Additive manufacturing surface topography prediction method and system based on multi-source molten pool feature fusion

Abstract

The invention discloses an additive manufacturing surface topography prediction method and system based on multi-source molten pool feature fusion. The method comprises the steps that a visible light image and temperature field data of a molten pool in the laser additive manufacturing process are collected; key molten pool features are extracted according to the collected visible light images and temperature field data; fusing the extracted key molten pool feature data through a grid mapping method to generate a feature matrix; scanning the surface of the formed part to obtain surface topography data, and processing the surface topography data through a grid interpolation segmentation method to generate a label matrix with the same size as the feature matrix; and using a deep learning network based on an adaptive multi-source model, taking the feature matrix as an input, taking the label matrix as an output, carrying out training, and constructing a surface topography prediction model. By means of the scheme, the surface appearance in the laser additive manufacturing process is accurately predicted in real time, and an effective monitoring and predicting means is provided for quality control of additive manufacturing.

Description

Technical Field

[0001] The present invention relates to the field of additive manufacturing technology, and in particular to a method and system for predicting surface morphology of additive manufacturing by fusion of multi-source molten pool features. Background Art

[0002] Laser additive manufacturing (LAM) is a novel manufacturing method that constructs three-dimensional structures by stacking materials layer by layer. It has widespread applications in aerospace, biomedicine, mold manufacturing, and other fields. In LAM, surface topography is a key indicator of build quality, directly impacting product performance and service life.

[0003] Traditional surface topography measurement methods primarily perform offline testing after manufacturing is complete, failing to achieve real-time monitoring and prediction during the process. In recent years, researchers have begun attempting to predict forming quality by monitoring the morphological characteristics of the melt pool. However, information from a single sensor is often insufficient to accurately describe the complex forming process. Existing technologies present the following challenges:

[0004] 1. The data dimension of a single sensor is limited and cannot fully reflect the characteristics and changes of the molten pool;

[0005] 2. Lack of effective multi-source data fusion methods, making it difficult to integrate the advantages of different sensors;

[0006] 3. The existing prediction models have limited accuracy and cannot meet the needs of high-precision manufacturing;

[0007] 4. The lack of standardized processing procedures leads to insufficient reliability and stability of prediction results.

[0008] Therefore, developing a method and system that can fuse multi-source sensor data and accurately predict the surface morphology in the laser additive manufacturing process has important theoretical significance and practical value. Summary of the Invention

[0009] The present invention provides a method and system for predicting surface morphology in additive manufacturing by fusing multi-source melt pool features. By fusing the melt pool features collected by a coaxial CMOS camera and a rangefinder near-infrared camera, combined with deep learning technology, real-time and accurate prediction of the surface morphology in the laser additive manufacturing process is achieved.

[0010] To this end, the present invention provides the following technical solutions:

[0011] A method for predicting surface morphology in additive manufacturing by fusion of multi-source melt pool features, the method comprising:

[0012] Step 1: Collect visible light images and temperature field data of the molten pool during the laser additive manufacturing process;

[0013] Step 2: Extract key melt pool features based on the collected visible light image and temperature field data;

[0014] Step 3: The extracted key melt pool feature data are fused through the grid mapping method to generate a feature matrix;

[0015] Step 4: Scan the surface of the formed part to obtain surface topography data, and process the surface topography data using a grid interpolation segmentation method to generate a label matrix with the same size as the feature matrix;

[0016] In step 5, a deep learning network based on an adaptive multi-source model is used to train the feature matrix as input and the label matrix as output to construct a surface morphology prediction model.

[0017] Optionally, in step 1, a coaxial CMOS camera and a rangefinder near-infrared camera are used to synchronously capture visible light images and temperature field data of the molten pool during the laser additive manufacturing process.

[0018] Optionally, in step 2, the extracted key molten pool features include the area and perimeter of the molten pool in the visible light image, the area and perimeter of the high temperature zone in the temperature field data, and the vertical change of the molten pool centroid; wherein, through image segmentation and edge detection algorithms, the total number of pixels in the molten pool area in each frame image is calculated to obtain the area of ​​the molten pool in the visible light image; the pixel length of the molten pool edge is calculated to obtain the perimeter of the molten pool in the visible light image; the area where the temperature exceeds the threshold is selected, and the area of ​​the high temperature zone in the temperature field data is calculated; the length of the high temperature zone edge is calculated to obtain the perimeter of the high temperature zone in the temperature field data; and the vertical position change of the molten pool centroid is calculated to obtain the vertical change of the molten pool centroid.

