An intelligent representation and life assessment method for structural surface defects of an aero-engine

Abstract

The application discloses an aero-engine structure surface defect intelligent characterization and life evaluation method, and relates to the technical field of aero-engine structure health management.The method fuses and collects defect data through multi-source sensing technology and constructs a knowledge graph; utilizes a multi-branch deep learning model to realize automatic identification of defect types and intelligent extraction of key geometric features; determines standard defect sizes based on a statistical distribution model of group defect data; quickly calculates defect stress concentration coefficients through a trained AI agent model, realizes mechanical equivalence of irregular defects; and finally implants equivalent defects into a component digital twin model to dynamically simulate and predict the evolution path and residual life thereof.The application solves the technical problems of experience-dependent characterization, isolated analysis links and static conservative evaluation, realizes intelligent processing of individualized life prediction from precise digital characterization of engine surface defects, and provides support for on-condition maintenance and predictive health management of engines.

Description

Technical Field

[0001] This application relates to the field of aero-engine structural health management technology, and in particular to a method for intelligent characterization and life assessment of surface defects in aero-engine structures. Background Technology

[0002] During manufacturing, assembly, and service, critical components of aero-engines, such as blades, disks, and casings, are prone to developing minute defects like scratches, pits, and corrosion due to factors such as tool impacts, foreign object impacts, and corrosive environments. These defects are potential sources of fatigue crack initiation, seriously threatening the engine's structural safety and service life. Therefore, accurate characterization and evaluation are essential in structural integrity analysis.

[0003] In current engineering practice, defect characterization methods are overly empirical and simplistic, generally relying on manual inspection and judgment, and using regular geometric shapes (such as semi-cylinders and hemispheres) to approximate complex real defect morphologies. This simplification ignores the key local features of defects, leading to distorted input conditions for subsequent mechanical analysis, and evaluation results that are often overly conservative or pose safety hazards. In addition, defect data analysis is in a "siloed" state, with defect detection, measurement, classification, and life assessment processes disconnected from each other. The massive amount of detection data has not been effectively integrated and mined, making it impossible to extract statistically significant standard defect spectra for guiding design. Moreover, the safety assessment process is static. Traditional defect tolerance analysis is based on a fixed initial defect size and cannot simulate the dynamic evolution of defects under real complex loads. Therefore, it cannot predict the remaining life of individual components and their specific defects, resulting in a lack of foresight in maintenance decisions.

[0004] In conclusion, there is still much room for improvement in the accuracy and realism of current characterization of surface defects in aero-engine structures. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for intelligent characterization and life assessment of surface defects in aero-engine structures to address the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification:

[0007] This specification provides a method for intelligent characterization and life assessment of surface defects in aero-engine structures, including:

[0008] Acquire multimodal defect data of the target aero-engine surface; and construct a knowledge graph of surface defects of the target aero-engine by fusing contextual information from the multimodal defect data.

[0009] The trained multi-branch deep convolutional neural network model is used to identify multimodal defect data and obtain the geometric feature parameters of surface defects of aero-engines.

[0010] Based on the geometric feature parameters of aero-engine surface defects, a trained AI proxy model is used to characterize the surface defects of the target aero-engine as standard equivalent cracks with the same mechanical effects; wherein, the AI ​​proxy model is a fast mapping model from defect geometric parameters to theoretical stress concentration coefficient Kt based on a lightweight neural network.

[0011] Based on standard equivalent cracks and knowledge graphs, dynamic simulation is performed using digital twin models of aero-engine components and combined with the actual load spectrum of the aero-engine components to predict the evolution path of surface defects in the target aero-engine and assess its remaining service life.

[0012] Furthermore, the contextual information of the multimodal defect data includes the component model, specific location, material grade, service history, and historical maintenance records of the defect.

[0013] Furthermore, the construction of the target aero-engine surface defect knowledge graph includes:

[0014] By associating entity nodes of surface defects of the target aero-engine with component entity nodes, material entity nodes, and operating condition entity nodes through relational edges, an interconnected knowledge graph of surface defects of the target aero-engine is constructed.

