A scanning path planning method for electron beam powder bed points based on GA-BP neural network

Abstract

The application discloses a kind of based on GA-BP neural network's electron beam powder bed point melt scanning path planning method, belong to metal additive manufacturing technical field.This method first uses Latin hypercube sampling to determine the parameter combination of different beam and residence time, obtains the corresponding melt pool aspect ratio, cooling time and heat affected distance of each combination by single point point melt physical simulation, constructs and trains neural network model.In actual path planning, the current beam and residence time are input into the model to obtain the corresponding melt pool aspect ratio, cooling time and heat affected distance, and then obtain the powder melting degree under the current process parameters, the return time of point and the number of point jump points, finally guide the adjustment of process parameters and give the scanning path planning of point melt.The application drives multi-point path decision by combining single-point physical simulation with neural network, can significantly reduce the risk of heat accumulation, ensure the stability of melt pool during multi-point scanning, and improve the forming quality and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of metal additive manufacturing, specifically relating to an electron beam powder bed spot melting scanning path planning method based on GA-BP neural network. Background Technology

[0002] In recent years, the demand for manufacturing complex structural components in aerospace, biomedicine, and other fields has been increasing, and traditional processing methods are struggling to meet the requirements. Electron beam powder bed melting (EBM), as a type of additive manufacturing technology, employs a "layer-by-layer, bottom-up" forming method, offering significant advantages in manufacturing complex structures. However, traditional electron beam line scanning suffers from high overlap of the heat-affected zones of adjacent scan lines, leading to continuous heat accumulation along the scanning path. This can easily cause localized overheating, uncontrolled molten pool, and consequently, warping or even forming failure. Unlike line scanning, EBM equipment uses magnetic lenses and deflection coils to control the electron beam, enabling microsecond-level beam spot switching. Based on this, the spot melting scanning strategy discretizes continuous scanning into a "jump-dwell-re-jump" pattern. The electron beam only pauses at the target position to form a single molten pool before rapidly jumping to the next position, fundamentally cutting off the continuous path of heat accumulation. By selectively jumping, it actively avoids high-temperature areas, effectively suppressing localized overheating and thermal stress concentration. Therefore, rational path planning is crucial in spot melting scanning. Currently, the key challenge of electron beam spot melting scanning technology lies in the fact that spot melting process parameters and path planning need to be formulated based on physical quantities such as the melting state of the powder, the cooling time of the molten pool, and the heat-affected zone. These physical quantities have a certain nonlinear relationship with the spot melting process parameters and path planning parameters, making it difficult to quickly determine them through experience or simple rules, and there is a lack of efficient path planning methods. Summary of the Invention

[0003] This invention addresses the problems in existing technologies by providing an electron beam powder bed spot melting scanning path planning method based on a GA-BP neural network. The aim is to solve the problem of inefficient spot melting scanning path planning while ensuring complete powder melting. The method specifically includes the following steps:

[0004] Step 1: Within the set process parameter range, generate multiple sets of different combinations of beam current and residence time parameters;

[0005] Step 2: Obtain the molten pool aspect ratio, cooling time, and heat-affected distance for the corresponding parameter combination through single-point spot melting simulation, and construct a training sample set with the beam current and residence time as inputs and the molten pool aspect ratio, cooling time, and heat-affected distance as outputs;

[0006] Step 3: Use the training sample set to train and obtain the GA-BP neural network model;

[0007] Step 4: Input the current beam current and residence time into the GA-BP neural network model to obtain predicted information on the molten pool aspect ratio, cooling time, and heat-affected zone.

[0008] Step 5: Adjust the residence time of the beam and the point according to the predicted aspect ratio of the molten pool;

[0009] Step 6: Based on the adjusted residence time of the beam and the point, as well as the corresponding cooling time and heat-affected zone, perform point melting scanning path planning;

[0010] Step 7: Perform multi-point simulation verification based on the path planning results.

[0011] Furthermore, in step 1, the parameter combination generation adopts the Latin hypercube sampling method.

[0012] Furthermore, in step 2, the single-point melting simulation uses ANSYS APDL parametric analysis language to perform finite element simulation of the temperature field, including establishing the geometric models of the powder bed and the substrate, assigning corresponding material properties to the powder bed and substrate regions, and applying the heat source load. The heat source formula is:

[0013]

[0014] In the formula, Q is the heat flux density, η is the thermal efficiency of the heat source, r is the distance between the heated node and the center of the heat source, and R is the radius of the electron beam spot.

[0015] Furthermore, in step 3, in the GA-BP neural network model, all the hyperparameters of the basic BP neural network are obtained automatically by Bayesian hyperparameter optimization search.

[0016] Furthermore, the GA-BP neural network model described in step 3 is obtained by using a genetic algorithm (GA) to perform a global optimal search for the weights of the basic BP neural network model and then training it based on the optimal weights.