[0019] Optionally, in step 3, the key melt pool feature data is divided into n×m grids according to the number of printing layers and length, and mapped to a p×q standard grid to generate a feature matrix; specifically, the following steps are performed:

[0020] 31) Establish a mapping relationship between each cell in the p×q standard grid and the corresponding four locations in the original n×m grid;

[0021] 32) Perform different mapping calculations on the sequence numbers in the original grid according to the different directions of the odd and even layers;

[0022] 33) Read the original feature data of the four corresponding positions and generate a new matrix according to the prescribed arrangement;

[0023] 34) Normalize the generated matrix to ensure consistent data distribution;

[0024] 35) Save the normalized matrix as a standardized data file.

[0025] Optionally, in step 4, when performing network interpolation, the steps specifically include:

[0026] 41) Read the original surface topography data and obtain the X, Y coordinates and Z values;

[0027] 42) The entire surface is divided into 50 layers vertically, and each layer is divided into 50 segments horizontally;

[0028] 43) Generate a 10×10 dot matrix using linear interpolation for each point in the grid area;

[0029] 44) Set a point threshold detection mechanism. When the number of points in a grid area is less than the preset threshold, the system identifies it as a possible formed collapse area and uses a special interpolation strategy to avoid inaccurate data matching;

[0030] 45) All interpolation data are normalized to generate standardized grid data files.

[0031] Optionally, in step 5, a mapping relationship between feature data and surface topography is established and prediction is performed based on a deep learning network of an adaptive multi-source model, specifically including:

[0032] 51) Use adaptive pooling layers to process input matrices from different sources to ensure size consistency;

[0033] 52) Use channel attention mechanism and spatial attention mechanism to enhance feature expression;

[0034] 53) Use the Transformer block with multi-head self-attention mechanism to process feature sequences;

[0035] 54) Fusion of features from different sources through cross-attention modules;

[0036] 55) All training data are normalized to have a mean of 0 and a standard deviation of 1 to ensure that all features are trained at the same scale;

[0037] 56) The cosine annealing learning rate scheduling strategy is used to optimize the training process.

[0038] Optionally, in step 56, the training process uses a weighted mean square error loss function, assigns different weights according to the magnitude and direction of the error, and increases the penalty for specific error intervals; when performing training control, an early stopping strategy is adopted, and the training process is stopped when the verification loss does not improve within a continuous preset period.

[0039] A multi-source melt pool feature fusion additive manufacturing surface morphology prediction system, the system comprising:

[0040] Image acquisition module, which collects visible light images and temperature field data of the molten pool during the laser additive manufacturing process;

[0041] Feature extraction module, which extracts key melt pool features based on the collected visible light images and temperature field data;

[0042] Grid mapping module, which uses grid mapping method to standardize and fuse the extracted key melt pool feature data rows;

[0043] Surface scanning module, which scans the surface of the formed part and obtains surface topography data;

[0044] The grid interpolation module segments and interpolates surface topography data and has an anti-collapse data processing mechanism;

[0045] The deep learning prediction module establishes a mapping relationship between feature data and surface morphology and performs predictions.

[0046] Optionally, the deep learning prediction module adopts an adaptive multi-source model architecture, including a feature extraction network, an attention mechanism, a Transformer block and a cross-attention fusion layer.

[0047] A computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the method for predicting surface morphology of additive manufacturing by fusion of multi-source melt pool features.

[0048] The present invention provides a method and system for predicting surface topography in additive manufacturing by fusing data from a multi-source melt pool feature. By fusing data from a coaxial CMOS camera and a rangefinder near-infrared camera, the system obtains more comprehensive melt pool feature information, thereby improving the accuracy of the prediction. The present invention also proposes a grid mapping method that effectively solves the spatial alignment and fusion problems of data from different sensors; designs a deep learning network based on an adaptive multi-source model that can process input data from different sources and sizes; and achieves high-precision prediction of surface topography in the laser additive manufacturing process. 2 The correlation coefficient reaches 0.95.

