Special-shaped storage tank welding path planning system

Abstract

The present invention relates to the technical field of welding of special-shaped storage tanks, and discloses a welding path planning system for special-shaped storage tanks, which includes a surface data acquisition module, a geometric feature extraction module, a path planning module, a path optimization module, and a welding instruction generation module, and may also include a welding execution module. The surface data acquisition module acquires surface point cloud data through a laser scanner; the geometric feature extraction module performs three-dimensional reconstruction and feature segmentation on the data to obtain a geometric feature set; the path planning module performs initial welding path planning based on the set; the path optimization module performs smoothing optimization and collision detection on the initial path segment set to obtain an optimized path segment set; and the welding instruction generation module generates welding control instructions based on the optimized path segment set. The welding execution module is composed of a welding robot and a position sensor, and can execute instructions and perform real-time monitoring. The system improves the accuracy, efficiency, and stability of special-shaped storage tank welding and is suitable for the welding process of special-shaped storage tanks in industrial manufacturing.

Description

Technical Field

[0001] The invention relates to the technical field of special-shaped storage tank welding, in particular to a special-shaped storage tank welding path planning system. Background Art

[0002] Judging from the current state of technological development, the existing welding path planning system has many shortcomings when dealing with special-shaped storage tanks. On the one hand, the extraction of geometric features of the tank surface is not accurate enough. Traditional point cloud processing technology often uses simple threshold segmentation or region growing algorithms, which cannot effectively deal with complex features such as sudden curvature changes, depressions, and protrusions that may exist on the surface of special-shaped storage tanks, resulting in a lack of reliable geometric basis for subsequent path planning. On the other hand, the path planning and optimization process lacks sufficient consideration of the spatial dimension. Most of the existing path planning algorithms are based on two-dimensional planes or simple three-dimensional straight line models, ignoring the continuity and smoothness requirements of the welding path in space during the welding process of special-shaped storage tanks. Problems such as excessively sharp path turns and unstable movement of the welding robot are prone to occur, which not only affects welding efficiency, but may also lead to a decline in welding quality due to motion impact.

[0003] Traditional systems lack effective management of path accuracy and dynamic weighting during the generation and execution of welding instructions. Welding instructions are often generated based on fixed path parameters and cannot be dynamically adjusted based on path accuracy differences and spatial position relationships during the actual welding process. This makes it difficult for welding robots to accurately track and optimize paths when executing instructions, and welding position deviations are prone to occur. Furthermore, existing systems lack real-time monitoring capabilities for the welding process, failing to collect welding position data in a timely manner and align it with the planned path. This makes it difficult to achieve closed-loop control of the welding process, making it difficult to ensure stable welding quality.

[0004] With the development of intelligent manufacturing and industrial automation, the requirements for welding quality and efficiency of special-shaped storage tanks are increasing. There is an urgent need for a welding path planning system that can accurately extract geometric features, optimize welding paths, dynamically generate welding instructions, and achieve real-time monitoring. The limitations of existing technologies for welding path planning of special-shaped storage tanks make the development of a more efficient and accurate system a pressing technical challenge for the industry. Summary of the Invention

[0005] The purpose of the present invention is to provide a welding path planning system for special-shaped storage tanks to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a welding path planning system for special-shaped storage tanks, the system comprising:

[0007] Surface data acquisition module, used to obtain surface point cloud data of special-shaped storage tanks through laser scanner;

[0008] A geometric feature extraction module, configured to perform three-dimensional reconstruction and feature segmentation on the surface point cloud data to obtain a geometric feature set;

[0009] a path planning module, configured to perform initial welding path planning based on the set of geometric features to obtain a set of initial path segments;

[0010] a path optimization module, configured to perform smoothing optimization and collision detection on the set of initial path segments to obtain a set of optimized path segments;

[0011] A welding instruction generation module is used to generate welding control instructions according to the set of optimized path segments.

[0012] Preferably, the geometric feature extraction module includes: a point cloud segmentation unit, used to divide the surface point cloud data into a preset grid to obtain a set of point cloud units; a feature extraction unit, used to input the set of point cloud units into a deep learning-based feature extractor to obtain the geometric feature set.

[0013] Preferably, the path optimization module includes: a path segment reconstruction unit, which reorganizes the initial path segments along the spatial dimension to obtain a set of initial path vectors; a path smoothness calculation unit, which is used to calculate the geometric correlation between adjacent vectors in the set of initial path vectors to generate a path smoothness matrix; a weight correction unit, which is used to perform weight distribution correction on the path smoothness matrix according to the spatial relationship between adjacent vectors in the set of initial path vectors to obtain a path constraint matrix; and a path optimization unit, which is used to perform multi-dimensional convolution fusion on the path constraint matrix and the set of initial path vectors to obtain the optimized path segment.

[0014] Preferably, the path smoothness calculation unit includes: a spatial mapping subunit, used to map each vector in the set of initial path vectors to Euclidean space to obtain a set of mapped initial path vectors; a geometric association analysis subunit, used to calculate the cosine similarity between adjacent mapping vectors in the set of mapped initial path vectors to generate the path smoothness matrix composed of multiple path smoothness values.

[0015] Preferably, the path optimization unit is specifically implemented as follows: performing deep convolution processing on the path constraint matrix to obtain a path constraint feature matrix; inputting the set of initial path vectors and the path constraint feature matrix into a gated encoding network to obtain a set of path context vectors; and performing tensor stacking on the set of path context vectors to obtain the optimized path segment.

