Intelligent welding feature recognition method and system for complex components

Abstract

This invention discloses an intelligent welding feature recognition method and system for complex components, belonging to the field of welding technology. The method acquires multi-source sensing data of the complex component to be welded; performs time synchronization, coordinate unification, and preprocessing on the multi-source sensing data; establishes or invokes a prior model of the complex component; extracts welding feature candidates from the preprocessed data and limits the search range by combining robot pose data; registers and fuses the welding feature candidates with the prior model to obtain a fusion score; updates the confidence of the welding feature candidates based on the fusion score to determine valid welding features; and outputs the weld start point, weld end point, weld centerline, local posture reference, and recognition confidence based on the valid welding features. When local features are missing, the missing segment is compensated according to the continuity constraints of the prior model, and a compensation mark is output. This invention improves the stability, continuity, and accuracy of welding feature recognition, providing a reliable basis for welding robots.

Description

Technical Field

[0001] This invention belongs to the field of welding automation technology, specifically relating to a method and system for intelligent welding feature recognition of complex components. Background Technology

[0002] In the automated welding of complex components, welding robots typically need to accurately identify the weld centerline, groove boundary, joint endpoint, and local attitude reference before or during welding in order to generate welding trajectories and maintain the stability of the welding torch attitude.

[0003] In existing technologies, weld seam identification is often performed using a single visual image or a single laser profile. When components have assembly deviations, surface reflections, occlusions, spatter contamination, gap variations, or curved surface transitions, a single sensor is prone to edge false detections, centerline interruptions, or endpoint drift, which can lead to welding trajectory deviations.

[0004] Other solutions, while introducing multiple sensors, simply overlay them without incorporating the nominal geometry of the components, assembly tolerances, and weld layout information as prior constraints into the recognition process. This results in difficulty in consistently outputting recognition results that can be used for welding planning when there are many candidate features or when some features are missing.

[0005] Therefore, it is necessary to propose a complex component welding feature recognition scheme that combines multi-source sensing data with prior models to improve the accuracy and robustness of welding feature extraction. Summary of the Invention

[0006] To address the problems of insufficient stability, poor continuity, and significant influence from assembly deviations in the identification of welding features of complex components in existing technologies, this invention provides an intelligent welding feature identification method and system for complex components. The aim is to output welding feature results that can be directly used for welding robot trajectory planning by using multi-source sensing data as a basis and the component prior model as a constraint, through feature candidate construction, registration and fusion, confidence update and missing segment compensation.

[0007] To achieve the above objectives, the specific solution of the present invention is as follows:

[0008] A method for intelligent welding feature recognition of complex components includes the following steps:

[0009] S1. Acquire multi-source sensing data of the complex component to be welded, wherein the multi-source sensing data includes at least image data, contour data and robot pose data;

[0010] S2. Perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system;

[0011] S3. Establish or call the prior model of the complex component to be welded. The prior model shall include at least one or more of the following: nominal weld centerline, weld start and end position, joint boundary information, and assembly tolerance range.

[0012] S4. Extract welding feature candidates from the preprocessed data and limit the candidate search range by combining the robot pose data;

[0013] S5. Register and fuse the welding feature candidates with the prior model to obtain the fusion score of each welding feature candidate;

[0014] S6. Update the confidence of welding feature candidates based on the fusion score within the continuous sampling period, and determine the valid welding features according to the preset judgment rules.

[0015] S7. Based on the effective welding features, output the weld start point, weld end point, weld center line, local attitude reference, and identification confidence level. When local features are missing, compensate for the missing segment according to the continuity constraint of the prior model and output the missing segment compensation mark.

[0016] Furthermore, the image data in step S1 is acquired by an industrial camera, which has an industrial camera coordinate system; the contour data is acquired by a contour sensor, which has a contour sensor coordinate system; and the robot pose data is robot end-effector pose data, which is output by a robot pose encoder, which has a robot end-effector coordinate system.

