Adaptive intelligent welding robot system and method based on multi-source information fusion

Abstract

This invention discloses an adaptive intelligent welding robot system and method based on multi-source information fusion, relating to the fields of intelligent welding and robot control technology. The system integrates a multi-source heterogeneous sensing module, a spatiotemporal registration and synchronization module, a three-level information fusion module, a digital twin and incremental learning module, a trajectory-parameter collaborative control module, and an execution drive module. It synchronously collects multi-dimensional welding data through multiple sensors, including laser 3D vision. After time synchronization and spatial registration, it completes three-level information fusion using Kalman filtering, hybrid feature selection, and D-S evidence theory. Combined with multiphysics twin simulation and federated incremental learning, a fractional-order PID controller synchronously generates trajectory correction and optimized process parameters. Adaptive welding is then performed by a six-axis robotic arm and other execution components. This invention significantly improves welding accuracy, stability, and adaptability, meeting the high-precision and high-efficiency welding requirements of fields such as bridges and new energy equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent welding and robot control technology, and in particular to an adaptive intelligent welding robot system and method based on multi-source information fusion. Background Technology

[0002] With the deep penetration of intelligent manufacturing technology into the industrial field, welding, as a key processing technology, has been widely used in many industries such as construction, automobiles, and shipbuilding, playing a decisive role in the structural strength and reliability of products. Especially in fields such as large-scale infrastructure and new energy equipment, the requirements for high precision and consistency in welding quality are becoming increasingly stringent. Traditional manual welding and semi-automatic welding equipment can no longer meet the efficiency and quality demands of modern production. Welding robots, with their advantages of automation and precision, have become core equipment for industry upgrading.

[0003] However, existing arc welding robot systems still face significant technical bottlenecks under complex working conditions. On the one hand, the welding process is affected by environmental interference such as strong arc light, spatter, and light fluctuations. Single sensing methods (such as vision or electrical sensing) are prone to data distortion, resulting in large weld positioning errors and high defect detection miss rates. On the other hand, traditional path planning algorithms are mostly based on preset programs, making it difficult to adapt to dynamic changes such as workpiece thermal deformation and inconsistent bevel dimensions in real time. The robotic arm's trajectory is prone to abrupt changes, leading to defects such as incomplete welds and over-welds. At the same time, the information fusion level of existing systems is limited, mostly remaining at the data level or feature level for simple integration. It has not formed a complete closed-loop system of perception-decision-control, and cannot achieve coordinated adaptive adjustment of trajectory and process parameters.

[0004] Furthermore, existing arc welding systems suffer from poor model adaptability and weak quality traceability. Most fusion models and control parameters rely on offline data training, lacking self-evolution capabilities for different materials and welding conditions, requiring frequent manual calibration and adjustment. Moreover, the correlation between multi-source data during welding is not fully explored, making it difficult to deduce process optimization directions through data analysis, resulting in high material waste and manual maintenance costs in production. These problems not only restrict the improvement of welding efficiency but also affect the stability of product quality, becoming key factors hindering the implementation of intelligent manufacturing. Therefore, there is an urgent need to develop an adaptive intelligent welding technology that integrates multi-source sensing, deep fusion, and dynamic optimization. Summary of the Invention

[0005] The purpose of this invention is to propose an adaptive intelligent welding robot system and method based on multi-source information fusion. By constructing a three-level information fusion architecture, integrating digital twin and incremental learning technologies, and designing a fractional-order PID collaborative control strategy, this invention solves the problems of insufficient perception robustness, weak adaptive ability, and lack of self-evolution in existing arc welding robots. It achieves high-precision tracking of welding trajectories under complex working conditions, dynamic optimization of process parameters, and online closed-loop control of welding quality.