[0016] Preferably, the welding instruction generation module includes: an optimized path dimensionality reduction unit, which is used to perform mean pooling processing on the length dimension of each segment in the set of optimized path segments to obtain a set of optimized path vectors; a path dynamic accuracy calculation unit, which is used to calculate the accuracy value of each vector in the set of optimized path vectors to generate a path accuracy set; an instruction reference center determination unit, which is used to select the optimized path vector corresponding to the maximum accuracy value in the path accuracy set as the initial instruction center vector; an instruction dynamic weight allocation unit, which is used to calculate the dynamic instruction weight of each vector in the set of optimized path vectors and the initial instruction center vector based on the geometric distance between each vector and the initial instruction center vector and the accuracy value of each vector to generate a dynamic weight set; a control instruction reconstruction unit, which is used to use the dynamic weight set to perform weighted fusion on the set of optimized path vectors to generate the welding control instruction.

[0017] Preferably, the path dynamic accuracy calculation unit is specifically implemented as follows: calculating the median vector and standard deviation vector of the optimized path vector; performing element-by-element difference calculation between the optimized path vector and the median vector, and squaring the difference result to obtain a path difference vector; calculating the overall median value of the path difference vector; multiplying the median value by the square value of the standard deviation vector, and inputting the result into a normalization function to obtain the accuracy value.

[0018] Preferably, the instruction dynamic weight allocation unit is specifically implemented as follows: multiplying the accuracy value of the optimized path vector and the accuracy value of the initial instruction center vector by a first correction coefficient to obtain a first dynamic instruction factor; multiplying the Euclidean distance between the optimized path vector and the initial instruction center vector by a second correction coefficient to obtain a second dynamic instruction factor; and performing a product operation on the first dynamic instruction factor and the second dynamic instruction factor to obtain the dynamic instruction weight.

[0019] Preferably, the welding instruction generation module is specifically implemented as follows: inputting the welding control instruction into a command signal generator based on a decision tree to obtain a welding execution control signal, and the welding execution control signal is used to indicate the start and stop strategy of the welding process.

[0020] Preferably, the system also includes a welding execution module, which is composed of a welding robot and a position sensor, wherein: the welding robot is equipped with a multi-axis robotic arm for executing the welding control instructions; the position sensor is arranged on the surface of the storage tank, for collecting welding position data in real time, and generating real-time control feedback after position correction with the set of optimized path segments; the real-time control feedback is transmitted to the central controller via the industrial bus to complete welding process monitoring.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The system uses a laser scanner through the surface data acquisition module to acquire surface point cloud data of irregular-shaped tanks, laying a precise data foundation for subsequent work. The geometric feature extraction module performs 3D reconstruction and feature segmentation on the surface point cloud data, accurately obtaining a set of geometric features. Compared with traditional methods, this greatly improves the accuracy and completeness of the geometric features acquired on the surface of irregular-shaped tanks.

[0023] The path planning module performs initial welding path planning based on a set of geometric features to obtain a set of initial path segments. This planning method based on actual geometric features makes the initial path more consistent with the structural characteristics of special-shaped storage tanks and avoids the unreasonable path problems caused by traditional fixed-mode planning. The path optimization module performs smoothing optimization and collision detection on the initial path segment set. The path segment reconstruction unit reorganizes the initial path segments along the spatial dimension to obtain an initial path vector set. The path smoothness calculation unit then calculates the geometric correlation of adjacent vectors to generate a path smoothness matrix. The weight correction unit performs weight distribution correction to obtain a path constraint matrix. Finally, the path optimization unit obtains the optimized path segment through multi-dimensional convolution fusion. This series of operations effectively improves the smoothness of the path and avoids vibration during the execution of the welding robot. At the same time, collision detection ensures the safety of the path and reduces the risk of equipment collision.

[0024] The welding instruction generation module generates welding control instructions based on a set of optimized path segments. The multiple units it contains work together. The optimized path dimensionality reduction unit performs length dimension mean pooling, the path dynamic precision calculation unit calculates the precision values ​​of each vector, the instruction reference center determination unit selects the vector corresponding to the maximum precision value as the initial instruction center vector, the instruction dynamic weight allocation unit calculates the dynamic instruction weight based on the geometric distance and precision value, and the control instruction reconstruction unit uses the dynamic weights to perform weighted fusion to generate instructions. This method generates more accurate and flexible welding control instructions, better adapting to the welding requirements of special-shaped storage tanks and improving welding quality and efficiency.

[0025] The system also includes a welding execution module, consisting of a welding robot and position sensors. The welding robot, equipped with a multi-axis robotic arm, executes welding control commands. The position sensor collects welding position data in real time, performs position corrections on the optimized path segment set, and generates real-time control feedback. This data is transmitted to a central controller via an industrial bus to monitor the welding process. This enables real-time monitoring and adjustment of the welding process, ensuring reliability and stability, enabling timely identification and resolution of problems, and further improving welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a working principle diagram of the special-shaped storage tank welding path planning system of the present invention;

[0027] Figure 2 This is the working principle diagram of the path optimization module;

[0028] Figure 3 Generate working principle diagram for welding instruction module;

[0029] Figure 4 This is the working principle diagram of the path dynamic accuracy calculation unit. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] See also Figures 1-4 The present invention relates to a welding path planning system for special-shaped storage tanks. The system comprises: a surface data acquisition module, a geometric feature extraction module, a path planning module, a path optimization module, and a welding instruction generation module. The specific implementation method is as follows:

[0032] The surface data acquisition module uses a laser scanner to acquire surface point cloud data of irregular-shaped tanks. This laser scanner can be a high-precision 3D laser scanner. Its laser beam rapidly scans the tank surface, acquiring 3D coordinate information point by point based on a preset sampling frequency and scanning range. This generates surface point cloud data containing a large number of discrete points that accurately reflect the geometric shape and spatial distribution of the irregular-shaped tank surface.

