An online detection system for milk powder tank weld quality and a control method thereof

Abstract

The application provides a milk powder tank weld quality online detection system and a control method, relates to the weld quality measurement and control technical field, and comprises an adjusting module, which is used for intercepting section data corresponding to a tin layer poor mutation section from a phase angle step distribution spectrum according to a weld quality detection result, performing point-by-point accumulation on phase angle difference values in the section data, obtaining a phase angle difference value accumulation amplitude, determining a pressure compensation coefficient according to the phase angle difference value accumulation amplitude, and generating a feeding pressure dynamic adjustment signal of an electrode wheel according to the pressure compensation coefficient. The application can reduce the influence of the tank body camber reflection on the micro-morphology mutation recognition.

Description

Technical Field

[0001] This invention relates to the field of weld quality measurement and control technology, and in particular to an online detection system for the weld quality of milk powder cans and its control method. Background Technology

[0002] In automated milk powder can manufacturing lines, the longitudinal seams of the can body are mostly formed by resistance welding. The overlapping edges are then immersed in a liquid tin bath to form a protective tin layer, ensuring the can's airtightness and food safety. To guarantee the quality of the weld, the production line needs to monitor the forming status of the tin layer at the overlapping edges of the longitudinal seams in real time and conduct online evaluation of the surface morphology of the liquid tin bath cooling section.

[0003] In practice, a high-speed milk powder can production line uses laser contour scanning to perform non-contact measurements on the overlapping edges of longitudinal seams, attempting to determine the quality of the tin layer coverage by acquiring changes in the surface height of the section. However, when the liquid tin bath cools, the overlapping edges may exhibit abnormally poor tin layer conditions and tiny depressions due to differences in surface tension and cooling rate. These depressions, along with adjacent smooth transition sections, exhibit subtle curvature changes and small spatial scales. During rapid scanning, laser contour scanning is hampered by reflections from the can's curved surface and in-situ vibrations, making it difficult to reliably separate these minute curvature abrupt changes from the height data. Consequently, some tin-deficient depressions are not detected in time.

[0004] On milk powder can production lines, electrode wheels are mounted on specialized resistance seam welding machines to weld the overlapping edges of rolled cylindrical steel plates. However, existing detection methods cannot effectively extract the phase angle step characteristics representing abrupt changes in the morphology of the tin layer cooling section from the impedance information of the electrode wheels themselves. This results in the electrode wheel pressure not being able to be adjusted online, and the recessed areas not being adequately compensated in real time. Consequently, weak points in the sealing of the can are prone to appear during the rolling process, posing a risk of hidden leakage. Therefore, existing solutions have limitations in identifying minute morphological changes in the cooling section of the liquid tin pool at the longitudinal seam overlapping edge, and lack a closed-loop mechanism to effectively correlate morphological change information with real-time adjustment of the electrode wheel feed pressure. Consequently, the online control of the consistency of the weld tin layer quality needs further improvement. Summary of the Invention

[0005] This invention provides an online inspection system and control method for the weld quality of milk powder cans, which can reduce the impact of reflection from the curved surface of the can on the identification of minute morphological changes.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: An online inspection system for the weld quality of milk powder cans includes: The acquisition module is used to acquire the impedance phase angle sampling point array when the electrode wheel sweeps across the overlapping edge of the longitudinal seam of the milk powder can, and obtain the phase angle step distribution spectrum based on the phase angle difference between adjacent sampling points. The module is used to construct a three-dimensional discrete point cloud based on the phase angle step distribution spectrum, and to construct a local weighted scatter matrix based on the spatial coordinates of the target discrete point and its neighboring discrete points. The detection module is used to perform feature decomposition on the local weighted scatter matrix to obtain the main direction morphology bending index and Gaussian morphology index. Based on the main direction morphology bending index and Gaussian morphology index, a local morphology undulation characterization map is generated. Based on the local morphology undulation characterization map, the liquid tin pool cooling section is divided into a smooth transition section and a tin layer depletion abrupt section. The tin layer depletion abrupt section is marked as a weld quality abnormality, and the weld quality detection result is obtained. The adjustment module is used to extract the segment data corresponding to the tin layer depletion abrupt change segment from the phase angle step distribution spectrum according to the weld quality inspection results, accumulate the phase angle difference value in the segment data point by point to obtain the cumulative amplitude of the phase angle difference value, determine the pressure compensation coefficient according to the cumulative amplitude of the phase angle difference value, and generate a dynamic adjustment signal for the feed pressure of the electrode wheel according to the pressure compensation coefficient.

[0007] A control method for an online inspection system for the weld quality of milk powder cans includes: The impedance phase angle sampling point array is obtained when the electrode wheel sweeps across the longitudinal seam overlap edge of the milk powder can. The phase angle difference between adjacent sampling points is calculated based on the impedance phase angle sampling point array, and the phase angle difference is arranged along the spatial extension direction of the longitudinal seam overlap edge to obtain the phase angle step distribution spectrum. A three-dimensional discrete point cloud is constructed based on the phase angle step distribution spectrum, and a local weighted scatter matrix is ​​constructed based on the spatial coordinates of the target discrete point and its neighboring discrete points. The local weighted scatter matrix is ​​subjected to eigenvalue decomposition to obtain the main direction morphological curvature index and Gaussian morphological index. Based on the main direction morphological curvature index and Gaussian morphological index, a local morphological undulation characterization map is generated. Based on the local morphological fluctuation characterization map, the cooling section of the liquid tin pool is divided into a smooth transition section and a tin-poor abrupt change section. The tin-poor abrupt change section is marked as a weld quality abnormality, and the weld quality inspection results are obtained. Based on the weld quality inspection results, the data corresponding to the tin layer depletion abrupt change segment is extracted from the phase angle step distribution spectrum. The phase angle difference values ​​in the segment data are accumulated point by point to obtain the cumulative amplitude of the phase angle difference value. The pressure compensation coefficient is determined based on the cumulative amplitude of the phase angle difference, and a dynamic adjustment signal for the feed pressure of the electrode wheel is generated based on the pressure compensation coefficient.

[0008] The above-described solution of the present invention has at least the following beneficial effects: This application collects the impedance phase angle when the electrode wheel sweeps across the overlapping edge of the longitudinal seam of a milk powder can, and calculates the phase angle difference between adjacent sampling points. This transforms the local impedance change caused by tin depletion into phase angle step information distributed along the overlapping edge of the longitudinal seam. The detection process does not rely on the optical reflection intensity of the can surface, thus reducing the impact of can surface reflection on the identification of subtle morphological changes. Based on the phase angle step distribution spectrum, a three-dimensional discrete point cloud of morphology is constructed. Combined with the local weighted scatter matrix, the principal direction morphological curvature index, and the Gaussian morphological index, a local morphological undulation characterization map is formed. This allows for the identification of tin using neighborhood spatial relationships. The local morphological differences between the tin-deficient depression and the smooth transition region reduce the interference of fluctuations at individual sampling points and in-situ jitter on the detection results, thereby improving the reliability of online identification of tin-deficient abrupt segments. By accumulating the phase angle difference values ​​corresponding to the tin-deficient abrupt segments point by point, the cumulative amplitude of the phase angle difference value characterizing the degree of abnormality can be obtained. Based on this, the pressure compensation coefficient can be determined and the dynamic adjustment signal of the electrode wheel feed pressure can be generated. This allows the weld quality detection results to be directly used for electrode wheel pressure adjustment, thereby improving the welding quality consistency at the longitudinal seam overlap of the milk powder can and reducing the risk of weak sealing points caused by tin-deficient depressions. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of an online inspection system for the weld quality of milk powder cans provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of the control method of an online inspection system for weld quality of milk powder cans provided by an embodiment of the present invention.