[0015] Furthermore, the multi-branch deep convolutional neural network model specifically includes:

[0016] The PointNet++ branch used for processing 3D point cloud data;

[0017] CNN branch used for processing two-dimensional image data;

[0018] Connectivity layer for feature fusion;

[0019] Output layer used for feature output.

[0020] Furthermore, based on the trained multi-branch deep convolutional neural network model, multimodal defect data is identified to obtain the geometric feature parameters of aero-engine surface defects, including:

[0021] For 3D point cloud data, feature extraction and downsampling are performed using the Set Abstraction (SA) module of PointNet++; upsampling and feature fusion are performed using the Feature Propagation (FP) module to obtain the global feature vector of surface defects of aero-engines.

[0022] For two-dimensional images, feature vectors are extracted using CNN branches;

[0023] The global feature vector output by the PointNet++ branch is concatenated with the feature vector output by the CNN branch to obtain the fused feature vector of the surface defects of the aero-engine.

[0024] The fused feature vector is passed through a fully connected layer with an output dimension equal to the number of defect categories and a Softmax function to obtain the probability distribution of defect types;

[0025] The fused feature vector is passed through a fully connected layer whose output dimension is the number of required geometric features to obtain the feature values ​​of specific aero-engine surface defects.

[0026] Furthermore, the surface defects of the aero-engine specifically include scratch defects, pit defects, and corrosion defects.

[0027] Furthermore, the fast mapping model from defect geometric parameters to theoretical stress concentration factor Kt based on a lightweight neural network specifically includes:

[0028] The Kt values ​​of defect models with different geometric shapes are calculated in batches using the parametric finite element method, generating a large number of defect geometry-Kt value sample pairs.

[0029] Using the sample pairs as the training set, a lightweight neural network regression model is trained, which learns the nonlinear mapping relationship from defect geometric parameters to Kt values.

[0030] Furthermore, the prediction of the evolution path of surface defects on the target aero-engine and the assessment of its remaining service life specifically include:

[0031] A multiphysics coupling model was constructed based on the SIMULIA Abaqus platform by coupling structural dynamics, fatigue damage mechanics, and corrosion evolution physics.

[0032] The crack is implanted into the multiphysics coupling model, and the crack propagation amount under each load cycle is calculated using the Paris formula or the Forman formula based on the fatigue crack propagation theory of fracture mechanics.

[0033] The remaining life of the crack is obtained when the crack propagation reaches the critical crack size.

[0034] Furthermore, the construction of the intelligent defect representation method also includes:

[0035] Based on knowledge graphs, in-depth data on multiple defects of the same type were obtained;

[0036] Statistical analysis was performed on the in-depth data of multiple defects of the same type extracted;

[0037] A hybrid probability model is used to fit the depth data of multiple defects of the same type, and the standard characterization size range of each type of defect is determined according to a preset coverage criterion.

[0038] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0039] This invention replaces the traditional, extensive representation method that relies on human experience and regular geometric approximations by using high-precision 3D scanning and multi-source data fusion. It automatically extracts key geometric features of defects, such as depth and curvature, using a deep learning model, effectively avoiding subjective errors from human judgment. This results in a more realistic and accurate digital description of defects, providing a reliable data foundation for subsequent analysis.

[0040] Furthermore, the fully automated analysis chain constructed in this invention, from intelligent identification and feature extraction to statistical modeling, significantly reduces manual intervention and repetitive labor. In particular, through the trained AI proxy model, the finite element mechanics analysis that originally took several hours is shortened to the second level, realizing rapid mechanical quantification of massive defects, improving analysis efficiency by orders of magnitude, and enhancing the accuracy of its characterization and prediction. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0042] Figure 1 This document provides a flowchart of a method for intelligent characterization and life assessment of surface defects in aero-engine structures. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0044] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart illustrating a method for intelligent characterization and life assessment of surface defects in aero-engine structures, as described in this specification. The method includes the following steps:

[0046] S101: By using high-precision 3D scanning, industrial CT and digital imaging technology, multimodal defect data of the surface of key engine components are obtained, and the contextual information of the defects is fused to construct an interconnected defect knowledge graph.