[0017] Further, the time synchronization in step S2 adopts a unified trigger pulse synchronization method or a timestamp-based alignment method; the coordinate unification includes transforming the sensor coordinate system for acquiring image data, the sensor coordinate system for acquiring contour data, and the sensor coordinate system for acquiring robot pose data to the workpiece reference coordinate system or the robot base coordinate system; the preprocessing includes preprocessing the image data, contour data, and robot pose data respectively, wherein the preprocessing of the image data includes at least one of median filtering, grayscale equalization, and region of interest cropping; the preprocessing of the contour data includes at least one of outlier removal, smoothing filtering, and contour resampling; and the preprocessing of the pose data includes at least one of outlier removal and smoothing processing.

[0018] Furthermore, the prior model described in step S3 is used to provide nominal positional constraints and continuity constraints for welding features, and consists of at least two of the following: CAD model, offline teaching trajectory, nominal weld layout, joint type parameters, and assembly tolerance parameters.

[0019] Furthermore, the extraction of welding feature candidates in step S4 includes at least one of the following: extracting candidate boundaries based on straight line fitting or curve fitting of image edges, extracting weld center candidates based on contour depth valleys, extracting groove boundary candidates based on contour curvature abrupt change points, and defining the candidate search area based on pose sliding windows.

[0020] Furthermore, the prior model described in step S3 is used to provide nominal positional constraints and continuity constraints for welding features. The prior model consists of at least two of the following: CAD model, offline teaching trajectory, nominal weld layout, joint type parameters, and assembly tolerance parameters.

[0021] Furthermore, the extraction of welding feature candidates in step S4 includes at least one of the following: extracting candidate boundaries based on straight line fitting or curve fitting of image edges, extracting weld center candidates based on contour depth valleys, extracting groove boundary candidates based on contour curvature abrupt change points, and defining the candidate search area based on pose sliding windows.

[0022] Further, the registration and fusion step in step S5 is as follows: first, coarse registration is performed on the welding feature candidate and the prior model based on the robot pose data and the preset tooling reference; then, local matching is performed based on the local morphological differences to obtain the deviation of the welding feature candidate relative to the prior model.

[0023] The formula for calculating the fusion score is as follows:

[0024] F = w1·G + w2·P + w3·M,

[0025] In the formula, G represents the image feature continuity score extracted from the image data; P represents the contour feature saliency score extracted from the contour data; M represents the weighted sum of the matching scores between the welding feature candidate and the prior model; w1, w2, and w3 represent preset weights or weights adaptively adjusted according to the signal quality of the current sampling period.

[0026] Furthermore, the confidence update formula in step S6 is as follows:

[0027] C k =α·C k-1 +(1-α)·F,

[0028] In the formula, C k C represents the confidence level for the current sampling period. k-1 α is the confidence level of the previous sampling period; α is a smoothing coefficient between 0 and 1; when the updated confidence level is higher than the preset threshold, the corresponding welding feature candidate is determined as a valid welding feature.

[0029] Furthermore, the local attitude reference mentioned in step S7 includes at least one of the weld seam normal, weld seam tangent, and section normal, and is used for welding robot torch attitude adjustment or welding trajectory correction.

[0030] Furthermore, the local feature loss mentioned in step S7 includes:

[0031] When the image features extracted from the image data are missing, but the fusion score corresponding to the contour features extracted from the contour data meets the preset judgment rule, the corresponding welding features are retained.

[0032] When both the image features extracted from the image data and the contour features extracted from the contour data are missing, and there are valid welding features before and after the missing segment, the missing segment is interpolated based on the continuity constraints of the prior model and the valid welding features before and after.

[0033] A complex component intelligent welding feature recognition system for implementing the method includes a multi-source sensing unit, a preprocessing and synchronization unit, a priori model unit, a fusion recognition unit, a recognition output unit, and a welding execution interface;

[0034] A multi-source sensing unit is used to acquire multi-source sensing data of a complex component to be welded, wherein the multi-source sensing data includes at least image data, contour data and robot pose data.

[0035] The preprocessing and synchronization unit is used to perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system.

