A method for quantitative characterization and performance prediction of additive manufacturing steel organization based on sparse label self-evolution

Abstract

The application discloses a kind of based on sparse label self-evolution's additive manufacturing steel organization quantitative characterization and performance prediction method, the original image of additive manufacturing steel is cut and batch generates weak mask, forms weak label image block;Select the most representative part image to carry out interactive pixel-level fine-tuning, obtain high-precision fine-tuning mask;Attention U-Net with attention gate is constructed to semantic segmentation network, utilize deep layer semantic feature as gate signal, self-adaptively suppress martensite false boundary and activate real grain boundary feature;In turn through weak supervision preheating, expert knowledge injection fine-tuning, and confidence weighted semi-supervised evolution, obtain the teacher model with high precision;Teacher model is used to weak label image block is re-inferred, generates high-quality pseudo label, initializes a brand-new, Attention U-Net network without pre-training as student model, again using the above weak label image block and its generated high-quality pseudo label, the student model is completely trained from zero, realize the global generalization of expert knowledge;For grain topological quantification and yield strength cross-scale prediction;Realize the forward prediction from micro digital topology to macro yield strength.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision, materials intelligence and additive manufacturing technology, and specifically relates to a method for quantitative characterization and performance prediction of additive manufacturing steel microstructure based on sparse tag self-evolution. Background Technology

[0002] In the research and quality inspection of high-performance metallic materials, the topological structure and grain size distribution of in-situ austenite grain boundaries are core physical factors determining the macroscopic yield strength of the material. However, under the rapid non-equilibrium solidification conditions of additive manufacturing (such as laser wire additive manufacturing), complex acicular substructures such as high-density lath martensite are generated inside alloy steel. The boundaries of these substructures are highly similar to the real in-situ austenite grain boundaries, leading to the following problems in existing additive manufacturing microstructure characterization and performance prediction technologies, including traditional metallographic image processing methods such as watershed algorithms and adaptive thresholding:

[0003] (1) Severe interference from complex substructures: The rapid non-equilibrium solidification of laser additive manufacturing (LWAM) will produce substructures such as high-density martensite laths in steel. Their visual characteristics are highly similar to the real original austenite grain boundaries (PAGBs), which makes traditional image segmentation algorithms prone to boundary fractures and oversegmentation.

[0004] (2) High cost of expert annotation and bottleneck of small sample size: High-precision grain boundary annotation is extremely dependent on the knowledge of metallurgical experts, and the sample size is often less than 20 images. With such a small sample size, traditional fully supervised deep learning is prone to overfitting, and the massive amount of unlabeled data is seriously wasted.

[0005] (3) Disconnect between visual indicators and mechanical properties: Existing AI models mostly focus on optimizing computer vision indicators (such as IoU) and fail to establish a physical causal chain from microscopic geometric features to macroscopic mechanical properties (such as yield strength), making it difficult to achieve non-destructive mechanical assessment in industry.

[0006] Therefore, in view of the above-mentioned defects in the existing technology, a method is needed to solve the above problems in order to obtain high-fidelity grain geometry topology data. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method for quantitative characterization and performance prediction of additive manufacturing steel microstructure based on sparse tag self-evolution. The technical solution adopted by this invention is as follows:

[0008] A quantitative characterization and property prediction method for additive manufacturing steel based on sparse tag self-evolution includes the following steps:

[0009] S1: Obtain the original image of the additively manufactured steel and crop it into microstructure image blocks. Generate weak masks in batches from the original microscopic images to form weak label image blocks.

[0010] Select n of the most representative images for interactive pixel-level retouching to obtain a high-precision retouching mask, which will serve as the seed dataset.

[0011] S2: Construct an Attention U-Net semantic segmentation network with attention gates, embed attention gate modules at the skip connections of the Attention U-Net encoder-decoder, and use deep semantic features as gating signals to adaptively suppress martensitic pseudo-boundaries and activate real grain boundary features.

[0012] S3: Perform three-stage progressive training, successively through weak supervision warm-up, expert knowledge injection fine-tuning, and confidence-weighted semi-supervised evolution, to obtain a teacher model with high accuracy;

[0013] S4: Teacher-student knowledge distillation. The teacher model is used to re-infer the weakly labeled image patches generated in S1 to generate high-quality pseudo-labels. A brand-new, untrained Attention U-Net network is initialized as the student model. The student model is then trained from scratch using the aforementioned weakly labeled image patches and the high-quality pseudo-labels it generates, achieving global generalization of expert knowledge.

