3D printing intelligent orientation method based on virtual experiment
By constructing a 3D printing orientation virtual simulation environment and reinforcement learning algorithms, the construction direction of parts is optimized, solving the problem of poor construction direction selection in existing technologies, and achieving a reduction in support structures and a shortening of the printing cycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANMUSHAN LABORATORY
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-16
AI Technical Summary
Existing 3D printing technologies struggle to achieve globally optimal results when determining the direction of component construction, and existing automated methods are ineffective when handling complex components, resulting in excessively large support structures and contact areas, which increases printing costs and time.
A virtual experiment-based intelligent orientation method is adopted. By constructing a 3D printing orientation virtual simulation environment model and combining reinforcement learning algorithms, the construction direction of parts is optimized to generate the optimal support structure. Iterative optimization is then carried out through the interaction between the orientation agent and the virtual simulation environment.
By reducing the volume and contact area of the support structure, the additive manufacturing cycle is shortened, and printing quality and efficiency are improved, without the need for frequent printing of actual objects for verification.
Smart Images
Figure CN121973449B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent manufacturing technology, specifically relating to an intelligent orientation method for 3D printing based on virtual experiments. Background Technology
[0002] As the application of metal and polymer additive manufacturing (3D printing) in aerospace, medical, and mold-making fields continues to deepen, the requirements for component quality are also becoming increasingly stringent. Determining the build direction of a component is the first and most crucial step in the additive manufacturing process, as it has a decisive impact on the component's support requirements, surface quality, residual stress, and printing time. Incorrect or suboptimal build directions often result in excessively large support structures and contact areas, increased post-processing costs, or even failure to form the component due to build volume constraints. In complex curved surfaces and thin-walled structures, the coupling between direction, support, thermal field, and path planning makes it difficult to obtain a stable and effective solution for the component's build direction through experience or iteration of a single indicator alone. The choice of build direction simultaneously affects support topology, surface roughness, build time, and cost, representing a key bottleneck in the additive manufacturing preparation stage.
[0003] Traditional printing technology typically doesn't optimize the build orientation when processing small parts. It often prints according to the initial orientation in the model file or manually selects the orientation based on experience. This method usually doesn't yield the most material-efficient printing orientation, but it saves time, making it common for customized printing. However, for mass production, failing to optimize the build orientation increases printing costs, while relying on experience is both time-consuming and not always optimal. Therefore, automated methods for orienting the build orientation before printing are gradually replacing the traditional method. Existing automation technologies either rely on repeated simulation experiments to select an optimal printing orientation or on heuristic search algorithms to output orientation results. While these have achieved some success, they cannot guarantee consistently globally optimal results and are time-consuming, potentially delaying production schedules. Heuristic search methods typically conduct orientation experiments on a single target in a specific environment, considering the characteristics of the part, to evaluate and optimize the printing orientation. Although automated, this approach only considers a single target and specific environment, which differs from actual production requirements. Furthermore, this method primarily handles simple parts and is ineffective with large, complex components.
[0004] Therefore, there is an urgent need to propose a 3D printing intelligent orientation method based on virtual experiments. In a simulation environment, combined with the printing space and forming axis constraints of the equipment, the advanced algorithm of reinforcement learning is used to quickly evaluate and iteratively optimize the relevant indicators of candidate orientations, so as to reduce the amount of support, control costs, and shorten the printing preparation cycle. Summary of the Invention
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0006] A 3D printing intelligent orientation method based on virtual experiments, comprising:
[0007] S1. Construct a 3D printing orientation virtual simulation environment model. The 3D printing orientation virtual simulation environment model includes the original part mesh model, the printing equipment workspace model, and the constraint relationship between the original part mesh model and the printing equipment workspace model.
[0008] S2. Construct the component support rule module to generate the support structure required for 3D printing of components, and calculate the volume of the support structure and the contact area between the support structure and the component in the corresponding component construction direction;
[0009] S3. Design a directional agent that includes an evaluation network and a target network. The directional agent calculates the long-term reward after executing the action selected according to the action selection strategy, updates the network parameters of the evaluation network and the target network, selects the action that can obtain the maximum long-term reward among all actions, and performs intelligent orientation optimization of components.