[0017] Furthermore, in step 4, the GA-BP neural network model can predict the aspect ratio of the molten pool, cooling time, and heat-affected zone based on the input beam current and residence time.

[0018] Furthermore, the point melting scanning path planning described in step 6 mainly involves planning the number of jump points based on the heat-affected distance and determining the return time of the points based on the cooling time.

[0019] The advantages and beneficial effects of this invention are as follows: This invention provides an electron beam powder bed spot melting scanning path planning method based on a GA-BP neural network. This method is an offline spot melting scanning path planning method. By inputting specific spot melting process parameters into the GA-BP neural network model trained based on simulation data, the degree of powder melting under the current process parameters can be obtained. The process parameters can be adjusted in a timely manner to ensure complete powder melting. Under this condition, the planning information of the spot melting scanning path is provided, providing a decision basis for path planning. This method organically integrates physical simulation and neural network prediction, solving the problem of inefficient spot melting scanning path planning. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the invention method flow;

[0021] Figure 2 A finite element model for single-point simulation;

[0022] Figure 3 The curves represent the convergence curves during the training of the GA-BP neural network model; among them, Figure 3 (a) is the convergence curve of the weighted search. Figure 3 (b) is the convergence curve of the loss function;

[0023] Figure 4 The degree of data fit of the GA-BP neural network model on all training samples;

[0024] Figure 5 The morphology of the molten pool corresponds to three different degrees of powder melting; among them, Figure 5 (a) is in a state of insufficient melting. Figure 5 (b) is in a fully molten state. Figure 5 (c) is in the overmelted state;

[0025] Figure 6 Simulated molten pool cloud diagrams corresponding to three different powder melting degrees; among them... Figure 6 (a) is in a state of insufficient melting. Figure 6 (b) is in a fully molten state. Figure 6 (c) is in the overmelted state;

[0026] Figure 7 The simulation shows the relationship between the aspect ratio of the molten pool and the residence time and beam current; among which, Figure 7 (a) shows the relationship between the depth-to-width ratio of the molten pool and the residence time. Figure 7 (b) shows the relationship between the aspect ratio of the molten pool and the beam current.

[0027] Figure 8 For multi-point simulated cloud maps;

[0028] Figure 9This is the temperature-time history curve of a point in a multi-point simulation; where points A, B, and C are the scan points in adjacent scan sequences, and point D is the scan point adjacent to point A after the corresponding return time. Detailed Implementation

[0029] To better understand the technical solution of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Figure 1 The diagram shows a flowchart of an electron beam powder bed spot melting scanning path planning method based on a GA-BP neural network proposed in this invention, which is implemented according to the following steps:

[0031] Step 1: Select the residence time of the beam and points within a certain range, and use the Latin hypercube sampling method to generate multiple sets of different combinations of beam and residence time parameters. The number of sampling sets is selected to be 150 sets.

[0032] Step 2: Use the ANSYS APDL parametric analysis language to build a single-point simulation model in ANSYS software. For example... Figure 2 As shown, the dimensions of the powder layer and the substrate are 1000μm×1000μm×100μm and 1000μm×1000μm×300μm, respectively. During mesh generation, the mesh is refined in certain areas of the powder layer and the substrate. Both the powder layer and the substrate are made of TC4 titanium alloy in this invention. When applying the heat source, it is applied to the origin of the moving local coordinate system, and the heat source function is:

[0033]

[0034] In the formula, Q is the heat flux density, η is the heat source thermal efficiency, r is the distance between the heated node and the center of the heat source, and R is the electron beam spot radius, which is currently set to 180 μm.

[0035] Once the single-point simulation model is built, the beam current and residence time are input into the solution to obtain information such as the aspect ratio of the molten pool, cooling time, and heat-affected zone. This information will be used to construct a training sample set.

[0036] Step 3: Use the training sample set to train and obtain the GA-BP neural network model.

[0037] When partitioning the data, the ratio of the training set, test set, and validation set was 7:2:1. The hyperparameters of the basic BP neural network model were obtained by automatic search using Bayesian hyperparameter optimization, with an initial number of sampling points and 30 optimization iterations.

[0038] After training the basic BP neural network model, a genetic algorithm (GA) is used to perform a global optimal search for the weights of the basic BP neural network model. After obtaining the optimal weights, the model is then trained based on these weights. Figure 3 (a) and (b) show the convergence curve during weight search and the loss function curve during training, respectively. The genetic algorithm was set to a population size of 80, a generation count of 30, and a mutation rate of 0.2.

[0039] like Figure 4 It can be seen that the trained GA-BP neural network model has a good fit on all sample data.

[0040] Step 4: Input the current beam current and residence time into the GA-BP neural network model to obtain predicted information on the molten pool aspect ratio, cooling time, and heat-affected zone.