[0014] S5: Quantitative analysis of grain topology and cross-scale prediction of yield strength; the segmentation results are skeletonized and segmented into instances, the median grain area is automatically extracted and substituted into the Hall-Petch constitutive model to achieve positive prediction from microscopic digital topology to macroscopic yield strength.

[0015] Furthermore, all images are batch-processed using a computer vision pipeline to generate low-precision weak masks containing fracture and erosion artifacts.

[0016] Furthermore, pixel-level refinement: white pixel brushes are used to erase false boundaries and noise introduced by chemical etching, and black pixel brushes are used to fill in broken martensite laths and original austenite grain boundaries. After cropping, high-precision image blocks are obtained.

[0017] Furthermore, in the semantic segmentation network Attention U-Net, the deep semantic features of the decoder are used as gating signals to perform pixel-by-pixel nonlinear weight redistribution on the shallow spatial features; the calculation formula is expressed as:

[0018] ;

[0019] in, Attention coefficient It is the Sigmoid activation function. The linear transformation weight matrix is ​​the feature of skip connections. The linear transformation weight matrix of the gated signal. The transformation weights are those before nonlinear activation. The output features are reweighted; for regions identified as "martensite pseudo-boundaries", the attention coefficient is reduced to suppress them; for regions identified as "original austenite grain boundaries", the attention coefficient is increased to enhance activation.

[0020] Furthermore, the three-stage progressive training steps for S3 are as follows:

[0021] Phase 1: Initial training of the Attention U-Net constructed by S2 using weakly labeled image patches;

[0022] The second stage involves using high-precision refined masks for transfer learning, employing an asymmetric learning rate strategy. This involves freezing the encoder layer by reducing its learning rate, and increasing the learning rates of the decoder and attention gates to fully activate them.

[0023] The third stage involves using the model trained in the second stage to generate pseudo-labels for the remaining unlabeled images, and calculating the entropy or probability value of the predicted probability distribution as a confidence index. The pseudo-labeled data with the highest confidence is selected and mixed with a high-precision refined mask to form a high-quality labeled subset. This subset is then used for semi-supervised fine-tuning. Tversky Loss is introduced to penalize false negatives, and a confidence-weighted pseudo-label strategy is used to guide the model to converge smoothly on the remaining unlabeled data, thus obtaining the teacher model.

[0024] Furthermore, the S5 student model: the original image of the unknown microstructure of additively manufactured steel is obtained as the input of the student model, and the direct output of the student model is a single-channel pixel-level segmentation probability map, representing the probability that each pixel belongs to a grain boundary.

[0025] Furthermore, in S5, hole filling, small connected component removal, and skeletonization are performed on the segmentation probability map, and the watershed algorithm based on distance transformation is used to complete the instance segmentation of independent grains.

[0026] Furthermore, Hall-Petch macroscopic performance prediction: converting the median grain area to the effective grain diameter. Where C is the irregularity shape factor and A_med is the median grain area.

[0027] Furthermore, the Hall-Petch constitutive model is denoted as: ;

[0028] in, This is the predicted theoretical yield strength. The baseline yield strength, This is the grain boundary strengthening coefficient.

[0029] Beneficial effects

[0030] (1) Extremely high data efficiency: Relying on only a small number (e.g. 10) expert-refined images, the self-evolutionary framework successfully generalizes expert knowledge to tens of thousands of unlabeled image blocks, completely eliminating the dependence of traditional deep learning on massive manual annotation.

[0031] (2) Excellent anti-interference and feature extraction capabilities: The introduced attention mechanism and progressive training effectively suppress pseudo-edge noise such as martensitic laths and twins, and the terminal model achieves an extremely high accuracy of mIoU 0.966.

[0032] (3) Achieved high-fidelity cross-scale prediction: opened up the closed loop of “pixel → grain → macroscopic performance”. The relative error between the theoretically predicted yield strength and the measured value of the same micro tensile test is only 17.4%, providing an efficient and low-cost non-destructive mechanical evaluation scheme for the advanced manufacturing field. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall framework of the method of the present invention, showing the closed-loop process from sample preparation, evolution segmentation, feature quantization to cross-scale performance prediction.

