A method of fast screening scan paths to reduce local heat accumulation in additive manufacturing

Abstract

The application provides a method for quickly screening a scanning path to reduce local heat accumulation in additive manufacturing, and relates to the technical field of additive manufacturing.The application simulates temperature field data of different deposition sequences through a finite element method, and then the deposition sequences and the temperature field data corresponding to the deposition sequences constitute a sample set, so that the details and change rules of the temperature field can be accurately captured.The application trains a deep regression model through a training set in the sample set, and through the training and application of the deep regression model, a large amount of temperature field data under different working conditions can be quickly generated, and the calculation efficiency is greatly improved.The application only needs to input the deposition sequence into the deep regression model, so that the temperature field prediction results under different working conditions can be quickly obtained, the complexity and difficulty of operation are greatly reduced, the demand for computing resources is reduced, and the generation and analysis of the temperature field data can be automatically performed, so that the labor cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of additive manufacturing technology, specifically a method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing. Background Technology

[0002] Currently, the exploration of additive manufacturing processes still requires extensive simulations and experiments. However, if certain parameters change, both the overall model and experimental samples need to be re-prepared, which significantly limits the speed of optimization and iteration in additive manufacturing processes. It is necessary to develop a rapid predictive model, and machine learning models show great potential for application in additive manufacturing. Machine learning methods do not require creating and solving complex regression equations for heat transfer models. By training the model using data extracted from finite element simulation models, complex engineering problems can be readily solved.

[0003] For additive manufacturing processes, data acquisition is a major challenge for machine learning algorithms. Currently, the common approach is to extract data from experimental processes to build datasets. Two current application directions are molten pool morphology detection based on data acquired using coaxial infrared cameras and defect detection based on online acoustic signals. Currently available data types can be categorized as follows:

[0004] (1) Due to the rapid solidification of materials during additive manufacturing, gas porosity and poor fusion are caused. The main reason for the formation of porosity is the mismatch between powder supply and energy input. When the powder supply is high but the energy input is insufficient, a large amount of powder will be contained in the molten pool without being completely melted. These unmelted particles will greatly affect the mechanical properties.

[0005] (2) Dimensional errors in the additive manufacturing process are mainly caused by the shrinkage of the material after solidification. During the deposition process, the material shrinks as the molten pool cools. This process is constrained by the already deposited area, causing stress, and these changes will result in significant changes in geometry.

[0006] (3) The size of the molten pool is an important parameter in the additive manufacturing process. It is closely related to process parameters such as laser power and scanning speed, and is one of the main factors used to evaluate the microstructure. Regression models that use process parameters to obtain molten pool characteristics can better predict the microstructure and mechanical properties of parts.

[0007] (4) The thermal history of the deposited layer is the main means of analyzing the heat transfer mechanism in the additive manufacturing process. Controlling heat accumulation can effectively reduce deformation and crack formation. If the temperature near the molten pool is too low, the overlap between molten pools will be insufficient, leading to the formation of pores nearby. Increasing the scanning speed will result in a larger thermal gradient between the starting and ending points, leading to increased residual stress and crack risk. Excessive heat accumulation near the molten pool will cause microstructure coarsening, resulting in decreased mechanical properties.

[0008] Machine learning offers significant advantages in data processing, enabling regression and classification of large datasets collected during experiments. Based on the data type and quantity within the machine learning model, models can be categorized into supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. The first three methods differ primarily in whether the data is pre-labeled, while reinforcement learning employs a pre-designed reward / punishment mechanism for training. For additive manufacturing process optimization, offline instance-based data prediction and analysis are commonly used, with pre-labeled data. Supervised learning is generally suitable for this type of problem. In supervised learning, the model's classification or regression problem is considered. Classification methods are suitable for crack or microscopic metallographic identification, while regression methods are used for numerical prediction problems such as temperature field prediction.