[0047] Multimodal data acquisition: 3D point cloud data of the pits were acquired using the ATOS Q 3D scanner (measurement accuracy ±5μm) from GOM GmbH, Germany, with a point spacing of 0.05mm. Simultaneously, 2D true-color images of the pits were acquired using a Keyence VHX-6000 digital microscope at 200x magnification.

[0048] Contextual information fusion: Extract the blade's serial number, material certificate number (corresponding to Ti-6Al-4V), and total service time (5000 equivalent flight cycles) from the engine history database, and associate them with the current inspection mission number.

[0049] Knowledge graph construction: The Neo4j graph database is used for construction, including the creation of nodes and relationships.

[0050] Create nodes: (d:Defect {id: 'D001', type: 'To be categorized'}), (p:Part {name: 'HPC_Stage3_Blade', SN: 'BLD-XXXX'}), (m:Material {name: 'Ti-6Al-4V',YieldStrength: 930MPa}), (u:Usage {TotalFC: 5000});

[0051] Create relations: (d)-[r1:LOCATED_ON]->(p), (p)-[r2:MADE_OF]->(m), (p)-[r3:HAS_HISTORY]->(u);

[0052] This results in a machine-readable and searchable defect knowledge graph.

[0053] S102: Construct and utilize a multi-branch deep convolutional neural network model to automatically identify and classify multimodal defect data, and extract key geometric feature parameters of defects.

[0054] Network Model Structure: The multi-branch deep learning model is implemented in the PyTorch framework. The 3D branch uses PointNet++, and its network layers include: a Set Abstraction layer (number of points: 1024->256->64) and a FeaturePropagation layer for upsampling. The 2D branch uses ResNet-50, removing its last fully connected layer to output a 1024-dimensional feature vector. The feature vectors from the two branches are concatenated and fed into a fusion module consisting of two fully connected layers (dimensions 512 and 256 respectively), and finally connected to a Softmax classifier and a regression layer.

[0055] Model Training: The network was trained using a dataset containing 5000 labeled defect samples (including 3D point clouds and 2D images). The optimizer used was Adam with an initial learning rate of 0.001, and the loss function was a weighted sum of cross-entropy loss (classification) and smoothed L1 loss (regression).

[0056] Application: Input the 3D point cloud of the crater in this example (downsampled to 2048 points) and the cropped 2D image (scaled to 224x224 pixels) into the trained model. The model outputs a classification vector of [0.985, 0.010, 0.005,...], corresponding to the "impact crater" category; the regression layer output vector is [0.13, 1.2, 0.9], corresponding to depth, major axis, and minor axis (unit: mm), respectively.

[0057] S103: Perform statistical analysis on the extracted group defect feature parameters, fit them using a mixed probability model, and determine the standard characterization size range for each type of defect according to the preset coverage criterion.

[0058] Data preparation: Depth data of 255 similar pits were retrieved from the knowledge graph, forming a dataset X = {x1, x2, ..., x255}.

[0059] Mixture model fitting: The GaussianMixture model from Python's scikit-learn library is used, with n_components=2 (i.e., two Gaussian distributions). The model is fitted to the dataset X to obtain the weights of the two Gaussian components. mean and standard deviation ;

[0060] Standard size range calculation: Based on the fitted GMM model, its cumulative distribution function is calculated. The depths d90, d5, and d95 corresponding to cumulative probabilities of 90%, 5%, and 95% are solved using numerical methods (such as the bisection method). The final standard characterization depth range is determined to be [d5, d95], and in this embodiment, the calculation result is [0.10, 0.20] mm.

[0061] S104: By training a lightweight neural network, a fast mapping model from defect geometric parameters to theoretical stress concentration coefficient Kt is established. This model is used to equate real defects to standard equivalent cracks with the same mechanical effects.

[0062] Sample database generation: Parametric finite element models of the pits were built in Abaqus software, with parameters including depth H (0.05-0.3 mm), major axis L (0.5-2.0 mm), and minor axis W (0.5-2.0 mm). Calculations were submitted in batches using a Python script to extract the maximum Von Mises stress for each model and calculate Kt = σmax / σnominal. A total of 15,000 (H, L, W, Kt) sample pairs were generated.

