A visual detection system for weld defects of a brazed water-cooled plate

Abstract

The application discloses a kind of weld defect visual inspection systems of brazing water-cooled plate, it is related to welding seam visual inspection technical field, the system includes: topological design module, identification registration module, iteration alignment module and defect discrimination module;Its technical key points are: construct X scattering topological network;Obtain first X scattering image, construct registration cost function;Obtain the mechanical execution end vibration signal and welding infrared temperature signal of known defect event, generate correlation matrix in combination with overlapping shielding coefficient, obtain the registration parameter range of each class of defect event;By constructing control grid tree, introduce rule engine, determine the correction parameter of the region to be detected by minimizing registration cost function, determine second X scattering image;Traverse control grid tree, determine third X scattering image by screening comparison, output defect recognition result in combination with registration parameter range;The application reduces artifact false positive rate, realizes the automatic identification of seam defect.

Description

Technical Field

[0001] This invention relates to the field of weld visual inspection technology, specifically a visual inspection system for weld defects in brazed water-cooled plates. Background Technology

[0002] Water-cooled plates are key components of electronic devices, mainly used for efficient heat dissipation and temperature control. They are suitable for precise temperature control of high-power equipment such as power batteries and motors in energy storage systems and new energy vehicles. Most water-cooled plates are formed by brazing. With the development of high-voltage fast charging technology for new energy vehicles, brazed needle-shaped fins have become the preferred structure for extreme heat dissipation scenarios. The internal fin density of this structure is extremely high. After being brazed at high temperature, the staggered needle-like structure inside causes severe spatial overlap and physical obstruction of the root weld seam in X-ray scattering imaging.

[0003] Traditional weld seam visual inspection techniques struggle to eliminate interference, leading to frequent missed detections and false alarms. They also have the following limitations: First, the weld bonding rate inside cold-rolled steel plates cannot be manually determined; it can only be observed by cutting the product open using gas testing and wire cutting to check for internal defects. Furthermore, flux crystals may remain in the internal flow channels, causing blockages and affecting product usability. Second, industrial visual judgment typically relies on single two-dimensional radiometric image pixel features, which are prone to false alarms or missed detections under complex structural occlusion. For example, scattered highlights from densely overlapping heat dissipation pins can obscure the true weld quality. Incomplete weld penetration, porosity, or cracks can lead to image spatial slippage and distortion caused by X-ray scattering, resulting in inaccurate feature alignment and ultimately severely insufficient defect localization accuracy and recognition confidence. On the other hand, during the welding process, the water-cooled plate will experience local warping and nonlinear deformation due to uneven temperature distribution. Traditional methods only correct the overall translation, rotation, and scaling, failing to handle local deformation differences. This results in severe misalignment between the weld seam and the topology network in densely packed needle-pillar areas, leading to artifact misjudgments. Especially in areas where needle-pillars overlap, the artifact misjudgment rate remains consistently high, making it difficult to quickly and accurately identify weld defects in brazed water-cooled plates. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a visual inspection system for weld defects in brazed water-cooled plates. By extracting the three-dimensional layout of the water-cooled plate, the pin posts and weld scattering channels are abstracted into an X-scattering topology network composed of nodes and edges. By simultaneously fusing mechanical vibration and infrared temperature signals from the welding site, a dynamic registration parameter range is constructed using an association matrix. By introducing a rule engine, the first X-scattering image is corrected while minimizing the registration cost function. By traversing the control mesh tree to solve for artifact costs, path merging and noise blocking are performed on the second X-scattering image to generate a high-fidelity third X-scattering image. Finally, the defect identification result is output based on the registration parameter range, thus solving the problems mentioned in the background technology.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This application provides a visual inspection system for weld defects in brazed water-cooled plates, the system comprising:

[0009] The topology design module obtains the water-cooled plate design layout of the area to be inspected, extracts the needle columns and abstracts them into a set of feature nodes, extracts the weld scattering channels between adjacent needle columns and abstracts them into a set of edges, and constructs an X-scattering topology network; where each feature node corresponds to the three-dimensional coordinates of the needle column, and the weight of each edge is determined by the overlap occlusion coefficient of the weld scattering channel.

[0010] The identification and registration module acquires the first X-scattering image of the area to be detected and constructs a registration cost function by combining it with the X-scattering topology network. At the same time, it acquires the mechanical execution end vibration signal and welding infrared temperature signal of the known defect event, extracts the first feature parameter, and performs mapping association by combining the overlap occlusion coefficient to generate an association matrix, thereby obtaining the registration parameter range for each type of defect event.

[0011] The iterative alignment module constructs a control grid tree based on the feature node set and the overlap occlusion coefficient, introduces a rule engine, determines the correction parameters of the region to be detected by minimizing the registration cost function, and transforms the first X-scattering image into the second X-scattering image.

[0012] The defect identification module traverses the control grid tree to calculate the artifact cost and merges the paths of the second X-ray image to determine the third X-ray image. It then extracts the second feature parameters and outputs the defect identification result by combining the registration parameter range.

[0013] Furthermore, the steps for determining the overlap occlusion coefficient include:

[0014] Analyze the design layout of the water-cooled plate and establish projection rays from N0 X-rays pointing to the weld contour of the water-cooled plate;

[0015] The collision detection algorithm is used to count the intersection frequency of each projected ray with the needle column to determine the number of X-ray overlaps, and the physical thickness value of the needle column is accumulated along the penetration trajectory.

[0016] Based on the number of X-ray overlaps and the accumulated physical thickness value of the weld scattering channel, a first dynamic curve is established and the peak region is identified. By extracting the X-ray energy value and penetration intensity value, a second dynamic curve is established. The DTW algorithm is used to determine the curve type and match the proportional weight combination. The overlap and occlusion coefficient is determined by exponential weighting.

[0017] Furthermore, the registration cost function is constructed, including:

[0018] Based on the first X-ray image, the first pixel coordinates of the weld contour of each water-cooled plate are extracted; the three-dimensional coordinates of the feature node set in the X-ray topology network are identified, the second pixel coordinates corresponding to the three-dimensional coordinates are retrieved from the database, the minimum Euclidean distance between the two pixel coordinates is calculated, the feature nodes with a minimum Euclidean distance less than the standard distance threshold are selected, and a registration mapping chain is established to identify pin-column node pairs.

[0019] The feature matching residual is constructed based on the sum of the squared minimum Euclidean distances between each pair of needle and column nodes, and the data terms of the registration cost function are determined. At the same time, the smoothing term of the registration cost function is determined based on the absolute value of the difference in edge weights between any two adjacent edges sharing the same feature node under the defect event. The dynamically updated first and second weights are retrieved, and the data terms and smoothing terms are linearly weighted and summed to construct the registration cost function.

[0020] Furthermore, the registration parameter range for each type of defect is obtained, including:

[0021] The first characteristic parameters include vibration characteristics and temperature distribution characteristics:

[0022] Vibration characteristics are determined based on the vibration signal at the mechanical actuator end, and the first change vector is determined by time-series differential analysis.