[0009] The challenge of applying machine learning in additive manufacturing is establishing usable datasets. Current research relies on large amounts of real experimental data, which requires significant time and resources. Numerical modeling, as a mature computational method, can serve as a good data source containing physical information about the heat transfer process of laser energy in the deposited layer. Hosseini et al. used an unsupervised learning strategy to solve the heat transfer equation based on a single-pass deposition model, obtaining temperature curves under different combinations of process conditions. Li et al. established a Physically Neutral Network (PINN) model for single-pass deposition, which includes a basic training part and a transfer learning part, achieving accurate and efficient prediction of the temperature field. Zhao et al. constructed a dataset based on data from numerical prediction models and established a multilayer perceptron (MLP) deep learning model, achieving bidirectional prediction of melt pool size and process parameters.

[0010] In the deposition of large-sized parts, the temperature field under different scanning modes is also a crucial factor to consider during the manufacturing process. Roy et al. established a multi-layer deposition surrogate model based on machine learning, which can accurately predict the temperature field distribution of the GCode, deposition area, and heat-affected zone with low computational cost. Demir et al. and Ren et al. used a finite element model to establish an arbitrary path temperature field dataset, enabling the prediction of in-plane temperature fields under different scanning modes. Ogoke et al. proposed a deep reinforcement learning framework to obtain optimized scanning speed and power parameters to reduce melt pool variations during the scanning process. Ren et al. developed a temperature mode recurrent neural network model and used different scanning mode pairings between layers to build a dataset to select the optimal path planning to reduce local heat. Mozaffar et al. proposed a graph-based neural network model to predict the temperature field distribution of different part shapes during deposition.

[0011] Research on predicting temperature fields in additive manufacturing using deep learning methods, both domestically and internationally, has encountered some shortcomings and problems, along with the reasons for these problems. The following are some key points:

[0012] (1) Challenges in data acquisition: Deep regression models rely heavily on high-quality datasets, which are costly and challenging to obtain in the field of metal additive manufacturing. This is because data acquisition in the additive manufacturing process requires high-precision sensors and monitoring tools, which are often expensive and difficult to deploy and maintain in high-temperature, dynamic manufacturing environments.

[0013] (2) Accuracy of real-time prediction: Although deep regression models have made some progress in predicting temperature fields, their predictive performance is limited under real-time changing conditions. This is because the temperature field changes very rapidly and complexly during additive manufacturing, and the model may have difficulty capturing all the subtle changes, leading to a decrease in prediction accuracy.

[0014] (3) Model generalization ability: Deep regression models may perform well under specific conditions, but their generalization ability may be insufficient under different process parameters, geometries and deposition patterns. This is because model training is often based on specific datasets, and the model's predictive performance may decline when faced with new or unseen situations.

[0015] (4) Computational resources and time: While physics-based computational models offer accuracy, they are typically very time-consuming and unsuitable for real-time prediction and online control in iterative design scenarios. Deep regression models, while offering speed advantages, require significant computational resources and time to train.

[0016] (5) Model interpretability: Deep regression models are often considered "black box" models, as their decision-making process lacks transparency and interpretability. This is particularly problematic in additive manufacturing, as understanding and interpreting temperature field predictions is crucial for optimizing processes and improving product quality.

[0017] (6) Challenges of physical information fusion: Although the Physical Information Neural Network (PINN) framework combines physical principles and deep learning to accurately predict the temperature and dynamics of the molten pool, this method of fusion of physical information may face challenges in practical applications because it requires an accurate physical model and a large amount of labeled data to train the network.

[0018] (7) Accuracy of Defect Detection: Although deep learning methods have shown potential in defect detection, exhibiting high recognition ability for anomalies caused by porosity defects (accuracy exceeding 90%), the accuracy of defect detection in practical applications can be affected by various factors, including the complexity of molten pool features, environmental noise, and the training data of the model. In summary, while deep learning has made some progress in predicting temperature fields in additive manufacturing, it still faces challenges such as data acquisition, real-time prediction accuracy, model generalization ability, computational resources, model interpretability, and physical information fusion. Solving these problems requires interdisciplinary research efforts, including collaboration in sensor technology, data science, materials science, and computer science.

