Multi-sensor dynamic path planning method and system of large iron tower component free teaching welding robot

Abstract

The application relates to the technical field of path planning, and discloses a multi-sensor dynamic path planning method and system for a large iron tower component free-teaching welding robot, wherein the method comprises the following steps: collecting multi-sensor synchronous data through a laser sensor, a visual sensor and an electric arc sensor and solving segmented calibration parameters; combining the segmented calibration parameters and the multi-sensor synchronous data to calculate three-dimensional position data of a weld; performing error compensation on the three-dimensional position data of the weld at a segmented switching point to obtain a weld track position; performing curve fitting on the weld track position as a control point and setting a constraint condition at a segmented boundary to obtain a target welding path. The method performs error prediction and real-time correction at the segmented switching point, effectively suppresses cumulative positioning deviation, and further realizes high-precision free-teaching welding of a long-distance weld of a large iron tower component, and solves the technical problems of low precision and poor efficiency of a traditional method.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to a multi-sensor dynamic path planning method and system for a teach-free welding robot for large iron tower components. Background Technology

[0002] Traditional welding of large iron tower components relies primarily on manual teaching programming and single sensor guidance. However, in long-distance weld seam tracking, due to the large size of the components and the complex and variable weld seam paths, manual teaching is not only inefficient but also struggles to guarantee the accuracy and consistency of the welding path. Existing welding robot systems often employ fixed path planning when handling long weld seams, failing to adapt to real-time changes during the welding process, leading to unstable weld quality. Furthermore, during long-distance welding, factors such as sensor coordinate system drift and accumulated thermal deformation cause the end-effector positioning error to gradually amplify. Traditional overall calibration methods cannot effectively address the problem of segmented error compensation, ultimately resulting in low welding accuracy. Summary of the Invention

[0003] This invention provides a multi-sensor dynamic path planning method and system for a teachless welding robot for large iron tower components. This invention performs error prediction and real-time correction at segment switching points, effectively suppressing cumulative positioning deviations, thereby realizing high-precision teachless welding of long-distance welds in large iron tower components, and solving the technical problems of low accuracy and poor efficiency of traditional methods.

[0004] In a first aspect, the present invention provides a multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components, the multi-sensor dynamic path planning method for the teach-free welding robot for large iron tower components comprising:

[0005] The system acquires synchronous data from multiple sensors, including laser, vision, and arc sensors, and solves for segmented calibration parameters.

[0006] The three-dimensional position data of the weld are calculated by combining the segmented calibration parameters and the multi-sensor synchronous data;

[0007] Error compensation is performed on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position;

[0008] The weld trajectory position is used as a control point for curve fitting, and constraints are set at the segment boundaries to obtain the target welding path.

[0009] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of acquiring synchronous data from multiple sensors—a laser sensor, a vision sensor, and an arc sensor—and solving for segmented calibration parameters includes:

[0010] Using the laser sensor in the welding robot, line laser scanning is performed on the weld area of ​​large iron tower components to obtain laser point cloud data;

[0011] Using a vision sensor in a welding robot, weld images of the weld area are acquired, and visual feature data is extracted based on the weld images;

[0012] Arc current data during the welding process of the large iron tower components are collected using an arc sensor in a welding robot.

[0013] Based on the unified reference clock signal issued by the main controller, the laser point cloud data, the visual feature data and the arc current data are synchronized to obtain multi-sensor synchronized data;

[0014] Based on the multi-sensor synchronous data, segmented calibration points are set at preset intervals on the weld seam, and the segmented calibration parameters are solved.

[0015] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of setting segmented calibration points at preset intervals on the weld seam based on the multi-sensor synchronous data and solving for the segmented calibration parameters includes:

[0016] Based on the laser point cloud data in the multi-sensor synchronous data, the spatial trajectory of the weld is analyzed and calibration point positions are set along the spatial trajectory to obtain a segmented calibration point sequence.

[0017] At each calibration point in the segmented calibration point sequence, extract the homogeneous coordinate data of the laser sensor and the feature point coordinate data of the vision sensor, and record the corresponding robot end pose matrix to obtain the coordinate correspondence data at the calibration point.

[0018] The coordinate correspondence data is subjected to singular value decomposition, and the rotation matrix and translation vector are solved. The rotation matrix and the translation vector are used to form the coordinate transformation matrix of each segment.

[0019] Segment calibration parameters are generated based on the coordinate transformation matrix of each segment, the length information of the corresponding segment, and the position of the calibration point.

[0020] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the multi-sensor dynamic path planning method for the teach-free welding robot for large iron tower components further includes:

[0021] Based on the length information of the corresponding segment in the segment calibration parameters, the temperature gradient data of each segment is monitored by an infrared temperature sensor.

[0022] Based on the temperature gradient data, the thermal deformation rate of each segment during the welding process is calculated using the material thermal expansion coefficient of the large iron tower components.

[0023] Multiply the thermal deformation rate of each segment by the length information of the corresponding segment, calculate the thermal deformation accumulation term, and set the error attenuation coefficient to calculate the attenuation and propagation of the error of the previous segment, so as to obtain the error accumulation calculation parameters.

[0024] Based on the error accumulation calculation parameters, a recursive relationship is established that the current segment error equals the previous segment error multiplied by the attenuation coefficient plus the current segment thermal deformation accumulation term, thus obtaining the error accumulation model.

[0025] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of calculating the three-dimensional position data of the weld seam by combining the segmented calibration parameters and the multi-sensor synchronous data includes:

[0026] The coordinate transformation matrix of each segment in the segmented calibration parameters is used to perform coordinate system transformation on the laser point cloud data and visual feature data in the multi-sensor synchronous data to obtain the laser point cloud coordinates and visual feature point coordinates.

[0027] A target state vector is constructed based on the laser point cloud coordinates and the visual feature point coordinates. The target state vector is then input into the upper-level main filter for Kalman filtering state estimation to obtain the main filter output result.

[0028] The arc current data in the multi-sensor synchronization data is compared with the set reference arc current value by performing difference integration to obtain the difference integration sequence. The difference integration sequence is then input into the lower-level sub-filter for lateral deviation filtering to obtain the sub-filter output result.

[0029] Global optimal estimation is performed based on the output results of the main filter and the sub-filter to obtain the three-dimensional position data of the weld.

[0030] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of performing a global optimal estimation based on the output results of the main filter and the sub-filter to obtain the three-dimensional position data of the weld seam includes:

[0031] Weld position and velocity information are extracted from the output of the main filter, and arc lateral deviation information is extracted from the output of the sub-filter to obtain dual-layer filtered state information.