[0023] The temperature distribution characteristics are determined based on the welding infrared temperature signal, and the second change vector is determined by similarity comparison.

[0024] Retrieve the overlapping occlusion coefficients, match time labels for the overlapping occlusion coefficients according to the time series, including the first time label and the second time label; establish the correlation matrices between the first change vector, the second change vector and the overlapping occlusion coefficients respectively;

[0025] Extract all time tags that trigger the defect event, take the first recorded time tag as the intermediate time, and extract two step lengths in a bidirectional symmetrical manner. Extract the step lengths corresponding to the first change vector and the second change vector respectively, combine them to form the defect time period, sort them in chronological order to form a dynamic time window and process them equally to generate multiple sub-time period dynamic score segments, including the first dynamic score segment and the second dynamic score segment.

[0026] An adaptive clustering algorithm is used to select different dynamic score segments to divide the first change vector and the second change vector clusters, and output the registration parameter range for each type of defect event; wherein, the defect events include the first defect event and the second defect event.

[0027] Furthermore, the adaptive selection clustering algorithm based on different dynamic score segments includes:

[0028] If there is a remainder in the equal division result, determine the first dynamic score segment and perform manifold clustering: extract the corresponding first and second change vectors to form a fused feature vector, extract the monotonic intervals and identify the extreme points as initial seed points, preset the topological connected neighborhood radius, perform manifold expansion on the fused feature vector, filter out seed points larger than the topological connected neighborhood radius, and merge them into an irregular manifold defect cluster, which is determined as the first defect event;

[0029] If the equal division result has no remainder, determine the second dynamic score segment and perform ellipsoidal clustering: extract the corresponding first and second change vectors as observation vectors, use the expectation-maximization algorithm for iterative training, extract the mean and covariance matrix, select observation vectors that fall into the oblique ellipsoidal interval consisting of the major axis interval and the minor axis interval, and merge them into oblique ellipsoidal defect clusters, which are determined as the second defect event.

[0030] Further, the correction parameters for the area to be detected are determined, including:

[0031] A rule engine is introduced, which has built-in geometric constraints and convergence constraints for iterative solution of the control grid tree; among them, the geometric constraints include the maximum displacement vector threshold of the grid nodes; the convergence constraints include the control gradient step size and the convergence error threshold.

[0032] Under the rule engine, gradient descent is used to iteratively solve the registration cost function. When the registration cost function reaches a minimum value, the state vector of the current iteration is output as a correction parameter.

[0033] Furthermore, the costs of obtaining artifacts include:

[0034] A first spatial scale is preset, the displacement vector of the current needle-pillar node pair in the state vector is calculated, and the difference vector between it and the historical displacement vector of the previous adjacent needle-pillar node pair in the state vector is calculated. The variance of the angle between the continuous difference vectors is calculated to determine the spatial stability index. At the same time, the X-ray scattering direction is preset, the projection component of the difference vector on the X-ray scattering direction vector is extracted, and quantified into an artifact index.

[0035] A second spatial scale is preset, and the displacement vectors of several consecutive needle-column node pairs are linearly regressed to determine the trend change rate; the artifact cost is generated by weighted aggregation of spatial stability index, artifact index and trend change rate.

[0036] The minimum spanning tree algorithm is adopted. The unconnected pin nodes corresponding to the minimum projection cost are expanded iteratively and added to the X-scattering topology network in turn to update the control grid tree.

[0037] Furthermore, a control mesh tree is constructed by combining the overlap occlusion coefficient, including:

[0038] The envelope boundary is identified based on the feature node set as the root node, and the initial coordinate system of the root node is established. By performing hierarchical spatial recursion, in each iteration, the current parent mesh unit is equally divided into four child mesh units, and the corresponding mesh nodes are generated.

[0039] Map and calculate all overlapping occlusion coefficients within the area covered by the current sub-grid cell. Extract multiple quantiles and take the arithmetic mean. If the calculated result is greater than or equal to the standard blocking threshold, terminate the iteration; if the calculated result is less than the standard blocking threshold, continue the iteration to construct the control grid tree. The quantiles include the first quartile, the second quartile, and the third quartile.

[0040] Further, the third X-ray scattering image is determined, including:

[0041] Determine overlapping paths based on control grid trees;

[0042] Compare the artifact cost with the standard cost threshold: If the artifact cost is less than the standard cost threshold, perform path merging: perform image pixel-level weighted average fusion on the corresponding overlapping paths, and use the merged image as the third X-scatter image; if the artifact cost is greater than or equal to the standard cost threshold, keep the second X-scatter image and continue path traversal.

[0043] (III) Beneficial Effects

[0044] This invention provides a visual inspection system for weld defects in brazed water-cooled plates, which has the following advantages:

[0045] 1. This invention abstracts the three-dimensional geometric coordinates of the pin columns in the design layout of the water-cooled plate into a set of feature nodes and the scattering channels of the guide weld into a set of edges, thereby constructing an X-ray scattering topology network. By analyzing the number of X-ray overlaps and the accumulated physical thickness value, the first dynamic curve and the second dynamic curve are quantified. By screening and comparing different curve types, the severity of image degradation caused by structural overlap in each detection area inside the water-cooled plate can be accurately and quantitatively evaluated, providing core data flow basis for the subsequent adaptive correction of the control grid tree and accurate defect judgment.

[0046] 2. This invention achieves cross-domain fusion of the acquired first X-ray scattering image with the X-ray scattering topology network and the monitoring signals during the welding process. The introduction of mechanical execution end vibration signals and welding infrared temperature signals can significantly improve the accuracy of defect event judgment. By constructing time labels, the first change vector and the second change vector are respectively associated and aligned with the overlap and occlusion coefficients of the X-rays under the time label on the time axis to form a corresponding correlation matrix, thereby obtaining the registration parameter range for each type of defect event. This method realizes cross-modal spatiotemporal coupling of dynamic monitoring signals and static images, and adaptively adjusts the registration parameter range for different defect events, effectively enhancing the accurate positioning and classification of subsequent defect events and improving the robustness of the model.

[0047] 3. This invention constructs a control grid tree based on a set of feature nodes and an overlap occlusion coefficient. The constructed hierarchical control grid tree can achieve fine local registration at different resolutions. Regions with high overlap occlusion coefficients are automatically assigned finer control grids. By introducing a rule engine, constraint rules are assigned to the displacement vector of each grid node, and gradient descent is used to iteratively optimize the registration cost function to avoid getting trapped in local optima. While ensuring sub-pixel level registration accuracy, the iteration convergence speed is greatly improved. The first X-scattering image is converted into a second X-scattering image to achieve high-precision correction of the water-cooled plate, reduce misjudgments in subsequent defect identification, and improve the accuracy of identification.

