A deep learning-based metal 3D printing parameter prediction method and system
By constructing forward and backward prediction models based on deep learning, the process parameters of metal 3D printing can be monitored and adjusted in real time, solving the problem that traditional models cannot be adjusted in real time, thus improving printing quality and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-17
- Publication Date
- 2026-04-10
AI Technical Summary
In the existing metal 3D printing process, traditional unidirectional prediction models cannot control process parameters in real time, making it difficult to control microscopic defects and affecting printing quality and stability.
By collecting monitoring videos of the molten pool, we construct forward and backward prediction models based on deep learning, extract the interrelationship features of historical printing points using time series, establish the interrelationship length between laser path points, optimize model weight parameters, and realize real-time monitoring and control of molten pool area and process parameters.
It improves the stability of the internal structure of 3D printed parts, reduces the probability of defects, enhances prediction accuracy and printing efficiency, and adapts to material/process characteristics.
Smart Images
Figure CN121118709B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of additive manufacturing, in particular to a metal 3D printing parameter prediction method and system based on deep learning. BACKGROUND
[0002] The metal additive manufacturing process is a multi-physical and multi-scale process, including material-laser interaction at the microscale, melt pool dynamics at the mesoscale, and thermal-mechanical coupling at the macroscale. However, the metal additive manufacturing technology generally faces the severe challenge of micro-defects being difficult to control in practical applications. Taking the additive manufacturing of turbine blades for aerospace engines as an example, under continuous high temperature and high pressure conditions, stress concentration effects caused by internal keyhole porosities in the part will lead to rapid crack propagation, and interlayer un-melted defects will directly affect the air tightness and structural strength of the blade. The formation of these micro-defects is closely related to the dynamic behavior of the melt pool during the processing: when the laser power density is too high or the scanning speed is too fast, the melt pool surface will produce violent vaporization phenomena, forming a keyhole structure with an abnormal depth-width ratio, and the gas wrapped by liquid metal during the cooling process cannot escape in time; on the contrary, when the energy input is insufficient, the powder particles cannot be completely melted, resulting in un-melted interfaces between adjacent scanning passes.
[0003] In order to solve the problem of micro-defects being difficult to control, it is necessary to conduct in-depth research on the parameter setting of the 3D printing process. Currently, a forward prediction model of a deep learning network is generally constructed by collecting melt pool information to predict the melt pool area. Although the forward prediction model performs well in melt pool area prediction, it only supports the prediction of process parameters to melt pool area, and it is difficult to realize real-time prediction and active reverse control of process parameters, which limits its practical printing application. In addition, process parameters will be adjusted as the printing requirements change (such as changes in internal structural characteristics of the part, printing heat accumulation, etc.), which in turn leads to changes in the required melt pool area. The melt pool area will directly affect the internal characteristics of the printed part, resulting in an increase in internal structural defects and a decrease in stability during 3D printing. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application aims to provide a metal 3D printing parameter prediction method and system based on deep learning, which aims to solve the technical problem that the traditional one-way prediction model in the prior art cannot realize real-time reverse control of process parameters.
[0005] The present application is directed to a metal 3D printing parameter prediction method based on deep learning, which comprises:
[0006] Collect monitoring videos of the molten pool in the additive manufacturing process under different process parameters, extract the interrelation features of the historical printing points according to the time sequence, establish the interrelation length between the laser path points, construct the mapping data set between the physical parameters of the laser path points and the molten pool area, and the physical parameters of the laser path points include process parameters and historical path parameters;
[0007] Based on the mapping data set, a forward prediction model is constructed by deep learning with the physical parameters of the laser path points as the input, and the predicted molten pool area is output;
[0008] The molten pool area and the historical path parameters are taken as the input, a backward prediction model is constructed by deep learning, the initial process parameters are output, the forward prediction model is taken as the optimizer to train the backward prediction model, the weight parameters of the forward prediction model and the backward prediction model are iteratively optimized, and the intermediate iterative process parameters are output;
[0009] The interrelation length is adjusted through an optimization algorithm, and then the forward prediction model and the backward prediction model are repeatedly trained according to the adjusted interrelation length, until the optimal interrelation length and the optimal process parameters are output.