[0019] Currently used methods primarily rely on finite element method (FEM) simulation to model the temperature field of sedimentary layers. While traditional FEM has been validated over a long period and proven to provide relatively accurate temperature field simulations, the entire model needs to be recalculated if conditions change. Each model calculation takes several hours to several days, exhibiting computational efficiency bottlenecks that prevent its application to predicting a large number of sedimentary paths. Traditional temperature field simulation methods involve complex steps such as model building, mesh generation, and boundary condition setting, which are cumbersome and prone to errors. Furthermore, traditional temperature field simulation methods require significant computational resources and the involvement of specialized personnel, resulting in high costs. Summary of the Invention

[0020] To address the shortcomings of existing technologies, the present invention aims to provide a method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing, comprising:

[0021] Step 1: Construct a finite element model and obtain multiple random deposition sequences. For each deposition sequence, simulate the additive manufacturing process based on the finite element model to obtain temperature field data corresponding to each deposition sequence. The deposition sequence includes multiple deposition positions arranged in time.

[0022] Step 2: Use the deposition sequence as the input sample and the temperature field data corresponding to the deposition sequence as the output sample. The input sample and the output sample together form a training sample. Multiple training samples form a sample set. Divide the sample set into a training set and a validation set according to a preset ratio.

[0023] Step 3: Train the initial deep regression model using the training set, and validate the trained deep regression model using the validation set to obtain the trained deep regression model;

[0024] Step 4: Obtain multiple random deposition sequences, input each deposition sequence into the trained deep regression model, obtain the temperature field data corresponding to each deposition sequence, obtain the temperature uniformity of each deposition sequence based on the temperature field data, obtain the minimum value among all temperature uniformities, and take the deposition sequence corresponding to the minimum value as the optimal deposition sequence.

[0025] Optionally, step 1 specifically includes:

[0026] The finite element model irradiates the deposition site with a laser beam and simultaneously sprays powder onto the deposition site. Then, at the moment the powder solidifies, the temperature field data of the deposition site is extracted. Then, starting from the first deposition site in the deposition sequence, the above operation is repeated to obtain the temperature field data corresponding to each deposition site, and thus obtain the temperature field data corresponding to each deposition sequence.

[0027] Optionally, the deep regression model described in step 3 includes a max pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, upsampling, and skip connections.

[0028] Optionally, step 3 specifically includes:

[0029] The input samples are fed into the initial deep regression model. The input samples are encoded, decoded, and jump-connected to obtain the predicted temperature field data. Based on the predicted temperature field data and the output samples, the parameters in the initial deep regression model are updated. The trained deep regression model is validated through the validation set until the model converges, and the trained deep regression model is obtained.

[0030] Optionally, the input samples are encoded, decoded, and jump-connected to obtain the predicted temperature field data, including:

[0031] The deposition sequence is sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the first data. The first data is then sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the second data. The second data is then passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, a 3D convolutional layer, a batch normalization layer, an activation function, and upsampled to obtain the third data. The third data and the second data are then connected by a skip connection to obtain the fourth data. The fourth data is then upsampled to obtain the upsampled fourth data. The upsampled fourth data and the first data are then connected by a skip connection to obtain the fifth data. The upsampled fifth data is then upsampled to obtain the fifth data. The upsampled fifth data is then connected by a skip connection to the deposition sequence to obtain the predicted temperature field data.

[0032] Optionally, the third data and the second data have the same dimension, the upsampled fourth data and the first data have the same dimension, and the upsampled fifth data has the same dimension as the deposition order.

[0033] Optionally, in step 4, the temperature uniformity of each deposition sequence is obtained based on the temperature field data, including:

[0034] For the temperature field data of each deposition location in the deposition sequence, the average temperature variance is calculated, and the average temperature variance is used as the temperature uniformity.

[0035] The beneficial effects of adopting the above technical solution are as follows:

[0036] This invention simulates temperature field data with different deposition sequences using the finite element method. The deposition sequence and its corresponding temperature field data form a sample set, accurately capturing the detailed features and variation patterns of the temperature field, providing high-quality input data for subsequent training of a deep regression model. During training, the deep regression model learns the complex mapping relationships between a large number of temperature field data under different operating conditions, further improving the accuracy and quality of the generated temperature field data. This invention trains the deep regression model using the training set within the sample set, resulting in a trained deep regression model. Through the training and application of this model, a large amount of temperature field data under different operating conditions can be quickly generated, significantly improving computational efficiency. This enables rapid prediction and analysis of temperature fields under different operating conditions in practical engineering applications, providing strong support for engineering design and optimization. Compared to traditional temperature field simulation methods that require complex model building, mesh generation, and boundary condition settings, this invention only requires inputting the deposition sequence into the trained deep regression model to quickly obtain predicted temperature field data under different operating conditions. This greatly reduces the complexity and difficulty of operation, enabling non-technical personnel to easily predict and analyze temperature fields. It also reduces the demand for computing resources and automates the generation and analysis of temperature field data, thus reducing labor costs. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing, as described in an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of the data processing flow of the deep regression model in an embodiment of the present invention;