[0036] A priori model unit is used to store a priori models of complex components to be welded. The priori models include at least one or more of the following: nominal weld centerline, weld start and end positions, joint boundary information, and assembly tolerance range.

[0037] The fusion recognition unit is used to extract welding feature candidates from the preprocessed data, register and fuse the welding feature candidates with the prior model, calculate the fusion score of each welding feature candidate using the fusion scoring model F=w1·G +w2·P+w3·M, and update the confidence model C. k =α·C k-1 +(1-α)·F performs confidence updates and determines valid welding features based on preset judgment rules;

[0038] The identification output unit is used to output the weld start point, weld end point, weld center line, local attitude reference, and identification confidence level based on the effective welding features. When local features are missing, the missing segment is compensated according to the continuity constraint of the prior model, and a missing segment compensation mark is output.

[0039] The welding execution interface is used to transmit the information output by the identification output unit to the welding robot.

[0040] Advantages of the present invention

[0041] 1. This invention achieves a fusion score by multi-source fusion of image data, contour data, and robot pose data, and by weighting the image feature continuity score, contour feature saliency score, and prior model matching score. This comprehensive assessment of whether candidate features belong to real welding features reduces false detections caused by interference from factors such as reflection and spatter from a single sensor, thereby improving the stability of welding feature recognition.

[0042] 2. This invention introduces a priori model of the complex component to be welded, providing nominal positional and continuity constraints for welding features. When local features are missing, the missing segment is interpolated or the corresponding welding features are retained based on the continuity constraints of the priori model and the effective welding features before and after it, thereby maintaining the continuity of the identification results under assembly deviations or local feature loss conditions.

[0043] 3. The identification results output by this invention include the weld start point, weld end point, weld centerline, and local attitude reference, wherein the local attitude reference includes at least one of the weld normal, weld tangent, and cross-section normal. The welding robot can automatically generate a welding trajectory based on this, or fine-tune the original trajectory during the welding process, while simultaneously adjusting the welding torch attitude using the local attitude reference, thereby achieving automated control of the welding process.

[0044] 4. The overall solution of this invention revolves around welding feature recognition. From multi-source sensing data acquisition, data preprocessing, prior model establishment, welding feature candidate extraction, registration and fusion, confidence update to recognition result output, each step serves the accurate recognition of welding features, which facilitates the establishment of a clear correspondence with welding automation scenarios. Attached Figure Description

[0045] Figure 1 This is a flowchart of the intelligent welding feature recognition method for complex components according to the present invention.

[0046] Figure 2 This is a structural block diagram of the intelligent welding feature recognition system for complex components according to the present invention.

[0047] Figure 3 This is a schematic diagram illustrating the principle of multi-source sensing data and prior model fusion recognition in this invention.

[0048] Figure 4 This is a schematic diagram showing the weld centerline coordinate sequence, weld start point, weld end point, local attitude reference, identification confidence level, missing segment compensation mark, and welding planning application of the output of this invention. Detailed Implementation

[0049] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments are not intended to limit the scope of the present invention.

[0050] like Figure 1 As shown in the figure, this specific embodiment provides a method for intelligent welding feature recognition of complex components, which includes the following steps:

[0051] S1. Acquire multi-source sensing data of complex components to be welded.

[0052] The multi-source sensing data includes at least image data acquired by an industrial camera, contour data acquired by a contour sensor, and robot end-effector pose data output by a robot pose encoder; the industrial camera has an industrial camera coordinate system; the contour sensor has a contour sensor coordinate system; and the robot pose encoder has a robot end-effector coordinate system.

[0053] S2. Perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system;

[0054] Specifically, the time synchronization adopts a unified trigger pulse synchronization method or a timestamp-based alignment method; the coordinate unification includes transforming the sensor coordinate system for acquiring image data, the sensor coordinate system for acquiring contour data, and the sensor coordinate system for acquiring robot pose data to the workpiece reference coordinate system or the robot base coordinate system; the preprocessing includes preprocessing the image data, contour data, and robot pose data respectively, wherein the preprocessing of the image data includes at least one of median filtering, grayscale equalization, and region of interest cropping; the preprocessing of the contour data includes at least one of outlier removal, smoothing filtering, and contour resampling; and the preprocessing of the pose data includes at least one of outlier removal and smoothing.