[0032] Based on the differences in measurement accuracy among the laser sensor, vision sensor, and arc sensor, information allocation factors are set for the output results of the main filter and the sub-filter, respectively, to obtain the sensor weight allocation coefficients;

[0033] The inverse of the covariance matrix of the main filter output and the inverse of the covariance matrix of the sub-filter output are multiplied by the corresponding sensor weight allocation coefficients and then summed with weights to obtain the global information matrix.

[0034] The global covariance matrix is ​​calculated based on the global information matrix, and the optimal state estimate is calculated by weighted fusion in combination with the dual-layer filtered state information to obtain the three-dimensional position data of the weld.

[0035] In conjunction with the first aspect, in the sixth implementation of the first aspect of the present invention, the step of performing error compensation on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position includes:

[0036] Based on the error accumulation model, the error prediction data of the current segment is calculated at the segment switching point;

[0037] Calculate the Jacobian transformation matrix of each coordinate component in the three-dimensional position data of the weld at the segmented switching point based on the error prediction data;

[0038] The error compensation vector is obtained by performing an inverse matrix operation on the Jacobian transformation matrix and multiplying it with the error prediction data. The error compensation vector is then added to the three-dimensional position data of the weld to obtain the target weld position data.

[0039] The target weld position data is smoothed at the segment boundaries to obtain the weld trajectory position.

[0040] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of using the weld trajectory position as a control point for curve fitting and setting constraint conditions at the segment boundaries to obtain the target welding path includes:

[0041] The weld trajectory positions are sequentially arranged according to the preset control point spacing as control points of a cubic B-spline curve, and the initial fitting path is obtained by fitting the curve using cubic B-spline basis functions.

[0042] Based on the initial fitted path, constraints are set at the boundaries of each segment and constraint optimization is performed to obtain the path control points.

[0043] Calculate the curvature value of the path control point and compare it with the maximum executable curvature of the welding robot;

[0044] When the curvature value is detected to exceed the maximum executable curvature, the initial fitted path is replanned locally to avoid obstacles and the control points are updated to obtain the target welding path.

[0045] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the multi-sensor dynamic path planning method for the teach-free welding robot for large iron tower components further includes:

[0046] The tracking error vector is calculated based on the target welding path and the current position of the welding robot, and the basic data for sliding mode control is generated based on the tracking error vector.

[0047] Based on the aforementioned sliding mode control fundamental data, the equivalent control term is calculated using the system state matrix, control gain matrix, and feedback matrix.

[0048] The adaptive gain coefficient is calculated based on the ratio of the current welding length to the total welding length, and the switching control term is calculated in combination with the sliding surface sign function in the sliding mode control basic data.

[0049] The equivalent control term and the switching control term are added together and then the control continuity is processed at the segmented switching points to generate the trajectory tracking control command for the welding robot.

[0050] Secondly, the present invention provides a multi-sensor dynamic path planning system for a teach-free welding robot for large iron tower components, the multi-sensor dynamic path planning system for the teach-free welding robot for large iron tower components comprising:

[0051] The data acquisition module is used to acquire synchronous data from multiple sensors, including laser sensors, vision sensors, and arc sensors, and to solve for segmented calibration parameters.

[0052] The calculation module is used to calculate the three-dimensional position data of the weld by combining the segmented calibration parameters and the multi-sensor synchronous data;

[0053] An error compensation module is used to perform error compensation on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position;

[0054] The curve fitting module is used to perform curve fitting using the weld trajectory position as control points and set constraint conditions at the segment boundaries to obtain the target welding path.

[0055] The technical solution provided by this invention achieves synchronous hardware acquisition of laser, vision, and arc sensors through a unified reference clock signal, solving the problem of inconsistent data fusion time in traditional multi-sensor systems. A segmented calibration strategy and singular value decomposition are used to solve the coordinate transformation matrix, effectively controlling sensor coordinate system drift during long-distance welding. A hierarchical Kalman filter architecture with an upper-level main filter and lower-level sub-filters is constructed, and optimal fusion of multi-sensor information is achieved through a federated filtering structure, significantly improving weld position detection accuracy. Jacobian transform matrix error compensation technology based on an error accumulation model performs forward-looking error prediction and real-time correction at segmented switching points, effectively suppressing accumulated positioning deviations. A smooth welding path is generated using cubic B-spline curve fitting and continuity constraints, combined with an adaptive sliding mode controller to coordinate equivalent control terms and switching control terms, dynamically adjusting the control gain according to the welding progress. Ultimately, high-precision, teach-free welding of long-distance welds in large iron tower components is achieved, solving the technical problems of low accuracy and poor efficiency in traditional methods.

[0056] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of an embodiment of the multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components in this invention.

[0059] Figure 2 This is a schematic diagram of an embodiment of the multi-sensor dynamic path planning system for a teach-free welding robot for large iron tower components in this invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0062] To facilitate understanding of this embodiment, a multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components, as disclosed in this embodiment of the invention, will first be described in detail. For example... Figure 1 As shown, the multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components includes the following steps:

[0063] 101. Collect synchronous data from multiple sensors, including laser, vision, and arc sensors, and solve for segmented calibration parameters;

[0064] Specifically, a laser sensor mounted on the end effector of the welding robot performs line laser scanning on the weld area of ​​the large iron tower component. The laser sensor operates at a wavelength of 650nm, a scanning frequency of 200Hz, a scanning angle range of ±30°, and an accuracy of 0.1mm, acquiring high-density laser point cloud data covering the weld surface through continuous scanning. Simultaneously, a binocular vision sensor mounted on the robot acquires image information of the weld area at a rate of 30 frames per second and a resolution of 1920×1080 pixels. A preset visual processing algorithm extracts feature data characterizing the spatial morphology of the weld from the image. During welding, an arc sensor detects changes in the arc current in real time within a range of 50–300A at a sampling frequency of 1kHz. This current data is input into the data fusion module as auxiliary information reflecting welding stability and weld positional deviation. The main controller outputs a unified reference clock signal with a frequency of 1kHz. This signal is then used in hardware to drive the laser sensor to sample at a frequency division of 5, the vision sensor at 33, and the arc sensor at 1. At the data receiving end, linear interpolation is used to map data from different frequencies onto a unified time axis, keeping the time synchronization error within ±0.5ms. On the time-synchronized dataset, a calibration point is set at preset intervals (e.g., 1.5 meters) along the weld path, dividing the long weld into multiple independent segments. At each calibration point, laser point cloud, visual features, and robot end-effector pose data are simultaneously accessed. The calibration parameters for each segment are obtained through a coordinate transformation matrix solution method, effectively suppressing accuracy degradation caused by robotic arm posture drift, thermal deformation, and measurement delays during long-distance welding.