[0048] 4. By performing two spatial scale analyses on the second X-ray image, traversing the control grid tree, calculating the artifact cost, and performing intelligent fusion on the overlapping path images of multi-channel scanning, the continuous and complete shape of the weld is restored. False artifact edges are removed to establish the third X-ray image that truly reflects the solid structure of the water-cooled plate. By combining the registration parameter range for joint judgment, the defect identification result is output, reducing the false alarm rate caused by artifact interference. This achieves high-precision and high-reliability automatic identification of weld defects in brazed water-cooled plates. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] The core of this invention lies in extracting the three-dimensional layout of the water-cooled plate and abstracting the scattering channels of the pin column and weld seam into an X-scattering topology network composed of nodes and edges; by synchronously fusing mechanical vibration and infrared temperature signals from the welding site, a dynamic registration parameter range is constructed through an association matrix; by introducing a rule engine, the first X-scattering image is corrected while minimizing the registration cost function; by traversing the control mesh tree to solve for artifact costs, path merging and noise blocking are performed on the second X-scattering image to generate a high-fidelity third X-scattering image, and the defect identification result is output in combination with the registration parameter range, thereby achieving accurate qualitative analysis of minor defects and automatic filtering of artifact interference, realizing defect identification under spatial overlap of pin column structures, with extremely high robustness and automation level.

[0052] Example: This embodiment of the invention provides a visual inspection system for weld defects in brazed water-cooled plates; Figure 1 This is a schematic diagram of the module of the present invention; please refer to it. Figure 1 The system includes a topology design module, an identification and registration module, an iterative alignment module, and a defect discrimination module, and the topology design module, identification and registration module, iterative alignment module, and defect discrimination module are interconnected.

[0053] The following is an explanation of each module:

[0054] Topology design module: Obtain the water-cooled plate design layout of the area to be detected, extract the pins and abstract them into a set of feature nodes, extract the weld scattering channels between adjacent pins and abstract them into a set of edges, and construct an X-scattering topology network; where each feature node corresponds to the three-dimensional coordinates of the pin, and the weight of each edge is determined by the overlap occlusion coefficient of the weld scattering channel;

[0055] The steps for determining the overlap occlusion coefficient include: reading the physical coordinates of N0 X-rays emitted by the X-ray generator to the area to be detected, where the physical coordinates are defined as the three-dimensional coordinates and direction vectors of the starting point of the straight path of the N0 X-rays in three-dimensional Cartesian world coordinates; during the system initialization phase, the water-cooled plate design layout is analyzed by the CNC analysis engine, and the standardized three-dimensional CAD model of the brazed water-cooled plate to be detected is extracted by the meshing algorithm to extract the three-dimensional surface triangular piece set of all heat dissipation pin column entities of the water-cooled plate, and the internal pin columns are transformed into triangular mesh models in three-dimensional space. At the same time, the welding area of ​​the water-cooled plate is identified, and the weld contour of the water-cooled plate is extracted by the Canny algorithm.

[0056] Using the origin of the N X-rays at their starting 3D coordinates as the origin, the rays are extended step-by-step in 3D space along their direction vectors. When the extension line passes through the water-cooled plate and overlaps with the weld contour of the water-cooled plate, the ray segments in 3D space are collected to establish a projection ray pointing from N0 X-rays to the weld contour of the water-cooled plate. It should be noted that the calculated projection ray vector is usually stored in the temporary storage area of ​​the computer system in the form of a spatial 3D coordinate array or a floating-point vector array to ensure that the penetration path and the water-cooled plate needle column entity achieve a one-to-one linear mapping in 3D space. Based on the projection ray, the penetration trajectory between two adjacent needle columns is determined and collected into a welding scattering channel; where N0 is a positive integer.

[0057] The welding scattering channel is traversed, and a collision detection algorithm is used to count the intersection frequency of each projected ray with the needle column. Specifically, this includes: retrieving the triangular mesh models of all needle columns, constructing hierarchical bounding boxes, and converting the triangular mesh models into regular cubes for enclosure; by traversing each projected ray, determining whether the projected ray geometrically intersects with the bounding box corresponding to each needle column; if the result is no intersection, skipping directly and instantly eliminating unrelated needle columns; if the result is an intersection, counting the integer number of times each projected ray geometrically intersects with the bounding boxes corresponding to all needle columns, marking this as the determined X-ray overlap quantity; simultaneously, if a certain... When a projected ray intersects with a needle column, the three-dimensional coordinates of the entry point and exit point of the ray vector are immediately identified. The Euclidean distance between the two coordinates is calculated and marked as the penetration depth. Along the penetration trajectory where the ray intersects, the single penetration depths of all the needle columns that the ray contacts are linearly accumulated along the entire straight path of the projected ray to obtain the physical thickness value on the penetration trajectory. For example, if a ray passes through two overlapping needle columns, the number of X-ray overlaps is counted as 2. If the thickness of a single needle column is 1.5 mm, the accumulated physical thickness value is 3.0 mm.

[0058] Based on the number of X-ray overlaps and the accumulated physical thickness value of the weld seam scattering channel, the number of X-ray overlaps at a certain moment is used as the horizontal axis, and the accumulated physical thickness value at the same moment is used as the vertical axis. As the scan continues along the weld seam contour, the number of pins traversed by X-rays and the accumulated physical thickness change at different moments. The first dynamic curve is plotted by connecting these lines sequentially. It should be noted that the first dynamic curve represents the law of change of the number of X-ray overlaps and the superposition of physical thickness values ​​along the weld seam scattering channel with time step over a continuous scanning period. It is used to analyze the spatially compact state of the heat dissipation pins inside the water-cooled plate in three-dimensional space. In the subsequent analysis, the first dynamic curve needs to be smoothed and filtered to eliminate pseudo-fluctuations caused by mesh splicing errors that may exist in the triangular mesh model, so as to ensure the accuracy of the subsequent analysis.

[0059] The peak region is identified based on the first dynamic curve. Specifically, this involves first accurately determining how many needle-like entities each projected ray intersects with when penetrating the water-cooled plate, thus obtaining the accurate number of X-ray overlaps. Simultaneously, the physical thickness values ​​of all intersecting needle-like entities are accumulated along the penetration trajectory. Then, a point-by-point scan is performed on the smoothed and filtered first dynamic curve, marking points that are ΔH higher than their neighbors as valid peak points. Here, ΔH represents the peak amplitude extracted based on the first dynamic curve, expressed as the difference between the accumulated current peak value and the baseline. The peak region is obtained by aggregating the time periods covered by the valid peak points.

[0060] The X-ray energy value and penetration intensity value are extracted based on the peak region. The following is an explanation of the terms involved: X-ray energy value: The actual tube voltage emitted by the X-ray source at the corresponding moment in the peak region, which is directly read by the built-in X-ray generator; Penetration intensity value: The signal response value received by X-rays penetrating the weld contour of the water-cooled plate in the weld scattering channel, which is directly extracted by the built-in weld scattering channel; A second dynamic curve is established with X-ray energy value / penetration intensity value on the horizontal axis and X-ray energy value on the vertical axis; It should be noted that the X-ray energy value / penetration intensity value represents the energy attenuation of X-rays in the weld scattering channel, and the second dynamic curve represents the change of energy attenuation with X-ray energy value in the peak region; When the system recognizes that the first dynamic curve has entered the peak region, the control loop of the X-ray source will dynamically adjust the exposure tube voltage according to the current theoretical material stack thickness, that is, dynamically adjust the X-ray energy value, so that the X-ray energy value and penetration intensity value generate nonlinear dynamic feedback in continuous time-series scanning, and then map the time-series motion trajectory representing different physical structure morphologies in the two-dimensional coordinate system;

[0061] The DTW algorithm is used to determine the curve type, specifically including: based on the second dynamic curve, identifying the temporal change trajectory; the DTW algorithm constructs a distance accumulation matrix to find the optimal alignment path between the temporal change set of the second dynamic curve and the preset shape image on the time axis; by calculating the minimum bending distance between the two, the type corresponding to the shape image with the minimum distance is determined as the current curve type.