[0041] Step 5: Adjust the residence time of the beam and the point according to the predicted aspect ratio of the molten pool;

[0042] like Figure 5 (a), (b), and (c) show three different molten pool morphologies, representing different powder melting states: incomplete melting, complete melting, and over-melting. The corresponding simulation results are as follows: Figure 6 (a), (b), and (c) in the figure show that the melting state of the powder can be determined by the aspect ratio of the molten pool. Figure 7 (a) The aspect ratio of the molten pool for different residence times when the current is 4mA; Figure 7 (b) The aspect ratio of the molten pool for different beam currents with a residence time of 0.5 ms. This shows that there is a correlation between beam current and residence time and the aspect ratio of the molten pool. That is, the aspect ratio of the molten pool can be predicted by the beam current and residence time, and then the degree of powder melting can be judged and the process parameters adjusted according to the range of the aspect ratio. After parameter adjustment, the beam current of 4 mA and the residence time of 0.4 ms that can completely melt the powder are obtained, and the corresponding aspect ratio of the molten pool is 0.3115.

[0043] Step 6: Based on the adjusted residence time of the beam and the point, as well as the corresponding cooling time and heat-affected zone, perform point melting scanning path planning;

[0044] With a beam current of 4 mA and a residence time of 0.4 ms, the corresponding cooling time is 7.87 ms, and the heat-affected zone is 435.6 μm. Therefore, the distance corresponding to the number of jump points must be greater than or equal to 435.6 μm, and the return time should be greater than or equal to 7.87 ms.

[0045] Based on the current electron beam spot size and considering the impact of subsequent thermal diffusion, the number of jump points is set to 2, and the return time is set to 8ms.

[0046] Step 7: Perform multi-point simulation verification based on the path planning results.

[0047] Given the process parameters, number of jump points, and return time, a multi-point simulation experiment was designed to observe whether the molten pools affect each other under multi-point simulation conditions. Figure 8 The figure shows temperature cloud maps under multi-point simulation. It can be seen from the figure that the thermal influence between points is small; for example... Figure 9 The graph shows the temperature-time history curves for each point. Points A, B, and C are scan points in adjacent scan sequences, and point D is the scan point adjacent to point A after the corresponding return time. Figure 9 Looking at the temperature curves at points A, B, C, and D, the temperatures at each point are basically stable without significant increases, indicating that the temperatures of each molten pool are relatively independent.

Claims

1. A method for electron beam powder bed spot melting scanning path planning based on GA-BP neural network, characterized in that, Includes the following steps: Step 1: Within the set process parameter range, generate multiple sets of different combinations of beam current and residence time parameters; Step 2: Obtain the molten pool aspect ratio, cooling time, and heat-affected distance for the corresponding parameter combination through single-point spot melting simulation, and construct a training sample set with the beam current and residence time as inputs and the molten pool aspect ratio, cooling time, and heat-affected distance as outputs; Step 3: Use the training sample set to train and obtain the GA-BP neural network model; Step 4: Input the current beam current and residence time into the GA-BP neural network model to obtain predicted information on the molten pool aspect ratio, cooling time, and heat-affected zone. Step 5: Adjust the residence time of the beam and the point according to the predicted aspect ratio of the molten pool; Step 6: Based on the adjusted residence time of the beam and the point, as well as the corresponding cooling time and heat-affected zone, perform point melting scanning path planning; Step 7: Perform multi-point simulation verification based on the path planning results.

2. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network as described in claim 1, characterized in that: The parameter combination generation in step 1 uses the Latin hypercube sampling method.

3. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network as described in claim 1, characterized in that: The single-point melting simulation described in step 2 uses the ANSYS APDL parametric analysis language to perform finite element simulation of the temperature field. This includes establishing the geometric models of the powder bed and the substrate, assigning corresponding material properties to the powder bed and substrate regions, and applying the heat source load. The heat source formula is as follows: In the formula, Q is the heat flux density, η is the thermal efficiency of the heat source, r is the distance between the heated node and the center of the heat source, and R is the radius of the electron beam spot.

4. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network as described in claim 1, characterized in that: In the GA-BP neural network model described in step 3, all the hyperparameters of the basic BP neural network are obtained automatically through Bayesian hyperparameter optimization.

5. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network as described in claim 1, characterized in that: The GA-BP neural network model described in step 3 is obtained by using a genetic algorithm (GA) to perform a global optimal search for the weights of the basic BP neural network model and then training it based on the optimal weights.

6. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network according to claim 1, characterized in that: The GA-BP neural network model described in step 4 can predict the aspect ratio of the molten pool, cooling time, and heat-affected distance based on the input beam current and residence time.

7. The electron beam powder bed spot melting scanning path planning method based on GA-BP neural network according to claim 1, characterized in that: The point melting scanning path planning described in step 6 mainly involves planning the number of jump points based on the heat-affected distance and determining the return time of the points based on the cooling time.

Images