Intelligent collaborative control method and system for indexable welding and grinding robots

Abstract

This invention relates to the field of industrial control technology, specifically disclosing an intelligent collaborative control method and system for a welding and grinding indexable robot. The method includes: establishing a unified collaborative data body, which includes a workpiece coordinate chain, an indexable coordinate chain, a tool mounting chain, and a process state chain; during the welding stage, collecting weld point cloud data, robot joint data, and welding torch arc state data to generate a weld feature model and writing it into the collaborative data body; during the indexing stage, collecting a workpiece reference feature point set and a weld feature point set before indexing, and collecting the corresponding point sets again after indexing, calculating the reference transformation and weld transformation respectively, and synthesizing them sequentially to obtain an indexing superimposed compensation transformation to update the workpiece coordinate chain; during the grinding stage, generating a grinding trajectory based on the updated weld feature model and mapping it to the tool mounting chain. This invention achieves continuous operation of welding and grinding across processes through unified data collaboration, indexing superimposed compensation, and confidence-driven control.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and specifically discloses an intelligent collaborative control method and system for a welding and grinding indexable robot. Background Technology

[0002] Welding and post-weld grinding are critical steps in pressure vessel manufacturing to ensure load-bearing safety and surface quality. To reduce reliance on manual labor and improve consistency and efficiency, the industry is gradually adopting an integrated welding and grinding approach with a rotary robot. Through program control, multi-task collaboration, trajectory planning, process switching, and status monitoring are achieved, forming a closed-loop control system of "perception-decision-execution".

[0003] However, existing methods still tend to overlook two types of problems: Firstly, the micro-deformation under heat after welding and the micro-drift of repeated positioning of the indexing mechanism are amplified when switching processes, causing the grinding starting point and trajectory to deviate systematically under high precision requirements, resulting in the risk of over-grinding, under-grinding, or accidental grinding. Even if the deformation or positioning error is not significant when viewed individually, it will still fail.

[0004] Secondly, the vibrations from welding arc spatter, fumes, and grinding dust can cause periodic distortions in multi-source sensor data such as vision, laser, and force control. Existing collaborative control systems often assume the data is reliable, making it difficult to assess the reliability in a timely manner and dynamically adjust fusion, degradation, and re-registration strategies. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides an intelligent collaborative control method and system for a welding and grinding indexable robot to solve the problems in the prior art.

[0006] The technical solution adopted by this invention to solve its technical problem is: The intelligent collaborative control method for the welding and grinding indexable robot includes: Establish a unified collaborative data body, which includes a workpiece coordinate chain, a rotation coordinate chain, a tool installation chain, and a process status chain; During the welding stage, weld seam point cloud data, robot joint data, and welding torch arc status data are collected to generate a weld seam feature model and write it into the collaborative data body. Before the indexing stage, the workpiece reference feature point set and the weld feature point set are collected. After the indexing, the corresponding point set is collected again. The reference transformation and the weld transformation are calculated respectively, and the indexing superposition compensation transformation is synthesized in sequence to update the workpiece coordinate chain. During the grinding stage, a grinding trajectory is generated based on the updated weld feature model and mapped to the tool installation chain; During the welding and grinding stages, the process confidence level is calculated for the visual measurement channel, laser measurement channel, and torque measurement channel, respectively. The fusion model is selected according to the process confidence level, and the fusion pose and fusion contact baseline are output. The process state chain transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

[0007] Furthermore, the collaborative data body sets time index, process index and version number fields, and sets write locking rules for the workpiece coordinate chain, weld feature model, fused pose and fused contact baseline respectively; at the end of the welding stage, a new version is written to the weld feature model, at the end of the rotation stage, a new version is written to the workpiece coordinate chain, and during the grinding stage, the fused pose and fused contact baseline are continuously written to the same version sequence according to the time index.

[0008] Furthermore, the workpiece reference feature point set is composed of the fixture reference mark point set and the workpiece reference boundary point set, and the weld feature point set is composed of the weld centerline point set and the weld neighborhood surface point set; the reference transformation and weld transformation establish a correspondence through geometric constraint matching, and obtain the inner point set through random consistency sampling, and then obtain the rigid body transformation from the inner point set and write it into the collaborative data volume.

[0009] Furthermore, the indexing superposition compensation transformation is obtained by combining mechanical indexing transformation, reference drift correction transformation and thermal deformation transformation in the order of the operation. The mechanical indexing transformation is obtained by the encoder data of the indexing device and the kinematic model of the indexing device. The reference drift correction transformation is obtained by the reference feature point set of the workpiece before and after indexing. The thermal deformation transformation is obtained by the weld feature point set before and after indexing. The combined result of the three is used to update the workpiece coordinate chain and for grinding trajectory mapping.

[0010] Furthermore, the process confidence is composed of channel quality indicators and model consistency indicators. In the welding stage, the channel quality indicators are composed of arc saturation indicators, spatter outlier rate, and point cloud missing rate. In the grinding stage, the channel quality indicators are composed of dust obstruction rate, vibration disturbance index, and contact signal drift rate. The model consistency indicators are generated by the residual sequence of weld feature model predictions and channel observations, and together with the channel quality indicators, they generate the process confidence and are written into the collaborative data volume.

[0011] Furthermore, the fusion model consists of a multi-channel weighted filtering model and a contact constraint filtering model. The multi-channel weighted filtering model determines the observation weight matrix based on the process confidence and outputs the fused pose. The contact constraint filtering model constructs contact constraints based on the fused pose and torque measurement channel observations and outputs the fused contact baseline. When the process state chain enters the rotation switching window, the process confidence of the welding stage is mapped to the initial weight matrix, and the first segment of the grinding stage is updated to the weight matrix corresponding to the process confidence of the grinding stage.

[0012] Furthermore, the process state chain includes a perception reconstruction state, which is triggered by the process confidence meeting the mismatch criterion. When entering the perception reconstruction state, a perception verification sequence is executed. The perception verification sequence consists of a baseline feature rescan, a weld feature rescan, and a torque zero-point reset in sequence. After completing the perception verification sequence, the process confidence is recalculated and the version numbers of the fusion model and the fusion contact baseline are updated.

[0013] Furthermore, the transition trajectory is composed of a welding end trajectory segment, a rotation safety trajectory segment, and a grinding entry trajectory segment connected sequentially. All three trajectory segments are represented under the workpiece coordinate chain and the tool mounting chain. At the connection of adjacent trajectory segments, the trajectory endpoints are repositioned according to the rotation superposition compensation transformation, and the pose continuity constraint is generated by piecewise polynomial interpolation. The transition contact command is generated by the fusion contact baseline and the pose sequence of the connection segment and written into the collaborative data body.

[0014] Furthermore, the intelligent collaborative control method for the welding and grinding indexable robot also includes: calculating the confidence trend and marking the observability of the grinding trajectory after recording the process confidence, selecting the verification anchor point and writing it into the collaborative data volume; freezing the fusion contact baseline and performing a local rescan before reaching the anchor point, and updating the weight matrix and version number.

