A body perception-based welding robot system and intelligent route generation method

Abstract

The application discloses a kind of based on embodiment perception's welding robot system and intelligent route generation method, belong to robot welding technical field, including: embodiment perception module, fusion processing module, intelligent route generation module and execution feedback module, embodiment perception module is through multiple sensing unit Parallel acquisition multimodal data and pre-processing, fusion processing module is combined with three-dimensional attention mechanism and digital twin mapping realizes feature depth fusion, intelligent route generation module is planned-optimization-layered correction output adaptive path, execution feedback module constructs full closed loop dynamic adjustment mechanism.The system and method described above, accurately solve the problem that physical and virtual scene are not synchronized in complex weld welding, dynamic interference is not enough, path adaptability is poor, significantly improve the degree of fit and quality stability of welding path.

Description

Technical Field

[0001] This invention relates to the field of robotic welding technology, and in particular to a welding robot system based on embodied perception and an intelligent route generation method. Background Technology

[0002] In the field of robotic welding, dynamic algorithms have formed a mature application system in core aspects such as multimodal perception fusion, path planning, and feedback control. In complex weld welding scenarios, higher requirements are placed on the adaptability to dynamic disturbances, the synergy of multi-objective optimization, and the integrity of end-to-end feedback. Existing technologies still rely on traditional data fusion logic and local path correction modes.

[0003] However, in complex weld welding scenarios, it is difficult to synchronize the physical welding scenario with the virtual model in real time. When faced with dynamic interference such as temperature fluctuations and bevel deformation, the core features are easily obscured by redundant information during the multimodal data fusion process, resulting in insufficient data representation accuracy. This, in turn, affects the adaptability of path planning to actual working conditions, ultimately causing problems such as poor welding path fit and lack of quality stability. Summary of the Invention

[0004] The purpose of this invention is to provide a welding robot system and intelligent route generation method based on embodied perception, thereby solving the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides a welding robot system based on embodied perception, comprising:

[0006] An embodied perception module for collecting multimodal data of welding targets and robot body state data includes a basic multimodal sensing unit, a bionic compound eye vision module, a supplementary sensing unit, and a data preprocessing submodule. The basic multimodal sensing unit, the bionic compound eye vision module, and the supplementary sensing unit are all connected to the input end of the data preprocessing submodule to perform data preprocessing after parallel data acquisition and output preprocessed standardized data.

[0007] A fusion processing module for feature extraction and deep fusion of standardized data is provided. The input of the fusion processing module is electrically connected to the input of the data preprocessing submodule to receive standardized data. The fusion processing module includes a three-dimensional attention fusion submodule, a digital twin mapping submodule, and a feature fusion submodule connected in sequence to output a comprehensive feature set including weld features, molten pool state, and environmental adaptability.

[0008] An intelligent route generation module for outputting smooth and adaptive welding paths, wherein the intelligent route generation module is electrically connected to the feature fusion submodule to receive integrated feature set data, and the intelligent route generation module includes an initial path planning submodule, a path optimization submodule, and a hierarchical correction submodule connected in sequence.

[0009] The execution feedback module includes an execution unit, a control submodule, a quality monitoring submodule, and a feedback submodule. The control submodule is electrically connected to the execution unit and the quality monitoring submodule, respectively. The quality monitoring submodule is electrically connected to the feedback submodule, respectively. The feedback submodule establishes bidirectional communication with the embodied perception module, the fusion processing module, and the intelligent route generation module, respectively, to realize the feedback and dynamic adjustment of quality monitoring data.

[0010] Preferably, the basic multimodal sensing unit, the bionic compound eye vision module, and the supplementary sensing unit are all mounted on the robot arm, and their acquisition fields of view cover the welding target area. The basic multimodal sensing unit includes a high-resolution industrial camera, an infrared thermal imager, and a force sensor, and the signal output terminals of all three are connected to the data preprocessing submodule through a data interface. The supplementary sensing unit includes a three-axis accelerometer, a three-axis gyroscope, and a current and voltage sensor, and the signal output terminals of all three are electrically connected to the data preprocessing submodule. The three-axis accelerometer and the three-axis gyroscope are integrated into the end of the welding torch to synchronously acquire the acceleration and angular velocity of the welding torch in three axes. The current and voltage sensor is connected in series with the welding circuit to acquire welding current and voltage data during the welding process.

[0011] Preferably, the bionic compound eye vision module incorporates the YOLOv8n-cbam-seg model. This model integrates a convolutional attention module (CBAM) into the neck network of YOLOv8n-seg to output the weld ROI bounding box and mask map, which are synchronously transmitted to the data preprocessing submodule and the digital twin mapping submodule. This allows for the acquisition of 3D point clouds and 2D images of the weld through active-passive visual fusion, and the accurate positioning of the weld region is achieved after model segmentation.

[0012] Preferably, an intelligent route generation method for a welding robot system based on embodied perception includes the following steps:

[0013] S1. Collect multimodal data through the embodied perception module and preprocess it to obtain standardized data;

[0014] S2. Based on the standardized data in step S1, feature extraction and deep fusion are performed through the fusion processing module to obtain a comprehensive feature set that includes weld features, molten pool state and environmental adaptability.

[0015] S3. Based on the comprehensive feature set of step S2, a smooth and adaptive welding route is obtained through progressive processing of initial path planning, path optimization and hierarchical correction.

[0016] S4. The execution unit performs the welding operation according to the welding route in step S3. The quality monitoring submodule monitors the quality in real time and provides feedback for adjustment, so as to obtain a stable welding effect.

[0017] Preferably, step S1 includes the following specific steps:

[0018] S11. Through parallel acquisition of basic multimodal sensing unit, bionic compound eye vision module and supplementary sensing unit, the three-dimensional point cloud of weld, RGB image, temperature distribution image, contact force torque data, welding torch acceleration and angular velocity data, welding current and voltage data and molten pool image are obtained.

[0019] S12. Based on the data collected in step S11, the force data is processed by a low-pass filter, the three-dimensional point cloud is processed by an adaptive Gaussian filter, and the image contrast is enhanced by the CLAHE algorithm to obtain the clean data after denoising.

[0020] S13. Based on the clean data from step S12, perform coordinate transformation using a homogeneous transformation matrix to obtain regularized data in a unified base coordinate system.

[0021] S14. Based on the normalized data from step S13, data fusion is performed using the Kalman filter algorithm to obtain standardized data with unified dimensions and noise reduction. The calculation formula is as follows:

[0022] ;

[0023] ;

[0024] in, For the first State estimate at time; For the first Predict state values ​​at all times; Kalman gain; For the first Time-based observations; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix.

