates. The structural anchor point index table stores anchor point identifiers, corresponding workstation identifiers, pixel coordinates, and physical coordinate fields. The abnormal target index table stores abnormal identifiers, corresponding workstation identifiers, abnormal types, target coordinates, and observation orientation fields. The index tables are stored in a local database and the corresponding field contents are updated in real time as the system runs.

[0270] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A 3D contour scanning thickness measurement system for conveyor belts, applied to online thickness measurement operations comprising an upper three-dimensional contour scanning unit and a lower three-dimensional contour scanning unit arranged on a frame, and a speed measurement unit, characterized in that, include: The longitudinal displacement of the conveyor belt is tracked using the data from the speed measurement unit, and the upper three-dimensional contour scanning unit and the lower three-dimensional contour scanning unit are simultaneously triggered to acquire the upper contour point cloud and the lower contour point cloud, so as to construct a dual-sided point cloud field that eliminates the interference of time sampling misalignment. The lateral curvature caused by the support structure in the double-sided point cloud field is analyzed to extract the idler span and generate the idler phase characterizing the periodic sag of the idler support. Extract the centerline of the current measurement section corresponding to the dual-sided point cloud field, flatten the contour along the centerline of the section, and establish the width coordinates for correcting lateral deviation drift. The initial normal thickness is extracted along the centerline of the cross section in the local normal direction to construct an initial thickness field that reflects the true spatial thickness properties of the belt. A periodic disturbance model is constructed based on the idler roller phase to separate and remove the periodic deformation caused by the idler roller support from the initial thickness field, and to restore the true thickness field. The average bending curvature of the belt in the lateral direction is combined to perform local structural weight allocation on the true thickness field to generate a thickness feature set that reflects the structural wear tendency of the edge and the bottom of the groove; wherein, the thickness feature set includes a comprehensive thickness value and a phase thickness map; The historical reference thickness field of the previous measurement cycle is obtained. Based on the measurement residual of the real thickness field relative to the historical reference thickness field, adaptive recursion is performed. While filtering transient environmental noise, the thickness measurement output set and reference thickness field of the current cycle are updated and generated.

2. The 3D contour scanning thickness measurement system for conveyor belts according to claim 1, characterized in that, The construction of the bilateral point cloud field includes: The instantaneous speed of the conveyor belt is obtained through the speed measurement unit, and the longitudinal displacement corresponding to the current sampling moment is calculated based on the kinematic transformation relationship between speed and displacement to characterize the actual running mileage of the belt material; At the current sampling time, the upper three-dimensional contour scanning unit and the lower three-dimensional contour scanning unit are triggered to acquire the upper contour point cloud and the lower contour point cloud at the same physical cross-section, respectively; The longitudinal displacement is used as a physical positioning label and spatially bound to the corresponding upper contour point cloud and lower contour point cloud to generate the double-sided point cloud field that is continuously expanded based on the running mileage of the belt material, so as to eliminate the longitudinal spatial distortion interference introduced by belt speed fluctuation.

3. The 3D contour scanning thickness measurement system for conveyor belts according to claim 2, characterized in that, The generation of the idler phase includes: By fusing the height features of the upper contour point cloud and the lower contour point cloud at the same lateral coordinate, a centerline height representing the overall undulation state of the current measurement section is fitted and generated. Extract the curvature variation characteristics of the centerline height within the effective bandwidth of the conveyor belt and perform lateral spatial aggregation to evaluate the lateral curvature, which characterizes the degree of lateral bending deformation at each longitudinal displacement. The transverse curvature is matched with longitudinal morphological similarity along the conveyor belt running direction, and the cycle length corresponding to the peak value of the matching correlation is determined as the idler span value; The physical location corresponding to the peak value of the lateral curvature is used as the support reference point. Based on the periodic proportion of the longitudinal displacement relative to the support reference point in the idler span value, the idler phase representing the idler in the sag state of the idler support is generated.