[0063] AI Agent Model Training: A fully connected neural network with three hidden layers (64, 32, and 16 neurons respectively) was constructed, using ReLU activation. The input was (H, L, W), and the output was Kt. The sample database was divided into training and test sets in an 8:2 ratio, and training was performed using MSE loss and the Adam optimizer. The trained model achieved a determination coefficient R² > 0.98 on the test set.

[0064] Application and Equivalence: Inputting (0.13, 1.2, 0.9) into the AI ​​proxy model predicts Kt=2.8. In the predefined equivalent rule base, the semi-elliptical surface crack size corresponding to Kt=2.8 is found to be (a=0.15mm, 2c=0.6mm) by interpolation.

[0065] S105: Using standard equivalent cracks as the initial state, the model is embedded in the digital twin model of the engine component. Dynamic simulation is then performed in conjunction with the actual load spectrum of the component to predict the evolution path of the defect and assess the remaining service life of the structure.

[0066] Digital Twin and Load Spectrum: A parametric finite element model of the blade is created as a digital twin in ANSYS Workbench. The load spectrum is defined in tabular form, including the time / cycle number sequence and the corresponding rotational speed (percentage), centrifugal force load, and aerodynamic pressure.

[0067] Crack propagation simulation: An equivalent semi-elliptical crack (a=0.15, c=0.3) mm is implanted into the model. The crack propagation analysis function is invoked using the Franc3D plugin or through a self-written script. In each cycle i, the current stress intensity factor range ΔK is calculated, and the crack propagation amount Δa for this cycle is calculated according to the Paris formula for the material da / dN=2.0e-11*(ΔK)3.5 (unit: m / cycle, MPa√m), and the crack size a{i+1}=ai+Δa is updated.

[0068] Lifetime termination and output: The simulation stops when the crack depth 'a' reaches the critical crack size 'ac' = 1.5 mm calculated based on the material fracture toughness KIC. The total number of cycles, Ntotal = 3200, represents the predicted remaining lifetime.

[0069] S106: Maintenance decision support and feedback optimization.

[0070] Based on the remaining service life output in step S105, generate condition-based maintenance recommendations, including inspection intervals and maintenance timing;

[0071] When a component is decommissioned or undergoes detailed inspection, the actual expansion data of the defect is obtained and compared with the prediction result of step S105. The comparison result is used to calibrate and optimize the crack expansion parameters in the digital twin model or the prediction accuracy of the AI ​​proxy model.

[0072] In embodiments of the present invention, the specific implementation steps are as follows:

[0073] 1) The context information in step S101 above includes at least three of the following: the part model where the defect is located, the specific location, the material grade, the service history, and the historical maintenance records; the defect knowledge graph uses the defect entity as a node and is associated with the component entity, material entity, and operating condition entity through relational edges.

[0074] In this embodiment, the specific method for collecting and fusing the context information is as follows:

[0075] Part number: Obtained from the engine product structure tree (BOM), specifically "Model: GENX-1B, Part: HPC Stage 5 Blade".

[0076] Specific location: In the 3D scanning software, by marking the centroid coordinates (x, y, z) on the 3D point cloud of the defect and associating it with the parametric CAD model of the blade, its location is described as "Location: Leading Edge, 50% Span".

[0077] Material grade: The material certificate database for the component shows "Material:TC4, Heat Treat:AMS5662".

[0078] Service history: Extract the total number of flight cycles "Total Flight Cycles: 12000" and the total number of flight hours "Total Flight Hours: 28000" from the engine flight record database.

[0079] Historical maintenance records: The maintenance management system shows that the part was last inspected on August 1, 2023. The last item was found to be a minor scrape. The last item was not taken.

[0080] The defect knowledge graph is constructed using the graph database Neo4j. Its specific structure is as follows:

[0081] Node creation:

[0082] Create a defect node: CREATE (d:Defect {id: 'D-2024-001', type: 'Pitting_Corrosion'})

[0083] Create a part node: CREATE (p:Part {name: 'HPC_5_Blade', model: 'GENX-1B'})

[0084] Create a material node: CREATE (m:Material {name: 'IN718', UTS: 1350})

[0085] Create a cycle condition node: CREATE (c:Condition {total_cycles: 12000})

[0086] Relationship establishment:

[0087] (d)-[:LOCATED_ON {coordinates: '50% Span'}]->(p)

[0088] (p)-[:MADE_OF]->(m)

[0089] (p)-[:OPERATED_UNDER]->(c)

[0090] This graph structure enables complex semantic queries, such as: "Find all corrosion defects on parts made of IN718 material that have been in service for more than 10,000 cycles".

