High-precision intelligent welding system adapting to complex working conditions

Abstract

The invention provides a high-precision intelligent welding system adapting to complex working conditions, and relates to the technical field of laser welding. The high-precision intelligent welding system adapting to the complex working conditions comprises the steps of S1, based on a high-resolution visual sensor and a laser scanner, S2, based on preprocessed point cloud data, S3, based on a workpiece three-dimensional model, S4, based on an optimized welding path and a material attribute database, S5, based on a dynamic welding parameter set, S6, based on a real-time welding track, and S7, based on multi-sensor monitoring data. S8, based on a defect detection report; and S9, based on a repaired welding result and historical welding data. According to the invention, the geometric data of the workpiece are collected through the high-resolution visual sensor and the laser scanner, the filtering algorithm is applied to remove noise to generate high-quality point cloud data, and a subsequent processing foundation is laid; workpiece features are identified by using a point cloud processing algorithm and a machine learning method based on the point cloud data, and an accurate digital model is generated through a three-dimensional reconstruction algorithm.

Description

Technical Field

[0001] This invention relates to the field of laser welding technology, specifically a high-precision intelligent welding system adapted to complex working conditions. Background Technology

[0002] Laser welding technology is an advanced manufacturing technology that uses a high-energy-density laser beam as a heat source to precisely melt and join metallic or non-metallic materials. Its core lies in using an optical system to focus the laser to generate extremely high temperatures, causing the material to melt locally and instantaneously, forming a strong weld. This technology features high precision, high efficiency, low thermal deformation, non-contact processing, and strong automation compatibility, and is widely used in industries such as automotive manufacturing, aerospace, electronics, and medical devices. Among these applications, a high-precision intelligent welding system adapted to complex working conditions refers to intelligent equipment that integrates traditional welding processes with laser welding technology. It achieves autonomous decision-making and optimization of the welding process through the fusion of sensors, algorithm control, and actuators. Its purpose is to improve the consistency of welding quality and meet the high-precision processing requirements under complex working conditions.

[0003] Current laser welding technology relies on manual experience to set welding parameters in practice. The parameter selection lacks precision and is difficult to dynamically adapt to changes in material properties or environment, resulting in unstable and inconsistent welding quality. Monitoring mechanisms mostly use single sensors or offline detection methods, which cannot achieve real-time comprehensive status tracking. Defect identification lags, increasing repair costs and time delays. Path planning is based on pre-programmed or simple algorithms, which do not fully optimize obstacle avoidance and heat accumulation, resulting in low efficiency and potential operational risks. Execution control is relatively fixed and lacks dynamic adjustment capabilities. When dealing with complex workpiece shapes, the trajectory accuracy is insufficient and prone to deviation. Defect repair mainly relies on manual intervention, which has a slow response speed and low operational consistency, and may introduce secondary defects that affect the overall quality. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a high-precision intelligent welding system that adapts to complex working conditions. It solves the problem that existing laser welding technologies rely on manual experience to set welding parameters in actual operation, resulting in a lack of precision in parameter selection, difficulty in dynamically adapting to changes in material properties or environment, and thus unstable and inconsistent welding quality.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a high-precision intelligent welding system adaptable to complex working conditions, comprising the following steps: S1: Based on a high-resolution vision sensor and laser scanner, the workpiece geometric data is acquired using 3D scanning technology. Noise and outliers are removed by filtering algorithms, and data preprocessing is performed to generate preprocessed point cloud data. S2: Based on preprocessed point cloud data, point cloud processing algorithms and machine learning methods are used to identify workpiece features, and a precise digital model is generated through a 3D reconstruction algorithm to generate a 3D model of the workpiece. S3: Based on the 3D model of the workpiece, the welding path is autonomously planned using a reinforcement learning algorithm, taking into account obstacle avoidance, heat accumulation minimization and efficiency optimization, and the feasibility of the path is verified through a simulation environment to generate an optimized welding path; S4: Based on the optimized welding path and material property database, a deep learning algorithm is used to predict and optimize laser welding parameters in real time, and the network weights are adjusted through the backpropagation algorithm to generate a dynamic welding parameter set; S5: Based on a dynamic welding parameter set, it uses a PID control algorithm and robot motion control to execute the laser welding process. It fine-tunes the robot arm trajectory through real-time feedback to ensure accurate path following and generate a real-time welding trajectory. S6: Based on real-time welding trajectory, it integrates multiple sensors, uses data fusion algorithm to synchronously monitor welding status, and performs real-time data alignment and noise reduction to generate multi-sensor monitoring data; S7: Based on multi-sensor monitoring data, machine learning algorithms are used to detect welding defects in real time, and image and signal data are analyzed through feature extraction methods to generate a defect probability map and obtain a defect detection report. S8: Based on the defect detection report, the reinforcement learning algorithm is used to optimize the repair strategy, control the auxiliary laser head to perform local remelting or repair welding, and adjust the repair parameters through real-time control methods to minimize the heat-affected zone and generate the repaired welding result; S9: Based on the repaired welding results and historical welding data, a statistical analysis algorithm is used to evaluate the overall welding quality, and the system model parameters are updated through feedback loops for future welding process optimization, generating a final quality assessment report.

[0006] Preferably, the generation of preprocessed point cloud data based on S1 includes the following steps: S101: Based on a high-resolution laser scanner, the workpiece surface is scanned using triangulation to collect raw three-dimensional coordinate data and generate raw point cloud data; S102: Based on the original point cloud data, a statistical outlier removal algorithm is used to filter out noise points, and a voxel grid downsampling method is used to reduce data density and improve the efficiency of subsequent processing to generate denoised point cloud data. S103: Based on denoised point cloud data, Gaussian filtering algorithm is used for smoothing, and the surface is reconstructed by moving least squares method to enhance data quality and generate preprocessed point cloud data.