[0015] Furthermore, the aforementioned lightweight, age-appropriate exoskeleton control system based on gait analysis includes: A collaborative data module establishes a unified collaborative data body, which includes a workpiece coordinate chain, an indexing coordinate chain, a tool installation chain, and a process status chain. The weld seam modeling module collects weld seam point cloud data, robot joint data, and welding torch arc state data during the welding stage, generates a weld seam feature model, and writes it into the collaborative data body. The indexing compensation module collects the workpiece reference feature point set and the weld feature point set before indexing, and collects the corresponding point set again after indexing. It calculates the reference transformation and the weld transformation respectively, and synthesizes them in sequence to obtain the indexing superposition compensation transformation to update the workpiece coordinate chain. The trajectory mapping module generates a grinding trajectory based on the updated weld feature model during the grinding stage and maps it to the tool mounting chain. The confidence fusion module calculates the process confidence level for the visual measurement channel, laser measurement channel, and torque measurement channel during the welding and grinding stages, respectively, selects the fusion model according to the process confidence level, and outputs the fusion pose and fusion contact baseline. The switching continuous module transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence through the process state chain, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

[0016] The beneficial effects of this invention are: This invention achieves continuous operation of welding and grinding across processes through unified data collaboration, rotational superposition compensation, and confidence-driven fusion control, improving positioning consistency and perception reliability. By managing the time index, process index, and version number of the collaborative data body, along with write locking and release mechanisms, this invention ensures that the workpiece coordinate chain, weld feature model, fused pose, and fused contact baseline maintain consistent semantics during cross-process transmission, reducing the risk of data misuse and state breakage. By collecting workpiece reference feature point sets and weld feature point sets before and after rotation, geometric constraint matching and random consistency sampling are used to obtain reference transformation and weld transformation, and a rotational superposition compensation transformation is synthesized to update the workpiece coordinate chain, reducing the grinding start point offset caused by the superposition of thermal micro-deformation and repetitive positioning micro-drift. The mechanical rotational transformation, reference drift correction transformation, and thermal deformation transformation are combined in process sequence into a composite compensation, and this composite result is used simultaneously for coordinate chain updates and grinding trajectory mapping, extending compensation from single-point correction to a traceable chain update, enhancing trajectory mapping stability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the intelligent collaborative control method for the welding and grinding indexable robot in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the preventive collaborative control driven by the verification anchor point in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the intelligent collaborative control system of the welding and grinding indexable robot in Embodiment 3 of the present invention. Detailed Implementation

[0018] The present invention will now be described and illustrated in detail with reference to the accompanying drawings.

[0019] Example 1 like Figure 1As shown, an embodiment of the present invention provides an intelligent collaborative control method for a welder and grinder indexable robot, the method comprising: S1: Establish a unified collaborative data body, which includes a workpiece coordinate chain, a rotation coordinate chain, a tool installation chain, and a process status chain; S2: During the welding stage, collect weld point cloud data, robot joint data and welding torch arc status data, generate weld feature model and write it into the collaborative data body; S3: During the indexing stage, collect the workpiece reference feature point set and weld feature point set before indexing, and collect the corresponding point set again after indexing. Calculate the reference transformation and weld transformation respectively, and synthesize them in sequence to obtain the indexing superposition compensation transformation to update the workpiece coordinate chain. S4: During the grinding stage, a grinding trajectory is generated based on the updated weld feature model and mapped to the tool installation chain; S5: Calculate the process confidence level for the visual measurement channel, laser measurement channel, and torque measurement channel respectively during the welding and grinding stages. Select the fusion model according to the process confidence level and output the fusion pose and fusion contact baseline. S6: The process state chain transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

[0020] In this embodiment, when the controller is executed, it manages the collaborative data body as a structured data area with a unified header field. The collaborative data body consists of a meta-information area and a data area: the meta-information area has a fixed time index, process index and version number; the data area stores the workpiece coordinate chain, weld feature model, fused pose and fused contact baseline by object, and maintains a writer identifier, locking status and release mark for each object to implement the write locking rules.

[0021] The time index is generated by the controller's time base module and is represented by a monotonically increasing time value. When a write operation occurs, the write task reads the time base value once at the entry point into the critical section and writes the time index of the record. If the fused pose and the fused contact baseline are written simultaneously within the same control cycle, the two records share the same time index, forming a one-to-one pair of records. The process index is given by the process state chain. The process state chain maintains the current state of the welding stage, the rotation stage, and the grinding stage using a state machine, and updates the process index during state transitions. The write task must not rewrite the process index itself, but rather reads it and writes it along with the record to ensure that the process attribution of the data is consistent.

[0022] Version numbers are used to delineate the write boundaries between snapshot-type objects and sequence-type objects. The workpiece coordinate chain and weld feature model are snapshot-type objects, and their version numbers correspond to stable snapshots after a single write operation. The fused pose and fused contact baseline are sequence-type objects, and their version numbers correspond to a version sequence formed during the grinding phase. The controller sets a version manager to save the current version number and the version number to be submitted for each object. After the write task is successfully locked, new data is written according to the version number to be submitted. When submitting, the version number to be submitted is promoted to the current version number and a release flag is set for subsequent processes to read.

[0023] Write locking rules are defined separately for each object and implemented by a lock table. The lock table records the object identifier and the lock holder identifier, which consists of the process index and the task number. The write lock for the weld feature model is held by the welding stage task, while the indexing and grinding stages only read the published version. The write lock for the workpiece coordinate chain is held by the indexing stage task, while the welding stage only reads the initial version, and the grinding stage only reads the published version submitted after the indexing stage. The write lock for the fused pose and fused contact baseline is held by the grinding stage task, and these two types of sequences are not written during the welding and indexing stages. Lock acquisition and release are driven by the process state chain: when the process state chain enters a certain stage, it issues a write authorization token to the corresponding task. The token contains the current process index and a list of objects that can be written. A task can only apply for a write lock when it holds a token that matches the current process index, avoiding cross-stage writes.

[0024] At the end of the welding phase, the welding task performs the weld feature model version writing: after completing the summary of weld point cloud data and arc state data, a write lock is requested; after obtaining the lock, the version number to be submitted is incremented and written to the meta information area, and then the main data of the weld feature model is written; after the writing is completed, a consistency check is performed, which includes the integrity of the model index and the consistency of the time index landing point. After passing the check, a release flag is set and the write lock is released, so that the weld feature model is solidified with the new version snapshot.

[0025] At the end of the transposition phase, the transposition task executes the workpiece coordinate chain version writing: after the transposition action is completed and the corresponding calculation of the reference feature point set and weld feature point set before and after transposition is completed, a write lock is requested; after the lock is obtained, the link node relationship is updated based on the current version and the transposition superposition compensation transformation obtained in this phase, and the new version number is written; the writing adopts the new version area writing plus pointer switching method, first write to the pending area, then the version manager switches the current version pointer and sets the release flag, and finally releases the write lock to ensure that the reader does not read the semi-finished product.

[0026] During the polishing phase, the polishing task continuously writes the fused pose and fused contact baseline to the same version sequence according to the time index: upon entering the polishing phase, it requests write locks for both at once and obtains the current sequence version number from the version manager; within each control cycle, it first generates a fused pose record, then generates a fused contact baseline record. The two records share the same time index and the same sequence version number, and are appended to the sequence storage area in time index order; the sequence storage area adopts a log-style append layout and sets a segment header. The segment header records the sequence version number and the start time index, and the records within the segment are arranged in ascending order of time index; when the segment switching condition is met, only the segment number is switched without changing the sequence version number, until the process status chain issues a polishing phase end event, after which the write lock is released and the version sequence ends.

[0027] For example: In a certain processing, the welding stage ends at 10:15:30 on February 14, 2026, and the weld feature model is updated from version 2 to version 3 and released; the indexing stage ends at 10:18:05, and the workpiece coordinate chain is updated from version 5 to version 6 and released; the grinding stage starts at 10:18:06, the grinding task obtains sequence version 6, and writes three pairs of records consecutively at 10:18:06, 10:18:06 plus one control cycle, and 10:18:06 plus two control cycles. Each pair of records has the same time index and the same version 6. The fused pose of one pair of records can be recorded as "position is several coordinate values, attitude is several angle values", and the corresponding fused contact baseline can be recorded as "contact baseline is several torque reference values".

[0028] In the transposition stage of this embodiment, the controller constructs a workpiece reference feature point set and a weld feature point set respectively to obtain the reference transformation and weld transformation, and writes the two types of transformations into a collaborative data body for subsequent coordinate chain updates.