[0055] S3. Create or call up a priori model of the complex component to be welded.

[0056] The prior model is used to provide nominal positional and continuity constraints for welding features, and consists of at least two of the following: CAD model, offline teaching trajectory, nominal weld layout, joint type parameters, and assembly tolerance parameters. The prior model includes at least one or more of the following: nominal weld centerline, weld start and end positions, joint boundary information, and assembly tolerance range.

[0057] S4. Extract welding feature candidates from the preprocessed data and limit the candidate search range by combining the robot pose data;

[0058] The extraction of welding feature candidates includes at least one of the following: extracting candidate boundaries based on Hough line fitting or curve fitting based on image edges; extracting weld center candidates based on contour depth valleys; extracting bevel boundary candidates based on contour curvature abrupt change points; and defining the candidate search area based on pose sliding windows.

[0059] S5. Register and fuse the welding feature candidates with the prior model to obtain the fusion score of each welding feature candidate;

[0060] The registration and fusion process involves: first, coarse registration is performed between the welding feature candidates and the prior model based on the robot pose data and a preset tooling reference to eliminate overall component positional deviations; then, local matching is performed based on local morphological differences to compensate for local assembly deformations, thereby obtaining the deviation of the welding feature candidates relative to the prior model. The prior model registration can employ rigid transformation registration or local elastic compensation registration.

[0061] The formula for calculating the fusion score is as follows:

[0062] F = w1·G + w2·P + w3·M,

[0063] In the formula, G represents the image feature continuity score extracted from the image data; P represents the contour feature saliency score extracted from the contour data; M represents the weighted sum of the matching scores between the welding feature candidate and the prior model; w1, w2, and w3 represent preset weights or weights adaptively adjusted according to the signal quality of the current sampling period.

[0064] like Figure 3 As shown, the image edge features extracted from the image data, the contour depth features extracted from the contour data, and the robot pose data are weighted according to preset weights or weights adaptively adjusted according to the signal quality of the current sampling period to obtain a fusion score F. The current recognition result is then obtained through confidence updates. This method ensures stable welding features are output even when complex components have partial occlusion, surface contamination, or assembly deviations.

[0065] S6. The confidence level of the welding feature candidates is updated based on the fusion score within the continuous sampling period. The confidence update formula is as follows:

[0066] C k =α·C k-1 +(1-α)·F,

[0067] In the formula, C k C represents the confidence level for the current sampling period. k-1α is the confidence level of the previous sampling period; α is a smoothing coefficient between 0 and 1; when the updated confidence level is higher than the preset threshold, the corresponding welding feature candidate is determined as a valid welding feature. The threshold value range is [0, 1]. In this embodiment, the preset threshold is set to 0.7.

[0068] S7. Based on the effective welding features, output the weld start point, weld end point, weld centerline, local attitude reference, and recognition confidence level.

[0069] The local attitude reference includes at least one of the weld seam normal, weld seam tangent, and section normal, and is used for welding robot torch attitude adjustment or welding trajectory correction.

[0070] When local features are missing, the missing segments are compensated according to the continuity constraints of the prior model, and a missing segment compensation label is output.

[0071] The local feature loss includes:

[0072] When the image features extracted from the image data are missing, but the fusion score corresponding to the contour features extracted from the contour data meets the preset judgment rule, the corresponding welding features are retained; the preset judgment rule is: when the updated confidence level is ≥ the preset threshold, it is determined to be valid, otherwise it is not determined; the threshold value range is [0,1], and the preset threshold in this embodiment is set to 0.7.

[0073] When both the image features extracted from the image data and the contour features extracted from the contour data are missing, and there are valid welding features before and after the missing segment, the missing segment is interpolated based on the continuity constraints of the prior model and the valid welding features before and after.

