A molten pool profile temporal migration prediction method, system, medium and device

Abstract

The present application relates to the technical field of intelligent manufacturing, and provides a molten pool profile time sequence migration prediction method, system, medium and equipment, including: acquiring a titanium alloy source domain data set, training a gated residual cross-modal bridging network and saving the bridging weight and source domain standardization statistics; acquiring target domain time sequence process data, completing data set division with complete molten channel as a unit, and respectively generating standardized bridging input and original dimension main body input; loading a source domain pre-training molten pool time sequence prediction main body model, generating pseudo visual embedding by the bridging network in a no-image target domain migration scene and fusing the main body process parameter channel for prediction, injecting a low-rank adaptive layer in layers in a cross-material migration scene and superimposing Malan Goni curvature physical constraints; selecting an optimal model through a verification set, and predicting a molten pool profile on a test set. Only input of process parameters can generate pseudo visual features consistent with image embedding distribution, and improve time sequence migration prediction reliability in cross-condition or cross-material scenes.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, and in particular relates to a method, system, medium and equipment for predicting the temporal migration of molten pool contours. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Laser-directed energy deposition (L-DED) relies on laser melting of metal powder to form parts layer by layer. The molten pool profile, weld width (W), and weld height (H) directly determine the forming accuracy, internal defects, and mechanical properties of the parts, making them core monitoring indicators for process control. Currently, laser additive manufacturing forming prediction systems generally suffer from industry shortcomings such as strong material specificity, weak model generalization, and a lack of cross-material transfer prediction capabilities. Existing molten pool prediction models are all single-material-specific fitting models, only adaptable to the single metal material used for training, and cannot achieve the transfer and reuse of forming features between dissimilar materials. Furthermore, most existing molten pool prediction models rely on multimodal visual images such as infrared and X-rays as model input, using image visual embedding to complete molten pool feature fitting. When there is smoke, strong arc light, splash obstruction, or when high-speed cameras cannot be deployed in industrial settings, visual modal data is completely missing, and the model cannot reason normally, further exacerbating the application limitations under multi-material conditions.

[0004] Existing technology has three major flaws: Cross-modal mapping suffers from insufficient nonlinear adaptation and severe distribution mismatch: When the target domain has no image and only process parameters, traditional linear mapping cannot fit the complex nonlinear manifold relationship between process parameters and pre-trained high-dimensional visual embeddings. Simple linear projection causes feature distribution discontinuities and significantly reduces prediction accuracy. Furthermore, standardization operations often leak statistical information from the test set and future time series data, leading to model overfitting and decreased generalization ability. Cross-material transfer fine-tuning is prone to catastrophic forgetting and has poor robustness with small samples: when transferring from the TC4 titanium alloy source domain pre-trained model to the 316L stainless steel target domain, full fine-tuning will destroy the general features of the molten pool learned in the source domain; the unified rank LoRA cannot match the encoder, decoder, and cross-attention layer differentiated adaptation requirements, and the low-level general features and high-level material-specific features cannot be optimized in layers, and it is very easy to overfit after training with a small number of 316L samples. Purely data-driven models violate the laws of metal thermodynamics and predict distorted profiles: Deep learning without physical constraints uses only MSE and MAPA as losses, and the trained molten pool profiles have severe concavity and abrupt inflection points, which do not conform to the smooth morphological characteristics of liquid metal driven by Marangoni convection; the second-order curvature of the profile curve is not penalized, and the model outputs a large number of non-physical invalid samples, resulting in low reliability of process control decisions.

[0005] In summary, the existing solutions fail to address the three major challenges of graph-free cross-modal mapping, hierarchical migration fine-tuning, and physical prior constraints simultaneously, and cannot meet the requirements for high-precision time-series prediction of molten pool contours under high-power, vision-free, and cross-material L-DED conditions. Summary of the Invention

[0006] To address the technical problems mentioned above, this invention provides a method, system, medium, and device for predicting the temporal migration of molten pool contours. It nonlinearly fits process parameters with a visual embedded manifold distribution, generating pseudo-visual features consistent with the image embedded distribution simply by inputting process parameters. It can still stably output the molten pool contour even under conditions of no camera or strong smoke and dust obstruction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for predicting the temporal migration of a melt pool profile, comprising: Obtain the titanium alloy source domain dataset, use process parameters as input and visual embedding as fitting labels, train the gated residual cross-modal bridging network, and save the standardized statistics of the trained bridging network weights and source domain process parameters. Acquire time-series process data in the target domain, divide the dataset into training, validation, and test sets using complete melt channels as units; reuse the standardized statistics or use the training set statistics to standardize the inputs of the bridging branches to obtain standardized parameters; retain the original physical dimensions of the inputs of the main branches, and match the parameter input dimensions of the main model by zero padding. Load the molten pool temporal prediction backbone model pre-trained with a titanium alloy source domain dataset; in the target domain transfer scenario without image, input the standardized parameters into the bridging network to generate pseudo-visual embeddings, and superimpose learnable temporal position encodings as the visual channel input of the backbone model, and use the target domain process parameters that retain the original physical dimensions as the process parameter channel input of the backbone model; in the cross-material transfer scenario, inject low-rank adaptation layers into the backbone model layer by layer, and superimpose the Marangoni curvature regularization loss on the basic combined prediction loss; The bridging network and backbone model are validated using the validation set to obtain the optimal migration prediction model. The optimal migration prediction model is then used for blind inference using the test set to output the predicted results of the melt pool profile, melt width, or melt height.

[0008] Furthermore, the Marangoni curvature regularization loss is: ; Among them, the curvature weights in the e-th training round , This represents the maximum value of the curvature weight. The curvature weight increases linearly from 0 to The number of preheating rounds, B represents the batch sample size, and n represents the number of contour sampling points or label dimension; the one-dimensional second-order difference calculated from three adjacent predicted contour points in the b-th sample. ; This represents a numeric truncation function, which restricts the input value to between a preset lower and upper bound. , and Both represent the predicted contours after numerical truncation.

