Nickel-based single crystal alloy creep property prediction system based on AE-CNN-PINN

Abstract

The invention discloses a nickel-based single crystal alloy creep property prediction system based on AE-CNN-PINN, and belongs to the technical field of alloy creep prediction. In the prediction system, through combination of multi-modal data fusion and physical constraint, acoustic emission (AE) signal real-time feature extraction, microscopic structure depth feature analysis of a convolutional neural network (CNN) architecture and physical consistency optimization of a physical information neural network (PINN) model are utilized, so that the prediction accuracy is improved. And an efficient intelligent solution is provided for accurate prediction of the creep life and the creep behavior of the nickel-based single crystal alloy.

Description

Technical Field

[0001] This invention relates to the field of creep prediction technology for nickel-based single-crystal alloys, specifically to a system and method for predicting the creep life of nickel-based single-crystal alloys based on acoustic emission (AE), convolutional neural network (CNN), and physical information neural network (PINN). Background Technology

[0002] Nickel-based single-crystal alloys are widely used in critical components in aerospace, energy, and industrial fields due to their excellent high-temperature strength and creep resistance. However, these components undergo complex creep processes during long-term service under high-temperature and high-stress conditions, leading to microstructure evolution, dislocation accumulation, and eventual failure. Therefore, accurately predicting the creep properties of nickel-based single-crystal alloys (including creep life, creep stage division, and strain rate curves) is of great significance for assessing their service safety.

[0003] Currently, creep performance prediction mainly relies on the following methods:

[0004] (1) Empirical model: Empirical formulas are constructed based on experimental data, such as the Norton creep equation. This type of method relies on a large amount of experimental data, has a limited scope of application, and is difficult to consider the complex interactions of multimodal data.

[0005] (2) Numerical simulation: The finite element method is used to simulate creep behavior, but the computational complexity is high and it is difficult to handle large-scale multimodal data.

[0006] (3) Machine learning models: In recent years, machine learning has been widely used in the field of material property prediction. However, traditional machine learning methods are difficult to effectively integrate microstructure characteristics and creep signal dynamics, and lack the support of physical constraints, resulting in insufficient model interpretability and prediction accuracy.

[0007] The creep behavior of nickel-based single-crystal alloys is influenced by a variety of factors, including microstructure characteristics (such as...). The characteristics of microstructure and macroscopic properties, as well as the high dynamics of acoustic emission signals, pose higher requirements for creep performance prediction. These include phase area ratio, dislocation density, creep conditions (such as temperature and stress), and acoustic emission signal characteristics (such as event rate and energy).

[0008] Therefore, combining material microstructure characteristics with acoustic emission dynamic signals, utilizing physical laws to enhance model consistency, and optimizing prediction performance through efficient deep learning methods have become important directions in creep performance prediction research. Against this backdrop, this invention proposes to construct an intelligent creep performance prediction system based on AE-CNN-PINN by fusing AE signal features and microstructure characteristics, and introducing a Physical Information Neural Network (PINN). This provides a new approach for the creep behavior analysis and performance evaluation of nickel-based single-crystal alloys. Summary of the Invention

[0009] One objective of this invention is to provide a creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN, such as... Figure 1 The creep performance prediction system shown consists of a multimodal creep data preprocessing module (100), a CNN-based microstructure feature extraction module (200), an AE-CNN multimodal data fusion module (300), a creep physical constraint-based feature generation module (400), a PINN-based model training module (500), and an AE signal-based self-learning optimization module (600). The creep performance prediction system of this invention can accurately predict the creep life of nickel-based single-crystal alloys under different operating conditions.

[0010] The multimodal creep data preprocessing module (100) is used to perform normalization preprocessing on nickel-based single-crystal alloy sample data to obtain preprocessed numerical information (denoted as F100), which is then output to the CNN-based microstructure feature extraction unit (200) and the AE-CNN multimodal data fusion module (300). This module serves as the data entry and standardization unit, aligning, cleaning, and normalizing the input raw creep experimental data (time-series data) and its strictly corresponding microstructure image data (spatial data) to generate preprocessed numerical information F100 with a uniform format and standardized scale, providing a consistent input basis for all subsequent analysis modules.