[0011] Figure 3 This is a flowchart illustrating the acquisition module provided in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the described embodiments are used to illustrate this application and are not intended to limit the scope of protection of this application. In the embodiments of this application, "first" and "second" are used to distinguish different objects and do not indicate the order or importance of the objects. "Multiple" refers to two or more. "Including" indicates an open inclusion relationship. The same reference numerals in the figures indicate the same or similar components.

[0013] The longitudinal seam overlap edge referred to in this application embodiment refers to the overlap area of ​​the two side plates extending along the axial direction of the can body after the milk powder can body plate is rolled into a cylindrical shape; the electrode wheel applies feed pressure and welding current to the longitudinal seam overlap edge, and the original tin layer in the overlap area is heated to form a local liquid tin pool, which enters a cooling state after the electrode wheel leaves; the liquid tin pool cooling section refers to the spatial section where the liquid tin pool solidifies and forms a tin layer along the longitudinal seam overlap edge; the tin layer depletion abrupt change section refers to the section in the cooling section where the tin layer coverage is relatively insufficient and forms a continuous abrupt change feature in the local morphology; the impedance phase angle is the phase difference between the voltage signal and the current signal in the AC detection circuit formed by the electrode wheel and the longitudinal seam overlap edge, and its value can be expressed in angle or radian, with the same unit used for the same batch of tests.

[0014] The impedance phase angle is affected by the contact state between the electrode wheel and the edge of the longitudinal seam, the material resistance of the overlapping area, and the distribution of the surface tin layer. When the local tin layer is poor and a depression is formed, the contact area and current path change with the spatial position, and the impedance phase angle change between adjacent sampling points will show continuous or paired step changes. If only a single impedance phase angle sampling value is observed, electrode wheel vibration, tank roundness deviation and slow drift of the detection circuit may also cause baseline changes, making it difficult to separate local abrupt changes from overall drift. This application first forms a phase angle step distribution spectrum, then uses the spatial relationship of neighboring points to extract the morphological bending features, and converts the degree of anomaly into a dynamic adjustment signal of feed pressure, thereby establishing a processing chain between the detection results and the electrode wheel pressure control.

[0015] System Implementation Examples See Figure 1 The online inspection system for the weld quality of milk powder cans provided in this application embodiment can be installed in the control cabinet of the resistance seam welding machine, or it can be implemented by an industrial computer and the seam welding machine controller located next to the seam welding machine. The system includes an acquisition module, a construction module, a detection module and an adjustment module. The acquisition module is communicatively connected to the impedance acquisition circuit and the position encoder. The adjustment module is communicatively connected to the servo driver of the electrode wheel pressure actuator. The construction module and the detection module can be implemented by a processor executing a program in the memory, or they can be implemented by a digital signal processing circuit, a programmable logic device or a combination thereof.

[0016] In one feasible deployment, the impedance acquisition circuit includes an AC detection signal source, a voltage sampling channel, a current sampling channel, and a phase detection channel. The AC detection signal source applies a detection signal to the detection loop formed by the overlap edge of the electrode wheel and the longitudinal seam. The voltage sampling channel and the current sampling channel synchronously acquire the loop voltage and loop current. The phase detection channel determines the phase positions of the voltage signal and the current signal within the same sampling period and calculates the difference between the two phase positions to obtain the impedance phase angle sampling value. The detection signal can be superimposed on the detection frequency band of the welding current or applied within the non-sampling window of the welding current. The controller reduces the influence of the welding current on the phase detection through frequency band separation or timing separation. Since the voltage signal and the current signal are sampled using the same clock, the obtained impedance phase angle can correspond to the spatial position of the electrode wheel.

[0017] A position encoder outputs pulses corresponding to the displacement of the milk powder can or electrode wheel. The acquisition module determines the longitudinal position according to the encoder pulses and the lateral sampling position according to the arrangement of the detection channels in the width direction of the longitudinal seam overlap edge. In one possible implementation, multiple impedance detection channels are arranged side by side in the lateral direction to obtain multiple lateral sampling values ​​at the same longitudinal position. In another possible implementation, a single impedance detection channel reciprocates with the lateral micro-displacement mechanism. The acquisition module determines the lateral sampling position based on the position feedback of the lateral micro-displacement mechanism and aligns the sampling values ​​at different scanning times to a common longitudinal sampling row through longitudinal position interpolation. Both of the above methods can form an impedance phase angle sampling point matrix with longitudinal and lateral order.

[0018] The acquisition module includes a sampling unit and a difference calculation unit. The sampling unit synchronously receives impedance phase angle sampling values, longitudinal position, and lateral sampling position. Sampling values ​​with the same longitudinal position or located in the same longitudinal sampling interval are arranged in a row, and sampling values ​​with the same lateral sampling position or located in the same lateral sampling interval are arranged in a column to obtain the impedance phase angle sampling matrix. When the position corresponding to the encoder pulse does not coincide with the preset sampling position, the sampling unit can select the closest effective sampling values ​​on both sides of the target position and perform linear interpolation based on the relative distance between the target position and the sampling positions on both sides. The phase angle obtained by interpolation is kept within the range limited by the effective sampling values ​​on both sides, thereby reducing the matrix misalignment caused by the fluctuation of the conveying speed.

[0019] The sampling unit can also perform phase continuity processing on the impedance phase angle. When the difference between adjacent original phase angles crosses the endpoint of the phase representation interval, the sampling unit adds or subtracts a complete phase cycle to the next original phase angle, so that the difference between the corrected adjacent phase angles is the smaller representation value of the same physical phase change. If the voltage or current amplitude of a sampling point is lower than the effective detection lower limit of the acquisition circuit, the sampling unit marks the sampling point as an invalid point and does not use the point to directly participate in the difference calculation. When there are valid points on both sides of the invalid point and the spatial distance between the two sides does not exceed the allowable interpolation distance, the sampling unit fills in the sampling value according to the relative distance. Otherwise, the invalid mark is retained and the point is skipped in the subsequent neighborhood construction, thereby avoiding phase wrap-around and false steps caused by low signal-to-noise ratio sampling.