[0010] Compared with the prior art, the beneficial effects of the present application are that, by the metal 3D printing parameter prediction method based on deep learning provided by the present application, the interrelation features of the historical printing points are extracted according to the time sequence, the interrelation length between the laser path points is established, the heat accumulation of the previous historical printing points is captured to affect the molten pool effect, then the forward prediction model learns the influence of the heat accumulation on the molten pool, and provides a reliable verification benchmark for the backward prediction model, then the forward prediction model is taken as the optimizer of the backward prediction model, and the backward prediction model is verified based on the forward prediction model, so as to avoid the deviation of the theoretical process parameters from the actual thermal field, adjust the next process parameters in time according to the previous molten pool situation, realize online monitoring and regulation of the printing quality, effectively improve the stability of the internal structure of the printed part, and reduce the defect generation probability, finally the interrelation length is iteratively adjusted to adapt to the 3D printing material / process characteristics, and the prediction accuracy is improved, so that the technical problem that the traditional one-way prediction model cannot realize real-time reverse regulation of the process parameters in the prior art is solved.
[0011] According to one aspect of the above technical solution, the molten pool monitoring videos in the additive manufacturing process under different process parameters are collected, the interrelation features of the historical printing points are extracted according to the time sequence, and the mapping data set between the physical parameters of the laser path points and the molten pool area is constructed, and the physical parameters of the laser path points include process parameters and historical path parameters, and the steps specifically include:
[0012] Collecting the molten pool monitoring videos in the additive manufacturing process under different process parameters, and pre-processing the data;
[0013] The relationship features between historical print points are extracted from the preprocessed molten pool monitoring video based on the time series, as shown below:
[0014] ,
[0015] in, As a characteristic of mutual relationship, For the first The laser power of the current printing point For the first Before the current print point, the th Laser power of each historical printing point For the first The distance between the current print point and the current print point where the laser started. For the first The current print point and the previous print point The distance between historical print points;
[0016] Based on the aforementioned mutual characteristic relationships, the correlation lengths between laser path points are established, and a mapping dataset between the physical parameters of laser path points and the molten pool area is constructed. The physical parameters of the laser path points include process parameters and historical path parameters.
[0017] According to one aspect of the above technical solution, the preprocessing step specifically includes:
[0018] Perform interpolation and noise reduction on the data;
[0019] The data is normalized and standardized.
[0020] According to one aspect of the above technical solution, the forward prediction model is an MLP model, which includes an input layer, a fully connected layer, and an output layer connected in sequence. There are three fully connected layers, and each layer includes a ReLU activation function. The learning rate of the forward prediction model is 0.0001.
[0021] According to one aspect of the above technical solution, the positive loss function of the positive prediction model is:
[0022] ,
[0023] in, It is a positive prediction model. The first output of the positive prediction model The predicted melt pool area for the current print point. This represents the total number of print points currently being printed. For the first The physical parameters of the laser path point of the current printing point For the first The actual melt pool area of the current print point This represents the mean square error.
[0024] According to one aspect of the above technical solution, the reverse prediction model is an MLP model, and the structure of the reverse prediction model is the same as that of the forward prediction model. The input matrix of the reverse prediction model is:
[0025] ,
[0026] in, For the input matrix, For the first The actual melt pool area of the current print point For the first Before the current print point, the th Laser power of each historical printing point For the first The distance between the current print point and the current print point where the laser started. For the first The current print point and the previous print point The distance between historical print points.
[0027] According to one aspect of the above technical solution, the steps of taking the molten pool area and historical path parameters as input, constructing a reverse prediction model through deep learning, outputting initial process parameters, and using the forward prediction model as an optimizer to train the reverse prediction model, iteratively optimizing the weight parameters of the forward and reverse prediction models, and outputting intermediate iterative process parameters, specifically include:
[0028] Using the molten pool area and historical path parameters as input, a reverse prediction model is constructed through deep learning to output the initial process parameters;
[0029] The forward prediction model is used as an optimizer to train the backward prediction model. The initial process parameters and historical path parameters are input into the forward prediction model to obtain the new predicted molten pool area.