[0006] To achieve the above objectives, this invention proposes an adaptive intelligent welding robot system based on multi-source information fusion, comprising: a multi-source heterogeneous perception module, a spatiotemporal registration and synchronization module, a three-level information fusion module, a digital twin and incremental learning module, a trajectory-parameter collaborative control module, and an execution drive module; The multi-source heterogeneous sensing module includes a laser 3D vision sensor, a molten pool infrared thermal imager, an arc electrical acquisition unit, a welding acoustic sensor, an end-effector six-dimensional force sensor, and an inertial measurement unit, used to simultaneously acquire three-dimensional point clouds of weld grooves, molten pool temperature field distribution, arc current / voltage transient waveforms, welding acoustic signals, welding torch-workpiece contact force, and robot end-effector pose data. The spatiotemporal registration and synchronization module achieves nanosecond-level time synchronization of multi-sensor data based on EtherCAT bus hard triggering, and completes spatial registration of visual, force, and pose data through hand-eye calibration matrix and coordinate transformation. The three-level information fusion module includes a data-level fusion unit, a feature-level fusion unit, and a decision-level fusion unit. The data-level fusion unit uses Kalman filtering to reduce noise and complete multi-source data. The feature-level fusion unit uses a hybrid feature selection method based on Fisher filter and packager to select the optimal feature subset from high-dimensional sensor data and construct a welding state feature vector. The decision-level fusion unit uses DS evidence theory and combines the real-time confidence of each sensor mode to output a consistency decision on welding trajectory deviation, penetration state, and process parameter correction. The digital twin and incremental learning module is used to construct a multi-physics twin model of the welding process, predict the effect of process parameter adjustment through online simulation, and realize incremental learning of welding data under multiple working conditions based on the federated learning framework to update the fusion model and control parameters. The trajectory-parameter collaborative control module is based on a fractional-order PID controller, receives the output of the decision-level fusion unit, and synchronously generates the robot's six-degree-of-freedom trajectory correction and welding process parameter optimization values. The execution drive module includes a six-axis robotic arm, a swing actuator, and a digital welding machine, which are used to execute trajectory correction and parameter adjustment commands to complete adaptive welding.

[0007] Preferably, the feature-level fusion unit selects the optimal feature subset from the high-dimensional sensor data to construct a welding state feature vector, and the selection of the optimal feature subset satisfies: ; in, The evaluation score for the optimal feature subset. For the optimal feature subset, F For the subset of features to be selected, For the original high-dimensional feature set, The trace of the matrix, The matrix represents the inter-class divergence. It is a matrix representing the scattering within the class.

[0008] Preferably, the real-time confidence calculation formula for the decision-level fusion unit is as follows: ; in, For the first m Each sensing mode in t Confidence weight at each moment For the first m Real-time data variance for each sensing mode For the first m Reference variance of each sensing mode M For the total number of sensing modes, For the first k Real-time data variance for each sensing mode For the first k Reference variance of each sensing mode m , k This is an index for the sensing mode.

[0009] Preferably, the multiphysics twin model constructed by the digital twin and incremental learning modules has a molten pool temperature field distribution that satisfies the following heat conduction equation: ; in, Density of the parent material For isobaric specific heat capacity, For spatial points r exist t Temperature field at any moment Thermal conductivity at temperature The heat flux density of the welding heat source, For Hamiltonian operators; Welding heat source heat flux density Using a double ellipsoidal heat source model, the formula is as follows: ; in, Q For welding power, For thermal efficiency, The geometric parameters of the double ellipsoidal heat source are... These are the spatial coordinates in the coordinate system of the heat source center.

[0010] Preferably, the trajectory-parameter coordinated control module is based on a fractional-order PID controller, and the control law is as follows: ; in, To control the output, , , These are proportional, integral, and differential gains, respectively. To control error, For fractional integral operators, For fractional differential operators, , For the order of integration, It represents the order of the differential.

