A coating life prediction method based on multi-modal contrast and attention fusion
By employing a multimodal comparison and attention fusion approach, the heterogeneous fusion problem of process/composition and micromorphology data was solved, enabling high-precision coating lifetime prediction, reducing data acquisition costs, and providing physical interpretability verification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANMUSHAN LABORATORY
- Filing Date
- 2026-06-26
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies struggle to effectively integrate heterogeneous data on process/composition (tables) and microstructure (images), resulting in low accuracy in predicting the lifespan of thermal barrier coatings with small samples and high data acquisition costs.
A multimodal contrastive and attention fusion approach is adopted. By using a dual-tower feature extraction network and an attention-based feature fusion module, combined with the InfoNCE contrastive learning mechanism and cross-attention mechanism, a coating lifetime prediction model is constructed to achieve semantic alignment and high-precision prediction of table features and image features.
It significantly improves the accuracy of coating lifetime prediction and the robustness of the model, lowers the data acquisition threshold, supports single-modal reasoning without image input, and provides physical interpretability verification.
Smart Images
Figure CN122452389A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the fields of high-temperature material performance prediction for aero-engines and artificial intelligence, specifically to a coating lifetime prediction method based on multimodal comparison and attention fusion. Background Technology
[0002] Thermal barrier coatings (TBCs) are indispensable protective materials for modern aero-engines and gas turbine blades. However, when engines operate in environments containing sand, volcanic ash, or industrial dust, inhaled calcium-magnesium-aluminum-silicate (CMAS) particles melt at high temperatures and deposit on the coating surface. The molten CMAS penetrates along the coating pores, causing thermochemical corrosion and phase transformation, leading to premature coating peeling and failure, seriously threatening flight safety. Therefore, accurately predicting the residual life of TBCs under CMAS erosion is crucial for engine operation and maintenance decisions.
[0003] Currently, methods for predicting the lifetime of corrosion cracking (TBC) are mainly divided into physical experimental methods, numerical simulation methods, and traditional data-driven methods. Traditional physical experiments simulate the corrosion environment through high-temperature thermal shock experiments. Although the results are reliable, they are time-consuming, costly, and difficult to exhaustively account for all CMAS components and process combinations. Numerical simulation mainly uses the finite element method (FEM) to simulate the stress field. However, CMAS corrosion involves complex chemical penetration and mechanical coupling, making it extremely difficult to establish an accurate multiphysics constitutive model. Furthermore, simulations are usually based on idealized microstructures, making it difficult to reflect the true heterogeneity of the coating.
[0004] In recent years, machine learning has been applied to material lifetime prediction. Most existing techniques are single-modal, either based solely on tabular process parameters (ignoring the actual molding quality of the microstructure) or solely on microscopic images (SEM) (ignoring the quantitative influence of process parameters and CMAS chemical composition). When processing multimodal data, existing methods typically employ simple feature concatenation, ignoring the physical semantic misalignment between tabular and image features.
[0005] In summary, how to effectively integrate heterogeneous data of process / composition (tables) and microstructure (images), and achieve physical alignment of features and high-precision lifetime prediction with small samples, is a technical challenge that urgently needs to be solved. Summary of the Invention
[0006] In response, this disclosure provides a coating lifetime prediction method based on multimodal contrast and attention fusion.
[0007] This disclosure provides a coating lifetime prediction method based on multimodal contrast and attention fusion, including the following steps:
[0008] A multimodal dataset of thermal barrier coating samples was obtained. The multimodal dataset includes tabular data and microstructure image data after failure. The tabular data includes CMAS composition parameters and process parameters, and is labeled with thermal cycling failure lifetime.
[0009] A lifetime prediction model for thermal barrier coatings was constructed. The model architecture includes a dual-tower feature extraction network, a feature fusion module based on an attention mechanism, and a predictor based on a fully connected layer.
[0010] During the training phase, the dual-tower feature extraction network is optimized through contrastive learning to improve the table feature vectors. With image feature vectors Semantic alignment in the latent space;
[0011] Using a multimodal dataset, the parameters of the prediction model are trained by calculating the mean squared error loss, ranking loss of the predictor, and contrastive loss of the feature extraction network, thus obtaining the trained prediction model.