[0025] Preferably, step S2 includes the following specific steps:

[0026] S21. Based on the standardized data from step S14, the enhanced key features are obtained by weighting the features using the channel attention mechanism and spatial attention mechanism of the 3D attention fusion submodule. The formula for the channel attention mechanism is:

[0027] ;

[0028] The formula for spatial attention mechanism is:

[0029] ;

[0030] in, Input feature map; It is the sigmoid activation function; It is a multilayer perceptron; for Convolution operation; For average pooling; For max pooling;

[0031] S22. Based on the enhanced features from step S21 and the 3D point cloud data of the weld seam from step S11, the point cloud is aligned with the image space through the projection mapping model of the digital twin mapping submodule, resulting in a dense 3D digital twin map of the welding area. The formula is as follows:

[0032] ;

[0033] in, Point cloud coordinates; These are the coordinate normalization coefficients; This is the camera intrinsic parameter matrix; This is the extrinsic parameter matrix; These are pixel coordinates;

[0034] S23. Based on the digital twin map from step S22 and the molten pool image data from step S1, the weld centerline features are extracted using the RANSAC algorithm, and the molten pool features are extracted using a convolutional neural network. Then, the attention weights of the feature fusion submodule are calculated to obtain a comprehensive feature set. The calculation formula is as follows:

[0035] ;

[0036] in, The output is a multi-dimensional comprehensive feature set; For visual feature weights; The mapped visual feature vector contains weld and molten pool features; This is the mapped arc feature vector, containing current and voltage data.

[0037] Preferably, the specific steps of step S3 are as follows:

[0038] S31. Based on the weld centerline features of step S23, the welding space is divided by a preset layer height to obtain N layers of discrete mesh nodes, and the endpoint of the upper layer coincides with the starting point of the lower layer.

[0039] S32. Based on the grid node data from step S31, the path evaluation quantification standard is obtained by defining an evaluation function. The evaluation function formula is defined as follows:

[0040] ;

[0041] in, For nodes The total cost of the assessment; These are the welding space mesh nodes currently to be evaluated; From the starting point to the node The actual cost; This is a heuristic function used in path planning to guide the search for the optimal path;

[0042] S33. Based on the evaluation function in step S32, an optimized heuristic function is constructed by incorporating distance, direction, and temperature factors to obtain a path guidance basis that better fits the welding scenario. The formula is as follows:

[0043] ;

[0044] in, , , These are the weighting coefficients; This represents the straight-line distance from the node to the endpoint. This measures the difference between the welding direction and the direction pointing to the endpoint. The temperature weight is assigned to the location of the node. The target node for path planning;

[0045] S34. Based on the heuristic function and grid node data from step S33, the grid nodes are traversed and the parent nodes are backtracked through open and closed lists to obtain the initial path of each layer, and then connected sequentially to obtain the global initial path.

[0046] Preferably, step S3 further includes a path optimization process S35, which specifically includes:

[0047] S351. Based on the global initial path in step S34 and the real-time sensor data in step S11, a reinforcement learning framework is constructed through a deep Q-network. The initial path and real-time sensor data are used as the state space, and the path offset and welding parameter adjustment are used as the action space, thus obtaining the basis for reinforcement learning optimization.

[0048] S352. Based on the reinforcement learning framework of step S351, a multi-objective reward function is designed to obtain the incentive basis for path optimization. The formula is as follows:

[0049] ;

[0050] in, To enhance the immediate reward value for learning; The weight is the path error penalty weight; This represents the deviation between the actual path and the ideal path. Weighting of welding quality rewards; As an indicator for evaluating welding quality; Weighting is assigned to welding efficiency rewards; For welding efficiency;

[0051] S353. Based on the optimized path node data from step S352, the path nodes are smoothed using the B-spline curve formula to obtain a smooth path without inflection points. The formula is:

[0052] ;

[0053] in, Let be the coordinates of any point on the B-spline curve; This represents the number of nodes in the initial path. For the first indivual B-spline basis functions; Let be the degree of the B-spline basis function; For the first The coordinates of each control node;

[0054] S354. Based on the smoothed path data from step S353, the final optimized path is obtained by integrating reinforcement learning optimization and B-spline smoothing results.

[0055] Preferably, step S3 further includes a layer correction process S36, which includes correction of deviations within the same layer and correction of overlaps between different layers. Specific steps include:

[0056] S361, Based on the state estimate from step S14 The welding torch posture data and weld feature data from step S23 are used to construct a weld deviation prediction model, and the current layer height of the welding space layering in step S31 is input. Reference floor height Actual bevel width Expected bevel width The welding torch posture deviation value obtained by KF optimization of the data in step S14. The predicted deviation of the weld was obtained. ;

[0057] S362. Weld prediction deviation based on step S361 The final optimized path data in step S354 is used to obtain the corrected welding path for the same layer by adjusting the path node coordinate offset.

[0058] S363. Based on the corrected path data of the same layer in step S362, the search is expanded outward using the spiral equation of the spiral scan algorithm to obtain candidate overlapping positions for the next layer. The spiral equation is:

[0059] ;

[0060] in, The three-dimensional coordinates of the next layer's overlapping node; Correct the 3D coordinates of the path nodes in the current layer; The radius of the helix; The polar angle of the helix; The pitch of the helix;

[0061] S364. Based on the candidate overlap position data from step S363, adjust... , , The value of is adjusted to ensure that the overlap rate meets the welding quality requirements of 30%~50%, resulting in the corrected layered welding path.

[0062] Preferably, step S4 includes the following specific steps:

[0063] S41. Based on the corrected welding path data from step S364 and the state estimate from step S14. Real-time sensor data is used to execute the welding path via the execution unit. The control submodule adjusts the welding parameters based on a combination of PID control formulas and fuzzy control algorithms to obtain welding current, welding speed, and welding torch posture adapted to the real-time scenario. The calculation formula is as follows:

[0064] ;

[0065] in, This is the output value of the PID controller; For proportional gain; for Real-time quality deviation; For integral gain; For integration variables; This is the differential gain; for Rate of change of time error;

[0066] S42. Based on the welding process data from step S41 and the multimodal raw data from step S11, the quality monitoring submodule performs data fusion and quantitative analysis to obtain welding quality assessment indicators, including weld width. weld depth Average temperature of the molten pool and surface roughness ;

[0067] S43. Based on the quality assessment index data from step S42, the assessment index is compared with the preset standard through the feedback submodule. If... , Deviation from the preset range ±10% or If this is triggered, a second path correction will be achieved, which will output data acquisition parameter adjustment instructions to the embodied perception module, feature weight update instructions to the fusion processing module, and path node correction instructions to the intelligent route generation module, thereby obtaining the closed-loop optimized welding effect.