[0065] 102. Calculate the three-dimensional position data of the weld by combining the segmented calibration parameters and multi-sensor synchronous data;

[0066] Specifically, the coordinate transformation matrices corresponding to each segment in the segmented calibration parameters are invoked to accurately transform the laser point cloud data and visual feature data from the multi-sensor synchronous data from the sensor's own coordinate system to the welding robot's global coordinate system, obtaining laser point cloud coordinates and visual feature point coordinates with a unified spatial reference. Based on the laser point cloud coordinates and visual feature points, a target state vector containing weld position and velocity information is constructed. This target state vector is then input into the upper-level main filter for Kalman filtering state estimation. The upper-level main filter effectively suppresses single-sensor noise and improves 3D positioning accuracy during the fusion of laser and visual information, outputting the main filter estimation result reflecting the global geometric position and dynamic changes of the weld. Simultaneously, arc current data is extracted from the multi-sensor synchronous data. The difference between the arc current data and a preset reference arc current value is calculated, and the difference is integrated over time to obtain a difference integral sequence, reflecting the cumulative offset trend of the weld's lateral position during welding. This difference integral sequence is input into the lower-level sub-filter for filtering of the lateral deviation, yielding the sub-filter's output result. The outputs of the main filter and the sub-filter are fused together in the global estimation module according to the preset information weights to achieve the global optimal estimation and obtain the three-dimensional position data of the weld.

[0067] 103. At the segment switching point, perform error compensation on the three-dimensional position data of the weld to obtain the weld trajectory position;

[0068] Specifically, based on the error accumulation model, the system comprehensively considers the positioning error accumulated during the previous welding process, the current segment length, and thermal deformation prediction data, combined with system noise, to calculate the error prediction data for the current segment in real time at the segment switching point. Based on the error prediction data, the sensitivity of each coordinate component in the weld's 3D position data to changes at the segment switching point is analyzed, and a corresponding Jacobian transformation matrix is ​​constructed using numerical differentiation. Each element of the matrix represents the degree of influence of different error components on the coordinate components of the weld position. The Jacobian transformation matrix is ​​inversely multiplied and then multiplied by the error prediction data to obtain the error compensation vector, reflecting the 3D correction amount to be applied to the weld position data. The error compensation vector is then added to the original weld 3D position data to obtain the compensated target weld position data. To ensure the continuity and smoothness of the trajectory at the segment switching point, a smoothing process is introduced at the segment boundary based on the target weld position data. A smooth transition function is used to optimize the continuity of the trajectory near the boundary, ensuring that the weld trajectory has no significant abrupt changes in space and maintains good executability, outputting the weld trajectory position.

[0069] 104. Use the weld trajectory position as a control point to perform curve fitting and set constraints at the segment boundaries to obtain the target welding path.

[0070] Specifically, the compensated weld trajectory positions are sequentially arranged according to the preset control point spacing. These sequential spatial points are then used as control points for a cubic B-spline curve, inputting into the curve generation module. The control point sequence is fitted using cubic B-spline basis functions to obtain an initial fitted path with continuous and smooth characteristics. This initial fitted path smoothly follows the changes in the weld spatial distribution. Continuity constraints are introduced at the boundaries of each segment of the initial fitted path. These constraints include positional continuity, as well as the continuity of the first and second derivatives, ensuring that the direction and curvature changes of the path remain stable at the segment junctions. A constraint optimization solution method is then used to adjust the control point positions while satisfying the continuity conditions, resulting in a constraint-optimized set of path control points. The curvature values ​​of the path control points in the path parameter space are calculated and compared with the maximum executable curvature limit that the welding robot can execute. When the curvature value at any position exceeds the executable limit, the local obstacle avoidance path replanning process is triggered. A new path segment that meets the curvature limit is generated in the local area through an improved sampling planning algorithm. The new local path segment is seamlessly spliced ​​with the original path, and the corresponding control point positions are updated to obtain the target welding path.

[0071] In one specific embodiment, the process of performing step 101 may specifically include the following steps:

[0072] Using the laser sensor in the welding robot, line laser scanning is performed on the weld area of ​​large iron tower components to obtain laser point cloud data;

[0073] Using the vision sensors in the welding robot, weld images of the weld area are acquired, and visual feature data is extracted based on the weld images;

[0074] Arc sensors in welding robots are used to collect arc current data during the welding process of large iron tower components;

[0075] Based on the unified reference clock signal issued by the main controller, the laser point cloud data, visual feature data and arc current data are synchronized to obtain multi-sensor synchronized data.

[0076] Based on the synchronous data from multiple sensors, segmented calibration points are set at preset intervals on the weld seam, and the segmented calibration parameters are solved.

[0077] Specifically, a laser line scanning sensor mounted on the end effector of the welding robot performs a full-area line laser scan of the weld area of ​​the target component. The laser line scanning sensor operates at a wavelength of 650nm, a scanning frequency of 200Hz, and has a scanning angle range of ±30° and a measurement accuracy of 0.1mm. By continuously moving along the weld trajectory and emitting a line laser beam in real time, it acquires high-density three-dimensional laser point cloud data containing the geometric features of the weld surface in a very short time. Simultaneously, a binocular vision sensor mounted on the robot's end effector operates at a frame rate of 30 frames per second and a resolution of 1920×108 pixels. The system acquires full-color stereo images of the weld area with a resolution of 0 pixels. The optical structure uses imaging parameters of a 12mm focal length and a 120mm baseline distance. Stereo matching and feature extraction algorithms extract 2D and 3D visual feature data reflecting the weld's position, shape, and boundary characteristics from the acquired images. Throughout the welding process, the arc sensor monitors the current fluctuations of the welding arc in real time with a sampling frequency of 1kHz and a measurement range of 50–300A, achieving a measurement accuracy of ±1A. The arc current data reflects the stability of the heat input during welding and is correlated with minute shifts in the weld position. The main controller outputs a unified reference clock signal with a frequency of 1kHz, controlling the sampling time of various sensors through a hardware triggering mechanism. The laser sensor responds at a frequency division of 5, the vision sensor at 33, and the arc sensor at 1. At the acquisition end, linear interpolation and other time axis reconstruction algorithms are applied to the data acquired at different frequencies to map the raw data from all sensors to the same time reference, thereby controlling the time synchronization error within ±0.5ms. Based on time-aligned multi-sensor synchronized data, the entire long weld seam is divided into several calibration segments according to a strategy of setting segmented calibration points at preset intervals (1.5 meters) along the weld seam path. At each calibration point, synchronized laser point cloud data, visual feature data, and real-time pose data of the welding robot's end effector are simultaneously invoked. The coordinate transformation matrix corresponding to the segment is calculated using spatial geometric registration and least squares optimization methods. To offset the geometric drift caused by thermal expansion and material stress release during welding, the coordinate transformation matrix is ​​dynamically corrected based on a thermal deformation model, taking into account the real-time temperature distribution measurement results at the segment locations.