[0009] Furthermore, the bridging network is a gated residual cross-modal bridging network: it completes the process parameter dimension mapping through a linear projection layer, extracts high-order interactive features of process parameters by combining residual blocks, dynamically adjusts the residual feature weights by relying on adaptive gated branches, and completes feature distribution calibration through a normalization layer.

[0010] Furthermore, the process of dividing the dataset into complete melt channels as units includes: dividing the target domain time-series process data into groups based on complete melt channels as the smallest independent unit. Each melt channel contains a fixed length of continuous time-series process samples. Only the overall sequence number of the melt channels is shuffled to complete the division of the training set, validation set, and test set, without splitting the time-series data within a single melt channel.

[0011] Furthermore, the hierarchical injection of the low-rank adaptation layer includes: freezing the basic weights of the melt pool temporal prediction backbone model, configuring a low-rank LoRA adaptation layer for the mid-level encoder network, configuring a high-rank LoRA adaptation layer for the high-level cross-modal attention layer and decoder network, and only the low-rank adaptive parameters and the reserved adaptation parameters participate in the update.

[0012] Furthermore, during parameter optimization, a base learning rate is set for the trainable layers of the backbone model, and an independent learning rate is set for the bridging network, temporal position encoding, or low-rank LoRA adaptation layer. No weight decay is applied to the temporal position encoding.

[0013] Furthermore, for the task of predicting the molten pool profile, the basic combined prediction loss is used and the average crossover ratio is used as the verification index; for the task of predicting the molten width or molten height, the weighted mean square error loss is used and the average absolute percentage accuracy is used as the verification index.

[0014] A second aspect of the present invention provides a melt pool profile timing migration prediction system, comprising: The source domain bridging pre-training module is configured to: acquire the titanium alloy source domain dataset, use process parameters as input, use visual embeddings as fitting labels, train the gated residual cross-modal bridging network, and save the standardized statistics of the trained bridging network weights and source domain process parameters. The time-series data dual-path construction module is configured to: acquire time-series process data in the target domain, divide the dataset into training, validation and test sets by using complete melt channels as units; reuse the standardized statistics or use the training set statistics to standardize the input of the bridging branch to obtain standardized parameters; retain the original physical dimensions of the input of the main branch and match the parameter input dimension of the main model by zero padding. The fine-tuning module is configured to: load a molten pool temporal prediction backbone model pre-trained using a titanium alloy source domain dataset; in image-free target domain transfer scenarios, input standardized parameters into a bridging network to generate pseudo-visual embeddings, superimpose learnable temporal positional encodings as visual channel inputs to the backbone model, and use target domain process parameters that retain their original physical dimensions as process parameter channel inputs to the backbone model; in cross-material transfer scenarios, inject low-rank adaptation layers into the backbone model layer by layer, and superimpose Marangoni curvature regularization loss on the basic combined prediction loss; The evaluation and output module is configured to validate the bridging network and backbone model based on the validation set, obtain the optimal migration prediction model, and perform blind inference on the optimal migration prediction model using the test set to output the predicted results of the melt pool profile, melt width, or melt height.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the melt pool profile timing migration prediction method described above.

[0016] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps of the melt pool profile timing migration prediction method described above.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention uses nonlinear fitting of process parameters and visual embedding manifold distribution to generate pseudo-visual features consistent with image embedding distribution by simply inputting process parameters. It can still stably output the molten pool contour even under conditions of no camera or strong smoke and dust obstruction.

[0018] This invention uses complete melt channels as the dividing unit, without breaking up the time series samples of a single melt channel. Standardization relies solely on the statistics of the training set, and the dual-branch separation takes into account both cross-domain distribution alignment and the input requirements of the physical dimensions of the original model, thus solving the problem of test set information leakage caused by traditional single-path standardization.

[0019] This invention features a layered and differentiated LoRA rank configuration. The bottom layer's general melt pool features are completely frozen, while the middle and upper layers are finely adjusted in terms of rank. Only a small number of adaptation parameters participate in the update, accounting for less than 5% of the total parameters.

[0020] This invention constructs a second-order curvature penalty loss based on the Marangoni convection mechanism of liquid metal, constrains the prediction profile to be smooth and continuous, and eliminates abrupt changes and non-physical curves in the output of unconstrained models. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of a melt pool profile timing migration prediction method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of an application scenario of Embodiment 1 of the present invention; Figure 3 This is a comparison diagram of the first cross-sectional profile of Embodiment 1 of the present invention; Figure 4 This is a comparison diagram of the first melt width and melt height in Embodiment 1 of the present invention; Figure 5 This is a comparison diagram of the second cross-sectional profile of Embodiment 1 of the present invention; Figure 6 This is a comparison diagram of the second melt width and melt height in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 To avoid ambiguity caused by English abbreviations and code module identifiers, the terms involved in this invention have the following meanings: TC4 is the grade of Ti-6Al-4V titanium alloy, 316L is the grade of austenitic stainless steel, L-DED is laser-directed energy deposition, CSV is a comma-separated data table, GatedResidualBridge is a gated residual cross-modal bridging network, SiLU is a sigmoid linear unit activation function, Sigmoid is a sigmoid function, LayerNorm is layer normalization, MSE is mean squared error, AdamW is an adaptive moment estimator optimizer with decoupled weight decay, LoRA is low-rank adaptive, and Hier-LoRA is a layered low-rank adaptive optimizer. Rank adaptive, mIoU is the mean intersection-union ratio, MAE is the mean absolute error, MAPA is the mean absolute percentage accuracy calculated as 1-MAPE (prediction accuracy), JSON is the structured diagnostic data file, W / H represent the melt width / melt height respectively; the code identifiers input_embedding, encoder, cross_modal_attn, and decoder correspond to the input embedding layer, encoder, cross-modal attention layer, and decoder respectively, and img_embed_dim, param_size, and seq_len represent the visual embedding dimension, the backbone model parameter input dimension, and the temporal length respectively.