[0091] 2) The multi-branch deep convolutional neural network described in step S102 above includes a PointNet++ branch for processing three-dimensional point cloud data and a CNN branch for processing two-dimensional image data. The two branches perform feature fusion at the back end of the network. The key geometric feature parameters include depth and bottom radius for scratch defects, depth for pit defects, and depth and opening diameter for corrosion defects.

[0092] In this embodiment, the specific architecture and data processing flow of the multi-branch deep convolutional neural network are as follows:

[0093] Input data processing: For the PointNet++ branch: The 3D point cloud data is downsampled to 2048 points, with each point's features being (x, y, z) coordinates. This data is first processed through a series of Set Abstraction (SA) modules for feature extraction and downsampling, then upsampled and fused using the Feature Propagation (FP) module, and finally max pooling is used to obtain a 256-dimensional global feature vector.

[0094] For the CNN branch (using ResNet-50 as an example): the 2D image is scaled to 224x224 pixels and normalized. This image is then input into the modified ResNet-50 (with the last classification layer removed), and a 1024-dimensional feature vector is extracted from the avgpool layer.

[0095] Feature fusion: The 256-dimensional vector output by the PointNet++ branch and the 1024-dimensional vector output by the CNN branch are concatenated to obtain a 1280-dimensional fused feature vector.

[0096] Output layer:

[0097] Classification task: The fused feature vector is passed through a fully connected layer (output dimension is the number of defect categories) and a Softmax function to output the probability distribution of defect types.

[0098] Regression task: The fused feature vector is passed through another fully connected layer (the output dimension is the number of required geometric features) to output specific feature values.

[0099] Feature extraction for different defects:

[0100] For scratch defects, the regression layer outputs [depth, bottom radius]. The bottom radius is calculated by fitting the curvature of the bottom contour of the scratch.

[0101] For pit defects, the regression layer outputs [depth].

[0102] For corrosion defects, the regression layer outputs [depth, opening diameter]. The opening diameter is estimated by fitting the minimum circumcircle to the edge of the corrosion pit opening.

[0103] 3) If the mixed probability model in step S103 above is a Gaussian mixture model; the preset coverage criterion is 90% coverage, and on this basis, a size range with a specific confidence level is determined as the standard characterization size range.

[0104] Model selection and fitting: The Gaussian mixture model was fitted using the GaussianMixture class from the Scikit-learn library. Assuming 300 scratch depth data points were collected, the Bayesian Information Criterion (BIC) was used for model selection. It was found that the BIC value was minimized when n_components=2, therefore a mixture model containing two Gaussian distributions was chosen. The Expectation-Maximization (EM) algorithm was used to estimate the model parameters (weights, mean, variance).

[0105] Standard size range calculation:

[0106] 90% Coverage: Calculate the cumulative distribution function of the GMM model and find the quantile d90 corresponding to a cumulative probability of 90%. This means that 90% of the field defect depths are less than or equal to d90.

[0107] Confidence Intervals: To further enhance engineering reliability, Bootstrap sampling was employed. A sample set of 1000 samples with replacement was drawn from the original 300 data points. A Gaussian Mixture Model (GMM) was fitted to each sample set, and its d90 was calculated. These 1000 d90 values ​​formed a distribution, and its 2.5% and 97.5% quantiles were used as the lower and upper bounds of the 95% confidence interval for d90.

[0108] Ultimately, the standard characterization size range was determined to be [d_90_lower, d_90_upper]. For example, the calculation result is [0.08mm, 0.14mm]. This means that there is a 95% confidence that the upper limit of the depth covering 90% of field defects is between 0.08mm and 0.14mm.

[0109] 4) The specific method for establishing the fast mapping model in step S104 above is as follows:

[0110] The Kt values ​​of defect models with different geometric shapes are calculated in batches using the parametric finite element method, generating a large number of defect geometry-Kt value sample pairs.