[0062] If the time-series change trajectory is a straight upward line, similar to the side running trajectory of a shopping mall escalator rising at a constant speed, it indicates that the number of X-ray overlaps is a single layer and the physical thickness value is equal to the reference thickness. It is determined that the current area to be detected is unobstructed and the curve type is identified as the first curve type.

[0063] If the time-series change trajectory is elliptical and spiral-shaped, similar to the tilted visual projection after the small spring inside a ballpoint pen is gently stretched, when X-rays continuously sweep across the regularly arranged needle columns, the physical thickness value shows a periodic fluctuation of thick, thin, thick, thin, indicating that the number of X-ray overlaps is double and the physical thickness value fluctuates periodically in a step-like manner with the spatial arrangement of the heat dissipation needle columns. The penetration intensity shows a periodic characteristic distortion, and the physical structure of the current area to be detected is determined to be a regular needle column, and the curve type is identified as the second curve type.

[0064] If the time-series change trajectory shows an exponential decay, similar to the rapid descent of a high-altitude ski jump and the rapid flattening of the profile edge, it indicates that the number of X-ray overlaps has abruptly changed and the physical thickness value at the water-cooled plate welding interface shows a gradient unidirectional and drastic increase, resulting in a unidirectional and drastic deterioration of the X-ray penetration intensity. The curve type of the current area to be detected is determined to be the third curve type.

[0065] If the trajectory of the time change is hyperbolic, similar to the halo edge line formed by a flashlight shining obliquely on a flat wall at night, its trajectory will have a steep, large-angle smooth reversal in a very short local area, and then extend infinitely to both sides in a trumpet shape along the boundary. This indicates that the number of X-ray overlaps has reached the maximum value and the physical thickness has reached the maximum cross-sectional limit thickness. High-energy rays are completely shielded here, and the penetration intensity loses linear sensitivity. The physical structure of the current area to be detected is determined to be a multi-layered composite heavy blind zone, and the curve type is identified as the fourth curve type.

[0066] In addition, if the time-series change trajectory does not present the above image state, such as random sawtooth or cliff-like step shape, an emergency alarm signal is issued directly, and it is determined that there are pores and severe welding cracks in the current area to be detected, which is identified as a severe defect level and marked as a defect event.

[0067] The proportional weight combination is matched by curve type, including the first proportional weight combination, the second proportional weight combination, the third proportional weight combination and the fourth proportional weight combination, and the first proportional weight combination corresponds to the first curve type, the second proportional weight combination corresponds to the second curve type, the third proportional weight combination corresponds to the third curve type and the fourth proportional weight combination corresponds to the fourth curve type.

[0068] The X-ray overlap quantity and physical thickness values ​​are normalized to ensure they range from 0 to 1, eliminating dimensionality and giving the calculation results physical meaning. This process will not be elaborated further. An overlap / obstruction coefficient is determined by combining the normalized X-ray overlap quantity and the accumulated physical thickness of the needle column with corresponding proportional weights using an exponential weighting method. The X-ray overlap quantity is used as the base, and the first proportional weight in the weighting combination is used as the exponent of that base. The physical thickness value is used as the base, and the second proportional weight in the weighting combination is used as the exponent of that base. The first and second proportional weights are calculated, and the results of the two calculations are added together to obtain the overlap occlusion coefficient. The values ​​of the first proportional weight and the second proportional weight are in the range of 0 to 1, and the sum of the first proportional weight and the second proportional weight is 1. It should be noted that when the number of X-ray overlaps or the physical thickness value increases linearly in a single dimension, the subsequent first scattering image degradation usually shows a non-linear accelerated deterioration trend. The overlap occlusion coefficient quantifies the severity of image degradation caused by structural overlap. The larger the value, the more severe the image degradation caused by it. The smaller the overlap occlusion coefficient, the more severe the image degradation caused by it.

[0069] By abstracting the three-dimensional geometric coordinates of the pin columns in the water-cooled plate design layout into a set of feature nodes, and abstracting the scattering channels of the guide weld seam into a set of edges, an X-ray scattering topology network is constructed. By analyzing the number of X-ray overlaps and the accumulated physical thickness values, the first dynamic curve and the second dynamic curve are quantified. By screening and comparing different curve types, the severity of image degradation caused by structural overlap in each detection area inside the water-cooled plate is accurately and quantitatively evaluated, providing core data flow basis for subsequent adaptive correction of the control mesh tree and accurate defect judgment.

[0070] The identification and registration module acquires the first X-scattering image of the area to be detected and constructs a registration cost function by combining it with the X-scattering topology network. At the same time, it acquires the mechanical execution end vibration signal and welding infrared temperature signal of the known defect event, extracts the first feature parameter, and performs mapping association by combining the overlap occlusion coefficient to generate an association matrix and obtain the registration parameter range for each type of defect event.

[0071] An X-ray generator emits N0 X-rays to the area to be detected. A digital detector performs photoelectric conversion and sampling, outputting a two-dimensional grayscale matrix as the first X-ray image. A registration cost function is constructed through registration matching, including: morphological filtering and centroid extraction of the first X-ray image to quickly locate the first pixel coordinates of the center of each pin weld contour in the first X-ray image; identifying the three-dimensional coordinates of the feature node set in the X-ray topology network; retrieving the corresponding second pixel coordinates from the database; calculating the minimum Euclidean distance between the two pixel coordinates; if the minimum Euclidean distance is greater than or equal to a standard distance threshold, it indicates a weak correlation between the feature node and the weld contour, and the match is deemed unsuccessful; conversely, feature nodes with a minimum Euclidean distance less than the standard distance threshold are selected, indicating a strong correlation between the feature node and the weld contour, and the match is deemed successful. This initial screening between the first X-ray image and the X-ray topology network is achieved. By statically binding the initially screened feature nodes with the corresponding pin weld contour points, a registration mapping chain is established to form pin node pairs.