[0033] The geometric feature extraction module performs 3D reconstruction and feature segmentation on the surface point cloud data to obtain a geometric feature set. During the 3D reconstruction process, the spatial coordinate information of the point cloud data is used to connect discrete points into a continuous surface or mesh using a specific algorithm to construct a 3D model of the irregular-shaped tank. Feature segmentation is then performed to divide the 3D model into areas with different geometric features, such as planes, curved surfaces, corners, and edges. This generates a geometric feature set that contains the geometric feature information for each part of the tank surface.

[0034] The path planning module performs initial welding path planning based on the geometric feature set, generating a set of initial path segments. Based on the shape, position, and size of each geometric feature in the set, combined with the welding process requirements and specifications, a preliminary welding path is planned. For example, for a flat surface, a straight or parallel welding path might be planned; for a curved surface, a path that follows the curvature of the surface might be planned, thus forming a series of initial path segments that together form the initial path segment set.

[0035] The path optimization module performs smoothing optimization and collision detection on the initial path segments to obtain a set of optimized path segments. During the smoothing optimization process, the initial path segments are processed to make them smoother and reduce sudden changes and corners in the path, thereby improving the stability and quality of the welding process. Collision detection is also performed to check whether the path will collide with other parts of the tank or welding equipment, ensuring the feasibility and safety of the path. Through smoothing optimization and collision detection, the initial path segments are adjusted and optimized to obtain a set of optimized path segments.

[0036] The welding instruction generation module generates welding control instructions based on the set of optimized path segments. This module converts the set of optimized path segments into specific instructions that the welding robot can execute, including parameters such as the welding start position, end position, welding speed, welding current, and welding voltage. This module then controls the welding robot to perform welding operations according to the planned path.

[0037] The present invention will be further described below in conjunction with Examples 1 to 5:

[0038] Example 1: The system's geometric feature extraction module includes a point cloud segmentation unit and a feature extraction unit. When processing surface point cloud data, the point cloud segmentation unit must divide it according to a preset grid to obtain a collection of point cloud units. The preset grid here is not fixed, and its size and shape need to be set based on a comprehensive consideration of multiple factors. For example, when the density of surface point cloud data is high, a smaller grid size can be set to achieve more precise point cloud segmentation, ensure that each point cloud unit contains a relatively reasonable number of points, and more accurately reflect the characteristics of the local area. For areas with more complex surface geometries of irregular storage tanks, such as those with many corners or drastic surface changes, the shape and size of the grid may need to be appropriately adjusted to accommodate their complex structures. For example, an irregular grid division method may be used to ensure that each point cloud unit can better cover points with similar geometric features. Through this method of division according to the preset grid, the originally large amount of discrete surface point cloud data is divided into multiple independent point cloud units with a certain spatial range. The collection of these point cloud units provides a more orderly and targeted data structure for subsequent feature extraction, allowing the feature extraction process to be performed based on these small units.

[0039] The feature extraction unit inputs the set of point cloud units generated by the point cloud segmentation unit into a deep learning-based feature extractor, thereby obtaining a set of geometric features. The deep learning-based feature extractor employed here can be a convolutional neural network (CNN) or other network architecture suitable for processing spatial data. This feature extractor requires extensive training data, including point cloud data of various types of irregular-shaped tank surfaces and their corresponding geometric feature annotations. During training, the feature extractor continuously adjusts network parameters to learn to extract effective geometric features from the point cloud units. Once the set of point cloud units is input to the feature extractor, the network performs multiple layers of processing on each point cloud unit. For example, a convolutional neural network (CNN) first extracts features from the input point cloud units through a convolutional layer. The convolution kernel slides across the point cloud unit to extract local features such as shape and texture. Next, a pooling layer reduces the dimensionality of the features, preserving key feature information while reducing data size and improving computational efficiency. Subsequently, multiple convolutional and pooling layers may be used to alternate between layers to extract deeper and more abstract features. During this process, the feature extractor not only extracts local features from each point cloud unit, such as the curvature and flatness of a small area, but also, through the network's deep connections and global receptive field, captures the global feature relationships between point cloud units, such as the connection method between different areas and the overall shape trend. Ultimately, after processing by the feature extractor, the output geometric feature set can comprehensively and accurately describe the surface geometric features of the irregular-shaped storage tank. These features include but are not limited to the shape features, boundary features, and curvature features of each area, providing a rich and accurate geometric information foundation for the subsequent path planning module, enabling path planning to be rationally planned based on these features to meet the path requirements of the welding process.