[0012] During the model deployment phase, only tabular data is input, the feature fusion module degenerates from cross-attention to a self-attention mechanism, and the predictor outputs lifetime predictions.
[0013] In some embodiments, the multi-construction thermal barrier coating lifetime prediction model includes:
[0014] Construct a dual-tower feature extraction network to extract table feature vectors from the tabular data. Image feature vectors of image data In this case, the two feature vectors have the same dimension;
[0015] Construct an attention-based feature fusion module to use table feature vectors. As a query Image feature vectors as keys Sum Multimodal fusion feature vectors are generated through an attention mechanism. ;
[0016] Construct a predictor based on fully connected layers to fuse multimodal feature vectors. This is mapped to a predicted coating lifetime.
[0017] In some embodiments, during the model deployment phase, only tabular data is input, the feature fusion module degenerates from cross-attention to a self-attention mechanism, and the predictor outputs lifetime predictions, including:
[0018] During the model deployment phase, only tabular data is input, and the input source is switched from cross-attention during the training phase to self-attention, using tabular feature vectors. Simultaneously serving as a query ,key Sum Calculate the fused feature vector The predictor outputs the lifetime prediction value.
[0019] In some embodiments, during the training phase, the dual-tower feature extraction network is optimized through contrastive learning to improve the table feature vector. With image feature vectors Semantic alignment in the latent space includes:
[0020] Construct positive and negative sample pairs:
[0021] Positive sample pairs: tabular feature vectors of the same experimental sample With image feature vectors ;
[0022] Negative sample pair: The tabular feature vector of any test sample in the current batch. Image feature vectors corresponding to other test samples To form negative sample pairs, where, ;
[0023] The InfoNCE loss function is used, and the formula is as follows:
[0024]
[0025] in, The cosine similarity function is used. For temperature coefficient, The first in the batch The tabular feature vector of each sample, The first in the batch Image feature vectors of each sample The first in the batch Image feature vectors of each sample This refers to the batch size.
[0026] In some embodiments, the step of using a multimodal dataset and training the parameters of the prediction model by calculating the mean squared error loss, ranking loss, and contrastive loss of the feature extraction network to obtain the trained prediction model includes:
[0027] The total loss function, which includes mean squared error loss, ranking loss, and comparison loss, is constructed as follows:
[0028]
[0029] in, These are hyperparameter weights used to balance the importance of each loss term.
[0030] Mean square error loss Used to monitor the accuracy of lifetime predictions, defined as:
[0031]
[0032] In the formula, For actual lifespan, To predict lifetime values, Batch size;
[0033] Ranking loss To maintain consistency in lifetime ranking among samples, for any two samples and If the actual lifespan Then the constrained predicted value satisfies ,in The ranking loss is defined as follows, where the preset boundary threshold is used:
[0034] .
[0035] In some embodiments, obtaining the multimodal dataset of thermal barrier coating samples includes:
[0036] Preparation of thermal barrier coating specimens;
[0037] CMAS powders with different component contents were prepared and coated onto the surface of the test piece;
[0038] High-temperature thermal shock tests were conducted on CMAS-coated samples at different test temperatures.
[0039] Microstructural image data of failed samples were collected, tabular data including CMAS component content, CMAS viscosity and test temperature were compiled, and a multimodal dataset was constructed using thermal cycling failure lifetime as a label.
[0040] In some embodiments, prior to the model deployment phase, the method further includes:
[0041] The interpretability of the model was verified using Grad-CAM (Grad-Weighted Class Activation Mapping).
[0042] The beneficial effects of this disclosure are that, compared with the prior art, this disclosure has the following advantages:
[0043] 1. During the model training phase, a multimodal data fusion architecture is constructed by combining macroscopic process parameters with post-failure SEM images. By introducing the InfoNCE contrastive learning mechanism, the structural features of the microscopic images are distilled into tabular features, enabling the model to learn the intrinsic physical correspondence between process parameters and post-failure microstructures. Combined with a cross-attention mechanism, the model can automatically focus on the most relevant microstructural regions in the SEM images, including crack propagation paths and penetration depths, based on CMAS composition and process parameters. These mechanisms enable the model to more comprehensively characterize the degradation state of thermal barrier coatings under CMAS erosion, significantly improving the accuracy of lifetime prediction while enhancing feature robustness and the model's generalization ability on small sample data.