[0030] The loss value is calculated by combining the new predicted molten pool area with the molten pool area input to the inverse prediction model using the inverse loss function.
[0031] Continue repeating the above steps to iteratively optimize the weight parameters of the forward prediction model and the backward prediction model until the iteration converges or reaches the preset number of iterations, and output the intermediate iterative process parameters.
[0032] According to one aspect of the above technical solution, the formula for calculating the inverse loss function is as follows:
[0033] ,
[0034] in, for a reverse prediction model, for a forward prediction model the first new predicted molten pool area of the current printing point, representing the total number of current printing points, for the real molten pool area of the printing point, representing the mean square error.
[0035] According to an aspect of the above technical solution, the step of repeatedly training the forward prediction model and the reverse prediction model until the optimal mutual correlation length and the optimal process parameters are output, specifically includes:
[0036] The forward prediction model and the reverse prediction model are repeatedly trained by the optimization model until the performance of the optimization model is optimized, and the optimal mutual correlation length and the optimal process parameters are output.
[0037] Another aspect of the present application is to provide a deep learning-based metal 3D printing parameter prediction system, which is used to implement the above deep learning-based metal 3D printing parameter prediction method, and the system comprises:
[0038] A data set construction module is configured to collect molten pool monitoring videos in an additive manufacturing process under different process parameters, extract interrelation features of historical printing points according to time series, establish mutual correlation lengths between laser path points, and construct a mapping data set between laser path point physical parameters and molten pool areas, wherein the laser path point physical parameters include process parameters and historical path parameters.
[0039] A forward prediction model construction module is configured to construct a forward prediction model based on the mapping data set by deep learning with the laser path point physical parameters as input, and output a predicted molten pool area.
[0040] A reverse prediction model construction module is configured to construct a reverse prediction model by deep learning with the molten pool area and the historical path parameters as input, and output initial process parameters, and train the reverse prediction model with the forward prediction model as an optimizer, iteratively optimize the weight parameters of the forward prediction model and the reverse prediction model, and output intermediate iteration process parameters.
[0041] A result output module is configured to adjust the mutual correlation length by an optimization algorithm, repeatedly train the forward prediction model and the reverse prediction model according to the adjusted mutual correlation length, and output the optimal mutual correlation length and the optimal process parameters until the optimal mutual correlation length and the optimal process parameters are output. BRIEF DESCRIPTION OF DRAWINGS
[0042] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings:
[0043] Figure 1 The downward trend graph of the forward loss function in the training process of the forward prediction model in Embodiment One of the present application;
[0044] Figure 2 The downward trend graph of the forward loss function in the training process of the reverse prediction model in Embodiment One of the present application;
[0045] Figure 3 The comparison graph of the predicted molten pool area by the forward prediction model in Embodiment One of the present application and the molten pool area.
[0046] Figure 4 The comparison graph of the new predicted molten pool area by the reverse prediction model in Embodiment One of the present application and the molten pool area. DETAILED DESCRIPTION
[0047] In order to make the objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Several embodiments of the present application are given in the drawings. However, the present application can be realized in many different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.
[0048] Embodiment One
[0049] The first embodiment of the present application provides a metal 3D printing parameter prediction method based on deep learning, which comprises steps S10-S13:
[0050] Step S10, collect molten pool monitoring videos in the additive manufacturing process under different process parameters, extract the mutual relationship features of historical printing points according to the time sequence, establish the mutual correlation length between laser path points, and construct the mapping data set between the physical parameters of laser path points and the molten pool area, wherein the physical parameters of laser path points include process parameters and historical path parameters;
[0051] Specifically, collect molten pool monitoring videos in the additive manufacturing process under different process parameters, and pre-process the data;
[0052] By way of example but not limitation, for example, experimental data and collected data can be fused through simulation data, Comsol simulation software is used to set up a three-dimensional heat transfer model, material parameters, process parameters, initial conditions, boundary conditions and grid conditions, and the molten pool area under different process parameters is calculated. The simulation data is used to make up for the deficiency of the experimental data, save the test cost and time. At the same time, the molten pool data corresponding to different printing process parameters can also be obtained in the simulation environment, which greatly enriches the data diversity, so as to obtain a better model.