[0078] In one specific embodiment, the process of setting segmented calibration points at preset intervals on the weld seam based on multi-sensor synchronous data and solving for the segmented calibration parameters can specifically include the following steps:

[0079] Based on the laser point cloud data in the multi-sensor synchronous data, the spatial trajectory of the weld is analyzed and the calibration point positions are set along the spatial trajectory to obtain a segmented calibration point sequence.

[0080] Extract homogeneous coordinate data of the laser sensor and feature point coordinate data of the vision sensor at each calibration point in the segmented calibration point sequence, and record the corresponding robot end pose matrix to obtain the coordinate correspondence data at the calibration point.

[0081] The coordinate correspondence data is decomposed by singular value decomposition and the rotation matrix and translation vector are solved. The rotation matrix and translation vector are then used to form the coordinate transformation matrix of each segment.

[0082] Segment calibration parameters are generated based on the coordinate transformation matrix of each segment, the length information of the corresponding segment, and the position of the calibration point.

[0083] Specifically, point cloud information collected by the laser sensor is extracted from multi-sensor synchronous data. Spatial geometric analysis is performed on the high-density 3D point cloud data. Utilizing the continuity of weld surface height variations, morphological curves, and boundary features, the overall trajectory of the weld in 3D space is calculated. Calibration point positions are then uniformly distributed along this trajectory or adaptively set based on path complexity, resulting in a segmented calibration point sequence covering the entire weld path. At each calibration point in the segmented calibration point sequence, the original measurement results from the laser sensor are retrieved and converted into homogeneous coordinates. Simultaneously, spatial feature point coordinates corresponding to the laser point cloud are extracted from the weld area image collected by the vision sensor at the same calibration point. The actual pose data of the robot's end effector at the calibration point is recorded as a 4×4 homogeneous pose matrix, thus forming a coordinate correspondence dataset at each calibration point consisting of laser homogeneous coordinates, visual feature point coordinates, and robot end effector pose. The coordinate correspondence data established at each weld path calibration point is input into the matrix solving module. Singular value decomposition (SVD) is used to decompose and calculate the spatial mapping relationship between the corresponding data. A covariance matrix is ​​constructed from the correspondence between the homogeneous coordinate data and the end-effector pose matrix. SVD is then performed on the covariance matrix to obtain the orthogonal matrix factors representing the rotation. The rotation matrix of each segment is directly obtained through matrix multiplication. Simultaneously, the translation vector is calculated based on the center position differences of the calibration points. The rotation matrix and translation vector together constitute the coordinate transformation matrix mapping from the sensor measurement coordinate system to the robot's global coordinate system. To convert the coordinate transformation matrix into directly callable calibration parameters, the length information of each weld segment and the spatial position index of each calibration point are introduced. By associating the segment length, calibration point index, and corresponding coordinate transformation matrix, a set of segment calibration parameters containing segment number, length, start and end calibration point positions, and corresponding rotation matrix and translation vector is generated.

[0084] In one specific embodiment, the multi-sensor dynamic path planning method for performing teachless welding of large iron tower components by a robot further includes the following steps:

[0085] Based on the length information of the corresponding segment in the segment calibration parameters, the temperature gradient data of each segment is monitored by an infrared temperature sensor.

[0086] Based on temperature gradient data and the thermal expansion coefficient of the materials of large iron tower components, the thermal deformation rate of each segment during the welding process is calculated.

[0087] Multiply the thermal deformation rate of each segment by the length information of the corresponding segment, calculate the thermal deformation accumulation term, and set the error attenuation coefficient to calculate the attenuation and propagation of the error of the previous segment, so as to obtain the error accumulation calculation parameters.

[0088] Based on the error accumulation calculation parameters, a recursive relationship is established that the current segment error equals the previous segment error multiplied by the attenuation coefficient plus the current segment thermal deformation accumulation term, thus obtaining the error accumulation model.

[0089] Specifically, the entire weld seam is discretized along its length dimension based on a segmented calibration strategy. A mapping between segment indexes and spatial positions is established according to the calibrated segment length information. Infrared temperature sensors are installed at unobstructed locations on the robot's front end or in the bypass pose. Real-time thermal imaging and point-series temperature sampling are performed on the welding area of ​​each segment. Time alignment with data from other sensors is achieved through a unified 1kHz clock and frequency division triggering mechanism. Fixed temperature measurement windows are set before and after the welding torch, and a moving average and outlier rejection strategy is used to suppress interference from spatter, strong reflections, and smoke on infrared measurements. Simultaneously, sampling density is increased during rapid heat input changes such as weld seam arc initiation, arc termination, and direction reversal to fully capture the temperature gradient distribution along the welding direction and perpendicular to the thickness direction. The thermal expansion coefficients of large iron tower components are retrieved from a materials database, and combined with the temperature gradient of each segment under the current thermal field, the instantaneous thermal deformation rate of each segment during welding is calculated. To avoid systematic deviations caused by batch differences in materials and environmental heat exchange, a small-scale trial welding calibration is performed before welding to finely correct the theoretical coefficients using measured temperature and deformation, ensuring that the deformation rate estimate is consistent with the actual component. Based on the thermal deformation rate of each segment, the cumulative thermal deformation term is obtained by multiplying it by the corresponding segment length information. To reflect the "transmission-attenuation-re-accumulation" mechanism of errors in long-distance welding, an error attenuation coefficient reflecting geometric stiffness and process constraints is set for the model, set at the order of 0.02 per meter, so that the residual error of the previous segment decays exponentially with the length and continues to be transmitted with a smaller weight when entering the next segment. In practical implementation, at the segment switching moment, the error estimate at the end of the previous segment, the thermal deformation accumulation term driven by the temperature gradient of the current segment, and the statistics of the operating noise are simultaneously input into the error recursion module. The error recursion module refreshes the error state at a fixed frequency, ensuring that the relationship "the current segment error equals the attenuated portion of the previous segment error plus the thermal deformation accumulation term of the current segment" remains continuously effective within the control cycle. The geometric state quantity output by the upper-level main filter is used for constraint matching of the temperature gradient, and the stability signal on the arc side is used as a confidence adjustment factor to lightly weight the thermal deformation rate. When abnormal heat input or sudden changes in cooling conditions are detected, the suppression strength of the error attenuation coefficient is temporarily increased to prevent the abnormal segment error from being excessively amplified and spreading to subsequent segments. The calculation loop of the error accumulation model runs in parallel with the path planning and control loop, employing a 100Hz update rate and a smooth transition mechanism at segment boundaries, ensuring that the error state is fully estimated and used for online compensation before entering the next segment.

[0090] In one specific embodiment, the process of performing step 102 may specifically include the following steps:

[0091] The coordinate transformation matrix of each segment in the segmented calibration parameters is used to transform the coordinate system of the laser point cloud data and visual feature data in the multi-sensor synchronous data to obtain the laser point cloud coordinates and visual feature point coordinates.