[0011] The CNN-based microstructure feature extraction module (200) inputs the F100 information into a convolutional neural network (CNN) architecture for fully connected convolution-pooling operations to extract creep feature information (denoted as F200) of the nickel-based single-crystal alloy sample, and outputs it to the AE-CNN multimodal data fusion module (300). This module focuses on mining deep physical information in microstructure images. It receives image data from F100 and performs automated feature learning through a carefully designed convolutional neural network (CNN). This CNN abstracts microstructure features related to creep damage (such as F200) from the original pixels layer by layer through multiple layers of "convolution-activation-pooling" operations. (Phase size, generalization direction, pore density and distribution, etc.), ultimately outputting a highly condensed creep characteristic information vector F200 that characterizes the microstructure.

[0012] The AE-CNN multimodal data fusion module (300) fuses the F100 information with the F200 information to generate multimodal fusion information (denoted as F300) characterizing creep behavior, and outputs it to the feature generation module (400) based on creep physical constraints. This module is key to achieving information complementarity. It receives macroscopic experimental data features (the numerical part of F100) from module 100 and microscopic organizational features (F200) from module 200. Through deep fusion using feature concatenation or cross-modal attention mechanisms, it generates multimodal fusion information F300 that can simultaneously reflect the synergistic effect of macroscopic conditions and microscopic states.

[0013] The feature generation module (400) based on creep physical constraints generates global physical feature information (denoted as F400) of creep behavior by introducing the Norton creep equation into the F300 information, and outputs it to the PINN-based model training unit (500). It introduces classical creep physical equations (such as the Norton equation) as strong constraints based on the multimodal fusion information F300. By deriving new feature variables (such as theoretical creep rate, stress exponent related terms, etc.) from the physical equations, and combining these physical guiding features with F300, it generates global physical feature information F400 that contains both data information and physical laws.

[0014] The PINN-based model training module (500) inputs the F300 and F400 information into the Physical Information Neural Network (PINN) architecture for training, obtaining the training result information (denoted as F500), and outputting it to the self-learning optimization module (600) based on the AE signal. This module achieves the organic unity of data and physics. It constructs a Physical Information Neural Network (PINN), whose inputs are the fused information F300 and the physical features F400. The unique feature of PINN is its loss function design: it not only includes a "data loss" that measures the deviation between the predicted and experimental values, but also adds a "physical loss" term, which is obtained by substituting the predicted output of PINN into the creep physics equation to calculate the residual. By jointly optimizing these two losses, the trained model (outputting the training result information F500) not only fits the data but also adheres to the preset physical laws, significantly improving the reliability of the model.

[0015] The self-learning optimization module (600) based on the AE signal uses the F500 information to perform self-learning optimization on the test sample information F700, obtaining the creep performance prediction information of nickel-based single crystal alloy (denoted as F600). This module endows the system with dynamic evolution capabilities. After the system is deployed, when faced with a new test sample information F700, it first uses the trained PINN model (F500) for initial prediction. If the reconstruction error of the acoustic emission signal AE exceeds the threshold, it indicates that the current sample may exceed the original empirical range of the model. The system then triggers a lightweight online fine-tuning process, using the small amount of initial data that may already exist for the sample to quickly optimize the PINN model, and finally outputs a more accurate and personalized creep performance prediction information F600 for the current sample.

[0016] In this invention, a computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor; the processor executes the computer program to implement the operation of a creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN; and the steps of implementing a creep performance prediction method for nickel-based single-crystal alloys based on AE-CNN-PINN. The computer program is implemented using Python programming.

[0017] In this invention, a computer-readable storage medium stores a computer program thereon; when the computer program is executed by a processor, it implements the operation of a creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN; and implements the steps of a creep performance prediction method for nickel-based single-crystal alloys based on AE-CNN-PINN.

[0018] In this invention, a computer program product includes a computer program; when executed by a processor, the computer program implements the operation of a creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN; and implements the steps of a creep performance prediction method for nickel-based single-crystal alloys based on AE-CNN-PINN.

