A method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber

Abstract

The present invention discloses a method for reducing the inclusion rate of metal particles in additive manufacturing of aerospace propulsion chambers, comprising the following steps: Step 1: constructing a print layer dataset; Step 2: constructing an initial multi-scale graph structure; Step 3: outputting an inclusion risk probability map of the current print layer based on a multi-scale graph growth generative network; Step 4: inputting the inclusion risk probability map into a mimicking hummingbird optimization algorithm to generate a path optimization target space; Step 5: performing a path search using the mimicking hummingbird optimization algorithm and outputting a final scanning path; Step 6: controlling a laser to perform a print scanning operation based on the final scanning path; Step 7: inputting process feedback data into a multi-scale graph growth generative network and outputting an updated inclusion risk probability map for the next print layer; Step 8: repeating steps 4 to 7 until all print layers have completed printing operations. The present invention integrates a multi-scale graph growth generative network with the mimicking hummingbird optimization algorithm to effectively reduce the inclusion rate of metal particles.

Description

Technical Field

[0001] The present invention relates to the technical field of metal additive manufacturing, and in particular to a method for reducing the inclusion rate of metal particles in additive manufacturing of aerospace propulsion chambers. Background Art

[0002] With the widespread application of metal additive manufacturing technology in the aerospace field, especially the growing demand for integrated printing of complex propulsion chamber structures, how to reduce inclusion defects and improve material density and mechanical properties during the additive manufacturing process has become a research focus. Existing quality control methods for aerospace propulsion chamber additive manufacturing mainly rely on fixed scanning path design and post-printing inspection methods to identify and repair defects such as inclusions. However, the following problems are common in practical applications:

[0003] The scanning path fails to fully consider the spatial evolution laws of material powder distribution, thermal field changes during the printing process, and inclusion generation, resulting in no targeted control of inclusion risk areas in path planning, which easily induces metal particle aggregation and non-uniform fusion in complex structural areas; existing path optimization methods are mostly based on static graphic analysis or heuristic strategies, lack the ability to dynamically model voxel-level risk areas, and are difficult to adapt to the real-time evolution of cross-hierarchical structural complexity and layer-by-layer printing status; at the same time, the printing process monitoring data is not effectively integrated into the path planning process, and traditional control methods find it difficult to establish an effective feedback path between melting behavior and inclusion risk, resulting in delayed control system response and coarse adjustment granularity, which limits the active avoidance of inclusion risks and real-time optimization capabilities.

[0004] Therefore, how to provide a method for reducing the inclusion rate of metal particles in additive manufacturing of aerospace propulsion chambers is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of the present invention is to propose a method for reducing metal particle inclusions in additive manufacturing of aerospace propulsion chambers. This method integrates a multi-scale graph neural network with a mimicking hummingbird optimization algorithm to construct a dynamic adjustment mechanism for the printing path, targeting inclusion risk. This mechanism enables precise prediction and avoidance control of inclusion risk areas during metal additive manufacturing. By iteratively updating the graph structure using feedback data from the printing process, the adaptive capability of path planning and printing stability are enhanced. This method offers the advantages of high prediction accuracy, strong path continuity, and significant inclusion control. It is particularly suitable for improving quality and implementing closed-loop control during the additive manufacturing process of complex aerospace components with high density and high reliability requirements.

[0006] A method for reducing metal particle inclusions in additive manufacturing of a space propulsion chamber according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain a 3D model of the space propulsion chamber component to be printed and slice it to obtain 2D cross-sectional information of each printing layer and construct a printing layer dataset;

[0008] Step 2: constructing an initial multi-scale graph structure based on the printed layer dataset;

[0009] Step 3: Input the initial multi-scale graph structure into a multi-scale graph growth generation network, combine it with historical printing data to evolve the graph structure, and output the inclusion risk probability map of the current printing layer;

[0010] Step 4: Using the inclusion risk probability map as input to the mimic hummingbird optimization algorithm, constructing a hummingbird individual path population, and setting the low-risk areas in the inclusion risk probability map as target path areas, and the high-risk areas as disturbance control areas, to generate a path optimization target space;

[0011] Step 5: Based on the path optimization target space, the mimic hummingbird optimization algorithm is used to search the path and output the final scanning path;

[0012] Step 6: Control the laser to perform a print scanning operation of the current layer according to the final scanning path, and collect the molten pool boundary contour, thermal image distribution data and printing layer status information in real time to generate process feedback data;

[0013] Step 7: Input the process feedback data into a multi-scale graph growth generation network, combine it with the inclusion risk probability map of the current printing layer, and output an updated inclusion risk probability map for the next printing layer;

[0014] Step 8: Repeat steps 4 to 7 until all printing layers are completed, thereby achieving control over the reduction of inclusions in the aerospace propulsion chamber components during the printing process.

[0015] Optionally, the two-dimensional cross-sectional information of the printed layer includes the outline boundary coordinates, inner outline structural features, pore distribution positions, scanning area segmentation information, and connection topological relationships between the printed layer and adjacent upper and lower layers.