[0026] This embodiment provides a method for predicting the temporal migration of molten pool profiles.

[0027] This embodiment provides a molten pool contour temporal migration prediction method for scenarios without visual images, which integrates gated residual cross-modal bridging, hierarchical low-rank adaptation, and Marangoni thermodynamic physical constraints.

[0028] This embodiment provides a method for predicting the temporal migration of molten pool contours. It can generate equivalent visual embeddings by replacing visual images with process parameters alone. It achieves cross-material migration of small samples through hierarchical low-rank adaptive methods and introduces Marangoni curvature loss constraints to ensure the physical rationality of the predicted contours. This improves the prediction accuracy and engineering practicality of molten pool temporal contours, melt width and melt height under visual conditions.

[0029] like Figure 1 and Figure 2 The melt pool contour temporal migration prediction method in this embodiment is applicable to two scenarios.

[0030] Scenario 1: In high-power laser additive manufacturing processes, where the visual camera is obstructed by strong arc light, metal spatter, and smoke, resulting in complete loss of image modalities and the inability of traditional models to predict these scenarios, a gated residual bridging network pre-trained on a source domain image dataset is used to achieve temporal migration prediction of molten pool contour, melt width, and melt height driven purely by process parameters. This embodiment provides a molten pool contour temporal migration prediction method, including the following steps: Step 101: Pre-training and solidifying parameter statistics of the source domain gated residual bridging network.

[0031] (1) Select the TC4 titanium alloy additive manufacturing process data with low power and normal visual image acquisition as the source domain dataset, which covers multiple sets of laser power, scanning speed, powder feeding rate process parameters and corresponding molten pool visual embedding features.

[0032] (2) Construct a gated residual cross-modal bridging network (GatedResidualBridge), complete the process parameter dimension mapping through a linear projection layer, extract high-order interactive features of process parameters by combining the S-type linear unit activation function (SiLU) residual block, dynamically adjust the residual feature weights by relying on the S-type function (Sigmoid) adaptive gated branch, and complete the feature distribution calibration through the layer normalization layer.

[0033] Specifically, the gated residual bridging network includes a linear projection layer, an S-shaped linear unit residual block, an S-shaped gated branch, and a layer normalization layer. The forward computation flow is as follows: The standardized process parameters are initially mapped by a linear projection layer to obtain the basic embedding. That is, the initial feature representation of process parameters in the visual embedding space; High-order interactive residual features are extracted from two layers of SiLU residual blocks. ; Same-dimensional Sigmoid gated layer outputs adaptive weights Dynamically adjust the contribution ratio of residual characteristics; The basic embedding and the weighted residual are summed element-wise and then distributed stably by LayerNorm to output the equivalent visual embedding.

[0034] (3) Using the standardized source domain process parameters as input and the source domain visual embedding features as fitting labels, the mean squared error (MSE) loss function is used to complete the iterative training of the bridging network, fully learning the nonlinear mapping relationship between the process parameters and the high-dimensional visual features of the molten pool. After training, the weights of the bridging network are fixed, and the mean and standard deviation of the source domain process parameters are saved. The above bridging weights are used to initialize the target domain bridging network in step 103, and the above mean and standard deviation are used to standardize the target domain bridging input in step 102, thus serving as the distribution alignment benchmark for cross-condition migration.

[0035] The steps for pre-training the gated residual cross-modal bridging network are as follows: Data loading: Read the source domain comma-separated data table source_data.csv, where source_data.csv represents the source domain sample file; the columns in this file are divided into process parameter columns and visual embedding columns output by the pre-trained backbone model, and are converted into numerical arrays respectively; Standardized calculation: Calculate the mean value based only on all process parameters in the source domain. (i.e., param_mean), standard deviation (i.e., param_std), to complete parameter standardization and scaling; Network initialization: Instantiate the gated residual cross-modal bridging network, with the input dimension equal to the number of process parameters and the output dimension matching the visual embedding dimension img_embed_dim of the pre-trained backbone model; where img_embed_dim is the length of the visual embedding feature in the model configuration; Training configuration: An adaptive moment estimation optimizer (AdamW) with decoupled weight decay and a cosine annealing learning rate scheduler are used. The loss function is the mean squared error loss, and the iteration is 200 rounds. In each round, the fitting loss between the bridging output and the source domain visual embedding is calculated, and the optimal checkpoint is saved when the loss decreases. The checkpoint synchronously stores the bridging network state parameters, the mean of the source domain process parameters (param_mean), and the standard deviation of the source domain process parameters (param_std) for loading and reuse during target domain transfer.

[0036] The computation process of the gated residual cross-modal bridging network is represented as follows, following the order of standardization, linear projection, residual branching, gated branching, and layer normalization in the code: (1); (2); (3); (4); (5); (6); In formulas (1) to (6), x p,i This represents the original process parameter vector for the i-th sample, which includes configuration columns such as laser power POWER, scanning speed SCAN_SPEED, and powder feed rate POWDER_FEED_RATE. This represents the standardized bridging network input; The mean value of the source domain process parameters can be represented as param_mean; The standard deviation of the source domain process parameters can be represented by the standard deviation. Add to To avoid division by zero, it can be represented as param_std; e0 represents the basic embedding obtained by linear projection; r represents the higher-order residual features obtained by two linear layers and the SiLU activation function; g represents the Sigmoid gate weight; the symbol ⊙ represents element-wise multiplication; z represents the pseudo-visual embedding output by the bridging network; v represents the pseudo-visual embedding of the i-th sample; i The source domain visual embedding fits the labels; N represents the number of source domain training samples; W p W1, W2, W g and b p b1, b2, b g All of these are learnable parameters for bridging networks.