[0068] Therefore, the beneficial effects of the above-mentioned embodied perception-based welding robot system and intelligent route generation method are as follows:

[0069] 1. By parallel acquisition of basic multimodal sensing units, bionic compound eye vision modules, and supplementary sensing units, combined with low-pass filtering, adaptive Gaussian filtering, CLAHE algorithm for noise reduction and enhancement, homogeneous transformation matrix coordinate unification, and Kalman filter data fusion preprocessing, the limitations of single sensor data are overcome. This enables comprehensive capture of multi-source data such as weld seam 3D point cloud, RGB image, temperature distribution, contact force torque, welding torch posture, and welding electrical signals. It effectively suppresses noise interference and forms standardized data with unified dimensions and purification, providing high-precision and high-reliability data support for subsequent feature extraction, path planning, and parameter adjustment, ensuring the comprehensiveness and accuracy of the entire system's perception layer.

[0070] 2. Based on the channel and spatial attention mechanism of the 3D attention fusion submodule, the projection mapping model of the digital twin mapping submodule, and the attention weight calculation of the feature fusion submodule, a hierarchical and progressive feature processing system was constructed. This system not only strengthens the representation of key features through the attention mechanism, but also achieves spatial alignment between point clouds and images and dense 3D mapping of the welding area with the help of digital twin technology. Then, the RANSAC algorithm and convolutional neural network are used to extract core features and fuse multi-dimensional information such as vision and arc. This efficiently integrates the complementary value of heterogeneous data, suppresses the interference of redundant information, and generates a comprehensive feature set that includes weld features, molten pool state, and environmental adaptability. This significantly improves the accuracy and scene adaptability of feature fusion, and provides a comprehensive and reliable decision basis for intelligent route generation.

[0071] 3. A progressive design of initial path planning, path optimization, and layered correction is adopted. By pre-setting the layer height to divide the grid nodes and incorporating optimization heuristic functions based on distance, direction, and temperature factors, a global initial path is quickly generated. Then, a deep Q-network reinforcement learning framework (with B-spline curve smoothing) is used to achieve intelligent path optimization. Finally, the deviation of the same layer is corrected through a weld deviation prediction model. Combined with a spiral scanning algorithm, the overlap rate between layers is precisely controlled at 30%~50%. This ensures the efficiency and fit of path planning, and eliminates problems such as path inflection points and interlayer overlap defects through optimization and correction. The generated welding route is smooth and adaptive, fully adapting to the complex requirements of welding scenarios and significantly improving the rationality and accuracy of the path.

[0072] 4. Through a closed-loop design of the execution unit, control submodule, quality monitoring submodule, and feedback submodule, the entire process of welding execution, quality monitoring, and dynamic adjustment is linked. The control submodule can dynamically adjust the welding current, speed, and welding torch posture based on real-time data. The quality monitoring submodule quantitatively analyzes core indicators such as weld width, depth, molten pool temperature, and surface roughness. The feedback submodule triggers adjustments to data acquisition parameters, updates feature weights, and performs secondary path corrections based on indicator deviations. This allows for real-time adaptation to changes in the welding scenario, timely correction of deviations, and effective avoidance of problems such as weld size exceeding tolerances and surface roughness exceeding standards. This ensures the stability and consistency of welding results and improves welding quality and production efficiency.

[0073] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0074] Figure 1 A connection block diagram of a welding robot system based on embodied perception provided by the present invention;

[0075] Figure 2 A flowchart of an intelligent route generation method for a welding robot system based on embodied perception, provided by the present invention.

[0076] Figure Labels

[0077] 1. Embodied Perception Module; 11. Basic Multimodal Sensing Unit; 12. Bionic Compound Eye Vision Module; 13. Supplementary Perception Unit; 14. Data Preprocessing Submodule; 2. Fusion Processing Module; 21. 3D Attention Fusion Submodule; 22. Digital Twin Mapping Submodule; 23. Feature Fusion Submodule; 3. Intelligent Route Generation Module; 31. Initial Path Planning Submodule; 32. Path Optimization Submodule; 33. Hierarchical Correction Submodule; 4. Execution Feedback Module; 41. Execution Unit; 42. Control Submodule; 43. Quality Monitoring Submodule; 44. Feedback Submodule. Detailed Implementation

[0078] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0079] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0080] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0081] While existing technologies have employed dynamic techniques for welding robot path planning and welding control to achieve welding path planning and parameter adjustment, significant shortcomings remain: Monocular laser and other sensor solutions have data limitations, making it difficult to support accurate judgment in complex scenarios; multi-sensor fusion, though applied, relies solely on general filtering and attention convolution logic, failing to specifically enhance core features such as the weld centerline and molten pool, and lacks deep integration with digital twin technology, thus failing to effectively address dynamic interference such as temperature fluctuations and bevel deformation; path optimization focuses on a single smoothing objective, failing to achieve multi-objective coordination of quality, efficiency, and energy consumption, and layered correction only addresses local deviations, making it difficult to adapt to the full-process requirements of complex welds; feedback links are limited to parameter adjustment and local path correction, failing to form a closed-loop linkage with feature fusion and global path planning; quality assessment does not fully integrate key dimensions such as molten pool trailing angle and temperature gradient, ultimately leading to poor path fit and parameter adaptability during complex weld welding, making it difficult to guarantee the stability and consistency of welding quality.

[0082] Based on the above analysis, this invention is designed, see appendix. Figure 1-2 A welding robot system based on embodied perception, comprising:

[0083] An embodied perception module 1 is used to collect multimodal data of the welding target and state data of the robot body. The embodied perception module 1 includes a basic multimodal sensing unit 11, a bionic compound eye vision module 12, a supplementary sensing unit 13, and a data preprocessing submodule 14. The basic multimodal sensing unit 11, the bionic compound eye vision module 12, and the supplementary sensing unit 13 are all connected to the input end of the data preprocessing submodule 14 to perform data preprocessing after parallel data acquisition and output preprocessed standardized data.

[0084] A fusion processing module 2 is used for feature extraction and deep fusion of standardized data. The input end of the fusion processing module 2 is electrically connected to the input end of the data preprocessing submodule 14 to receive standardized data. The fusion processing module 2 includes a three-dimensional attention fusion submodule 21, a digital twin mapping submodule 22 and a feature fusion submodule 23 connected in sequence to realize hierarchical extraction and deep fusion of features, and output a comprehensive feature set including weld features, molten pool state and environmental adaptability.