[0016] Optionally, the constructing of the initial multi-scale graph structure is specifically as follows:

[0017] Setting each voxel unit in the printing layer data set as a node in a graph structure, wherein the node attributes include the spatial coordinates of the voxel unit, the distribution state of metal particles, the impurity information of metal particles, the initial temperature value, the laser scanning energy, the local geometric characteristics and the interlayer residual stress information;

[0018] Establish edge connections between nodes with spatial adjacency, and the edge weight is the Euclidean distance between the nodes;

[0019] A two-dimensional intra-layer graph structure is constructed based on node attributes and edge connections, and cross-layer edge connections are constructed through the topological mapping relationship of voxel nodes between the upper and lower printing layers to form an initial multi-scale graph structure.

[0020] Optionally, the step three: inputting the initial multi-scale graph structure into a multi-scale graph growth generation network, performing graph structure evolution in combination with historical printing data, and outputting an inclusion risk probability map of the current printing layer, specifically:

[0021] The multi-scale graph growth generation network includes a graph feature encoding module, a graph structure evolution module, a node state update module and a risk probability output module, which is used to dynamically model the inclusion generation behavior in the metal additive manufacturing process;

[0022] The graph feature encoding module receives the initial multi-scale graph structure, performs feature embedding and dimension normalization on node attributes, and generates a node state vector;

[0023] The graph structure evolution module uses the graph structures and evolution trajectories of each layer contained in the historical printing data as supervision data, adopts a graph convolutional network to aggregate the state vectors between adjacent nodes, and combines edge weight embedding values ​​for information transmission. It simulates the diffusion trend and structural aggregation behavior caused by thermal disturbances during the inclusion formation process through iterative propagation.

[0024] The node state update module updates the state vector of each node after each graph structure evolution, introduces the evolution trajectory of the node at the same spatial position in the historical printing process as a time series input, and splices it with the current node state vector;

[0025] The risk probability output module performs probability modeling on all updated node state vectors, outputs the inclusion risk probability value of the node corresponding to the current printing layer through a multi-layer perceptron, and reconstructs the inclusion risk probability map of the printing layer according to the spatial position of the voxel unit;

[0026] The node set whose mixed risk probability value is less than a preset low risk threshold is defined as a low risk area, and the node set whose mixed risk probability value is greater than a preset high risk threshold is defined as a high risk area;

[0027] The inclusion risk probability map is a probability distribution map corresponding to the printing layer space, and is used to characterize the inclusion formation risk area at the voxel level.

[0028] Optionally, the inclusion risk probability map is used as the input of the mimic hummingbird optimization algorithm to construct a hummingbird individual path population, specifically: according to the inclusion risk probability map, the printing layer is divided into multiple path candidate blocks using a spatial grid division method, and multiple initial path individuals are generated in each path candidate block. The initial path individual starts from the printing starting point and ends at the boundary of the path candidate block. The path shape is a continuous multi-segment vector trajectory, each segment of the vector trajectory corresponds to a scanning step and a direction angle, and the hummingbird individual path population is constructed based on the initial path individual set generated in multiple path candidate blocks.

[0029] Optionally, the step five: performing path search using a mimic hummingbird optimization algorithm based on the path optimization target space, and outputting a final scanning path, specifically including a forward flight strategy, a hovering perturbation strategy, and a regression strategy;

[0030] The forward flight strategy is to use a set of directions within a preset step size and direction angle range as candidate path directions in a low-risk area based on the starting point of the current hummingbird individual path, calculate the average inclusion risk probability value of the area corresponding to each candidate path direction, select the direction with the smallest average inclusion risk probability value and the continuity of the scanning path, and generate a main path segment as the backbone of the hummingbird individual path;

[0031] The hovering perturbation strategy triggers a path perturbation mechanism at the boundary of a high-risk area, generating multiple local perturbation path segments within the current position and its neighborhood with a set perturbation angle range and perturbation step size. The local perturbation path segments form a zigzag shape within a two-dimensional cross-section. The local perturbation path segments are sequentially spliced ​​with the current main path segments to form a perturbation extension segment of the path.

[0032] The regression strategy is to perform a path reversal operation when the average inclusion risk probability value of all candidate path segments in the path advancement direction is greater than a preset upper threshold, or the path continuity is interrupted, rewind the current hummingbird individual path to the previous node, reselect the direction set and evaluate the new forward path;

[0033] In each round of optimization iteration, the fitness value of the hummingbird individual path is calculated, which is calculated based on a weighted combination of the path length and the inclusion risk probability value of the nodes along the path;

[0034] When the number of optimization iterations reaches the preset maximum value, or the fitness value improvement rate between two consecutive iterations is lower than the preset change threshold, the iteration is terminated and the final scanning path is output;

[0035] The final scanning path is formed by sequentially splicing the main path segments and the disturbance path segments generated in each path candidate block according to the search execution order.

[0036] Optionally, the step six: controlling the laser to perform a print scanning operation on the current layer according to the final scanning path, and collecting the molten pool boundary contour, thermal image distribution data, and print layer status information in real time to generate process feedback data, specifically:

[0037] Sending a scanning trajectory instruction to the laser control system according to the final scanning path, controlling the laser to scan in sequence along the spliced ​​path segments to process the metal powder material of the current printing layer;

[0038] During the scanning process, the melt pool contour detection device, infrared thermal imaging sensor and layer status imaging unit set above or on the side of the printing area are used to synchronously observe the scanning area, collect the melt pool boundary contour, thermal image distribution data and printing layer status information in real time, and synchronously record them in time series. The data is bound to the corresponding scanning path segment to form a process feedback data set.