[0044] 2. In the model prediction stage, it supports image-free inference without image input, and can achieve high-precision lifetime prediction based solely on CMAS components and process parameters, which greatly reduces the data acquisition threshold for industrial deployment.
[0045] 3. To address the lifetime prediction problem, Ranking Loss is introduced. Compared to simply pursuing numerical regression, maintaining the relative ranking of lifetime lengths among samples has stronger practical guiding significance for material selection and process optimization.
[0046] 4. Physical interpretability is achieved through Grad-CAM. The model can locate the main areas causing failure in SEM images, verify whether the model has learned the correct CMAS corrosion mechanism, and provide intuitive feedback for material design and process improvement. Attached Figure Description
[0047] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.
[0048] Figure 1 A schematic flowchart of a coating lifetime prediction method based on multimodal contrast and attention fusion provided in this disclosure embodiment;
[0049] Figure 2 A schematic diagram illustrating the process of high-temperature thermal shock testing and multimodal data acquisition provided in this embodiment of the disclosure;
[0050] Figure 3 A schematic diagram of a TBC test piece provided in an embodiment of this disclosure;
[0051] Figure 4 A schematic diagram of the high-temperature thermal shock testing apparatus provided in the embodiments of this disclosure;
[0052] Figures 5-6This is a SEM microstructure image of a failed sample cross-section provided in an embodiment of this disclosure;
[0053] Figure 7 This is a network architecture diagram of a coating lifetime prediction method based on multimodal contrast and attention fusion provided in an embodiment of the present disclosure.
[0054] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation
[0055] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present disclosure more clearly, and should not be used to limit the scope of protection of the present disclosure.
[0056] like Figure 1 As shown, this embodiment provides a coating lifetime prediction method based on multimodal contrast and attention fusion, including the following steps:
[0057] S1. Obtain a multimodal dataset of thermal barrier coating samples. The multimodal dataset includes tabular data and microstructure image data after failure. The tabular data includes CMAS composition parameters and process parameters, and uses thermal cycling failure lifetime as a label.
[0058] like Figure 2 As shown, step S1 specifically includes the following sub-steps:
[0059] S11, Prepare thermal barrier coating test pieces;
[0060] Using a high-temperature alloy as a substrate, a bonding layer was prepared on the substrate surface using a multi-arc ion plating process, and a YSZ ceramic layer was deposited on the bonding layer surface using an electron beam physical vapor deposition (EB-PVD) process to obtain a thermal barrier coating specimen.
[0061] For example, a thermal barrier coating specimen is prepared with a thickness of 5 mm and a diameter of 25 mm. The coating surface is chamfered to reduce premature peeling due to edge effects. Grooves are provided along the circumference of the specimen's side to fix it to the sample stage. The shape and dimensions of the specimen are as follows... Figure 3 As shown. The adhesive layer was prepared using a multi-arc ion plating process, using NiCrAlY material, with a thickness of 40–60 μm; the ceramic layer was prepared using electron beam physical vapor deposition (EB-PVD), using YSZ material, with a ceramic layer thickness of 200 μm.
[0062] S12, prepare CMAS powders with different component contents and coat them on the surface of the test piece;
[0063] Formulate CMAS powders with different chemical compositions, including CaO, MgO, , , , , , , The ten major elements MnO were used to obtain CMAS materials with different composition ranges by adjusting the molar ratio of each element. The prepared CMAS powder was dissolved in anhydrous ethanol, uniformly coated on the surface of the thermal barrier coating test piece, and placed in a high-temperature furnace for adhesion heat treatment to ensure that the CMAS is stably adhered to the coating surface.