[0053] Furthermore, the preprocessing steps specifically include:
[0054] The data is interpolated and denoised; interpolation completes missing data to ensure data continuity, and denoising filters out interference (such as laser noise and parameter fluctuation noise) to improve data purity.
[0055] The data is normalized and standardized. Normalization and standardization unify the units of measurement to facilitate deep learning in subsequent forward and backward prediction models.
[0056] Next, the relationship features of historical printing points are extracted from the preprocessed molten pool monitoring video based on the time series, as shown below:
[0057] ,
[0058] in, As a characteristic of mutual relationship, For the first The laser power of the current printing point For the first Before the current print point, the th Laser power of each historical printing point For the first The distance between the current print point and the current print point where the laser started. For the first The current print point and the previous print point The distance between historical print points.
[0059] It should be noted that the heat from previous scan points (previous historical printing points) will affect the current molten pool. It is necessary to capture the spatiotemporal effect of heat accumulation in 3D printing. That is, by defining the number of historical printing points that the current printing point needs to refer to through the interrelated length (such as 25 points), the time series features are transformed into a structured input that the model can learn, providing training data with historical memory for the subsequent bidirectional model.
[0060] Based on the aforementioned mutual characteristic relationships, the correlation lengths between laser path points are established, and a mapping dataset between the physical parameters of laser path points and the molten pool area is constructed. The physical parameters of the laser path points include process parameters and historical path parameters.
[0061] Examples, not limitations, are used to define 25 points of interrelated length. The input to the mapping dataset includes laser power, laser power of historical printing points before the current printing point, distance from the current printing point to the current printing point where the laser started, and distances from the current printing point to the previous 25 historical printing points. The label is the melt pool area.
[0062] Step S11: Based on the mapping dataset, a positive prediction model is constructed using deep learning with the physical parameters of the laser path points as input, and the predicted molten pool area is output.
[0063] Specifically, the positive prediction model is an MLP model, which includes an input layer, a fully connected layer, and an output layer connected in sequence. There are three fully connected layers, and each layer includes a ReLU activation function. The learning rate of the positive prediction model is 0.0001.
[0064] Among them, the multilayer perceptron (MLP) model is used to process structured numerical inputs (process parameters and historical path parameters), and three fully connected layers and ReLU activation function are used to achieve nonlinear fitting and capture complex interactions between parameters.
[0065] For example, not a limitation, the input layer has 77 nodes, a ReLU activation function, and an L2 regularization of 0.001; the first fully connected layer has 128 nodes, a ReLU activation function, an L2 regularization of 0.001, and a dropout rate of 0.2; the second fully connected layer has 256 nodes, a ReLU activation function, an L2 regularization of 0.001, and a dropout rate of 0.2; the third fully connected layer has 64 nodes, a ReLU activation function, an L2 regularization of 0.001, and a dropout rate of 0.2; and the output layer has 1 node.
[0066] Furthermore, the positive loss function of the positive prediction model is:
[0067] ,
[0068] in, It is a positive prediction model. The first output of the positive prediction model The predicted melt pool area for the current print point. This represents the total number of print points currently being printed. For the first The physical parameters of the laser path point of the current printing point For the first The actual melt pool area of the current print point This represents the mean square error.
[0069] It should be noted that the input includes process parameters and historical path parameters. The forward prediction model learns the impact of thermal accumulation on the molten pool, that is, the residual heat from previous historical printing points changes the current temperature field, thereby changing the molten pool area. The output is the predicted molten pool area, and the deviation between the predicted molten pool area and the actual molten pool area is quantified by mean square error. This drives the forward prediction model to learn accurate mapping, providing a reliable validation benchmark for the reverse prediction model.