[0074] For example, if a segment of an image is obscured by strong reflections or splashes, causing the loss of image edge features, but the fusion score corresponding to the contour features still meets the preset judgment rules, then the recognition result of that segment is retained; if both image edge features and contour features are missing, but there are high-confidence valid features in both the preceding and following segments, then interpolation is performed according to the continuity constraints of the prior model.

[0075] like Figure 4 As shown, the final output recognition results include the weld centerline coordinate sequence, weld start point, weld end point, local attitude reference, recognition confidence level, and missing segment compensation markers. The welding robot can automatically generate a welding trajectory based on this, or fine-tune the original trajectory during the welding process.

[0076] To further verify the effectiveness of the above method, an application test was conducted using a type of bent box-shaped component as an example.

[0077] The area to be welded in this bent box-shaped component exhibits both a reflective surface and variations in assembly gaps.

[0078] First, an industrial camera acquires an image of the area to be welded, a contour sensor outputs a height curve corresponding to the weld cross-section, and a robot controller provides the real-time pose of the sensor in the workpiece coordinate system.

[0079] Secondly, multi-source data is mapped to a unified coordinate system through time synchronization and calibration transformation.

[0080] Finally, candidate weld regions are extracted based on image edges and contour depth valleys, and matched with the nominal weld model of the box-type component. For local image edge loss areas caused by reflection, compensation is performed using contour features and prior model continuity to output complete weld centerlines and endpoints.

[0081] Experimental results show that, compared with a single image recognition method, the above method can significantly reduce centerline interruptions and false endpoint detections.

[0082] like Figure 2 As shown, a complex component intelligent welding feature recognition system that implements the above method is provided, including a multi-source sensing unit, a preprocessing and synchronization unit, a priori model unit, a fusion recognition unit, a recognition output unit, and a welding execution interface;

[0083] The multi-source sensing unit is used to acquire multi-source sensing data of the complex component to be welded. The multi-source sensing data includes at least image data, contour data, and robot pose data.

[0084] Specifically, the image data is acquired by an industrial camera to obtain the texture, edge, and brightness distribution of the weld area; the contour data is acquired by a contour sensor to obtain the height changes, gap changes, and bevel morphology near the weld; and the robot pose data is the pose information of the robot end effector, used to establish the correspondence between the sensor coordinate system and the workpiece coordinate system.

[0085] The preprocessing and synchronization unit is used to perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system.

[0086] Specifically, time synchronization adopts a unified trigger pulse or timestamp alignment method; coordinate unification is used to transform the coordinates of the industrial camera, contour sensor, and robot end effector to the same reference coordinate system; the preprocessing includes preprocessing the image data, contour data, and robot pose data respectively, wherein the preprocessing of the image data includes at least one of median filtering, grayscale equalization, and region of interest cropping; the preprocessing of the contour data includes at least one of outlier removal, smoothing filtering, and contour resampling; and the preprocessing of the pose data includes at least one of outlier removal and smoothing.

[0087] A priori model unit is used to store a priori models of complex components to be welded. The priori models include at least one or more of the following: nominal weld centerline, weld start and end positions, joint boundary information, and assembly tolerance range.

[0088] The fusion recognition unit is used to extract welding feature candidates from the preprocessed data, register and fuse the welding feature candidates with the prior model to obtain the fusion score of each welding feature candidate, update the confidence of the welding feature candidates according to the fusion score in a continuous sampling period, and determine the effective welding features according to a preset judgment rule.

[0089] Specifically, the extraction of welding feature candidates includes: extracting image edge features from the image data based on image gradient, brightness continuity or texture difference; extracting contour features from the contour data based on contour depth abrupt change, curvature change or gap change; and then establishing a spatial correspondence between candidate features and prior model by combining robot pose data.

[0090] In one embodiment, the fusion recognition unit employs the following fusion scoring model:

[0091] F = w1·G + w2·P + w3·M,

[0092] In the formula, G represents the image feature continuity score, P represents the contour feature saliency score, M represents the prior model matching score, and w1, w2, and w3 are preset weights, with w1+w2+w3=1. This fusion scoring model can comprehensively measure whether candidate features belong to real welding features.