[0064] The CMAS coating concentration was strictly controlled at 10 mg / cm³ during the experiment. 2 To simulate deposition under harsh conditions, CMAS-coated specimens were placed in a high-temperature furnace at 1200°C for heat treatment. A 5-minute timer was started when the furnace temperature returned to 1200°C to ensure stable CMAS adhesion to the coating surface.
[0065] This process effectively prevents the CMAS from detaching during the heat shock process and ensures the stability of the CMAS during the test.
[0066] S13, set different test temperatures to conduct high-temperature thermal shock tests on CMAS-coated samples;
[0067] The CMAS-coated sample was fixed on the gas burner test bench fixture, and multiple different test temperatures were set. During the test, the sample was subjected to high-temperature thermal shock by a spray gun flame, and repeated heating to the preset test temperature and quenching cycle were performed to simulate the CMAS deposition conditions under the service environment of an aero-engine. When the area of coating peeling reached or exceeded the preset threshold, the sample was deemed to have failed, and the number of thermal shock cycles at this time was recorded as the actual life of the sample.
[0068] In this embodiment, coating failure is defined as when the peeling area reaches or exceeds 20%. Test parameters are set on a high-temperature thermal shock test bench, with three test temperatures set at 1200℃, 1250℃, and 1300℃. A complete thermal shock cycle includes three stages: heating, holding, and cooling. The heating and holding times are controlled to be 380s, with a holding time of 300s and a rapid cooling time of 60s. Back-side compressed air cooling is initiated when the surface temperature of the coating reaches 800℃. Figure 4 This shows the actual working state of the spray gun during the high-temperature thermal shock test.
[0069] This embodiment includes 10 CMAS component contents × 3 temperature conditions, for a total of 30 different test conditions. To eliminate random errors, each test condition was repeated 5 times, and a total of 150 sets of valid test data were collected.
[0070] S14. Collect microstructure image data of failed samples, organize tabular data including CMAS component content, CMAS viscosity and test temperature, and construct a multimodal dataset using thermal cycling failure lifetime as a label.
[0071] The failed samples were sequentially subjected to cold mounting, cutting, grinding and polishing. Microstructure images of the coating cross-section were acquired using a scanning electron microscope, and representative microstructure (SEM) images that could reflect the corrosion state of the coating were selected as image data. Figure 5 , Figure 6 SEM images of the thermal barrier coating are shown and its characteristic morphology is labeled.
[0072] The CMAS major element content, CMAS viscosity, and thermal shock test temperature of each sample were compiled into tabular data. The tabular data were normalized, and the image data were subjected to resolution unification and data augmentation. The tabular data and image data of the same sample were associated with the corresponding thermal shock cycle number label to construct a multimodal dataset.
[0073] S2. Construct a thermal barrier coating lifetime prediction model. The model architecture includes a dual-tower feature extraction network, an attention-based feature fusion module, and a predictor based on a fully connected layer.
[0074] Step S2 specifically includes the following sub-steps:
[0075] S21, Construct a dual-tower feature extraction network to extract the table feature vector of the table data and the image feature vector of the image data respectively, wherein the two feature vectors have the same dimension;
[0076] In this embodiment, the table feature extraction tower (Tower A) employs a multilayer perceptron (MLP). Normalized table data is taken as input, and after processing through several fully connected layers, batch normalization, and ReLU activation function, the output table feature vector is determined. , here The feature embedding dimension.
[0077] The image feature extraction tower (Tower B) employs a residual network (ResNet). Taking a corrupted SEM image as input, it removes the fully connected layers of the original ResNet, retaining only the convolutional layers and the global average pooling layer, outputting an image feature vector. .
[0078] S22, Construct a feature fusion module based on attention mechanism, using table feature vectors as queries and image feature vectors as keys and values, and generate multimodal fusion feature vectors through attention mechanism;
[0079] Specifically, cross-attention is used to deeply fuse information from the two modalities.
[0080] Using table feature vectors As the query vector Q (Query), the image feature vector As the key vector K (Key) and value vector V (Value), they are defined as follows:
[0081] Q = K = V =
[0082] in, , The projection matrix is learnable;
[0083] Subsequently, the multimodal fusion feature vector is calculated. :
[0084]
[0085] In the formula, To query the similarity between keys, it represents the correlation between table feature vectors and image feature vectors; It is a scaling factor to prevent the gradient from vanishing due to excessively large values.