[0029] The workpiece reference feature point set consists of the fixture reference mark point set and the workpiece reference boundary point set. The fixture reference mark point set refers to the set of coordinates of mark points arranged on the fixture, which are geometrically stable and can be reliably identified by sensors. Markers can be coded corner points of nameplates, center points of pin holes, and reference points of ball heads, etc. The controller first completes the mark area segmentation in the acquired point cloud, and then performs geometric fitting or template matching on each mark to obtain the three-dimensional coordinates of the mark points and form a point set. The workpiece reference boundary point set refers to the set of boundary geometric points on the workpiece used to limit the posture. It is usually taken from the outer contour edge, edge of the reinforcing ring, or outer boundary of the bevel that does not participate in the weld formation after clamping. The controller performs edge extraction on the boundary in the point cloud, sorts the points according to the connectivity of the boundary, and removes burr points and isolated points when necessary to obtain the boundary point set. The two types of point sets are acquired once before the indexing and once again after the indexing according to the same recognition rule to form paired data of "point set before indexing" and "point set after indexing".

[0030] The weld feature point set consists of the weld centerline point set and the weld neighborhood surface point set. The weld centerline point set refers to the set of point coordinates distributed along the weld direction that can characterize the geometric principal axis of the weld. The controller takes the weld point cloud as input, first identifies the continuous region of the weld, and then obtains the position corresponding to the axis of symmetry or the maximum height direction according to the weld cross-section direction. The centerline point set is extracted and smoothed along the weld length direction. The weld neighborhood surface point set refers to the set of surface points located in the defined neighborhood on both sides of the centerline. It is used to supplement the local geometric constraints outside the centerline. The controller takes the centerline as a reference, extracts the surface points on both sides from the point cloud according to a fixed neighborhood selection rule, calculates the local normal consistency of the surface points, and removes outliers that are inconsistent with the weld neighborhood surface to form the neighborhood surface point set. This weld feature point set is also formed before and after the rotation, and serves as the input for the "weld transformation".

[0031] The datum transformation and weld transformation establish correspondences through geometric constraint matching. Geometric constraint matching does not directly pair all points one by one, but first uses geometric invariants and structural relationships to screen out candidate correspondences, and then enters consistency sampling. For the fixture datum mark point set, if the mark has its own number, the correspondence is determined by the number; if the mark does not have a number, candidate correspondences are established by the distance set between marks, the angle relationship, and the coplanar constraint, and the candidate correspondences are required to maintain the same adjacency relationship before and after the rotation. For the workpiece reference boundary point set, the controller uses the sequence parameters of the boundary points as an index, combined with the curvature change points, corner points, and the boundary length ratio to determine the anchor points, and then establishes correspondences in the neighborhood of the anchor points. For the weld centerline point set, the controller uses the arc length parameter or key inflection points as constraints to establish correspondences, and requires the corresponding points to maintain a consistent front-to-back order in the tangential direction of the centerline. For the weld neighborhood surface point set, in addition to the position residual, a local normal angle constraint is also introduced to avoid incorrectly pairing reverse surface points to the same side neighborhood.

[0032] After obtaining candidate correspondences, the controller performs random consistency sampling on the datum transformation and weld transformation respectively to obtain the interior point set. The interior point set refers to the set of corresponding point pairs whose residuals satisfy the consistency criterion under the same rigid body transformation assumption. Specifically, the controller extracts a set of corresponding point pairs that satisfy the minimum solution condition from the candidate correspondences and calculates a candidate rigid body transformation; applies the candidate transformation to the point set before the rotation to obtain the predicted point position after the rotation; compares it with the corresponding points of the actual point set after the rotation, calculates the spatial residual of each pair, and determines whether to include it in the interior point set; repeats this sampling and verification process to select a set of interior points whose number of interior points and residual statistics simultaneously satisfy the criterion; to avoid the boundary points being sensitive to the local deformation of the weld, the interior point evaluation of the datum transformation is mainly based on the clamp datum mark point pairs and supplemented by the boundary point pairs; the interior point evaluation of the weld transformation is constrained by the centerline point pairs and the neighboring surface point pairs to ensure that the transformation has local geometric consistency with the weld.

[0033] After obtaining the set of interior points, the controller calculates the rigid body transformation from the interior point set and writes it into the collaborative data volume. The rigid body transformation here refers to the spatial pose transformation of mapping the pre-rotation point set to the post-rotation point set, which consists of two parts: rotation and translation. The controller performs closed-form least squares solution on the interior point set: first, it calculates the centroids of the pre-rotation and post-rotation interior points, then constructs the correlation matrix for the decentered point pairs, obtains the rotation matrix through singular value decomposition, and obtains the translation vector from the centroid relationship. The obtained rigid body transformations are denoted as datum transformation and weld transformation, respectively. Subsequently, the controller writes the two types of transformations as structured records into the data area of ​​the collaborative data volume, and writes the corresponding time index and process index into the meta-information area. At the same time, it updates the version number and release tag of the relevant objects according to the version management rules, so that subsequent steps can distinguish the transformation records from datum features and weld features when reading the workpiece coordinate chain update basis.

[0034] Closed-form least squares solutions include:

[0035] In the formula, For optimal rotation, For optimal translation, the rigid body transformation is obtained from the set of interior points and written into the collaborative data volume; R is the rotation matrix, corresponding to the "rotation part of the rigid body transformation", and is subject to constraints. (Represents an orthogonal rotation with a positive determinant); t is the translation vector, corresponding to the "translation part of a rigid body transformation"; The index number in the set of interior points. Let it be the set of interior points; For the first The weights of corresponding point pairs are used to incorporate "primary constraints" and "secondary constraints" into the solution process: When solving for datum transformation, the weights of fixture datum mark point pairs can be set higher than those of workpiece reference boundary point pairs to implement "mark point pairs as primary and boundary point pairs as secondary". When solving for weld transformation, weld centerline point pairs and weld neighborhood surface point pairs can participate simultaneously, and the contributions of the two to rotation and translation can be balanced by weights to implement "common constraints". For the first point in the pre-inversion set The three-dimensional coordinates of an interior point can be derived from one of the following: the fixture reference mark point, the workpiece reference boundary point, the weld centerline point, or the weld neighboring surface point. For the AND of the point set after transposition The corresponding number The three-dimensional coordinates of an interior point originate from... similar; is the Euclidean norm, used to measure the spatial residual of corresponding point pairs; and To calculate the weighted centroid, we first calculate the centroids of the interior points before and after the transposition. That is, the rotation matrix is ​​constructed by piecing together the results of singular value decomposition.

[0036] For example: Four circular holes are arranged on the fixture as reference marks. The controller performs circle fitting on the circular holes before and after the rotation to obtain the coordinates of the four center points, forming a set of reference mark points for the fixture; a set of reference boundary points is obtained by boundary extraction of the outer edge of the workpiece; the center line point set is extracted from the weld point cloud and a neighborhood surface point band is taken on each side of the center line; during the random consistency sampling process, the candidate solution of the reference transformation is extracted from the center point pairs, and the candidate solution of the weld transformation is extracted from the center line point pairs and the neighborhood surface point pairs; the record finally written into the collaborative data body contains two transformations and their time index, process index and version number, which are called by the synthesis step of the rotation superposition compensation transformation.

[0037] Before the end of the indexing stage in this embodiment, the controller generates and solidifies the indexing superposition compensation transformation to reliably transition the geometric information formed in the welding stage to the grinding stage. The indexing superposition compensation transformation refers to unifying three types of pose changes with different sources and physical meanings into a composite transformation that can be directly used for coordinate chain updates according to the process sequence. Its components are mechanical indexing transformation, reference drift correction transformation and thermal deformation transformation in sequence. The combination according to the process sequence is not a simple multiplication, but clearly defines the reference coordinates and the object of action corresponding to each type of transformation, so that the result of the combination can be consumed by the workpiece coordinate chain under the same coordinate semantics, avoiding the hidden danger of "transformations from different coordinate systems being directly superimposed".

[0038] The mechanical indexing transformation is obtained from the encoder data and kinematic model of the indexing device. The encoder data refers to the angle or displacement feedback sequence output by the indexing device during the indexing action, reflecting the actual motion state of the indexing axis. The kinematic model of the indexing device is a pose mapping relationship established by the geometric parameters of the indexing mechanism and the joint connection relationship. It is used to convert the encoder reading into the pose change of the indexing end relative to the base. In implementation, the controller reads the initial encoder value once before the indexing starts and reads the final encoder value after the indexing ends and the mechanism is in place and stable. The difference between the two readings is taken as the indexing input. Then, this input is substituted into the kinematic model to calculate the pose change of the workpiece mounting interface relative to the equipment base before and after the indexing, forming the mechanical indexing transformation. In order to reduce the impact of dynamic jitter on the final value, the controller can perform a consistency check on the encoder reading within the positioning window. After confirming that the reading is stable, the final value is locked. Once the mechanical indexing transformation is generated, it will not be changed by subsequent algorithm updates and will be used as the basis for the two types of corrections.