[0093] To suppress misjudgments caused by transient disturbances, confidence levels can be updated for the fusion results across consecutive sampling periods. In one embodiment, the current recognition confidence level is calculated as follows:

[0094] C k =α·C k-1 +(1-α)·F,

[0095] In the formula, C k C represents the confidence level for the current sampling period. k-1 α represents the confidence level of the previous sampling period; α is a smoothing coefficient between 0 and 1.

[0096] When the current confidence level is higher than the preset threshold, the corresponding candidate feature is determined as a valid welding feature; when the current confidence level is lower than the preset threshold but the prior model has continuous constraints, the missing segment can be interpolated based on the valid feature points before and after to output a continuous weld centerline.

[0097] The identification output unit is used to output the weld start point, weld end point, weld centerline, local attitude reference, and identification confidence level based on the effective welding features. The local attitude reference can be determined by the weld normal, tangential direction, or section normal, and is used for subsequent welding torch attitude control. When a local feature is missing, the missing segment is compensated according to the continuity constraints of the prior model, and a missing segment compensation mark is output.

[0098] The welding execution interface is used to send the recognition results to the welding robot controller to generate an initial welding trajectory or to correct an existing welding trajectory.

[0099] To verify the effectiveness of the above methods and systems, an aluminum alloy bent box with strong reflectivity and assembly gaps was used as the implementation object.

[0100] Image data of the area to be welded is acquired using an industrial camera. A 3D contour sensor outputs contour curve data corresponding to the weld cross-section. The robot controller provides the sensor's pose data in the workpiece coordinate system, which is then mapped to a unified workpiece reference coordinate system via a hand-eye calibration matrix. The acquired image, contour, and pose data are synchronously mapped to the unified reference coordinate system before fusion calculation. For high-reflectivity conditions, the fusion model F = w1·G + w2·P + w3·M is set with weights of w1 = 0.25, w2 = 0.55, and w3 = 0.20; α = 0.75. Preset judgment rules and thresholds are set: In this embodiment, as a preferred solution, when the confidence level is updated ≥ 0.7 for three consecutive cycles, it is determined to be a valid feature; if local features are missing, interpolation is performed based on the CAD prior model. As shown in Table 1:

[0101] Table 1:

[0102]

[0103] Experiments show that, compared with single vision, single contour and single pose data schemes, the feature recognition accuracy of this embodiment increases to 94.8%, the center line error decreases to 0.18mm, and the maximum break length is 0.01mm, which significantly reduces the occurrence of center line interruption and false detection of endpoints, and improves the stability, continuity and accuracy of welding feature recognition.

[0104] The above are merely preferred embodiments of the present invention. Those skilled in the art can make various modifications or substitutions without departing from the concept of the present invention, and all such modifications or substitutions should fall within the protection scope of the present invention.

Claims

1. A method for intelligent welding feature recognition of complex components, characterized in that, Includes the following steps: S1. Acquire multi-source sensing data of the complex component to be welded, wherein the multi-source sensing data includes at least image data, contour data and robot pose data; S2. Perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system; S3. Establish or call the prior model of the complex component to be welded. The prior model includes at least one or more of the following: nominal weld centerline, weld start and end positions, joint boundary information, and assembly tolerance range. The prior model is used to provide nominal position constraints and continuity constraints of welding features. The prior model consists of at least two of the following: CAD model, offline teaching trajectory, nominal weld layout, joint type parameters, and assembly tolerance parameters. S4. Extract welding feature candidates from the preprocessed data and limit the candidate search range by combining the robot pose data; S5. Register and fuse the welding feature candidates with the prior model to obtain the fusion score of each welding feature candidate; The registration and fusion steps are as follows: first, coarse registration is performed on the welding feature candidates and the prior model based on the robot pose data and the preset tooling reference; then, local matching is performed based on the local morphological differences to obtain the deviation of the welding feature candidates relative to the prior model. The formula for calculating the fusion score is as follows: F = w1·G + w2·P + w3·M, In the formula, G represents the image feature continuity score extracted from the image data; P represents the contour feature saliency score extracted from the contour data; M represents the weighted average of the matching scores between the welding feature candidates and the prior model; w1, w2, and w3 represent preset weights or weights that are adaptively adjusted according to the signal quality of the current sampling period. S6. Update the confidence of welding feature candidates based on the fusion score within the continuous sampling period, and determine the valid welding features according to the preset judgment rules. The confidence update formula is as follows: C k =α·C k-1 +(1-α)·F, In the formula, C k is the confidence of the current sampling period; C k-1 is the confidence of the previous sampling period; a is a smoothing coefficient between 0 and 1; when the updated confidence is higher than a preset threshold, the corresponding welding feature candidate is determined as an effective welding feature; S7. Based on the effective welding features, output the weld start point, weld end point, weld center line, local attitude reference, and identification confidence level. When local features are missing, compensate for the missing segment according to the continuity constraint of the prior model and output the missing segment compensation mark.