[0111] Using sample pairs as the training set, a lightweight neural network regression model is trained, which learns a nonlinear mapping relationship from defect geometry parameters to Kt values.

[0112] The specific implementation method of this embodiment is as follows:

[0113] Parametric Finite Element Method and Sample Generation:

[0114] In ANSYS Workbench, use DesignModeler to create a parametric pit model, setting the depth d, length l, and width w as input parameters.

[0115] Set up a static structural analysis in ANSYS Mechanical and apply typical blade working stresses.

[0116] Using the Central Composite Design experimental design method of ANSYS DesignXplorer, 500 sampling points were generated in the parameter space (d: 0.05-0.3mm, l: 0.5-2.0mm, w: 0.5-2.0mm).

[0117] A batch script automatically runs 500 finite element analyses, extracts the maximum equivalent stress from each result, and calculates the Kt value. The final result is a CSV file containing 500 rows of data, with each row containing (d, l, w, Kt).

[0118] Training a lightweight neural network regression model:

[0119] A fully connected neural network is built using the PyTorch framework. The network structure is as follows: Input layer (3 nodes, corresponding to d, l, w) -> Hidden layer 1 (32 nodes, ReLU activation) -> Hidden layer 2 (16 nodes, ReLU activation) -> Output layer (1 node, linear activation, corresponding to Kt).

[0120] The 500 samples were divided into a training set and a test set in an 8:2 ratio.

[0121] Compile the model, use the Adam optimizer, and set the loss function to Mean Squared Error (MSE).

[0122] The model is trained on the training set and validated on the test set. The model is considered successfully trained when the coefficient of determination (R²) between the model's predictions on the test set and the actual Kt values ​​calculated by the FEM is greater than 0.97. This model can then be used for Kt value prediction at the second level.

[0123] 5) The digital twin model in step S105 above is a multi-physics coupling model that integrates structural dynamics, fatigue damage mechanics and corrosion evolution physics; the dynamic simulation is based on the fatigue crack propagation theory of fracture mechanics, and calculates the crack propagation amount under each load cycle through the Paris formula or the Forman formula.

[0124] The specific implementation method of this embodiment is as follows:

[0125] Multiphysics Coupled Model: This digital twin was built on the SIMULIA Abaqus platform. It is a comprehensive model that couples structural dynamics (simulating inertial effects through implicit dynamic analysis steps), fatigue damage mechanics (through integrated Franc3D crack propagation analysis), and corrosion evolution physics (defining the degradation of the material surface under specific temperature and humidity conditions through user subroutines).

[0126] Dynamic simulation and crack propagation:

[0127] Load spectrum: The typical flight profile of the engine (takeoff-cruise-landing) is discretized into 10 time points, and each time point is assigned a corresponding rotational speed centrifugal load and aerodynamic pressure.

[0128] Crack propagation theory: The Forman formula da / dN=[C*(ΔK)m] / [(1-R)*Kc-ΔK] is used for calculation. This formula is particularly suitable for high load ratios R. The material parameters C, m, and Kc can be found in material handbooks.

[0129] Simulation workflow: After completing a static analysis of a flight profile in Abaqus, the Franc3D plugin is invoked. Franc3D reads the current crack lead ΔK, calculates the propagation Δa for this cycle according to the Forman formula, automatically updates the crack shape, and sends the new crack mesh back to Abaqus for the next round of analysis. This process is repeated until the crack reaches the critical size.

[0130] 6) Maintenance decision support and feedback optimization;

[0131] (6.1) Based on the remaining service life output in step 5), generate condition-based maintenance recommendations, including inspection intervals and maintenance timing;

[0132] (6.2) When a component is decommissioned or undergoes detailed inspection, the actual expansion data of the defect is obtained and compared with the prediction results of step S500. The comparison results are used to calibrate and optimize the crack propagation parameters in the digital twin model or the prediction accuracy of the AI ​​agent model.

[0133] The specific implementation of this embodiment is as follows:

[0134] Maintenance Decision Support

[0135] Input: Step 5) Output remaining lifetime = 3200 cycles.