[0092] A target state vector is constructed based on the coordinates of laser point cloud and visual feature points. The target state vector is then input into the upper-level main filter for Kalman filtering state estimation to obtain the output result of the main filter.

[0093] The arc current data in the multi-sensor synchronization data is compared with the set reference arc current value by performing difference integration to obtain the difference integration sequence. The difference integration sequence is then input into the lower-level sub-filter for lateral deviation filtering to obtain the sub-filter output result.

[0094] Global optimal estimation is performed based on the output results of the main filter and the sub-filter to obtain the three-dimensional position data of the weld.

[0095] Specifically, based on the coordinate transformation matrices of each segment, the observations from the sensor's own coordinate system in the multi-sensor synchronous data are uniformly mapped to the robot's global coordinate system: the point cloud frames continuously output by the laser line scan sensor at a wavelength of 650nm, a scanning frequency of 200Hz, a scanning angle of ±30°, and a measurement accuracy of 0.1mm, together with the feature points extracted by the binocular vision at 30 frames per second, 1920×1080 resolution, 12mm focal length, and 120mm baseline, are transformed frame by frame and point by point according to the coordinate transformation matrix corresponding to their respective segments, resulting in spatiotemporally consistent laser point cloud coordinates and visual feature point coordinates. Adhering to the principle of "equal emphasis on geometric position and motion state," the transformed laser point cloud coordinates and visual feature point coordinates are used to construct a target state vector, simultaneously representing the position of the weld in three-dimensional space and its instantaneous change trend along the welding direction. The target state vector is then input into the upper-level main filter for Kalman filtering estimation. The upper-level main filter complements and fuses the sensitivity of laser to contour sections and height steps with the stable recognition capability of vision for texture boundaries and shape lines. This results in a more robust main filter estimation of weld geometry while suppressing single-source noise and combating local occlusion. Simultaneously, the arc current recorded by the arc sensor at a sampling frequency of 1 kHz within a range of 50–300 A and an accuracy of ±1 A is differentially calculated with the process-defined reference current and integrated over time to form a differential integral sequence reflecting the cumulative effect of lateral drift. This differential integral sequence is input into the lower-level sub-filter, specifically targeting the coupling characteristics of lateral deviation and weld stability, to obtain the sub-filter output. Within the federated fusion framework, the outputs of the main and sub-filters are globally optimized based on information content: when the laser and vision data quality is high, the contribution of the geometric channel is increased to lock the 3D position; when arc stability indicates a weld offset trend, the influence of the sub-filter on lateral correction is moderately amplified. In a typical configuration, the information weights for laser, vision, and arc are 0.5, 0.3, and 0.2, respectively, and are adaptively fine-tuned during operation based on measurement variance. The fusion module runs continuously within the segment at an update rate of no less than 100Hz and shares the state covariance with the error compensation module at the segment boundary to avoid jumps introduced by segment switching, and outputs the three-dimensional position data of the weld.

[0096] In one specific embodiment, the process of performing a global optimal estimation based on the output results of the main filter and the sub-filter to obtain the three-dimensional position data of the weld can specifically include the following steps:

[0097] The weld position and velocity information are extracted from the output of the main filter, and the arc lateral deviation information is extracted from the output of the sub-filter to obtain the double-layer filtered state information.

[0098] Based on the differences in measurement accuracy among laser sensors, vision sensors, and arc sensors, information allocation factors are set for the output results of the main filter and the sub-filters respectively, thus obtaining the sensor weight allocation coefficients;

[0099] The inverse of the covariance matrix of the main filter output and the inverse of the covariance matrix of the sub-filter output are multiplied by the corresponding sensor weight allocation coefficients and then summed with weights to obtain the global information matrix.

[0100] The global covariance matrix is ​​calculated based on the global information matrix, and the optimal state estimate is calculated by weighted fusion of the state information from the two-layer filtering to obtain the three-dimensional position data of the weld.

[0101] Specifically, the two-stage outputs are structured under a unified time reference: the main filter is driven by the laser and vision channels, with inputs consisting of laser point cloud coordinates and visual feature point coordinates unified to the robot's global coordinate system via a piecewise coordinate transformation matrix; the output includes two types of information: weld position and velocity status. The sub-filter is driven by the arc channel, with inputs consisting of the integral sequence of the difference between the arc current and the reference current, sampled at 1kHz, with a range of 50–300A and an accuracy of ±1A; the output is arc lateral deviation information, which is highly sensitive to lateral offset. The "position-velocity" status of the main filter and the "lateral deviation" status of the sub-filter are combined at the same time to form a double-layer filtered state information, and their respective residual statistics and covariance information are retained for downstream fusion. Information allocation factors are set for the two-stage output based on the differences in sensor measurement accuracy: laser is most sensitive to changes in geometric cross-section and steps, vision has stable recognition of texture and contour continuity, and electric arc is extremely sensitive to lateral deviations but is easily affected by instantaneous operating condition fluctuations. Therefore, typical weights are set as 0.5 for laser, 0.3 for vision, and 0.2 for electric arc, with adaptive fine-tuning allowed during operation based on measurement variance. A chi-square threshold test is performed on the residuals of each cycle. If any channel is suspected of being an outlier, its allocation factor is temporarily reduced until the residuals recover to within the statistical confidence interval. After the weights are determined, the fusion module uses information form for numerically robust summation: the inverses of the estimated covariance matrices of the main filter and sub-filter are taken as information matrices, multiplied with the corresponding weight coefficients, and then weighted and summed to obtain the global information matrix. To avoid numerical instability caused by direct inversion, matrix operations preferentially adopt the lower triangular decomposition strategy such as Cholesky, and when necessary, a small diagonal regularization term is used to suppress ill-conditioned cases. At the same time, consistency constraints and minimum eigenvalue truncation are applied to the covariance to ensure that the uncertainty after fusion is physically interpretable. The global covariance matrix is ​​obtained by robustly inverting the global information matrix, and then used together with the state information from the two-layer filtering to calculate the optimal state estimate. The fusion unit operates continuously at a frequency of at least 100Hz, sharing the state and covariance with the error compensation module at segment boundaries. A smooth transition function is used to avoid state jumps introduced by segment switching. When a significant deterioration in arc stability or visual occlusion causing information degradation is detected, the weight of the geometric channel is temporarily increased or the influence of the arc channel is reduced. When a fault self-check is triggered, the remaining channels are used to maintain degenerate operation. Through the above process, the three-dimensional position data of the weld is fused and output.