[0039] Optionally, the seventh step is to input the process feedback data into a multi-scale graph growth generation network, combine it with the inclusion risk probability map of the current printing layer, and output an updated inclusion risk probability map of the next printing layer, specifically:

[0040] Inputting the process feedback data set into a multi-scale graph growth generative network, performing feature extraction and fusion processing on the melt pool boundary contour, thermal image distribution data, and printing layer status information in the process feedback data set to form a node feature update vector reflecting the actual printing status of the current printing layer;

[0041] The node feature update vector is concatenated with the corresponding node state vector in the inclusion risk probability graph of the current printing layer, and the concatenated vector is input into the graph structure evolution module and the node state update module of the multi-scale graph growth generation network. The updated graph node state vector is obtained through graph convolution operation and state iterative update.

[0042] The updated graph node state vector is input into the risk probability output module, and the inclusion risk probability value of the node corresponding to the next printing layer is output through probability modeling calculation, thereby forming an updated inclusion risk probability graph for the printing and scanning operation of the next printing layer.

[0043] The beneficial effects of the present invention are:

[0044] This paper addresses the issues of inaccurate inclusion risk area identification, unadjustable paths, and laser scanning control lag during the additive manufacturing of aerospace propulsion chambers by integrating a multiscale graph growth generative network with a mimetic hummingbird optimization algorithm. By employing multiscale graph structure modeling and inclusion risk probability distribution prediction for print layer voxel nodes, this paper combines historical print evolution trajectories with process feedback data to construct a dynamic graph evolution mechanism. During the path generation phase, a bionic path search strategy combining target guidance in low-risk areas with disturbance avoidance in high-risk areas is introduced, enabling the scanning path to actively avoid high-inclusion areas and ablate thermal disturbances. During the path execution phase, multi-source monitoring data is collected to construct a spatiotemporally aligned process feedback dataset. Node state vector updates and graph convolution propagation mechanisms are used to iteratively correct the inclusion risk prediction graph layer by layer, improving the timeliness of risk modeling and the response accuracy of path planning. Ultimately, this approach enables real-time modeling, prediction updates, and path-level control closed-loop feedback for metal particle inclusion formation during the printing process, effectively reducing the probability of inclusion aggregation and improving the forming quality and microstructural uniformity of aerospace components. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0046] Figure 1 This is an overall flow chart of a method for reducing metal particle inclusions in additive manufacturing of aerospace propulsion chambers proposed by the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of a multi-scale graph growth generation network for a method of reducing metal particle inclusions in additive manufacturing of aerospace propulsion chambers proposed by the present invention;

[0048] Figure 3 This is a flowchart of the execution of the mimicking hummingbird optimization algorithm for the method of reducing the inclusion rate of metal particles in additive manufacturing of aerospace propulsion chambers proposed by the present invention. DETAILED DESCRIPTION

[0049] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0050] refer to Figure 1-Figure 3 A method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chambers comprises the following steps:

[0051] Step 1: Obtain a 3D model of the space propulsion chamber component to be printed and slice it to obtain 2D cross-sectional information of each printing layer and construct a printing layer dataset;

[0052] Step 2: constructing an initial multi-scale graph structure based on the printed layer dataset;

[0053] Step 3: Input the initial multi-scale graph structure into a multi-scale graph growth generation network, combine it with historical printing data to evolve the graph structure, and output the inclusion risk probability map of the current printing layer;

[0054] Step 4: Using the inclusion risk probability map as input to the mimic hummingbird optimization algorithm, constructing a hummingbird individual path population, and setting the low-risk areas in the inclusion risk probability map as target path areas, and the high-risk areas as disturbance control areas, to generate a path optimization target space;

[0055] Step 5: Based on the path optimization target space, the mimic hummingbird optimization algorithm is used to search the path and output the final scanning path;

[0056] Step 6: Control the laser to perform a print scanning operation of the current layer according to the final scanning path, and collect the molten pool boundary contour, thermal image distribution data and printing layer status information in real time to generate process feedback data;

[0057] Step 7: Input the process feedback data into a multi-scale graph growth generation network, combine it with the inclusion risk probability map of the current printing layer, and output an updated inclusion risk probability map for the next printing layer;

[0058] Step 8: Repeat steps 4 to 7 until all printing layers are completed, thereby achieving control over the reduction of inclusions in the aerospace propulsion chamber components during the printing process.

[0059] The present invention provides a holistic control process for inclusion issues in the additive manufacturing of metal aerospace propulsion chambers, constructing a fully closed-loop path optimization system from three-dimensional modeling, inclusion risk modeling, path generation, printing control, to feedback regulation. By deeply coupling inclusion risk prediction with path generation, the traditional model of static printing path design and post-quality inspection response is broken, and real-time assessment and active avoidance of inclusion risks in the additive manufacturing process are achieved. It has the advantages of process adaptation, refined spatial modeling, and intelligent path control. It is suitable for printing tasks of complex geometric components and aerospace structures with high reliability requirements. It can effectively improve the consistency and structural uniformity of formed parts, significantly reduce the inclusion defect rate and rework costs, and has significant engineering practicality and system scalability.