[0049] Compared with the prior art, the additive manufacturing surface morphology prediction system and method based on multi-source melt pool feature fusion provided by the present invention has the following advantages:

[0050] Real-time monitoring: Through real-time analysis of the melt pool characteristics, the changing trend of the surface morphology can be monitored during the manufacturing process, and abnormalities can be detected in time;

[0051] Quality control: Based on the prediction results, process parameters (such as laser power, scanning speed, etc.) can be adjusted to achieve closed-loop control and improve forming quality;

[0052] Cost reduction: reduces the workload of post-surface treatment, saving time and cost;

[0053] Wide applicability: The method is applicable to a variety of additive manufacturing processes and has good scalability;

[0054] Easy to implement: The equipment and techniques used are easy to implement in an industrial environment and do not require special experimental conditions.

[0055] In summary, the present invention provides a method and system for predicting surface morphology in additive manufacturing that is highly practical and can monitor and predict multi-source melt pool characteristics in real time, providing an effective means for quality control of additive manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0057] Figure 1 A schematic diagram of the equipment used in the embodiment of the present invention;

[0058] Figure 2 A flow chart of the surface morphology prediction method for additive manufacturing using multi-source melt pool feature fusion according to the present invention;

[0059] Figure 3 Schematic diagram of a grid mapping method according to an embodiment of the present invention;

[0060] Figure 4 A schematic diagram of a grid interpolation method according to an embodiment of the present invention;

[0061] Figure 5 Schematic diagram of a deep learning network structure based on an adaptive multi-source model in an embodiment of the present invention;

[0062] Figure 6 The figure is a graph showing the change in loss function and accuracy during the training process in an embodiment of the present invention;

[0063] Figure 7 A scatter plot comparing the predicted results and the true values ​​in an embodiment of the present invention;

[0064] Figure 8 This is an example diagram of a surface morphology prediction heat map in an embodiment of the present invention;

[0065] Figure 9 Schematic diagram of the structure of the additive manufacturing surface morphology prediction system with multi-source melt pool feature fusion of the present invention. DETAILED DESCRIPTION

[0066] The following describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0067] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0068] like Figure 2 As shown, the overall process of the additive manufacturing surface morphology prediction method based on multi-source melt pool feature fusion provided by the present invention includes five main steps: data acquisition, feature extraction, data fusion, label generation and deep learning prediction.

[0069] Step 1: Data collection.

[0070] A coaxial CMOS camera and a rangefinder near-infrared camera are used to synchronously capture visible light images and temperature field data of the melt pool during laser additive manufacturing. The coaxial CMOS camera, mounted along the laser optical path, directly observes the melt pool, capturing its shape and dimensions. The rangefinder near-infrared camera observes the melt pool from the side, capturing its temperature distribution. Both cameras simultaneously capture data at a rate of 50 frames per second, ensuring temporal consistency.

[0071] Step 2: Feature extraction.

[0072] Five key melt pool characteristics are extracted from the acquired raw data:

[0073] 1. The area of ​​the melt pool in the visible light image: The total number of pixels in the melt pool area in each frame of the image is calculated through image segmentation and edge detection algorithms;

[0074] 2. Perimeter of the melt pool in the visible light image: Calculate the pixel length of the edge of the melt pool;

[0075] 3. Area of ​​high temperature zone in temperature field data: select the area where the temperature exceeds the threshold (such as 1500°C) and calculate its area;

[0076] 4. Perimeter of the high temperature area in the temperature field data: calculate the length of the edge of the high temperature area;

[0077] 5. Vertical change of the center of mass of the molten pool: Calculate the position change of the center of mass of the molten pool in the vertical direction (Z axis).

[0078] Step 3: Data fusion.

[0079] like Figure 3 As shown in Figure 1, the extracted feature data is fused using a grid mapping method. This method can effectively solve the spatial matching problem of data collected by different sensors and ensure the accurate fusion of feature data. The specific steps are as follows:

[0080] 1. Create a mapping relationship between the original n×m grid and the new p×q grid to achieve data dimensionality reduction and spatial matching;

[0081] 2. Considering that the scanning path is "S"-shaped during the additive manufacturing process, and the scanning directions of odd and even layers are different, different mapping rules are used for odd layers (from left to right) and even layers (from right to left);

[0082] 3. For each new grid position, find the data of the four corresponding positions in the original grid. These data points are spatially adjacent and can jointly represent the characteristics of the area.