[0039] Figure 3 The graphs are the training accuracy curves of the SC-3DCAE model in the embodiments of the present invention, where (a) is the training model loss function curve and (b) is the model accuracy curve.

[0040] Figure 4 The diagrams show the comparison between finite element simulation results and model prediction results for several deposition sequences with the smallest variance in the embodiments of the present invention. (a) is a comparison between finite element simulation results and model prediction results for the first deposition sequence, (b) is a comparison between finite element simulation results and model prediction results for the second deposition sequence, and (c) is a comparison between finite element simulation results and model prediction results for the third deposition sequence. Detailed Implementation

[0041] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0042] Localized heat accumulation is a common technical challenge in additive manufacturing, especially in techniques such as laser-directed energy deposition. Heat accumulation can lead to overheating and deformation of the material, affecting the geometric accuracy and mechanical properties of the formed part. This heat accumulation can not only cause dimensional deviations but also potentially trigger defects such as cracks and stress concentrations, severely impacting the performance and reliability of the part.

[0043] To address this issue, this patent proposes a rapid screening method for optimizing scanning paths, utilizing a deep regression model for fast prediction and selection. Traditional scanning path planning methods typically rely on empirical rules and trial-and-error approaches, which are inefficient and ill-suited to complex manufacturing environments. In contrast, the deep regression model, trained on extensive historical data, can accurately predict heat accumulation under different scanning paths, thereby quickly selecting the optimal scanning path. This method not only improves the efficiency of scanning path planning but also significantly reduces heat accumulation, minimizing residual stress and deformation during the forming process. Furthermore, the scalability and flexibility of the deep regression model allow it to adapt to different materials and process parameters, offering broad application prospects. This optimization method can significantly improve the quality and performance of additively manufactured parts, promoting the application of additive manufacturing technology in aerospace, medical, and other fields.

[0044] For predicting temperature fields under different deposition paths in additive manufacturing, the use of deep regression models can significantly reduce computation time, enabling more efficient completion of tasks such as temperature field prediction. Deep regression models rely on large amounts of data and exhibit better performance in data classification and regression. Deep regression models have several advantages: First, they are highly capable, automatically extracting data features and avoiding tedious manual intervention and adjustments. Second, deep regression models have a large number of network layers, allowing for complex model function mapping within current computing power. Third, many ready-made frameworks are available for deep regression models, enabling computational functionality across multiple platforms. Finally, deep regression models can achieve rapid prediction; although training requires significant resources, predictions can be made quickly after training, making them ideal for deposition path selection required in this patent. Therefore, this invention provides a method for rapidly selecting scanning paths to reduce local heat accumulation in additive manufacturing, combined with... Figure 1 This may include the following steps:

[0045] Step 1: Construct a finite element model to obtain multiple random deposition sequences. For each deposition sequence, simulate the additive manufacturing process based on the finite element model to obtain the temperature field data corresponding to each deposition sequence. Figure 1 The temperature field simulation results in the data, wherein the deposition sequence includes multiple deposition locations arranged in chronological order;

[0046] The finite element model irradiates the deposition site with a laser beam and simultaneously sprays powder onto the deposition site. Then, at the moment the powder solidifies, the temperature field data of the deposition site is extracted. Then, starting from the first deposition site in the deposition sequence, the above operation is repeated to obtain the temperature field data corresponding to each deposition site, and thus the temperature field data corresponding to each deposition sequence. That is, after the above operation is performed at the first deposition site, the laser beam is irradiated at the second deposition site, and powder is also sprayed onto the second deposition site. Then, at the moment the powder solidifies, the temperature field data of the second deposition site is extracted, and the above operation is performed sequentially according to the deposition sequence.