[0060] In this embodiment, the two-dimensional cross-sectional information of the printed layer includes the outline boundary coordinates, inner contour structural features, pore distribution positions, scanning area segmentation information, and connection topological relationships between adjacent upper and lower layers of the printed layer.

[0061] By slicing and analyzing the component to be printed, extracting the cross-sectional structural features of each layer and constructing a printed layer dataset, this not only lays a geometric foundation for mapping but also significantly enhances the structural sensitivity of risk modeling. Cross-sectional information integrates information from multiple dimensions, enabling more fine-grained risk zoning and path region identification. Furthermore, the cross-sectional structure ensures the complete representation of the component's local structure, facilitating dynamic perception of critical areas in multi-scale modeling, thereby improving path planning's responsiveness to microstructural changes. This approach is particularly applicable to the printing of aerospace components involving multi-layer transitions, complex channels, or heterogeneous topologies.

[0062] In this embodiment, the initial multi-scale graph structure is constructed as follows:

[0063] Setting each voxel unit in the printing layer data set as a node in a graph structure, wherein the node attributes include the spatial coordinates of the voxel unit, the distribution state of metal particles, the impurity information of metal particles, the initial temperature value, the laser scanning energy, the local geometric characteristics and the interlayer residual stress information;

[0064] Establish edge connections between nodes with spatial adjacency, and the edge weight is the Euclidean distance between the nodes;

[0065] A two-dimensional intra-layer graph structure is constructed based on node attributes and edge connections, and cross-layer edge connections are constructed through the topological mapping relationship of voxel nodes between the upper and lower printing layers to form an initial multi-scale graph structure.

[0066] By introducing spatial voxel units as graph structure nodes, and using the spatial coordinates of voxel units, metal particle distribution status, metal particle impurity information, initial temperature value, laser scanning energy, local geometric characteristics and interlayer residual stress information as node attributes, a multi-scale graph structure with a high degree of physical reality and strong correlation with printing behavior is established. It accurately depicts the state evolution characteristics of metal particles during the printing process and provides a scalable spatial semantic foundation for inclusion risk modeling. Through cross-layer edge connections, the timing and structural coupling between printed layers is enhanced, allowing the prediction model to perceive the spatial cumulative effect of inclusions between upper and lower layers, providing key support for the evolution and path decision-making of graph neural networks. Compared with the traditional modeling method based on single-layer features, this step has higher modeling accuracy and more complete timing dependence.

[0067] In this embodiment, the step three is: inputting the initial multi-scale graph structure into the multi-scale graph growth generation network, combining the historical printing data to evolve the graph structure, and outputting the inclusion risk probability map of the current printing layer, specifically:

[0068] The multi-scale graph growth generation network includes a graph feature encoding module, a graph structure evolution module, a node state update module and a risk probability output module, which is used to dynamically model the inclusion generation behavior in the metal additive manufacturing process;

[0069] The graph feature encoding module receives the initial multi-scale graph structure, performs feature embedding and dimension normalization on node attributes, and generates a node state vector;

[0070] The graph structure evolution module uses the graph structures and evolution trajectories of each layer contained in the historical printing data as supervision data, adopts a graph convolutional network to aggregate the state vectors between adjacent nodes, and combines edge weight embedding values ​​for information transmission. It simulates the diffusion trend and structural aggregation behavior caused by thermal disturbances during the inclusion formation process through iterative propagation.

[0071] The node state update module updates the state vector of each node after each graph structure evolution, introduces the evolution trajectory of the node at the same spatial position in the historical printing process as a time series input, and splices it with the current node state vector;

[0072] The risk probability output module performs probability modeling on all updated node state vectors, outputs the inclusion risk probability value of the node corresponding to the current printing layer through a multi-layer perceptron, and reconstructs the inclusion risk probability map of the printing layer according to the spatial position of the voxel unit:

[0073] ;

[0074] in, Indicates the The inclusion risk probability value of each node, represents the Sigmoid function, represents the activation function, and represents the weight matrix, Indicates the An updated node state vector, and represents the bias term;

[0075] The node set whose mixed risk probability value is less than a preset low risk threshold is defined as a low risk area, and the node set whose mixed risk probability value is greater than a preset high risk threshold is defined as a high risk area;

[0076] The inclusion risk probability map is a probability distribution map corresponding to the printing layer space, and is used to characterize the inclusion formation risk area at the voxel level.

[0077] By constructing a multi-scale graph growth generative network consisting of a graph feature encoding module, a graph structure evolution module, a node state update module, and a risk probability output module, dynamic modeling of inclusion risks in metal additive manufacturing is achieved. The multi-scale graph growth generative network not only utilizes spatial adjacency information for feature propagation but also incorporates time series evolution logic based on historical printing data, effectively improving the modeling capabilities for thermal disturbance propagation and impurity accumulation. The output is an inclusion risk probability map for each node in the print layer space, which can be used for high-precision path scheduling and regional risk shielding. The multi-scale graph growth generative network exhibits excellent portability and scalability, can be generalized and operated under a variety of component types and process parameters, and can be iteratively optimized based on feedback data to improve risk prediction accuracy.