[0085] The intelligent route generation module 3 is used to output a smooth and adaptive welding path. The input end of the intelligent route generation module 3 is electrically connected to the output end of the feature fusion submodule 23 to receive comprehensive feature set data. The intelligent route generation module 3 includes an initial path planning submodule 31, a path optimization submodule 32, and a hierarchical correction submodule 33 connected in sequence to complete the entire process of path planning, optimization, and correction.

[0086] The execution feedback module 4 includes an execution unit 41, a control submodule 42, a quality monitoring submodule 43, and a feedback submodule 44. The control submodule 42 is electrically connected to the execution unit 41 and the quality monitoring submodule 43, respectively. The quality monitoring submodule 43 is electrically connected to the feedback submodule 44, respectively. The feedback submodule 44 establishes bidirectional communication with the embodied perception module 1, the fusion processing module 2, and the intelligent route generation module 3, respectively, to realize the feedback and dynamic adjustment of quality monitoring data.

[0087] Specifically, the present invention is based on a embodied perception welding robot system using a six-axis industrial robot as a carrier. Each sensing unit of the embodied perception module 1 communicates with the robot controller through a standardized interface. The core algorithms of the fusion processing module 2 and the intelligent route generation module 3 are deployed on the edge computing unit. The execution feedback module 4 forms a linkage control with the robot execution unit 41 and the welding machine to ensure that the data transmission delay is ≤20ms, which meets the real-time control requirements.

[0088] In a specific embodiment of the basic multimodal sensing unit 11 in this invention, both the basic multimodal sensing unit 11 and the bionic compound eye vision module 12 are fixed to the end of the robot arm by a customized bracket; the three-axis accelerometer and three-axis gyroscope of the supplementary sensing unit 13 are integrated into the sensor integration base at the end of the welding torch, the current and voltage sensors are connected in series in the output circuit of the welding machine, and the acoustic sensor and humidity sensor are fixed to the side of the welding workbench by a magnetic bracket.

[0089] The high-resolution industrial camera and infrared thermal imager of the basic multimodal sensing unit 11 are connected to the data preprocessing submodule 14 via a USB 3.0 interface, and the force sensor transmits data via an EtherCAT bus; the bionic compound eye vision module 12 is connected to the data preprocessing submodule 14 and the digital twin mapping submodule 22 via a gigabit Ethernet interface; the three-axis accelerometer and three-axis gyroscope of the supplementary sensing unit 13 output data via an SPI interface, and the current and voltage sensors, acoustic sensors, and humidity sensors communicate with the data preprocessing submodule 14 via a CAN bus. All interfaces are equipped with anti-interference shielding layers to adapt to the electromagnetic environment of the welding site.

[0090] Specifically, the high-resolution industrial camera selected is the Basler ace acA2500 - 14uc model, with a resolution of 2592×1944 and a frame rate of 30fps. It is equipped with an 8mm fixed-focus lens, and a welding-specific filter is installed at the front of the lens to acquire RGB images of the weld and molten pool. The image data is transmitted to the data preprocessing submodule 14 in RAW format.

[0091] The infrared thermal imager selected is the FLIR A65 model, with a temperature measurement range of -40℃ to 1500℃, a resolution of 640×512, and a thermal sensitivity of ≤0.03℃. After emissivity calibration, it acquires real-time images of the temperature distribution in the welding area, with the data format being a 16-bit grayscale image.

[0092] The force sensor selected is the ATI Omega160 model, with a range of ±50N and ±5N·m and a sampling rate of 1kHz. It is used to monitor the contact force and torque changes between the welding torch and the workpiece, and output three-dimensional force and torque data.

[0093] The bionic compound eye vision module 12 incorporates the YOLOv8n-cbam-seg model, which is deployed on the edge computing unit. The training dataset contains 5,000 labeled images of different weld types, including straight welds, fillet welds, and curved welds, with ROI bounding boxes and mask images. The training iterations are 500 rounds, and the inference speed is ≥30fps. The model enhances the weld region features by integrating the CBAM attention module into the neck network. The output weld ROI bounding box positioning error is ≤±0.5mm, and the mask image segmentation accuracy is ≥95%. The data is synchronously transmitted to the data preprocessing submodule 14 for coordinate transformation and the digital twin mapping submodule 22 for spatial alignment.

[0094] The three-axis accelerometer and gyroscope are model MPU6050, with an accelerometer range of ±16g and a gyroscope range of ±2000° / s. The sampling rate is 1kHz. They communicate with the data preprocessing submodule 14 through the I2C interface to synchronously collect the acceleration and angular velocity data of the welding torch in the X, Y, and Z axes for calculating the welding torch attitude.

[0095] The current and voltage sensors selected are the LEMLA55-P current sensor and the LV25-P voltage sensor, with a current measurement range of 0~500A and a voltage measurement range of 0~50V, an accuracy of ±0.5%, and a sampling rate of 1kHz. They are connected in series in the circuit between the welding machine and the welding torch to collect real-time data of welding current and voltage.

[0096] The acoustic sensor selected is the PCB Piezotronics 378B02 model, with a frequency response of 20Hz~20kHz and a sensitivity of 10mV / Pa. It is used to monitor the acoustic frequency changes of the welding arc and capture abnormal acoustic signals when the arc is unstable.

[0097] The humidity sensor selected is model SHT30, with a measurement range of 0~100%RH, an accuracy of ±2%RH, and a sampling rate of 1Hz. It is installed 1m above the welding area to collect ambient humidity data in real time, providing an environmental adaptation basis for adjusting welding parameters.

[0098] The data preprocessing submodule 14 uses an STM32H743ZI microcontroller, which integrates a multi-channel data acquisition interface and supports parallel reception of data from various sensing units. The software is based on the FreeRTOS operating system and implements multi-threaded data processing: a low-pass filter is used to denoise the force data, an adaptive Gaussian filter is used to denoise the 3D point cloud, and the CLAHE algorithm is used to enhance the contrast of the image data. All preprocessed data is converted into a unified format and output to the fusion processing module 2 through an Ethernet interface, with a data transmission bandwidth of ≥100Mbps.

[0099] In a specific embodiment of the fusion processing module 2 in this invention, the core algorithm of the fusion processing module 2 is deployed on the edge computing unit. It receives standardized data output by the data preprocessing submodule 14 through the gigabit Ethernet interface. The three submodules transmit data through the internal PCIe 4.0 bus in the serial order of 3D attention fusion submodule 21 → digital twin mapping submodule 22 → feature fusion submodule 23 to ensure the real-time performance of feature processing.

[0100] In a specific embodiment of the intelligent route generation module 3 in this invention, the intelligent route generation module 3 and the fusion processing module 2 are deployed in the same edge computing unit. They receive comprehensive feature set data through internal shared memory. The three sub-modules work in series, and the final output welding path data is transmitted to the robot controller through the EtherCAT bus. The path data format conforms to the robot motion control requirements and includes the angles, motion speeds, and acceleration parameters of each joint.