[0070] In addition, the forward prediction model is pre-trained before the reverse prediction model, and after being trained, serves as an optimizer of the reverse prediction model.
[0071] In step S12, the pool area and the historical path parameters are taken as inputs, a reverse prediction model is constructed through deep learning, initial process parameters are output, the forward prediction model is taken as an optimizer to train the reverse prediction model, weight parameters of the forward prediction model and the reverse prediction model are iteratively optimized, and intermediate iteration process parameters are output.
[0072] Specifically, the reverse prediction model is an MLP model, the structure of the reverse prediction model is the same as that of the forward prediction model, and the input matrix of the reverse prediction model is:
[0073] ,
[0074] wherein, is the input matrix, is a real pool area of an i-th current printing point, is a laser power of an i-1-th historical printing point before the i-th current printing point, is a distance between the i-th current printing point and a current printing point at which the laser starts, is a distance between the i-th current printing point and an i-1-th historical printing point.
[0075] Further, the reverse prediction model also includes an input layer, a full connection layer and an output layer connected in sequence. The symmetry of bidirectional mapping is ensured, the model design complexity is reduced, and the feature extraction capability of the forward model is utilized.
[0076] Specifically, the pool area and the historical path parameters are taken as inputs, a reverse prediction model is constructed through deep learning, and initial process parameters are output.
[0077] wherein, the inputs are the pool area and the historical path parameters, the reverse prediction model is used for backstepping, and the initial process parameters are output.
[0078] Then, the forward prediction model is taken as an optimizer to train the reverse prediction model, the initial process parameters and the historical path parameters are input into the forward prediction model, and new predicted pool areas are obtained.
[0079] The new predicted pool areas and the pool areas input into the reverse prediction model are used to calculate loss values through a reverse loss function.
[0080] Continue repeating the above steps to iteratively optimize the weight parameters of the forward prediction model and the backward prediction model until the iteration converges or reaches the preset number of iterations, and output the intermediate iterative process parameters.
[0081] The formula for calculating the inverse loss function is as follows:
[0082] ,
[0083] in, It is a reverse prediction model. For positive prediction model The output of the first The new predicted melt pool area for the current print point. This indicates the total number of print points currently being printed. For the first The actual melt pool area of each print point. This represents the mean square error.
[0084] It should be noted that the closed-loop training of reverse → forward → error feedback ensures that the process parameters output by the reverse prediction model are effective in real printing, avoids the deviation between theoretical process parameters and actual thermal field, and adjusts the next process parameters in a timely manner according to the previous melt pool condition, realizing online monitoring and control of printing quality, effectively improving the stability of the internal structure of the printed parts and reducing the probability of defects.
[0085] Furthermore, the forward prediction model, as an optimizer for the backward prediction model, can, to some extent, resolve the complex many-to-one relationship between the molten pool area and the printing process parameters, solve the fuzziness problem of "many-to-one" (the same molten pool corresponds to multiple parameters), select the optimal solution, greatly reduce printing costs, and significantly improve printing efficiency.
[0086] according to Figures 1-2 As can be seen, the solid and dashed lines represent the changes in loss on the training and validation sets, respectively. Both lines show a downward trend, indicating that the forward and backward prediction models are gradually learning the patterns in the data and performing well on both the training and validation sets, demonstrating good generalization ability.
[0087] according to Figure 3 As can be seen, when comparing the predicted molten pool area obtained by the positive prediction model with the actual molten pool area, the points in the figure are all relatively close to the midline, which indicates that the prediction effect is good.
[0088] according to Figure 4It can be seen that the initial process parameters are first predicted according to the molten pool area by using the reverse prediction model, then the initial process parameters are taken as the input of the forward prediction model to obtain the new predicted molten pool area, and finally the new predicted molten pool area is compared with the real molten pool area to obtain the result. The points in the figure are close to the center line, which can indicate that the initial process parameters obtained by the reverse prediction model are close to the real process parameters. Because the process parameters and the molten pool area are one-to-one under the condition that all conditions remain unchanged, and the reverse prediction model can learn this relationship well. Therefore, it can be proved that the prediction effect of the reverse prediction model is good.