[0010] S4. Design a reinforcement learning algorithm for intelligent orientation optimization of construction direction under multiple objectives. The orientation agent interacts with the 3D printing orientation virtual simulation environment model in real time and outputs the optimal component construction direction under multiple objectives.
[0011] S5. Based on the component construction direction output by S4, print the actual component model, compare the printing quality of the actual component model with the original component mesh model, and verify and provide feedback on the effect of intelligent orientation optimization reinforcement learning.
[0012] The present invention has the following beneficial effects:
[0013] This invention constructs a directional virtual simulation environment model that includes the spatial motion relationship between 3D printing equipment and component models. Simultaneously, it establishes calculation rules for the necessary supports for component printing, accurately generates the support structure for the components, and calculates corresponding indicators such as the volume and supported area of the support structure. Through the interaction between a designed reinforcement learning-based directional agent and the 3D printing directional virtual simulation environment, intelligent orientation optimization obtains the optimal printing direction for the components. This invention, by constructing a reinforcement learning-based directional agent and a 3D printing directional virtual simulation environment, and using the rewards obtained from the interaction feedback between the two, trains the directional agent to intelligently optimize the orientation of components. This invention can reduce the volume of the support structure and the contact area between the support structure and the component without requiring frequent printing of actual objects for verification. It achieves efficient and accurate intelligent orientation optimization for complex 3D printed components, which helps to shorten the additive manufacturing cycle while ensuring the optimization effect of the construction direction. Attached Figure Description
[0014] Figure 1 The flowchart shows the intelligent orientation method for 3D printing based on virtual experiments according to the present invention.
[0015] Figure 2 This is a flowchart illustrating the construction of the component support rule module of the present invention;
[0016] Figure 3 This is a schematic diagram of the training process for the direction-oriented optimization reinforcement learning algorithm of the present invention;
[0017] Figure 4 This is a diagram of the original component mesh model for intelligent directional optimization according to the present invention;
[0018] Figure 5 This is a schematic diagram showing the results of the intelligent orientation method for 3D printing based on virtual experiments on the original component mesh model according to the present invention;
[0019] Figure 6 This is a schematic diagram showing the comparison results of the support structure volume and the contact area between the support structure and the component before and after the intelligent orientation optimization of the present invention; wherein, (a) is a schematic diagram of the support structure volume and the contact area between the support structure and the component of the original component mesh model before intelligent orientation optimization, and (b) is a schematic diagram of the support structure volume and the contact area between the support structure and the component of the actual component model after intelligent orientation optimization.
[0020] Figure 7This is a schematic diagram showing the quality comparison results of the component models before and after intelligent orientation optimization in an embodiment of the present invention; wherein, (a) is the front view and isometric view of the actual original component after intelligent orientation optimization in 3D printing, and (b) is the front view and isometric view of the actual original component before intelligent orientation optimization in 3D printing.
[0021] Figure 8 The above are heatmaps showing the surface geometric errors of the component models before and after intelligent orientation optimization in this embodiment of the invention. Among them, (a) is a heatmap showing the surface geometric errors of the scanned mesh model obtained by 3D scanning of the physical component after removing the support structure and the original component mesh model after intelligent orientation optimization, and (b) is a heatmap showing the surface geometric errors of the scanned mesh model obtained by 3D scanning of the physical component after removing the support structure and the original component mesh model before intelligent orientation optimization. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0023] This invention provides a virtual experiment-based intelligent orientation method for 3D printing. It calculates the required volume of the support structure and the contact area between the support structure and the part in a specific build direction and build space. Utilizing artificial intelligence technology, based on reinforcement learning and interaction with a virtual simulation environment, the printing direction of the part is iteratively optimized. This invention, based on reinforcement learning by an orientation agent in a virtual 3D printing space, reduces the volume of the support structure and the contact area between the support structure and the part without requiring frequent printing of actual objects for verification. This achieves efficient and accurate intelligent orientation optimization for complex 3D printed parts, which helps shorten the additive manufacturing cycle while ensuring the optimization effect of the build direction.
[0024] like Figure 1 As shown, the 3D printing intelligent orientation method based on virtual experiment of the present invention includes the following steps:
[0025] S1. Construct a 3D printing orientation virtual simulation environment model. The 3D printing orientation virtual simulation environment model includes the original part mesh model, the printing equipment workspace model, and the constraint relationship between the original part mesh model and the printing equipment workspace model.