[0083] 4. Combine the data at the four locations into a new matrix according to a specific arrangement, preserving the spatial relationship of the original data while achieving data dimensionality reduction;

[0084] 5. Generate a standardized data matrix for each feature data to facilitate subsequent deep learning network processing.

[0085] The core idea of ​​this grid mapping algorithm is to achieve dimensionality reduction through two-dimensional matrix transformation while preserving the spatial topology of the original data. For each cell (i, j) in the p×q standard grid, the algorithm calculates its corresponding four locations in the original n×m grid: the lower left and lower right corners for odd-numbered layers, and the upper left and upper right corners for even-numbered layers. Given the alternating scanning directions in additive manufacturing, the algorithm inverts the column indices for even-numbered layers to ensure correct spatial correspondence.

[0086] Step 4: Label generation.

[0087] like Figure 4 As shown, an industrial 3D scanner is used to scan the surface of the formed part to obtain surface morphology data, which is then processed using a grid interpolation segmentation method with an anti-collapse mechanism. The specific steps are as follows:

[0088] 1. Read the original surface topography data and obtain the X, Y coordinates and Z value (surface height data);

[0089] 2. Divide the entire surface into p layers vertically, and each layer into q segments horizontally, forming a p×q grid structure that matches the feature data;

[0090] 3. Use linear interpolation to generate a 10×10 dot matrix for each point in the grid area to improve data resolution;

[0091] 4. Key Improvement: A point count threshold detection mechanism is now implemented. When the number of points in a grid area is less than a preset threshold (e.g., 4 points), the system identifies it as a possible collapsed area. A special processing strategy is then implemented: first, reference values ​​are obtained from surrounding non-collapsed areas; missing values ​​are filled using nearest neighbor interpolation; and a smooth transition algorithm is applied to the filled areas to avoid data gaps.

[0092] 5. Normalize the generated grid data to ensure the data range is consistent;

[0093] 6. Save the processed grid data as a standardized data file.

[0094] The core innovation of the grid interpolation segmentation method lies in its anti-collapse data processing mechanism. In the additive manufacturing process, forming collapse may occur due to various reasons (such as insufficient materials, uneven energy distribution, etc.), resulting in missing or abnormal surface data in some areas. Traditional grid segmentation methods often have problems with inaccurate matching when processing such data. The grid interpolation segmentation method of the present invention has an anti-collapse mechanism. When the number of points in the scanned grid area is less than the preset threshold, the system can identify it as a possible forming collapse area and use a special interpolation strategy to ensure the accuracy of data matching. The method of the present invention can effectively identify and process collapsed areas by setting point threshold detection and special interpolation strategies, ensuring the accuracy of data segmentation, and laying a solid foundation for subsequent deep learning predictions.

[0095] Step 5: Deep learning prediction.

[0096] like Figure 5 As shown in Figure 1, the deep learning network structure based on the adaptive multi-source model includes the following key components:

[0097] 1. Feature extraction network: A convolutional neural network is used to extract features from each input source. This network uses a multi-layer convolutional structure to gradually extract complex features from simple features. Specifically, it includes the following layers: Input layer: Receives melt pool feature matrices from different sources; Convolution layer group: Contains multiple convolution layers, each using a 3×3 convolution kernel to extract local features; Batch normalization layer: Added after each convolution layer to accelerate network convergence and prevent overfitting; Activation function: Uses the ReLU activation function to increase the nonlinear expression capability of the network; Pooling layer: Uses max pooling to reduce the size of the feature map and extract the main features;

[0098] 2、Adaptive Pooling Layer: This is an important innovation of the invention, which can handle input matrices of different sizes and ensure the consistency of the output. The core principle of adaptive pooling is: not to limit the fixed size of the pooling window, but to dynamically adjust it according to the input size; no matter how the input size changes, it can output a feature map of a specified size; An(X) = Y, where X is an arbitrary size input, Y is a fixed size output, and An is an adaptive pooling function;

[0099] 3、Attention Mechanism: Including channel attention and spatial attention, enhancing the expression of key features: Channel Attention: Calculate the importance weight of each channel of the feature map, highlighting important feature channels; Spatial Attention: Calculate the importance weight of the spatial position of the feature map, focusing on important spatial regions; The two attention mechanisms complement each other and improve the feature expression ability together;