[0019] The second objective of this invention is to propose a method for predicting the creep performance of nickel-based single-crystal alloys based on AE-CNN-PINN. The input information used in the prediction process includes acoustic emission (AE) signals and microstructure characteristics of nickel-based single-crystal alloy samples. By fusing multi-source nickel-based single-crystal alloy sample information with an architecture based on convolutional neural networks (CNN) and physical information neural networks (PINN), accurate prediction of the creep performance of nickel-based single-crystal alloys is achieved. This invention overcomes the limitations of traditional empirical models and purely data-driven models in creep behavior prediction, and features strong physical consistency, high prediction accuracy, and good model interpretability.

[0020] (A) Compared with the prior art, the creep performance prediction system and method of the present invention can solve the following problems existing in the prior art:

[0021] (1) Difficulty in multimodal data fusion: Existing methods are difficult to effectively integrate the dynamic characteristics of AE signals with the static properties of microscopic tissue features, resulting in limited model prediction performance.

[0022] (2) Lack of physical constraints: Traditional machine learning methods rely solely on experimental data for training and lack support for the physical laws of materials (such as Norton's creep equation), which affects the credibility of the model.

[0023] (3) Inaccurate division of creep stages: Current prediction models have difficulty in accurately distinguishing the initial stage, steady-state stage and acceleration stage of creep, resulting in large deviations in lifetime prediction results.

[0024] (B) Compared with the prior art, the creep performance prediction system and method of the present invention have the following innovative features:

[0025] (a) Multimodal data fusion:

[0026] i) Extract dynamic characteristics of the creep process (such as energy and event frequency) based on AE signals;

[0027] ii) Using CNN to extract microscopic tissue features (e.g. Phase area ratio, dislocation density);

[0028] iii) Through the multimodal data fusion module (AE and CNN), PCA and Gini weight optimization are used to achieve global correlation modeling of creep behavior features.

[0029] Multimodal data fusion and feature extraction of nickel-based single-crystal alloys has always been a technical challenge in traditional methods. This invention incorporates the spatiotemporal correlation between AE signals and microstructural features into creep performance modeling, establishing a multimodal data input framework that integrates dynamic signals (such as energy and frequency) with microstructural features (such as... Correlation analysis of phase distribution and dislocation density comprehensively characterizes the creep behavior of materials. By fusing AE signal features and microstructure features extracted by CNN, combined with the physical feature matrix, the system can output high-precision creep life prediction values.

[0030] (b) Physical Information Neural Network (PINN):

[0031] (i) Introduce the Norton creep equation as a physical constraint term and embed physical laws into the deep learning model;

[0032] (ii) Construct the total loss function ,and , For the data-driven loss term in the CNN architecture, This is the physical consistency loss term in the PINN architecture.

[0033] In this invention, on the one hand, the reliability and physical consistency of model predictions are improved by introducing physical driving constraints, thus compensating for the shortcomings of purely data-driven models. On the other hand, the Norton creep equation is introduced as a physical constraint, and combined with the PINN model, the model prediction results are corrected through a physical consistency loss term to ensure that the predicted values ​​are consistent with the material creep law, thereby significantly improving the reliability of the model in practical engineering applications.

[0034] (c) Model performance optimization and feedback mechanism:

[0035] (i) Based on AE information, assign weight coefficients to the features of conventional descriptors to further improve the prediction accuracy of the model;

[0036] (ii) Supports GPU and Dask parallel computing frameworks, optimizes computing efficiency, and is suitable for large-scale material data;

[0037] iii) Supports dynamic optimization and large-scale parallel computing, significantly improving prediction efficiency and applicability, and is suitable for full-cycle analysis of complex creep behavior. Attached Figure Description

[0038] Figure 1 This is a structural block diagram of the nickel-based single-crystal alloy creep performance prediction system based on AE-CNN-PINN of the present invention.