[0136] Logic: The system's default safety factor is 2.0. Therefore, the recommended check interval = remaining service life / safety factor = 3200 / 2 = 1600 cycles.

[0137] Output: An automatic work order is generated in the maintenance management system, with the following content: "Work order type: borehole inspection; target component: engine serial number XX, HPC stage 5 blade; execution timing: within 1600 flight cycles from this inspection."

[0138] Feedback optimization and model calibration

[0139] Data Acquisition: When the blade undergoes borehole inspection or decomposition after 1600 cycles, accurately measure the current size of the defect. Assume the current crack depth is measured to be a_actual = 0.25 mm.

[0140] Data comparison: Compare the actual data (N=1600, a_actual=0.25) with the predicted curve from step S500. Assume the original predicted curve has a predicted size of a_predicted=0.22mm when N=1600.

[0141] Model calibration:

[0142] Calibration of the digital twin model: A Bayesian update method was employed. The parameters C and m in the Paris formula were treated as random variables, and their probability distributions were updated using actual measurement data (1600, 0.25). After the update, the expected values ​​of C and m will be closer to reality, which can be used for subsequent lifespan predictions of similar components, thus making the predictions more accurate.

[0143] Calibration of the AI ​​agent model: Add the actual measurement data (d,l,w,Kt_FEM) to the sample database of step S400, and retrain or fine-tune the lightweight neural network. This can correct the prediction bias of the model in specific parameter regions and improve its generalization ability.

[0144] A computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above; the computer program implementing the above methods S101 to S106 (e.g., written in Python and relying on libraries such as PyTorch, Scikit-learn, and Neo4j driver) is stored on a physical medium such as a USB flash drive, hard disk, or optical disc, and when the program in the medium is read and executed by the processor of one or more computers, the method described in this invention can be implemented.

[0145] The technical effects and advantages of this invention include:

[0146] (1) This invention replaces the traditional extensive representation method that relies on human experience and regular geometric approximation by using high-precision three-dimensional scanning and multi-source data fusion. By using a deep learning model to automatically extract key geometric features such as the depth and curvature of defects, the invention effectively avoids subjective errors in human judgment, making the digital description of defects more realistic and accurate, and providing a reliable data foundation for subsequent analysis.

[0147] (2) The fully automated analysis chain constructed by this invention, from intelligent identification and feature extraction to statistical modeling, greatly reduces manual intervention and repetitive labor. In particular, through the trained AI proxy model, the finite element mechanical analysis that originally took several hours is shortened to the second level, realizing rapid mechanical quantification of massive defects and improving analysis efficiency by orders of magnitude.

[0148] (3) This invention is based on statistical distribution modeling of group defect data and determines the standard defect size range according to the coverage criteria approved by the engineering, so that the input conditions on which the defect tolerance design is based are more in line with the actual statistical laws. This can avoid structural redundancy and weight costs caused by overly conservative design, and also prevent safety risks caused by insufficient characterization, thus achieving a balance between safety and economy.

[0149] (4) By embedding the defect equivalent model into a digital twin that integrates multiphysics fields and applying a load spectrum for dynamic simulation, this invention can predict the evolution path and expansion rate of a specific defect under real service conditions. This enables engineers to proactively assess the remaining service life of individual components, realizing a shift from "post-inspection" to "pre-prediction".

[0150] (5) The quantitative characterization parameters and individualized remaining life prediction results output by this invention can directly serve the engine maintenance and support decision-making. Based on the prediction results, engineers can scientifically formulate inspection intervals and maintenance plans, thereby realizing the transformation from the traditional "scheduled maintenance" to an efficient and accurate "condition-based maintenance" mode, effectively improving fleet uptime and reducing the total life cycle maintenance cost.