[0089] In step S13, the inter-association length is adjusted by an optimization algorithm, and then the forward prediction model and the reverse prediction model are repeatedly trained according to the adjusted inter-association length, until the optimal inter-association length and the optimal process parameters are output.
[0090] Specifically, the inter-association length is adjusted by an optimization algorithm, and then the forward prediction model and the reverse prediction model are repeatedly trained by the optimization model according to the adjusted inter-association length determined each time, until the performance of the optimization model is optimized, and the optimal inter-association length and the optimal process parameters are output.
[0091] In this embodiment, the performance of the optimization model is optimal when the inter-association length is 25 points. If the inter-association length is too short, the key heat accumulation will be ignored, which will lead to molten pool prediction bias, and if the inter-association length is too long, redundant information will be introduced, which will increase the model complexity.
[0092] Further, through the iteration of adjusting the length, retraining the model, and evaluating the precision, the optimal association length that adapts to the current printing material / process is found, so that the model can more accurately capture the heat field dynamics.
[0093] Compared with the prior art, by using the deep learning-based metal 3D printing parameter prediction method shown in the embodiment, the inter-relationship features of the historical printing points are extracted through time series, the inter-association length between the laser path points is established, the heat accumulation of the previous historical printing points is captured to affect the molten pool effect, then the forward prediction model is used to deeply learn the influence of the heat accumulation on the molten pool to provide a reliable verification benchmark for the reverse prediction model, then the forward prediction model is used as the optimizer of the reverse prediction model, and the reverse prediction model relies on the verification of the forward prediction model to avoid the deviation between the theoretical process parameters and the actual heat field, and the next process parameter is adjusted in time according to the previous molten pool situation, so as to realize online monitoring and regulation of the printing quality, effectively improve the stability of the internal structure of the printed part, and reduce the probability of defect generation. Finally, the inter-association length is iteratively adjusted to adapt to the 3D printing material / process characteristics, and the prediction accuracy is improved, thereby solving the technical problem that the traditional one-way prediction model in the prior art cannot realize real-time reverse regulation of the process parameters.
[0094] Embodiment two
[0095] The second embodiment of the present application provides a deep learning-based metal 3D printing parameter prediction system, which comprises:
[0096] A data set construction module is configured to collect molten pool monitoring videos in an additive manufacturing process under different process parameters, extract correlation features of historical printing points according to a time sequence, establish a mutual correlation length between laser path points, and construct a mapping data set between laser path point physical parameters and molten pool area, wherein the laser path point physical parameters include process parameters and historical path parameters.
[0097] A forward prediction model construction module is configured to construct a forward prediction model by deep learning based on the mapping data set, taking the laser path point physical parameters as input, and output a predicted molten pool area.
[0098] A reverse prediction model construction module is configured to take the molten pool area and the historical path parameters as input, construct a reverse prediction model by deep learning, output an initial process parameter, and train the reverse prediction model by taking the forward prediction model as an optimizer, iteratively optimize the weight parameters of the forward prediction model and the reverse prediction model, and output an intermediate iteration process parameter.
[0099] A result output module is configured to adjust the mutual correlation length by an optimization algorithm, and repeatedly train the forward prediction model and the reverse prediction model according to the adjusted mutual correlation length until an optimal mutual correlation length and an optimal process parameter are output.