[0047] In this invention, PyCharm software is used to incorporate each deposition sequence into the finite element model, enabling the execution of the aforementioned operations according to the deposition positions within the deposition sequence. The resulting model is... Figure 1 The input scanning strategy in the process.

[0048] In the laser-directed energy deposition additive manufacturing process, the laser and input powder are focused onto corresponding locations on the substrate, accompanied by melting and solidification processes. Heat source movement and material addition are implemented according to a predetermined deposition path procedure. To maintain consistency with actual thermophysical processes, the material parameters used are varied with temperature. To reduce computation time, a denser mesh arrangement is used near the deposited layer. The material parameters and boundary conditions have been described in previous work. The developed model can collect temperature field data at predetermined times and locations.

[0049] Step 2: Use the deposition sequence as the input sample and the temperature field data corresponding to the deposition sequence as the output sample. The input sample and the output sample together form the training sample. Multiple training samples form a sample set, i.e. Figure 1 The temperature dataset in the dataset is divided into a training set and a validation set according to a preset ratio.

[0050] Step 3: According to the scanning strategy, i.e. the scanning order, train the initial deep regression model based on the training set, and validate the trained deep regression model using the validation set to obtain the trained deep regression model.

[0051] The deep regression model includes a max pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, upsampling, and skip connections.

[0052] The model incorporates max-pooling downsampling and transposed convolutional upsampling to extract key features between temperature field data and the scanning pattern. This developed model can replace the finite element model for predicting temperature field data in additive manufacturing. The input scanning pattern is represented in three dimensions to match the size of the temperature field data. The encoding process primarily involves progressively extracting features from the input temperature field through a series of convolutional and pooling layers. Convolutional layers extract local features, while pooling layers reduce the spatial resolution of the feature map, thereby reducing computation and extracting higher-level abstract features. This process allows the network to capture crucial information from the data. Through multiple downsampling operations, the encoder gradually reduces the spatial size of the input temperature field data. This not only reduces computation but also allows the network to better handle the overall structure and contextual information of the image. The decoding process primarily involves progressively restoring the spatial resolution of the feature map through upsampling operations, gradually restoring the temperature field size to match the original input data. During decoding, skip connections are used to fuse the features extracted by the encoder with those in the decoder. This design allows the decoder to combine low-level detail information with high-level semantic information, thus restoring image details while maintaining an understanding of the overall structure. The model consists of three-dimensional convolutional layers, downsampling layers, and upsampling layers. Each convolutional layer is followed by a batch normalization layer and an activation function, namely the Corrected Linear Unit (ReLU). Input data undergoes encoding and decoding stages. In the encoding stage, convolutional layers progressively extract high-dimensional features, while downsampling layers increase the receptive field, capturing more feature representations with fewer parameters. After convolution, the model enhances its ability to represent the temperature field features within the deposition layer. The decoding stage consists of convolutional layers and upsampling layers. In the first two upsampling and convolutional layers, the number of feature channels is halved. After the last upsampling layer, the number of feature channels is further reduced to match the size of the temperature field data. The upsampling features obtained from each layer are merged with the corresponding downsampling features. The fused features continue to be input into the sampling process, preserving data details to achieve higher accuracy.

[0053] Based on this, in specific implementation, the present invention combines Figure 2 Specifically, this includes: inputting input samples into an initial deep regression model; encoding, decoding, and performing skip connections on the input samples to obtain predicted temperature field data; updating the parameters in the initial deep regression model based on the predicted temperature field data and output samples; validating the trained deep regression model using a validation set; and continuing until the model converges. Figure 1 The comparison and verification in the model yields the trained deep regression model, where the initial deep regression model can be the SC-3DCAE model.

[0054] The input samples are encoded, decoded, and jump-connected to obtain the predicted temperature field data, including:

[0055] The deposition sequence is sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the first data. The first data is then sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the second data. The second data is then passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, a 3D convolutional layer, a batch normalization layer, an activation function, and upsampled to obtain the third data. The third data and the second data are then connected by a skip connection to obtain the fourth data. The fourth data is then upsampled to obtain the upsampled fourth data. The upsampled fourth data and the first data are then connected by a skip connection to obtain the fifth data. The upsampled fifth data is then upsampled to obtain the fifth data. The upsampled fifth data is then connected by a skip connection to the deposition sequence to obtain the predicted temperature field data.