[0086] Through this mechanism, the model can automatically focus on the microstructural regions (such as crack propagation paths and penetration depths) in the SEM image that are most relevant to the CMAS composition and process parameters.
[0087] S23, Construct a predictor based on a fully connected layer to map the multimodal fusion feature vector into the coating lifetime prediction value;
[0088] In this embodiment, the predictor uses a regression MLP network, containing two fully connected layers, with no activation function in the output layer. Features are fused. Input predictor, output scalar value This refers to the predicted thermal barrier coating (TBC) lifetime.
[0089] S3, during the training phase, the dual-tower feature extraction network is optimized through contrastive learning to semantically align the table feature vector with the image feature vector in the latent space.
[0090] Construct positive and negative sample pairs:
[0091] 1) Positive sample pairs: tabular feature vectors of the same experimental sample With image feature vectors ;
[0092] 2) Negative sample pair: The tabular feature vector of any test sample in the current batch. Image feature vectors of other test samples To form negative sample pairs, where, .
[0093] The InfoNCE loss function is used, and the formula is as follows:
[0094]
[0095] in, The cosine similarity function is used. For temperature coefficient, The first in the batch The tabular feature vector of each sample, The first in the batch Image feature vectors of each sample The first in the batch Image feature vectors of each sample This is the batch size.
[0096] This step enables the model to learn the intrinsic physical relationship between CMAS composition / process parameters and the microstructure of the failed samples.
[0097] S4 uses a multimodal dataset and trains the parameters of the prediction model by calculating the mean squared error loss, ranking loss of the predictor and the contrastive loss of the feature extraction network, thus obtaining the trained prediction model.
[0098] To balance prediction accuracy, feature alignment, and lifetime ranking consistency, a total loss function is constructed, comprising mean squared error loss, ranking loss, and contrast loss, as follows:
[0099]
[0100] in, These are hyperparameter weights used to balance the importance of each loss term.
[0101] Mean square error loss Used to monitor the accuracy of lifetime predictions, defined as:
[0102]
[0103] In the formula, For actual lifespan, To predict lifetime values, This is the batch size.
[0104] Ranking loss To consider the monotonicity of the material failure process, a ranking loss is introduced to maintain consistency in lifetime ranking among samples. For any two samples... and If the actual lifespan Then the constrained predicted value satisfies ,in This is a preset boundary threshold. The specific form of the ranking loss can be defined as:
[0105]
[0106] S5, during the model deployment phase, only tabular data is input, the feature fusion module degenerates from cross-attention to self-attention mechanism, and the predictor outputs lifetime prediction values.
[0107] After the model completes training, it enters the deployment phase. During actual prediction, the system only needs to input tabular data, namely CMAS component parameters and process parameters, without the need to collect SEM images.
[0108] During the prediction process, the dual-tower feature extraction network only activates the table feature extraction tower, outputting the table feature vector. The feature fusion module maintains the original attention mechanism structure unchanged (i.e., it retains the original projection matrix parameters). , However, the input source is changed from cross-attention during the training phase to self-attention: in the form of tabular feature vectors. Simultaneously serving as a query ,key Sum Calculate fusion features .
[0109] Subsequently, this fusion feature The predictor performs forward calculations and outputs the lifetime prediction value. .
[0110] It is understandable that the cross-attention mechanism and the self-attention mechanism use the same set of learnable projection matrices. This ensures that the attention weight parameters used by the model remain consistent during the training and inference phases.
[0111] like Figure 7The model network architecture diagram is shown. This method achieves image-free inference. During the deployment and prediction phase, users only need to input tabular data such as CMAS major element content, CMAS viscosity, and process temperature. Without inputting SEM images, the model can output high-precision thermal barrier coating lifetime predictions through a self-attention smoothing degradation path.