[0100] Compared with the prior art, the deep learning-based metal 3D printing parameter prediction system shown in the embodiment can extract the correlation features of historical printing points according to a time sequence by the data set construction module, establish a mutual correlation length between laser path points, capture the heat accumulation of previous historical printing points affecting the molten pool effect, and then let the forward prediction model learn the influence of heat accumulation on the molten pool by the forward prediction model construction module, so as to provide a reliable verification benchmark for the reverse prediction model. Then, the forward prediction model is taken as an optimizer of the reverse prediction model by the reverse prediction model construction module, and the reverse prediction model is verified based on the forward prediction model to avoid the deviation of the theoretical process parameter from the actual thermal field. The next process parameter is adjusted in time according to the previous molten pool situation to realize online monitoring and regulation of printing quality, effectively improve the stability of the internal structure of the printed part, and reduce the probability of defect generation. Finally, the mutual correlation length is iteratively adjusted by the result output module to adapt to the 3D printing material / process characteristics and improve the prediction accuracy, thereby solving the technical problem that the traditional one-way prediction model in the prior art cannot realize real-time reverse regulation of the process parameter.
[0101] The technical features of each of the above embodiments can be combined in any manner. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not contradict each other, they should be considered as falling within the scope of the present application.
[0102] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0103] The above described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.
Claims
1.A deep learning based metal 3D printing parameter prediction method, characterized in that, The method comprises: Collecting molten pool monitoring videos in the additive manufacturing process under different process parameters, extracting the correlation features of historical printing points according to the time sequence, establishing the mutual correlation length between laser path points, and constructing a mapping data set between the physical parameters of the laser path points and the molten pool area, wherein the physical parameters of the laser path points include process parameters and historical path parameters, including: Collecting molten pool monitoring videos in the additive manufacturing process under different process parameters, and preprocessing the data, According to the time sequence, the correlation features of historical printing points are extracted from the preprocessed molten pool monitoring video, and are expressed as follows: ， wherein, is a mutual relationship feature, is a first laser power of a current print point, is a second laser power of a previous print point of the current print point, is a third laser power of a previous history print point of the current print point, is a fourth distance of the current print point from a start of the laser, is a fifth distance of the current print point from the previous history print point, According to the correlation features, the mutual correlation length between laser path points is established, and a mapping data set between the physical parameters of the laser path points and the molten pool area is constructed, wherein the physical parameters of the laser path points include process parameters and historical path parameters; Based on the mapping data set, a forward prediction model is constructed by deep learning, taking the physical parameters of the laser path points as input, and outputting the predicted molten pool area; Taking the molten pool area and the historical path parameters as input, a backward prediction model is constructed by deep learning, and the initial process parameters are outputted, and the forward prediction model is used as an optimizer to train the backward prediction model, and the weight parameters of the forward prediction model and the backward prediction model are iteratively optimized, and the intermediate iteration process parameters are outputted, including: Taking the molten pool area and the historical path parameters as input, a backward prediction model is constructed by deep learning, and the initial process parameters are outputted, The forward prediction model is used as an optimizer to train the backward prediction model, the initial process parameters and the historical path parameters are inputted into the forward prediction model, and a new predicted molten pool area is obtained, The loss value is calculated by a backward loss function based on the new predicted molten pool area and the molten pool area inputted into the backward prediction model, The above steps are repeatedly continued, the weight parameters of the forward prediction model and the backward prediction model are iteratively optimized until the iteration converges or a preset number of times is reached, and the intermediate iteration process parameters are outputted; The mutual correlation length is adjusted by an optimization algorithm, and the forward prediction model and the backward prediction model are repeatedly trained according to the adjusted mutual correlation length until the optimal mutual correlation length and the optimal process parameters are outputted, including: The mutual correlation length is adjusted by an optimization algorithm, and the forward prediction model and the backward prediction model are repeatedly trained according to the adjusted mutual correlation length until the optimal mutual correlation length and the optimal process parameters are outputted. 