[0020] The difference calculation unit selects two adjacent impedance phase angle sampling values ​​sequentially from each column of the impedance phase angle sampling point array along the spatial extension direction of the longitudinal seam overlap edge. The impedance phase angle sampling value with the change direction is obtained by subtracting the impedance phase angle sampling value with the change direction from the impedance phase angle sampling value with the change direction in the spatial order. The difference calculation unit associates the obtained phase angle difference value with the midpoint between the two sampling points and arranges all phase angle difference values ​​according to the longitudinal position and the transverse sampling position to obtain the phase angle step distribution spectrum. The difference between adjacent sampling points can weaken the common baseline that changes slowly along the longitudinal seam, while the phase abrupt change corresponding to the local tin layer morphology change is still retained in the phase angle step distribution spectrum, thus providing input that is insensitive to slow drift for subsequent local morphology analysis.

[0021] In some embodiments, the difference calculation unit can also scale the phase angle difference based on the actual distance between adjacent sampling points. Specifically, it first determines the ratio of the actual distance between adjacent sampling points to the nominal sampling interval, and then divides the phase angle difference by the ratio to obtain the phase angle difference corresponding to the nominal sampling interval. When the actual distance is zero, the position order is reversed, or the actual distance exceeds the allowable interval, the difference calculation unit marks the corresponding difference as an invalid value. This process makes the phase angle difference under different conveying speeds or a small number of missed samples comparable and avoids division with zero distance.

[0022] The construction module includes a point cloud construction unit and a matrix construction unit. The point cloud construction unit reads the longitudinal and lateral sampling positions corresponding to each effective phase angle difference value from the phase angle step distribution spectrum and converts the phase angle difference value into a shape height characterization value. In one implementation, the point cloud construction unit divides the longitudinal position by the longitudinal nominal sampling interval to obtain the longitudinal dimensionless coordinates, divides the lateral sampling position by the lateral nominal sampling interval to obtain the lateral dimensionless coordinates, and divides the phase angle difference value by the reference phase angle difference scale to obtain the shape height characterization value. The reference phase angle difference scale is determined by the statistical distribution of the absolute values ​​of phase angle differences in qualified weld samples and is a non-zero value. The longitudinal dimensionless coordinates, lateral dimensionless coordinates, and shape height characterization value constitute discrete point spatial coordinates. All discrete point spatial coordinates are collected to obtain a three-dimensional shape discrete point cloud.

[0023] In another implementation, a calibration relationship between phase angle difference and morphological height can be established using calibration samples with known variations in morphological height. The calibration process involves collecting phase angle difference values ​​at multiple known height levels, establishing a piecewise linear mapping according to the order of phase angle difference and known height, and using a point cloud construction unit to determine the calibration interval containing the target phase angle difference. Then, based on the relative distance from the phase angle difference to the two ends of the interval and the corresponding height values ​​at the two ends, the morphological height characterization value is determined. When the phase angle difference exceeds the calibration range, the point cloud construction unit takes the nearest endpoint height value, or marks the point as out-of-range and uses the endpoint height value in subsequent calculations. This calibration method unifies the three coordinate components into length quantities or dimensionless quantities normalized to the same scale, avoiding distortion of neighborhood distances caused by dimensional inconsistencies.

[0024] The matrix construction unit sequentially determines the effective discrete points in the 3D topographic discrete point cloud as target discrete points, and selects neighboring discrete points within a preset neighborhood centered on the target discrete point. The preset neighborhood can be a spherical neighborhood with coordinate distances not exceeding the neighborhood radius, or a rectangular neighborhood with longitudinal and lateral index differences not exceeding the corresponding window widths. The neighborhood radius or window width is set according to the minimum spatial scale that can cover the edge of the tin-poor depression, and ensures that the target discrete point in the normal point cloud can obtain at least three non-collinear neighboring discrete points. The target discrete point located at the point cloud boundary uses actual neighboring discrete points, without expanding the neighborhood by copying boundary points, thereby preserving the true spatial relationship at the boundary.

[0025] For each neighborhood discrete point, the matrix construction unit subtracts the corresponding coordinate components of the target discrete point from the vertical coordinate, horizontal coordinate, and topographic height of the neighborhood discrete point to obtain a coordinate difference vector. Then, it squares each of the three coordinate components and sums them, taking the square root of the sum to obtain the spatial distance from the neighborhood discrete point to the target discrete point. The matrix construction unit determines the neighborhood weights according to the rule that the neighborhood weights do not increase as the spatial distance increases. One implementation method is to use the ratio of the spatial distance to the neighborhood radius to represent the relative distance, and then subtract the square of the relative distance to obtain the neighborhood weight. Points outside the neighborhood boundary have zero weights. Another implementation method is to use the square of the ratio of the spatial distance to a preset scale to determine the attenuation amount, and then determine the neighborhood weights based on the negative exponent of the attenuation amount. The weight of the target discrete point can be set to the maximum weight, and all weights are non-negative values.

[0026] The matrix construction unit constructs a 3x3 locally weighted scatter matrix based on the coordinate difference vector and neighborhood weights. Specifically, the squares of the vertical, horizontal, and topographic height components of each coordinate difference vector are multiplied by their corresponding neighborhood weights and summed to obtain the three elements on the main diagonal of the matrix. The products of the vertical and horizontal components, the vertical and topographic height components, and the horizontal and topographic height components are multiplied by their corresponding neighborhood weights and summed to obtain three cross-dimensional elements. The same cross-dimensional elements are used for positions symmetrical about the main diagonal of the matrix. The sum of all effective neighborhood weights is then divided by the above summed results to obtain the locally weighted scatter matrix. Since the contribution of each point to the matrix decreases with increasing spatial distance, topographic changes far from the target discrete point are less likely to obscure small abrupt changes near the target discrete point.

[0027] When there are fewer than three effective neighborhood discrete points, or when the effective neighborhood discrete points are collinear in the coordinate space, or when the sum of the weights of all neighborhoods is zero, the matrix construction unit marks the target discrete point as a point with insufficient neighborhood, expands the neighborhood range once, and reselects the point; if the expanded neighborhood still does not meet the conditions, the target discrete point will not participate in the segment abrupt change discrimination, and its joint morphology label will be obtained by interpolation of the nearest effective target discrete points along the longitudinal distance; this boundary processing can avoid local directional instability caused by insufficient rank of the scatter matrix.

[0028] The detection module includes an eigenvalue decomposition unit, a map generation unit, and a segmentation unit. The eigenvalue decomposition unit performs eigenvalue decomposition on the local weighted scatter matrix, which is applicable to real symmetric matrices, to obtain three non-negative eigenvalues ​​and unit eigenvectors corresponding to the three eigenvalues ​​respectively. The eigenvalue decomposition unit arranges the eigenvalues ​​in descending order, determines the unit eigenvector corresponding to the largest eigenvalue as the first local direction of the target discrete point, determines the unit eigenvector corresponding to the second largest eigenvalue as the second local direction, and determines the unit eigenvector corresponding to the smallest eigenvalue as the local normal direction. The first local direction and the second local direction are located in the local tangent plane of the target discrete point and are orthogonal to each other.