[0011] This invention also provides an adaptive intelligent welding robot method based on multi-source information fusion, comprising the following steps: Step S1: The three-dimensional point cloud of the weld groove, the temperature field of the molten pool, the electric arc electrical signal, the welding acoustic text, the end force and pose data are collected synchronously through the multi-source heterogeneous sensing module. The time synchronization and spatial registration are completed by the spatiotemporal registration and synchronization module to obtain multi-source sensing data under a unified spatiotemporal coordinate system. Step S2: Use Kalman filtering to reduce noise and fill in missing values ​​in the registered multi-source data to obtain preprocessed standardized data; Step S3: Based on the hybrid feature selection method, extract groove geometry features, weld pool morphology features, arc stability features, acoustic spectrum features, and force contact features from the standardized data, screen the optimal feature subset, and construct the welding state feature vector. Step S4: Calculate the real-time confidence weight of each sensing mode, use DS evidence theory to fuse the feature vectors, and output the weld trajectory deviation, penetration status label and process parameter correction amount. Step S5: Input the process parameter correction amount into the digital twin model and simulate and predict the welding forming effect; if the effect meets the quality requirements, proceed to step S6; otherwise, re-optimize the process parameters based on the simulation results until the quality requirements are met. Step S6: Input the weld trajectory deviation and optimized process parameters into the fractional-order PID controller to generate robot trajectory correction instructions and welding machine parameter adjustment instructions, and drive the execution module to complete adaptive welding. Step S7: Upload the multi-source data, control parameters, and quality detection results to the federated learning framework to complete the incremental update of the fusion model and the twin model, and realize the self-evolution of the model.

[0012] Preferably, in step S4, the formula for calculating the weld trajectory deviation is as follows: ; in, For weld track deviation, For the first m Trajectory deviation calculated independently for each sensing mode. For the first m Real-time confidence weights for each sensing mode M This represents the total number of sensing modes.

[0013] Preferably, in step S6, the optimized process parameters include welding current. Welding voltage Welding speed Welding torch oscillation amplitude and oscillation frequency The updated formula is as follows: ; in, , , , , These are the initial process parameters. , , , , The parameter correction amount is the output of the decision-level fusion. , , , , These are the weighting coefficients.

[0014] Preferably, in step S7, incremental learning uses the federated averaging algorithm, and the model parameter update formula is as follows: ; in, For the updated model parameters, These are the model parameters before the update. For learning rate, The number of robot nodes participating in federated learning. For the first i The amount of local data on each node. The total amount of data across all nodes. For the first i The gradient of the local loss function of each node.i For node indexing.

[0015] Therefore, this invention proposes an adaptive intelligent welding robot system and method based on multi-source information fusion, the beneficial effects of which are as follows: (1) This invention integrates "electro-optical-thermal-shape-force" multi-source heterogeneous sensing to construct a three-level information fusion architecture. Combined with a confidence-driven dynamic weighted fusion strategy, it effectively solves the limitations of a single sensor under complex working conditions such as strong arc light and thermal deformation. (2) The present invention uses digital twin technology to realize online pre-optimization of process parameters. Combined with the trajectory-parameter collaborative control of fractional PID, it can quickly adapt to real-time working condition changes such as bevel deviation and workpiece thermal deformation, and avoid defects such as weld burn-through and lack of fusion.

[0016] (3) The present invention is based on the incremental learning framework of federated learning. Under the premise of protecting data privacy, it realizes the model update of welding data under multiple working conditions. There is no need to manually recalibrate the parameters, which greatly reduces the operation and maintenance costs and improves the versatility and life cycle of the system.

[0017] 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

[0018] Figure 1 This is an architecture diagram of an adaptive intelligent welding robot system based on multi-source information fusion according to the present invention. Figure 2 This is a flowchart of an adaptive intelligent welding robot method based on multi-source information fusion according to the present invention. Detailed Implementation

[0019] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0021] This embodiment addresses the welding requirements of Φ18mm main bars and Φ8mm stirrups in the context of rebar cage welding. It solves complex problems such as illumination fluctuations (200-1500 lux), rebar cage deformation due to its own weight (maximum deformation of 1.5mm), and bevel gap deviation (±0.8mm) during field construction. It requires weld trajectory tracking accuracy ≤±0.05mm, penetration state prediction accuracy ≥95%, and a single-shift rebar cage production capacity increase of more than 3 times.