[0056] The third data and the second data have the same dimension, the upsampled fourth data and the first data have the same dimension, and the upsampled fifth data has the same dimension as the deposition order.

[0057] Step 4: Obtain multiple random deposition sequences, that is, randomly generate a large number of deposition sequences. Figure 1 A large amount of random scan path data is generated. Each deposition sequence is input into the trained depth regression model to obtain the temperature field data corresponding to each deposition sequence. Based on the temperature field data, the temperature uniformity of each deposition sequence is obtained. The minimum value among all temperature uniformities is obtained, and the deposition sequence corresponding to the minimum value is taken as the optimal deposition sequence.

[0058] For the temperature field data of each deposition location in the deposition sequence, the average temperature variance is calculated, and the average temperature variance is used as the temperature uniformity.

[0059] In other words, the optimal scanning path is selected and the predicted path is optimized based on the criterion of minimizing the temperature distribution variance at various locations within the sedimentary layer. To obtain lower heat accumulation and a more uniform temperature distribution, a large number of random paths need to be generated and input into the model. Existing models can quickly obtain the corresponding thermal data. The predicted temperature field is output, the average temperature variance is calculated, and the average temperature variance is used as the temperature uniformity. All temperature uniformities are then sorted. Figure 1 The temperature distribution variance is used for ranking, where a smaller variance indicates better uniformity, which can significantly reduce local heat accumulation in the sedimentary layer. The sedimentation sequence corresponding to the smallest temperature uniformity is taken as the optimal sedimentation sequence. Figure 1 The optimized scan path is obtained in the process.

[0060] Based on the above scheme, the present invention conducted the following experiments:

[0061] The temperature field data was divided into training and test sets in a 3:1 ratio. The training set was used to train the model and fit the data distribution pattern. The test set was used to evaluate the model's generalization ability and accuracy. The training loss curve and test accuracy curve on the test set during the training phase are shown below. Figure 3 As shown in the figure, (a) is the loss function curve of the training model, and (b) is the accuracy curve of the model. It can be seen that after the model training is completed and reaches a stable state, the final accuracy is calculated to be at a high level when compared with the test set that did not participate in the training process. This indicates that the prediction model can match the temperature field data well and has achieved a relatively stable training process.

[0062] This invention selects several scanning paths with the lowest variance, i.e., deposition sequences, as the scanning paths recommended by the model. The predicted results are then compared with the simulation results of the finite element model. Figure 4 As shown, (a) is a comparison diagram of the finite element simulation results and model prediction results for the first deposition sequence, (b) is a comparison diagram of the finite element simulation results and model prediction results for the second deposition sequence, and (c) is a comparison diagram of the finite element simulation results and model prediction results for the third deposition sequence. By comparing the data from the prediction model and the finite element simulation using contour plots, the accuracy of the prediction can be intuitively understood. It can be seen that the temperature distribution shown by the model prediction and the finite element simulation is quite close. A uniform temperature distribution not only reduces part deformation but also obtains a finer and more uniform microstructure distribution. The developed model can perform a simulation prediction for a specified deposition sequence within one millisecond. Compared to several hours for a single finite element model, this greatly improves efficiency and makes scan mode optimization possible.

[0063] This demonstrates that the present invention, through the finite element method, simulates temperature field data under different operating conditions, accurately capturing the detailed features and variation patterns of the temperature field, providing high-quality input data for the subsequent training of the deep regression model. During training, the deep regression model learns the complex mapping relationships between a large amount of temperature field data under different operating conditions, further improving the accuracy and quality of the generated temperature field data. In terms of efficiency, it offers faster computation speed: traditional temperature field simulation methods, such as finite element analysis, typically require significant computational resources and time, especially for large-scale complex operating condition simulations, where the calculation process may take hours or even days. This patent, through the training and application of the deep regression model, can quickly generate a large amount of temperature field data under different operating conditions, greatly improving computational efficiency. This allows for rapid prediction and analysis of temperature fields under different operating conditions in practical engineering applications, providing strong support for engineering design and optimization. Furthermore, this patent simplifies the operational process of temperature field prediction through the training and application of the deep regression model. Traditional temperature field simulation methods require complex steps such as model building, mesh generation, and boundary condition setting, which are cumbersome and prone to errors. This technology allows for rapid generation of predicted temperature field data under different operating conditions simply by inputting the deposition sequence into a trained depth regression model. This significantly reduces the complexity and difficulty of the operation, enabling even non-technical personnel to easily predict and analyze temperature fields. This patented method effectively reduces the cost of temperature field prediction. Traditional temperature field simulation methods require substantial computing resources and the involvement of professional personnel, resulting in high costs. Using this technology, however, the demand for computing resources is reduced, and the generation and analysis of temperature field data can be automated, lowering labor costs.