[0029] To avoid directional jumps caused by the equivalence of positive and negative signs in the unit eigenvectors of adjacent target discrete points, the eigenvalue decomposition unit compares the first local direction of the current target discrete point with the first local direction of the previous effective target discrete point in the longitudinal direction. When the sum of the products of the corresponding components of the two directions is less than zero, the components of the current first local direction are inversely represented. The second local direction and the local normal direction are processed using the same direction continuity method. When the difference between the largest eigenvalue and the second largest eigenvalue is less than the direction stability threshold, the eigenvalue decomposition unit projects the first local direction of the previous effective target discrete point onto the current local tangent plane and normalizes the projection result as the current first local direction. This reduces the pseudo-bending caused by the interchange of the two tangent directions when the eigenvalues ​​are close.

[0030] The feature decomposition unit determines the main direction topography curvature index based on the change in the local direction of adjacent target discrete points along the longitudinal seam overlap edge. Specifically, for the current target discrete point, the previous effective target discrete point and the next effective target discrete point in the longitudinal direction are obtained respectively. The angle between the previous first local direction and the current first local direction, and the angle between the current first local direction and the next first local direction are calculated. The sum of the two angles is divided by the longitudinal distance between the previous and next target discrete points to obtain the first direction curvature change per unit length. The second local direction curvature change is calculated in the same way, and the larger of the two curvature changes is determined as the main direction topography curvature index. Target discrete points located on the longitudinal boundary are calculated using one-sided adjacent points. When the longitudinal distance is zero, the division is not performed and the point is marked as a positional anomaly.

[0031] In an implementation that preserves the concavity / convexity direction, the eigenvalue decomposition unit can also determine the sign of the bending change by utilizing the change in adjacent local normal directions. Specifically, the change component of the local normal direction in the first local direction is determined as the first directional bending change, and the change component of the local normal direction in the second local direction is determined as the second directional bending change. The eigenvalue decomposition unit determines the product of the first and second directional bending changes as the Gaussian morphological index, or records the product of their absolute values ​​and the direction sign together as the Gaussian morphological index. A product close to zero indicates that the local morphology is close to flat in at least one direction, a positive product indicates that the two orthogonal directions are bending in the same direction, and a negative product indicates that the two orthogonal directions are bending in opposite directions. Since tin-depleted depressions usually cause normal changes in both directions, the main direction morphological bending index and the Gaussian morphological index can jointly characterize the morphological changes at the edge and inside of the depression.

[0032] The atlas generation unit maps the main direction morphology bending index and the Gaussian morphology index to a preset value range, which can be a dimensionless range from zero to one. For the main direction morphology bending index, the atlas generation unit obtains the lower limit reference value and the upper limit reference value from the qualified weld calibration sample. It subtracts the lower limit reference value from the index to be mapped to obtain the first difference. It subtracts the lower limit reference value from the upper limit reference value to obtain the first range width. Then, it divides the first difference by the first range width to obtain the main direction bending characterization value. The result less than zero is restricted to zero, and the result greater than one is restricted to one. For the Gaussian morphology index, the atlas generation unit can use the same mapping process to obtain the Gaussian morphology characterization value for its absolute value and retain the positive and negative direction marks of the Gaussian morphology index separately.

[0033] The lower and upper reference values ​​can be determined using qualified weld calibration samples. The controller arranges the corresponding indices of the calibration samples according to their numerical values, and uses the value corresponding to the statistical position that is close to the low end and can exclude occasional minimum values ​​as the lower reference value, and uses the value corresponding to the statistical position that is close to the high end and can exclude occasional maximum values ​​as the upper reference value. When the upper reference value equals the lower reference value, the spectrum generation unit sets the corresponding characterization value to zero and outputs a recalibration prompt. By using the statistical range of the calibration samples for mapping, the indices under different batches of materials, detection frequencies, or sampling scales can enter a common discrimination interval, while avoiding a single extreme value from determining the entire mapping range.

[0034] The map generation unit determines the joint topography marker based on the main direction bending characterization value and the Gaussian morphology characterization value. One implementation method is to combine the main direction bending characterization value, the Gaussian morphology characterization value, the directional marker of the Gaussian morphology index, and the effective state of the target discrete point into a multi-channel joint topography marker. Another implementation method is to weight the main direction bending characterization value and the Gaussian morphology characterization value according to the first topography weight and the second topography weight respectively, and use the sum of the two weighted results as the joint topography intensity. The two topography weights are both non-negative values ​​and are determined by calibration samples with qualified markers and tin-depleted markers. The map generation unit arranges the joint topography markers according to the longitudinal position and transverse sampling position of each target discrete point in the liquid tin pool cooling section to obtain a local topography undulation characterization map. Since the map retains the spatial position of the target discrete point, isolated noise and continuous tin-depleted topography can be distinguished by spatial continuity.

[0035] The segmentation unit compares the main direction bending characterization value with the first discrimination threshold and the Gaussian shape characterization value with the second discrimination threshold. The first and second discrimination thresholds can be determined by calibration samples. Specifically, two types of characterization values ​​are collected from qualified weld calibration samples, and the characterization values ​​are arranged from smallest to largest. The value corresponding to the upper statistical position that can cover the normal fluctuations of qualified samples and exclude a small number of outliers is set as the initial threshold. Then, the initial threshold is verified using calibration samples with tin-deficient markers. When the number of missed samples exceeds the allowable number, the corresponding threshold is gradually reduced. When the number of false positive samples exceeds the allowable number, the corresponding threshold is gradually increased until the preset missed and false positive constraints are met. The calibration process only determines the thresholds and does not change the processing chain during online detection.

[0036] When the principal direction curvature characterization value of the target discrete point is greater than or equal to the first discrimination threshold, and the Gaussian morphology characterization value is greater than or equal to the second discrimination threshold, the segment division unit determines that the target discrete point simultaneously satisfies the principal direction curvature abrupt change condition and the Gaussian morphology abrupt change condition, and identifies it as an abrupt change discrete point; for implementations that need to distinguish between depressions and protrusions, the segment division unit may also require that the direction mark of the Gaussian morphology index be consistent with the calibrated tin layer depletion depression direction; when the characterization value is exactly equal to the threshold, the target discrete point is assigned to the side that satisfies the threshold condition, thereby keeping the boundary assignment determined.

[0037] The segment division unit arranges abrupt discrete points according to their longitudinal position. Multiple abrupt discrete points that are continuous in the longitudinal position and adjacent in the lateral position are merged into the same abrupt connected region. The minimum and maximum longitudinal positions covered by the same abrupt connected region are then determined as spatial boundaries. The longitudinal seam segments within the spatial boundaries are determined as tin-depleted abrupt segments. In one possible implementation, continuity means that the longitudinal index difference between adjacent abrupt discrete points is no greater than one sampling step and the lateral index difference is no greater than one sampling step. In another possible implementation, non-abrupt discrete points not exceeding a preset gap length are allowed between two abrupt discrete points, and the gap is merged when the joint morphology intensity of the abrupt discrete points on both sides of the gap is higher than a threshold. The remaining segments in the liquid tin pool cooling section that are not covered by tin-depleted abrupt segments are determined as smooth transition segments.