[0007] Preferably, generating a 3D model of the workpiece based on S2 includes the following steps: S201: Based on preprocessed point cloud data, a density-based clustering method is used to segment the different components of the workpiece and generate segmented point cloud clusters. S202: Based on the segmented point cloud clusters, principal component analysis is used to calculate the normal vector and curvature features of each cluster, identify the geometric features of planes and cylinders, and generate geometric feature descriptors; S203: Based on geometric feature descriptors, a region growing algorithm is used to reconstruct the surface, and a closed triangular mesh model is generated through the Poisson reconstruction algorithm to generate the initial triangular mesh model; S204: Based on the initial triangular mesh model, the mesh quality is optimized using the Laplace smoothing algorithm, and the iterative nearest point algorithm is used to register with the standard CAD model to correct deviations and generate a 3D model of the workpiece.

[0008] Preferably, generating an optimized welding path based on S3 includes the following steps: S301: Based on the 3D model of the workpiece, the A* search algorithm is used to generate an initial sequence of welding path points on the model surface to ensure coverage of all areas to be welded. S302: Based on the initial path point sequence, a deep Q-network is trained in a simulation environment. The path point sequence and robot posture are optimized through a reward function to avoid collisions and generate an optimized welding path.

[0009] Preferably, generating a dynamic welding parameter set based on S4 includes the following steps: S401: Based on the optimized welding path, extract the curvature and distance features of each point on the path from the previous welding point to generate a path geometric feature vector; S402: Based on the path geometric feature vector and material database, a convolutional neural network is used to predict the basic welding parameters of each path point and generate a set of basic welding parameters. S403: Based on the basic welding parameter set, a genetic algorithm is used for multi-objective optimization to find the Pareto optimal solution set and generate a dynamic welding parameter set.

[0010] Preferably, the generation of real-time welding trajectories based on S5 includes the following steps: S501: Based on the dynamic welding parameter set, the cubic spline interpolation algorithm is used to convert discrete path points into smooth and continuous robot joint space trajectories, generating joint space trajectory commands. S502: Based on joint space trajectory commands, a six-axis industrial robot is driven by feedforward control and PID control algorithms, and the trajectory is tracked in real time through encoder feedback to generate theoretical trajectory feedback data. S503: Based on theoretical trajectory feedback data and real-time welding trajectory, the Kalman filter algorithm is used to fuse the data, estimate the optimal trajectory state, and fine-tune the control commands to generate the real-time welding trajectory.

[0011] Preferably, generating multi-sensor monitoring data based on S6 includes the following steps: S601: During the welding process, based on the real-time welding trajectory, the vision camera is synchronously triggered to acquire the molten pool image, and the median filtering algorithm is used for image preprocessing to generate molten pool image data. S602: Simultaneously, an infrared thermal imager is used to collect temperature field data of the molten pool and heat-affected zone, and a non-uniformity correction algorithm is used to process the raw thermal data to generate temperature field data; S603: The acoustic emission sensor is used to collect the acoustic signal during the welding process, and the wavelet transform algorithm is used to perform time-frequency analysis, extract the characteristic frequency, and generate acoustic emission signal characteristics. S604: Based on molten pool image data, temperature field data, and acoustic emission signal characteristics, an extended Kalman filter algorithm is used for time synchronization and spatial registration, which are then fused into a unified multi-dimensional data stream to generate multi-sensor monitoring data.

[0012] Preferably, generating a defect detection report based on S7 includes the following steps: S701: Based on multi-sensor monitoring data, texture features are extracted from image data using the gray-level co-occurrence matrix method, gradient features are extracted from temperature data, and amplitude features are extracted from sound signals to generate multimodal feature vectors. S702: Based on multimodal feature vectors, a support vector machine classifier is used for real-time reasoning to determine whether there are defects in the current solder joint and the type of defects, and to generate a real-time defect probability distribution; S703: Based on real-time defect probability distribution, it uses time series analysis to track the generation and development of defects, and generates a structured report containing defect location, type and confidence level through a rule reasoning engine, thus generating a defect detection report.

[0013] Preferably, the generated repaired welding result based on S8 includes the following steps: S801: Based on the defect detection report, a path replanning algorithm is used to generate a local compensation welding path that covers the defect area at the defect location, thus generating a local repair path; S802: Based on local repair paths and defect types, the Q-learning algorithm is used to query the optimal repair parameters from historical data and generate a repair parameter strategy; S803: Based on the repair parameter strategy, control the robot and laser to perform repair welding, and use the monitoring process of S6 to monitor the repair process in real time and generate repair process monitoring data; S804: Based on the monitoring data of the repair process, support vector machine is used again to detect defects at the repair points, verify the repair effect, and generate the welding results after repair.

[0014] Preferably, generating the final quality assessment report based on S9 includes the following steps: S901: Based on the welding results after repair and the defect detection report, the key quality indicators are calculated using statistical process control methods to generate a quality indicator dataset. S902: Based on the quality index dataset and dynamic welding parameter set, a mapping model between process parameters and quality indicators is established using multiple linear regression analysis. The significance of parameter influence is analyzed, and a process-quality correlation model is generated. S903: Based on the process-quality correlation model, the gradient descent algorithm is used to back-optimize the weight parameters of the convolutional neural network in S4, so as to realize the system's self-iterative improvement and generate model optimization parameters; S904: Integrates the quality indicator dataset, process-quality correlation model, and all metadata of this welding task, and automatically generates a structured evaluation report using natural language generation technology, thus generating the final quality evaluation report.