[0064] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing, characterized in that, include: Step 1: Construct a finite element model and obtain multiple random deposition sequences. For each deposition sequence, simulate the additive manufacturing process based on the finite element model to obtain temperature field data corresponding to each deposition sequence. The deposition sequence includes multiple deposition positions arranged in time. Step 2: Use the deposition sequence as the input sample and the temperature field data corresponding to the deposition sequence as the output sample. The input sample and the output sample together form a training sample. Multiple training samples form a sample set. Divide the sample set into a training set and a validation set according to a preset ratio. Step 3: Train the initial deep regression model using the training set, and validate the trained deep regression model using the validation set to obtain the trained deep regression model; Step 4: Obtain multiple random deposition sequences, input each deposition sequence into the trained deep regression model, obtain the temperature field data corresponding to each deposition sequence, obtain the temperature uniformity of each deposition sequence based on the temperature field data, obtain the minimum value among all temperature uniformities, and take the deposition sequence corresponding to the minimum value as the optimal deposition sequence.

2. The method for rapidly screening scanning paths to reduce local heat accumulation in additive manufacturing according to claim 1, characterized in that, Step 1 specifically includes: The finite element model irradiates the deposition site with a laser beam and simultaneously sprays powder onto the deposition site. Then, at the moment the powder solidifies, the temperature field data of the deposition site is extracted. Then, starting from the first deposition site in the deposition sequence, the above operation is repeated to obtain the temperature field data corresponding to each deposition site, and thus obtain the temperature field data corresponding to each deposition sequence.

3. The method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing according to claim 1, characterized in that, The deep regression model described in step 3 includes a max pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, upsampling, and skip connections.

4. The method for rapidly screening scanning paths to reduce local heat accumulation in additive manufacturing according to claim 1, characterized in that, Step 3 specifically includes: The input samples are fed into the initial deep regression model. The input samples are encoded, decoded, and jump-connected to obtain the predicted temperature field data. Based on the predicted temperature field data and the output samples, the parameters in the initial deep regression model are updated. The trained deep regression model is validated through the validation set until the model converges, and the trained deep regression model is obtained.

5. The method for rapidly screening scanning paths to reduce local heat accumulation in additive manufacturing according to claim 4, characterized in that, The input samples are encoded, decoded, and jump-connected to obtain the predicted temperature field data, including: The deposition sequence is sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the first data. The first data is then sequentially passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, and an activation function to obtain the second data. The second data is then passed through a max-pooling layer, a 3D convolutional layer, a batch normalization layer, an activation function, a 3D convolutional layer, a batch normalization layer, an activation function, and upsampled to obtain the third data. The third data and the second data are then connected by a skip connection to obtain the fourth data. The fourth data is then upsampled to obtain the upsampled fourth data. The upsampled fourth data and the first data are then connected by a skip connection to obtain the fifth data. The upsampled fifth data is then upsampled to obtain the fifth data. The upsampled fifth data is then connected by a skip connection to the deposition sequence to obtain the predicted temperature field data.

6. The method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing according to claim 5, characterized in that, The third data has the same dimension as the second data, the upsampled fourth data has the same dimension as the first data, and the upsampled fifth data has the same dimension as the deposition order.

7. The method for rapidly screening scanning paths to reduce localized heat accumulation in additive manufacturing according to claim 1, characterized in that, Step 4 involves obtaining the temperature uniformity of each deposition sequence based on the temperature field data, including: For the temperature field data of each deposition location in the deposition sequence, the average temperature variance is calculated, and the average temperature variance is used as the temperature uniformity.

Images