[0040] In practical applications, effective data exchange and collaboration between the point cloud segmentation unit and the feature extraction unit are required. The point cloud unit set output by the point cloud segmentation unit must be input into the feature extractor in an appropriate format and structure to ensure that the feature extractor can correctly read and process this data. Furthermore, the geometric feature set output by the feature extractor must be understandable and usable by the subsequent path planning module. Therefore, the feature representation and organization must meet the requirements of path planning. For example, each feature in the geometric feature set may need to include information about its location on the tank surface, its type, and its specific parameters, so that the path planning module can use this information to plan an appropriate welding path. Furthermore, to improve the performance of the entire geometric feature extraction module, the preset grid partitioning parameters, the feature extractor network structure, and the training parameters can be optimized based on the actual application scenario and requirements. For example, when dealing with different types of irregular-shaped tanks, the preset grid size and shape may need to be resized to accommodate the different surface characteristics of the tanks. Alternatively, the feature extractor training objectives may need to be adjusted to better meet the requirements of welding path planning based on the precision requirements of the welding process.

[0041] Example 2: The path optimization module consists of a path segment reconstruction unit, a path smoothness calculation unit, a weight correction unit and a path optimization unit. When processing the initial path segment, the path segment reconstruction unit needs to reorganize the segments along the spatial dimension, and then obtain a set of initial path vectors. The spatial dimension here covers all directions in the three-dimensional space. During the reorganization process, it is necessary to operate according to the specific position and direction of the initial path segment in the space. For each initial path segment, it is necessary to determine its starting point and end point coordinates in the three-dimensional space. A vector can be determined through these two points. The direction of the vector represents the extension direction of the path segment, and the length of the vector represents the length of the path segment. In this way, all initial path segments are converted into vector form, thereby forming a set of initial path vectors. This method of converting path segments into a set of vectors can more conveniently process and analyze the path mathematically.

[0042] The path smoothness calculation unit calculates the geometric correlation between adjacent vectors in the initial path vector set to generate a path smoothness matrix. This unit consists of a spatial mapping subunit and a geometric correlation analysis subunit. The spatial mapping subunit maps each vector in the initial path vector set to Euclidean space, thereby obtaining a set of mapped initial path vectors. Euclidean space is a common geometric space in which vector operations and geometric relationships have clear definitions and rules, facilitating mathematical calculations and analysis. This mapping operation converts path vectors, which may have been represented in other spaces, into Euclidean space, enabling more accurate subsequent geometric correlation analysis. The geometric correlation analysis subunit calculates the cosine similarity between adjacent mapped vectors in the mapped initial path vector set, thereby generating a path smoothness matrix composed of multiple path smoothness values. Cosine similarity is a metric used to measure the directional similarity between two vectors, and its value ranges from -1 to 1. When two vectors have exactly the same direction, the cosine similarity is 1; when they have completely opposite directions, the cosine similarity is -1; and when they are perpendicular, the cosine similarity is 0. In path planning, the higher the cosine similarity between adjacent path vectors, the smaller the directional variation and the smoother the path. By calculating the cosine similarity of each pair of adjacent mapping vectors, a series of path smoothness values ​​are obtained. These values ​​are arranged in order to form a path smoothness matrix. This matrix comprehensively reflects the smoothness of the entire set of initial path vectors.

[0043] The weight correction unit performs weighted corrections on the path smoothness matrix based on the spatial relationships between adjacent vectors in the initial set of path vectors, ultimately generating a path constraint matrix. The spatial relationship between adjacent vectors is a comprehensive concept that includes factors such as the distance between adjacent vectors and the angle between them. For example, when the distance between adjacent vectors is close and the angle variation is large, it indicates that the path at that location is experiencing significant path variation and may require more rigorous smoothing, thus requiring a higher weighting. Conversely, when the distance between adjacent vectors is large and the angle variation is small, it indicates that the path is relatively smooth and the weighting can be appropriately reduced. The weight correction unit analyzes and evaluates the spatial relationship between each pair of adjacent vectors and adjusts the weights of the corresponding elements in the path smoothness matrix based on the evaluation results. This adjustment can be performed by using a function or algorithm to convert the various spatial relationship factors into weight coefficients, which are then multiplied by the original path smoothness values ​​to obtain the corrected weights. After this weighted correction, the path smoothness matrix is ​​converted into a path constraint matrix that better meets the actual path constraint requirements and more accurately reflects the smoothness constraints that must be met at different path locations.

[0044] The path optimization unit's task is to fuse the path constraint matrix with the initial set of path vectors through multi-dimensional convolution to obtain optimized path segments. The specific implementation process of this unit is as follows: First, the path constraint matrix is ​​subjected to deep convolution to obtain a path constraint feature matrix. Deep convolution is a common technique in deep learning. Through multiple layers of convolution operations, it can extract deeper and more abstract features from the path constraint matrix, uncovering the complex relationships and patterns implicit in the path constraints. These deep features can more comprehensively describe the path constraints, providing richer information for subsequent path optimization. The initial set of path vectors and the path constraint feature matrix are then input into a gated encoding network to obtain a set of path context vectors. The gated encoding network is a neural network with a gating mechanism that selectively encodes and adjusts the initial path vectors based on the input path constraint feature matrix. The gating mechanism controls which information is retained and which is ignored, allowing the output path context vector to better account for path contextual relationships, namely the connections and influences between previous and subsequent paths. Finally, the set of path context vectors is tensor-stacked to obtain the optimized path segment. Tensor stacking combines multiple path context vectors in a specific order and dimension to form a complete three-dimensional tensor, which represents the optimized path segment. Through this multidimensional convolution fusion process, the initial set of path vectors is optimized within the constraints of the path constraint matrix. The resulting optimized path segment exhibits significant improvements in smoothness and feasibility, better meeting welding process requirements.