[0101] In a specific embodiment of the execution feedback module 4 in this invention, the execution unit 41 is a six-axis industrial robot, and the control submodule 42 is integrated into the robot controller and communicates with the execution unit 41 and the quality monitoring submodule 43 via an EtherCAT bus. The sensors of the quality monitoring submodule 43 share some equipment with the embodied perception module 1, and a new surface roughness sensor is installed at the end of the robot arm. The feedback submodule 44 establishes bidirectional communication with the embodied perception module 1, the fusion processing module 2, and the intelligent route generation module 3 via gigabit Ethernet, with a feedback delay of ≤10ms.

[0102] A method for intelligent route generation for a welding robot system based on embodied perception, according to the above embodiments, includes the following steps:

[0103] S1. Collect multimodal data through the embodied perception module 1 and preprocess it to obtain standardized data;

[0104] The specific steps of step S1 include:

[0105] S11. Through parallel acquisition by the basic multimodal sensing unit 11, the bionic compound eye vision module 12 and the supplementary sensing unit 13, the three-dimensional point cloud of the weld, RGB image, temperature distribution image, contact force torque data, welding torch acceleration and angular velocity data, welding current and voltage data and molten pool image are obtained.

[0106] Specifically, a high-resolution industrial camera acquires RGB images of the weld seam, an infrared thermal imager acquires temperature distribution images, and a force sensor acquires contact force and torque data; the bionic compound eye vision module 12: outputs the weld seam ROI bounding box and mask map based on the YOLOv8n-cbam-seg model, and simultaneously acquires the weld seam 3D point cloud and 2D image; a three-axis accelerometer and gyroscope acquire the welding torch's three-dimensional acceleration and angular velocity, a current and voltage sensor acquires the welding current and voltage, an acoustic sensor acquires the arc sound frequency, and a humidity sensor acquires the ambient humidity;

[0107] S12. Based on the data collected in step S11, the force data is processed by a low-pass filter, the three-dimensional point cloud is processed by an adaptive Gaussian filter, and the image contrast is enhanced by the CLAHE algorithm to obtain the clean data after denoising.

[0108] The low-pass filter noise reduction formula is:

[0109] ;

[0110] in, The data after noise reduction; This is the original input data; The filtering time constant is 0.01s. The current moment;

[0111] The adaptive Gaussian filtering denoising formula is:

[0112] ;

[0113] in, For the first The coordinates of each point after filtering; For the first The neighborhood point set of each point; For the first The point and its neighborhood The weight of each point; For the neighboring region The original coordinates of each point;

[0114] S13. Based on the clean data from step S12, perform coordinate transformation using a homogeneous transformation matrix to obtain regularized data in a unified base coordinate system; the coordinate transformation formula is:

[0115] ;

[0116] in, This refers to standardized coordinate data under a unified base coordinate system after transformation. This is the transformation matrix from the base coordinate system to the sensor's original coordinate system; The raw coordinate data collected by the sensor;

[0117] S14. Based on the normalized data from step S13, data fusion is performed using the Kalman filter algorithm to obtain standardized data with unified dimensions and noise reduction. The calculation formula is as follows:

[0118] ;

[0119] ;

[0120] in, For the first State estimate at time; For the first Predict state values ​​at all times; Kalman gain; For the first Time-based observations; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix.

[0121] S2. Based on the standardized data in step S1, feature extraction and deep fusion are performed through the fusion processing module 2 to obtain a comprehensive feature set that includes weld features, molten pool state and environmental adaptability.

[0122] The specific steps of step S2 include:

[0123] S21. Based on the standardized data from step S14, the enhanced key features are obtained by weighting the features using the channel attention mechanism and spatial attention mechanism of the 3D attention fusion submodule 21. The formula for the channel attention mechanism is:

[0124] ;

[0125] The formula for spatial attention mechanism is:

[0126] ;

[0127] in, Input feature map; It is the sigmoid activation function; It is a multilayer perceptron; for Convolution operation; For average pooling; For max pooling;

[0128] ;

[0129] in, This is the enhanced key feature map; Let be the channel attention weight coefficient, and ; Let be the spatial attention weight coefficient, and ; This refers to the number of channels in the feature map.

[0130] S22. Based on the enhanced features from step S21 and the 3D point cloud data of the weld seam from step S11, the point cloud is aligned with the image space through the projection mapping model of the digital twin mapping submodule 22, resulting in a dense 3D digital twin map of the welding area. The formula is as follows:

[0131] ;

[0132] in, Point cloud coordinates; These are the coordinate normalization coefficients; This is the camera intrinsic parameter matrix; This is the extrinsic parameter matrix; These are pixel coordinates;

[0133] S23. Based on the digital twin map from step S22 and the molten pool image data from step S1, the weld centerline features are extracted using the RANSAC algorithm, and the molten pool features are extracted using a convolutional neural network. Then, the attention weights of the feature fusion submodule 23 are used to calculate the comprehensive feature set. The calculation formula is as follows:

[0134] ;

[0135] in, The output is a multi-dimensional comprehensive feature set; For visual feature weights, and , The cosine similarity function; The mapped visual feature vector contains weld and molten pool features; This is the mapped arc feature vector, containing current and voltage data.

[0136] S3. Based on the comprehensive feature set of step S2, a smooth and adaptive welding route is obtained through progressive processing of initial path planning, path optimization and hierarchical correction.

[0137] The specific steps of step S3 are as follows:

[0138] S31. Based on the weld centerline characteristics of step S23, the welding space is divided by a preset layer height. The value range is 0.8~1.2mm, which not only matches the millimeter-level positioning and attitude feedback accuracy of the body perception module 1, but also meets the requirements of high-residue control and interlayer fusion of multiple welding layers in the precision welding industry. Then, N layers of discrete mesh nodes are obtained, each node represents the candidate position of the welding gun, and the end point of the previous layer coincides with the starting point of the next layer.