[0039] The reference drift correction transformation is obtained from the workpiece reference feature point set before and after indexing. Reference drift here refers to the deviation of the workpiece relative to the actual reference of the fixture caused by factors such as the elastic rebound of the fixture, slight slippage of the positioning surface, and slight loosening of fasteners, even though the indexing device completes the action according to the command. This deviation is not necessarily reflected in the encoder reading, therefore it must be identified and corrected using the workpiece reference feature point set. The construction and acquisition method of the workpiece reference feature point set is consistent with the aforementioned embodiment, that is, the actual positional relationship of the workpiece on the fixture is jointly defined by the fixture reference mark point set and the workpiece reference boundary point set; the controller in the rotation... The point set is collected before and after the positioning, and the set of interior points is obtained through geometric constraint matching and random consistency sampling. The rigid body transformation is then obtained from the set of interior points. This rigid body transformation expresses "the remaining deviation of the workpiece datum relative to the fixture after deducting the influence of the mechanical rotation ideal". Therefore, it is defined as the datum drift correction transformation. In implementation, when the controller solves this transformation, it will take the fixture datum marker point as the priority constraint to stabilize the overall attitude. Then, the workpiece reference boundary point is used to supplement the constraint to eliminate small angular degrees of freedom around a specific axis, so that the correction transformation is repeatable and avoids the boundary local noise dominating the overall solution.

[0040] The hot deformation transformation is obtained from the weld feature point set before and after the rotation. Hot deformation is not equivalent to the overall rigid body pose change, but rather to the local geometric changes caused by welding heat input and the springback changes resulting from cooling over time. There is a time interval between welding and grinding, and the rotation action alters the stress and support conditions; therefore, the geometry near the weld may undergo considerable but uneven changes. To bring these changes within a controllable range, the controller uses the weld feature point set for identification. The weld feature point set consists of the weld centerline point set and the weld neighborhood surface point set, including both the mainline constraints along the weld direction and the surface constraints of the lateral neighborhood. The device collects weld feature point sets before and after rotation, and establishes corresponding relationships through geometric constraint matching. Then, it obtains an interior point set through random consistency sampling, and calculates the rigid body transformation from the interior point set. This rigid body transformation extracts the equivalent pose offset of the part that is most sensitive to the grinding start section and most prone to amplifying errors, which is used to realign the weld features at the coordinate chain level. To avoid treating local plastic protrusions or spatter residues as deformation features, the controller introduces normal consistency screening when constructing the weld neighborhood surface point set, and introduces centerline sequence constraints during consistency sampling, so that the obtained thermal deformation transformation is closer to the overall drift trend of the weld body geometry.

[0041] The three types of transformations are combined sequentially to obtain a superimposed compensation transformation for the indexing process. The combination relationship is as follows: first, a mechanical indexing transformation is used to transfer the coordinate relationships of the welding stage to the theoretical position after indexing; then, a reference drift correction transformation is used to eliminate the remaining deviation of the "workpiece relative to the fixture reference"; finally, a thermal deformation transformation is used to realign the weld features under the updated workpiece coordinate semantics. The controller stores the three types of transformations and their generation time indices and process indices in the collaborative data body, and displays the version numbers of the three referenced records when generating the composite result, ensuring that the composite result corresponds to the same indexing process and avoiding cross-process mixing. After the composite overlay compensation transformation is written into the collaborative data body, it is used to update the workpiece coordinate chain: the controller reads the relevant node relationships of the current workpiece coordinate chain and applies the composite transformation to the segment of the workpiece coordinate chain from the fixture side node to the workpiece side node, so that the workpiece coordinate chain expresses the consistent relationship between "the actual position of the workpiece and the actual geometry of the weld" after the overlay. After the update is completed and published, the trajectory mapping in the grinding stage is based on the workpiece coordinate chain, and the grinding trajectory generated by the weld feature model is semantically stably mapped from the welding stage to the tool mounting chain in the grinding stage, avoiding the starting point offset caused by the overlay error.

[0042] For example: The indexing device is a single-axis rotary table. The encoder records the angles before and after indexing, and the controller calculates the mechanical indexing transformation based on the kinematic model. After indexing, it is found that the center position obtained by fitting the reference mark point of the fixture deviates from the position predicted by the mechanical indexing before indexing. The controller obtains the reference drift correction transformation accordingly. At the same time, the key inflection point of the weld centerline obtained by rescanning after indexing has an overall translational trend relative to before indexing. Combined with the normal consistency constraint of the neighboring surface points, the thermal deformation transformation is obtained. The three are combined in the above order and written into the collaborative data body, and the workpiece coordinate chain is updated accordingly. Subsequently, the grinding trajectory mapping directly references the updated workpiece coordinate chain to complete the coordinate transformation.

[0043] In this embodiment, the controller generates process confidence scores during both the welding and grinding stages and records them continuously as process data in the collaborative data body according to the time index. This data is used to constrain the subsequent selection of the fusion model and the update of the fusion weights. The process confidence score does not evaluate whether a sensor is "good" or "bad". Instead, it unifies the availability, stability and model conformity of each measurement channel under the same process into a calculable and verifiable quantity. It is composed of channel quality indicators and model consistency indicators and is written into the collaborative data body with the same time index and process index, so that it corresponds to the fusion pose and fusion contact baseline one by one.

[0044] Channel quality metrics are defined separately for the welding and grinding stages because the interference mechanisms in the two stages are different. Directly using the same set of metrics could easily misjudge environmental disturbances as algorithm errors. The channel quality metrics for the welding stage consist of arc saturation flag, spatter outlier rate, and point cloud missing rate. The arc saturation flag is used to characterize the occupancy of the dynamic range of the visual measurement channel and the laser measurement channel by the arc light: the controller performs saturation discrimination for each frame observation, and the discrimination criteria adopt image grayscale overflow statistics, laser echo intensity overflow statistics, or similar sampling upper limit trigger conditions; frames that meet the criteria are marked as valid, and a saturation percentage is formed within the time window to reflect the saturation rate. The window observes whether the arc light is suppressed; the spatter outlier rate is used to reflect the degree of damage to the geometric observation caused by spatter: the controller uses the neighborhood defined by the weld feature model as a reference to jointly screen the distance distribution, normal consistency and connectivity of the point cloud, and marks the points that cannot be consistent with the main point cluster as outliers, and forms the outlier rate by the ratio of outliers to valid points; the point cloud missing rate is used to reflect the return gap: the controller determines the set of sampling indexes that should be generated in the window according to the scanning trajectory and sampling cycle, and then counts the number of valid echo indexes. The proportion of missing parts is the point cloud missing rate, thus unifying non-return and invalid return to the same standard.

[0045] The channel quality indicators during the polishing stage consist of dust occlusion rate, vibration disturbance index, and contact signal drift rate. The dust occlusion rate characterizes the degree of dust occlusion on the visual measurement channel and the laser measurement channel: the controller calculates occlusion criteria for each frame, which are composed of observable measurements such as the degree of fragmentation of effective pixel connected regions, image contrast attenuation, and laser echo loss bandwidth; frames that meet the criteria are recorded as occluded frames, and the proportion of occluded frames within the window is the dust occlusion rate. The vibration disturbance index reflects the interference intensity of polishing vibration on pose observation and trajectory tracking: the controller... The encoder feedback, end-effector pose feedback, or acceleration measurement data are projected onto the coordinate semantics corresponding to the tool mounting chain. High-frequency disturbance components related to the contact process are extracted, and a vibration disturbance index is formed by the statistical value of the disturbance amplitude within the window. The contact signal drift rate is used to reflect the baseline stability of the torque measurement channel: the controller uses the fused contact baseline as a reference to form a sequence of torque observations, extracts low-frequency trend terms, and uses the rate of change of the trend terms to characterize the drift rate, which is used to distinguish between the actual contact changes and the slow baseline drift, and to prevent the drift from being written into the fused contact baseline.