[0078] In this embodiment, the inclusion risk probability map is used as the input of the mimic hummingbird optimization algorithm to construct a hummingbird individual path population. Specifically, according to the inclusion risk probability map, the print layer is divided into multiple path candidate blocks using a spatial grid division method, and multiple initial path individuals are generated in each path candidate block. The initial path individual starts from the printing start point and ends at the boundary of the path candidate block. The path shape is a continuous multi-segment vector trajectory, and each vector trajectory corresponds to a scanning step size and a direction angle. The hummingbird individual path population is constructed based on the initial path individual set generated in multiple path candidate blocks.

[0079] The inclusion risk probability map is used to divide the entire print layer into multiple candidate path blocks, and an initial population of individual paths is generated within each block, achieving structured partitioning of the print layer space and individualized path initialization. This step, based on the risk map-driven path layout, not only improves the regional specificity of the initial path search but also provides a diverse path solution space for the mimicking hummingbird optimization algorithm, improving the algorithm's global search performance. Furthermore, by encoding individual paths in the form of multi-segment vectors, they possess greater directional flexibility and structural expression capabilities, laying the foundation for strategy evolution and local disturbance control, significantly improving the stability and pinch avoidance performance of the path solution.

[0080] In this embodiment, the step five: based on the path optimization target space, a mimicking hummingbird optimization algorithm is used to perform path search and output a final scanning path, which specifically includes a forward flight strategy, a hovering perturbation strategy, and a regression strategy;

[0081] The forward flight strategy is to use a set of directions within a preset step size and direction angle range as candidate path directions in a low-risk area based on the starting point of the current hummingbird individual path, calculate the average inclusion risk probability value of the area corresponding to each candidate path direction, select the direction with the smallest average inclusion risk probability value and the continuity of the scanning path, and generate a main path segment as the backbone of the hummingbird individual path;

[0082] The hovering perturbation strategy triggers a path perturbation mechanism at the boundary of a high-risk area, generating multiple local perturbation path segments within the current position and its neighborhood with a set perturbation angle range and perturbation step size. The local perturbation path segments form a zigzag shape within a two-dimensional cross-section. The local perturbation path segments are sequentially spliced ​​with the current main path segments to form a perturbation extension segment of the path.

[0083] The regression strategy is to perform a path reversal operation when the average inclusion risk probability value of all candidate path segments in the path advancement direction is greater than a preset upper threshold, or the path continuity is interrupted, rewind the current hummingbird individual path to the previous node, reselect the direction set and evaluate the new forward path;

[0084] In each round of optimization iteration, the fitness value of the hummingbird individual path is calculated, which is calculated based on a weighted combination of the path length and the inclusion risk probability value of the nodes along the path;

[0085] ;

[0086] in, represents the fitness value of the hummingbird individual path, and represents the weight coefficient, represents the path length, Indicates the total number of voxel unit nodes passed on the path, Indicates the The inclusion risk probability value of each node;

[0087] When the number of optimization iterations reaches the preset maximum value, or the fitness value improvement rate between two consecutive iterations is lower than the preset change threshold, the iteration is terminated and the final scanning path is output;

[0088] The final scanning path is formed by sequentially splicing the main path segments and the disturbance path segments generated in each path candidate block according to the search execution order.

[0089] By integrating three biomimetic strategies into path search—forward flight, hovering perturbation, and regression—the mimicking hummingbird optimization mechanism achieves active perception and path fine-tuning of inclusion risk areas. It maximizes path continuity in low-risk areas, triggers a perturbation mechanism at high-risk boundaries to achieve zigzag traversal and ablation correction, and implements intelligent rollback to avoid invalid path advancement when path advancement is interrupted. The overall path strikes a balance between global optimal search and local obstacle avoidance fine-tuning, significantly reducing heat accumulation and impurity remelting risks during laser scanning, effectively improving the overall structural consistency and defect resistance of the component.

[0090] In this embodiment, the step six is ​​to control the laser to perform the printing scanning operation of the current layer according to the final scanning path, and collect the molten pool boundary contour, thermal image distribution data and printing layer status information in real time to generate process feedback data, specifically:

[0091] Sending a scanning trajectory instruction to the laser control system according to the final scanning path, controlling the laser to scan in sequence along the spliced ​​path segments to process the metal powder material of the current printing layer;

[0092] During the scanning process, the melt pool contour detection device, infrared thermal imaging sensor and layer status imaging unit set above or on the side of the printing area are used to synchronously observe the scanning area, collect the melt pool boundary contour, thermal image distribution data and printing layer status information in real time, and synchronously record them in time series. The data is bound to the corresponding scanning path segment to form a process feedback data set.

[0093] By combining the final scanning path to control the laser to perform scanning operations in real time, and introducing the collection of molten pool boundary contours, thermal image distribution data and printing layer status information during the printing process, the collected data are spatially bound and temporally aligned to form a structured process feedback data set, which can achieve full-dimensional recording of the corresponding state of each laser trajectory. It not only provides a real-time data basis for the next step of inclusion risk modeling, but also has the ability to identify anomalies, re-scheduling paths and subsequent quality traceability expansion, significantly enhancing the controllability and traceability of the additive manufacturing process.