[0026] S2. Construct a component support rule module to generate the support structure required for 3D printing of components, and calculate the volume of the support structure and the contact area between the support structure and the component in the corresponding component construction direction. This serves as the basis for subsequent support required for 3D printing of components and for evaluating the effect of intelligent orientation optimization.
[0027] S3. Design an oriented agent that includes an evaluation network and a target network. Define the action space, state space, and reward function. The oriented agent is based on the orientation of the parts in the 3D printing oriented virtual simulation environment model constructed in S1 and the volume of the support structure and the contact area between the support structure and the parts calculated in S2. Using the evaluation network and the target network, calculate the long-term reward of the oriented agent after performing the action selected according to the action selection strategy in the 3D printing oriented virtual simulation environment model, and update the network parameters of the evaluation network and the target network.
[0028] Based on the evaluation network's assessment of the long-term reward after performing an action, the directed agent adopts the principle of maximum expected reward as the action selection strategy, selecting the action that can obtain the maximum long-term reward among all actions and performing intelligent orientation optimization of components.
[0029] S4. Design a multi-objective intelligent orientation optimization reinforcement learning algorithm to construct an intelligent orientation agent that interacts in real time with a 3D printed orientation virtual simulation environment model; wherein, the orientation agent performs time steps t according to the action selection strategy. Action selection This is applied to the 3D printing orientation virtual simulation environment model. Based on the S2 component support rule module, the 3D printing orientation virtual simulation environment model updates the orientation of the components and the volume of the support structure and the contact area between the support structure and the components under the corresponding component construction direction. After calculating the reward, it is fed back to the orientation agent. The orientation agent learns the parameters and evaluates the expected reward through the evaluation network and the target network, updates the network parameters, and selects the next action. This process is repeated iteratively until the action selection strategy tends to converge, and the optimal component construction direction under multiple objectives is output.
[0030] S5. Based on the component construction direction output by S4, print the actual component model, compare the printing quality of the actual component model with the original component mesh model, and verify and provide feedback on the effect of intelligent orientation optimization reinforcement learning.
[0031] Furthermore, S1 includes:
[0032] S11. Based on the geometric features of the 3D printed parts, construct a triangular mesh model of the parts. The triangular mesh information of the triangular mesh model of the parts includes the vertex coordinates and normal vectors of the triangular facets.
[0033] S12. Based on the constraint relationship that the parts need to always be contained within the workspace of the 3D printing equipment, limit the range of coordinate changes of the parts in the workspace of the 3D printing equipment, construct a 3D printing orientation virtual simulation environment model, and ensure that the movement range of the parts is always within the workspace of the 3D printing equipment.
[0034] The S13 3D printing orientation virtual simulation environment model takes the vertex coordinates and normal vectors of the triangular facets of the triangular mesh model of the parts under different states as input, calculates and outputs the vertex coordinates and normal vectors of the triangular facets after rotation and translation, and realizes the position information, translation and rotation motion mapping and real-time update of the state of the triangular mesh model of the parts in the 3D printing orientation virtual simulation environment model.
[0035] In S12, given the rotation matrix and displacement t Under the given conditions, the constraint relationship is:
[0036] ;
[0037] in, For rotation matrix and displacement t The outer envelope of the component obtained from the action. For the workspace of the printing equipment, A suitable distance is reserved to ensure a safe working space and prevent collisions between parts and the work area. It is the intersection of the working space of the printing equipment and the safe working space.
[0038] via rotation matrix and displacement normal vector of the surface of the component after the action for:
[0039] ;
[0040] in, For displacement, This is the rotation matrix obtained by the Rodrigues rotation formula; This is the normal vector of the surface of the component before the action.
[0041] Rotation matrix obtained from Rodrigues' rotation formula for:
[0042] ;
[0043] in, The unit vector of the rotation axis. For any unit vector Angle of rotation It is a third-order identity matrix. , , Unit vectors of rotation axis exist , , The components of the axis.
[0044] In S2, a component support rule module is constructed based on the triangular mesh model of the component, the forming axis direction, the preset critical overhang angle threshold, and the platform position coordinates. This module includes determining the support area and generating the support structure required for 3D printing of the component.