[0015] This invention provides a high-precision intelligent welding system adapted to complex working conditions. It has the following beneficial effects: This invention acquires workpiece geometric data using a high-resolution vision sensor and laser scanner, applies filtering algorithms to remove noise and generate high-quality point cloud data, laying the foundation for subsequent processing; based on the point cloud data, it uses point cloud processing algorithms and machine learning methods to identify workpiece features, and generates an accurate digital model through a 3D reconstruction algorithm, achieving automatic feature extraction and modeling; it employs reinforcement learning algorithms to autonomously plan welding paths, considering obstacle avoidance and minimizing heat accumulation to optimize path efficiency, and verifies its feasibility through simulation; it combines a material property database with deep learning algorithms to predict and optimize welding parameters in real time, and achieves dynamic parameter adaptation by adjusting network weights through backpropagation; during execution, it uses PID control algorithms and robot motion control, and fine-tunes the trajectory through real-time feedback. To ensure precise path following and achieve high-precision welding, the system integrates multiple sensors and data fusion algorithms to synchronously monitor the welding status, providing comprehensive monitoring data through data alignment and noise reduction. Based on the monitoring data, machine learning algorithms are applied to detect defects in real time, and defect reports are generated through feature extraction and analysis for rapid identification. Reinforcement learning algorithms are used to optimize repair strategies, controlling the auxiliary laser head to perform local remelting or repair welding, and adjusting repair parameters to minimize the heat-affected zone for automatic repair. Based on the repair results and historical data, statistical analysis algorithms are used to evaluate quality, and the system continuously optimizes by updating model parameters through feedback loops. The overall process improves welding accuracy, efficiency, and quality consistency, reduces manual intervention, enhances adaptability to complex working conditions, and achieves fully automated intelligent welding. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the main steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a detailed schematic diagram of S6 of the present invention; Figure 8 This is a detailed schematic diagram of S7 of the present invention; Figure 9 This is a detailed schematic diagram of S8 of the present invention; Figure 10 This is a detailed schematic diagram of S9 of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0018] like Figure 1-10 As shown, this embodiment of the invention provides a high-precision intelligent welding system adapted to complex working conditions, comprising the following steps: S1: Based on a high-resolution vision sensor and laser scanner, the workpiece geometric data is acquired using 3D scanning technology. Noise and outliers are removed by filtering algorithms, and data preprocessing is performed to generate preprocessed point cloud data. S2: Based on preprocessed point cloud data, point cloud processing algorithms and machine learning methods are used to identify workpiece features, and a precise digital model is generated through a 3D reconstruction algorithm to generate a 3D model of the workpiece. S3: Based on the 3D model of the workpiece, the welding path is autonomously planned using a reinforcement learning algorithm, taking into account obstacle avoidance, heat accumulation minimization and efficiency optimization, and the feasibility of the path is verified through a simulation environment to generate an optimized welding path; S4: Based on the optimized welding path and material property database, a deep learning algorithm is used to predict and optimize laser welding parameters in real time, and the network weights are adjusted through the backpropagation algorithm to generate a dynamic welding parameter set; S5: Based on a dynamic welding parameter set, it uses a PID control algorithm and robot motion control to execute the laser welding process. It fine-tunes the robot arm trajectory through real-time feedback to ensure accurate path following and generate a real-time welding trajectory. S6: Based on real-time welding trajectory, it integrates multiple sensors, uses data fusion algorithm to synchronously monitor welding status, and performs real-time data alignment and noise reduction to generate multi-sensor monitoring data; S7: Based on multi-sensor monitoring data, machine learning algorithms are used to detect welding defects in real time, and image and signal data are analyzed through feature extraction methods to generate a defect probability map and obtain a defect detection report. S8: Based on the defect detection report, the reinforcement learning algorithm is used to optimize the repair strategy, control the auxiliary laser head to perform local remelting or repair welding, and adjust the repair parameters through real-time control methods to minimize the heat-affected zone and generate the repaired welding result; S9: Based on the repaired welding results and historical welding data, a statistical analysis algorithm is used to evaluate the overall welding quality, and the system model parameters are updated through feedback loops for future welding process optimization, generating a final quality assessment report.

[0019] The generation of preprocessed point cloud data based on S1 includes the following steps: S101: Based on a high-resolution laser scanner, the workpiece surface is scanned using triangulation to collect raw three-dimensional coordinate data and generate raw point cloud data; Based on a high-resolution laser scanner, triangulation is used to scan the workpiece surface. This process involves a laser emitter projecting a beam of light onto the workpiece surface. For example, in scanning a car door panel, the laser moves with a resolution of 0.1 mm. A sensor receives the reflected light, and the distance is calculated by measuring the difference between the incident angle and the reflection angle of the beam, using a formula... Where c is the speed of light, taken as 3e8m / s. Assuming a time difference of 1e-9 seconds between measurement points, the distance d is calculated to be 0.15 meters. The corresponding coordinates x and y are recorded by the scanner's position encoder, and z is converted from the distance d. The scanning area is 10cm x 10cm, generating approximately 10,000 points. The coordinates of each point are... Stored as an array, for example, points 50.0, 30.0, 0.15 This indicates that at a certain location on the panel, data is accumulated through continuous scanning, raw 3D coordinate data is collected, and raw point cloud data is generated.

[0020] S102: Based on the original point cloud data, a statistical outlier removal algorithm is used to filter out noise points, and a voxel grid downsampling method is used to reduce data density and improve the efficiency of subsequent processing to generate denoised point cloud data. Based on the original point cloud data, a statistical outlier removal algorithm is used to filter out noisy points. This algorithm calculates the average distance between each point and its neighbors, and sets a distance threshold based on the standard deviation. For example, taking 100 point samples from the point cloud, the calculated average distance μ is 0.12 mm, the standard deviation σ is 0.03 mm, and the threshold is set to... That is, 0.18 mm. Points with a distance greater than 0.18 mm are considered noise and should be removed. 50.1, 30.2, 0.25 Points within 0.22 mm of a neighboring point are filtered out as they exceed the threshold. Subsequently, a voxel grid downsampling method is used to reduce data density, dividing the point cloud into voxel cubes with a side length of 0.5 mm. The centroid of each point within a voxel is used as the representative point. (The voxel contains points...) 50.0, 30.0, 0.15) 50.1, 30.1, 0.16 The centroid is calculated as follows 50.05, 30.05, 0.155 The number of points is reduced from 10,000 to 5,000, improving subsequent processing efficiency and generating noise-reduced point cloud data.

[0021] S103: Based on denoised point cloud data, Gaussian filtering algorithm is used for smoothing, and the surface is reconstructed by moving least squares method to enhance data quality and generate preprocessed point cloud data.

[0022] Based on the denoised point cloud data, a Gaussian filtering algorithm is used for smoothing. This algorithm applies a Gaussian kernel function to weighted average the point locations, with the kernel size set to a 3x3 grid and the standard deviation σ set to 0.5 mm. 50.05, 30.05, 0.155 Calculate the weights of neighboring points ,in The distance between points, such as millimeter The new position is calculated as a weighted average; for example, the z-coordinate of a point is adjusted to a new value of 0.154. The surface is then reconstructed using the moving least squares method. This method fits a polynomial surface to the neighborhood of each point, with a neighborhood radius of 1 mm. A quadratic polynomial is used for fitting, and the coefficients are solved using least squares. (The fitted surface is shown in the image.) The coefficients a to f are obtained by solving a system of linear equations. The surface shape is updated based on point data to enhance data quality and generate preprocessed point cloud data.