[0045] The various units of the path optimization module require close collaboration and data transfer. The initial path vector set output by the path segment reconstruction unit must be accurately transmitted to the path smoothness calculation unit for subsequent smoothness calculation. The path smoothness matrix generated by the path smoothness calculation unit must be corrected by the weight correction unit to obtain an accurate path constraint matrix. The path constraint matrix and the initial path vector set are then input to the path optimization unit for multidimensional convolution fusion to ultimately generate the optimized path segments. Furthermore, to improve the performance of the path optimization module, the parameters and algorithms of each unit can be adjusted and optimized based on the actual welding requirements and the characteristics of special-shaped tanks. For example, during path segment reconstruction, the vector generation method can be adjusted based on the tank shape and welding process requirements. When calculating path smoothness, the cosine similarity calculation method or weight correction rules can be adjusted according to actual needs. During deep convolution processing and gated coding, the neural network structure and parameters can be adjusted to better meet the needs of path optimization.

[0046] Example 3: The welding instruction generation module includes an optimized path dimensionality reduction unit, a path dynamic accuracy calculation unit, an instruction reference center determination unit, an instruction dynamic weight allocation unit, and a control instruction reconstruction unit. The optimized path dimensionality reduction unit is required to perform mean pooling processing on each segment in the set of optimized path segments in the length dimension to obtain a set of optimized path vectors. When performing mean pooling processing in the length dimension, for each optimized path segment, it is necessary to first determine its numerical distribution in the length direction. The specific operation of mean pooling is to take the average value of all data points in the length dimension of the path segment. In this way, the dimensionality of the data is reduced while retaining the main information of the path segment in the length dimension. For example, suppose an optimized path segment has multiple data points in the length dimension, each data point corresponding to a different position or parameter. After mean pooling processing, these data points will be replaced by an average value, thereby simplifying the representation of the path segment in the length dimension to a single value. Multiple such values ​​constitute the set of optimized path vectors. This dimensionality reduction process not only reduces the amount of data and improves the efficiency of subsequent calculations, but also can eliminate noise in the data to a certain extent, so that the optimized path vector can better reflect the overall characteristics of the path.

[0047] The path dynamic accuracy calculation unit calculates the accuracy of each vector in the set of optimized path vectors, thereby generating a set of path accuracy. The unit's specific implementation process is as follows: First, the median vector and standard deviation vector of the optimized path vectors are calculated. The median vector is a vector formed by taking the median of the values ​​in each dimension of the set of optimized path vectors. It reflects the central tendency of the set in each dimension and, compared to the mean, is more effective in avoiding the influence of extreme values. The standard deviation vector is a vector formed by calculating the standard deviation of the values ​​in each dimension. It reflects the degree of dispersion of the vectors in the set in each dimension. Next, the optimized path vector is element-wise subtracted from the median vector and the resulting difference is squared to obtain a path difference vector. Element-wise subtraction involves subtracting the value in each dimension of each optimized path vector from the value in the corresponding dimension of the median vector to obtain the difference in each dimension. These differences are then squared to amplify larger differences and make them more apparent. After that, the overall median value of the path difference vector is calculated. The median value is calculated by sorting the values ​​of all elements in the path difference vector and taking the value in the middle position. It can also avoid the influence of extreme values ​​on the results and more robustly reflect the overall level of the path difference vector. Finally, the median value is multiplied by the square value of the standard deviation vector, and the result is input into the normalization function to obtain the accuracy value. Let the optimized path vector set be ,in Indicates the optimization path vectors, each of which has dimensions. Median vector No. The dimensions are ,in For the The first vector Dimension values. Standard deviation vector No. The dimensions are Path difference vector No. The dimensions are The overall median value is .Precision value The calculation formula is:

[0048]

[0049] in, It is the maximum value in the normalization process, which is used to map the result to an appropriate range. Through such calculation, the accuracy value of each optimized path vector is obtained. These accuracy values ​​constitute the path accuracy set, which reflects the accuracy of each vector.

[0050] The instruction reference center determination unit selects the optimized path vector with the maximum precision value in the path precision set as the initial instruction center vector. Each optimized path vector in the path precision set has a corresponding precision value. A larger precision value indicates a higher precision. Using this vector as the initial instruction center vector provides a more reliable reference point for subsequent instruction generation, ensuring that subsequent instruction generation is based on the more accurate path vector.

[0051] The instruction dynamic weight allocation unit calculates the dynamic instruction weight for each vector in the set of optimized path vectors based on the geometric distance between each vector and the initial instruction center vector and the precision of each vector, thereby generating a dynamic weight set. This unit is specifically implemented as follows: First, the precision of the optimized path vector and the precision of the initial instruction center vector are multiplied by a first correction coefficient to obtain a first dynamic instruction factor. The first correction coefficient is an adjustable parameter whose value can be determined based on actual welding process requirements and experience. It is used to adjust the influence of the precision on the dynamic instruction weight, so that the proportion of the precision in the weight calculation meets actual requirements. Then, the Euclidean distance between the optimized path vector and the initial instruction center vector is multiplied by a second correction coefficient to obtain a second dynamic instruction factor. The second correction coefficient is also an adjustable parameter that adjusts the influence of geometric distance on the dynamic instruction weight. The Euclidean distance measures the actual distance between two vectors in space. The closer the distance, the higher the correlation between the optimized path vector and the initial instruction center vector. Finally, the first dynamic instruction factor is multiplied by the second dynamic instruction factor to obtain the dynamic instruction weight. Through such calculation, the accuracy value of the optimized path vector and the geometric distance from the initial instruction center vector are comprehensively considered to obtain the dynamic instruction weight of each optimized path vector, and these weights constitute the dynamic weight set.