[0072] Standard distance threshold: First X-ray images of multiple batches of water-cooled plates under normal operation are acquired; the center of the pin weld contour in the image is located using an edge extraction algorithm to obtain a set of candidate pixels; simultaneously, the Euclidean distance between the two pixels is calculated and aggregated using the coordinates of the second pixel; the distance set is traversed and statistically analyzed to construct a frequency distribution histogram; based on this histogram, a probability density function is obtained by fitting a Gaussian kernel density estimate; through iterative training, the intersection point of the false negative rate and false positive rate curves is calculated, and the Euclidean distance corresponding to the intersection point is set as the standard distance threshold;

[0073] Based on the successfully matched needle-pillar node pairs identified by the registration mapping chain, feature matching residuals are constructed based on the minimum Euclidean distance sum of squares for each needle-pillar node pair, which are then determined as data terms of the registration cost function. Specifically, the data terms are used to constrain the overall geometric fit between the first X-ray image and the X-ray topology network. Accurate registration is achieved by collecting the minimum Euclidean distance sum of squares. Simultaneously, a smoothing term for the cost function is determined based on the absolute value of the difference in the overlap occlusion coefficients of the edge set under defect events. Specifically, during normal welding, the structure has continuity, and the absolute value of the difference in edge weights between adjacent edges approaches 0. When defects such as porosity or local incomplete penetration exist in the area to be detected, grayscale heterogeneity in the first X-ray image occurs, causing changes in the edge weights within the area to be detected. By introducing the absolute value of the difference in edge weights between adjacent edges of the same shared node into the smoothing term and accumulating them as a whole, a continuous smoothing constraint is applied to the X-ray topology network distortion caused by potential defect events during the registration process, thereby determining the smoothing term of the registration cost function.

[0074] The system retrieves preset first and second weighting coefficients, performs a linear weighted summation on the data terms and smoothing terms, and constructs a registration cost function: multiply the data terms by the first weighting coefficient, multiply the smoothing terms by the second weighting coefficient, and then sum the results to construct the registration cost function; wherein, the dynamic first and second weighting coefficients range from 0 to 1, and the sum of the first and second weighting coefficients is 1; it should be noted that the combination of data terms and smoothing terms can achieve accurate topological visualization registration of X-ray scattering images under high noise or multi-layer structure overlap interference, providing data and technical support for the location and quantitative analysis of internal defects in brazed water-cooled plates;

[0075] It should be noted that the first and second weight coefficients are determined based on the particle swarm optimization algorithm. The specific steps are as follows: In an S-dimensional target space, an initial population containing multiple particles is randomly generated. Each particle represents a combination of the first and second weight coefficients, and the state of each particle is characterized by its current velocity vector and position vector. The initial position of each particle is set as its historical best position. The current position vector of each particle is substituted into the registration cost function as a weight, and registration iteration is performed using a known registration mapping chain. After registration convergence, the root mean square error is calculated. A fitness function negatively correlated with the root mean square error is constructed. By comparing the fitness value obtained from the current fitness function of each particle with the fitness value of the historical best position, the larger value is selected to update the historical best position. All historical best positions of the current particle swarm are traversed, and the position with the absolute highest fitness value is found and defined as the global best position of the current generation. After reaching the maximum number of iterations, the iterative training is stopped, and the final first and second weight coefficients are output.

[0076] The registration parameter range for each type of defect is obtained, including: acquiring vibration signals from the mechanical end effector using a triaxial piezoelectric accelerometer. The triaxial piezoelectric accelerometer is adaptively mounted and typically fixed to the pressure mechanism or the outer surface of the spindle bearing housing at the end effector of the welding robot. During normal brazing welding, the mechanical end effector experiences stable force. However, if incomplete penetration occurs inside the water-cooled plate, transient nonlinear vibrations will occur at the end effector. This three-dimensional dynamic acceleration vibration signal is captured in real-time using the triaxial piezoelectric accelerometer. Time-frequency domain transformation is performed, including sliding window segmentation and Fourier transform: the acquired vibration signals from the mechanical end effector are segmented by a sliding window to extract time-domain features; Fourier transform is performed on the acquired vibration signals from the mechanical end effector to extract frequency-domain features; and the time-domain and frequency-domain features are concatenated to determine the vibration characteristics.

[0077] The infrared temperature signal of welding is collected by an infrared thermal analyzer. The infrared thermal analyzer is deployed synchronously with the mechanical end effector. Its working band is preferably 1.0μm to 1.7μm short-wave infrared or 8μm to 14μm long-wave infrared. It is usually deployed above the oblique side of the mechanical end effector to collect infrared temperature signals. The temperature gradient distribution is extracted by calculating the gradient of the execution area to determine the temperature distribution characteristics.

[0078] Based on vibration and temperature distribution characteristics, a first feature parameter is determined, and a first-order temporal difference matrix is ​​extracted to determine a first change vector, which is used to capture the peak transients caused by instantaneous obstruction or pressure change at the end of mechanical actuation. Simultaneously, the cosine similarity between the first feature parameter and the standard defect-free feature parameters stored in the database is calculated to determine a second change vector, which is used to characterize the slow global drift of the temperature field. In this way, multi-source and nonlinear monitoring signals are successfully transformed into highly sensitive and time-aligned change vectors. The overlap and occlusion coefficient is retrieved, and a first time label is established for the overlap and occlusion coefficient according to the time series, constructing an association matrix between the first change vector and the overlap and occlusion coefficient. The overlap and occlusion coefficient is retrieved again, and a second time label is established for the overlap and occlusion coefficient according to the time series, constructing an association matrix between the second change vector and the overlap and occlusion coefficient.

[0079] Based on the correlation matrix, all time tags that trigger defect events are identified and extracted. The first recorded time tag is taken as the intermediate moment, and two step sizes are extracted bidirectionally and symmetrically. The time sequence corresponding to the first change vector and the second change vector is extracted respectively and combined to form a defect time period. The time sequence is sorted to form a dynamic time window. Through equal division, a dynamic score segment containing multiple sub-time periods is determined, including the first dynamic score segment and the second dynamic score segment.

[0080] Cluster analysis is performed within different dynamic score segments to identify defect events, including first and second defect events. If the equal division results have remainders, the first dynamic score segment is determined. At this time, the data exhibits irregular transient steps, and manifold clustering is performed. The corresponding first and second change vectors are extracted and formed into a fused feature vector through feature concatenation. Based on the fused feature vector, the first derivative is calculated, and monotonically increasing or decreasing monotonic intervals are extracted. By identifying extreme points as initial seed points, the radius of the topologically connected neighborhood is preset, and manifold expansion is performed on the fused feature vector. Starting from the seed point, the distance between the seed point and other surrounding feature points is calculated. Seed points whose distance is greater than the radius of the topologically connected neighborhood are selected and merged into irregular manifold defect clusters, which are determined as first defect events, such as large particle spatter, severe local blowout of welds, or hard slag inclusions.