[0038] The segmentation unit marks tin-deficient abrupt transitions as weld quality anomalies and outputs weld quality inspection results including the milk powder can identifier, the starting longitudinal position and ending longitudinal position of the tin-deficient abrupt transition, the lateral coverage, the set of main direction bending characterization values, and the set of Gaussian morphology characterization values. When there are no spatially continuous abrupt transition discrete points, the segmentation unit outputs a normal weld quality mark and retains the local morphological fluctuation characterization map for traceability. By using two morphological conditions and spatial continuity for joint discrimination, the possibility of fluctuations at a single sampling point being directly extended into an abnormal segment can be reduced.

[0039] The adjustment module includes a cumulative amplitude calculation unit and a pressure adjustment unit. The cumulative amplitude calculation unit reads the spatial boundary of the tin-deficient abrupt change segment in the weld quality inspection results and extracts segment data from the phase angle step distribution spectrum according to the same position index. When the abrupt change segment covers multiple sampling positions laterally, the cumulative amplitude calculation unit can form segment data for each lateral sampling position, or select the effective phase angle difference with the largest absolute value at each longitudinal position to form a representative sequence. Since the extraction process uses the phase angle step distribution spectrum before point cloud construction, the cumulative amplitude retains the original amplitude information of impedance phase change and is not limited by the spectrum normalization range.

[0040] The cumulative amplitude calculation unit determines the spatial boundary based on the starting and ending longitudinal positions of the tin-deficient abrupt change segment, and extends the spatial boundary forward and backward by a preset number of sampling points to obtain a preset accumulation segment. The extended portion is used to cover the rising and falling edges of the phase angle step at the abrupt change edge. If the extended boundary exceeds the liquid tin pool cooling segment, the cumulative amplitude calculation unit restricts it to the effective boundary of the liquid tin pool cooling segment. The cumulative amplitude calculation unit reads the effective phase angle difference values ​​within the preset accumulation segment according to the spatial extension order of the longitudinal seam overlap edge, takes the absolute value of each phase angle difference value, and adds it to the previous accumulation result point by point. The accumulation result corresponding to the last effective sampling point is the cumulative amplitude of the phase angle difference value. If there is no effective phase angle difference value within the preset accumulation segment, no pressure compensation coefficient is generated and a sampling anomaly mark is output.

[0041] When there are multiple phase angle difference sequences in the lateral direction, one approach is to first accumulate each lateral sampling position point by point, and then take the maximum value of each lateral accumulation result as the phase angle difference accumulation amplitude, so that the pressure compensation responds to the most severe local depletion. Another approach is to determine the lateral weight according to the coverage ratio of each lateral sampling position in the width of the electrode wheel pressure action, multiply the lateral weight by the corresponding lateral accumulation result and sum them to obtain the phase angle difference accumulation amplitude. The sum of the lateral weights is one and each lateral weight is non-negative, so that the accumulation amplitude reflects both the abnormal intensity and the lateral coverage range.

[0042] The pressure regulation unit pre-stores the correspondence between the cumulative amplitude of the phase angle difference and the candidate pressure compensation coefficients. This correspondence can be a one-dimensional piecewise mapping table. The mapping table is established through process calibration. The process calibration selects multiple pressure levels within the allowable pressure range and performs test soldering on calibration tanks with different tin layer depletion degrees. The cumulative amplitude of the phase angle difference and the weld quality re-inspection results are recorded at each pressure level. Among the pressure levels that meet the weld quality requirements, the level with the smaller pressure change is selected, and the pressure compensation ratio corresponding to this level is written into the corresponding amplitude range. Each amplitude range is arranged in ascending order of cumulative amplitude. The candidate pressure compensation coefficient does not decrease as the amplitude range increases. Therefore, the pressure compensation amount remains or increases when the degree of abnormality increases.

[0043] The pressure regulating unit determines the target amplitude range where the cumulative amplitude of the phase angle difference lies. When the cumulative amplitude is located at the common boundary of two amplitude ranges, it is assigned to the amplitude range with the larger value. Then, it obtains the candidate pressure compensation coefficient corresponding to the target amplitude range from the preset correspondence. The pressure compensation coefficient can represent the dimensionless increment ratio relative to the reference feed pressure, or it can represent the increment ratio of the control quantity corresponding to the pressure actuator. If the target pressure corresponding to the candidate pressure compensation coefficient is higher than the upper limit of the allowable pressure of the electrode wheel, the pressure regulating unit sets the target pressure to the upper limit of the allowable pressure. If the target pressure is lower than the lower limit of the allowable pressure, the target pressure is set to the lower limit of the allowable pressure. The pressure compensation coefficient is then calculated back based on the restricted target pressure, thereby avoiding dynamic compensation from exceeding the allowable working range of the electrode wheel, tank plate, or servo actuator.

[0044] To ensure the detection position corresponds to the pressure application position, the system determines the distance along the longitudinal seam between the effective sampling position of the impedance phase angle and the maximum pressure application position of the electrode wheel during production calibration. This distance is then divided by the current conveying speed to obtain the trigger delay, or the distance can be directly converted into the number of position encoder pulses. The pressure adjustment unit writes the pressure compensation coefficient, the spatial position of the tin layer depletion abrupt change section, the trigger encoder count, the pressure rise slope, the pressure holding distance, and the pressure recovery slope into the feed pressure dynamic adjustment signal. When the encoder count indicates that the corresponding position of the tin layer depletion abrupt change section has reached the pressure application position, the servo drive adjusts the feed pressure to the target pressure according to the pressure rise slope, maintains the target pressure within the pressure holding distance, and returns to the reference feed pressure according to the pressure recovery slope. The position trigger does not depend on the constant conveying speed, which can reduce the compensation position offset caused by speed fluctuations.

[0045] In one possible mechanical arrangement, the effective sampling position of the impedance phase angle is located on the entry side of the electrode wheel contact arc, and the maximum pressure application position is located in the middle of the contact arc. The spatial distance between the entry side and the middle provides a response advance for the pressure actuator. When the inherent response time of the pressure actuator is greater than this advance, the controller can generate a pre-trigger signal based on the phase angle step change trend of the continuous sections before the current tin layer depletion abrupt change section, and correct the target pressure after the abrupt change condition is confirmed. If it is impossible to complete the compensation for the current abrupt change section at the time of confirmation, the system retains the current section as a weld quality abnormality and does not record the pressure adjustment of the subsequent sections as the current section has been repaired, thereby keeping the detection results consistent with the actual control state.

[0046] In some embodiments, the pressure regulating unit also sets a pressure change rate limit and an adjacent regulating segment merging rule. When the spatial interval between two tin-depleted abrupt change segments is less than the distance required for the pressure actuator to complete one recovery, the pressure regulating unit merges the two regulating segments and uses the larger of the two pressure compensation coefficients or the corresponding coefficient according to the position segment within the merged segment. When the difference between the pressure sensor feedback value and the target pressure continues to exceed the allowable tracking error, the pressure regulating unit stops increasing the pressure and outputs an actuator abnormality flag. These processes reduce servo oscillations caused by frequent pressure increases and decreases and avoid continuing to issue pressure boosting commands when the pressure feedback fails.