[0023] The steps for generating a 3D model of a workpiece based on S2 are as follows: S201: Based on preprocessed point cloud data, a density-based clustering method is used to segment the different components of the workpiece and generate segmented point cloud clusters. Based on preprocessed point cloud data, a density-based clustering method is used to segment the different components of the workpiece. This method groups points by calculating the distance density between each point and its neighbors, setting the neighborhood radius ε to 2 mm and the minimum number of points MinPts to 15. For a point p(50.05, 30.05, 0.155) in the point cloud... Calculate its Euclidean distance to surrounding points, such as the distance to point q(50.10, 30.10, 0.160). If the distance is 0.071 mm less than ε, then q belongs to the neighborhood of p. If the number of points in the neighborhood of p reaches MinPts, then a new cluster is created. All points are traversed, and points with connected densities are grouped into the same cluster. For example, in the point cloud of a car door, the panel area has a higher point density and forms a main cluster, while the edge bolts have a lower point density and form an independent cluster, thus generating a segmented point cloud cluster.

[0024] S202: Based on the segmented point cloud clusters, principal component analysis is used to calculate the normal vector and curvature features of each cluster, identify the geometric features of planes and cylinders, and generate geometric feature descriptors; Based on the segmented point cloud clusters, principal component analysis is used to calculate the normal vector and curvature features of each cluster. For the point set of a single cluster, the covariance matrix is ​​calculated, and eigenvalues ​​are obtained through eigenvalue decomposition. And the eigenvectors, the eigenvector corresponding to the smallest eigenvalue is the normal vector direction, and the curvature is calculated as follows: A curvature threshold of 0.05 is set; values ​​less than 0.05 are classified as planes, and values ​​greater than 0.3 are classified as cylinders. For example, the eigenvalues ​​of the covariance matrix of the panel cluster points are used. 2. Curvature δ=0.02<0.05 is identified as a plane, bolt cluster. It belongs to the transition region, but is judged to be a cylindrical feature based on the direction of the feature vector, and a geometric feature descriptor is generated.

[0025] S203: Based on geometric feature descriptors, a region growing algorithm is used to reconstruct the surface, and a closed triangular mesh model is generated through the Poisson reconstruction algorithm to generate the initial triangular mesh model; Based on geometric feature descriptors, a region growing algorithm is used for surface reconstruction. Points with curvature less than 0.1 are used as seed points. The angle between the normal vectors of neighboring points is compared, with an angle threshold of 15 degrees. For example, the normal vector of the seed point... 0.01, 0.02, 0.99 Neighborhood point normal vector 0.02, 0.01, 0.99 Angles calculated to be 1.2 degrees, which are less than the threshold, are included in the same region. All points are traversed to complete the region segmentation. Then, a closed triangular mesh model is generated by the Poisson reconstruction algorithm. This algorithm constructs an indicator function based on the point cloud normal vectors, solves the Poisson equation, and sets the isosurface extraction to 0.5 to generate the initial triangular mesh model.

[0026] S204: Based on the initial triangular mesh model, the mesh quality is optimized using the Laplace smoothing algorithm, and the iterative nearest point algorithm is used to register with the standard CAD model to correct deviations and generate a 3D model of the workpiece.

[0027] Based on the initial triangular mesh model, the Laplace smoothing algorithm is used to optimize the mesh quality. The position of each vertex is adjusted to the average position of its adjacent vertices, and the smoothing factor is set to 0.2. For example, the original position of vertex V is... 50.0, 30.0, 0.15 Average position of adjacent points 50.1, 30.1, 0.16 The new position is calculated as 0.8V + 0.2 average position. 50.02, 30.02, 0.152 The iterative nearest point algorithm is used to register the model with the standard CAD model. The distance between the points is calculated, and the distance threshold is set to 0.1 mm. The rotation matrix R and translation vector t are solved by singular value decomposition. For example, after 5 iterations, the registration error is reduced from the initial 0.3 mm to 0.05 mm. The deviation is corrected and the 3D model of the workpiece is generated.

[0028] The generation of optimized welding paths based on S3 includes the following steps: S301: Based on the 3D model of the workpiece, the A* search algorithm is used to generate an initial sequence of welding path points on the model surface to ensure coverage of all areas to be welded. Based on the 3D model of the workpiece, the A* search algorithm is used to generate an initial sequence of welding path points on the model surface. This algorithm evaluates the cost function of each point. Perform path search, where h(n) represents the actual cost from the starting point to the current point. To estimate the cost from the current point to the target, set the grid resolution to 1 millimeter, with the bottom left corner of the door panel as the reference point. 0,0,0 Starting from the top right corner 100,50,0 For the target point, calculate the f-value for each grid point, for example, point... 10,10,0 The g value is 14.14 mm, which is the straight-line distance from the starting point; the h value is 94.34 mm, which is the straight-line distance to the target point; and the f value is 108.48 mm. The point with the smallest f value is selected and added to the path, and then extended to the target point in sequence to form a sequence of about 150 points, ensuring that all areas to be welded are covered, thus generating the initial path point sequence.

[0029] S302: Based on the initial path point sequence, a deep Q-network is trained in a simulation environment. The reward function optimizes the path point order and robot posture to avoid collisions and generate an optimized welding path. In practice, historical welding path datasets are preferred for training.

[0030] Based on the initial path point sequence, a deep Q-network is trained in a simulation environment. The state space is set as the path point coordinates and robot joint angles, and the action space is set as moving to an adjacent point or adjusting posture. The reward function is... The weight Distance to reward ,in Collision penalty for distance traveled That is, if the distance between the robot and the workpiece is less than 2 millimeters, heat accumulation penalty Where T is the cumulative time during which the local temperature exceeds 300°C, and the network input status. Output the Q value and select the action with the maximum Q value. Update target Q value With a discount factor γ=0.9 and a learning rate α=0.001, after 1000 iterations of training, the number of path points was optimized from 150 to 120, the number of posture adjustments was reduced by 40%, and an optimized welding path was generated.