[0081] If the equal division results in no remainder, determine the second dynamic fraction segment. At this point, the data exhibits a stationary distribution, and ellipsoidal clustering is performed: the corresponding first and second change vectors are extracted as observation vectors, and the expectation-maximization algorithm is used for iterative training: the probability value of the observation vector belonging to a Gaussian distribution is calculated, and the mean and covariance matrix of the distribution are recalculated and dynamically updated based on the probability value; the eigenvalues ​​of the covariance matrix are decomposed, and the first and second eigenvalues ​​of the top two ranked eigenvalues ​​are extracted using principal component analysis, and the square roots of the corresponding eigenvalues ​​are calculated, including the first and second square roots; adjustment... A confidence coefficient, typically 2, is used. The first square root is multiplied by the confidence coefficient to obtain the radius of the major axis interval. The second square root is multiplied by the confidence coefficient to obtain the radius of the minor axis interval. Observation vectors that fall within the oblique ellipsoidal interval jointly defined by the major and minor axis intervals are selected and merged into oblique ellipsoidal defect clusters, which are then identified as the second defect event. For example, continuous large-area shallow incomplete penetration caused by a continuously low overall temperature of the brazing furnace or a systematically thin solder coating. The registration parameter range for each type of defect event is divided and output based on the first and second defect events.

[0082] The first X-ray scattering image acquired is fused across domains with the X-ray scattering topology network and the monitoring signals during the welding process. The introduction of mechanical execution end vibration signals and welding infrared temperature signals can significantly improve the accuracy of defect event judgment. By constructing time labels, the first and second change vectors are correlated and aligned with the overlap and occlusion coefficients of the X-rays under the time label on the time axis to form corresponding correlation matrices, thereby obtaining the registration parameter range for each type of defect event. This approach realizes cross-modal spatiotemporal coupling of dynamic monitoring signals and static images, and adaptively adjusts the registration parameter range for different defect events, effectively enhancing the accurate localization and classification of subsequent defect events and improving the robustness of the model.

[0083] Iterative alignment module: Constructs a control grid tree based on feature node set and overlap occlusion coefficient, introduces a rule engine, determines the correction parameters of the region to be detected by minimizing the registration cost function, and transforms the first X-scattering image into the second X-scattering image;

[0084] Constructing a control mesh tree by combining the overlap occlusion coefficient, including:

[0085] The envelope boundary is determined based on the feature node set and defined as the root node of the control mesh tree. The geometric center of the feature node set is used as the reference to establish the initial coordinate system of the root node. Based on the root node, hierarchical spatial recursion is performed. The specific steps are as follows: quadtree partitioning is used. In each iteration, the current parent mesh unit is bisected in the initial coordinate system at equal intervals in both horizontal and vertical directions, and cut into four child mesh units. Corresponding mesh nodes are generated at the intersection and boundary of the partition.

[0086] The system maps and calculates all overlapping occlusion coefficients within the area covered by the current sub-grid cell. It extracts the first quartile, second quartile, and third quartile and calculates their arithmetic mean. If the arithmetic mean is greater than or equal to the standard blocking threshold, it indicates that the grid cell structure is complex and excessive segmentation will lead to severe image registration artifacts. The rule engine immediately triggers a blocking operation, terminating further splitting of the grid cell to maintain a large grid size. The current grid cell is locked as a leaf node of the control grid tree. If the mean is less than the standard blocking threshold, it indicates that the image of the grid cell is clear and the pixels are reliable. The system allows it to continue to recursively split downwards to generate dense grid nodes and continue iterating to build the control grid tree.

[0087] Standard blocking threshold: Based on historical data, analyze defect events that occur during the welding process of brazed water-cooled plates, collect all overlap occlusion coefficient data under a certain defect event, extract the first quartile, second quartile, and third quartile, and calculate their arithmetic mean; understand the range of arithmetic mean during quadtree segmentation, extract all arithmetic means that did not trigger blocking operations, and through statistical analysis, set the standard blocking threshold to the average of the corresponding arithmetic means plus a standard deviation of 2 or 3 times; it should be noted that the specific value of the multiple is just an example, and the specific setting should be based on the actual situation, which will not be elaborated here.

[0088] The correction parameters for the region to be detected are determined, including: introducing a rule engine, which has built-in geometric constraints on the control grid tree and convergence constraints on the iterative solution; among which, the geometric constraints include the maximum displacement vector threshold of the grid nodes; the convergence constraints include the control gradient step size and the iteration error; the following is an explanation of some of the terms involved: maximum displacement vector threshold: constrains the maximum drift distance of any grid node on the control grid tree in a single iteration or the entire iteration process; control gradient step size: the learning rate; convergence error threshold: sets the range of the change difference of the registration cost function over three consecutive iterations;

[0089] Under the rule engine, gradient descent is used to iteratively solve the registration cost function. When the value of the registration cost function reaches a minimum, the state vector of the current iteration is output as a correction parameter. It should be noted that the state vector is specifically a parameter vector indexed by the control grid tree. The specific steps to obtain it include: traversing all unblocked grid nodes in the entire control grid tree, including branch nodes and leaf nodes at each level, and assigning a unique two-dimensional dynamic displacement independent variable pair (x_lm, y_lm) to each grid node in the initial coordinate system of the root node, where x and y represent coordinates. 'l' represents the node's level depth in the control grid tree, and 'm' represents the node's topology index at that level. All grid nodes are concatenated and stitched together in the DFS order of root node, sub-level, and leaf node to form a state vector. When the registration cost function reaches its minimum convergence, the state vector is directly calculated as the correction parameter for the region to be detected and output in layers. By reversing the search of the control grid tree, the state vector can be directly mapped back to the corresponding grid cell. Through local non-rigid interpolation, the first X-scattering image with geometric distortion is transformed into a standard-aligned second X-scattering image.

[0090] A control grid tree is constructed based on the feature node set and the overlap occlusion coefficient. The constructed hierarchical control grid tree can achieve local fine registration at different resolutions. For areas with high overlap occlusion coefficients, a finer control grid is automatically assigned. By introducing a rule engine, constraint rules are assigned to the displacement vector of each grid node, and gradient descent is used to iteratively optimize the registration cost function to avoid getting trapped in local optima. While ensuring sub-pixel level registration accuracy, the iteration convergence speed is greatly improved. The first X-scattering image is transformed into a second X-scattering image to achieve high-precision correction of the water-cooled plate, reduce misjudgments in subsequent defect identification, and improve the accuracy of identification.

[0091] Defect identification module: Traverses the control grid tree to calculate the artifact cost, merges the paths of the second X-scattering image to determine the third X-scattering image, extracts the second feature parameters, and outputs the defect identification result by combining the registration parameter range;

[0092] The cost of obtaining artifacts includes: setting a first spatial scale, calculating the displacement vector of the current pin-pillar node pair in the state vector, and subtracting it from the historical displacement vector of the previous adjacent pin-pillar node pair in the state vector to obtain the difference vector between the two; continuously acquiring the difference vectors and calculating the angular variance between these difference vectors, which reflects the disorder of local mesh deformation, and quantifying it into a spatial stability index by taking the reciprocal of the angular variance; it should be noted that the smaller the value of the spatial stability index, the greater the probability of surface deformation.

[0093] Simultaneously, a preset X-ray scattering direction is used to extract the projection component of the difference vector onto the X-ray scattering direction vector, which is then quantified as an artifact index. The dot product between the difference vector and the preset X-ray scattering direction vector is calculated to obtain the algebraic projection value of the difference vector onto the X-ray scattering direction. The absolute value of this algebraic projection value is then taken as the projection component magnitude. Subsequently, the ratio between the projection component magnitude and the magnitude of the X-ray scattering direction vector itself is calculated, and this ratio is used as the artifact index for the current pin-column node pair. It should be noted that a higher artifact index value indicates a higher degree of fit between the displacement deformation trend of the current mesh node and the fringe orientation of the X-ray scattering artifact.