[0047] See Figure 2 , Figure 2 This application illustrates a control method for an online inspection system for the weld quality of milk powder cans, which can be executed by the system and includes steps S201 to S206. Figure 2 The order shown illustrates the data dependencies. Without breaking these dependencies, some data validation and caching actions can be performed in parallel.

[0048] S201: Obtain the impedance phase angle sampling point array when the electrode wheel sweeps across the overlapping edge of the longitudinal seam of the milk powder can. Calculate the phase angle difference between adjacent sampling points based on the impedance phase angle sampling point array, and arrange the phase angle differences along the spatial extension direction of the overlapping edge of the longitudinal seam to obtain the phase angle step distribution spectrum. Specifically, the sampling unit determines the longitudinal position based on the output of the position encoder, determines the lateral sampling position based on the position of the detection channel or the feedback of the lateral micro-displacement mechanism, and writes the impedance phase angle sampling values ​​within the same spatial grid into the impedance phase angle sampling point array. The difference calculation unit sequentially subtracts the longitudinally adjacent impedance phase angle sampling values ​​in each lateral sampling column and associates the difference values ​​with the corresponding spatial positions. The phase angle step distribution spectrum obtained in this step is used as the input for step S202.

[0049] See Figure 3 , Figure 3 The diagram illustrates a process for the acquisition module to execute step S201. The sampling unit first synchronously receives the impedance phase angle sampling value and the sampling point, and then completes phase continuity, invalid point marking, position alignment and dot matrix arrangement according to the spatial order of the sampling points. The difference calculation unit then selects spatially adjacent sampling values ​​from the dot matrix to calculate the phase angle difference, and arranges the phase angle difference along the spatial extension direction of the longitudinal seam overlap edge. By adopting the order of completing position alignment first and then calculating the difference, the change in sampling spacing caused by the change in conveying speed can be avoided as a phase angle step.

[0050] S202: Construct a three-dimensional discrete point cloud based on the phase angle step distribution spectrum, and construct a local weighted scatter matrix based on the spatial coordinates of the target discrete point and its neighboring discrete points. Specifically, the point cloud construction unit uses the longitudinal position, lateral sampling position, and topography height characterization value corresponding to each phase angle difference to form the spatial coordinates of the discrete point. The matrix construction unit selects a preset neighborhood for each effective discrete point as the target discrete point, calculates the coordinate difference vector and spatial distance between each neighboring discrete point and the target discrete point, allocates neighborhood weights according to the rule that the weight decreases as the distance increases, and then performs weighted accumulation and weighted sum normalization on the product of the same-dimensional coordinate components and the product of the cross-dimensional coordinate components to obtain the local weighted scatter matrix. The local weighted scatter matrix obtained in this step is used as the input for step S203.

[0051] S203: Perform eigenvalue decomposition on the local weighted scatter matrix to obtain the principal direction topography curvature index and Gaussian morphology index. Generate a local topography undulation characterization map based on the principal direction topography curvature index and Gaussian morphology index. Specifically, the eigenvalue decomposition unit determines two local tangential directions and one local normal direction according to the magnitude of the eigenvalues. The principal direction topography curvature index is determined based on the change in the local direction of adjacent target discrete points in the longitudinal seam extension direction. The Gaussian morphology index is determined based on the directional curvature change and directional relationship of the two mutually orthogonal local directions. The map generation unit maps the two indices to a common preset value range to form a joint topography marker for each target discrete point. The joint topography marker is then arranged according to the spatial position of the target discrete points. The local topography undulation characterization map obtained in this step is used as the input for step S204.

[0052] S204: Based on the local morphological undulation characterization map, the liquid tin pool cooling section is divided into a smooth transition section and a tin-depleted abrupt change section. The tin-depleted abrupt change section is marked as a weld quality anomaly, and the weld quality inspection result is obtained. Specifically, the section division unit compares the main direction bending characterization value with the first discrimination threshold and the Gaussian morphology characterization value with the second discrimination threshold. The target discrete point that simultaneously satisfies both abrupt change conditions is determined as the abrupt change discrete point. Multiple abrupt change discrete points with continuous spatial positions are merged into a tin-depleted abrupt change section, and the remaining sections are determined as smooth transition sections. The weld quality inspection result obtained in this step includes the anomaly mark and the spatial boundary of the tin-depleted abrupt change section, and is used as the input for step S205.

[0053] S205: Based on the weld quality inspection results, extract the segment data corresponding to the tin-deficient abrupt change segment from the phase angle step distribution spectrum, and accumulate the phase angle difference values ​​in the segment data point by point to obtain the cumulative amplitude of the phase angle difference value; specifically, the cumulative amplitude calculation unit determines the preset accumulation segment according to the spatial boundary of the tin-deficient abrupt change segment, reads the effective phase angle difference value in the spatial extension sequence along the longitudinal seam overlap edge, takes the absolute value of each phase angle difference value and adds it to the previous accumulation result, and determines the final accumulation result as the cumulative amplitude of the phase angle difference value; the cumulative amplitude of the phase angle difference value obtained in this step characterizes the overall intensity of the phase step in the abnormal segment and is used as the input for step S206.

[0054] S206, the pressure compensation coefficient is determined based on the cumulative amplitude of the phase angle difference, and a dynamic adjustment signal for the feed pressure of the electrode wheel is generated based on the pressure compensation coefficient. Specifically, the pressure adjustment unit determines the target amplitude range where the cumulative amplitude of the phase angle difference is located, reads the candidate pressure compensation coefficient from the preset correspondence, limits the target pressure corresponding to the candidate pressure compensation coefficient to the allowable pressure adjustment range of the electrode wheel, and associates the limited pressure compensation coefficient with the spatial position of the tin layer depletion abrupt change section, and generates a dynamic adjustment signal for the feed pressure by combining the position encoder count. Since the adjustment signal contains both the pressure target and the spatial triggering condition, the pressure actuator can enter the pressure compensation state at the corresponding position and restore the reference feed pressure after the section leaves the pressure application position.

[0055] In a complete online processing example, when the milk powder can enters the welding station, the system reads the milk powder can identifier and clears the spatial index of the previous can. The sampling unit continuously fills the impedance phase angle sampling point matrix as the position encoder counts. The difference calculation unit updates the phase angle step distribution spectrum line by line. The point cloud construction unit and matrix construction unit only update the three-dimensional topography discrete point cloud and local weighted scatter matrix for target discrete points that have complete preceding and following neighborhoods. The detection module then updates the local topography undulation characterization map and confirms the abrupt discrete points. When the continuous abrupt discrete points reach the segment merging condition, the segment division unit outputs the starting position of the tin-deficient abrupt segment. After the abrupt condition ends, it outputs the ending position. The adjustment module calculates the cumulative amplitude of the phase angle difference and generates a pressure adjustment command associated with the encoder position. For the end data of the next neighboring point that has not yet been obtained, the system can use unilateral calculation and complete the final confirmation before the milk powder can leaves the station.