[0112] In a preferred embodiment, before step S5, the following steps are also included:
[0113] Step S5a, verification of physical interpretability based on Grad-CAM;
[0114] During the laboratory validation phase, the Gradient Weighted Class Activation Mapping (Grad-CAM) technique was used to analyze whether the model learned the correct physical laws, specifically including:
[0115] S5a1, calculate the feature map contribution weights;
[0116] For the ResNet residual network used in the image feature extraction tower, the predicted lifetime is calculated with the model parameters fixed after training. The gradients of each feature map in the last convolutional layer of ResNet are calculated, and global average pooling is performed on the gradients to obtain the contribution weight of each feature map to the prediction result.
[0117] S5a2, generate Grad-CAM heat map;
[0118] The feature maps are weighted and summed according to their contribution weights, and then processed by the ReLU activation function to generate the Grad-CAM heatmap.
[0119] S5a3, overlaying visualization and mechanism verification;
[0120] The generated heatmap is overlaid on the original SEM image, highlighting the microstructural regions that contribute most to lifetime prediction. This allows for the intuitive location of key areas leading to coating lifetime degradation, including crack propagation paths and CMAS penetration layers. By comparing the degree of agreement between the highlighted areas and known failure mechanisms, the model's ability to learn correct failure patterns is verified.
[0121] Example 1
[0122] This embodiment demonstrates the specific process of predicting the TBC lifetime under high temperature conditions based on the method described in this disclosure for specific CMAS component content.
[0123] (1) The 150 sets of data were randomly divided into 80% training set (120 sets) and 20% test set (30 sets). During the division process, it was ensured that the test set covered all 10 CMAS types with different principal component contents to verify the generalization ability of the model.
[0124] (2) A prediction model is constructed based on the dual-tower structure proposed in this disclosure. A 3-layer MLP (input layer → 128 → 256 → 512) is used to build the table tower, the activation function is ReLU, and the output feature dimension is mapped to 512.
[0125] (3) ResNet-50 is used as the backbone network to build the image tower, and pre-trained weights based on the ImageNet dataset are loaded; the original fully connected layers of ResNet-50 are removed, the first 3 stages are frozen, and only the last 2 stages are fine-tuned, and the output feature dimension is mapped to 512.
[0126] (4) Set the temperature coefficient of InfoNCE loss =0.07; The cross-attention module uses single-head attention (Head=1).
[0127] (5) Use the Adam optimizer with an initial learning rate of 1×10⁻⁶. -4 The batch size is set to 16 (to accommodate small samples), and the number of training epochs is set to 200. The loss weights are set to... =1.0(MSE), =0.1(InfoNCE), =0.5 (Rank).
[0128] (6) After the joint training of the models is completed, single-modal inference evaluation without image input is performed on 30 test sets. Specifically, the input of the image tower (ResNet) is disconnected during the prediction stage, and only the CMAS components and process parameters (tabular data) in the test set are input into the table tower (MLP) to extract features. At this stage, the cross-attention module degenerates into self-attention computation (i.e., using...). Simultaneously, as Q, K, and V, fused features are generated and the predicted lifetime is output.
[0129] (7) Offline verification of the physical interpretability of the model. Take any test sample containing the corresponding SEM image, input it completely into the dual-tower network to calculate the gradient, and output the Grad-CAM heatmap. By comparing the highlighted areas with the actual failure physical observation results, determine whether the model has learned the correct CMAS corrosion mechanism.
[0130] This disclosure also proposes an electronic device, including: a memory for storing computer programs;
[0131] The processor is used to execute the program stored in the memory to implement the steps of the above embodiment of the coating lifetime prediction method based on multimodal contrast and attention fusion.
[0132] For details on the specific implementation of each step and related explanations, please refer to the aforementioned embodiment of a coating lifetime prediction method based on multimodal contrast and attention fusion, which will not be repeated here.
[0133] The memory of the electronic device mentioned in the embodiments of this disclosure may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.
[0134] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0135] This disclosure also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described embodiment of a coating lifetime prediction method based on multimodal contrast and attention fusion. For specific implementation details and explanations of each step, please refer to the foregoing method embodiments; further elaboration is not provided here.
[0136] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0137] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure.