[0094] A second spatial scale is preset, the control grid tree is traversed, and the displacement vectors of several consecutive needle-column node pairs are linearly regressed to extract the slope of the fitting process, which is then marked as the trend change rate. The artifact cost is generated through weighted aggregation of spatial stability indicators, artifact indicators, and trend change rates. The specific steps are as follows:

[0095] First, the three indicators are normalized and mapped to the 0-1 interval. The spatial stability indicator uses min-max normalization, while the artifact indicator and trend rate of change are standardized using Z-score and then mapped to the 0-1 interval using the Sigmoid function. These are then concatenated into an input feature vector: [spatial stability indicator, artifact indicator, trend rate of change]. A first feature transformation is performed on the input feature vector, using a learnable weight matrix to map the 3D input features into an 8D intermediate feature vector. The transformed feature vector is then compared element-wise with the original input feature vector after dimensional expansion. The features are summed and non-linearly mapped using the ReLU activation function, outputting an 8-dimensional first intermediate feature vector. A second feature transformation is performed on the first intermediate feature vector, using a learnable weight matrix to map the 8-dimensional first intermediate feature vector into a 4-dimensional second intermediate feature vector. The transformed feature vector is then added element-wise to the feature vector of the first intermediate feature vector after dimensionality compression, and non-linearly mapped using the Sigmoid activation function, outputting a 4-dimensional second intermediate feature vector. Global average pooling is then performed on the second intermediate feature vector, and the arithmetic mean of all elements of the 4-dimensional feature vector is taken as the artifact cost of the grid node.

[0096] The minimum spanning tree algorithm is adopted. By iteratively expanding the unconnected pin nodes corresponding to the minimum projection cost, the unconnected pin nodes are added to the X-ray scattering topology network in turn, and the control grid tree is updated. All pin nodes in the first X-ray transmission image are defined as the vertex set of an undirected graph, and the connecting edges between adjacent nodes are defined as the edge set. First, the pre-stored overlap occlusion coefficient between the current pin node pairs is retrieved as the basic edge weight. The calculated artifact cost is used as a dynamic adjustment factor. The overlap occlusion coefficient and the artifact cost are linearly mapped to generate the final edge weight. The pin node with the lowest artifact cost, that is, the highest registration confidence and the clearest image, is retrieved and located. It is added to the X-ray transmission topology network as the initial known node. In each iteration, all adjacent edges of the currently established X-ray scattering topology network are traversed to find the unconnected pin node with the minimum artifact cost. The unconnected pin node is added to the X-ray transmission topology network in turn until all pin nodes are covered.

[0097] Determining the third X-ray scattering image includes: retrieving the artifact cost generated for the pin nodes in the current control grid tree, and comparing the artifact cost with the standard cost threshold.

[0098] The overlapping path is determined based on the control grid tree. The overlapping path is used to represent the topological connection edge between each adjacent pin column node in the updated control grid tree. After performing non-rigid coordinate transformation mapping in the two-dimensional coordinate system of the image, the corresponding local image overlapping area or pixel alignment trajectory is obtained. Each overlapping path is rigidly associated with its corresponding artifact cost.

[0099] If the artifact cost is less than the standard cost threshold, the local region to which the path belongs is determined to have high registration confidence, and path merging is performed: weighted average fusion at the image pixel level is performed along the corresponding overlapping paths to complete the path merging, and all image data after path merging and pixel fusion are matrix stitched and memory-stabilized, and the merged and fused image is output as the final third X-ray scattering image.

[0100] If the artifact cost is greater than or equal to the standard cost threshold, the local region to which the path belongs is determined to have low registration confidence, and the path traversal continues: keep the current second X-scatter image data without pixel-level fusion, drive the hierarchical topology pointer of the control grid tree by traversing the artifact cost, retrieve the overlapping path and its artifact cost corresponding to the next adjacent topological connection edge, and return to perform a new round of artifact cost and standard cost threshold comparison and judgment until all overlapping paths in the entire tree have been traversed;

[0101] Standard cost threshold: The system performs multiple traversal calculations on normal water-cooled plates and water-cooled plates with defect events in the sample library to obtain the artifact cost probability density functions of the two types of samples. The overlapping intersection point of the two distributions is calculated using Bayesian decision theory, and this point is set as the standard cost threshold. For example, if the artifact cost of 1.51 is compared with the standard cost threshold of 2.00 in real time, and 1.51 is less than 2.00, path merging is triggered. Image pixel-level weighted average fusion is performed on the overlapping pixels of the overlapping path to smooth and suppress artifacts. Finally, the fused image data is established as a high-definition third X-ray scattering image. Conversely, if the artifact cost calculated at another needle-pillar node pair reaches or exceeds the standard cost threshold of 2.00, the rule engine determines that there is a real nonlinear material defect fracture at this point. The system will adaptively terminate the weighted average fusion operation, keep the original image edge data of the original path unchanged, and directly continue to execute the traversal of the subsequent paths.

[0102] Based on the third X-ray scattering image, a second feature parameter is extracted, and the time label of the feature extraction is obtained to match the corresponding dynamic score segment. The registration parameter range corresponding to each defect event under the dynamic score segment is retrieved, and the second feature parameter and the registration parameter range are compared and verified for consistency. If the second feature parameter falls within the registration parameter range of a specific defect event, it is determined that the structural anomaly in the third X-ray scattering image is strongly correlated with the defect event, and the defect identification result containing the defect type and location is output. If the second feature parameter does not fall within the registration parameter range of any type of defect event under the dynamic score segment, it is determined that the anomaly is caused by process fluctuations or residual scattering energy artifacts, and automatic filtering is performed.

[0103] By performing two spatial scale analyses on the second X-ray image, traversing the control grid tree, calculating the artifact cost, and performing intelligent fusion on the overlapping path images of multi-channel scanning, the continuous and complete shape of the weld was restored. False artifact edges were removed to establish a third X-ray image that truly reflects the solid structure of the water-cooled plate. By combining the registration parameter range for joint judgment and outputting defect identification results, the false alarm rate caused by artifact interference was reduced, achieving high-precision and high-reliability automatic identification of weld defects in brazed water-cooled plates.