[0056] In another streaming processing method, the system does not wait for the complete collection of milk powder can data. Instead, it sets up a sliding buffer that can cover a preset neighborhood and a preset accumulation segment. Each time the acquisition module writes a new vertical sampling row, the construction module updates the local weighted scatter matrix of the central row of the buffer, the detection module updates the joint morphology label of the corresponding target discrete point, and the adjustment module generates a dynamic adjustment signal for the feed pressure after the spatial boundary meets the confirmation conditions. Before the sliding buffer is removed from the system, it only retains the weld quality inspection results and necessary traceability data, thereby reducing the storage occupation of online processing and the waiting time for the whole can.

[0057] Before the system is put into production, calibration can be performed using qualified weld samples of the same material, plate thickness, and overlap width, as well as weld samples with tin-poor characteristics. The calibration content includes the reference phase angle difference scale, neighborhood range, directional stability threshold, upper and lower limits of exponential mapping, first discrimination threshold, second discrimination threshold, amplitude range, candidate pressure compensation coefficient, allowable pressure adjustment range of electrode wheel, and encoder pulse count from sampling position to pressure application position. Different tank types or material batches use their own parameter sets. The controller loads the corresponding parameter set according to the tank type identifier and material batch identifier in the production task. If there is no matching parameter set, the system maintains the reference feed pressure and outputs a prompt that calibration is required, without directly applying parameter sets with inconsistent dimensions or scales.

[0058] For cases where the phase angle difference scale is zero, the exponential mapping range width is zero, the neighborhood weight sum is zero, the spatial distance between adjacent sampling points is zero, or there is a lack of effective segment data, the aforementioned units process the data according to the corresponding invalidation flag, neighborhood expansion, endpoint restriction, or stop adjustment rules. For cases where the phase angle difference exceeds the range of the acquisition circuit, the sampling unit marks continuous saturation points as acquisition anomalies and does not directly interpret saturation values ​​as the degree of tin layer depletion. For cases where eigenvalue calculation produces tiny negative values, if the absolute value of the negative value is within the numerical calculation tolerance, the eigenvalue decomposition unit restricts it to zero; if it exceeds the numerical calculation tolerance, the corresponding matrix is ​​discarded and a calculation anomaly flag is output. These boundary rules keep the division input, matrix input, and control output within the executable range.

[0059] After the system is powered on again, production tasks are switched, or the electrode wheel is replaced, the controller can first perform zero-point calibration on the reference sample or the qualified starting section set by the production line to obtain the phase angle baseline and noise range of the current detection loop, and then allow the section division unit to output abnormal results. If the baseline drift exceeds the calibration allowable range, the controller prompts to check the surface condition of the electrode wheel, conductive connection and impedance acquisition loop, and keeps the pressure regulation module in the prohibited automatic pressurization state. This calibration does not change the processing method of the differential between adjacent sampling points, but can prevent the detection loop fault from being mistaken for a process abnormality and triggering pressure regulation.

[0060] The sampling unit and difference calculation unit in the acquisition module correspond to the sampling processing and difference processing in step S201, respectively. The point cloud construction unit and matrix construction unit in the construction module correspond to step S202. The feature decomposition unit, map generation unit and segmentation unit in the detection module correspond to steps S203 and S204, respectively. The cumulative amplitude calculation unit and pressure adjustment unit in the adjustment module correspond to steps S205 and S206, respectively. Each module or unit can be deployed independently according to the above functional boundaries, or they can be combined and deployed in the same processor without changing the input, processing relationship and output purpose.

[0061] In one implementation, the online inspection system for weld quality of milk powder cans includes a processor, a memory, an impedance acquisition interface, a position encoder interface, and a pressure control interface. The memory stores an impedance phase angle sampling point matrix, a phase angle step distribution spectrum, a three-dimensional topographic discrete point cloud, a local weighted scatter matrix, a local topographic undulation characterization map, a calibration parameter set, and a computer program. The processor executes the computer program to complete steps S201 to S206. The impedance acquisition interface receives impedance phase angle sampling values, the position encoder interface receives sampling point position information, and the pressure control interface sends a dynamic adjustment signal for feed pressure to the servo driver. The processor can be a central processing unit, a digital signal processor, a microcontroller, a programmable logic device, or a combination of at least two of these devices.

[0062] In one chip implementation, the chip includes a processing circuit and an interface circuit. The interface circuit is used to receive the impedance phase angle sampling value and sampling point position, and output the feed pressure dynamic adjustment signal. The processing circuit is used to perform sampling point array arrangement, phase angle difference calculation, point cloud construction, local weighted scatter matrix construction, eigenvalue decomposition, morphology index determination, map generation, segment division, cumulative amplitude calculation, and pressure compensation coefficient determination. The chip can be combined with external memory, impedance acquisition circuit, position encoder, and pressure actuator to form an online inspection system for the weld quality of milk powder cans.

[0063] This application embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it causes the processor to perform steps S201 to S206. The computer-readable storage medium can be a read-only memory, random access memory, flash memory, solid-state memory, magnetic storage medium, or optical storage medium. This application embodiment also provides a computer program product, which includes a computer program. When the computer program runs on a processor, it causes the processor to perform the above-described control method. The input, processing, and output relationships of each program instruction are consistent with the aforementioned method embodiments.

[0064] In the absence of logical conflicts, the technical features in the above embodiments or implementations can be combined with each other. Data verification, caching, and interface transmission in the method steps can be executed in parallel or in combination according to the hardware structure. However, the progressive processing relationship between the phase angle step distribution spectrum, the three-dimensional topography discrete point cloud, the local weighted scatter matrix, the local topography undulation characterization map, the weld quality inspection result, the cumulative amplitude of the phase angle difference, the pressure compensation coefficient, and the feed pressure dynamic adjustment signal remains unchanged. The modules or units in the device can be implemented by software, hardware, firmware, or a combination thereof. The scope of protection of this application shall be determined by the scope defined above.

Claims

1. An online inspection system for the weld quality of milk powder cans, characterized in that, include: The acquisition module is used to acquire the impedance phase angle sampling point array when the electrode wheel sweeps across the overlapping edge of the longitudinal seam of the milk powder can, and obtain the phase angle step distribution spectrum based on the phase angle difference between adjacent sampling points. The module is used to construct a three-dimensional discrete point cloud based on the phase angle step distribution spectrum, and to construct a local weighted scatter matrix based on the spatial coordinates of the target discrete point and its neighboring discrete points. The detection module is used to perform feature decomposition on the local weighted scatter matrix to obtain the main direction morphology bending index and Gaussian morphology index. Based on the main direction morphology bending index and Gaussian morphology index, a local morphology undulation characterization map is generated. Based on the local morphology undulation characterization map, the liquid tin pool cooling section is divided into a smooth transition section and a tin layer depletion abrupt section. The tin layer depletion abrupt section is marked as a weld quality abnormality, and the weld quality detection result is obtained. The adjustment module is used to extract the segment data corresponding to the tin layer depletion abrupt change segment from the phase angle step distribution spectrum according to the weld quality inspection results, accumulate the phase angle difference value in the segment data point by point to obtain the cumulative amplitude of the phase angle difference value, determine the pressure compensation coefficient according to the cumulative amplitude of the phase angle difference value, and generate a dynamic adjustment signal for the feed pressure of the electrode wheel according to the pressure compensation coefficient.