[0022] Example 1 like Figure 1 As shown, the present invention provides an adaptive intelligent welding robot system based on multi-source information fusion, including: a multi-source heterogeneous perception module, a spatiotemporal registration and synchronization module, a three-level information fusion module, a digital twin and incremental learning module, a trajectory-parameter collaborative control module, and an execution drive module; The multi-source heterogeneous sensing module includes a laser 3D vision sensor, a molten pool infrared thermal imager, an arc electrical acquisition unit, a welding acoustic sensor, an end-effector six-dimensional force sensor, and an inertial measurement unit. The parameter configurations are as follows: Laser 3D vision sensor: Basler ace 2 series, resolution 1280×1024, point cloud accuracy ±0.02mm, sampling frequency 30Hz, laser wavelength 660nm, effective measurement distance 500-1500mm, used to acquire three-dimensional point cloud and geometric parameters of weld bevel.

[0023] Molten pool infrared thermal imager: FLIR A655sc, temperature measurement range 200-2000℃, accuracy ±5℃, frame rate 100Hz, infrared resolution 640×512, equipped with a high-temperature filter lens to avoid arc light interference, used to capture the temperature field distribution of the molten pool.

[0024] Arc electrical acquisition unit: Customized high-frequency acquisition module with a sampling frequency of 1MHz, current measurement range of 10-500A (accuracy ±0.5%), voltage measurement range of 10-50V (accuracy ±0.3%), supports real-time Ethernet transmission, and records transient waveforms of current and voltage.

[0025] Welding acoustic sensor: A MEMS microphone array with a frequency range of 20Hz-20kHz, sensitivity of -40dB, and signal-to-noise ratio of ≥60dB is selected and installed 15cm from the side of the welding torch to collect welding acoustic signals.

[0026] End-effector six-dimensional force sensor: adopts ATI Mini45, with a range of ±50N, accuracy of ±0.01N, and sampling frequency of 1000Hz. It is rigidly connected to the welding torch through a flange to measure the contact force between the welding torch and the workpiece.

[0027] Inertial Measurement Unit (IMU): BMI088 is selected, with a sampling frequency of 1000Hz, an angle accuracy of ±0.1°, and an acceleration accuracy of ±0.001g. It is integrated into the robot end effector to provide real-time feedback of pose information.

[0028] The spatiotemporal registration and synchronization module achieves nanosecond-level time synchronization of multi-sensor data based on EtherCAT bus hard triggering, as detailed below: Time synchronization: It adopts an EtherCAT bus master station, supports hard-triggered synchronization, and achieves a synchronization accuracy of 100ns. It realizes the alignment of timestamps of multi-sensor data through a distributed clock protocol.

[0029] Spatial registration: A 10×10cm high-contrast checkerboard calibration board was used. The camera intrinsic parameters were obtained through Zhang's calibration method. Combined with multi-pose sampling of the robot (6 different viewpoints), the hand-eye transformation matrix was calculated. The spatial registration error was ≤0.01mm.

[0030] The three-level information fusion module includes a data-level fusion unit, a feature-level fusion unit, and a decision-level fusion unit. Specific parameter settings are as follows: Data-level fusion: After repeated calibration and optimization using offline simulation and field measurement data of the steel cage welding scenario, the Kalman filter parameters were set as follows: process noise covariance matrix Q=diag([0.01,0.01,0.01]), observation noise covariance matrix R=diag([0.005,0.005,0.005]), with ≥20 iterations, achieving data denoising and completion.