[0045] The process of building the component support rule module is as follows: Figure 2 As shown. When the 3D printing orientation virtual simulation environment model needs to update the state of the part model in space, it first reads the triangular facet information of the model (the vertex coordinates and normal vectors of the triangular facets), then determines the facets that need to be supported according to the critical overhang angle threshold, then calculates the key parameters (including the volume of the support structure under the corresponding part construction direction and the contact area between the support structure and the part), and finally generates the support structure, that is, the support structure required for the 3D printing of the part.
[0046] The support area is determined using the following formula:
[0047] ;
[0048] in, The unit normal vector of the surface. To construct a directional unit vector, The angle of the cantilever face.
[0049] The following formula is used to generate the support structure required for 3D printing of parts and to calculate the volume of the support structure and the contact area between the support structure and the parts in the corresponding part construction direction:
[0050] ;
[0051] in, For the first The area of each triangular facet. , For the first The two edge vectors of a triangular facet. For the first The projected area of a triangular facet on the XOY plane. For the first The height from the vertices of each triangular facet to the platform of the 3D printing equipment. This is the average of the sum of the heights from all vertices of the triangular facet to the 3D printing platform. For the first The first triangular facet The y-coordinates of the vertices, For the first The first triangular facet The vertical coordinate of the surface contact point or 3D printing equipment platform contact point below each vertex that can provide support. The number of vertices of the triangular facet. , These are the volume of the supporting structure for the component and the contact area between the supporting structure and the component, respectively.
[0052] In S3, the motion space is a set of motion sequences where components rotate around or remain stationary on orthogonal coordinate axes, and the execution of an action is performed from the motion group. The product of the selected action most likely to lead to the optimal construction direction and the rotation step size, is the action group. The rotation step size is defined as follows:
[0053] ;
[0054] ;
[0055] in, For action groups, i When the value is between 0 and 6 The values are as shown in the above formula; , These represent the rotation angles at each time step for the next round and the current round, respectively. This is the current round number. Total number of rounds To find the function with the maximum value, This indicates taking the rotation angle 5 and the rotation angle of the current round. The maximum value in.
[0056] Based on the directed agent in the action group The rotation angle of a component at a given time step is obtained by multiplying the action most likely to lead to the optimal construction direction by the rotation angle of the current round.
[0057] The state space is a five-dimensional array containing the construction direction information of the components and the supporting structure information. The definition is as follows:
[0058] ;
[0059] in, For time steps, For each component, wind separately , , The angle of rotation of the axis , These represent the volume of the supporting structure for the component and the contact area between the supporting structure and the component, respectively; as long as the reinforcement learning model training process does not converge or terminate, the state of the component... It will be replaced with the corresponding new state as time steps change. , among them The time steps are respectively hour, The new state value.
[0060] The reward function is related to the reduction in the volume of the support structure and the contact area between the support structure and the component during the orientation process, and is defined as follows:
[0061] ;
[0062] in, For vector rewards, and These are constants whose different values are selected in the real number field based on experience (the range of values is set according to experience). for Scalar reward after scalarization and Within the range of values The constants selected in the different values, Rewards related to changes in the volume of the supporting structure. Rewards are related to changes in the contact area between the supporting structure and components. The volume of the supporting structure for the components after the action is performed. The contact area between the supporting structure and the component after the action is performed.
[0063] Based on the definition of action space, state space, and reward function, the directed agent calculates the action value function through two deep neural networks with identical structures: an evaluation network and a target network. The expected value is determined, and parameter iteration and intelligent directional optimization are performed accordingly. That is, the Q-value function, which is the value at the current time step. status Next action selection A function for estimating future value, numerically equal to the value of the directed agent at the current time step. status Next, select the action to execute. Then, the expected long-term cumulative reward. Its mathematical expression follows the Bellman equation:
[0064] ;
[0065] in, For components in time step state, In time step From the action group The selected rotation action, In time step The scalar reward provided by the environment after the action is performed (same as the definition in the reward section). This is the discount factor (constant). To find the function with the maximum value, For components in time step state, Choose from all possible actions. For random variables The mathematical expectation.
[0066] The evaluation network is used to output the selection probability or predictive value of each action in the current state. , To evaluate the set of all learnable parameters in the network, these learnable parameters are updated at each time step.