[0039] Figure 2 This is a comparison chart of the prediction results using the creep performance prediction method of this invention and the traditional prediction results. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings. The examples of the parameters listed are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0041] Multimodal creep data preprocessing module (100)

[0042] The multimodal creep data preprocessing module (100) of this invention is used to perform normalization preprocessing on nickel-based single-crystal alloy sample data to ensure the consistency and high quality of the input data. The normalization preprocessing employs any one of the Min-Max, Z-Score, and Sigmoid methods. All attributes of the acoustic emission signal (AE) are represented in the same unit of measurement and a common scale or range is used. Normalization attempts to assign equal weight to all data attributes, making comparison and aggregation between attributes easier and improving data convergence. It prevents attribute values ​​from dominating other attribute values. Furthermore, normalization helps prevent distorted results from machine learning algorithms that use distance metrics between attributes and improves the efficiency of data analysis. In this invention, the nickel-based single-crystal alloy sample data includes, but is not limited to, acoustic emission signal (AE) and microstructure image data of nickel-based single-crystal alloys disclosed in books, journals, conferences, academic papers, patents, etc.

[0043] (1) Acoustic emission signal preprocessing;

[0044] Parameters such as energy, event frequency, amplitude, rise time, and peak frequency are extracted from the acoustic emission (AE) signals of nickel-based single-crystal alloys to generate an acoustic emission signal training dataset, denoted as . The wavelet transform method is used to analyze the... Feature extraction in the time and frequency domains yields the time-frequency feature matrix, denoted as... The above Random number forest and decision tree models using Min-Max data normalization techniques are normalized to... The preprocessed acoustic emission signal is obtained from the interval, denoted as . .

[0045] (2) Preprocessing of microscopic tissue image data;

[0046]

[0047] (3) Multimodal data warehouse;

[0048] The With the Data is stored in the order of the numbering of nickel-based single-crystal alloy samples to obtain multimodal feature information. .

[0049] Through the processing of the multimodal creep data preprocessing module (100), the original data (nickel-based single-crystal alloy sample data) is standardized, denoised, and feature-extracted, ensuring the consistency, high quality, and usability of the multimodal data, providing reliable input data for subsequent modules (such as the CNN feature extraction unit and the multimodal fusion module). In summary, the multimodal creep data preprocessing unit (100) is used to perform normalized preprocessing on the acoustic emission signal (AE) and microstructure image data of the nickel-based single-crystal alloy samples, obtaining preprocessed numerical information (denoted as F100), and outputting it to the CNN-based microstructure feature extraction unit (200) and the AE-CNN multimodal data fusion module (300).

[0050]

[0051]

[0052] (1) Feature extraction of multi-scale convolution kernels;

[0053] First, use convolution kernels of different sizes. (like or or Multi-scale feature extraction was performed on microscopic tissue images, and... Convolution operation to capture Microscopic features such as phases and grain boundaries; then, multi-scale features are weighted. The microstructure feature matrix is ​​then fused and generated. Applications. The weighting coefficients are dynamically adjusted based on the training data until convergence is achieved.

[0054] (2) Enhanced spatial-temporal attention mechanisms;

[0055] First, spatial attention mechanism enhancement is performed by applying the Sigmoid activation function to spatially weight the microscopic tissue feature matrix, thereby strengthening... Phase and grain boundary characteristic regions; secondly, fully connected layer operations are used to enhance the time attention mechanism, introducing time weights into dynamic microstructure characteristics and strengthening the characteristic change patterns; finally, spatial attention features and time attention features are weighted. The data are then fused to generate the final microstructure depth feature matrix information, namely the creep feature information F200 of the nickel-based single crystal alloy sample. Spatial attention features These are the weighting coefficients for spatial attention features. For time attention features, These are the weighting coefficients for the time attention feature. These are the attention features after fusion.

[0056] In this invention, the CNN-based microscopic tissue feature extraction module (200) can achieve multi-scale feature extraction, that is, simultaneously capture the microscopic information of large-size and small-size feature regions, ensuring... Precise representation of features such as phases, grain boundaries, and dislocations. Simultaneously, attention mechanisms are utilized to enhance the extraction capability of key regions and dynamic features; the spatial-temporal joint attention mechanism improves this capability. In summary, the CNN-based microstructure feature extraction unit (200) inputs the F100 information into a convolutional neural network (CNN) architecture for convolution-pooling operations, extracting the creep feature information (denoted as F200) of the nickel-based single-crystal alloy sample, and outputting it to the AE-CNN multimodal data fusion module (300).