[0046] The model consistency index is generated from the residual sequence of the weld feature model predictions and channel observations. The weld feature model is a geometric representation of the weld that is written into the collaborative data volume and can be read. The predictions refer to the expected observations calculated by the controller using the weld feature model under given robot joint data and tool mounting chain conditions, such as the relative position of the weld centerline, the normal direction of the neighboring surface, or the geometric relationship between the tool and the weld surface. The controller establishes a prediction mapping of the same dimension as its observation for each measurement channel, transforms the predictions into the semantics of the channel observation, and subtracts the actual observations to obtain the residuals. The residuals are arranged by time index to form a residual sequence. The controller calculates consistency statistics within the time window, including steady-state bias, fluctuation range, and number of continuous out-of-bounds errors, and adopts corresponding discrimination logic in combination with the process index: the welding stage focuses on identifying short-term distortion caused by arc light and peak residuals caused by spatter, and the grinding stage focuses on identifying continuous distortion caused by dust obstruction and periodic residuals caused by vibration. Based on this, the model consistency index can screen out systematic biases that are difficult to explain by the channel quality index, and prevent erroneous observations from continuing to enter the fusion model.

[0047] The generation of process confidence is accomplished in two steps: The first step is for the controller to generate a channel quality score and a model consistency score for each measurement channel: the channel quality score is obtained by normalizing and mapping the three channel quality indicators defined for the process, and the mapping rules are configured by the controller and fixed into the process parameter set corresponding to the process state chain; the model consistency score is obtained by consistency mapping the residual sequence statistics, and a higher penalty is given for persistent inconsistencies during the mapping.

[0048] The second step involves the controller combining the two types of scores into the process confidence component for that channel, and writing each channel component and the summary component into the data area of ​​the collaborative data body. When writing, the time index and process index rules of the collaborative data body are followed to ensure that the process confidence record, the fused pose record, and the fused contact baseline record within the same control cycle form a one-to-one correspondence. The process confidence record in the collaborative data body also contains the weld feature model version number and the workpiece coordinate chain version number referenced in generating the record, which is used to trace the model and coordinate semantics on which the residual sequence is based, and avoids inconsistent misjudgments caused by cross-version mixing.

[0049] For example: During the welding stage, the proportion of arc saturation frames in the visual measurement channel is 12%, the point cloud outlier rate is 8%, and the point cloud missing rate is 5%. Based on this, the controller gives the channel quality score for that channel and generates a model consistency score by combining the number of times the peak residuals in the residual sequence persist. After synthesis, the process confidence component of that channel is obtained. After entering the grinding stage, the proportion of dust obstruction frames in the same channel increases to 20%, the vibration disturbance index increases, and the residual sequence shows continuous bias. Based on this, the controller lowers the process confidence component of that channel and writes the updated record into the collaborative data volume, so that the subsequent fusion model selection and weight update can be based on data within the process switching window.

[0050] In this embodiment, under the constraints of the collaborative data body and the process state chain, the controller adopts a fusion model composed of a multi-channel weighted filtering model and a contact constraint filtering model to uniformly process the multi-source observations of the welding and grinding stages, forming a fused pose and a fused contact baseline that can be directly used by the control sequence. The fusion model refers to the output of state quantities that meet the requirements of process switching continuity after aligning the multi-source observations such as the visual measurement channel, laser measurement channel, and torque measurement channel with the weld feature model and the workpiece coordinate chain under the same time index. The fused pose refers to the robot end pose estimate determined under the common semantics of the workpiece coordinate chain and the tool mounting chain. The fused contact baseline refers to the contact quantity benchmark used as a zero-point reference and slow variable tracking in contact control, which is used to offset the influence of sensor zero-point drift and process gradual change factors on contact judgment.

[0051] The multi-channel weighted filtering model determines the observation weight matrix based on process confidence and outputs the fused pose. In implementation, the controller first performs observation assimilation: transforming the poses or geometric quantities obtained from the visual measurement channel and the laser measurement channel into a unified observation semantic based on the coordinate relationship between the workpiece coordinate chain and the tool mounting chain, and attaching the corresponding time index and process index to each channel observation; subsequently, it reads the process confidence records under the same time index from the collaborative data volume to obtain the process confidence components of each channel. The observation weight matrix refers to the matrix-based weights used to describe the relative confidence level of each channel observation in the filter update. Its structure is consistent with the dimension of the filter state quantity; the larger the weight, the stronger the correction effect of the channel observation on the corresponding state component. The controller then... When mapping confidence components to observation weight matrices, a single proportional coefficient is not used directly. Instead, the mapping rules corresponding to the components are adopted. For example, the weights related to pose translation are mainly affected by the point cloud missing rate, dust occlusion rate, and residual steady-state bias, while the weights related to attitude are mainly affected by normal consistency, vibration disturbance index, and residual fluctuation range. The mapping rules are switched with the process index, so that the same channel can generate different weight distributions in the welding stage and the grinding stage, avoiding the incorrect transmission of the arc saturation effect in the welding stage to the grinding stage. After the mapping is completed, the multi-channel weighted filtering model uses the observation weight matrix to perform prediction and update on the multi-channel observations, outputs the fused pose, and writes the fused pose into the collaborative data volume for use by the contact constraint filtering model and control sequence.

[0052] The contact constraint filtering model constructs contact constraints and outputs a fused contact baseline by fusing pose and torque measurement channel observations. The contact constraint is not a simple threshold judgment, but rather associates the geometric relationship during the contact process with the torque observation as a constraint term that can be absorbed by the filter. This allows the contact baseline to converge slowly without disrupting trajectory following. In implementation, the controller reads the fused pose at each time index and substitutes it into the weld feature model and the mapping relationship between the grinding trajectory to obtain the geometric proximity of the tool end relative to the weld surface or the expected contact direction. Simultaneously, it reads the torque measurement channel observations and defines them according to the direction of the tool mounting chain. The torque observation is projected onto the contact-related components to form contact observations. The controller constructs contact constraints accordingly: when the geometric proximity quantity and the contact observation show a consistent contact trend within the time window, the contact constraint filtering model is allowed to perform small-step updates on the fused contact baseline; when there is a significant mismatch between the two, the contact constraint filtering model keeps the fused contact baseline unchanged or enters the reconstruction process, which is triggered by the process state chain for perception reconstruction state processing; the update result of the fused contact baseline, along with the torque observation and fused pose used, is written into the collaborative data body with a time index to ensure that the subsequent control sequence has traceability in reading the contact quantity.

[0053] When the process state chain enters the shift switching window, the controller performs cross-process migration of the weight matrix to avoid "weight zeroing" or "weight mutation" at the moment of switching, which could cause a jump in the fused pose. The shift switching window is a transition interval defined during the process state chain's migration from the welding stage to the grinding stage, including several control cycles before the end of the welding end trajectory segment, the completion of the shifting safety action, and the start of the grinding entry trajectory segment. At the beginning of this window, the controller freezes the last process confidence record of the welding stage and maps it to an initial weight matrix, which serves as the weight basis for the multi-channel weighted filtering model within the switching window. This initial weight matrix and the weights already generated in the collaborative data volume... The weld feature model version number and workpiece coordinate chain version number are bound and written to prevent inconsistent model semantics from being read within the window. After entering the first segment of the grinding stage, the controller does not immediately replace the weight matrix with the process confidence of the grinding stage. Instead, it first generates a new process confidence record according to the channel quality index and model consistency index of the grinding stage within the first segment, and then maps the record to the weight matrix of the grinding stage and completes the weight update. A smooth switching strategy is adopted during the update so that the weight transitions from the initial weight matrix of the welding stage to the weight matrix of the grinding stage, ensuring that the fused pose is continuous near the switching point and that the fused contact baseline is not pulled off by transient noise in the early stage of contact establishment.