[0067] The target network is used to generate target reward predictions. :
[0068] ;
[0069] in, For the target network Value function, In time step From the action group The selected rotation action, The set of all learnable parameters in the target network. and The update frequency is different. Updated every fixed time interval. Update at every time step.
[0070] The directed agent employs the expected reward principle as its action selection strategy. Specifically, at each time step... The directed agent calculates the values of all actions in the action space by evaluating the network. Value, and select to make Rotational motion that reaches its maximum value implement, To evaluate the set of all learnable parameters in the network.
[0071] The parameter updates of the directed agent minimize the predictive value of the evaluation network. The target reward prediction value of the target network Mean squared error loss function To achieve this. The formula is as follows:
[0072] ;
[0073] Use the loss calculation results to evaluate the parameters of the network. Perform gradient descent updates for the target network. Regularly synchronize from the evaluation network to ensure the convergence and stability of the intelligent targeted optimization process for components.
[0074] S4 includes: designing a multi-objective intelligent orientation optimization reinforcement learning algorithm for construction direction, real-time interaction between the orientation agent and the 3D printing orientation virtual simulation environment model, and realizing the training of the intelligent orientation optimization reinforcement learning algorithm for construction direction.
[0075] The real-time interaction process is as follows: at the beginning of each round, the parts to be printed are loaded into the environment; at the beginning of each time step in each round, the input state is initialized, and the state of the parts in the 3D printing orientation virtual simulation environment model is initialized; the orientation agent selects actions according to the action selection strategy and performs time step... Action selection This is applied to the 3D printing orientation virtual simulation environment model, where actions are performed on the parts, i.e., the parts are oriented. Following this, the 3D printing orientation virtual simulation environment model, based on the S2 part support rule module, updates the part's orientation, the volume of the support structure in the corresponding part's construction direction, and the contact area between the support structure and the part. Then, it calculates the reward for performing this action and feeds it back to the orientation agent. Internally, the orientation agent learns parameters and evaluates expected rewards based on the environmental feedback rewards through an evaluation network and a target network, updating the network parameters. Based on the updated network parameters, the evaluation network periodically synchronizes parameters with the target network; the orientation agent continues with time steps... Action selection Repeat the above real-time interaction process until the time step... Action selection When the action selection strategy tends to converge, that is, the component orientation results of this training round converge. If it does not converge, the above real-time interaction process is repeated, and the vector reward is applied. Input the orientation agent; if convergence occurs, exit the real-time interaction process. Based on the Pareto optimization method, select from the multi-objective optimization results, choosing the orientation result obtained in the last round as the optimal construction direction for output. The training process of the orientation optimization reinforcement learning algorithm is as follows: Figure 3As shown in Figure S5, based on the component construction direction output by S4, the actual component model after intelligent orientation optimization is obtained through 3D printing. The printing quality of the actual component model is compared with that of the original component mesh model to verify and provide feedback on the reinforcement learning effect of intelligent orientation optimization, so as to adjust the reinforcement learning parameters.
[0076] To comprehensively evaluate the applicability and superiority of the virtual experiment-based intelligent orientation method for 3D printing of the present invention, in such cases... Figure 4 The original component mesh model (a typical 3D printed component (support)) shown was used for intelligent orientation optimization. The simulation experiment of reinforcement learning intelligent orientation optimization was carried out on it. The intelligent orientation optimization was carried out with the goal of reducing the volume of the support structure and the contact area between the support structure and the component. The simulation and physical comparison analysis of the component before and after intelligent orientation optimization were performed.
[0077] First, the intelligent orientation method for 3D printing based on virtual experiments of this invention is used to intelligently optimize the orientation of the original part mesh model to obtain the Pareto front solution set and orientation results for the printing direction, such as... Figure 5 As shown, the horizontal axis represents the support volume, indicating the volume of the support structure of the component, and the vertical axis represents the contact area, indicating the contact area between the support structure and the component. The star symbol represents the Pareto front solution, and the dots represent all solutions. The solutions on the Pareto front represented by the star symbol are connected by dashed lines to form the Pareto front. The result of the last round of training after the reinforcement learning converges is used as the result of intelligent directional optimization.