[0139] S32. Based on the grid node data from step S31, the path evaluation quantification standard is obtained by defining an evaluation function. The evaluation function formula is defined as follows:

[0140] ;

[0141] in, For nodes The total cost of the assessment; These are the welding space mesh nodes currently to be evaluated; From the starting point to the node The actual cost, and , For nodes The parent node; This is a heuristic function used in path planning to guide the search for the optimal path;

[0142] S33. Based on the evaluation function in step S32, an optimized heuristic function is constructed by incorporating distance, direction, and temperature factors to obtain a path guidance basis that better fits the welding scenario. The formula is as follows:

[0143] ;

[0144] in, , , These are the weighting coefficients; This represents the straight-line distance from the node to the endpoint. This measures the difference between the welding direction and the direction pointing to the endpoint. The temperature weight is assigned to the location of the node. The target node for path planning;

[0145] Based on the dynamic characteristics of the welding scenario, namely temperature gradient and weld curvature, the weights are adjusted, and the gradient descent method is used to minimize the path planning cost error. The objective function is:

[0146] ;

[0147] in, For the first The ideal total evaluation cost for each node; The total number of nodes;

[0148] The weight update formula is:

[0149] ;

[0150] ;

[0151] in, For learning rate and ;

[0152] S34. Based on the heuristic function and grid node data from step S33, the grid nodes are traversed and the parent nodes are backtracked through open and closed lists to obtain the initial path of each layer, and then connected sequentially to obtain the global initial path.

[0153] Preferably, step S3 further includes a path optimization process S35, which specifically includes:

[0154] S351. Based on the global initial path in step S34 and the real-time sensor data in step S11, a reinforcement learning framework is constructed through a deep Q-network. The initial path and real-time sensor data are used as the state space, and the path offset and welding parameter adjustment are used as the action space, thus obtaining the basis for reinforcement learning optimization.

[0155] S352. Based on the reinforcement learning framework of step S351, a multi-objective reward function is designed to obtain the incentive basis for path optimization. The formula is as follows:

[0156] ;

[0157] in, To enhance the immediate reward value for learning; The weight is the path error penalty weight; This represents the deviation between the actual path and the ideal path. Weighting of welding quality rewards; As an indicator for evaluating welding quality; Weighting is assigned to welding efficiency rewards; For welding efficiency;

[0158] S353. Based on the optimized path node data from step S352, the path nodes are smoothed using the B-spline curve formula to obtain a smooth path without inflection points. The formula is:

[0159] ;

[0160] in, Let be the coordinates of any point on the B-spline curve; This represents the number of nodes in the initial path. For the first indivual The second-order B-spline basis function is used to control the degree of influence of nodes on the curve; Let be the degree of the B-spline basis function; here, we take it to be 3 to ensure that the curve is second-order continuous and smooth. For the first The coordinates of each control node;

[0161] S354. Based on the smoothed path data from step S353, the final optimized path is obtained by integrating reinforcement learning optimization and B-spline smoothing results.

[0162] Preferably, step S3 further includes a layer correction process S36, which includes correction of deviations within the same layer and correction of overlaps between different layers. Specific steps include:

[0163] S361, Based on the state estimate from step S14 The welding torch posture data and weld feature data from step S23 are used to construct a weld deviation prediction model, and the current layer height of the welding space layering in step S31 is input. Reference floor height Actual bevel width Expected bevel width The welding torch posture deviation value obtained by KF optimization of the data in step S14. The predicted deviation of the weld was obtained. The calculation formula is:

[0164] ;

[0165] in, ; ; ; The current floor height; For reference floor height; This refers to the actual bevel width; The desired bevel width; This refers to the welding torch attitude deviation angle;

[0166] KF optimization, or the Kalman filter algorithm used in step S14, uses the state estimate... Extract welding torch posture data and calculate the deviation;

[0167] S362. Weld prediction deviation based on step S361 The final optimized path data in step S354 is used to obtain the corrected welding path for the same layer by adjusting the path node coordinate offset.

[0168] S363. Based on the corrected path data of the same layer in step S362, the search is expanded outward using the spiral equation of the spiral scan algorithm to obtain candidate overlapping positions for the next layer. The spiral equation is:

[0169] ;

[0170] in, The three-dimensional coordinates of the next layer's overlapping node; Correct the 3D coordinates of the path nodes in the current layer; The radius of the helix; The polar angle of the helix; The pitch of the helix;

[0171] S364. Based on the candidate overlap position data from step S363, adjust... , , The value of is adjusted to ensure that the overlap rate meets the welding quality requirements of 30%~50%, resulting in the corrected layered welding path.

[0172] S4. The welding operation is performed by the execution unit 41 according to the welding route of step S3, and the quality monitoring submodule 43 monitors the quality in real time and provides feedback for adjustment to obtain a stable welding effect.

[0173] The specific steps of step S4 include:

[0174] S41. Based on the corrected welding path data from step S364 and the state estimate from step S14. Real-time sensor data is used to execute the welding path through execution unit 41. Control submodule 42 adjusts welding parameters based on a combination of PID control formula and fuzzy control algorithm to obtain welding current, welding speed, and welding torch posture adapted to the real-time scenario. The calculation formula is as follows:

[0175] ;

[0176] in, This is the output value of the PID controller; For proportional gain; for Real-time quality deviation; To increase integral gain, eliminate static error, and improve control accuracy; The variable is the integral variable, representing the time history; Differential gain is used to suppress overshoot and improve system stability; for The rate of change of error at any given time characterizes the trend of error change.

[0177] S42. Based on the welding process data from step S41 and the multimodal raw data from step S11, the quality monitoring submodule 43 performs data fusion and quantitative analysis to obtain welding quality assessment indicators, including weld width. weld depth Average temperature of the molten pool and surface roughness Calculation formula:

[0178] ;

[0179] ;

[0180] in, This refers to the weld width deviation. For surface roughness deviation; The standard roughness is 3.2. ;

[0181] S43. Based on the quality assessment index data from step S42, the assessment index is compared with the preset standard through the feedback submodule 44. If... , Deviation from the preset range ±10% or If this is triggered, a second path correction will be achieved, which will output data acquisition parameter adjustment instructions to the embodied perception module 1, feature weight update instructions to the fusion processing module 2, and path node correction instructions to the intelligent route generation module 3, thereby obtaining the closed-loop optimized welding effect.

[0182] In a specific embodiment based on the above system and method, this invention addresses the welding scenario of complex curved fillet welds on a Q355 steel structure with a thickness of 8mm and a bevel width of 10±1mm. During the welding process, there are temperature fluctuations of ±6℃ and dynamic interference from 0.4mm bevel deformation. A six-axis industrial robot is used as the carrier. The embodied perception module 1 utilizes a Basler ace acA2500-14uc industrial camera, a FLIR A65 infrared thermal imager, an ATI Omega160 force sensor, and an MPU6050 inertial sensor. The bionic compound eye vision module 12 incorporates a YOLOv8n-cbam-seg model. The fusion processing module 2 and the intelligent route generation module 3 are deployed on an edge computing unit. The execution feedback module 4 employs a PID + fuzzy control strategy with proportional gain... Integral gain Differential gain .