[0057] AE-CNN Multimodal Data Fusion Module (300)

[0058] The AE-CNN multimodal data fusion module (300) of this invention is the core module for achieving deep fusion of acoustic emission signal features and microscopic tissue features. It can jointly analyze the AE feature matrix and the CNN microscopic tissue feature matrix to generate a multimodal fusion feature matrix characterizing creep behavior, providing input for subsequent physical feature generation and creep performance prediction. Through principal component analysis (PCA), Gini weight analysis, and dynamic weighted optimization methods, this module fully explores the correlation between multimodal features.

[0059] The AE-CNN multimodal data fusion module (300) is used to fuse the F100 information with the F200 information to generate multimodal fusion information (denoted as F300) representing creep behavior, and output it to the feature generation module (400) based on creep physical constraints.

[0060] (1) Feature fusion and dimensionality reduction;

[0061] First, time-dimensional alignment is performed, aligning AE features with microscopic tissue features according to the sampling time series to ensure that AE data and microscopic tissue data at the same time point correspond. Second, data normalization is performed, normalizing the AE and CNN feature matrices to the same scale. Finally, principal component analysis (PCA) is used to reduce the dimensionality of the weighted AE and CNN feature matrices and extract feature components.

[0062] AE features normalized to , This is the currently sampled AE information. For the minimum AE information in a sampling period, For the maximum AE information in a sampling period, This is the normalized AE information.

[0063] Microstructure characteristics normalized to , This is the microscopic tissue feature information currently being sampled. This refers to the minimum microscopic tissue feature information within a sampling period. This represents the maximum microscopic tissue feature information within a sampling period. This represents the normalized microstructural characteristics.

[0064] (2) Input parameter feature weight output;

[0065] To enhance the model's adaptability in the absence of AE data, this invention introduces conventional descriptive features of nickel-based single-crystal alloys (such as alloy composition, temperature, stress, etc.) as a third type of input information (denoted as...). ), the It is incorporated into the multimodal feature fusion process. For those using... Formal normalization The subsequent alloy composition-temperature-stress correlation information is denoted as .

[0066] Regarding the After performing Gini weight analysis, the weight vector is obtained. Using the aforementioned With the Construct the weighted descriptor feature matrix as follows , This is a point-by-point multiplication operation.

[0067] The AE-CNN multimodal data fusion module (300) of this invention is used to achieve deep fusion of multimodal features, which can effectively combine dynamic AE signal features and static microscopic tissue features to improve the comprehensiveness of feature representation. It can utilize Gini weight analysis and dynamic weighted optimization to improve the adaptability of feature fusion. Furthermore, it reduces feature redundancy and computational complexity through PCA, providing support for large-scale data processing.

[0068] Feature generation module based on creep physical constraints (400)

[0069] The feature generation module (400) based on creep physical constraints of this invention is responsible for combining the multimodal fusion feature matrix with the physical constraints of material creep. By introducing the Norton creep equation and other creep physical models, it generates a global physical feature matrix of creep behavior, providing physically enhanced input features for subsequent PINN model training and creep performance prediction. This module generates creep stage features based on physical laws, making up for the shortcomings of purely data-driven models in terms of physical consistency and prediction reliability.

[0070] According to Norton's creep equation, the creep process is divided into three stages: the initial stage, the steady-state stage, and the acceleration stage.

[0071] The multimodal fusion information F300 of creep behavior is input into the Norton creep equation to obtain the characteristic values ​​of the creep rate calculation stage. .

[0072] The experimental conditions for setting up nickel-based single-crystal alloys include stress. ,temperature Sampling time Data on creep conditions.

[0073] Combining experimental operating parameters and the F300 information, creep physical characteristics are generated, including the generation of initial stage creep characteristics, steady-state stage creep characteristics, and accelerated stage creep characteristics.