[0052] The control instruction reconstruction unit uses a dynamic weight set to perform weighted fusion on the set of optimized path vectors to generate welding control instructions. The weighted fusion process multiplies each optimized path vector by the corresponding dynamic instruction weight, and then adds all the product results to obtain the final welding control instruction. This weighted fusion method ensures that optimized path vectors with high precision and close distance to the initial instruction center vector occupy a larger proportion in the welding control instruction, so that the generated welding control instruction can comprehensively consider the information of all optimized path vectors and tend to use accurate path vectors as a reference, ensuring that the welding control instruction can more accurately guide the welding operation of the welding robot, so that the welding robot can weld according to more reasonable paths and parameters.

[0053] The various units in the welding instruction generation module require precise data transmission and collaborative work. The optimized path vector set output by the optimized path dimensionality reduction unit must be accurately transmitted to the path dynamic accuracy calculation unit for accuracy calculation. The path accuracy set generated by the path dynamic accuracy calculation unit and the initial instruction center vector determined by the instruction reference center determination unit must be provided to the instruction dynamic weight allocation unit for dynamic instruction weight calculation. The dynamic weight set and optimized path vector set generated by the instruction dynamic weight allocation unit are then input to the control instruction reconstruction unit for weighted fusion to generate the final welding control instructions. Furthermore, to ensure that the generated welding control instructions are more consistent with actual welding process requirements, the parameters in each unit can be adjusted based on the different special-shaped tank structures and welding process parameters. For example, the parameters used in the mean pooling process, the values ​​of the first and second correction coefficients, and the maximum value of the normalization function can be adjusted to adapt to different welding scenarios. This ensures that the generated welding control instructions can accurately control the movements of the welding robot, achieve efficient and high-quality welding operations, and meet the various complex requirements of special-shaped tank welding.

[0054] Example 4: During implementation, the welding instruction generation module inputs welding control instructions into a decision-tree-based instruction signal generator, which then generates a welding execution control signal that indicates the start and stop strategy for the welding process. This example uses a specific welding scenario for a special-shaped storage tank as an example. Assume that the tank has a complex curved surface structure with a hemispherical top, cylindrical sides, and a conical bottom. Different start and stop strategies are required for circumferential and longitudinal seam welding to ensure welding quality and efficiency.

[0055] The welding control instructions contain various parameter information corresponding to the set of optimized path segments, such as the starting and ending coordinates, welding speed, welding current, and welding voltage for each optimized path segment. These parameters are determined based on the tank's geometric characteristics, welding process requirements, and path optimization results. For example, for an optimized path segment for a circular weld along a cylindrical area on the side of a tank, the starting coordinates might be at a specific angle on the side of the cylinder, and the ending coordinates would be the corresponding position after one full circle around the cylinder. The welding speed is set at 5 mm / s, the welding current is 200 amps, and the welding voltage is 25 volts.

[0056] The command signal generator based on a decision tree is a trained model whose decision tree is constructed based on a large amount of welding process data and actual welding cases. Each node in the decision tree represents a judgment of a characteristic attribute, such as the length of the welding path, the combination of welding parameters, the current state of the welding robot, etc.; each branch represents a judgment result; and each leaf node corresponds to the final start-stop strategy decision. During the training process, the decision tree needs to be provided with a large amount of input data. This data includes welding control command parameters for different tank structures and the corresponding optimal start-stop strategies. Through training, the decision tree learns the mapping relationship between parameters and strategies.

[0057] When a welding control command is input into the command signal generator, the decision tree begins at the root node and makes decisions at each node based on the parameters in the welding control command. Taking the girth weld path segment mentioned earlier as an example, the decision tree first determines the length of the path segment. Assuming the girth weld path segment is 3 meters long, a node in the decision tree determines whether the path length is greater than 2.5 meters. Since 3 meters is greater than 2.5 meters, the corresponding branch is entered. The tree then determines whether the welding speed is within the specified range. The welding speed for this path segment is 5 mm / s, which is within the specified range, so the tree proceeds to the next branch. The tree then determines whether the welding current and voltage combination meet the welding requirements for the tank material. A combination of 200 amps and 25 volts does meet the requirements, so the tree proceeds to the next branch. At this point, the tree may also consider whether the welding robot's current position is aligned with the starting position of the path segment. Assuming the welding robot has moved to the correct starting position and is functioning normally, the tree reaches a leaf node and generates the corresponding start strategy, which immediately initiates welding.

[0058] During the welding process, when one path segment is completed, the welding control command includes the parameters for the next path segment, for example, switching from the side circumferential weld to the longitudinal weld path segment in the bottom tapered area. This longitudinal weld path segment starts at the bottom edge of the cone and ends at the top. Its length is 1.5 meters, and the welding speed is adjusted to 4 mm / s, the welding current to 180 amps, and the welding voltage to 22 volts. After these parameters are input into the command signal generator, the decision tree re-evaluates. First, it determines if the path length is 1.5 meters, which is less than 2.5 meters, and then branches to another branch. It then determines whether the welding speed, current, and voltage combination meet the requirements for welding in the tapered area. If so, it determines whether the welding robot needs to adjust its posture to accommodate the tapered structure. If the robot needs to adjust its arm angle, but the adjustment is not yet complete, the decision tree will generate a pause strategy based on this state, waiting for the robot to complete the adjustment before continuing welding.