[0031] The generation of dynamic welding parameter sets based on S4 includes the following steps: S401: Based on the optimized welding path, extract the curvature and distance features of each point on the path from the previous welding point to generate a path geometric feature vector; Based on the optimized welding path, the curvature and distance features from the previous weld point are extracted from each point on the path. The curvature is calculated using a three-point method, taking the current point... and front and back points Calculate vector curvature For example, point 50,30,0 and adjacent points 49.8,30.1,0) 50.2,29.9,0 ,vector , Cross product modulus 0.02, dot product modulus 0.0447, curvature Distance features If the point spacing is 1.2 mm, these features are calculated for all 120 path points to form a feature vector. Generate path geometric feature vectors.

[0032] S402: Based on the path geometric feature vector and material database, a convolutional neural network is used to predict the basic welding parameters of each path point and generate a set of basic welding parameters. Based on path geometric feature vectors and a material database, a convolutional neural network is used to predict the basic welding parameters for each path point. The network input is a sequence of feature vectors, and the 1D convolutional kernel size is set to 3. The network has 32 output channels, ReLU activation function, maxpooling pooling layer, and fully connected layer outputs three parameters: power P, velocity v, and focal length f. A material database provides baseline values ​​for these parameters, such as stainless steel with a thermal conductivity of 16 W / mK, melting point of 1400℃, baseline power of 1500 W, velocity of 0.8 m / min, and focal length of 0 mm. Network training weights are adjusted based on material properties, for example, input features [1.78, 1.2]. Convolutional layers output feature maps, and fully connected layers perform calculations. , Output from the network, such as The predicted power is 1620W, and the predicted speed is 0.75m / min. The focal length is -0.2mm. Calculate for all points to generate a basic welding parameter set.

[0033] S403: Based on the basic welding parameter set, a genetic algorithm is used for multi-objective optimization to find the Pareto optimal solution set and generate a dynamic welding parameter set.

[0034] Based on a set of basic welding parameters, a genetic algorithm is used for multi-objective optimization. The population size is 100, and the iteration is performed for 50 generations. Each individual is encoded as a parameter adjustment amount. The mutation probability is 0.1, the crossover probability is 0.8, and the objective function includes melt depth H and heat input. ,efficiency Constraints Pareto ordering selects non-dominated solutions, such as individuals. Calculate P = 1570 W, v = 0.85 H is obtained through empirical formulas The estimated thickness is 1.2mm. The constraints are satisfied, and the parameters are sorted. Finally, the Pareto front solution set is selected, such as 10 sets of parameters [P=1600W, v=0.8m / min, f=-0.1mm], to generate a dynamic welding parameter set.

[0035] Generating a real-time welding trajectory based on S5 includes the following steps: S501: Based on the dynamic welding parameter set, the cubic spline interpolation algorithm is used to convert discrete path points into smooth and continuous robot joint space trajectories, generating joint space trajectory commands. Based on a dynamic welding parameter set, a cubic spline interpolation algorithm is used to transform discrete path points into smooth, continuous robot joint space trajectories. This algorithm uses 120 path points as nodes, with each node acting as a reference point. For example, the first derivative at the node is calculated using the continuity condition, the boundary condition is set to natural boundary, and the second derivative is zero. For each interval... The trajectory is represented as a cubic polynomial. The coefficients are determined by solving a system of linear equations, for example, at point... Interval, time interval 0.1 seconds, calculate coefficients 500 interpolation points are generated, each containing position and velocity information, which are then converted into the six joint angles of the robot. Through inverse kinematics calculations, such as the angle of joint 1 Generate joint space trajectory instructions.

[0036] S502: Based on joint space trajectory commands, a six-axis industrial robot is driven by feedforward control and PID control algorithms, and the trajectory is tracked in real time through encoder feedback to generate theoretical trajectory feedback data. Based on joint space trajectory commands, a six-axis industrial robot is driven using feedforward control and PID control algorithms. The feedforward control calculates the torque corresponding to the desired acceleration. Where M is the mass matrix, C is the Coriolis force matrix, and G is the gravity term. For the desired joint angle, the PID control calculates the error torque. ,in proportionality coefficient Integral coefficient Differential coefficients The encoder provides real-time feedback of the joint angle. For example, the expected angle of joint 1 is 30 degrees, the actual angle is 29.8 degrees, and the error is 0.2 degrees. Total torque Drive the motor and record the desired position. and actual location This generates theoretical trajectory feedback data.

[0037] S503: Based on theoretical trajectory feedback data and real-time welding trajectory, the Kalman filter algorithm is used to fuse the data, estimate the optimal trajectory state, and fine-tune the control commands to generate the real-time welding trajectory.

[0038] Based on theoretical trajectory feedback data and real-time welding trajectory, a Kalman filter algorithm is used for data fusion, and the state vector... Process noise covariance Measure noise covariance Prediction steps The state transition matrix Control matrix Update steps to calculate Kalman gain Status update For example, at time t=0.5 seconds, predict the state. Spend, degrees per second, observed value Spend, degrees / second, gain After the update Spend, degrees per second, fine-tuning control commands based on state estimates. Generates real-time welding trajectories.

[0039] Generating multi-sensor monitoring data based on S6 includes the following steps: S601: During the welding process, based on the real-time welding trajectory, the vision camera is synchronously triggered to acquire the molten pool image, and the median filtering algorithm is used for image preprocessing to generate molten pool image data. During the welding process, based on the real-time welding trajectory, a vision camera is synchronously triggered to acquire images of the molten pool. The camera sampling frequency is 1000Hz, and the exposure time is... The image was acquired at a resolution of 1024x768 pixels. The original image contained Gaussian noise and was processed using a median filter algorithm. A 3x3 filter window was selected, and the grayscale values ​​of the nine pixels within the window were sorted. The median value was used as the center pixel threshold, for example, the window pixel value. ; After sorting The intermediate value 121 replaces the original center value 255, effectively removing impulse noise and generating molten pool image data.