[0102] In one specific embodiment, the process of performing step 103 may specifically include the following steps:

[0103] The error prediction data for the current segment is calculated at the segment switching point based on the error accumulation model;

[0104] Calculate the Jacobian transformation matrix of each coordinate component in the three-dimensional position data of the weld at the segment switching point based on the error prediction data;

[0105] The error compensation vector is obtained by performing an inverse matrix operation on the Jacobian transformation matrix and multiplying it with the error prediction data. The error compensation vector is then added to the three-dimensional position data of the weld to obtain the target weld position data.

[0106] The target weld location data is smoothed at the segment boundaries to obtain the weld trajectory location.

[0107] Specifically, at the instant the robot transitions from one segment to the next along the weld seam, an error accumulation model based on temperature gradient and segment length is invoked. This model recursively synthesizes the residual deviation at the end of the previous segment with the length-weighted cumulative thermal deformation term of the current segment, and incorporates prior noise from on-site statistics. This ensures that the error prediction reflects both the attenuation effect of geometric stiffness and the deformation increment caused by heat input. The error attenuation coefficient is set to the order of 0.02 per meter, the thermally induced term is obtained by mapping the temperature gradient obtained from infrared thermography to the material's thermal expansion coefficient, and the noise standard deviation is initialized to 0.5 mm and adaptively updated during operation. This forms the error prediction data for the current segment at the segment switching point. Based on the prediction error, the sensitivity of the weld seam's three-dimensional position data is characterized. Specifically, small perturbations are applied to each coordinate component at the segment switching point, and numerical differentiation is used to calculate the sensitivity coefficient of the state to the error in 0.01 mm steps. A local Jacobian transformation matrix is ​​then constructed based on this coefficient. To ensure numerical stability, the differentiation process is performed under a unified time reference. Data-driven amplitude limiting and denoising strategies are used to suppress occasional spikes, resulting in a reliable coordinate-error mapping relationship under the current attitude and operating conditions. The Jacobian matrix serves as a linear mapper from error to coordinate correction. A numerically stable matrix decomposition method is employed to perform inverse operations to solve for the compensation direction and magnitude. Regularized decomposition is used to avoid ill-conditioned amplification, and threshold truncation is applied to the smallest eigenvalue. The obtained error compensation vector is then synthesized by vector addition with the original weld seam 3D position data to obtain the target weld seam position data. During the process, a physical upper limit is set for the compensation amount, and consistency is checked with the covariance of the upper-layer main filter to ensure that the correction does not exceed the limit and does not compromise the reliability of the state. Considering that the segment switching point is a weak link in path continuity, the target weld seam position data is smoothed in the boundary neighborhood to obtain the weld seam trajectory position. The compensation amount is gradually started and stopped by a smooth transition function in the time domain. An S-shaped transition is preferred to reduce abrupt changes in the first and second derivatives. In the spatial domain, several control points on both sides of the boundary are weighted and fused in a windowed manner to make the position, direction and curvature three levels synchronous and continuous, so as to avoid acceleration spikes in the trajectory at the physical execution level.

[0108] In one specific embodiment, the process of performing step 104 may specifically include the following steps:

[0109] The weld trajectory positions are sequentially arranged according to the preset control point spacing as control points of the cubic B-spline curve, and the initial fitting path is obtained by fitting the curve using the cubic B-spline basis function.

[0110] Based on the initial fitted path, constraints are set at the boundaries of each segment and constraint optimization is performed to obtain the path control points;

[0111] Calculate the curvature value of the path control point and compare it with the maximum executable curvature of the welding robot;

[0112] When the curvature value exceeds the maximum executable curvature, the initial fitted path is replanned locally to avoid obstacles and the control points are updated to obtain the target welding path.

[0113] Specifically, uniform sampling and denoising correction are performed on the weld trajectory position sequence. The trajectory points are sequentially arranged according to a preset control point spacing of 10mm. A combination of chord length parameterization and sliding window outlier removal is used to ensure that the control point distribution covers minor geometric fluctuations without introducing unnecessary oscillations. The processed control points are input into a cubic B-spline fitting module. The cubic B-spline basis function is used to generate an initial fitting path with smooth tangential and curvature characteristics. During the generation process, soft constraints are applied to the endpoint velocities and accelerations to ensure compatibility with the welding cycle. Since long-distance welds are managed in segments, to avoid geometric abrupt changes and dynamic peaks at segment boundaries, continuity constraints are introduced at each segment boundary. The constraints cover three levels: positional continuity, directional continuity, and curvature continuity. This is equivalent to simultaneously constraining the continuity of zero-order, first-order, and second-order geometric quantities at the boundaries, thereby fine-tuning the control point positions through constraint optimization. The optimizer employs quadratic programming with a trust region and superimposes a composite objective of smoothness and path length to ensure that the welding torch stroke is minimized and acceleration requirements are reduced while maintaining boundary continuity. After constraint optimization, curvature is calculated for the obtained path control points. Curvature evaluation uses uniform arc-length sampling and differential stabilization strategies to obtain the curvature distribution along the path. The curvature values ​​are then compared point-by-point with the maximum executable curvature limit of the welding robot, which is 0.5m. -1When curvature exceeds the limit or the rate of curvature change is too large at any location, a local replanning mechanism is triggered. Local replanning uses the neighborhood of the current control point as the search domain, with a radius of 50mm. An improved RRT* is used for rapid feasibility searching, with a maximum of 1000 iterations. Simultaneously, process accessibility constraints, nozzle-workpiece safety clearance constraints, and obstacle avoidance constraints near heat sources are applied to the search process. A heuristic cost function comprehensively considers path length, curvature, and rate of curvature change to ensure that a local replacement segment that meets the execution limits is obtained within a 20ms time budget. Continuity constraints on position, direction, and curvature are applied again to the replacement segment and the original path at the splicing boundary. An S-shaped transition is used to update control points and redistribute weights within a small range, avoiding sharp bends in the path at the numerical level. To improve overall stability and executability, the path generation module uses arc length reparameterization throughout the process, allowing subsequent velocity planning to directly complete feedforward velocity and acceleration allocation on equidistant sampling. This, combined with curvature upper limits and welding torch posture restrictions, achieves a "slow-fast-slow" rhythm. In the quality control stage, the length, average curvature and standard deviation of curvature of the initial fitted path and the target path after replanning are calculated and a comprehensive score is formed. When the score is lower than the threshold of 85 points, it automatically reverts to the previous stable solution and expands the local search domain to solve again until the upper limit of the process and the quality threshold are satisfied together, and the target welding path is obtained.

[0114] In one specific embodiment, the multi-sensor dynamic path planning method for performing teachless welding of large iron tower components by a robot further includes the following steps:

[0115] The tracking error vector is calculated based on the target welding path and the current position of the welding robot, and the basic data for sliding mode control is generated based on the tracking error vector.