[0031] Feature-level fusion: In the hybrid feature selection method, the Fisher filter inter-class / intra-class divergence ratio threshold is set to 3.0, and the wrapper adopts 5-fold cross-validation to select the 18-dimensional optimal feature subset, with a feature dimension compression rate of 77.5%.

[0032] Decision-level fusion: Employing DS evidence theory, for complex working conditions such as rebar cage welding, dynamic confidence weights are first calculated based on the real-time data variance of each sensing modality and the offline calibrated reference variance (laser vision 0.02, infrared thermography 0.025, arc electrical 0.015, acoustic 0.03, force sensing 0.02, IMU 0.018). The feature vectors of six sensing modalities, including laser 3D vision, molten pool infrared thermography, and arc electrical, are treated as independent evidence sources. A fusion judgment rule with a trust function threshold of 0.85 and a conflict coefficient threshold of 0.3 is set. Basic probability allocation is performed on decision propositions such as trajectory deviation, penetration state, and process parameter correction for each evidence source. Information conflicts between different sensing modalities are resolved through evidence combination rules. Finally, consistent decision results for welding trajectory deviation, penetration state labels, and process parameter corrections are output, effectively avoiding decision biases caused by interference from arc light, illuminance fluctuations, and thermal deformation in a single sensing modality.

[0033] The digital twin and incremental learning module is used to construct a multiphysics twin model of the welding process. It predicts the effects of process parameter adjustments through online simulation and uses a federated learning framework to perform incremental learning on welding data under multiple working conditions, updating the fused model and control parameters. The parameter settings for the multiphysics twin model and the federated learning framework are as follows: Multiphysics twin model: Base material parameters (6061 aluminum alloy): density ρ =2700kg / m 3 Specific heat capacity at constant pressure thermal conductivity (At 200℃); parameters of the double ellipsoidal heat source: a 1 = 0.015m, b =0.008m, c =0.005m, thermal efficiency η =0.75.

[0034] Federated Learning Framework: Number of Participating Nodes N =3, learning rate =0.005, local data volume =1.2×10⁵, n 2 = 1.0 × 10⁵ n 3 = 1.1 × 10⁵, total data volume N total =3.3×10⁵, the model update cycle is 24 hours.

[0035] The trajectory-parameter collaborative control module, based on a fractional-order PID controller, receives the output of the decision-level fusion unit and synchronously generates the robot's six-degree-of-freedom trajectory correction and welding process parameter optimization values; wherein, the fractional-order PID controller has a proportional gain... =5.2, integral gain =1.8, differential gain =2.5, Integral order =0.8, differential order =0.2, control cycle is 10ms.

[0036] The execution drive module includes a six-axis robotic arm, a swing actuator, and a digital welding machine, used to execute trajectory correction and parameter adjustment commands to complete adaptive welding. The six-axis robotic arm uses a KUKA KR18, with a repeatability of ±0.03mm, a maximum load of 18kg, and a maximum joint angular velocity of 150° / s. The swing actuator has a swing amplitude of 0-50mm, a frequency of 0-10Hz, a response time of ≤0.1s, and a sine curve swing trajectory. The digital welding machine uses a Megmeet MIG welding machine, with a current adjustment range of 10-500A, a voltage adjustment range of 10-50V, a response time of ≤0.5ms, and supports pulse welding mode.