[0037] Step 102: Zero-leakage partitioning and dual-input construction of image-free high-power target domain time-series dataset.

[0038] (1) Collect additive manufacturing process data under high power without visual image conditions. Take the complete melt channel as the smallest independent unit and divide the time sequence process data into groups. Each melt channel contains three consecutive time sequence process samples. Only the overall sequence number of the melt channel is shuffled to complete the division of training set, validation set and test set. The time sequence data inside the single melt channel is not split, so as to completely eliminate the problem of leakage of future time sequence information and test set statistical information.

[0039] For example, read the CSV process data table of the target domain of 316L stainless steel, and divide it into training, validation and test sets with the single melt channel 3 time series sample as the smallest unit. Only use the training set to statistically standardize the parameters.

[0040] Specifically, the dataset partitioning rules are as follows: the original data is grouped by melt channel, each melt channel contains 3 consecutive rows of time series data, the melt channel IDs are randomly shuffled instead of individual samples, the remaining melt channels are evenly distributed among the training set, validation set and test set, and the time series leakage of the same melt channel is completely eliminated.

[0041] The detailed steps are as follows: Raw data reading: Load the 316L high-power process CSV file. The parameter columns include laser power, scanning speed, and powder feeding rate. The label columns are the 20-point molten pool contour coordinates or molten width W and molten height H. Fuse grouping: Set the single fuse time length seq_len=3, the total number of valid fuses = total number of rows / 3, shuffle the fuse index to divide the training, validation and test fuse sets, and do not split the time samples within a single fuse.

[0042] (2) Construct two independent input branches: The mean and standard deviation of the source domain solidified in step 1 of bridging branch reuse are used to standardize the process parameters of high-power operating conditions, ensuring that the target domain input feature distribution is completely aligned with the source domain pre-training scenario. The main branch directly retains the original physical dimensions of the process parameters, and aligns the parameter input dimension param_size of the source domain pre-trained molten pool timing prediction main model loaded in step 103 by zero padding, thus preserving the true physical characteristics of the process parameters and adapting to the physical representation logic of the main model.

[0043] Specifically, the dual-input separation rule is as follows: the bridging input is standardized using the mean and standard deviation of the source domain process parameters param_mean and param_std to ensure cross-domain distribution alignment; the backbone input is not standardized, and the part that is insufficient for the input dimension param_size of the pre-trained backbone model parameters is padded with zeros at the end to match the input logic of the physical dimensions of the original model.

[0044] The detailed steps are as follows: Standardized isolation: Only the parameters within the training melt channel are extracted to calculate the mean and standard deviation; when there are no pre-trained bridging weights, the local training set statistics are used; when there are source domain bridging weights, the param_mean and param_std stored in the source domain are forcibly reused. Dual tensor generation: The bridging branch standardizes all original parameters and reshapes them into a tensor of [number of melt channels, 3, parameter dimension]; the main branch retains the original parameter values, and pads the end with zeros to align the input dimension param_size of the pre-trained main model parameters when the dimension is insufficient.

[0045] Dataset deployment: Label the melt flow and timing step of each sample, completely isolate the data of each set, and prevent leakage of timing and test set information.

[0046] Step 103: Loading the source domain pre-trained backbone model, freezing fine-tuning, and dual-path temporal fusion prediction.

[0047] (1) Load the source domain pre-trained melt pool temporal prediction backbone model, which is different from the bridging network in step 101. This backbone model is generated by the original model building interface and loaded with the optimal weights in the source domain. It includes an input embedding layer, an encoder, a cross-modal attention layer, and a decoder. For high-power scenarios without image adaptation, the input embedding layer and the first two layers of the encoder are frozen, while retaining their general melt pool forming characteristics. The remaining high-level backbone parameters, the gated residual bridging network, and the temporal position encoding are all fine-tuned.

[0048] Specifically, the input embedding layer (input_embedding) and the first two encoder layers (encoder.layers.0 and encoder.layers.1) are not updated; the high-level parameters of the backbone, the parameters of the bridging network, and the temporal position encoding are all trained parameters and participate in the optimization uniformly. The backbone model uses a base learning rate, while the bridging network and the temporal position encoding use independent learning rates. The temporal position encoding is not subject to weight decay to avoid the source domain's general melt pool features being corrupted by small sample data from the target domain.

[0049] (2) The process parameters after dual-path processing are input into the bridging network and the backbone network respectively: the standardized process parameters are input into the bridging network, outputting equivalent pseudo-visual embedding features, and added element-wise with the learnable temporal position code to form the visual channel input; the process parameters after retaining the original physical dimensions and zero-padding alignment are used as the process parameter channel input of the backbone network. The above visual channel input and process parameter channel input are simultaneously fed into the source domain pre-trained backbone model after being aligned one by one according to the same melt flow and the same temporal step. The two types of features are fused by the cross-modal attention layer and decoder inside the backbone model, and the two-dimensional geometric prediction results of the 20-point contour of the melt pool or the melt width W and melt height H of the last temporal step are output.

[0050] The detailed steps are as follows: Pre-trained backbone model loading: Read the model configuration and optimal weights best_model_wts.pth from the TC4 source domain pre-training directory to construct the melt pool time series prediction backbone model; Bridging weight loading: If a source domain bridging checkpoint exists, load the bridging network state parameters, source domain mean (param_mean), and source domain standard deviation (param_std); if no source domain bridging checkpoint exists, use only training set statistics to standardize the bridging input. Temporal position encoding initialization: Create a learnable tensor [1, seq_len, embed_dim], where seq_len is the temporal length and embed_dim is the embedding dimension. It is initialized using a truncated normal distribution and is only added and fused with the pseudo-visual embedding output by the bridging branch. Forward fusion logic: The pseudo-visual embedding of the bridging branch output, after being encoded with temporal position, is used as the visual channel input of the backbone model, and the original dimensional process parameter tensor is used as the process parameter channel input of the backbone model. Both are input into the cross-modal attention layer and decoder of the backbone model, and the final temporal step melt pool prediction result is output.