[0116] The equivalent control term is calculated based on the sliding mode control fundamental data using the system state matrix, control gain matrix, and feedback matrix.

[0117] The adaptive gain coefficient is calculated based on the ratio of the current welding length to the total welding length, and the switching control term is calculated in combination with the sliding surface sign function in the sliding mode control basic data.

[0118] The equivalent control term and the switching control term are added together and then the control continuity is processed at the segmented switching points to generate the trajectory tracking control command for the welding robot.

[0119] Specifically, the target welding path and the current pose of the welding robot are aligned in a unified global coordinate system, and the desired position, desired velocity, and desired attitude are obtained from the path by sampling with equal arc length. Within the current control cycle, the difference between the desired quantity and the measured quantity is calculated to form a tracking error vector including three-dimensional position error and three-axis attitude error. At the same time, the first-order rate of change of the error is filtered by differential filtering to suppress high-frequency noise, and the sliding mode control basic data is obtained. The sliding mode control basic data includes the error and the rate of change of the error, as well as the reachability mask derived from the path curvature and welding gun attitude constraints, the confidence weight derived from the multi-sensor fusion covariance estimation, and the amplitude upper limit used to prevent control saturation. Based on this, the controller calls the system state matrix, control gain matrix, and feedback matrix to solve for the equivalent control term. The design of the equivalent control term follows the principle of "position priority, attitude coordination." The position and attitude channels use diagonal gains to reduce coupling, i.e., the diagonal elements of the control gain matrix are [100,100,100,50,50,50], while the diagonal elements of the sliding surface parameter matrix are [5,5,5,2,2,2]. Fine-tuning is performed within a small range through online identification to adapt to actual working conditions with different thicknesses and joint types. Simultaneously, the state covariance output of the upper main filter is... As a weight injection, the equivalent control term becomes more sensitive to high-confidence state components. An adaptive switching control term related to welding progress is introduced, and the adaptive gain coefficient is calculated based on the ratio of the current welding length to the total welding length. The gain gradually increases from the initial value as the welding progresses. Typical parameters are an initial gain of 10 and an adjustment coefficient of 2, which provides stronger correction capability for cumulative deviations in the later stages of the weld. At the same time, the switching direction is generated based on the sliding surface sign function and a boundary layer is superimposed to reduce chattering. The boundary layer thickness converges exponentially in the time domain, with an initial thickness of 2 mm and a convergence rate coefficient of 0.1 s. -1The ideal sign function is replaced with a saturation function to limit high-frequency energy. The equivalent control term and the switching control term are added together to form the original control vector. The original control vector is subject to dual physical and technological limits (including joint torque, end velocity, heat input rate, and nozzle-workpiece gap, etc.) to ensure execution safety. The root mean square of the tracking error is monitored statistically. When the root mean square exceeds 1 mm, the corresponding gain of the position and attitude channels is increased proportionally to restore control margin. Considering the problem of state transition and command discontinuity at the switching points of segmented paths, the controller introduces a smooth transition mechanism before and after each segment switching: in the time domain, an S-shaped transition function is used to weight and fuse the old and new control vectors. The transition speed parameter is set to 50 to complete the soft switching within a few control cycles. In the spatial domain, the reference state in the switching neighborhood is slightly reparameterized to ensure that the expected sequence of position, velocity, and acceleration maintains a consistent derivative level before and after the switching, avoiding acceleration spikes and sudden changes in heat input in the drive layer. The entire closed loop operates at a 100Hz update rate and maintains tight time synchronization with a 1kHz unified clock, with synchronization errors controlled within ±0.5ms. In each cycle, the controller reads the quality flag from the multi-sensor fusion module. When visual occlusion, laser drift, or arc fluctuations cause a decrease in confidence, the controller reduces the contribution of the corresponding channel to the sliding surface and briefly increases the switching gain by one step to enhance robustness against external disturbances and modeling errors. When a fault self-check is triggered, the controller degenerates into a conservative mode, freezing the adaptive gain upper limit and widening the boundary layer thickness, sacrificing transient performance for stability. Through this process, trajectory tracking control commands rapidly approach within a segment without excessive chattering, transition continuously at segment boundaries without execution spikes, and are governed by both process and dynamic constraints throughout the entire process.

[0120] The above describes the multi-sensor dynamic path planning method for the teach-free welding robot of large iron tower components in the embodiments of the present invention. The following describes the multi-sensor dynamic path planning system for the teach-free welding robot of large iron tower components in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the multi-sensor dynamic path planning system for a teach-free welding robot for large iron tower components in this invention includes:

[0121] The acquisition module 201 is used to acquire synchronous data from multiple sensors, including a laser sensor, a vision sensor, and an arc sensor, and to solve for segmented calibration parameters.

[0122] Calculation module 202 is used to calculate the three-dimensional position data of the weld by combining segmented calibration parameters and multi-sensor synchronous data;

[0123] Error compensation module 203 is used to perform error compensation on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position;

[0124] The curve fitting module 204 is used to perform curve fitting with the weld trajectory position as control point and set constraint conditions at the segment boundary to obtain the target welding path.

[0125] Through the collaborative efforts of the aforementioned components, hardware triggering synchronization of the laser sensor, vision sensor, and arc sensor is achieved using a unified reference clock signal. This solves the fusion error problem caused by inconsistent data acquisition times in traditional multi-sensor systems, ensuring a unified time reference for multi-source sensor data. A segmented calibration strategy is proposed, dividing the long-distance weld into multiple independent calibration segments. The coordinate transformation matrix of each segment is solved through singular value decomposition, effectively addressing the accuracy degradation issue of traditional overall calibration methods in long-distance welding. This achieves precise transformation of the sensor coordinate system and segmented error control. A two-layer architecture is constructed, with an upper-level main filter processing laser and vision data, and a lower-level sub-filter processing arc data. Optimal fusion of multi-sensor information is achieved through a federated filtering structure, fully leveraging the measurement advantages of different sensors and significantly improving the accuracy and robustness of weld position detection. Based on an error accumulation model, forward-looking error prediction and real-time compensation are performed at segment switching points. Precise error correction is achieved through the Jacobian transformation matrix, effectively suppressing cumulative positioning deviations during long-distance welding and ensuring the consistency of accuracy throughout the welding path. A cubic B-spline curve was used to fit the weld trajectory position, and continuity constraints were set at the segment boundaries to generate a smooth and continuous welding path. Combined with curvature constraints and a local replanning mechanism, dynamic optimization and real-time adjustment of the path were achieved, ensuring the stability and continuity of the welding process. A sliding mode controller combining equivalent control terms and switching control terms was designed, adaptively adjusting the control gain according to the welding progress. Continuity processing was applied at segment switching points, enabling the robot to accurately track complex welding paths and improving the control accuracy and system stability of long-distance welding.