[0078] The results of intelligent orientation optimization were applied to components, and the comparison results of the support structure volume and the contact area between the support structure and the component before and after intelligent orientation optimization were shown below. Figure 6 As shown. Figure 6 (a) is a schematic diagram of the volume of the support structure and the contact area between the support structure and the component in the original component mesh model before intelligent directional optimization. Figure 6 (b) is a schematic diagram of the volume of the support structure and the contact area between the support structure and the component in the actual component model after intelligent orientation optimization; where V1 and A1 are the volume of the support structure and the contact area between the support structure and the component before intelligent orientation optimization, respectively, and V1' and A1' are the volume of the support structure and the contact area between the support structure and the component after intelligent orientation optimization, respectively. The volume of the support structure and the contact area between the support structure and the component are significantly improved after intelligent orientation optimization.
[0079] Second, the component models with the applied intelligent orientation optimization results are 3D printed to obtain physical components with the optimized construction direction. The printing quality of the optimized components and the unoptimized components is analyzed and compared. Specifically, bracket components without intelligent orientation optimization and bracket components optimized using this invention are printed, weighed using a precision electronic scale, and their mass is compared. The comparison results are as follows: Figure 7 (1-5mm is the scale; 30.64 represents the part mass after intelligent orientation optimization (30.64g), and 31.94 represents the part mass before intelligent orientation optimization (31.94g), with a difference of 1.30g.) The results show that the actual part model after intelligent orientation optimization has a significantly lower quality than the original part mesh model. Among these, Figure 7 (a) shows the front view and isometric view of the actual original part after intelligent orientation optimization in 3D printing. Figure 7 (b) is a front view and isometric view of the actual original part before intelligent orientation optimization for 3D printing.
[0080] Third, remove the support structures of the parts that were not intelligently oriented optimized and the parts that were intelligently oriented optimized based on this invention. Use a 3D scanner to scan the two actual part models to obtain scanned mesh models. Compare these two scanned mesh models with the original part mesh model, and compare the geometric errors of the two printed and post-processed parts with the original part mesh model. The comparison results are as follows: Figure 8 As shown. Figure 8 (a) is a heatmap comparing the surface geometric errors of the scanned mesh model obtained by 3D scanning of the physical component after removing the supporting structure, and the original component mesh model, after intelligent orientation optimization. Figure 8 (b) is a heat map comparing the surface geometric error of the scanned mesh model obtained by 3D scanning of the physical parts after removing the supporting structure before intelligent orientation optimization, and the original part mesh model; the comparison results show that the geometric error after intelligent orientation optimization is reduced compared with the original part mesh model.
[0081] The comparative results show that, compared with the parts without intelligent orientation optimization, the parts with intelligent orientation optimization using the present invention have a 12.8% reduction in support structure volume, a 21.3% reduction in contact area between the support structure and the parts, and a significant reduction in part mass and post-processing geometric error. The comparative results demonstrate that the present invention can significantly reduce the printing cost of parts and the accuracy loss caused by model post-processing while effectively balancing the trade-offs between multiple objectives.
[0082] The above description is merely an embodiment of the present invention and does not limit the scope of the invention. Any equivalent structural or procedural transformations made based on the description and drawings of this invention, or direct or indirect applications in other related system fields, are similarly included within the protection scope of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art.
Claims
1. A 3D printing intelligent orientation method based on virtual experiments, characterized in that, include: S1. Construct a 3D printing orientation virtual simulation environment model. The 3D printing orientation virtual simulation environment model includes the original part mesh model, the printing equipment workspace model, and the constraint relationship between the original part mesh model and the printing equipment workspace model. S2. Construct the component support rule module to generate the support structure required for 3D printing of components, and calculate the volume of the support structure and the contact area between the support structure and the component in the corresponding component construction direction; S3. Design a directional agent that includes an evaluation network and a target network. The directional agent calculates the long-term reward after executing the action selected according to the action selection strategy, updates the network parameters of the evaluation network and the target network, selects the action that can obtain the maximum long-term reward among all actions, and performs intelligent orientation optimization of components. S4. Design a reinforcement learning algorithm for intelligent orientation optimization of construction direction under multiple objectives. The orientation agent interacts with the 3D printing orientation virtual simulation environment model in real time and outputs the optimal component construction direction under multiple objectives. S5. Based on the component construction direction output by S4, print the actual component model, compare the printing quality of the actual component model with the original component mesh model, and verify and provide feedback on the effect of intelligent orientation optimization reinforcement learning.