[0074] (1) Initial stage creep characteristics are generated;

[0075] The strain growth characteristics in the initial stage are expressed as follows: , Indicates the duration of the initial phase. Based on... The generated initial stage feature matrix is .

[0076] (2) Creep characteristics are generated in the steady-state stage;

[0077] The strain rate in the steady state is expressed as The Satisfying Norton's creep equation, based on The steady-state phase feature matrix is ​​generated as follows .

[0078] (3) Creep characteristics are generated during the accelerated phase;

[0079] The strain growth rate during the acceleration phase is expressed as , represents the creep rate in Norton's creep equation. This refers to the sampling time point. Based on... The generated creep physical feature matrix is .

[0080] The combined physical characteristics of creep in these three stages are as follows .

[0081] In this invention, the feature generation module (400) based on creep physical constraints can combine multimodal fusion feature matrices and experimental condition parameters, providing physically enhanced inputs for subsequent PINN-based model training. Embedding material creep laws such as the Norton equation into the feature generation process improves the physical consistency of model predictions and increases the accuracy of creep stage segmentation. The feature generation formula dynamically adapts to different experimental conditions, supporting diverse creep behavior modeling. In summary, the feature generation module (400) based on creep physical constraints generates global physical feature information of creep behavior (denoted as F400) by introducing the Norton creep equation into the F300 information, and outputs it to the PINN-based model training unit (500).

[0082] PINN-based model training module (500)

[0083]

[0084] The PINN-based model training module (500) inputs the F300 and F400 information into the Physical Information Neural Network (PINN) architecture for training, obtains the training result information (denoted as F500), and outputs it to the self-learning optimization module (600) based on the AE signal.

[0085] (1) Network structure design;

[0086] First, the F300 information and the F400 information are introduced into the input layer.

[0087] If acoustic emission features are missing, the PINN-based model training unit (500) will automatically ignore them, and effective prediction can still be achieved through the F400 information.

[0088] Secondly, the hidden layer contains multiple fully connected layers with ReLU activation function.

[0089] Finally, the output layer's model output is the predicted creep life value. .

[0090] (2) Loss function design;

[0091] Data-driven loss term It can be used to minimize the error between model predictions and experimental data.

[0092] Physical consistency loss term It can be used to ensure that the model prediction results conform to the reference values ​​of the Norton creep equation.

[0093] Total loss function This is the weighted sum of the data-driven loss term and the physical consistency loss term, i.e. ,in, For data-driven loss terms The weighting coefficients, Physical consistency loss term The weighting coefficients are dynamically adjusted through experiments.

[0094] To ensure the appropriate weighting of each modality during training, the model simultaneously optimizes the fusion weights α and β during error backpropagation. When AE data is missing, the system automatically sets α to 0 to ensure the model still has predictive ability.

[0095] (3) Model training and optimization;

[0096] First, the Adam optimizer is used to iteratively update the parameters of the PINN model. Second, hyperparameters (such as learning rate and number of hidden layer nodes) are optimized using a grid search method. Here, training is stopped when the validation set loss does not decrease significantly in 3-5 consecutive iterations.

[0097] (4) Parallel computing acceleration;

[0098] Using CUDA to accelerate the PINN model on GPUs significantly improves training efficiency on large-scale data. Simultaneously, the Dask framework distributes the model training task across multiple computing nodes for parallel execution, reducing training time.

[0099] The Compute Unified Device Architecture (CUDA) and OpenComputing Language (OpenCL) GPUs were used, leveraging general GPU optimization methods, including memory access optimization, computation optimization, and data localization optimization. NVIDIA GPUs were employed in the actual model testing.

[0100] (5) Output information F500;

[0101] The output is the training result information F500, which includes the creep life prediction value and creep stage division, etc.

[0102] In this invention, the PINN-based model training module (500) combines data-driven approaches with physical principles. The PINN model incorporates physical constraints such as the Norton equation to ensure the physical consistency of the prediction results. Simultaneously, the total loss function considers both data accuracy and physical principles, improving the model's predictive reliability. Furthermore, training is accelerated using GPUs and distributed computing, making it suitable for large-scale, complex data. The PINN model trained using this module achieves high-precision prediction and comprehensive analysis of creep performance, providing crucial support for material service life assessment.