[0100] 4、Transformer Block: Using multi-head self-attention mechanism to process serialized features: reshape the feature map into a sequence for easy processing by Transformer; Multi-head self-attention layer: using 16 attention heads to capture the correlation information in different subspaces; Feedforward Neural Network: nonlinear transformation of attention output; Layer Normalization and Residual Connection: ensure stable gradient propagation and effective information transmission;

[0101] 5、Cross-Attention Module: Realize the interaction and fusion between different source features: treat different source features as different sequences and establish correlation through cross-sequence attention; Q, K, V transformation: use one source feature as query (Q), another source feature as key (K) and value (V); Through the calculation of the similarity matrix, realize the dynamic weight fusion between different source features;

[0102] 6、Fusion Layer: Weighted fusion of different source features: use soft attention mechanism to automatically learn the weight coefficients of each source; The fusion process uses nonlinear transformation to enhance the expression ability of the features; Contains layer normalization and Dropout mechanism to enhance the generalization ability of the model;

[0103] 7、Output Layer: Generate surface topography prediction results: Fully Connected Layer: Map the fused features to the output space; Reshaping operation: reshape the one-dimensional output to a two-dimensional surface topography matrix; Output activation: use appropriate activation functions to ensure that the output is within a reasonable range;

[0104] The training process uses the following strategies:

[0105] 1、Data Processing:

[0106] Data Normalization: Normalize all input features and output labels to have a mean of 0 and a standard deviation of 1, ensuring that various features participate in training on the same scale.

[0107] Data split: The training set and test set are split in an 8:2 ratio.

[0108] Batch processing: Use batch training of 32 samples.

[0109] 2. Optimization strategy:

[0110] Weighted mean square error loss function: assigns different weights to different error intervals, especially strengthening the penalty for large error samples.

[0111] Cosine annealing learning rate scheduling strategy: The initial learning rate is set to 0.001, a hot restart is performed every 10 cycles, and the learning rate decays in a cosine manner within the cycle.

[0112] Weight decay: set to 0.0001 to control model complexity and prevent overfitting.

[0113] Gradient clipping: Limit the gradient norm to no more than 1.0 to prevent gradient explosion.

[0114] 3. Training control:

[0115] Early stopping strategy: Set the patience parameter to 15 and stop training when the validation loss does not improve within 15 epochs.

[0116] Model saving: Save the model parameters with the lowest validation loss.

[0117] Training Monitoring: Real-time monitoring of the training and validation loss functions and mean absolute errors.

[0118] Through this deep learning framework, the adaptive multi-source model's deep learning network automatically learns and optimizes the weighting coefficients of each source data when fusing feature data from different sources, achieving the optimal feature combination. This invention can effectively fuse melt pool feature data from different sources, establish a mapping relationship between feature data and surface topography, and achieve high-precision surface topography prediction.

[0119] Example 1

[0120] like Figure 1 As shown, a thin-walled part 13 is placed on a substrate 14 and is additively manufactured using a laser cladding head 11. The infrared camera 12 is a paraxial near-infrared camera, and the visible light camera 10 is a coaxial CMOS camera. Visible light images and temperature field data of the molten pool during the laser additive manufacturing process are synchronously captured by the visible light camera 10 and the infrared camera 12. The surface roughness of the thin-walled part 13 is then determined through calculation and deep learning prediction using the method of the present invention.

[0121] Example 2

[0122] like Figure 9 As shown, Figure 9Schematic diagram of the structure of the additive manufacturing surface morphology prediction system based on multi-source molten pool feature fusion of the present invention, the additive manufacturing surface morphology prediction system based on multi-source molten pool feature fusion includes:

[0123] Image acquisition module 901, which collects visible light images and temperature field data of the molten pool during the laser additive manufacturing process;

[0124] Feature extraction module 902 extracts key melt pool features based on the collected visible light image and temperature field data;

[0125] A grid mapping module 903 is used to standardize and fuse the extracted key melt pool feature data rows using a grid mapping method;

[0126] Surface scanning module 904 scans the surface of the formed part to obtain surface topography data;

[0127] The grid interpolation module 905 segments and interpolates the surface topography data and has an anti-collapse data processing mechanism;

[0128] The deep learning prediction module 906 establishes a mapping relationship between feature data and surface topography and performs prediction.