[0059] The decision tree also needs to consider handling abnormal situations. For example, if a parameter in a welding control instruction exceeds the normal range, such as the welding current suddenly increasing to 300 amperes, the decision tree will identify this as an abnormality, generate a strategy to stop welding, and issue an alarm signal so that the operator can promptly inspect and address the problem. Alternatively, if the surface temperature of a storage tank is detected to be too high, potentially affecting weld quality, the decision tree will generate a strategy to suspend welding until the temperature returns to a reasonable range.

[0060] The command signal generator based on the decision tree needs to be flexibly adjusted and optimized based on the different special-shaped tank structures and welding process requirements. For example, the welding parameter requirements for tanks made of different materials, such as stainless steel and carbon steel, vary, and the judgment conditions and weights in the decision tree need to be adjusted accordingly. For complex tank structures, such as areas with bosses or grooves, the decision tree needs to add more judgment nodes to handle these special cases. Furthermore, as welding experience accumulates and new data is added, the decision tree model needs to be continuously updated and trained to improve its decision accuracy and adaptability.

[0061] Example 5: The system also includes a welding execution module, which consists of a welding robot and a position sensor. For example, consider welding a special-shaped storage tank at a chemical company. The tank has a cylindrical main body, an ellipsoidal head at the top, and a conical discharge port at the bottom. The tank also has multiple nozzles and bosses on the surface. The welding execution module must complete the welding of circumferential seams, longitudinal seams, and nozzle fillet welds on this complex structure.

[0062] The welding robot is equipped with a multi-axis robotic arm. Taking a six-axis robotic arm as an example, the degrees of freedom of its six joints enable it to flexibly adjust its posture in three-dimensional space. When welding the circumferential seam of the ellipsoidal head on the top of the tank, the robotic arm needs to start from the connection between the head and the cylindrical body and move along the circumferential seam trajectory according to the set of optimized path segments. For example, the initial optimized path segment instruction requires the welding gun at the end of the robotic arm to move at a speed of 8 mm per second, maintaining a posture perpendicular to the weld surface, and the welding current is set to 220 amperes. After receiving the welding control instruction, the robotic arm accurately executes the position and posture control of the welding gun through the coordinated movement of each joint, ensuring that the welding gun moves along the planned path to complete the welding of this section of the weld.

[0063] Position sensors are installed on the tank surface and can be laser displacement sensors or vision sensors. For example, when welding the longitudinal seam of the conical discharge port at the bottom of the tank, the laser displacement sensor is fixed close to the weld seam and collects real-time welding position data at a rate of 100 times per second. As the welding robot moves along the longitudinal seam, the sensor continuously measures the actual position of the welding gun relative to the weld seam. Suppose at a certain moment, the theoretical position of the optimized path segment requires the welding gun to be at a point on the conical surface (X=100mm, Y=200mm, Z=300mm), while the actual position recorded by the sensor is (X=102mm, Y=201mm, Z=300mm). At this point, the system performs a position correction between the actual position data and the set of optimized path segments. By calculating the deviation (+2mm in the X direction and +1mm in the Y direction), a real-time control feedback signal is generated and transmitted to the welding robot controller. The controller adjusts the motion parameters of the robotic arm to return the welding gun to the theoretical position and resume welding according to the planned path.

[0064] Real-time control feedback is transmitted to the central controller via an industrial bus, such as EtherCAT, which features high-speed data transmission and enables millisecond-level real-time communication. During the aforementioned tapered discharge port longitudinal seam welding process, the central controller receives real-time feedback from the position sensor, displaying parameters such as the current welding position, welding speed, and welding current. If the welding gun position deviation exceeds a preset threshold (e.g., an X or Y deviation exceeding 3mm), the central controller issues an alert and automatically adjusts the robot arm's motion. If the welding current fluctuates abnormally (e.g., exceeding ±10% of the set value), the central controller triggers a command to pause welding until the operator has inspected the equipment before continuing.

[0065] When welding the fillet welds between the tank pipe and the main body, the collaborative work of the welding execution module is even more critical. The connection between the pipe and the main body is a spatial curved surface, and the optimized path segment is circular around the root of the pipe. The multi-axis robotic arm of the welding robot needs to adjust its position and posture at the same time so that the welding gun always maintains the best welding angle (such as 45°) with the fillet weld. The position sensor monitors the position of the welding gun on the circular path in real time. For example, in a certain section of the path, the sensor detects that the distance between the welding gun and the pipe surface changes due to the motion error of the robotic arm. The central controller adjusts the robotic arm joint angle in real time based on the feedback data to ensure that the distance between the welding gun and the weld remains within the range of 5±1mm.

[0066] The position sensors in the welding execution module also require calibration with fixed points on the tank surface. For example, before welding, the operator manually teaches the welding robot's welding gun to touch three known coordinate points on the tank surface. The system then establishes a global coordinate system, aligning the data collected by the position sensors with the coordinate system of the optimized path segment. Once calibrated, the position sensors always use this global coordinate system as a reference when the welding robot moves across the tank surface, ensuring accurate weld position.

[0067] The welding execution module must handle complex path switching for the boss structures on the tank surface. For example, when the weld path transitions from the cylindrical surface of the tank body to the boss surface, the optimized path segment includes instructions for path turning. The welding robot adjusts the arm posture in advance according to these instructions. Position sensors monitor the actual position of the turning point in real time. The central controller receives feedback via the industrial bus to ensure a smooth transition of the welding trajectory at the turning point, avoiding weld deviation or incomplete penetration.