[0051] For the 316L cross-material migration scenario, the hierarchical low-rank adaptive structure is enabled in step 203. Its function is different from that of the gated residual bridging network: the gated residual bridging network is responsible for mapping process parameters to pseudo-visual embeddings, while the low-rank adaptive layer is responsible for completing cross-material adaptation with a small number of newly added parameters within the pre-trained backbone model in the source domain.

[0052] Step 104: Loss function configuration, training stability constraints, and early stopping data preservation.

[0053] (1) Configure the loss function according to the prediction task. For the 20-point melt pool profile prediction task, the combined prediction loss built into the source domain backbone model is adopted and mIoU is used as the verification index. For the two-dimensional prediction task of melt width W and melt height H, the weighted mean square error loss is adopted and MAPA is used as the verification index. For the 316L cross-material migration task, a one-dimensional Marangoni curvature regularization term is superimposed on the basic combined prediction loss, the smoothness of the profile is predicted by second-order difference constraint, and a 10-round preheating mechanism is set to gradually increase the physical constraint penalty weight.

[0054] Specifically, the basic combined prediction loss, Marangoni curvature regularization term, and evaluation index for each prediction task are calculated according to the following formulas: The basic portfolio prediction loss is: (7); (8); (9); Weighted mean square error loss for: (10); The mean absolute percentage accuracy is: (11); Combined loss of basic prediction loss and Marangoni second-order curvature penalty loss for: (12); (13); (14); (15); (16); In formulas (7) to (16), L base λ represents the basic combined loss used for 20-point profile prediction. MAE , λ MSE , λ IoU L represents the weighting coefficients of the mean absolute error term, mean square error term, and cross-union ratio constraint term in the basic combination loss, respectively; MAE L MSE and mIoU represent the mean absolute error loss, mean squared error loss, and mean intersection-union ratio, respectively; B represents the batch size, n represents the number of contour sampling points or label dimension, and y b,i This represents the true label of the i-th contour point in the b-th sample. L represents the prediction result of the i-th contour point of the b-th sample; WHω represents the weighted mean square error loss for the two-dimensional task involving weld width W and weld height H. W and ω H W represents the weighting coefficients for the melt width error term and the melt height error term, respectively. b and H b Let the actual melt width and actual melt height of the b-th sample be represented respectively. and Represent the predicted melt width and predicted melt height of the b-th sample, respectively; ε represents the zero-prevention threshold in MAPE calculation, used to avoid abnormal proportional error caused by the true label being close to zero, y b,i and These represent the true labels and the predicted results, respectively. MAPA is calculated as 1-MAPE; p b,i This represents the predicted contour points after numerical truncation, with a truncation range of [-1000, 1000]. `clip` represents the numerical truncation function, which restricts the input values ​​to a preset lower and upper bound. b,j Let λ(e) represent the second-order difference calculated from a one-dimensional convolution kernel [1,-2,1] or three adjacent predicted contour points and then truncated, used to characterize the local curvature of the predicted contour. The truncation range of the second-order difference is [-100, 100]; λ(e) represents the curvature weights in the e-th training round. max T represents the maximum curvature weight. warm This indicates that the curvature weight increases linearly from 0 to λ. max The number of preheating wheels is preferably 10; L c(e) denoted by Marangoni curvature regularization loss; L represents the total loss resulting from the superposition of the basic portfolio prediction loss and the Marangoni curvature regularization loss.

[0055] (2) Set the optimizer learning rate for each group, add gradient pruning and numerical truncation to prevent non-finite values, trigger early stopping based on validation set metrics and save the optimal weights.

[0056] Specifically, the optimizer grouping configuration is as follows: the pre-trained backbone trainable layers are set with a base learning rate; the gated bridging network, temporal positional encoding, or low-rank adaptive parameters are set with independent learning rates; no weight decay is applied to the temporal positional encoding. Each training cycle completes validation set evaluation; if there is no improvement in mIoU or MAPA for several consecutive cycles, early stopping is triggered to terminate training.

[0057] The detailed steps are as follows: Loss function construction: The 20-point contour task uses the basic combination prediction loss shown in formulas (7) to (9); the melt width W and melt height H tasks use the weighted mean square error loss shown in formula (10); the 316L cross-material task uses the combination loss of the basic prediction loss shown in formula (16) and the Marangoni second-order curvature penalty loss. The number of preheating rounds is 10 rounds, and the penalty weight in the first 10 rounds increases linearly from 0 to the preset maximum value. Optimizer grouping: The pre-trained backbone trainable layers use the base learning rate; the bridging network, temporal positional encoding, and LoRA low-rank adaptive parameters each use independent learning rates specified in the configuration file; weight decay is disabled in temporal positional encoding; Training stability constraints: Forward inference and autoregressive caching both truncate the predicted tensor values; backpropagation performs gradient pruning, and the maximum norm in the 316L stable version script is tightened to 0.3 to avoid gradient overflow and the generation of non-finite values; Early Stop and Save: In each round of the validation set, the mIoU or MAPA metric is calculated. If the metric is better than the historical best, the complete checkpoint is saved. For tasks without image bridging, the backbone model weights, bridging network weights, temporal position codes, and standardized statistics are saved. For the 316L cross-material task, the backbone model weights containing LoRA low-rank adaptive parameters are saved. Continuous P... stop If there is no improvement in a round, training is terminated, where P stop This indicates the preset number of early stop patience rounds.

[0058] In the training implementation, if any batch loss has a non-finite value, the abnormal batch is skipped; the validation metric is selected according to the task, either mIoU or MAPA, and the corresponding checkpoint is saved when the metric is better than the historical best. The checkpoints for the image-less bridging task include the backbone model weights, bridging network weights, temporal position encoding, and normalized statistics; the checkpoints for the 316L cross-material task include the backbone model weights injected with low-rank adaptive and training statistics.