[0126] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0127] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0128] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components, characterized in that, include: The system acquires synchronous data from multiple sensors, including laser, vision, and arc sensors, and solves for segmented calibration parameters. Specifically, this includes: using a laser sensor in the welding robot to perform line laser scanning on the weld area of ​​the large iron tower component to obtain laser point cloud data; using a vision sensor in the welding robot to acquire weld images of the weld area and extract visual feature data based on the weld images; using an arc sensor in the welding robot to acquire arc current data during the welding process of the large iron tower component; synchronizing the laser point cloud data, the visual feature data, and the arc current data based on a unified reference clock signal issued by the main controller to obtain multi-sensor synchronized data; and using the laser point cloud data from the multi-sensor synchronized data to obtain multi-sensor synchronized data. Cloud data analysis is used to determine the spatial trajectory of the weld seam and to set calibration point positions along the spatial trajectory, resulting in a segmented calibration point sequence. At each calibration point in the segmented calibration point sequence, homogeneous coordinate data from the laser sensor and feature point coordinate data from the vision sensor are extracted, and the corresponding robot end-effector pose matrix is ​​recorded, yielding coordinate correspondence data at the calibration points. Singular value decomposition is performed on the coordinate correspondence data, and the rotation matrix and translation vector are solved. The rotation matrix and translation vector are used to form the coordinate transformation matrix for each segment. Segmented calibration parameters are generated based on the coordinate transformation matrix of each segment, the length information of the corresponding segment, and the calibration point positions. The three-dimensional position data of the weld is calculated by combining the segmented calibration parameters and the multi-sensor synchronous data. Specifically, this includes: using the coordinate transformation matrix of each segment in the segmented calibration parameters to perform coordinate transformation on the laser point cloud data and visual feature data in the multi-sensor synchronous data to obtain the laser point cloud coordinates and visual feature point coordinates; constructing a target state vector based on the laser point cloud coordinates and the visual feature point coordinates, and inputting the target state vector into the upper-level main filter for Kalman filtering state estimation to obtain the main filter output result; performing difference integration calculation on the arc current data in the multi-sensor synchronous data and the set reference arc current value to obtain the difference integration sequence, and inputting the difference integration sequence into the lower-level sub-filter for lateral deviation filtering to obtain the sub-filter output result; and performing global optimal estimation based on the main filter output result and the sub-filter output result to obtain the three-dimensional position data of the weld. Error compensation is performed on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position. Specifically, this includes: calculating the error prediction data of the current segment at the segment switching point based on the error accumulation model; calculating the Jacobian transformation matrix of each coordinate component in the three-dimensional position data of the weld at the segment switching point based on the error prediction data; performing an inverse matrix operation on the Jacobian transformation matrix and multiplying it with the error prediction data to obtain an error compensation vector; performing a vector addition operation on the error compensation vector and the three-dimensional position data of the weld to obtain the target weld position data; and smoothing the target weld position data at the segment boundary to obtain the weld trajectory position. The target welding path is obtained by using the weld trajectory positions as control points for curve fitting and setting constraints at the segment boundaries. Specifically, this includes: sequentially arranging the weld trajectory positions according to a preset control point spacing as control points for a cubic B-spline curve and fitting the curve using cubic B-spline basis functions to obtain an initial fitted path; setting constraints at each segment boundary based on the initial fitted path and performing constraint optimization to obtain path control points; calculating the curvature value of the path control points and comparing it with the maximum executable curvature of the welding robot; when the curvature value exceeds the maximum executable curvature, performing local obstacle avoidance path replanning on the initial fitted path and updating the control points to obtain the target welding path.

2. The multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components according to claim 1, characterized in that, Also includes: Based on the length information of the corresponding segment in the segment calibration parameters, the temperature gradient data of each segment is monitored by an infrared temperature sensor. Based on the temperature gradient data, the thermal deformation rate of each segment during the welding process is calculated using the material thermal expansion coefficient of the large iron tower components. Multiply the thermal deformation rate of each segment by the length information of the corresponding segment, calculate the thermal deformation accumulation term, and set the error attenuation coefficient to calculate the attenuation and propagation of the error of the previous segment, so as to obtain the error accumulation calculation parameters. Based on the error accumulation calculation parameters, a recursive relationship is established that the current segment error equals the previous segment error multiplied by the attenuation coefficient plus the current segment thermal deformation accumulation term, thus obtaining the error accumulation model.

3. The multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components according to claim 1, characterized in that, The process of performing a global optimal estimation based on the output results of the main filter and the sub-filter to obtain the three-dimensional position data of the weld includes: Weld position and velocity information are extracted from the output of the main filter, and arc lateral deviation information is extracted from the output of the sub-filter to obtain dual-layer filtered state information. Based on the differences in measurement accuracy among the laser sensor, vision sensor, and arc sensor, information allocation factors are set for the output results of the main filter and the sub-filter, respectively, to obtain the sensor weight allocation coefficients; The inverse of the covariance matrix of the main filter output and the inverse of the covariance matrix of the sub-filter output are multiplied by the corresponding sensor weight allocation coefficients and then summed with weights to obtain the global information matrix. The global covariance matrix is ​​calculated based on the global information matrix, and the optimal state estimate is calculated by weighted fusion in combination with the dual-layer filtered state information to obtain the three-dimensional position data of the weld.

4. The multi-sensor dynamic path planning method for a teach-free welding robot for large iron tower components according to claim 1, characterized in that, Also includes: The tracking error vector is calculated based on the target welding path and the current position of the welding robot, and the basic data for sliding mode control is generated based on the tracking error vector. Based on the aforementioned sliding mode control fundamental data, the equivalent control term is calculated using the system state matrix, control gain matrix, and feedback matrix. The adaptive gain coefficient is calculated based on the ratio of the current welding length to the total welding length, and the switching control term is calculated in combination with the sliding surface sign function in the sliding mode control basic data. The equivalent control term and the switching control term are added together and then the control continuity is processed at the segmented switching points to generate the trajectory tracking control command for the welding robot.

5. A multi-sensor dynamic path planning system for a teach-free welding robot for large iron tower components, characterized in that, A multi-sensor dynamic path planning method for performing teachless welding of large iron tower components using a robot as described in any one of claims 1-4, comprising: The data acquisition module is used to acquire synchronous data from multiple sensors, including laser sensors, vision sensors, and arc sensors, and to solve for segmented calibration parameters. The calculation module is used to calculate the three-dimensional position data of the weld by combining the segmented calibration parameters and the multi-sensor synchronous data; An error compensation module is used to perform error compensation on the three-dimensional position data of the weld at the segment switching point to obtain the weld trajectory position; The curve fitting module is used to perform curve fitting using the weld trajectory position as control points and set constraint conditions at the segment boundaries to obtain the target welding path.

Images