[0037] Example 2 like Figure 2 As shown, this invention provides an adaptive intelligent welding robot method based on multi-source information fusion, the specific steps of which are as follows: Step S1: The three-dimensional point cloud of the weld groove, the temperature field of the molten pool, the electric arc electrical signal, the welding acoustic text, the end force and pose data are collected synchronously through the multi-source heterogeneous sensing module. The time synchronization and spatial registration are completed by the spatiotemporal registration and synchronization module to obtain multi-source sensing data under a unified spatiotemporal coordinate system. After the system is started, the multi-source heterogeneous sensing modules synchronously collect data: the laser 3D vision sensor acquires the bevel width (10±0.8mm), depth (5±0.5mm), and angle (60±3°); the molten pool infrared thermal imager captures the highest temperature (650±30℃) and temperature gradient (50℃ / mm) of the molten pool; the arc electrical acquisition unit records the current (220±10A) and voltage (28±2V); the acoustic sensor acquires the peak value of the acoustic spectrum (1.2kHz) and energy entropy (0.35); and the force sensor measures the contact force (mean 0.8N, variance 0.05N). 2 The IMU provides feedback on the robot's end-effector pose (position deviation ≤ 0.02 mm, attitude angle deviation ≤ 0.01°). Time synchronization is achieved via the EtherCAT bus, and all data is mapped to the robot's base coordinate system based on the hand-eye transformation matrix, forming a unified spatiotemporal dataset.

[0038] Step S2: Use Kalman filtering to reduce noise and fill in missing values ​​in the registered multi-source data to obtain preprocessed standardized data; Low-pass filtering (1kHz cutoff frequency) was applied to the arc current / voltage data, improving the signal-to-noise ratio by 35%. Missing values ​​in the laser point cloud data were filled to eliminate holes caused by strong arc light (hole area ≤3%). Outlier removal (3σ criterion) was performed on the force and acoustic data. Standardized data was output after preprocessing, including 12 types of basic data such as bevel geometry parameters, molten pool characteristic parameters, and arc stability parameters.

[0039] Step S3: Extract 80-dimensional original features from standardized data, covering geometric features, molten pool features, arc features, acoustic signature features, and force signature features. Using a hybrid feature selection method, first use a Fisher filter to filter out 32-dimensional candidate features, and then use a wrapper to further filter out an 18-dimensional optimal feature subset to construct a welding state feature vector.

[0040] Step S4: Calculate the real-time confidence weights for each sensing mode: Laser vision =0.012, =0.32; Infrared thermal imager =0.015, =0.28; Arc electrical =0.008, =0.40; Acoustics =0.022, =0.05; force sensation =0.014, =0.18; Inertial Measurement Unit (IMU) =0.010, =0.27.

[0041] The weld trajectory deviation is output by fusing feature vectors using DS evidence theory. The penetration status label L(t) = "Normal", and the process parameter correction amount. .

[0042] Step S5: Input the process parameter correction amount into the digital twin model, substitute it into the heat conduction equation and the double ellipsoidal heat source model, and simulate and predict the weld penetration depth of 4.2mm (target 4±0.3mm) and the heat-affected zone width of 3.5mm, which meets the quality requirements. Output the optimized process parameters: I=225A, U=28.8V, v=3.7mm / s, welding torch oscillation amplitude of 8mm, and oscillation frequency of 2Hz.

[0043] Step S6: Input the weld trajectory deviation and optimized process parameters into the fractional-order PID controller to generate the robot's six-axis joint angle correction amount. The robot arm moves along the corrected trajectory, and the welding machine adjusts the current and voltage in real time to achieve precise weld coverage.

[0044] Step S7: Upload the multi-source data, control parameters, and quality inspection results (qualified) of this welding to the federated learning framework, update the model parameters using the federated averaging algorithm, complete the incremental update of the fusion model and the twin model, and realize the self-evolution of the model.

[0045] Compared to traditional semi-automatic arc welding equipment that relies on manual positioning (welding point identification error rate of 15%, average trajectory tracking deviation of 0.12mm, penetration state prediction accuracy of 89.2%, thermal deformation compensation response time of 0.3s, weld pass rate of 85%, single-shift capacity of 2 sets, labor cost accounting for 40%, and material loss rate of 3%-5%), this embodiment reduces the welding point identification error rate to the 0.1mm level, the average trajectory tracking deviation to 0.035mm, the penetration state prediction accuracy to 96.8%, the thermal deformation compensation response time to 0.08s, the weld pass rate to over 95%, the single-shift capacity to 8 sets, the labor cost to 60%, the material loss rate to 0.8%, and the overall production efficiency to 400%, fully meeting the high-precision and high-efficiency welding requirements of rebar cages.