2. The intelligent orientation method for 3D printing based on virtual experiments according to claim 1, characterized in that, S1 includes: S11. Based on the geometric features of 3D printed parts, construct a triangular mesh model of the parts. The triangular mesh information of the triangular mesh model of the parts includes the vertex coordinates and normal vectors of the triangular facets. S12. Based on the constraint relationship that the parts need to be always contained within the workspace of the 3D printing equipment, limit the range of coordinate changes of the parts in the workspace of the 3D printing equipment, and construct a 3D printing orientation virtual simulation environment model. The S13 3D printing orientation virtual simulation environment model takes the vertex coordinates and normal vectors of the triangular facets of the triangular mesh model of the parts under different states as input, calculates and outputs the vertex coordinates and normal vectors of the triangular facets after rotation and translation, and realizes the position information, translation and rotation motion mapping and real-time update of the state of the triangular mesh model of the parts in the 3D printing orientation virtual simulation environment model.
3. The intelligent orientation method for 3D printing based on virtual experiments according to claim 2, characterized in that, In S12, under the condition of known rotation matrix and displacement, the constraint relationship is: the outer envelope of the component obtained by the rotation matrix and displacement action is contained in the intersection of the working space and safe working space of the printing equipment.
4. The intelligent orientation method for 3D printing based on virtual experiments according to claim 3, characterized in that, In S12, the normal vector of the surface of the component after the rotation matrix and the displacement action is: the displacement plus the dot product of the rotation matrix and the normal vector of the surface of the component before the action.
5. The intelligent orientation method for 3D printing based on virtual experiments according to claim 4, characterized in that, In S12, the rotation matrix includes a third-order identity matrix, a unit vector of the rotation axis, the angle of rotation of any unit vector, and the position of the unit vector of the rotation axis. , , The amount.
6. The intelligent orientation method for 3D printing based on virtual experiments according to claim 5, characterized in that, S2 includes: Based on the triangular mesh model of the parts, the forming axis direction, the preset critical overhang angle threshold, and the platform position coordinates, a parts support rule module is constructed, including determining the support area and generating the support structure required for 3D printing of the parts. The component support rule module includes: reading the vertex coordinates and normal vectors of the model's triangular facets, determining the support area based on the critical overhang angle threshold, calculating the volume of the support structure and the contact area between the support structure and the component in the corresponding component construction direction, and finally generating the support structure required for 3D printing of the component.
7. The intelligent orientation method for 3D printing based on virtual experiments according to claim 6, characterized in that, In S2, the support region is determined according to the rule that the negative of the dot product of the unit normal vector of the facet and the unit vector of the construction direction is greater than the cosine of the boundary angle of the overhanging facet; the volume of the support structure and the contact area between the support structure and the parts are calculated based on the triangular facet information to generate the support structure required for 3D printing of the parts.
8. The intelligent orientation method for 3D printing based on virtual experiments according to claim 7, characterized in that, S3 includes: defining the action space, state space, and reward function; the action space is a set of action sequences for components to rotate or remain stationary around orthogonal coordinate axes, and the state space is a six-dimensional array containing component construction direction information, support structure information, and time steps; the reward function is related to the reduction of support structure volume and contact area.
9. The intelligent orientation method for 3D printing based on virtual experiments according to claim 8, characterized in that, S3 includes: Based on the defined action space, state space, and reward function, the directed agent uses the orientation of the parts in the 3D printing directed virtual simulation environment model constructed by S1 and the volume of the support structure and the contact area between the support structure and the parts calculated by S2. Using the evaluation network and the target network, the agent calculates the long-term reward after performing the action selected according to the action selection strategy in the 3D printing directed virtual simulation environment model, and updates the network parameters of the evaluation network and the target network.
10. The intelligent orientation method for 3D printing based on virtual experiments according to claim 9, characterized in that, S4 includes: designing a multi-objective intelligent orientation optimization reinforcement learning algorithm for building direction, real-time interactive training between the orientation agent and the 3D printing orientation virtual simulation environment model until the action selection strategy converges, and outputting the optimal building direction according to the Pareto optimization method, thereby realizing the training of the intelligent orientation optimization reinforcement learning algorithm for building direction.