2.The deep learning-based metal 3D printing parameter prediction method of claim 1, wherein, The preprocessing step specifically comprises: Interpolation and noise reduction processing are performed on the data; Normalization and standardization processing are performed on the data. 3.The deep learning-based metal 3D printing parameter prediction method of claim 1, wherein, The forward prediction model is an MLP model, which comprises an input layer, a fully connected layer and an output layer connected in sequence, the fully connected layer has three layers, each layer includes a Relu activation function, and the learning rate of the forward prediction model is 0.0001. 4.The deep learning-based metal 3D printing parameter prediction method of claim 1, wherein, The forward loss function of the forward prediction model is: ， wherein, is a forward prediction model, is a predicted molten pool area of the i-th current printing point output by the forward prediction model, is a predicted molten pool area of the i-th current printing point output by the forward prediction model, is a total number of current printing points, is a laser path point physical quantity of the i-th current printing point, is a laser path point physical quantity of the i-th current printing point, is a true molten pool area of the i-th current printing point, is a true molten pool area of the i-th current printing point, denotes a mean squared error. 5.The deep learning-based metal 3D printing parameter prediction method of claim 3, wherein, The backward prediction model is an MLP model, the structure of the backward prediction model is the same as that of the forward prediction model, and the input matrix of the backward prediction model is: ， in, For the input matrix, For the first The actual melt pool area of the current print point. For the first Before the current print point, the th Laser power of each historical printing point For the first The distance between the current print point and the current print point where the laser started. For the first The current print point and the previous print point The distance between historical print points. 6.The deep learning-based metal 3D printing parameter prediction method of claim 5, wherein, The calculation formula of the backward loss function is as follows: ， wherein, is a backward prediction model, is a forward prediction model the output of a new predicted pool area for the current printing point, denotes the total number of current printing points, is the real pool area for the current printing point, denotes the mean square error. 7.A deep learning based metal 3D printing parameter prediction system, characterized in that, The system is used to realize the deep learning-based metal 3D printing parameter prediction method in any one of claims 1 to 6, and the system comprises: A data set construction module is configured to collect molten pool monitoring videos in an additive manufacturing process under different process parameters, extract the correlation features of historical printing points according to a time sequence, establish the interrelated length between laser path points, and construct a mapping data set between the physical parameters of the laser path points and the molten pool area, wherein the physical parameters of the laser path points include process parameters and historical path parameters, and the process parameters include: Collecting molten pool monitoring videos in an additive manufacturing process under different process parameters, and pre-processing the data, According to the correlation features, the interrelated length between laser path points is established, and a mapping data set between the physical parameters of the laser path points and the molten pool area is constructed, wherein the physical parameters of the laser path points include process parameters and historical path parameters; ， wherein, is a mutual relationship feature, is a first laser power of a current print point, is a second laser power of a previous print point of the current print point, is a third laser power of a previous historical print point of the current print point, is a fourth distance of the current print point from a current print point where the laser started, is a fifth distance of the current print point from the previous historical print point, A forward prediction model construction module is configured to construct a forward prediction model by deep learning based on the mapping data set, taking the physical parameters of the laser path points as input, and outputting a predicted molten pool area; A reverse prediction model construction module is configured to take the molten pool area and the historical path parameters as input, construct a reverse prediction model by deep learning, output initial process parameters, and train the reverse prediction model by taking the forward prediction model as an optimizer, iteratively optimize the weight parameters of the forward prediction model and the reverse prediction model, and output intermediate iteration process parameters, including: Taking the molten pool area and the historical path parameters as input, constructing a reverse prediction model by deep learning, and outputting initial process parameters, Taking the forward prediction model as an optimizer to train the reverse prediction model, inputting the initial process parameters and the historical path parameters into the forward prediction model to obtain a new predicted molten pool area, The loss value is calculated by a reverse loss function on the new predicted molten pool area and the molten pool area input into the reverse prediction model, The above steps are repeated to iteratively optimize the weight parameters of the forward prediction model and the reverse prediction model until the iteration converges or reaches a preset number of times, and an intermediate iteration process parameter is outputted; A result output module is configured to adjust the interrelated length by an optimization algorithm, repeatedly train the forward prediction model and the reverse prediction model according to the adjusted interrelated length, and output an optimal interrelated length and an optimal process parameter, including: The interrelated length is adjusted by an optimization algorithm, the forward prediction model and the reverse prediction model are repeatedly trained according to the adjusted interrelated length determined each time, and an optimal interrelated length and an optimal process parameter are outputted until the performance of the optimization model is optimized.