[0094] In this embodiment, the step seven is: inputting the process feedback data into the multi-scale graph growth generation network, combining it with the inclusion risk probability map of the current printing layer, and outputting the updated inclusion risk probability map of the next printing layer, specifically:

[0095] Inputting the process feedback data set into a multi-scale graph growth generative network, performing feature extraction and fusion processing on the melt pool boundary contour, thermal image distribution data, and printing layer status information in the process feedback data set to form a node feature update vector reflecting the actual printing status of the current printing layer;

[0096] The node feature update vector is concatenated with the corresponding node state vector in the inclusion risk probability graph of the current printing layer, and the concatenated vector is input into the graph structure evolution module and the node state update module of the multi-scale graph growth generation network. The updated graph node state vector is obtained through graph convolution operation and state iterative update.

[0097] The updated graph node state vector is input into the risk probability output module, and the inclusion risk probability value of the node corresponding to the next printing layer is output through probability modeling calculation, thereby forming an updated inclusion risk probability graph for the printing and scanning operation of the next printing layer.

[0098] By feeding process feedback data collected during the printing process into a multi-scale graph growth generative network, the inclusion risk probability map is modified in real time through state updates and feature fusion, enabling dynamic evolution and predictive correction capabilities during mid-printing. The integration of process feedback data with graph node states ensures the ability to detect local thermal field anomalies, scanning offsets, and unstable particle behavior, enhancing responsiveness to microscopic evolutionary processes and significantly improving closed-loop control and printing consistency during the manufacturing process.

[0099] Example 1:

[0100] To verify the feasibility of this invention, the entire metal additive manufacturing process for a propulsion chamber component of an aerospace engine was applied. This component, a thin-walled cooling structure cavity assembly, was printed using selective laser melting (SLM) technology. Made of GH3536 high-temperature alloy, the print dimensions measured 250 mm × 120 mm × 80 mm, totaling 320 layers, with a processing cycle of 72 hours. This component has a complex structure and frequently fluctuating wall thickness. This makes it susceptible to the formation of microscopic inclusions or metallurgical defects due to factors such as uneven local laser heat input, metal powder agglomeration, or the diffusion of material defects, seriously impacting the propulsion chamber's service stability and safety.

[0101] In a specific implementation, the CAD model of the propulsion chamber component is first acquired using 3D modeling software and input into a slicing engine for slicing. This process generates the boundary contours, structural pore information, and interlayer topological mapping relationships for each printed cross-section, thereby constructing a printed layer dataset. Next, according to the method of the present invention, voxel units are established as graph structure nodes, and attribute data including spatial coordinates, metal particle distribution, particle impurity ratio, initial preheat temperature, scanning energy density, local geometric change rate, and residual stress characteristics are extracted. Finally, an initial multi-scale graph structure with cross-layer connection edges is constructed.

[0102] Subsequently, a multi-scale graph growth generation model based on a graph neural network was used. While inputting the initial multi-scale graph structure, the historical printing process data of the past three batches of this model engine was called to extract its thermal disturbance development path, typical defect area distribution, and inclusion pattern evolution sequence as the graph structure evolution supervision trajectory. After 8 rounds of graph structure iterative training, the model was able to output an inclusion risk probability map at a voxel-level resolution per layer. Visual analysis of the 84th, 137th, and 219th layers as representative layers showed that their mean inclusion risk probabilities were concentrated around 0.31, 0.47, and 0.42, which were significantly higher than the preset threshold of 0.25, indicating that the graph model can effectively identify high-risk trend areas under complex thermal history.

[0103] Based on the above inclusion risk probability map, the present invention introduces the mimic hummingbird optimization algorithm to dynamically search for each layer path. The printing layer is divided into candidate blocks according to the 20mm×20mm spatial grid. 12 path individuals are generated for each block. The forward flight strategy (main path construction), hovering perturbation strategy (local obstacle avoidance reconstruction) and regression strategy (path fallback and reorganization) are respectively executed. Each round of update is combined with the fitness evaluation function. The experiment sets the maximum number of iterations to 30, and the path fitness value weight coefficient is =0.6, =0.4. During the iteration process, most paths converged in the 8th to 12th rounds, with the average fitness value decreasing from the initial 0.756 to 0.312, and the mean inclusion probability decreasing by 34.8%.

[0104] During the final laser scanning control phase, the printing process utilizes a proprietary path-following control system, performing layer-by-layer processing according to the final scan path output by the mimicking Hummingbird optimization algorithm. Real-time laser scanning trajectory and melt pool thermal imaging data are collected, and a feedback mechanism is used to further adjust the graph structure and risk modeling module. In comparative tests, two identical components were processed using both a traditional equidistant scanning path and the proposed path optimization strategy. The inclusion defect rate was then determined using X-ray CT scanning and metallographic analysis.

[0105] Table 1 Comparison of inclusion risk probability values ​​in key layers of the propulsion chamber

[0106]

[0107] Based on the data in Table 1 above, it can be seen that after applying the method described in the present invention, the probability of occurrence of high-risk inclusion areas during the printing process is significantly reduced. Taking the 137th layer as an example, the mean probability of inclusion risk under the original path is 0.472, which is significantly higher, indicating that the traditional path fails to effectively avoid high thermal disturbance areas and areas where material defects are prone to gather. After adopting the graph neural network evolution and mimetic hummingbird path planning method of the present invention, the risk probability is reduced to 0.284, a reduction of 39.8%. Other key layers, such as the 84th layer, the 219th layer, and the 276th layer, also achieved a risk value reduction of 37.1%, 38.0%, and 38.5%, respectively, indicating that the risk identification and path reconstruction mechanism have shown good universality and stability in multiple structural areas.