[0054] For example: At the end of the welding stage, the confidence level of a certain vision measurement channel decreases due to arc saturation. Accordingly, the controller reduces the weight of this channel on the attitude component in the initial weight matrix. After entering the first stage of the grinding stage, the confidence level of this channel is further reduced due to dust obstruction, while the laser measurement channel has stable echoes and high confidence. The controller increases the weight of the laser channel on the translation component in the weight matrix of the grinding stage and smoothly completes the weight migration within several control cycles. When the torque measurement channel establishes contact, the baseline slowly drifts. The contact constraint filtering model, combined with the geometric approach trend given by the fused pose, makes small-step updates to the fused contact baseline, so that the baseline read by the subsequent contact control remains stable.

[0055] In this embodiment, in addition to the welding stage, the indexing stage, and the grinding stage, the process state chain also includes a perception reconstruction state, which is used to handle the observation confidence gap that occurs during process switching and processing. The perception reconstruction state is not a shutdown branch after an abnormal alarm, but an executable state that is incorporated into the normal control closed loop. Its triggering, execution, and exit are all determined by the controller based on the process confidence recorded in the collaborative data body, and the data boundary before and after reconstruction is solidified by a version number mechanism to avoid the reconstruction action being limited to experience processing and difficult to verify.

[0056] The perception reconstruction state is triggered by the process confidence level meeting the mismatch criterion. The mismatch criterion refers to the judgment rule obtained by the controller after performing a consistency review on the process confidence level record. It is used to distinguish between short-term noise and staged failures. In implementation, the controller reads the latest process confidence level record from the collaborative data volume in each control cycle, and simultaneously reads the fused pose, fused contact baseline, and residual sequence statistics corresponding to the same time index of the record. The construction of the mismatch criterion follows two constraints: first, the decrease in process confidence level must be continuous to avoid single-frame occlusion or single splash triggering reconstruction; second, there must be an interrelationship between the decrease in confidence level and the deterioration of model consistency. Explainable correlations prevent false triggers due to channel quality fluctuations. To this end, the controller uses a sliding time window to statistically analyze the process confidence components. If a channel or aggregated component remains below the available threshold within the window, and the residual sequence shows continuous bias or continuous out-of-bounds behavior, and this phenomenon is consistent with the typical interference source corresponding to the current process index, such as an increase in the arc saturation ratio during welding, an increase in the dust obstruction rate during grinding, or a sudden increase in vibration disturbance indicators, then the mismatch criterion is met. The process state chain migrates from the current state to the perception reconstruction state. This migration event and the time index of the triggering time are written into the collaborative data body for subsequent traceability.

[0057] Upon entering the perception reconstruction state, the controller executes a perception verification sequence, which consists of a baseline feature rescan, a weld feature rescan, and a torque zero-point reset in sequence. The phrase "in sequence" emphasizes the execution order and their dependencies: the baseline feature rescan is used to first pull the coordinate semantics back to a stable baseline; the weld feature rescan updates the consistency of the local geometry of the weld on this basis; and the torque zero-point reset is executed last to avoid mistakenly writing transient contact as a zero point when the attitude is unstable. When the controller enters the perception reconstruction state, it first freezes the update permissions of the weight matrix of the current fusion model and the fusion contact baseline, and at the same time generates a safety motion trajectory to make the end tool withdraw from the contact area and enter a rescannable position. The start and end poses and execution time of this safety motion trajectory are also written into the collaborative data volume to ensure that the motion during reconstruction is verifiable.

[0058] The implementation of the reference feature rescan relies on the workpiece reference feature point set acquisition mechanism. The controller calls the same reference recognition process as the rotation stage, segments and geometrically fits the fixture reference mark point set, extracts and sorts the workpiece reference boundary point set to form a rescan point set, and performs geometric constraint matching and random consistency sampling with the previously published reference point set version to obtain a new set of interior points and corresponding rigid body transformation. This rigid body transformation is used to verify the workpiece coordinate chain: if the transformation shows that the workpiece has undergone a non-negligible drift relative to the fixture reference, the controller writes a new version to be submitted according to the workpiece coordinate chain update rules and publishes it after the rescan verification is completed, so that subsequent weld rescan and pose fusion are performed under the same coordinate semantics; if the transformation shows that the drift is within an acceptable range, the controller keeps the workpiece coordinate chain version unchanged and only records the rescan verification pass mark.

[0059] Weld feature rescanning is performed after the baseline semantics have stabilized. The controller performs rescanning acquisition on the weld centerline point set and the weld neighborhood surface point set, and performs consistency comparison between the rescanned point set and the current weld feature model according to the version management rules of the weld feature model. During the comparison, the controller does not pursue complete point-to-point overlap, but uses the consistency of the centerline arc length parameter, the consistency of the neighborhood surface normal, and the convergence of the residual statistics as the criteria: if the rescanning result shows that the predicted quantity and the observed residual of the original weld feature model have recovered to the normal form, the controller records the rescanning result as verified; if the rescanning result shows that the local geometry of the weld has undergone systematic shift, the controller generates a new version snapshot according to the writing rules of the weld feature model, writes it into the collaborative data volume and publishes it to ensure that the subsequent trajectory mapping and model consistency calculation refer to the updated weld feature model.

[0060] Torque zero-point reset is performed after geometric semantic verification. Torque zero-point reset refers to the controller reconstructing the zero point of the torque measurement channel after confirming that the tool has been removed from the contact area and the end effector is in a non-contact state. This ensures that the fused contact baseline has a reliable reference before the start of a new round of contact control. In this process, the controller first reads a segment of torque observation sequence after removal and performs stability judgment on the sequence, excluding moments where vibration, impact, or accidental contact still exists. Within the range that meets the stability conditions, the controller uses the statistical mean or equivalent reference quantity as the current zero point and writes it into the pending submission area of ​​the fused contact baseline. At the same time, the time index and process index associated with the zero-point reset action are recorded in the collaborative data body to form a traceable baseline reconstruction record.

[0061] After completing the perception verification sequence, the controller recalculates the process confidence and updates the version numbers of the fusion model and the fusion contact baseline. In practice, the controller takes the re-scanned reference point set, the re-scanned weld feature point set, and the baseline after torque zero-point reset as inputs to regenerate the channel quality index and model consistency index, and generates new process confidence records accordingly. At the same time, the controller updates the version number of the weight matrix referenced by the fusion model to the new version after reconstruction through the version manager, and switches the current version number of the fusion contact baseline to the new version sequence containing the zero-point reset record. The purpose of the version number update is to clarify the data boundaries before and after reconstruction: the fusion pose and fusion contact baseline records before reconstruction no longer participate in the fusion update after reconstruction, and the control sequence after reconstruction only reads the records with the new version number that have been published, thereby avoiding leaving erroneous observations during the mismatch period to subsequent closed loops.

[0062] For example: During the grinding stage, the dust obstruction rate continuously increases and the residual sequence shows a continuous bias. The process confidence summary component remains low for multiple control cycles. After the mismatch criterion is met, the process state chain enters the perception reconstruction state. The controller first removes the tool to the rescan pose, completes the rescan of the fixture reference mark point and the workpiece boundary, and verifies the workpiece coordinate chain. Then, it rescans the weld centerline and the neighboring surface to verify the weld feature model. Finally, it completes the torque zero point reset under non-contact conditions. After completion, the process confidence is recalculated and the version number of the fusion model and the fusion contact baseline is switched. The grinding task then returns to the normal state and continues to execute.

[0063] In this embodiment, after the process state chain enters the indexing switching window, the controller generates a transition trajectory and a transition contact instruction, which are written as process data in the collaborative data body to ensure the continuity of kinematics and contact control between the welding stage and the grinding stage. The transition trajectory is a sequence of trajectory segments that can be directly issued, which is formed by sequentially splicing the welding end trajectory segment, the indexing safety trajectory segment, and the grinding entry trajectory segment. All three trajectory segments are represented under the workpiece coordinate chain and the tool mounting chain, so that the coordinate semantics of the trajectory points are consistent with the subsequent trajectory mapping, and the break caused by the mixing of coordinates across processes is avoided.