[0034] Figure 2 This is a schematic diagram of the overall architecture of Attention U-Net and the structure of the Attention Gate (AG) module of the present invention.

[0035] Figure 3 This is the overall execution logic diagram for the three-stage progressive training strategy and iterative knowledge distillation framework.

[0036] Figure 4 This is a comparison chart of integrated results from the microstructure analysis pipeline, demonstrating the visualization effects from the original microscopic images to the boundary skeleton and final grain instance segmentation.

[0037] Figure 5 This is an engineering stress-displacement curve of a homologous micro tensile specimen used for mechanical verification. Detailed Implementation

[0038] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0039] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0040] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. While only preferred methods and materials have been described herein, any methods and materials similar or equivalent to those described herein may be used in the implementation or testing of this invention. All references to this specification are incorporated by way of citation to disclose and describe methods and / or materials associated with those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

[0041] Various modifications and variations can be made to the specific embodiments described in this specification without departing from the scope or spirit of the invention, as will be apparent to those skilled in the art. Other embodiments derived from this specification will also be obvious to those skilled in the art. This application specification and embodiments are merely exemplary.

[0042] The terms “include,” “including,” “have,” “contain,” etc., used in this article are all open-ended terms, meaning that they include but are not limited to.

[0043] The following is in conjunction with the appendix Figure 1-5 The present invention will be further described in detail with specific embodiments. This embodiment uses 10CrNi3MoV high-strength steel manufactured by laser wire additive manufacturing (LWAM) as the reference system.

[0044] Step S1: Hybrid data annotation and seed dataset construction.

[0045] Raw data acquisition: Metallographic polishing and chemical etching (4% nitric acid alcohol) were performed on 10CrNi3MoV steel samples prepared by LWAM. 500 raw images were acquired using an optical microscope and cropped to generate 19,811 microstructure image blocks.

[0046] Weak Mask Generation: All images are batch-processed using a computer vision pipeline to generate low-precision "weak masks" containing fractures and erosion artifacts. The computer vision pipeline includes processing algorithms such as erosion, dilation, and watershed.

[0047] Pixel-level interactive retouching: Ten of the most representative images were selected and retouched pixel-level using an interactive image overlay correction method. White pixel brushes were used to erase false boundaries and noise introduced by chemical etching, and black pixel brushes were used to fill in fractured martensite laths and original austenite grain boundaries. After cropping, 540 high-precision image blocks were obtained as the "seed dataset" for subsequent evolution training.

[0048] Step S2: Construct an Attention U-Net with embedded attention gates (AG).

[0049] For martensitic substructure interference, such as Figure 2 As shown, an attention gate (AG) module is embedded at the encoder-decoder jump connection of U-Net.

[0050] Its core operating mechanism is as follows: using the deep semantic features of the decoder as the gating signal, a pixel-by-pixel nonlinear weight redistribution is performed on the shallow spatial features. The calculation formula is:

[0051]

[0052] in, Attention coefficient It is the Sigmoid activation function. The linear transformation weight matrix is ​​the feature of skip connections. The linear transformation weight matrix of the gated signal. The transformation weights are those before nonlinear activation. This is the output feature after weight redistribution (i.e., the calibrated feature). For regions identified as "martensitic pseudo-boundaries", the attention coefficient is set to close to 0 to suppress them; for regions identified as "proto-austenitic grain boundaries", the coefficient is set to close to 1 to enhance activation.

[0053] Step S3: Three-stage progressive model evolution.

[0054] like Figure 3 As shown, in order to obtain a highly robust model with sparse labels, the following three-stage training is performed:

[0055] Phase 1 (Weakly Supervised Warm-up): Using these 19,811 image patches with "low-precision weak masks" as input and labels, the Attention U-Net constructed by S2 is initially trained. Dice Loss and BCE Loss are used for initial training to establish basic structure perception capability (Base Model).

[0056] The BCE Loss expression is:

[0057]

[0058] The Dice Loss expression is:

[0059]

[0060] in, This represents the total number of pixels in the image block. For pixel index; For the first The true label value of each pixel (for grain boundary pixels) For background pixels ); For the model to predict the first The probability value of a pixel belonging to a grain boundary (ranging from 0 to 1).