[0104] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0105] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0106] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A visual inspection system for weld defects in brazed water-cooled plates, characterized in that, The system includes: The topology design module obtains the water-cooled plate design layout of the area to be inspected, extracts the needle columns and abstracts them into a set of feature nodes, extracts the weld scattering channels between adjacent needle columns and abstracts them into a set of edges, and constructs an X-scattering topology network; where each feature node corresponds to the three-dimensional coordinates of the needle column, and the weight of each edge is determined by the overlap occlusion coefficient of the weld scattering channel. The identification and registration module acquires the first X-scattering image of the area to be detected and constructs a registration cost function by combining it with the X-scattering topology network. At the same time, it acquires the mechanical execution end vibration signal and welding infrared temperature signal of the known defect event, extracts the first feature parameter, and performs mapping association by combining the overlap occlusion coefficient to generate an association matrix, thereby obtaining the registration parameter range for each type of defect event. The iterative alignment module constructs a control grid tree based on the feature node set and the overlap occlusion coefficient, introduces a rule engine, determines the correction parameters of the region to be detected by minimizing the registration cost function, and transforms the first X-scattering image into the second X-scattering image. The defect identification module traverses the control grid tree to calculate the artifact cost and merges the paths of the second X-ray image to determine the third X-ray image. It then extracts the second feature parameters and outputs the defect identification result by combining the registration parameter range.

2. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, The steps for determining the overlap occlusion coefficient include: Analyze the design layout of the water-cooled plate and establish projection rays from N0 X-rays pointing to the weld contour of the water-cooled plate; The collision detection algorithm is used to count the intersection frequency of each projected ray with the needle column to determine the number of X-ray overlaps, and the physical thickness value of the needle column is accumulated along the penetration trajectory. Based on the number of X-ray overlaps and the accumulated physical thickness value of the weld scattering channel, a first dynamic curve is established and the peak region is identified. By extracting the X-ray energy value and penetration intensity value, a second dynamic curve is established. The DTW algorithm is used to determine the curve type and match the proportional weight combination. The overlap and occlusion coefficient is determined by exponential weighting.

3. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 2, characterized in that, Construct the registration cost function, including: Based on the first X-ray image, the first pixel coordinates of the weld contour of each water-cooled plate are extracted; the three-dimensional coordinates of the feature node set in the X-ray topology network are identified, the second pixel coordinates corresponding to the three-dimensional coordinates are retrieved from the database, the minimum Euclidean distance between the two pixel coordinates is calculated, the feature nodes with a minimum Euclidean distance less than the standard distance threshold are selected, and a registration mapping chain is established to identify pin-column node pairs. The feature matching residual is constructed based on the sum of the squared minimum Euclidean distances between each pair of needle and column nodes, and the data terms of the registration cost function are determined. At the same time, the smoothing term of the registration cost function is determined based on the absolute value of the difference in edge weights between any two adjacent edges sharing the same feature node under the defect event. The dynamically updated first and second weights are retrieved, and the data terms and smoothing terms are linearly weighted and summed to construct the registration cost function.

4. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, Obtain the registration parameter range for each type of defect, including: The first characteristic parameters include vibration characteristics and temperature distribution characteristics: Vibration characteristics are determined based on the vibration signal at the mechanical actuator end, and the first change vector is determined by time-series differential analysis. The temperature distribution characteristics are determined based on the welding infrared temperature signal, and the second change vector is determined by similarity comparison. Retrieve the overlapping occlusion coefficients, match time labels for the overlapping occlusion coefficients according to the time series, including the first time label and the second time label; establish the correlation matrices between the first change vector, the second change vector and the overlapping occlusion coefficients respectively; Extract all time tags that trigger the defect event, take the first recorded time tag as the intermediate time, and extract two step lengths in a bidirectional symmetrical manner. Extract the step lengths corresponding to the first change vector and the second change vector respectively, combine them to form the defect time period, sort them in chronological order to form a dynamic time window and process them equally to generate multiple sub-time period dynamic score segments, including the first dynamic score segment and the second dynamic score segment. An adaptive clustering algorithm is used to select different dynamic score segments to divide the first change vector and the second change vector clusters, and output the registration parameter range for each type of defect event; wherein, the defect events include the first defect event and the second defect event.

5. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 4, characterized in that, For different dynamic score ranges, adaptive clustering algorithms include: If there is a remainder in the equal division result, determine the first dynamic score segment and perform manifold clustering: extract the corresponding first and second change vectors to form a fused feature vector, extract the monotonic intervals and identify the extreme points as initial seed points, preset the topological connected neighborhood radius, perform manifold expansion on the fused feature vector, filter out seed points larger than the topological connected neighborhood radius, and merge them into an irregular manifold defect cluster, which is determined as the first defect event; If the equal division result has no remainder, determine the second dynamic score segment and perform ellipsoidal clustering: extract the corresponding first and second change vectors as observation vectors, use the expectation-maximization algorithm for iterative training, extract the mean and covariance matrix, select observation vectors that fall into the oblique ellipsoidal interval consisting of the major axis interval and the minor axis interval, and merge them into oblique ellipsoidal defect clusters, which are determined as the second defect event.

6. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, Determine the correction parameters for the area to be detected, including: A rule engine is introduced, which has built-in geometric constraints and convergence constraints for iterative solution of the control grid tree; among them, the geometric constraints include the maximum displacement vector threshold of the grid nodes; the convergence constraints include the control gradient step size and the convergence error threshold. Under the rule engine, gradient descent is used to iteratively solve the registration cost function. When the registration cost function reaches a minimum value, the state vector of the current iteration is output as a correction parameter.

7. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, The costs of obtaining artifacts include: A first spatial scale is preset, the displacement vector of the current needle-pillar node pair in the state vector is calculated, and the difference vector between it and the historical displacement vector of the previous adjacent needle-pillar node pair in the state vector is calculated. The variance of the angle between the continuous difference vectors is calculated to determine the spatial stability index. At the same time, the X-ray scattering direction is preset, the projection component of the difference vector on the X-ray scattering direction vector is extracted, and quantified into an artifact index. A second spatial scale is preset, and the displacement vectors of several consecutive needle-column node pairs are linearly regressed to determine the trend change rate; the artifact cost is generated by weighted aggregation of spatial stability index, artifact index and trend change rate. The minimum spanning tree algorithm is adopted. The unconnected pin nodes corresponding to the minimum projection cost are expanded iteratively and added to the X-scattering topology network in turn to update the control grid tree.

8. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, Constructing a control mesh tree by combining the overlap occlusion coefficient, including: The envelope boundary is identified based on the feature node set as the root node, and the initial coordinate system of the root node is established. By performing hierarchical spatial recursion, in each iteration, the current parent mesh unit is equally divided into four child mesh units, and the corresponding mesh nodes are generated. Map and calculate all overlapping occlusion coefficients within the area covered by the current sub-grid cell. Extract multiple quantiles and take the arithmetic mean. If the calculated result is greater than or equal to the standard blocking threshold, terminate the iteration; if the calculated result is less than the standard blocking threshold, continue the iteration to construct the control grid tree. The quantiles include the first quartile, the second quartile, and the third quartile.

9. The visual inspection system for weld defects in a brazed water-cooled plate according to claim 1, characterized in that, Determine the third X-ray scattering image, including: Determine overlapping paths based on control grid trees; Compare the artifact cost with the standard cost threshold: If the artifact cost is less than the standard cost threshold, perform path merging: perform image pixel-level weighted average fusion on the corresponding overlapping paths, and use the merged image as the third X-scatter image; if the artifact cost is greater than or equal to the standard cost threshold, keep the second X-scatter image and continue path traversal.

Images