[0040] S602: Simultaneously, an infrared thermal imager is used to collect temperature field data of the molten pool and heat-affected zone, and a non-uniformity correction algorithm is used to process the raw thermal data to generate temperature field data; Simultaneously, an infrared thermal imager was used to acquire temperature field data of the molten pool and heat-affected zone. The thermal imager had a resolution of 320x240 pixels, a temperature measurement range of 500-2500℃, and a sampling frequency of 100Hz. Since the raw thermal data contained non-uniformity errors, a two-point correction algorithm was used for processing. First, at a low temperature... and high temperature Acquire a uniform blackbody image and calculate the gain coefficient for each pixel. and offset coefficient Where V is the original output value of the pixel, for example, pixel 100,100 The output value is 1200 at 500℃ and 1800 at 2500℃. Therefore, G = (1800 - 1200) / 2000 = 0.3, O = 1200 - 0.3 * 500 = 1050. The corrected output... For example, if the current collected value is 1500, the corrected temperature T = (1500-1050) / 0.3 = 1500℃, and the temperature field data is generated.

[0041] S603: The acoustic emission sensor is used to collect the acoustic signal during the welding process, and the wavelet transform algorithm is used to perform time-frequency analysis, extract the characteristic frequency, and generate acoustic emission signal characteristics. Acoustic emission sensors were used to simultaneously acquire acoustic signals during the welding process. The sensors had a sensitivity of 100 mV / m / s, a frequency range of 50-400 kHz, and a sampling rate of 1 MHz. Raw sound pressure signals were acquired, and time-frequency analysis was performed using Daubechies wavelet transform. A 5-level decomposition was conducted using the db4 wavelet basis, and detail coefficients cD1-cD5 and approximation coefficients cA5 were extracted for each level. The energy of each level was then calculated. The maximum energy corresponds to the main frequency band of the layer. For example, the maximum energy of the cD3 layer is 0.45, which corresponds to a frequency of 125kHz. This frequency is the characteristic frequency of molten pool oscillation and generates acoustic emission signal characteristics.

[0042] S604: Based on molten pool image data, temperature field data, and acoustic emission signal characteristics, an extended Kalman filter algorithm is used for time synchronization and spatial registration, which are then fused into a unified multi-dimensional data stream to generate multi-sensor monitoring data.

[0043] Based on molten pool image data, temperature field data, and acoustic emission signal characteristics, an extended Kalman filter algorithm is used for time synchronization and spatial registration. Time synchronization is based on the welding trajectory time, and the data from each sensor are aligned through linear interpolation of timestamps. Spatial registration involves matching the image pixel coordinates. Mapped to the workpiece coordinate system via affine transformation The transformation matrix is ​​obtained through calibration, and the state vector... Includes temperature, molten pool area, acoustic emission frequency, and observation vector. Process noise covariance Observation noise covariance Prediction steps Update steps to calculate Kalman gain Status update These data are integrated into a unified multi-dimensional data stream, generating multi-sensor monitoring data.

[0044] Generating a defect detection report based on S7 includes the following steps: S701: Based on multi-sensor monitoring data, texture features are extracted from image data using the gray-level co-occurrence matrix method, gradient features are extracted from temperature data, and amplitude features are extracted from sound signals to generate multimodal feature vectors. Based on multi-sensor monitoring data, texture features were extracted from the image data using the gray-level co-occurrence matrix method, with a distance d=1 pixel and an angle... Calculate four features: contrast, correlation, energy, and homogeneity. For example, contrast... Where P is the gray-level co-occurrence matrix, i,j are gray levels, and gradient features are calculated from temperature data using the Sobel operator. The horizontal convolution kernel is [-1,0,1;-2,0,2;-1,0,1], and the vertical convolution kernel is [-1,-2,-1;0,0,0;1,2,1]. The gradient magnitude is... ,direction Calculate amplitude characteristics from acoustic signals, including root mean square. Peak factor This generates multimodal feature vectors.

[0045] S702: Based on multimodal feature vectors, a support vector machine classifier is used for real-time reasoning to determine whether there are defects in the current solder joint and the type of defects, and to generate a real-time defect probability distribution; Based on multimodal feature vectors, a support vector machine classifier is used for real-time inference. The kernel function is a radial basis function with a penalty coefficient C=1.0 and a kernel coefficient γ=0.1. The input feature vectors are normalized to [0,1]. The training data includes 1000 normal samples and 300 defective samples, with defects classified into three categories: porosity, cracks, and lack of fusion. The decision function is calculated during inference. The probability is determined by the sigmoid function. calculate; For example, feature vectors If a defect is identified, the probability of each category [porosity: 0.6, crack: 0.3, lack of fusion: 0.1] is calculated, and a real-time defect probability distribution is generated.

[0046] S703: Based on real-time defect probability distribution, it uses time series analysis to track the generation and development of defects, and generates a structured report containing defect location, type and confidence level through a rule reasoning engine, thus generating a defect detection report.

[0047] Based on real-time defect probability distribution, time series analysis is used to track the generation and development of defects. A sliding window of 10 points is used, and the mean probability μ and standard deviation σ within the window are calculated. Defects with μ > 0.6 and σ < 0.1 are defined as persistent defects. The rule-based inference engine is set with the condition: IF probability > 0.7 AND persist for 3 points THEN defect confirmation. The location is determined by the coordinates of the welding trajectory points, and the type is the category with the highest probability. The confidence level is calculated as: probability value 0.8 + duration coefficient 0.2. For example, for a point... 50,30,0 The probability of porosity is 0.75 for 5 consecutive points, and the confidence level is 0.75 + 0.8 + 0.9 + 0.2 = 0.78. A defect detection report is generated.

[0048] The steps for generating the repaired weld result based on S8 are as follows: S801: Based on the defect detection report, a path replanning algorithm is used to generate a local compensation welding path that covers the defect area at the defect location, thus generating a local repair path; Based on the defect detection report, a path replanning algorithm is used to generate a localized, compensated welding path covering the defect area at the defect location. A circular path with a radius of 2 mm is generated, centered on the defect point, with 8 points spaced at 45° intervals. The point coordinates are calculated as follows: For example, defect points 50,30,0 Generating points 52,30,0) 51.414,31.414,0 Then, generate a local repair path.