2. The method according to claim 1, characterized in that, The image data in step S1 is acquired by an industrial camera, which has an industrial camera coordinate system; the contour data is acquired by a contour sensor, which has a contour sensor coordinate system; the robot pose data is the robot end-effector pose data, which is output by a robot pose encoder, which has a robot end-effector coordinate system.

3. The method according to claim 1, characterized in that, Step S2, time synchronization, employs a unified trigger pulse synchronization method or a timestamp-based alignment method. Coordinate unification includes transforming the sensor coordinate system for acquiring image data, the sensor coordinate system for acquiring contour data, and the sensor coordinate system for acquiring robot pose data to the workpiece reference coordinate system or the robot base coordinate system. Preprocessing includes preprocessing the image data, contour data, and robot pose data respectively. The preprocessing of the image data includes at least one of median filtering, grayscale equalization, and region of interest cropping. The preprocessing of the contour data includes at least one of outlier removal, smoothing filtering, and contour resampling. The preprocessing of the pose data includes at least one of outlier removal and smoothing.

4. The method according to claim 1, characterized in that, The extraction of welding feature candidates in step S4 includes at least one of the following: extracting candidate boundaries based on straight line fitting or curve fitting of image edges, extracting weld center candidates based on contour depth valleys, extracting groove boundary candidates based on contour curvature abrupt change points, and defining the candidate search area based on pose sliding windows.

5. The method according to claim 1, characterized in that, The local attitude reference mentioned in step S7 includes at least one of the weld seam normal, weld seam tangent and section normal, and is used for welding robot torch attitude adjustment or welding trajectory correction.

6. A complex component intelligent welding feature recognition system for implementing the method of any one of claims 1 to 5, characterized in that, It includes a multi-source sensing unit, a preprocessing and synchronization unit, a priori model unit, a fusion recognition unit, a recognition output unit, and a welding execution interface; A multi-source sensing unit is used to acquire multi-source sensing data of a complex component to be welded, wherein the multi-source sensing data includes at least image data, contour data and robot pose data. The preprocessing and synchronization unit is used to perform time synchronization, coordinate unification and preprocessing on the multi-source sensing data to obtain preprocessed data under a unified reference coordinate system. A priori model unit is used to store a priori models of complex components to be welded. The priori models include at least one or more of the following: nominal weld centerline, weld start and end positions, joint boundary information, and assembly tolerance range. The fusion recognition unit is configured to extract welding feature candidates from the preprocessed data, perform registration fusion of the welding feature candidates and the prior model, calculate a fusion score of each welding feature candidate by using a fusion score model F = w1G + w2P + w3M, and perform confidence update by using a confidence update model C k = αC k-1 +(1-α)F, and determine effective welding features according to a preset determination rule. The identification output unit is used to output the weld start point, weld end point, weld center line, local attitude reference, and identification confidence level based on the effective welding features. When local features are missing, the missing segment is compensated according to the continuity constraint of the prior model, and a missing segment compensation mark is output. The welding execution interface is used to transmit the information output by the identification output unit to the welding robot.

Images