[0108] Table 2 Comprehensive comparison of final component printing effects

[0109]

[0110] Based on the analysis of the data in Table 2 above, it can be seen that after the application of the present invention, the overall quality of the components has been significantly improved. In terms of defect rate, under the traditional path strategy, the inclusion volume accounts for as high as 0.87%, while the path optimization strategy of the present invention reduces it to 0.36%, a decrease of 58.6%, effectively reducing the metallurgical impact of material defects. In terms of molten pool stability, the strategy of the present invention significantly reduces the fluctuation of the interlayer molten pool width during processing, and the fluctuation range is reduced from ±13μm to ±7μm, reflecting a more stable energy input control and molten pool morphology adjustment capability. In terms of dimensional accuracy, the component error range is optimized from ±0.21mm to ±0.13mm, an increase of about 38%, providing reliable protection for subsequent assembly and service safety. In addition, in terms of overall processing efficiency, although the optimization process introduced an intelligent path algorithm, it did not slow down the manufacturing process. Instead, it shortened the total printing time from 73.2 hours to 71.8 hours, an increase of 1.9%, indicating that the path is shorter and the energy consumption is more reasonable.

[0111] This example incorporates a multi-scale graph growth generative network and a mimetic hummingbird optimization algorithm into the metal additive manufacturing process for complex aerospace propulsion chamber components, constructing a closed-loop control system for dynamic inclusion risk identification and intelligent path adjustment. This method utilizes a graph neural network to spatially model voxel-level risks in the printed layer and dynamically evolves an inclusion risk map based on historical data and process feedback. Based on this, a biomimetic optimization strategy is employed to achieve local perturbation avoidance and global continuous optimization of the printing path, significantly improving the path's ability to detect and avoid high-risk areas. Experimental results demonstrate that this method reduces the component inclusion defect rate by over 50%, reduces melt pool width fluctuation by 46%, converges the dimensional error range to ±0.13 mm, and slightly shortens total printing time, significantly improving overall manufacturing quality and process stability. This example demonstrates the practicality and robustness of this technical solution for printing complex high-temperature alloy structures. It provides a practical and intelligent solution to the ubiquitous inclusion control challenge in additive manufacturing of critical aerospace components, demonstrating significant engineering value and application prospects.

[0112] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber, characterized in that: The steps include: Step 1: Obtain a 3D model of the space propulsion chamber component to be printed and slice it to obtain 2D cross-sectional information of each printing layer and construct a printing layer dataset; Step 2: constructing an initial multi-scale graph structure based on the printed layer dataset; Step 3: Input the initial multi-scale graph structure into a multi-scale graph growth generation network, combine it with historical printing data to evolve the graph structure, and output the inclusion risk probability map of the current printing layer; Step 4: Using the inclusion risk probability map as input to the mimic hummingbird optimization algorithm, constructing a hummingbird individual path population, and setting the low-risk areas in the inclusion risk probability map as target path areas, and the high-risk areas as disturbance control areas, to generate a path optimization target space; Step 5: Based on the path optimization target space, the mimic hummingbird optimization algorithm is used to search the path and output the final scanning path; Step 6: Control the laser to perform a print scanning operation of the current layer according to the final scanning path, and collect the molten pool boundary contour, thermal image distribution data and printing layer status information in real time to generate process feedback data; Step 7: Input the process feedback data into a multi-scale graph growth generation network, combine it with the inclusion risk probability map of the current printing layer, and output an updated inclusion risk probability map for the next printing layer; Step 8: Repeat steps 4 to 7 until all printing layers are completed, thereby achieving control over the reduction of inclusions in the aerospace propulsion chamber components during the printing process.

2. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: The two-dimensional cross-sectional information of the printed layer includes the outline boundary coordinates, inner outline structural features, pore distribution position, scanning area segmentation information, and connection topological relationship between the printed layer and adjacent upper and lower layers.

3. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: The construction of the initial multi-scale graph structure is specifically as follows: Setting each voxel unit in the printing layer data set as a node in a graph structure, wherein the node attributes include the spatial coordinates of the voxel unit, the distribution state of metal particles, the impurity information of metal particles, the initial temperature value, the laser scanning energy, the local geometric characteristics and the interlayer residual stress information; Establish edge connections between nodes with spatial adjacency, and the edge weight is the Euclidean distance between the nodes; A two-dimensional intra-layer graph structure is constructed based on node attributes and edge connections, and cross-layer edge connections are constructed through the topological mapping relationship of voxel nodes between the upper and lower printing layers to form an initial multi-scale graph structure.

4. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: The third step is to input the initial multi-scale graph structure into a multi-scale graph growth generation network, perform graph structure evolution in combination with historical printing data, and output an inclusion risk probability map of the current printing layer, specifically: The multi-scale graph growth generation network includes a graph feature encoding module, a graph structure evolution module, a node state update module and a risk probability output module, which is used to dynamically model the inclusion generation behavior in the metal additive manufacturing process; The graph feature encoding module receives the initial multi-scale graph structure, performs feature embedding and dimension normalization on node attributes, and generates a node state vector; The graph structure evolution module uses the graph structures and evolution trajectories of each layer contained in the historical printing data as supervision data, adopts a graph convolutional network to aggregate the state vectors between adjacent nodes, and combines edge weight embedding values ​​for information transmission. It simulates the diffusion trend and structural aggregation behavior caused by thermal disturbances during the inclusion formation process through iterative propagation. The node state update module updates the state vector of each node after each graph structure evolution, introduces the evolution trajectory of the node at the same spatial position in the historical printing process as a time series input, and splices it with the current node state vector; The risk probability output module performs probability modeling on all updated node state vectors, outputs the inclusion risk probability value of the node corresponding to the current printing layer through a multi-layer perceptron, and reconstructs the inclusion risk probability map of the printing layer according to the spatial position of the voxel unit; The node set whose mixed risk probability value is less than a preset low risk threshold is defined as a low risk area, and the node set whose mixed risk probability value is greater than a preset high risk threshold is defined as a high risk area; The inclusion risk probability map is a probability distribution map corresponding to the printing layer space, and is used to characterize the inclusion formation risk area at the voxel level.

5. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: The inclusion risk probability map is used as the input of the mimic hummingbird optimization algorithm to construct a hummingbird individual path population. Specifically, according to the inclusion risk probability map, the print layer is divided into multiple path candidate blocks using a spatial grid division method, and multiple initial path individuals are generated in each path candidate block. The initial path individuals start from the printing start point and end at the boundary of the path candidate block. The path shape is a continuous multi-segment vector trajectory, and each vector trajectory corresponds to a scanning step size and a direction angle. The hummingbird individual path population is constructed based on the initial path individual set generated in the multiple path candidate blocks.

6. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: Step 5: Using the mimic hummingbird optimization algorithm to search for a path based on the path optimization target space, and outputting a final scanning path, specifically including a forward flight strategy, a hovering perturbation strategy, and a regression strategy; The forward flight strategy is to use a set of directions within a preset step size and direction angle range as candidate path directions in a low-risk area based on the starting point of the current hummingbird individual path, calculate the average inclusion risk probability value of the area corresponding to each candidate path direction, select the direction with the smallest average inclusion risk probability value and the continuity of the scanning path, and generate a main path segment as the backbone of the hummingbird individual path; The hovering perturbation strategy triggers a path perturbation mechanism at the boundary of a high-risk area, generating multiple local perturbation path segments within the current position and its neighborhood with a set perturbation angle range and perturbation step size. The local perturbation path segments form a zigzag shape within a two-dimensional cross-section. The local perturbation path segments are sequentially spliced ​​with the current main path segments to form a perturbation extension segment of the path. The regression strategy is to perform a path reversal operation when the average inclusion risk probability value of all candidate path segments in the path advancement direction is greater than a preset upper threshold, or the path continuity is interrupted, rewind the current hummingbird individual path to the previous node, reselect the direction set and evaluate the new forward path; In each round of optimization iteration, the fitness value of the hummingbird individual path is calculated, which is calculated based on a weighted combination of the path length and the inclusion risk probability value of the nodes along the path; When the number of optimization iterations reaches the preset maximum value, or the fitness value improvement rate between two consecutive iterations is lower than the preset change threshold, the iteration is terminated and the final scanning path is output; The final scanning path is formed by sequentially splicing the main path segments and the disturbance path segments generated in each path candidate block according to the search execution order.

7. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: Step 6: Control the laser to perform a print scanning operation on the current layer according to the final scan path, and collect the molten pool boundary contour, thermal image distribution data, and print layer status information in real time to generate process feedback data, specifically: Sending a scanning trajectory instruction to the laser control system according to the final scanning path, controlling the laser to scan in sequence along the spliced ​​path segments to process the metal powder material of the current printing layer; During the scanning process, the melt pool contour detection device, infrared thermal imaging sensor and layer status imaging unit set above or on the side of the printing area are used to synchronously observe the scanning area, collect the melt pool boundary contour, thermal image distribution data and printing layer status information in real time, and synchronously record them in time series. The data is bound to the corresponding scanning path segment to form a process feedback data set.

8. The method for reducing metal particle inclusion in additive manufacturing of aerospace propulsion chamber according to claim 1, characterized in that: Step 7: Inputting the process feedback data into a multi-scale graph growth generation network, combining it with the inclusion risk probability map of the current printing layer, and outputting an updated inclusion risk probability map of the next printing layer, specifically: Inputting the process feedback data set into a multi-scale graph growth generative network, performing feature extraction and fusion processing on the melt pool boundary contour, thermal image distribution data, and printing layer status information in the process feedback data set to form a node feature update vector reflecting the actual printing status of the current printing layer; The node feature update vector is concatenated with the corresponding node state vector in the inclusion risk probability graph of the current printing layer, and the concatenated vector is input into the graph structure evolution module and the node state update module of the multi-scale graph growth generation network. The updated graph node state vector is obtained through graph convolution operation and state iterative update. The updated graph node state vector is input into the risk probability output module, and the inclusion risk probability value of the node corresponding to the next printing layer is output through probability modeling calculation, thereby forming an updated inclusion risk probability graph for the printing and scanning operation of the next printing layer.

Images