[0129] In some specific applications, the additive manufacturing surface morphology prediction system with multi-source melt pool feature fusion may also include a visualization module 907 for displaying the prediction results in the form of a heat map and generating a comparative analysis report between the prediction results and the actual values.

[0130] Example 3

[0131] During a single laser additive manufacturing process, 12,500 sets of melt pool data were collected, including visible light images from an on-axis CMOS camera and temperature field data from a rangefinder near-infrared camera. Five melt pool features were extracted from this data, and a standardized feature matrix was generated using a grid mapping method. Simultaneously, the surface of the formed part was scanned using an industrial 3D scanner, and surface topography labels corresponding to the feature matrix were generated using a grid interpolation method.

[0132] 80% of the data is used for training and 20% for testing. The model is trained for 100 epochs (cycles), and the early stopping mechanism is triggered when there is no improvement in 15 epochs. The loss function and accuracy change during training are as follows Figure 6 As shown in the figure, we can see that the training process converges steadily, and the loss value and mean absolute error of the final model on the validation set drop to about 0.12 and 0.26 respectively.

[0133] The performance evaluation results of the model on the test set are as follows:

[0134] Mean square error (MSE): 0.0027;

[0135] Root mean square error (RMSE): 0.0522;

[0136] R 2 :0.9024;

[0137] Mean absolute error (MAE): 0.0397;

[0138] Maximum error (Max_Error): 0.2483;

[0139] Correlation coefficient: 0.9516;

[0140] Normalized root mean square error (NRMSE): 0.0574.

[0141] Figure 7 The scatter plot of the predicted results and the true values ​​is shown, where R 2 The value reaches 0.9025, indicating that the model prediction value is highly consistent with the true value. The scatter density distribution shows a narrow distribution band, concentrated near the ideal prediction line (y=x), indicating that the prediction accuracy is high and stable.

[0142] Figure 8 A histogram of the error distribution is shown, showing that the prediction errors are normally distributed with a center close to zero, indicating that the model predictions have no significant systematic bias. Most prediction errors are within the range of ±0.1, further confirming the high accuracy of the model.

[0143] Example 4

[0144] In this embodiment, the method of the present invention can also be extended to other additive manufacturing processes, such as electron beam melting, arc additive manufacturing, etc. In an electron beam melting experiment, a similar sensor configuration and data processing process were used to predict the surface morphology of the formed part. The experimental results show that the method of the present invention is also applicable to the electron beam melting process, and the predicted R 2 It reaches 0.85 and the correlation coefficient reaches 0.92.

[0145] The proposed system and method for predicting surface topography in additive manufacturing (AM) using multi-source melt pool feature fusion, combined with deep learning techniques, achieves high-precision prediction of surface topography during laser AM. This method boasts real-time performance, accuracy, and wide applicability, providing an effective monitoring and prediction tool for AM quality control.

[0146] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0147] The present application also provides a storage medium, which is a computer readable storage medium, and a computer program is stored on the storage medium, and the computer program can execute Figure 2 part or all of the steps of the method shown. The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc. The storage medium can also include a non-volatile memory or a non-transitory memory, etc.

[0148] The above embodiments can be realized all or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized all or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable system. The computer instructions can be stored in a computer readable storage medium or transferred from one computer readable storage medium to another, for example, the computer instructions can be transferred from one website, computer, server or data provider to another website, computer, server or data provider through wired or wireless means.

[0149] The above describes the embodiments of the present application in detail, and the present application is described by applying specific implementation manners. The above embodiment description is only used to help understand the method and system of the present application, and it is only an embodiment of a part of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application, and the content of the specification should not be understood as a limitation of the present application. Therefore, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting surface morphology in additive manufacturing by fusion of multi-source melt pool features, characterized in that: The method comprises: Step 1: Collect visible light images and temperature field data of the molten pool during the laser additive manufacturing process; Step 2: Extract key melt pool features based on the collected visible light image and temperature field data; Step 3: The extracted key melt pool feature data are fused through the grid mapping method to generate a feature matrix; Step 4: Scan the surface of the formed part to obtain surface topography data, and process the surface topography data using a grid interpolation segmentation method to generate a label matrix with the same size as the feature matrix; In step 5, a deep learning network based on an adaptive multi-source model is used to train the feature matrix as input and the label matrix as output to construct a surface morphology prediction model.

2. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 1, characterized in that: In step 1, a coaxial CMOS camera and a rangefinder near-infrared camera are used to synchronously capture visible light images and temperature field data of the molten pool during the laser additive manufacturing process.

3. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 1, characterized in that: In step 2, the key melt pool features extracted include the area and perimeter of the melt pool in the visible light image, the area and perimeter of the high temperature zone in the temperature field data, and the vertical change of the melt pool centroid; among them, through image segmentation and edge detection algorithms, the total number of pixels in the melt pool area in each frame image is calculated to obtain the area of ​​the melt pool in the visible light image; the pixel length of the melt pool edge is calculated to obtain the perimeter of the melt pool in the visible light image; the area where the temperature exceeds the threshold is selected, and the area of ​​the high temperature zone in the temperature field data is calculated; the length of the high temperature zone edge is calculated to obtain the perimeter of the high temperature zone in the temperature field data; the vertical position change of the melt pool centroid is calculated to obtain the vertical change of the melt pool centroid.

4. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 1, characterized in that: In step 3, the key melt pool feature data is divided into n×m grids according to the number of printing layers and length, and mapped to a p×q standard grid to generate a feature matrix; Specifically include: 31) Establish a mapping relationship between each cell in the p×q standard grid and the corresponding four locations in the original n×m grid; 32) Perform different mapping calculations on the sequence numbers in the original grid according to the different directions of the odd and even layers; 33) Read the original feature data of the four corresponding positions and generate a new matrix according to the prescribed arrangement; 34) Normalize the generated matrix to ensure consistent data distribution; 35) Save the normalized matrix as a standardized data file.

5. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 1, characterized in that: In step 4, when performing network interpolation, it specifically includes: 41) Read the original surface topography data and obtain the X, Y coordinates and Z values; 42) The entire surface is divided into 50 layers vertically, and each layer is divided into 50 segments horizontally; 43) Generate a 10×10 dot matrix using linear interpolation for each point in the grid area; 44) Set a point threshold detection mechanism. When the number of points in a grid area is less than the preset threshold, the system identifies it as a possible formed collapse area and uses a special interpolation strategy to avoid inaccurate data matching; 45) All interpolation data are normalized to generate standardized grid data files.

6. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 1, characterized in that: In step 5, based on the deep learning network of the adaptive multi-source model, a mapping relationship between feature data and surface topography is established and prediction is performed, specifically including: 51) Use adaptive pooling layers to process input matrices from different sources to ensure size consistency; 52) Use channel attention mechanism and spatial attention mechanism to enhance feature expression; 53) Use the Transformer block with multi-head self-attention mechanism to process feature sequences; 54) Fusion of features from different sources through cross-attention modules; 55) All training data are normalized to have a mean of 0 and a standard deviation of 1 to ensure that all features are trained at the same scale; 56) The cosine annealing learning rate scheduling strategy is used to optimize the training process.

7. The surface morphology prediction method for additive manufacturing based on multi-source melt pool feature fusion according to claim 6, characterized in that: In step 56, the training process uses a weighted mean square error loss function, assigning different weights according to the magnitude and direction of the error to increase the penalty for specific error intervals; when performing training control, an early stopping strategy is adopted, and the training process is stopped when the verification loss does not improve within a continuous preset period.

8. A surface morphology prediction system for additive manufacturing based on multi-source melt pool feature fusion, characterized in that: The system comprises: Image acquisition module, which collects visible light images and temperature field data of the molten pool during the laser additive manufacturing process; Feature extraction module, which extracts key melt pool features based on the collected visible light images and temperature field data; Grid mapping module, which uses grid mapping method to standardize and fuse the extracted key melt pool feature data rows; Surface scanning module, which scans the surface of the formed part and obtains surface topography data; The grid interpolation module segments and interpolates surface topography data and has an anti-collapse data processing mechanism; The deep learning prediction module establishes a mapping relationship between feature data and surface morphology and performs predictions.

9. The additive manufacturing surface morphology prediction system based on multi-source melt pool feature fusion according to claim 8, characterized in that: The deep learning prediction module adopts an adaptive multi-source model architecture, including a feature extraction network, an attention mechanism, a Transformer block and a cross-attention fusion layer.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting surface morphology of additive manufacturing by fusion of multi-source melt pool features as claimed in any one of claims 1 to 7 are executed.

Images