[0059] Step 105: Export of blind evaluation and prediction reports for independent test sets.

[0060] Load the optimal save checkpoint, freeze the model, bridging network, and temporal coding weights, and all samples of the gradient-free inference test set; calculate the single-point MAE, global average MAE, and mIoU / MAPA accuracy for each contour; concatenate the original process parameters, actual W / H / contour, predicted values, and point-by-point absolute errors into a data table and output it as a CSV file; at the same time, generate a JSON format diagnostic report, recording the dataset partitioning scale, optimal validation metrics, and the full set of quantitative accuracy of the test set, for experimental analysis and offline process review.

[0061] (1) Gradient clipping and prediction numerical truncation are implemented during training to avoid gradient explosion and numerical anomalies during high-power data training. A group optimizer is used to differentiate the learning rate of network parameters.

[0062] (2) Trigger an early stopping mechanism by using the validation set mIoU and MAPA indices to save the optimal migration prediction model. That is, load the optimal model weights for inference on the completely isolated test set and calculate the mIoU, MAPA, and single-point MAE indices.

[0063] The mIoU and MAE metrics during the blind evaluation phase of the test set are as follows: } (17); (18); (19); (20); In formulas (17) to (20), τ represents the tag noise truncation threshold, used to set the amplitude of real or predicted contour points with absolute values ​​less than the threshold to zero, in order to reduce the impact of small noise on the cross-union ratio calculation; q b,i This represents the effective amplitude after taking the absolute value of the i-th true contour point of the b-th sample and filtering it by a threshold. q represents the effective amplitude of the corresponding predicted contour point; b and Let represent the true contour amplitude sequence and the predicted contour amplitude sequence of the b-th sample, respectively; trapz represents the trapezoidal integral along the direction of the contour sampling point. This represents the sequence of smaller values ​​at each point in the two contour amplitude sequences. Represents the area of ​​intersection; The area of ​​the union is represented by ε; ε represents a small positive number used to avoid a denominator of zero; MAE i The mean absolute error (MAE) represents the average absolute error of the i-th contour point. avg y represents the average MAE of all contour points; y represents the true label. This indicates the prediction result.

[0064] (3) Complete blind inference evaluation based on a completely isolated test set, quantify the model migration prediction accuracy under image-free conditions, output complete prediction results and error analysis reports, export process parameters, true values, predicted values ​​and prediction errors in CSV files, and quantitative index JSON diagnostic reports.

[0065] Scenario 2: Addressing the challenge of switching between dissimilar materials in laser additive manufacturing, this project tackles the industry pain point of requiring extensive re-experimentation, data collection, and model training from scratch when changing to 316L stainless steel. Based on the TC4 titanium alloy source domain pre-trained model, and combined with hierarchical transfer learning and physical constraint mechanisms, it achieves high-precision transfer prediction of the molten pool profile of 316L stainless steel in small samples. The specific implementation steps are as follows: Step 201: Basic preparation for cross-material migration.

[0066] Using the TC4 titanium alloy laser additive manufacturing dataset as the source domain, we pre-trained a gated residual bridging network and a molten pool prediction backbone model, learning the basic characteristics of general laser molten pool melting, flow, and forming, as well as the process-molten pool mapping rules. We also solidified the source domain network weights and standardized statistics of process parameters. We collected a small sample of 316L stainless steel laser additive manufacturing process and molten pool forming data, considering the differences in thermal properties, laser absorptivity, melt rheological properties, and Marangoni convection characteristics between 316L stainless steel and TC4 titanium alloy. This resulted in the construction of a dedicated target domain small sample dataset, covering the true contour and geometric dimension labels of the molten pool under different process parameters.

[0067] Steps 202 and 316 are used to construct the target domain dataset for normalization.

[0068] Continuing the zero-leakage data partitioning rule at the melt channel level, the time-series process data of 316L stainless steel is grouped by melt channel, with each complete melt channel as a unit to complete the dataset partitioning. This ensures that the training, validation, and test datasets are completely independent of each other, avoiding transfer overfitting issues caused by data leakage. The dual-input separation mechanism is also adopted. The bridging branch reuses TC4 source domain standardized statistics to align parameter distributions, achieving cross-material feature manifold adaptation. The main branch retains the original physical dimensions of the 316L process parameters, maintaining the model's ability to capture material-specific process features, and achieving input dimension matching through zero-padding.

[0069] Step 203: Layered Hier-LoRA cross-material adaptive fine-tuning.

[0070] To address the differences in characteristics between dissimilar materials, a layered differentiated low-rank adaptation strategy is adopted. The bottom layer of the model, representing the general molten pool feature, is frozen, preserving the basic laser molten pool forming mechanism learned from the source domain. Low-rank adaptation ranks are set in the encoder layer of the model to fine-tune the shallow process feature shifts caused by material property differences. High-rank adaptation ranks are set in the higher-level cross-modal attention layer and decoder layer to accurately learn the nonlinear characteristics of molten pool flow and contour forming specific to 316L stainless steel. For the adapted linear layers, the original weights and biases remain frozen, and only a new low-rank adaptive matrix is ​​added for training. Its calculation formula is as follows: (twenty one); (twenty two); In formulas (21) to (22), x represents the features input to the replaced linear layer; f0(x) represents the frozen output of the original linear layer; W0 and b0 represent the original weights and biases of the linear layer in the source domain pre-trained backbone model, which are not updated during transfer training; A and B represent the newly added low-rank adaptive matrices, where A is used to project the input features to the low-rank space and B is used to map the low-rank features back to the output space; r represents the dimension of the low-rank space. Indicates the scaling factor. represents the scaling factor of the scaled low-rank incremental weights or low-rank adaptive increments; h represents the linear layer output after LoRA injection; the middle layer encoder uses a lower rank value to adapt to the shallow process feature offset; the high-level cross-modal attention layer and decoder use a higher rank value to enhance the ability to express material-specific nonlinear forming features.