[0103] Self-learning optimization module based on AE signal (600)

[0104] The self-learning optimization module (600) based on AE signals of this invention is a crucial component in dynamically optimizing the AE-CNN-PINN creep performance prediction model. By real-time acquisition and processing of sample AE signals, an adaptive feedback mechanism is constructed to dynamically adjust model feature weights and parameters, thereby improving the accuracy, robustness, and adaptability of the prediction model. The core of the module lies in comprehensively evaluating the real-time AE signals and prediction results, and achieving real-time correction through a self-learning optimization strategy.

[0105] The reinforcement learning A3C algorithm is used to define the state, action, and reward functions, i.e., the state-action mapping method: the current state is the current AE signal features and model error; the action is adjusting feature weights, learning rate, and hidden layer nodes; the reward is a negative value based on prediction error and physical consistency loss. The parameters are iteratively optimized in the Adam optimizer to dynamically update predictions. The optimized features are then re-input into the AE-CNN-PINN model to iteratively update creep lifetime, stage division, and strain rate predictions.

[0106] In this invention, the self-learning optimization module (600) based on AE signals, combined with the real-time acquisition and feedback mechanism of AE signals, enables dynamic optimization of model parameters. Simultaneously, by utilizing reinforcement learning algorithms, an intelligent mechanism for adjusting AE signal weights is constructed, thereby achieving high-precision, multi-dimensional output of creep life, stage division, and strain rate, providing a reliable basis for material life prediction. Through the self-learning optimization strategy of this module, the characteristics of AE signals are fully utilized, providing higher accuracy and engineering practicality for creep performance prediction.

[0107] The method for predicting the creep performance of nickel-based single-crystal alloys using the AE-CNN-PINN-based nickel-based single-crystal alloy creep performance prediction system of this invention includes the following prediction steps:

[0108] Step 1: Obtain sample data;

[0109] The experimental data on the creep performance of the sample to be tested are input into the constructed creep performance prediction model of nickel-based single crystal alloy based on AE-CNN-PINN.

[0110] The experimental operating parameters (stress, temperature, time, etc.) of the sample to be tested are input into the model for predicting the creep performance of nickel-based single crystal alloys based on AE-CNN-PINN.

[0111] The acoustic emission signal characteristics, microstructure images, conventional descriptors (such as alloy composition, load level, etc.), and corresponding creep life and stage division data of the sample to be tested are input into the AE-CNN-PINN-based nickel-based single crystal alloy creep performance prediction model.

[0112] Step two: Data standardization and multimodal feature fusion;

[0113] The AE-CNN multimodal data fusion module (300) normalizes the acquired multi-source data, normalizing the AE and CNN feature matrices to the same scale; then, the principal component analysis (PCA) method is used to reduce the dimensionality of the weighted AE and CNN feature matrices and extract feature components.

[0114] Step 3: Generation of creep physical characteristics;

[0115] According to Norton's creep equation, the creep process is divided into three stages: the initial stage, the steady-state stage, and the acceleration stage.

[0116] Step 4: Output and Verification of Results;

[0117] The output prediction result is the predicted creep life value. Simultaneously, the mean squared error (MSE) and... (Called the coefficient of determination) is used to evaluate the accuracy of the model's prediction results.

[0118] The final output is an optimized prediction result, with weight coefficients dynamically adjusted according to deployment requirements to ensure model compatibility with scenarios lacking Advanced Effects (AEs), thus improving practicality and robustness. Evaluation metric: Re 2 Refer to "Python Data Analysis and Machine Learning for Beginners", published by Tsinghua University Press, Beijing, October 2021, author Wang Guoping, 10.2.5 R-squared and Case Studies, pp. 227-229. Example