Claims
1. A coating lifetime prediction method based on multimodal contrast and attention fusion, characterized in that, Includes the following steps: A multimodal dataset of thermal barrier coating samples was obtained. The multimodal dataset includes tabular data and microstructure image data after failure. The tabular data includes CMAS composition parameters and process parameters, and is labeled with thermal cycling failure lifetime. A lifetime prediction model for thermal barrier coatings was constructed. The model architecture includes a dual-tower feature extraction network, a feature fusion module based on an attention mechanism, and a predictor based on a fully connected layer. During the training phase, the dual-tower feature extraction network is optimized through contrastive learning to improve the table feature vectors. With image feature vectors Semantic alignment in the latent space; Using a multimodal dataset, the parameters of the prediction model are trained by calculating the mean squared error loss, ranking loss of the predictor, and contrastive loss of the feature extraction network, thus obtaining the trained prediction model. During the model deployment phase, only tabular data is input, the feature fusion module degenerates from cross-attention to a self-attention mechanism, and the predictor outputs lifetime predictions.
2. The method according to claim 1, characterized in that, The construction of the thermal barrier coating lifetime prediction model includes: Construct a dual-tower feature extraction network to extract table feature vectors from the tabular data. Image feature vectors of image data In this case, the two feature vectors have the same dimension; Construct an attention-based feature fusion module to use table feature vectors. As a query Image feature vectors as keys Sum Multimodal fusion feature vectors are generated through an attention mechanism. ; Construct a predictor based on fully connected layers to fuse multimodal feature vectors. This is mapped to a predicted coating lifetime.
3. The method according to claim 1, characterized in that, During the model deployment phase, only tabular data is input, the feature fusion module degenerates from cross-attention to a self-attention mechanism, and the predictor outputs lifetime prediction values, including: During the model deployment phase, only tabular data is input, and the input source is switched from cross-attention during the training phase to self-attention, using tabular feature vectors. Simultaneously serving as a query ,key Sum Calculate the fused feature vector The predictor outputs the lifetime prediction value.
4. The method according to claim 1, characterized in that, During the training phase, the dual-tower feature extraction network is optimized through contrastive learning to improve the table feature vectors. With image feature vectors Semantic alignment in the latent space includes: Construct positive and negative sample pairs: Positive sample pairs: tabular feature vectors of the same experimental sample With image feature vectors ; Negative sample pair: The tabular feature vector of any test sample in the current batch. Image feature vectors corresponding to other test samples To form a negative sample pair, where ; The InfoNCE loss function is used, and the formula is as follows: ； in, The cosine similarity function is used. For temperature coefficient, The first in the batch The tabular feature vector of each sample, The first in the batch Image feature vectors of each sample The first in the batch Image feature vectors of each sample This refers to the batch size.
5. The method according to claim 4, characterized in that, The process involves using a multimodal dataset and training the prediction model's parameters by calculating the mean squared error loss, ranking loss, and contrastive loss of the feature extraction network, resulting in a trained prediction model. This includes: The total loss function, which includes mean squared error loss, ranking loss, and comparison loss, is constructed as follows: ； in, These are hyperparameter weights used to balance the importance of each loss term. Mean square error loss Used to monitor the accuracy of lifetime predictions, defined as: ； In the formula, For actual lifespan, To predict lifetime values, Batch size; Ranking loss To maintain consistency in lifetime ranking among samples, for any two samples and If the actual lifespan Then the constrained predicted value satisfies ,in The ranking loss is defined as follows, where the preset boundary threshold is used: 。 6. The method according to claim 1, characterized in that, The multimodal dataset for obtaining thermal barrier coating samples includes: Preparation of thermal barrier coating specimens; CMAS powders with different component contents were prepared and coated onto the surface of the test piece; High-temperature thermal shock tests were conducted on CMAS-coated samples at different test temperatures. Microstructural image data of failed samples were collected, tabular data including CMAS component content, CMAS viscosity and test temperature were compiled, and a multimodal dataset was constructed using thermal cycling failure lifetime as a label.
7. The method according to claim 1, characterized in that, Prior to the model deployment phase, the following is also included: The interpretability of the model was verified using Grad-CAM (Grad-Weighted Class Activation Mapping).