[0071] Therefore, when crossing material scenarios, only the LoRA low-rank adaptive parameters and the adaptation parameters retained in the original model are updated, while the core basic weights are frozen throughout. This avoids catastrophic forgetting caused by cross-material fine-tuning and achieves efficient feature transfer in small sample scenarios.

[0072] Step 204: Layered fine-tuning of the Marangoni curvature physical constraints.

[0073] To address the Marangoni convection characteristics of the molten pool in 316L stainless steel, a one-dimensional curvature regularization loss function was employed to apply second-order difference smoothing constraints to the predicted pool profile. This eliminated distortion issues such as abrupt profile changes and non-physical turbulence that are prone to occur during small-sample training, ensuring that the prediction results conform to the forming law of 316L stainless steel liquid metal. The curvature regularization term was calculated according to formulas (12) to (16) in step 104, and combined with the preheating mechanism, hierarchical learning rate optimization, gradient clipping, numerical truncation, and abnormal batch skipping strategy, the cross-material transfer model was stabilized and fine-tuned.

[0074] Step 205: Export of blind evaluation and prediction reports for independent test sets.

[0075] After training, blind evaluation was conducted based on an independent test set to verify the model's time-series prediction accuracy for the molten pool profile, melt width, and melt height of 316L stainless steel. This enabled efficient cross-material migration prediction from TC4 titanium alloy to 316L stainless steel, significantly reducing the experimental cost and cycle of dissimilar material processes.

[0076] The process details using the process combination IDs shown in Table 1, based on the constructed TC4 database (laser power 1200~3200W, scanning speed 600~1200mm / min, powder feeding rate 12~18g / min), using the method of this embodiment, predict the high-power (8000W 2000mm / min 60g / min) TC4 forming cross-sectional profile without image or temperature as shown in the figure. Figure 3 As shown, the weld width and weld height are as follows Figure 4 As shown.

[0077] Table 1. Process Details of Process Combination ID

[0078] In high-power, image-free, and temperature-free cross-condition migration prediction, the optimal validation set mIoU (mean intersection-union ratio) of the method in this embodiment is 0.8412, the test set mIoU is 0.8613, and the average MAE (mean absolute error) of the contour points on the test set is 0.083397.

[0079] Using the process list with process combination IDs as shown in Table 2, the predicted cross-sectional profile is as follows using the method of this embodiment. Figure 5 As shown, the weld width and weld height are as follows Figure 6 As shown.

[0080] Table 2. Process List of Process Combination IDs

[0081] This embodiment provides a method for predicting the temporal migration of molten pool contours, which solves the problem of adapting to image-less conditions by bridging gated residuals across modes. It uses adaptive sigmoid gating to dynamically adjust the weights of high-order residual features of process parameters, and nonlinearly fits the process parameters with the visual embedding manifold distribution. By inputting only the process parameters, it can generate pseudo-visual features consistent with the image embedding distribution. It can still output the molten pool contour under conditions without a camera or with strong smoke and dust obstruction, and achieves a prediction result with an mIoU of 0.8613 on an independent test set.

[0082] This embodiment provides a method for predicting the temporal migration of melt pool contours, proposing a melt-track level temporal partitioning and a dual-input mechanism to prevent data leakage. The method uses complete melt tracks as partitioning units, without scattering single-track temporal samples; when using bridging statistics in the passive domain, standardization relies solely on training set statistics; when using bridging statistics in the active domain, the source domain mean and standard deviation are reused; the dual-branch separation balances cross-domain distribution alignment with the input requirements of the original model's physical dimensions, avoiding test set information leakage caused by traditional single-path standardization.

[0083] This embodiment provides a temporal migration prediction method for melt pool contours, proposing a hierarchical Hier-LoRA approach to achieve lightweight migration across materials and suppress catastrophic forgetting. The LoRA rank is configured hierarchically and differentially, with the bottom layer's general melt pool features completely frozen, while the middle and upper layers undergo rank fine-tuning, and only a small number of low-rank adaptive parameters participate in the update. The training process outputs total parameters, trainable parameters, and the percentage of trainable parameters to verify the lightweight nature of cross-material migration in a small sample of 316L melt pools, and MAPA is used as the evaluation index for melt width and melt height prediction accuracy.

[0084] This embodiment provides a method for predicting the temporal migration of molten pool contours, proposing Marangoni thermodynamic physical constraints to ensure the physical realism of the contours. A second-order curvature penalty loss is constructed based on the Marangoni convection mechanism of liquid metal, constraining the predicted contours to be smooth and continuous, eliminating abrupt changes and non-physical curves in the output of unconstrained models. A preheating mechanism balances the accuracy of basic predictions with the effect of physical constraints, and the process defect warning and parameter recommendation results are more consistent with the actual forming mechanism of L-DED.

[0085] This embodiment provides a method for predicting the temporal migration of melt pool contours, proposing an automated data processing, training, and evaluation output pipeline. The accompanying program automatically completes the construction of a CSV time-series dataset, model training, optimal weight saving, export of training / validation / test data partitioning tables, export of test prediction reports, and generation of JSON diagnostic reports, reducing the workload of manual data partitioning and indicator calculation.

[0086] Example 2 This embodiment provides a melt pool profile temporal migration prediction system, including: The source domain bridging pre-training module is configured to: load the source domain dataset with images, initialize the gated residual cross-modal bridging network, perform bridging network pre-training, and persistently save the bridging weights and standardized statistics (mean and standard deviation) of the source domain parameters.

[0087] The time-series data dual-path construction module is configured to: read the target domain image-free process data table CSV and the melt-sink level three-part dataset, divide it into training, validation and test sets, generate standardized bridging input tensors and original dimensional backbone input tensors respectively, and output the partitioned independent training / validation / test CSV files.