2. The online inspection system for weld quality of milk powder cans according to claim 1, characterized in that, The acquisition module includes a sampling unit and a difference calculation unit; The sampling unit is used to synchronously acquire the impedance phase angle sampling value and the sampling point corresponding to each impedance phase angle sampling value during the electrode wheel scanning process. The impedance phase angle sampling values ​​are arranged according to the spatial order of each sampling point on the longitudinal seam overlap edge to obtain the impedance phase angle sampling point array. The difference calculation unit is used to select spatially adjacent impedance phase angle sample values ​​from the impedance phase angle sampling point array, calculate the difference between the two selected impedance phase angle sample values, obtain the phase angle difference, and arrange the phase angle difference along the spatial extension direction of the longitudinal seam overlap edge to obtain the phase angle step distribution spectrum.

3. The online inspection system for weld quality of milk powder cans according to claim 2, characterized in that, The building module includes point cloud building units; The point cloud construction unit is used to obtain the longitudinal and lateral sampling positions corresponding to each phase angle difference in the phase angle step distribution spectrum. The phase angle difference is used as the topography height characterization value of the corresponding sampling position. Based on the longitudinal position, lateral sampling position and topography height characterization value, the corresponding discrete point spatial coordinates are generated, and the spatial coordinates of each discrete point are collected to obtain a three-dimensional topography discrete point cloud.

4. The online inspection system for weld quality of milk powder cans according to claim 3, characterized in that, The building module also includes matrix building units; The matrix construction unit is used to select neighborhood discrete points located in a preset neighborhood from the three-dimensional discrete point cloud with the target discrete point as the center, calculate the coordinate difference vector between each neighborhood discrete point and the target discrete point, determine the neighborhood weight according to the spatial distance between each neighborhood discrete point and the target discrete point, and determine the neighborhood weight according to the rule that the weight decreases as the spatial distance increases. The weighted summation of the product of the same dimension coordinate components and the product of the cross dimension coordinate components of each coordinate difference vector is performed to obtain the local weighted scatter matrix.

5. The online inspection system for weld quality of milk powder cans according to claim 4, characterized in that, The detection module includes a feature decomposition unit; The eigenvalue decomposition unit is used to sort the eigenvalues ​​of the local weighted scatter matrix, determine the eigenvectors corresponding to each eigenvalue as the local direction of the target discrete point, obtain the main direction morphology bending index based on the amount of change of the local direction of adjacent target discrete points along the longitudinal seam overlap edge extension direction, and obtain the Gaussian morphology index based on the bending change and directional relationship of two mutually orthogonal local directions.

6. The online inspection system for weld quality of milk powder cans according to claim 5, characterized in that, The detection module also includes a spectrum generation unit; The map generation unit is used to map the main direction morphology bending index and Gaussian morphology index to preset value ranges to obtain the main direction bending characterization value and Gaussian morphology characterization value. Based on the main direction bending characterization value and Gaussian morphology characterization value, the joint morphology label of each target discrete point is determined, and the joint morphology label is arranged according to the spatial position of each target discrete point in the liquid tin pool cooling section to obtain the local morphology undulation characterization map.

7. The online inspection system for weld quality of milk powder cans according to claim 6, characterized in that, The detection module also includes a segmentation unit; The segmentation unit is used to compare the main direction bending characterization value with the first discrimination threshold, and the Gaussian shape characterization value with the second discrimination threshold. Based on the comparison results, the target discrete point that simultaneously satisfies the main direction bending abrupt change condition and the Gaussian shape abrupt change condition is determined as the abrupt change discrete point. Multiple abrupt change discrete points with continuous spatial location are merged into the tin layer depletion abrupt change segment, and the remaining segments in the liquid tin pool cooling segment are determined as the smooth transition segment.

8. The online inspection system for weld quality of milk powder cans according to claim 7, characterized in that, The adjustment module includes a cumulative amplitude calculation unit; The cumulative amplitude calculation unit is used to extract the spatially continuous phase angle difference values ​​from the segment data according to the spatial boundary of the tin-deficient abrupt segment, obtain the abrupt segment feature data cluster, determine the preset accumulation segment based on the spatial boundary, and accumulate the absolute values ​​of each phase angle difference value in the abrupt segment feature data cluster point by point according to the spatial extension order of the longitudinal seam overlap edge within the preset accumulation segment to obtain the cumulative amplitude of the phase angle difference value.

9. The online inspection system for weld quality of milk powder cans according to claim 8, characterized in that, The adjustment module also includes a pressure adjustment unit; The pressure regulation unit is used to determine the target amplitude range where the cumulative amplitude of the phase angle difference is located. It obtains the candidate pressure compensation coefficient corresponding to the target amplitude range from the preset correspondence. The candidate pressure compensation coefficient increases as the target amplitude range increases. The candidate pressure compensation coefficient is limited to the allowable pressure regulation range of the electrode wheel to obtain the pressure compensation coefficient. The pressure compensation coefficient is then associated with the spatial position of the tin layer depletion abrupt change section to generate a dynamic adjustment signal for the feed pressure.

10. A control method for an online inspection system for the weld quality of milk powder cans as described in any one of claims 1 to 9, characterized in that, include: The impedance phase angle sampling point array is obtained when the electrode wheel sweeps across the longitudinal seam overlap edge of the milk powder can. The phase angle difference between adjacent sampling points is calculated based on the impedance phase angle sampling point array, and the phase angle difference is arranged along the spatial extension direction of the longitudinal seam overlap edge to obtain the phase angle step distribution spectrum. A three-dimensional discrete point cloud is constructed based on the phase angle step distribution spectrum, and a local weighted scatter matrix is ​​constructed based on the spatial coordinates of the target discrete point and its neighboring discrete points. The local weighted scatter matrix is ​​subjected to eigenvalue decomposition to obtain the main direction morphological curvature index and Gaussian morphological index. Based on the main direction morphological curvature index and Gaussian morphological index, a local morphological undulation characterization map is generated. Based on the local morphological fluctuation characterization map, the cooling section of the liquid tin pool is divided into a smooth transition section and a tin-poor abrupt change section. The tin-poor abrupt change section is marked as a weld quality abnormality, and the weld quality inspection results are obtained. Based on the weld quality inspection results, the data corresponding to the tin layer depletion abrupt change segment is extracted from the phase angle step distribution spectrum. The phase angle difference values ​​in the segment data are accumulated point by point to obtain the cumulative amplitude of the phase angle difference value. The pressure compensation coefficient is determined based on the cumulative amplitude of the phase angle difference, and a dynamic adjustment signal for the feed pressure of the electrode wheel is generated based on the pressure compensation coefficient.

Images