[0064] The welding end trajectory segment is taken from the tail segment of the welding control sequence, and its end pose corresponds to the tool end state when the welding task is completed; the grinding entry trajectory segment is taken from the first segment of the grinding control sequence, and its starting pose corresponds to the entry state before the grinding task enters contact; the indexing safety trajectory segment is generated by the controller in the indexing switching window, and is used to complete tool removal, avoidance, maintaining the indexing allowable posture, and returning to the approach area after indexing; this safety trajectory simultaneously satisfies the spatial constraints of the indexing mechanism and the posture constraints of the tool installation chain: the controller first reads the indexing allowable posture constraints from the safety rule base, including the distance requirement of the indexing axis interference zone, the allowable range of tool posture, and the boundary of the fixture avoidance zone; then, based on the workpiece coordinate chain, the constraints are mapped to the current workpiece spatial semantics, and the feasible path from the welding end pose to the vicinity of the grinding entry pose is obtained in the constraint set; after the feasible path is discretized, the indexing safety trajectory segment is formed, and a time index and a process index are added to each trajectory point for alignment with the transition contact command.

[0065] At the connection points of adjacent trajectory segments, the controller repositions the trajectory endpoints based on the inversion superposition compensation transformation to eliminate endpoint misalignment caused by changes in coordinate semantics before and after the inversion. The welding end point belongs to the pre-inversion semantics, while the grinding start point belongs to the post-inversion semantics. Direct splicing would cause the connection segment to use the incorrect endpoints as boundaries, leading to pose jumps or pullbacks. Therefore, before generating the connection segment, the controller reads the published inversion superposition compensation transformation from the collaborative data body, maps the inversion front point to the post-inversion semantics, or pushes the post-inversion endpoints back to the pre-inversion semantics, so that the endpoints are in the same coordinate chain semantics before connection. The repositioned endpoints and the version number of the referenced transformation are written into the collaborative data body to facilitate tracing the source of the endpoints and the basis of the transformation.

[0066] The connection segment is generated using piecewise polynomial interpolation based on pose continuity constraints. The controller first determines the boundary conditions, which are jointly provided by the pose of the repositioning endpoints, the velocity direction information of the endpoints of adjacent trajectory segments, and the timing of the switching window. Then, the position components are interpolated to ensure that the connection segment is positionally continuous with the adjacent trajectory segments at both ends and maintains a smooth velocity trend. The attitude components are interpolated using the attitude parameters defined by the tool mounting chain to ensure that the attitude change direction is consistent with the adjacent segments and to limit the rate of change, avoiding unnecessary flipping. After generation, the connection segment is constrained and verified, including the fixture avoidance zone boundary, the rotation interference zone, and the allowable range of tool attitude. If it fails, the connection segment is recalculated according to the same boundary conditions or the intermediate control points of the safe trajectory segment are adjusted until it passes. The connection segment and the three trajectory segments form a complete transition trajectory and are written into the collaborative data body according to the time index.

[0067] The transition contact command is generated by the fused contact baseline and the pose sequence of the connecting segment and written into the collaborative data volume. This command arranges the release, preparation and re-establishment of contact control within the switching window to reduce the risk of transient false contact when entering the grinding process. The controller first reads the latest published fused contact baseline in the collaborative data volume as the zero-point reference, and then determines whether the tool end is in the withdrawal zone, approach zone or contact establishment zone based on the pose sequence of the connecting segment. In the withdrawal zone, a contact release command is generated and the baseline is frozen. In the approach zone, the baseline is kept frozen and a contact preparation command is enabled, so that the torque observation is only used for abnormal contact monitoring. In the contact establishment zone, a contact start command is generated based on the fused contact baseline and a slow gain enable strategy is adopted, so that the contact constraint filtering model gradually takes over the baseline update under the condition of continuous pose. The command sequence and the trajectory points of the connecting segment are aligned with the same time index and written into the collaborative data volume to form a traceable pairing record of pose sequence and contact command sequence.

[0068] For example: When the welding end trajectory segment ends, the tool pose is near the end of the weld. The indexing safety trajectory segment first lifts the tool and moves it outside the fixture avoidance area. After the indexing is completed, it returns to the grinding entry position. After the controller reads the indexing superimposed compensation transformation, it repositions the welding end point to make it semantically consistent with the indexed point. Then, it uses piecewise polynomial interpolation to generate the connection segment to ensure that the position and attitude are continuous at the switching point. The transition contact command keeps the contact released in the lifting and withdrawal segment, keeps the baseline frozen in the approach segment, and starts the contact control with the fused contact baseline as a reference after entering the grinding contact establishment area. The command sequence is written into the collaborative data body with time index for direct call by the grinding control sequence.

[0069] Example 2 like Figure 2 As shown, this embodiment further improves the design based on embodiment 1. The difference is that in actual operation of embodiment 1, it was found that there was intermittent dust obstruction and surface reflection superposition near the grinding entry section, resulting in fragmented loss of the visual measurement channel and unilateral echo bias in the weld seam neighborhood of the laser measurement channel. This phenomenon may not trigger the mismatch criterion when the weight has been reduced, causing the fused pose to drift slowly around the weld seam neighborhood normal near the entry section. The contact constraint filter mistakenly absorbs this drift into the fused contact baseline before and after contact establishment, forming a baseline bias and propagating to subsequent control. It failed to stably maintain the consistency of the contact establishment position, contact direction and baseline of the grinding entry section without frequently entering the perception reconstruction state. Based on this, the intelligent collaborative control method of the welding and grinding indexable robot also includes: confidence trend prediction and verification anchor insertion steps, which are used to transform the "imminent confidence gap" into a controllable and plannable verification action in advance.

[0070] The specific implementation is as follows: After completing step S5 and obtaining the current process confidence record, the controller does not immediately use it only for updating the weight matrix in the current iteration. Instead, it reads the process confidence component and residual sequence statistics from a continuous time index in the collaborative data volume to calculate the confidence trend, which is used to characterize whether the confidence is "recovering," "stabilizing," or "continuously deteriorating." At the same time, the controller uses the grinding trajectory generated in step S4 to establish observability labels in the trajectory domain: mapping the trajectory points to the centerline position and neighborhood surface normal corresponding to the weld feature model. Combining the missing patterns, outlier patterns, and spatial distribution of residuals that have appeared at this position in the point cloud, the trajectory is divided into observational stable segments and observational vulnerable segments. Based on this, the controller selects a verification anchor point at the leading edge of the observational vulnerable segment and writes the verification anchor point and its triggering conditions into the collaborative data volume as constraint inputs when generating the transition trajectory and transition contact command in the subsequent step S6.

[0071] During step S6, the process state chain issues the transition trajectory and transition contact instructions simultaneously with the verification anchor point as the switching window and the execution plan for the first grinding stage. Before reaching the verification anchor point, the process state chain enters a preventative verification sub-state: the controller first freezes the fused contact baseline, restricts the writing of contact constraint filtering to the baseline, and then selects to execute either a baseline feature rescan or a weld feature rescan based on the observability label. The rescan is not equivalent to a complete reconstruction; it uses the local area near the anchor point as its scope and reuses the existing geometric constraint matching and random consistency sampling process. The purpose is simply to obtain a "verification transformation" or "verification residual correction". After the rescan is completed, the controller recalculates the process confidence using the verification results and updates the observation weight matrix of the multi-channel weighted filtering model accordingly. At the same time, the version manager switches the version number of the weight matrix referenced by the fusion model and the version number of the fusion contact baseline to ensure that the fusion pose before and after the anchor point forms a clear boundary with the fusion contact baseline, preventing the drift before the anchor point from continuing to "drive" the contact establishment process after the anchor point. Then the system exits the preventive verification sub-state and returns to the normal polishing state to continue execution.