[0046] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0047] Therefore, this invention provides an adaptive intelligent welding robot system and method based on multi-source information fusion. By constructing a three-level multi-source information fusion architecture of data-feature-decision, and integrating multimodal perception, digital twin simulation, fractional-order PID collaborative control and federated incremental learning technology, it effectively overcomes the perception limitations and adaptive shortcomings of traditional welding equipment under complex working conditions, significantly improves the accuracy, stability and adaptability of the welding process, and meets the core requirements of high precision, high efficiency and high consistency of welding processes in fields such as bridges and new energy equipment.

[0048] 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. An adaptive intelligent welding robot system based on multi-source information fusion, characterized in that, include: Multi-source heterogeneous sensing module, spatiotemporal registration and synchronization module, three-level information fusion module, digital twin and incremental learning module, trajectory-parameter collaborative control module and execution drive module; The multi-source heterogeneous sensing module includes a laser 3D vision sensor, a molten pool infrared thermal imager, an arc electrical acquisition unit, a welding acoustic sensor, an end-effector six-dimensional force sensor, and an inertial measurement unit, used to simultaneously acquire three-dimensional point clouds of weld grooves, molten pool temperature field distribution, arc current / voltage transient waveforms, welding acoustic signals, welding torch-workpiece contact force, and robot end-effector pose data. The spatiotemporal registration and synchronization module achieves nanosecond-level time synchronization of multi-sensor data based on EtherCAT bus hard triggering, and completes spatial registration of visual, force, and pose data through hand-eye calibration matrix and coordinate transformation. The three-level information fusion module includes a data-level fusion unit, a feature-level fusion unit, and a decision-level fusion unit; the data-level fusion unit uses Kalman filtering to complete the noise reduction and completion of multi-source data. The feature-level fusion unit uses a hybrid feature selection method combining Fisher filters and wrappers to select the optimal feature subset from high-dimensional sensor data and construct a welding state feature vector. The decision-level fusion unit adopts the DS evidence theory and combines the real-time confidence of each sensing mode to output a consistent decision on welding trajectory deviation, penetration state and process parameter correction. The digital twin and incremental learning module is used to construct a multi-physics twin model of the welding process, predict the effect of process parameter adjustment through online simulation, and realize incremental learning of welding data under multiple working conditions based on the federated learning framework to update the fusion model and control parameters. The trajectory-parameter collaborative control module is based on a fractional-order PID controller, receives the output of the decision-level fusion unit, and synchronously generates the robot's six-degree-of-freedom trajectory correction and welding process parameter optimization values. The execution drive module includes a six-axis robotic arm, a swing actuator, and a digital welding machine, which are used to execute trajectory correction and parameter adjustment commands to complete adaptive welding.

2. The adaptive intelligent welding robot system based on multi-source information fusion according to claim 1, characterized in that, The feature-level fusion unit selects the optimal feature subset from the high-dimensional sensor data to construct a welding state feature vector. The optimal feature subset selection satisfies the following: ； in, The evaluation score for the optimal feature subset. For the optimal feature subset, F For the subset of features to be selected, For the original high-dimensional feature set, The trace of the matrix, The matrix represents the inter-class divergence. It is a matrix representing the intra-class scatter.

3. The adaptive intelligent welding robot system based on multi-source information fusion according to claim 1, characterized in that, The real-time confidence calculation formula for the decision-level fusion unit is as follows: ； in, For the first m Each sensing mode in t Confidence weight at each moment For the first m Real-time data variance for each sensing mode For the first m Reference variance of each sensing mode M For the total number of sensing modes, For the first k Real-time data variance for each sensing mode For the first k Reference variance of each sensing mode m , k This is an index for the sensing mode.