[0049] S802: Based on local repair paths and defect types, the Q-learning algorithm is used to query the optimal repair parameters from historical data and generate a repair parameter strategy; Based on local repair paths and defect types, the Q-learning algorithm is used to query the optimal repair parameters from historical data, with state s= Defect type, location, size Action a = Power adjustment, speed adjustment, focal length adjustment Reward R = - Defect probability after repair +0.1 * heat input Discount factor γ=0.9, learning rate α=0.1, for example, porosity defects, actions in historical Q values. Power +50W, speed -0.1m / min The highest Q value is -0.2. Select this action to generate a repair parameter strategy.

[0050] S803: Based on the repair parameter strategy, control the robot and laser to perform repair welding, and use the monitoring process of S6 to monitor the repair process in real time and generate repair process monitoring data; Based on the repair parameter strategy, the robot and laser are controlled to perform repair welding. The laser power is increased by 50W, the speed is decreased by 0.1m / min, and the focal length remains unchanged. The S6 monitoring process is used to monitor the repair process in real time, collect images of the molten pool, temperature field and acoustic signals, process and generate monitoring data, and generate repair process monitoring data.

[0051] S804: Based on the monitoring data of the repair process, support vector machine is used again to detect defects at the repair points, verify the repair effect, and generate the welding results after repair.

[0052] Based on the monitoring data of the repair process, a support vector machine is used again to detect defects at the repair points. The feature vector is input into the trained classifier to calculate the defect probability. A threshold of 0.3 is set, and a probability below the threshold is considered a successful repair. For example, the feature vector of the repair point is [0.2, 0.1, 0.4, 0.3, 0.1, 0.2]. However, it decreased from 0.75 before the repair, resulting in the post-repair welding result.

[0053] Generating the final quality assessment report based on S9 includes the following steps: S901: Based on the welding results after repair and the defect detection report, the key quality indicators are calculated using statistical process control methods to generate a quality indicator dataset. Based on the repaired welding results and defect detection report, key quality indicators were calculated using statistical process control methods, including defect rate = number of defect points / total number of defect points, and weld width standard deviation. Melt depth variation coefficient For example, with a total of 120 points, 2 defect points, a defect rate of 1.67%, an average weld width of 2.0 mm, and a standard deviation of 0.1 mm, a quality index dataset is generated.

[0054] S902: Based on the quality index dataset and dynamic welding parameter set, a mapping model between process parameters and quality indicators is established using multiple linear regression analysis. The significance of parameter influence is analyzed, and a process-quality correlation model is generated. Based on a quality index dataset and a dynamic welding parameter set, a mapping model between process parameters and quality indicators is established using multiple linear regression analysis. The dependent variable is y = defect rate, and the independent variables are X = [power, speed, focal length]. The least squares method is used to solve for the coefficients, for example... Calculate the t-statistic to determine significance and generate a process-quality correlation model.

[0055] S903: Based on the process-quality correlation model, the gradient descent algorithm is used to back-optimize the weight parameters of the convolutional neural network in S4, so as to realize the system's self-iterative improvement and generate model optimization parameters; Based on the process-quality correlation model, the gradient descent algorithm is used to back-optimize the weight parameters of the convolutional neural network in S4, and the loss function is... Learning rate η = 0.001, weight update For example, weight w=0.5, gradient After the update, w = 0.5 - 0.001 * 0.2 = 0.4998, generating the model optimization parameters.

[0056] S904: Integrates the quality indicator dataset, process-quality correlation model, and all metadata of this welding task, and automatically generates a structured evaluation report using natural language generation technology, thus generating the final quality evaluation report.

[0057] By integrating the quality indicator dataset, the process-quality correlation model, and all metadata of this welding task, a structured evaluation report is automatically generated using natural language generation technology. The report includes sections such as welding time, workpiece ID, defect statistics, parameter impact analysis, and improvement suggestions. For example, the generated text is: Welding time 2023-10-10 10:00:00, workpiece ID -D001, total number of points 120, defect points 2, defect rate 1.67%, power influence coefficient on defect rate -0.0001, suggested power increase of 50W, generating the final quality evaluation report.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-precision intelligent welding system adaptable to complex working conditions, characterized in that, Includes the following steps: S1: Based on a high-resolution vision sensor and laser scanner, the workpiece geometric data is acquired using 3D scanning technology. Noise and outliers are removed by filtering algorithms, and data preprocessing is performed to generate preprocessed point cloud data. S2: Based on preprocessed point cloud data, point cloud processing algorithms and machine learning methods are used to identify workpiece features, and a precise digital model is generated through a 3D reconstruction algorithm to generate a 3D model of the workpiece. S3: Based on the 3D model of the workpiece, the welding path is autonomously planned using a reinforcement learning algorithm, taking into account obstacle avoidance, heat accumulation minimization and efficiency optimization, and the feasibility of the path is verified through a simulation environment to generate an optimized welding path; S4: Based on the optimized welding path and material property database, a deep learning algorithm is used to predict and optimize laser welding parameters in real time, and the network weights are adjusted through the backpropagation algorithm to generate a dynamic welding parameter set; S5: Based on a dynamic welding parameter set, it uses a PID control algorithm and robot motion control to execute the laser welding process. It fine-tunes the robot arm trajectory through real-time feedback to ensure accurate path following and generate a real-time welding trajectory. S6: Based on real-time welding trajectory, it integrates multiple sensors, uses data fusion algorithm to synchronously monitor welding status, and performs real-time data alignment and noise reduction to generate multi-sensor monitoring data; S7: Based on multi-sensor monitoring data, machine learning algorithms are used to detect welding defects in real time, and image and signal data are analyzed through feature extraction methods to generate a defect probability map and obtain a defect detection report. S8: Based on the defect detection report, the reinforcement learning algorithm is used to optimize the repair strategy, control the auxiliary laser head to perform local remelting or repair welding, and adjust the repair parameters through real-time control methods to minimize the heat-affected zone and generate the repaired welding result; S9: Based on the repaired welding results and historical welding data, a statistical analysis algorithm is used to evaluate the overall welding quality, and the system model parameters are updated through feedback loops for future welding process optimization, generating a final quality assessment report.