[0072] For example: When the dust occlusion rate near the grinding entry section shows an upward trend over several control cycles and the residual bias continues to accumulate, the trend is judged as continuous deterioration. The controller marks a section of the trajectory before the entry section as a vulnerable observation section and inserts a verification anchor point at the leading edge of this section. Before reaching the anchor point, the fusion contact baseline is frozen, the weld feature is rescanned to obtain the verification transformation and correct the residual, and then the fusion update is restored and the baseline version number is switched.

[0073] This way, even without triggering the perception refactoring state, the drift of the entry segment can be cut off outside the version boundary.

[0074] Example 3 like Figure 3As shown, based on the same inventive concept as the intelligent collaborative control method for a welding and grinding indexable robot in the foregoing embodiments, this application provides an intelligent collaborative control system for a welding and grinding indexable robot. The system and method embodiments in this application are based on the same inventive concept. The system includes: Collaborative Data Module: The collaborative data module establishes a unified collaborative data body, which includes a workpiece coordinate chain, a rotation coordinate chain, a tool installation chain, and a process status chain; Weld seam modeling module: During the welding stage, the weld seam modeling module collects weld seam point cloud data, robot joint data and welding torch arc status data, generates weld seam feature model and writes it into the collaborative data body; The indexing compensation module collects the workpiece reference feature point set and the weld feature point set before indexing, and collects the corresponding point set again after indexing. It calculates the reference transformation and the weld transformation respectively, and synthesizes them in sequence to obtain the indexing superposition compensation transformation to update the workpiece coordinate chain. Trajectory Mapping Module: During the grinding stage, the trajectory mapping module generates a grinding trajectory based on the updated weld feature model and maps it to the tool installation chain; Confidence Fusion Module: During the welding and grinding stages, the confidence fusion module calculates the process confidence for the visual measurement channel, laser measurement channel, and torque measurement channel, respectively. It selects the fusion model according to the process confidence and outputs the fusion pose and fusion contact baseline. Switching Continuous Module: The switching continuous module transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence through the process state chain, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

Claims

1. An intelligent collaborative control method for a welder and grinder indexable robot, characterized in that the method... include: Establish a unified collaborative data body, which includes a workpiece coordinate chain, a rotation coordinate chain, a tool installation chain, and a process status chain; During the welding stage, weld seam point cloud data, robot joint data, and welding torch arc status data are collected to generate a weld seam feature model and write it into the collaborative data body. Before the indexing stage, the workpiece reference feature point set and the weld feature point set are collected. After the indexing, the corresponding point set is collected again. The reference transformation and the weld transformation are calculated respectively, and the indexing superposition compensation transformation is synthesized in sequence to update the workpiece coordinate chain. During the grinding stage, a grinding trajectory is generated based on the updated weld feature model and mapped to the tool installation chain; During the welding and grinding stages, the process confidence level is calculated for the visual measurement channel, laser measurement channel, and torque measurement channel, respectively. The fusion model is selected according to the process confidence level, and the fusion pose and fusion contact baseline are output. The process state chain transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

2. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The collaborative data body sets time index, process index and version number fields, and sets write locking rules for the workpiece coordinate chain, weld feature model, fused pose and fused contact baseline respectively; at the end of the welding stage, a new version is written to the weld feature model, at the end of the transposition stage, a new version is written to the workpiece coordinate chain, and during the grinding stage, the fused pose and fused contact baseline are continuously written to the same version sequence according to the time index.

3. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The workpiece reference feature point set consists of the fixture reference mark point set and the workpiece reference boundary point set. The weld feature point set consists of the weld centerline point set and the weld neighborhood surface point set. The reference transformation and weld transformation establish a correspondence through geometric constraint matching, and obtain the inner point set through random consistency sampling. Then, the rigid body transformation is obtained from the inner point set and written into the collaborative data volume.

4. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The indexing superposition compensation transformation is obtained by combining mechanical indexing transformation, reference drift correction transformation and thermal deformation transformation in the order of the operation. The mechanical indexing transformation is obtained by the encoder data of the indexing device and the kinematic model of the indexing device. The reference drift correction transformation is obtained by the reference feature point set of the workpiece before and after indexing. The thermal deformation transformation is obtained by the weld feature point set before and after indexing. The combined result of the three is used to update the workpiece coordinate chain and for grinding trajectory mapping.

5. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The process confidence score is composed of channel quality indicators and model consistency indicators. In the welding stage, the channel quality indicators are composed of arc saturation indicators, spatter outlier rate, and point cloud missing rate. In the grinding stage, the channel quality indicators are composed of dust obstruction rate, vibration disturbance index, and contact signal drift rate. The model consistency indicator is generated by the residual sequence of weld feature model prediction and channel observation, and together with the channel quality indicators, it generates the process confidence score and writes it into the collaborative data volume.

6. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The fusion model consists of a multi-channel weighted filtering model and a contact constraint filtering model. The multi-channel weighted filtering model determines the observation weight matrix based on the process confidence and outputs the fused pose. The contact constraint filtering model constructs contact constraints based on the fused pose and torque measurement channel observations and outputs the fused contact baseline. When the process state chain enters the rotation switching window, the process confidence of the welding stage is mapped to the initial weight matrix, and the first segment of the grinding stage is updated to the weight matrix corresponding to the process confidence of the grinding stage.

7. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The process state chain includes a perception reconstruction state, which is triggered by the process confidence meeting the mismatch criterion. When entering the perception reconstruction state, a perception verification sequence is executed. The perception verification sequence consists of a baseline feature rescan, a weld feature rescan, and a torque zero-point reset in sequence. After completing the perception verification sequence, the process confidence is recalculated and the version numbers of the fusion model and the fusion contact baseline are updated.

8. The intelligent collaborative control method for the indexable welding and grinding robot according to claim 1, characterized in that, The transition trajectory consists of a welding end trajectory segment, a rotation safety trajectory segment, and a grinding entry trajectory segment connected sequentially. All three trajectory segments are represented under the workpiece coordinate chain and the tool mounting chain. At the connection of adjacent trajectory segments, the trajectory endpoints are repositioned according to the rotation superposition compensation transformation, and the pose continuity constraint is generated by piecewise polynomial interpolation. The transition contact command is generated by the fusion contact baseline and the pose sequence of the connection segment and written into the collaborative data body.

9. The intelligent collaborative control method for the welding and grinding indexable robot according to claim 1, characterized in that, Also includes: After recording the confidence level of the process, calculate the confidence level trend and mark the observability of the grinding trajectory. Select the verification anchor point and write it into the collaborative data volume. Before reaching the anchor point, freeze the fusion contact baseline and perform a local rescan. Update the weight matrix and version number.

10. A lightweight, age-friendly exoskeleton control system based on gait analysis, utilizing the intelligent collaborative control method of the welding and grinding indexable robot according to any one of claims 1-9, characterized in that the system include: A collaborative data module establishes a unified collaborative data body, which includes a workpiece coordinate chain, an indexing coordinate chain, a tool installation chain, and a process status chain. The weld seam modeling module collects weld seam point cloud data, robot joint data, and welding torch arc state data during the welding stage, generates a weld seam feature model, and writes it into the collaborative data body. The indexing compensation module collects the workpiece reference feature point set and the weld feature point set before indexing, and collects the corresponding point set again after indexing. It calculates the reference transformation and the weld transformation respectively, and synthesizes them in sequence to obtain the indexing superposition compensation transformation to update the workpiece coordinate chain. The trajectory mapping module generates a grinding trajectory based on the updated weld feature model during the grinding stage and maps it to the tool mounting chain. The confidence fusion module calculates the process confidence level for the visual measurement channel, laser measurement channel, and torque measurement channel during the welding and grinding stages, respectively, selects the fusion model according to the process confidence level, and outputs the fusion pose and fusion contact baseline. The switching continuous module transmits the rotation superposition compensation transformation, fusion pose and fusion contact baseline between the welding control sequence and the grinding control sequence through the process state chain, generates the transition trajectory and transition contact command within the rotation switching window, and applies pose continuity constraints and contact baseline continuity constraints to the transition trajectory.

Images