[0183] During welding, the embodied perception module 1 collects multimodal data in parallel, preprocesses it using low-pass filtering, the CLAHE algorithm, and Kalman filtering, and outputs standardized data. The fusion processing module 2 enhances key features through a 3D attention mechanism, combines it with digital twin mapping to generate a 3D dense map, and finally outputs a comprehensive feature set. The intelligent route generation module 3 divides the grid according to a 1.0mm layer height, optimizes it through reinforcement learning and smooths it with B-spline curves, and then achieves a 35%~45% interlayer overlap rate through a spiral scanning algorithm, outputting a corrected path. The execution feedback module 4 welds according to the path, and the quality monitoring submodule 43 monitors it in real time. When the weld width deviates from the preset range, the feedback submodule 44 triggers correction within 10ms, ultimately achieving a weld width deviation ≤ ±0.4mm, a depth deviation ≤ ±0.2mm, and a surface roughness of ≤ ±0.2mm. Welding efficiency was increased by 25%, and the quality pass rate under dynamic interference increased from 85% to 98%.

[0184] In summary, this invention effectively solves the problems of poor welding path fit and quality stability caused by existing technologies in complex weld welding scenarios, such as the inability of existing technologies to cope with dynamic interference, insufficient accuracy of multimodal data fusion, poor adaptability of path planning and optimization, and incomplete full-link feedback. This is achieved through multimodal parallel acquisition and multi-step preprocessing of the embodied perception module 1, hierarchical feature extraction and deep fusion of the fusion processing module 2, the progressive design of initial planning-optimization-hierarchical correction of the intelligent route generation module 3, and the full closed-loop dynamic adjustment of the execution feedback module 4. It not only realizes the comprehensive and accurate capture and efficient integration of welding-related multi-source data, strengthens the characterization of key features, and generates a smooth and adaptive welding route, but also ensures the stability and consistency of welding effect through real-time quality monitoring and dynamic adjustment, significantly improving welding quality and production efficiency.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent route generation in a welding robot system based on embodied perception, characterized in that, Includes the following steps: S1. Collect multimodal data through the embodied perception module and preprocess it to obtain standardized data; S2. Based on the standardized data in step S1, feature extraction and deep fusion are performed through the fusion processing module to obtain a comprehensive feature set that includes weld features, molten pool state and environmental adaptability. S3. Based on the comprehensive feature set of step S2, a smooth and adaptive welding route is obtained through progressive processing of initial path planning, path optimization and hierarchical correction. S4. The execution unit performs the welding operation according to the welding route in step S3, and the quality monitoring submodule monitors the quality in real time and provides feedback for adjustment to obtain a stable welding effect. The specific steps of step S1 include: S11. Through parallel acquisition of the basic multimodal sensing unit, bionic compound eye vision module and supplementary sensing unit of the embodied perception module, the three-dimensional point cloud of the weld, RGB image, temperature distribution image, contact force torque data, welding torch acceleration and angular velocity data, welding current and voltage data and molten pool image are obtained. S12. Based on the data collected in step S11, the force data is processed by a low-pass filter, the three-dimensional point cloud is processed by an adaptive Gaussian filter, and the image contrast is enhanced by the CLAHE algorithm to obtain the clean data after denoising. S13. Based on the clean data from step S12, perform coordinate transformation using a homogeneous transformation matrix to obtain regularized data in a unified base coordinate system. S14. Based on the normalized data from step S13, data fusion is performed using the Kalman filter algorithm to obtain standardized data with unified dimensions and noise reduction. The calculation formula is as follows: ； ； in, For the first State estimate at time; For the first Predict state values ​​at all times; Kalman gain; For the first Time-based observations; The observation matrix; The prediction error covariance matrix; To observe the noise covariance matrix; The specific steps of step S2 include: S21. Based on the standardized data from step S14, the enhanced key features are obtained by weighting the features using the channel attention mechanism and spatial attention mechanism of the 3D attention fusion submodule of the fusion processing module. The formula for the channel attention mechanism is: ； The formula for spatial attention mechanism is: ； in, Input feature map; It is the sigmoid activation function; It is a multilayer perceptron; for Convolution operation; For average pooling; For max pooling; S22. Based on the enhanced features from step S21 and the 3D point cloud data of the weld seam from step S11, the point cloud is aligned with the image space through the projection mapping model of the digital twin mapping submodule of the fusion processing module, resulting in a dense 3D digital twin map of the welding area. The formula is as follows: ； in, Point cloud coordinates; These are the coordinate normalization coefficients; This is the camera intrinsic parameter matrix; This is the extrinsic parameter matrix; These are pixel coordinates; S23. Based on the digital twin map from step S22 and the molten pool image data from step S1, the weld centerline features are extracted using the RANSAC algorithm, and the molten pool features are extracted using a convolutional neural network. Then, the attention weights of the feature fusion submodule are calculated to obtain a comprehensive feature set. The calculation formula is as follows: ； in, The output is a multi-dimensional comprehensive feature set; For visual feature weights; The mapped visual feature vector contains weld and molten pool features; This is the mapped arc feature vector, containing current and voltage data.

2. The intelligent route generation method for a welding robot system based on embodied perception according to claim 1, characterized in that: The basic multimodal sensing unit, the bionic compound eye vision module, and the supplementary sensing unit are all mounted on the robot arm, and their field of view covers the welding target area. The basic multimodal sensing unit includes a high-resolution industrial camera, an infrared thermal imager, and a force sensor, and the signal output terminals of all three are connected to the data preprocessing submodule via a data interface. The supplementary sensing unit includes a three-axis accelerometer, a three-axis gyroscope, and a current and voltage sensor, and the signal output terminals of all three are electrically connected to the data preprocessing submodule. The three-axis accelerometer and the three-axis gyroscope are integrated into the end of the welding torch to synchronously acquire the acceleration and angular velocity of the welding torch in three axes. The current and voltage sensor is connected in series with the welding circuit to acquire welding current and voltage data during the welding process.

3. The intelligent route generation method for a welding robot system based on embodied perception according to claim 2, characterized in that: The biomimetic compound eye vision module incorporates the YOLOv8n-cbam-seg model. This model integrates a convolutional attention module (CBAM) into the neck network of YOLOv8n-seg to output the weld ROI bounding box and mask map, which are synchronously transmitted to the data preprocessing submodule and the digital twin mapping submodule. This allows for the acquisition of 3D point clouds and 2D images of the weld through active-passive visual fusion, and the precise positioning of the weld region is achieved after model segmentation.