[0151] (6) By introducing in-service feedback data to calibrate and optimize the digital twin model and the AI ​​agent model, this invention forms an analysis system with self-learning and self-improvement capabilities. As the amount of data continues to accumulate, the accuracy of its representation and prediction will continue to improve, and the reliability and practicality of the system will be continuously enhanced.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for intelligent characterization and life assessment of surface defects in aero-engine structures, characterized in that, include: Acquire multimodal defect data on the surface of the target aero-engine; By integrating the contextual information of multimodal defect data, a knowledge graph of surface defects of the target aero-engine is constructed. The contextual information of the multimodal defect data includes the component model, specific location, material grade, service history, and historical maintenance records of the defect. The construction of the knowledge graph of surface defects of the target aero-engine includes associating the entity nodes of surface defects of the target aero-engine with component entity nodes, material entity nodes, and operating condition entity nodes through relational edges to construct an interconnected knowledge graph of surface defects of the target aero-engine. The trained multi-branch deep convolutional neural network model is used to identify multimodal defect data and obtain the geometric feature parameters of surface defects of aero-engines. Based on the geometric feature parameters of aero-engine surface defects, a trained AI proxy model is used to characterize the surface defects of the target aero-engine as standard equivalent cracks with the same mechanical effects; wherein, the AI ​​proxy model is a fast mapping model from defect geometric parameters to theoretical stress concentration coefficient Kt based on a lightweight neural network. Based on standard equivalent cracks and knowledge graphs, dynamic simulation is performed using digital twin models of aero-engine components and combined with the actual load spectrum of the aero-engine components to predict the evolution path of surface defects of the target aero-engine and assess its remaining service life. The prediction of the evolution path of surface defects in the target aero-engine and the assessment of its remaining service life specifically include: Based on the SIMULIA Abaqus platform, a multiphysics coupling model is constructed by coupling structural dynamics, fatigue damage mechanics, and corrosion evolution physics. The standard equivalent crack is implanted into the multiphysics coupling model, and the crack propagation amount under each load cycle is calculated using the Paris formula or the Forman formula based on the fatigue crack propagation theory of fracture mechanics. The remaining life of the crack is obtained when the crack propagation reaches the critical crack size.

2. The intelligent characterization and life assessment method for surface defects in aero-engine structures as described in claim 1, characterized in that, The multi-branch deep convolutional neural network model specifically includes: The PointNet++ branch used for processing 3D point cloud data; CNN branch used for processing two-dimensional image data; Connectivity layer for feature fusion; Output layer used for feature output.

3. The intelligent characterization and life assessment method for surface defects of aero-engine structures as described in claim 1, characterized in that, Based on the trained multi-branch deep convolutional neural network model, multimodal defect data is identified to obtain the geometric feature parameters of aero-engine surface defects, including: For 3D point cloud data, feature extraction and downsampling are performed using the Set Abstraction (SA) module of PointNet++; upsampling and feature fusion are performed using the Feature Propagation (FP) module to obtain the global feature vector of surface defects of aero-engines. For two-dimensional images, feature vectors are extracted using CNN branches; The global feature vector output by the PointNet++ branch is concatenated with the feature vector output by the CNN branch to obtain the fused feature vector of the surface defects of the aero-engine. The fused feature vector is passed through a fully connected layer with an output dimension equal to the number of defect categories and a Softmax function to obtain the probability distribution of defect types; The fused feature vector is passed through a fully connected layer whose output dimension is the number of required geometric features to obtain the feature values ​​of specific aero-engine surface defects.

4. The intelligent characterization and life assessment method for surface defects of aero-engine structures as described in claim 1, characterized in that, The surface defects of the aero-engine specifically include scratches, pits, and corrosion.

5. The intelligent characterization and life assessment method for surface defects of aero-engine structures as described in claim 1, characterized in that, The fast mapping model from defect geometric parameters to theoretical stress concentration factor Kt, based on a lightweight neural network, specifically includes: The Kt values ​​of defect models with different geometric shapes are calculated in batches using the parametric finite element method, generating a large number of defect geometry-Kt value sample pairs. Using the sample pairs as the training set, a lightweight neural network regression model is trained, which learns the nonlinear mapping relationship from defect geometric parameters to Kt values.

6. The intelligent characterization and life assessment method for surface defects of aero-engine structures as described in claim 1, characterized in that, The construction of intelligent defect representation methods also includes: Based on knowledge graphs, in-depth data on multiple defects of the same type were obtained; Statistical analysis was performed on the in-depth data of multiple defects of the same type extracted; A hybrid probability model is used to fit the depth data of multiple defects of the same type, and the standard characterization size range of each type of defect is determined according to a preset coverage criterion.

Images