Aero-engine gear remaining life prediction method based on multi-model fusion and Gaussian process regression

Abstract

The invention discloses an aero-engine gear remaining life prediction method based on multi-model fusion and Gaussian process regression, and belongs to the technical field of aerospace engineering and machine learning cross fusion. The method is realized through combination of a code module and a visual identification module, input parameters comprise damage information extracted from a gear image, working condition parameters and gear inherent characteristics, and output is a residual life prediction value and a confidence interval with hour as a unit. The method realizes standardization and automation of the evaluation process. Visual damage characteristics, dynamic working conditions and material attributes are fused, and through double mechanisms of physical modeling and GPR calibration, a prediction result is comprehensive and reliable. The GPR algorithm is innovatively introduced, and a stable and effective prediction result can still be realized under the condition of small samples. The prediction result is not limited to life point estimation, the confidence range is output, and risk control and maintenance strategy making are facilitated. And a modular design is adopted, so that the method has good engineering iteration capability.

Description

Technical Field

[0001] This invention relates to a method for predicting the remaining life of aero-engine gears based on multi-model fusion and Gaussian process regression, belonging to the field of interdisciplinary integration technology of aerospace engineering and machine learning. Background Technology

[0002] Aero-engine gears are critical components in the engine's power transmission system. Operating under extreme conditions such as high temperature, high pressure, and high speed, they are highly susceptible to various types of surface damage, including pitting, spalling, abrasion, and indentation. This damage not only weakens the gear's transmission performance but can also lead to serious accidents such as gear fracture and meshing failure, directly threatening flight safety.

[0003] In existing technologies, gear life assessment methods mainly rely on manual interpretation or empirical formulas, which have the following prominent problems:

[0004] (1) High subjectivity and inconsistent standards: Damage assessment relies heavily on manual visual inspection, lacks unified standards, and the assessment results vary greatly, making it difficult to be objective and repeatable;

[0005] (2) Low efficiency and difficulty in scaling up: The evaluation process relies heavily on professional personnel, making it difficult to achieve efficient batch processing and not adapting to the needs of modern aerospace manufacturing for intelligence and automation;

[0006] (3) Small sample size restricts modeling: The actual failure samples of gears are extremely limited, making it difficult to effectively apply traditional deep learning methods that rely on big data, resulting in poor generalization ability of the prediction model;

[0007] (4) Damage coupling has not been effectively modeled: Gear damage is mostly multi-type composite and co-evolved. Existing studies mostly target single damage types and cannot accurately describe its life evolution law.

[0008] In recent years, machine vision technology has been introduced for the identification and quantification of static damage to gears. Although it can obtain accurate damage geometry information, how to effectively integrate these static, offline image features with the dynamic service conditions and physical degradation mechanisms of gears to establish a reliable life prediction model remains a key technical problem that has not been fully solved in this field. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a method for predicting the life of aero-engine gears that can integrate multi-source information, support small sample modeling, and possess the dual advantages of physical interpretability and prediction accuracy, so as to achieve intelligent, automated, and quantifiable life assessment.

[0010] The technical solution adopted in this invention is: a method for predicting the remaining life of aero-engine gears based on multi-model fusion and Gaussian process regression. This method is implemented by combining MATLAB code module and visual recognition module. The input parameters include damage information extracted from gear images, operating condition parameters and inherent characteristics of gears. The output is the predicted remaining life value and confidence interval in hours.

[0011] The method includes the following steps:

[0012] S1. Multi-source information acquisition and fusion

[0013] S1.1 comprehensively collects the overall attribute parameters of the gear system through industrial sensors, digital design drawings, and experimental measurements.

[0014] S1.2 uses a deep learning-based visual recognition module to automatically detect and recognize the gear surface, directly extracting structured damage information from high-resolution images and converting the original image information into numerical features that can be directly used by the computational model.