[0061] Based on the above loss, the Combo Loss expression is constructed as follows:

[0062]

[0063] in, For binary cross-entropy loss, For the focus of loss, For Dice's loss; , , These are the weight coefficients of the corresponding loss function, and they satisfy...

[0064] + + =1.

[0065] Phase Two (Expert Knowledge Injection): 540 high-precision refined masks are introduced for transfer learning. An asymmetric learning rate strategy is adopted: to prevent the model from forgetting the basic micro-texture, an extremely low learning rate (almost frozen, with a range set to...) is assigned to the encoder layer. to ), while assigning higher learning rates (e.g., setting them to ) to the decoder and attention gate (AG) modules. This fully activates the model, forcing it to focus on recalibrating attention weights and spatial reconstruction. This stage results in a vertical leap in model accuracy (mIoU).

[0066] The Tversky Loss is introduced in the second and third stages primarily to address the problem of extreme class imbalance and to penalize false negatives (missed detections). This loss function is designed to force the model to pay attention to subtle grain boundary fragments, ensuring that the grain boundaries in the network output have topological continuity.

[0067] The third stage (semi-supervised evolution): Using the model trained in the second stage, pseudo-labels are generated from the remaining unlabeled images through inference. The entropy or probability value of the predicted probability distribution is calculated as a confidence index. The pseudo-labeled data with the highest confidence is selected and mixed with 540 high-precision refined masks to form a high-quality labeled subset accounting for approximately 20% of the total data. This subset is then used to fine-tune the model in a semi-supervised manner. Tversky Loss is introduced to penalize false negatives (missed detections), and a confidence-weighted pseudo-label strategy guides the model to converge smoothly on the remaining unlabeled data, resulting in the Teacher Model.

[0068] Step S4: Teacher-Student Knowledge Distillation

[0069] To completely eliminate systematic biases in weak labels and achieve global generalization of knowledge, knowledge distillation is performed:

[0070] The high-fidelity "teacher model" from step S3 is used to re-infer all 19,811 image patches with "low-precision weak masks" to generate high-quality binarized pseudo-labels (soft targets).

[0071] Subsequently, a "Student Model" with the same network architecture as the teacher model was initialized and trained from scratch under the supervision of a full set of high-quality pseudo-labels (constructing a student model with the same architecture but randomly initialized weights, and performing complete end-to-end training under the supervision of a full set of high-quality pseudo-labels). The final generated student model achieved an mIoU of 0.966 and a standard deviation of only 0.019, reaching extremely high expert-level segmentation stability.

[0072] Step S5: Grain instance segmentation and cross-scale yield strength prediction.

[0073] The original image of the unknown microstructure of additively manufactured steel is obtained and input into a trained student model. The direct output of the student model is a single-channel pixel-level probability map, representing the probability that each pixel belongs to a grain boundary. Figure 4 As shown, this approach breaks away from purely visual output and applies the segmentation probability map output by the student model to physical quantization:

[0074] Grain topology quantization: The segmentation probability map (the direct output map of the student model inference) undergoes hole filling, small connected component removal, and skeletonization. A watershed algorithm based on distance transform is used to segment individual grains. Key descriptor extraction: Median grain area (A_med) is used to avoid statistical contamination of the arithmetic mean by extremely large grains.

[0075] Hall-Petch macroscopic performance prediction: converting median grain area to effective grain diameter Where C is the irregularity shape factor and A_med is the median grain area, used to avoid statistical contamination of the average value by extremely large grains.

[0076] Combining the lattice friction stress (including solid solution strengthening and dislocation strengthening) independently determined by physical metallurgical principles with the material strengthening coefficient, we substitute these values ​​into the Hall-Petch equation for forward derivation:

[0077]

[0078] in, This is the predicted theoretical yield strength.

[0079] Experimental verification: To verify the physical fidelity of the model, such as... Figure 5 As shown, micro-samples were extracted from adjacent deposition areas for homologous uniaxial tensile tests. The baseline measured tensile yield strength was 713.3 MPa, while the theoretically predicted value for the pipeline in this invention was 837.5 MPa. The relative error between the two was only 17.4%, which mainly stemmed from the inherent internal stress and micro-yield mechanism of additive manufacturing. This error is well within the acceptable threshold range for industrial non-destructive evaluation, demonstrating the high reliability of this method in the exploration of smart materials.