[0119] Taking the creep test data of a certain model (DD6-x) nickel-based single crystal alloy as an example, the creep prediction system of this invention is embedded in the control unit (a computer terminal) of a universal testing machine. The data resources of the control unit are read. Creep performance is predicted using the creep prediction system of this invention. The training data includes: (a) 30 sets of small sample data, mainly used for acoustic emission signal feature analysis; (b) 100 sets of large sample data, mainly used for PINN model training. Full-cycle creep behavior prediction is performed using the AE-CNN-PINN creep performance prediction system, and the output result is the predicted creep life value. The trend between the predicted result and the true value is significantly improved, R... 2 The statistic increased from 0.35 to 0.88, such as Figure 2As shown, the efficiency and reliability of the method of the present invention in creep performance prediction are verified.

[0120] The creep performance prediction system for nickel-based single crystal alloys based on AE-CNN-PINN of the present invention combines the advantages of acoustic emission signal analysis, microstructure feature extraction and physical driving neural network, and can achieve high-precision prediction of creep life, which has important engineering application value.

Claims

1. A creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN, characterized in that... Including: The multimodal creep data preprocessing module (100) is used to perform normalization preprocessing on nickel-based single crystal alloy sample data to obtain preprocessed numerical information F100, and output it to the CNN-based microstructure feature extraction unit (200) and the AE-CNN multimodal data fusion module (300). The microstructure feature extraction module (200) based on CNN inputs the F100 information into the convolutional neural network (CNN) architecture for fully connected convolution-pooling operations to extract the creep feature information (denoted as F200) of the nickel-based single crystal alloy sample, and outputs it to the AE-CNN multimodal data fusion module (300). The AE-CNN multimodal data fusion module (300) is used to fuse the F100 information with the F200 information to generate multimodal fusion information F300 representing creep behavior, and output it to the feature generation module (400) based on creep physical constraints. The feature generation module (400) based on creep physical constraints generates global physical feature information F400 of creep behavior by introducing the Norton creep equation into the F300 information, and outputs it to the PINN-based model training unit (500). The PINN-based model training module (500) inputs the F300 information and the F400 information into the Physical Information Neural Network (PINN) architecture for training, obtains the training result information F500, and outputs it to the self-learning optimization module (600) based on the AE signal. The self-learning optimization module (600) based on the AE signal uses the F500 information to perform self-learning optimization on the test sample information F700, and obtains the creep performance prediction information F600 of nickel-based single crystal alloy.

2. The creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN according to claim 1, characterized in that: A convolutional neural network (CNN) architecture is introduced into the CNN-based microscopic tissue feature extraction module (200), and convolution operations are used as... Multi-scale feature extraction was performed on microscopic tissue images to capture Microscopic features such as phases and grain boundaries are analyzed, and then multi-scale features are weighted. The data is then fused to generate a preliminary matrix of microscopic tissue characteristics. For the first Layer convolution features This is the identifier for the convolutional layer. For neurons The Layer convolution features This refers to the horizontal position in the image. This refers to the vertical position in the image. For convolution kernel weights, For bias, For neurons Input pixels, For neurons pixels, For neurons Pixels; This is the feature of the third convolutional layer. The weights of the first convolution kernel are... This is the feature of the 5th convolutional layer. The weights of the second convolution kernel. The features are from the 7th convolutional layer. The weights of the third convolution kernel. These are the fused convolutional features.

3. The creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN according to claim 1, characterized in that: The Norton creep equation is introduced in the feature generation module (400) based on creep physical constraints to divide the creep process into three stages: the initial stage, the steady-state stage and the acceleration stage.

4. The creep performance prediction system for nickel-based single-crystal alloys based on AE-CNN-PINN according to claim 1, characterized in that: The Physical Information Neural Network (PINN) architecture is introduced into the PINN-based model training module (500), with fully connected hidden layers and ReLU activation function.

5. A computer device, comprising: Memory, processor, and computer programs stored in memory and capable of running on the processor; The processor executes the computer program to implement the operation of the nickel-based single-crystal alloy creep performance prediction system based on AE-CNN-PINN according to any one of claims 1-4; and to implement the method for predicting the creep performance of nickel-based single-crystal alloys based on AE-CNN-PINN.

Images