4. A 3D contour scanning thickness measurement system for conveyor belts according to claim 3, characterized in that, The establishment of the width coordinates includes: Based on the centerline height, local slope features are extracted, and the curve path length is accumulated inward along the effective data boundary of the current measurement section to calculate and obtain the arc length coordinates of the centerline of the section after eliminating the groove bending deformation. Determine the total arc length parameter after the centerline of the cross section is fully unfolded, and proportionally map the arc length coordinates to the two end boundaries to convert them into the width coordinates that are standardized within a specified reference interval; The discrete points of the upper contour point cloud and the lower contour point cloud are resampled using the width coordinates. The bilateral height data in the fixed equipment coordinate system are transformed into the width coordinate system composed of the longitudinal displacement and the width coordinates to correct the thickness measurement position drift error caused by lateral snake deviation.

5. A 3D contour scanning thickness measurement system for conveyor belts according to claim 4, characterized in that, The construction of the initial thickness field includes: In the width coordinate system, the resampled bilateral height data is reconstructed into a spatial surface to generate an upper surface and a lower surface that reflect longitudinal displacement and width changes, respectively, and the spatial mean layer of the two is extracted as the mid-surface surface. Orthogonal tangential features are extracted along the longitudinal and transverse directions of the mid-surface surface, and the mid-surface normal vector representing the local real spatial attitude of the band body is determined based on the orthogonal normal constraint. Extract the spatial difference vector between the upper surface and the lower surface, and project the spatial difference vector into the orientation direction of the mid-surface normal vector to separate the initial normal thickness perpendicular to the local strip surface; The initial normal thickness is assembled with the corresponding width coordinates and the idler phase into an attribute matrix to construct the initial thickness field, thereby eliminating the geometric measurement error caused by the vertical height difference deviating from the true normal thickness.

6. A 3D contour scanning thickness measurement system for conveyor belts according to claim 5, characterized in that, The restoration yields the true thickness field, including: The initial normal thickness in the initial thickness field is decoupled into three independent physical components: the true thickness field, the phase disturbance field representing the periodic deformation of the idler roller, and the occasional measurement noise. With the constraints of balancing the complexity of the disturbance model and the convergence of the residual, the optimal order for characterizing the deformation features of the idler support is determined by adaptive optimization. At the optimal order, the initial normal thickness at different width coordinates is expanded in the time-frequency domain using periodic basis functions to extract the phase disturbance field that is synchronous with the phase of the idler roller. The phase disturbance field is stripped from the initial normal thickness, and the periodic deformation caused by the mid-span sagging of the idler roller and rubber indentation is eliminated, restoring the true thickness field that reflects the physical wear state.

7. A 3D contour scanning thickness measurement system for conveyor belts according to claim 6, characterized in that, The generation of the thickness feature set includes: The spatial variation rates of the true thickness field in the longitudinal displacement and width coordinates are extracted and fused to generate a thickness gradient characterizing the local wear gradient of the belt. Spatial smoothing operation is performed on the mid-surface along the specified longitudinal measurement range to extract the average groove line of the mid-surface that reflects the transverse morphology of the solid support structure of the conveyor belt. The second derivative of the average groove line of the mid-surface with respect to the width coordinate is extracted as the average bending curvature. Based on this, a width weight is constructed to increase the weight of structural failure assessment in bending areas such as the junction of the vane rollers. The thickness gradient is spatially weighted and aggregated using the width weight to generate the comprehensive thickness value in a single-point dimension; and the real thickness field is redistributed in two-dimensional space according to the idler phase and the width coordinate to generate the phase thickness map characterizing the periodic stress coupling wear law.

8. A 3D contour scanning thickness measurement system for conveyor belts according to claim 7, characterized in that, The process of updating and generating the thickness measurement output set and reference thickness field for the current period includes: By comparing the actual thickness field of the current period with the historical reference thickness field of the previous measurement period, the measurement residual reflecting the characteristics of the newly added thickness deviation is extracted. The relationship between the thickness fluctuation parameters of historical periods and the measurement residuals of the current period is comprehensively evaluated, and the feedback gain is adjusted to generate update coefficients to resist transient anomaly interference. The measurement residual is attenuated and adjusted using the update coefficient, and the compensation is superimposed on the historical reference thickness field to complete the smoothing and denoising recursion of the reference thickness field for the current period. The thickness fluctuation parameter is simultaneously corrected by using the update coefficient and the measurement residual, and the thickness measurement error field characterizing the stability of the current thickness value is iteratively output; the smoothed and denoised multidimensional attribute parameters are packaged and integrated to generate the thickness measurement output set.

Images