[0080] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for quantitative characterization and property prediction of additive manufacturing steel microstructure based on sparse tag self-evolution, characterized in that, Includes the following steps: S1: Obtain the original image of the additively manufactured steel and crop it into microstructure image blocks. Generate weak masks in batches from the original microscopic images to form weak label image blocks. Select n of the most representative images for interactive pixel-level retouching to obtain a high-precision retouching mask, which will serve as the seed dataset. S2: Construct an Attention U-Net semantic segmentation network with attention gates, embed attention gate modules at the skip connections of the Attention U-Net encoder-decoder, and use deep semantic features as gating signals to adaptively suppress martensitic pseudo-boundaries and activate real grain boundary features. S3: Perform three-stage progressive training, successively through weak supervision warm-up, expert knowledge injection fine-tuning, and confidence-weighted semi-supervised evolution, to obtain a teacher model with high accuracy; S4: Teacher-student knowledge distillation. The teacher model is used to re-infer the weakly labeled image patches generated in S1 to generate high-quality pseudo-labels. A brand-new, untrained Attention U-Net network is initialized as the student model. The student model is then trained from scratch using the aforementioned weakly labeled image patches and the high-quality pseudo-labels it generates, achieving global generalization of expert knowledge. S5: Quantitative analysis of grain topology and cross-scale prediction of yield strength; the segmentation results are skeletonized and segmented into instances, the median grain area is automatically extracted and substituted into the Hall-Petch constitutive model to achieve positive prediction from microscopic digital topology to macroscopic yield strength.

2. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, All images are batch-processed using a computer vision pipeline to generate low-precision weak masks containing fracture and erosion artifacts.

3. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, Pixel-level retouching: White pixel brushes are used to erase false boundaries and noise introduced by chemical etching, and black pixel brushes are used to fill in broken martensite laths and original austenite grain boundaries. After cropping, high-precision image blocks are obtained.

4. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, In the semantic segmentation network Attention U-Net, the deep semantic features of the decoder are used as the gating signal to perform pixel-by-pixel nonlinear weight redistribution on the shallow spatial features; the calculation formula is expressed as: ； in, Attention coefficient It is the Sigmoid activation function. The linear transformation weight matrix is ​​the feature of skip connections. The linear transformation weight matrix of the gated signal. The transformation weights are those before nonlinear activation. The output features are reweighted; for regions identified as "martensite pseudo-boundaries", the attention coefficient is reduced to suppress them; for regions identified as "original austenite grain boundaries", the attention coefficient is increased to enhance activation.

5. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, The three-stage progressive training steps for S3 are as follows: Phase 1: Initial training of the Attention U-Net constructed by S2 using weakly labeled image patches; The second stage: using high-precision refined masks for transfer learning, and adopting an asymmetric learning rate strategy, the encoder layer is frozen by reducing the learning rate of the encoder layer. Increase the learning rate of the decoder and attention gate to fully activate them; The third stage involves using the model trained in the second stage to generate pseudo-labels for the remaining unlabeled images, and calculating the entropy or probability value of the predicted probability distribution as a confidence index. The pseudo-labeled data with the highest confidence is selected and mixed with a high-precision refined mask to form a high-quality labeled subset. This subset is then used for semi-supervised fine-tuning. Tversky Loss is introduced to penalize false negatives, and a confidence-weighted pseudo-label strategy is used to guide the model to converge smoothly on the remaining unlabeled data, thus obtaining the teacher model.

6. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, S5 Student Model: The original image of the unknown microstructure of additively manufactured steel is obtained as the input of the student model. The direct output of the student model is a single-channel pixel-level segmentation probability map, representing the probability that each pixel belongs to a grain boundary.

7. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, In S5, hole filling, small connected component removal, and skeletonization are performed on the segmentation probability map, and the watershed algorithm based on distance transformation is used to complete the instance segmentation of independent grains.

8. The method for quantitative characterization and performance prediction of additive manufacturing steel based on sparse tag self-evolution as described in claim 1, characterized in that, Hall-Petch macroscopic performance prediction: converting median grain area to effective grain diameter Where C is the irregularity shape factor and A_med is the median grain area.

9. A quantitative characterization and performance prediction method for additive manufacturing steel microstructure based on sparse tag self-evolution characterized by the Hall-Petch constitutive model, as described in claim 8.

Images