[0088] The transfer fine-tuning module is configured to: load the source domain pre-trained melt pool prediction backbone model, gated residual bridging network and temporal position encoding in the high-power scene without image, and construct a dual-path forward inference pipeline; inject low-rank adaptive layers into the source domain pre-trained backbone model in a hierarchical manner in the 316L cross-material scene; and configure the AdamW optimizer and learning rate scheduler.

[0089] The physical constraint evaluation and output module is configured to: integrate the Marangoni curvature hybrid loss function and calculate evaluation indicators such as mIoU, MAPA, and MAE in real time according to formulas (12) to (20); after training, load the optimal weights to perform blind inference on the test set and export the prediction error data table and the quantitative indicator diagnosis JSON.

[0090] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0091] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the melt pool profile timing migration prediction method described in Embodiment 1 above.

[0092] Example 4 This embodiment provides a computer device, such as... Figure 7 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps in the melt pool contour timing migration prediction method described in Embodiment 1 above.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the temporal migration of a molten pool profile, characterized in that, include: Obtain the titanium alloy source domain dataset, use process parameters as input and visual embedding as fitting labels, train the gated residual cross-modal bridging network, and save the standardized statistics of the trained bridging network weights and source domain process parameters. Acquire time-series process data in the target domain, divide the dataset into training, validation, and test sets using complete melt channels as units; reuse the standardized statistics or use the training set statistics to standardize the inputs of the bridging branches to obtain standardized parameters; retain the original physical dimensions of the inputs of the main branches, and match the parameter input dimensions of the main model by zero padding. Load the molten pool time-series prediction backbone model pre-trained with titanium alloy source domain dataset; in the target domain transfer scenario without image, input the standardized parameters into the bridging network to generate pseudo-visual embeddings, superimpose the learnable temporal position encoding as the visual channel input of the backbone model, and use the target domain process parameters that retain the original physical dimensions as the process parameter channel input of the backbone model. In cross-material migration scenarios, a low-rank adaptation layer is injected into the backbone model, and a Marangoni curvature regularization loss is superimposed on the basic combined prediction loss. The bridging network and backbone model are validated using the validation set to obtain the optimal migration prediction model. The optimal migration prediction model is then used for blind inference using the test set to output the predicted results of the melt pool profile, melt width, or melt height.

2. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, The Marangoni curvature regularization loss is: ； Among them, the curvature weights in the e-th training round , This represents the maximum value of the curvature weight. The curvature weight increases linearly from 0 to The number of preheating rounds, B represents the batch sample size, and n represents the number of contour sampling points or label dimension; the one-dimensional second-order difference calculated from three adjacent predicted contour points in the b-th sample. ; This represents a numerical truncation function. , and Both represent the predicted contours after numerical truncation.

3. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, The bridging network is a gated residual cross-modal bridging network: it completes the dimensional mapping of process parameters through a linear projection layer, extracts high-order interactive features of process parameters by combining residual blocks, dynamically adjusts the residual feature weights by relying on adaptive gated branches, and completes feature distribution calibration through a normalization layer.

4. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, The process of dividing the dataset into complete melt channels includes: dividing the target domain time-series process data into groups based on complete melt channels as the smallest independent unit. Each melt channel contains a fixed length of continuous time-series process samples. Only the overall sequence number of the melt channel is shuffled to complete the division of the training set, validation set, and test set, without splitting the time-series data within a single melt channel.

5. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, The hierarchical injection of low-rank adaptation layers includes: freezing the basic weights of the melt pool temporal prediction backbone model, configuring low-rank LoRA adaptation layers for the mid-level encoder network, configuring high-rank LoRA adaptation layers for the high-level cross-modal attention layer and decoder network, and only low-rank adaptive parameters and reserved adaptation parameters participate in the update.

6. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, When optimizing parameters, a base learning rate is set for the trainable layers of the backbone model, and an independent learning rate is set for the bridging network, temporal position encoding, or low-rank LoRA adaptation layer. No weight decay is applied to the temporal position encoding.

7. The method for predicting the temporal migration of a molten pool profile as described in claim 1, characterized in that, For the task of predicting the molten pool profile, the basic combined prediction loss is used and the average crossover ratio is used as the verification index; for the task of predicting the molten width or molten height, the weighted mean square error loss is used and the average absolute percentage accuracy is used as the verification index.

8. A melt pool profile time-series migration prediction system, characterized in that, include: The source domain bridging pre-training module is configured to: acquire the titanium alloy source domain dataset, use process parameters as input, use visual embeddings as fitting labels, train the gated residual cross-modal bridging network, and save the standardized statistics of the trained bridging network weights and source domain process parameters. The time-series data dual-path construction module is configured to: acquire time-series process data in the target domain, divide the dataset into training, validation and test sets by using complete melt channels as units; reuse the standardized statistics or use the training set statistics to standardize the input of the bridging branch to obtain standardized parameters; retain the original physical dimensions of the input of the main branch and match the parameter input dimension of the main model by zero padding. The fine-tuning module is configured to: load the molten pool time-series prediction backbone model pre-trained with the titanium alloy source domain dataset; in the image-free target domain transfer scenario, input the standardized parameters into the bridging network to generate pseudo-visual embeddings, superimpose the learnable temporal position encoding as the visual channel input of the backbone model, and use the target domain process parameters that retain the original physical dimensions as the process parameter channel input of the backbone model. In cross-material migration scenarios, a low-rank adaptation layer is injected into the backbone model, and a Marangoni curvature regularization loss is superimposed on the basic combined prediction loss. The evaluation and output module is configured to validate the bridging network and backbone model based on the validation set, obtain the optimal migration prediction model, and perform blind inference on the optimal migration prediction model using the test set to output the predicted results of the melt pool profile, melt width, or melt height.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the melt pool profile timing migration prediction method as described in any one of claims 1-7.

10. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the melt pool profile timing migration prediction method as described in any one of claims 1-7.

Images