2. The high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: The generation of preprocessed point cloud data based on S1 includes the following steps: S101: Based on a high-resolution laser scanner, the workpiece surface is scanned using triangulation to collect raw three-dimensional coordinate data and generate raw point cloud data; S102: Based on the original point cloud data, a statistical outlier removal algorithm is used to filter out noise points, and a voxel grid downsampling method is used to reduce data density and improve the efficiency of subsequent processing to generate denoised point cloud data. S103: Based on the denoised point cloud data, a Gaussian filtering algorithm is used for smoothing, and the surface is reconstructed by the moving least squares method to improve data quality and generate preprocessed point cloud data.

3. The high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: The steps for generating a 3D model of a workpiece based on S2 are as follows: S201: Based on preprocessed point cloud data, a density-based clustering method is used to segment the different components of the workpiece and generate segmented point cloud clusters. S202: Based on the segmented point cloud clusters, principal component analysis is used to calculate the normal vector and curvature features of each cluster, identify the geometric features of planes and cylinders, and generate geometric feature descriptors; S203: Based on geometric feature descriptors, a region growing algorithm is used to reconstruct the surface, and a closed triangular mesh model is generated through the Poisson reconstruction algorithm to generate the initial triangular mesh model; S204: Based on the initial triangular mesh model, the mesh quality is optimized using the Laplace smoothing algorithm, and the iterative nearest point algorithm is used to register with the standard CAD model to correct deviations and generate a 3D model of the workpiece.

4. The high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: The generation of optimized welding paths based on S3 includes the following steps: S301: Based on the 3D model of the workpiece, the A* search algorithm is used to generate an initial sequence of welding path points on the model surface to ensure coverage of all areas to be welded. S302: Based on the initial path point sequence, a deep Q-network is trained in a simulation environment. The path point order and robot posture are optimized through the reward function to avoid collisions and generate an optimized welding path.

5. A high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: The generation of dynamic welding parameter sets based on S4 includes the following steps: S401: Based on the optimized welding path, extract the curvature and distance features of each point on the path from the previous welding point to generate a path geometric feature vector; S402: Based on the path geometric feature vector and material database, a convolutional neural network is used to predict the basic welding parameters of each path point and generate a set of basic welding parameters. S403: Based on the basic welding parameter set, a genetic algorithm is used for multi-objective optimization to find the Pareto optimal solution set and generate a dynamic welding parameter set.

6. The high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: Generating a real-time welding trajectory based on S5 includes the following steps: S501: Based on the dynamic welding parameter set, the cubic spline interpolation algorithm is used to convert discrete path points into smooth and continuous robot joint space trajectories, generating joint space trajectory commands. S502: Based on joint space trajectory commands, a six-axis industrial robot is driven by feedforward control and PID control algorithms, and the trajectory is tracked in real time through encoder feedback to generate theoretical trajectory feedback data. S503: Based on theoretical trajectory feedback data and real-time welding trajectory, the Kalman filter algorithm is used to fuse the data, estimate the optimal trajectory state, and fine-tune the control commands to generate the real-time welding trajectory.

7. A high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: Generating multi-sensor monitoring data based on S6 includes the following steps: S601: During the welding process, based on the real-time welding trajectory, the vision camera is synchronously triggered to acquire the image of the molten pool, and the median filtering algorithm is used for image preprocessing to generate molten pool image data. S602: Simultaneously, an infrared thermal imager is used to collect temperature field data of the molten pool and heat-affected zone, and a non-uniformity correction algorithm is used to process the raw thermal data to generate temperature field data; S603: The acoustic emission sensor is used to collect the acoustic signal during the welding process, and the wavelet transform algorithm is used to perform time-frequency analysis, extract the characteristic frequency, and generate acoustic emission signal characteristics. S604: Based on molten pool image data, temperature field data, and acoustic emission signal characteristics, an extended Kalman filter algorithm is used for time synchronization and spatial registration, which are then fused into a unified multi-dimensional data stream to generate multi-sensor monitoring data.

8. A high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: Generating a defect detection report based on S7 includes the following steps: S701: Based on multi-sensor monitoring data, texture features are extracted from image data using the gray-level co-occurrence matrix method, gradient features are extracted from temperature data, and amplitude features are extracted from sound signals to generate multimodal feature vectors. S702: Based on multimodal feature vectors, a support vector machine classifier is used for real-time reasoning to determine whether there are defects in the current solder joint and the type of defects, and to generate a real-time defect probability distribution; S703: Based on real-time defect probability distribution, it uses time series analysis to track the generation and development of defects, and generates a structured report containing defect location, type and confidence level through a rule reasoning engine, thus generating a defect detection report.

9. A high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: The steps for generating the repaired weld result based on S8 are as follows: S801: Based on the defect detection report, a path replanning algorithm is used to generate a local compensation welding path that covers the defect area at the defect location, thus generating a local repair path; S802: Based on local repair paths and defect types, the Q-learning algorithm is used to query the optimal repair parameters from historical data and generate a repair parameter strategy; S803: Based on the repair parameter strategy, control the robot and laser to perform repair welding, and use the monitoring process of S6 to monitor the repair process in real time and generate repair process monitoring data; S804: Based on the monitoring data of the repair process, support vector machine is used again to detect defects at the repair points, verify the repair effect, and generate the welding results after repair.

10. A high-precision intelligent welding system adaptable to complex working conditions according to claim 1, characterized in that: Generating the final quality assessment report based on S9 includes the following steps: S901: Based on the welding results after repair and the defect detection report, the key quality indicators are calculated using statistical process control methods to generate a quality indicator dataset. S902: Based on the quality index dataset and dynamic welding parameter set, a mapping model between process parameters and quality indicators is established using multiple linear regression analysis. The significance of parameter influence is analyzed, and a process-quality correlation model is generated. S903: Based on the process-quality correlation model, the gradient descent algorithm is used to back-optimize the weight parameters of the convolutional neural network in S4, so as to realize the system's self-iterative improvement and generate model optimization parameters; S904: Integrates the quality indicator dataset, process-quality correlation model, and all metadata of this welding task, and automatically generates a structured evaluation report using natural language generation technology, thus generating the final quality evaluation report.

Images