[0068] Throughout the welding process, the central controller, through a monitoring system built on an industrial bus, records welding parameters and position data in real time, generating a welding process log. For example, this log records the start and end time of each optimized path segment, the actual welding speed, and current and voltage fluctuations. This data can be used for subsequent welding quality traceability. If a weld defect is discovered, the operator can use the log to analyze parameter changes and position deviations during the welding process to pinpoint the cause.

[0069] The multi-axis robotic arm in the welding execution module also features collision protection. If, due to programming errors or sensor failure, the welding torch collides with the tank surface during movement, the arm's force sensor immediately detects the abnormal resistance and triggers an emergency stop signal, which is transmitted to the central controller via the industrial bus, preventing equipment damage and safety incidents.

[0070] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0071] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Special-shaped tank welding path planning system, characterized by: include: Surface data acquisition module, used to obtain surface point cloud data of special-shaped storage tanks through laser scanner; A geometric feature extraction module, configured to perform three-dimensional reconstruction and feature segmentation on the surface point cloud data to obtain a geometric feature set; a path planning module, configured to perform initial welding path planning based on the set of geometric features to obtain a set of initial path segments; a path optimization module, configured to perform smoothing optimization and collision detection on the set of initial path segments to obtain a set of optimized path segments; A welding instruction generation module for generating welding control instructions according to the set of optimized path segments; The path optimization module includes: a path segment reconstruction unit for reorganizing the initial path segments along the spatial dimension to obtain a set of initial path vectors; a path smoothness calculation unit for calculating the geometric correlation between adjacent vectors in the set of initial path vectors to generate a path smoothness matrix; a weight correction unit for performing weight distribution correction on the path smoothness matrix based on the spatial relationship between adjacent vectors in the set of initial path vectors to obtain a path constraint matrix; and a path optimization unit for performing multi-dimensional convolution fusion on the path constraint matrix and the set of initial path vectors to obtain the optimized path segment. The geometric feature extraction module includes: a point cloud segmentation unit for dividing the surface point cloud data into a preset grid to obtain a set of point cloud units; a feature extraction unit for inputting the set of point cloud units into a deep learning-based feature extractor to obtain the geometric feature set; The path smoothness calculation unit includes: a spatial mapping subunit for mapping each vector in the set of initial path vectors to Euclidean space to obtain a set of mapped initial path vectors; a geometric association analysis subunit for calculating the cosine similarity between adjacent mapped vectors in the set of mapped initial path vectors to generate the path smoothness matrix composed of multiple path smoothness values; The path optimization unit is specifically implemented as follows: performing deep convolution processing on the path constraint matrix to obtain a path constraint feature matrix; inputting the set of initial path vectors and the path constraint feature matrix into a gated encoding network to obtain a set of path context vectors; performing tensor stacking on the set of path context vectors to obtain the optimized path segment; The welding instruction generation module includes: an optimized path dimension reduction unit, which is used to perform mean pooling processing on each segment in the set of optimized path segments in the length dimension to obtain a set of optimized path vectors; a path dynamic accuracy calculation unit, which is used to calculate the accuracy value of each vector in the set of optimized path vectors to generate a path accuracy set; an instruction reference center determination unit, which is used to select the optimized path vector corresponding to the maximum accuracy value in the path accuracy set as the initial instruction center vector; an instruction dynamic weight allocation unit, which is used to calculate the dynamic instruction weight of each vector in the set of optimized path vectors and the initial instruction center vector according to the geometric distance between each vector and the initial instruction center vector and the accuracy value of each vector to generate a dynamic weight set; a control instruction reconstruction unit, which is used to perform weighted fusion on the set of optimized path vectors using the dynamic weight set to generate the welding control instruction; The path dynamic accuracy calculation unit is specifically implemented as follows: calculating the median vector and the standard deviation vector of the optimized path vector; performing element-by-element difference calculation between the optimized path vector and the median vector, and squaring the difference result to obtain a path difference vector; calculating the overall median value of the path difference vector; performing a product operation on the median value and the square value of the standard deviation vector, and inputting the result into a normalization function to obtain the accuracy value; The instruction dynamic weight allocation unit is specifically implemented as follows: multiplying the accuracy value of the optimized path vector and the accuracy value of the initial instruction center vector by a first correction coefficient to obtain a first dynamic instruction factor; multiplying the Euclidean distance between the optimized path vector and the initial instruction center vector by a second correction coefficient to obtain a second dynamic instruction factor; and performing a product operation on the first dynamic instruction factor and the second dynamic instruction factor to obtain the dynamic instruction weight.

2. The special-shaped tank welding path planning system according to claim 1 is characterized in that: The welding instruction generation module is specifically implemented as follows: inputting the welding control instruction into the instruction signal generator based on the decision tree to obtain a welding execution control signal, and the welding execution control signal is used to indicate the start and stop strategy of the welding process.

3. The special-shaped tank welding path planning system according to claim 2 is characterized in that: It also includes a welding execution module, which is composed of a welding robot and a position sensor, wherein: the welding robot is equipped with a multi-axis robotic arm for executing the welding control instructions; the position sensor is arranged on the surface of the tank, for collecting welding position data in real time, and generating real-time control feedback after position correction with the set of optimized path segments; the real-time control feedback is transmitted to the central controller via the industrial bus to complete the welding process monitoring.

Images