4. The intelligent route generation method for a welding robot system based on embodied perception according to claim 1, characterized in that: The specific steps of step S3 are as follows: S31. Based on the weld centerline features of step S23, the welding space is divided by a preset layer height to obtain N layers of discrete mesh nodes, and the endpoint of the upper layer coincides with the starting point of the lower layer. S32. Based on the grid node data from step S31, the path evaluation quantification standard is obtained by defining an evaluation function. The evaluation function formula is defined as follows: ； in, For nodes The total cost of the assessment; These are the welding space mesh nodes currently to be evaluated; From the starting point to the node The actual cost; This is a heuristic function used in path planning to guide the search for the optimal path; S33. Based on the evaluation function in step S32, an optimized heuristic function is constructed by incorporating distance, direction, and temperature factors to obtain a path guidance basis that better fits the welding scenario. The formula is as follows: ； in, , , These are the weighting coefficients; This represents the straight-line distance from the node to the endpoint. This measures the difference between the welding direction and the direction pointing to the endpoint. The temperature weight is the location of the node. The target node for path planning; S34. Based on the heuristic function and grid node data from step S33, the grid nodes are traversed and the parent nodes are backtracked through open and closed lists to obtain the initial path of each layer, and then connected sequentially to obtain the global initial path.

5. The intelligent route generation method for a welding robot system based on embodied perception according to claim 4, characterized in that: Step S3 also includes a path optimization process S35, the specific steps of which include: S351. Based on the global initial path in step S34 and the real-time sensor data in step S11, a reinforcement learning framework is constructed through a deep Q-network. The initial path and real-time sensor data are used as the state space, and the path offset and welding parameter adjustment are used as the action space, thus obtaining the foundation for reinforcement learning optimization. S352. Based on the reinforcement learning framework of step S351, a multi-objective reward function is designed to obtain the incentive basis for path optimization. The formula is as follows: ； in, To enhance the immediate reward value for learning; The weight is the path error penalty weight; This represents the deviation between the actual path and the ideal path. Weighting of welding quality as a reward; As an indicator for evaluating welding quality; Weighting is assigned to welding efficiency rewards; For welding efficiency; S353. Based on the optimized path node data from step S352, the path nodes are smoothed using the B-spline curve formula to obtain a smooth path without inflection points. The formula is: ； in, Let be the coordinates of any point on the B-spline curve; This represents the number of nodes in the initial path. For the first indivual B-spline basis functions; Let be the degree of the B-spline basis function; For the first The coordinates of each control node; S354. Based on the smoothed path data from step S353, the final optimized path is obtained by integrating reinforcement learning optimization and B-spline smoothing results.

6. The intelligent route generation method for a welding robot system based on embodied perception according to claim 5, characterized in that: Step S3 also includes a layered correction process S36, which includes correction of deviations within the same layer and correction of overlaps between different layers. Specific steps include: S361, Based on the state estimate from step S14 The welding torch posture data and weld feature data from step S23 are used to construct a weld deviation prediction model, and the current layer height of the welding space layering in step S31 is input. Reference floor height Actual bevel width Expected bevel width The welding torch posture deviation value obtained by KF optimization of the data in step S14. The predicted deviation of the weld was obtained. ; S362. Weld prediction deviation based on step S361 The final optimized path data in step S354 is used to obtain the corrected welding path for the same layer by adjusting the path node coordinate offset. S363. Based on the corrected path data of the same layer in step S362, the search is expanded outward using the spiral equation of the spiral scan algorithm to obtain candidate overlapping positions for the next layer. The spiral equation is: ； in, The three-dimensional coordinates of the next layer's overlapping node; Correct the 3D coordinates of the path nodes in the current layer; The radius of the helix; The polar angle of the helix; The pitch of the helix; S364. Based on the candidate overlap position data from step S363, adjust... , , The value of is adjusted to ensure that the overlap rate meets the welding quality requirements of 30%~50%, resulting in the corrected layered welding path.

7. The intelligent route generation method for a welding robot system based on embodied perception according to claim 6, characterized in that: The specific steps of step S4 include: S41. Based on the corrected welding path data from step S364 and the state estimate from step S14. Real-time sensor data is used to execute the welding path via the execution unit. The control submodule adjusts the welding parameters based on a combination of PID control formulas and fuzzy control algorithms to obtain welding current, welding speed, and welding torch posture adapted to the real-time scenario. The calculation formula is as follows: ； in, This is the output value of the PID controller; For proportional gain; for Real-time quality deviation; This is the integral gain; For integration variables; This is the differential gain; for Rate of change of time error; S42. Based on the welding process data from step S41 and the multimodal raw data from step S11, the quality monitoring submodule performs data fusion and quantitative analysis to obtain welding quality assessment indicators, including weld width. weld depth Average temperature of the molten pool and surface roughness ; S43. Based on the quality assessment index data from step S42, the assessment index is compared with the preset standard through the feedback submodule. If... , Deviation from the preset range ±10% or If this is triggered, a second path correction will be achieved, which will output data acquisition parameter adjustment instructions to the embodied perception module, feature weight update instructions to the fusion processing module, and path node correction instructions to the intelligent route generation module, thereby obtaining the closed-loop optimized welding effect.

8. The intelligent route generation method for a welding robot system based on embodied perception according to claim 7, characterized in that: Embodied perception welding robot systems based on the above methods include: An embodied perception module for collecting multimodal data of welding targets and robot body state data includes a basic multimodal sensing unit, a bionic compound eye vision module, a supplementary sensing unit, and a data preprocessing submodule. The basic multimodal sensing unit, the bionic compound eye vision module, and the supplementary sensing unit are all connected to the input end of the data preprocessing submodule to perform data preprocessing after parallel data acquisition and output preprocessed standardized data. A fusion processing module for feature extraction and deep fusion of standardized data is provided. The input of the fusion processing module is electrically connected to the input of the data preprocessing submodule to receive standardized data. The fusion processing module includes a three-dimensional attention fusion submodule, a digital twin mapping submodule, and a feature fusion submodule connected in sequence to output a comprehensive feature set including weld features, molten pool state, and environmental adaptability. An intelligent route generation module for outputting smooth and adaptive welding paths, wherein the intelligent route generation module is electrically connected to the feature fusion submodule to receive integrated feature set data, and the intelligent route generation module includes an initial path planning submodule, a path optimization submodule, and a hierarchical correction submodule connected in sequence. The execution feedback module includes an execution unit, a control submodule, a quality monitoring submodule, and a feedback submodule. The control submodule is electrically connected to the execution unit and the quality monitoring submodule, respectively. The quality monitoring submodule is electrically connected to the feedback submodule, respectively. The feedback submodule establishes bidirectional communication with the embodied perception module, the fusion processing module, and the intelligent route generation module, respectively, to realize the feedback and dynamic adjustment of quality monitoring data.

Images