4. The adaptive intelligent welding robot system based on multi-source information fusion according to claim 1, characterized in that, The multiphysics twin model constructed by the digital twin and incremental learning modules satisfies the following heat conduction equation for the molten pool temperature field distribution: ； in, Density of the parent material For isobaric specific heat capacity, For spatial points r exist t Temperature field at any moment Thermal conductivity at temperature The heat flux density of the welding heat source, For Hamiltonian operators; Welding heat source heat flux density Using a double ellipsoidal heat source model, the formula is as follows: ； in, Q For welding power, For thermal efficiency, The geometric parameters of the double ellipsoidal heat source are... These are the spatial coordinates in the coordinate system of the heat source center.

5. The adaptive intelligent welding robot system based on multi-source information fusion according to claim 1, characterized in that, The trajectory-parameter coordinated control module is based on a fractional-order PID controller, and the control law is as follows: ； in, To control the output, , , These are proportional, integral, and differential gains, respectively. To control error, For fractional integral operators, For fractional differential operators, , For the order of integration, It represents the order of the differential.

6. An adaptive intelligent welding robot method based on multi-source information fusion, applied to the adaptive intelligent welding robot system based on multi-source information fusion as described in any one of claims 1-5, characterized in that, Includes the following steps: Step S1: The three-dimensional point cloud of the weld groove, the temperature field of the molten pool, the electric arc electrical signal, the welding acoustic text, the end force and pose data are collected synchronously through the multi-source heterogeneous sensing module. The time synchronization and spatial registration are completed by the spatiotemporal registration and synchronization module to obtain multi-source sensing data under a unified spatiotemporal coordinate system. Step S2: Use Kalman filtering to reduce noise and fill in missing values ​​in the registered multi-source data to obtain preprocessed standardized data; Step S3: Based on the hybrid feature selection method, extract groove geometry features, weld pool morphology features, arc stability features, acoustic spectrum features, and force contact features from the standardized data, screen the optimal feature subset, and construct the welding state feature vector. Step S4: Calculate the real-time confidence weight of each sensing mode, use DS evidence theory to fuse the feature vectors, and output the weld trajectory deviation, penetration status label and process parameter correction amount. Step S5: Input the process parameter correction amount into the digital twin model to simulate and predict the welding forming effect; If the effect meets the quality requirements, proceed to step S6; otherwise, re-optimize the process parameters based on the simulation results until the quality requirements are met. Step S6: Input the weld trajectory deviation and optimized process parameters into the fractional-order PID controller to generate robot trajectory correction instructions and welding machine parameter adjustment instructions, and drive the execution module to complete adaptive welding. Step S7: Upload the multi-source data, control parameters, and quality detection results to the federated learning framework to complete the incremental update of the fusion model and the twin model, and realize the self-evolution of the model.

7. The adaptive intelligent welding robot method based on multi-source information fusion according to claim 6, characterized in that, In step S4, the formula for calculating the weld trajectory deviation is as follows: ； in, For weld track deviation, For the first m Trajectory deviation calculated independently for each sensing mode. For the first m Real-time confidence weights for each sensing mode M This represents the total number of sensing modes.

8. The adaptive intelligent welding robot method based on multi-source information fusion according to claim 6, characterized in that, In step S6, the optimized process parameters include welding current. Welding voltage Welding speed Welding torch oscillation amplitude and oscillation frequency The updated formula is as follows: ； in, , , , , These are the initial process parameters. , , , , The parameter correction amount is the output of the decision-level fusion. , , , , These are the weighting coefficients.

9. The adaptive intelligent welding robot method based on multi-source information fusion according to claim 6, characterized in that, In step S7, incremental learning uses the federated averaging algorithm, and the model parameter update formula is as follows: ； in, For the updated model parameters, These are the model parameters before the update. For learning rate, The number of robot nodes participating in federated learning. For the first i The amount of local data on each node. The total amount of data across all nodes. For the first i The gradient of the local loss function of each node. i For node indexing.

Images