[0015] S1.3 In the data fusion stage, the above-collected working condition parameters, geometric attribute parameters and damage features obtained by visual recognition are fused to form a unified comprehensive input feature set;

[0016] S2. Damage Type Encoding and Model Calling Mechanism

[0017] S2.1 uses a four-bit binary encoding scheme to uniquely identify the type of gear damage, with each bit corresponding to a specific type of damage;

[0018] S2.2 In the scenario of complex damage identification, a bitwise OR operation is used to generate a combined code;

[0019] S2.3 Based on the above coding mechanism, the system can automatically match and call the corresponding single damage calculation module in the physical life prediction model library. In the case of multiple damages, multiple related models are called in parallel to form a combination scheme for basic life calculation.

[0020] S3. Multiphysics Model-Based Lifetime Prediction

[0021] S3.1 For the four typical damage types of gear systems, namely crushing, scuffing, pitting and spalling, select the corresponding public physical life prediction model and independently calculate the basic life value under each single damage condition.

[0022] S3.2 Combines the four-bit binary encoding mechanism from step S2 to characterize the composite state of multiple damages, and implements parallel invocation of multiple physical models under the encoding drive, generating a vector-based output containing the basic lifetimes corresponding to each damage. ;

[0023] S4. To address the issues of ideal model bias and sample scarcity, a GPR correction mechanism is employed.

[0024] S4.1 uses the lifetime vector obtained from the multi-physics model's basic lifetime prediction as the main physical feature, and fuses it with the damage features obtained by the visual recognition module and the collected operating condition parameters to form a structured high-dimensional input feature vector.

[0025] In terms of learning strategy, S4.2 learns the residual distribution between the physical model's predicted lifetime and the actual lifetime. Through residual prediction, GPR accurately captures the impact of factors that are not considered or are difficult to quantify in the physical model on lifetime.

[0026] S4.3 The final lifetime prediction result is obtained by weighted superposition of the minimum lifetime value output by the physical model and the GPR residual correction amount;

[0027] In terms of uncertainty estimation, S4.4 utilizes the inherent predictive variance output capability of GPR and integrates it with the discrete characteristics of the basic lifetime vector of the multiphysics model to construct a confidence interval for the prediction of composite damage lifetime.

[0028] Furthermore, in step S1.1, the comprehensive attribute parameters cover working condition parameters, geometric structure parameters, and material physical properties, including torque, rotational speed, load variation characteristics, center distance, module, tooth width, number of teeth, hardness, elastic modulus, and fatigue limit.

[0029] Furthermore, in step S1.2, the damage information includes damage type and key feature parameters. The damage type is pressure mark, abrasion, pitting, and exfoliation. The key feature parameters are damage area percentage, depth range, and location distribution.

[0030] Furthermore, in step S1.3, the data in the comprehensive input feature set undergoes feature standardization, normalization, and encoding preprocessing to ensure that the multi-source information maintains dimensional consistency and numerical stability in subsequent model calls and fusion calculations.

[0031] Furthermore, in step S4.3, the final lifetime prediction result is:

[0032]

[0033] in, These are the weighting coefficients. This is the minimum lifetime value output by the physical model. This is the GPR residual correction amount.

[0034] A composite damage life prediction system that implements the above-mentioned methods;

[0035] The system comprises an image acquisition and visual recognition module, a damage binary encoding module, a multi-physics model parallel computing module, a GPR residual correction module, and a result output and visualization module. It also achieves communication and data sharing among the various parts through a modular interface design.

[0036] Furthermore, in the image acquisition and visual recognition module, the system uses a deep learning visual recognition algorithm to automatically detect and classify damage on the gear surface, outputting structured data containing damage type and key dimensional features. The recognition results are directly input into the damage binary encoding module, which generates the corresponding damage state code according to a predefined four-bit binary encoding scheme, and uses the code as a control signal to drive the multi-physics model parallel computing module. This module automatically calls the corresponding single-damage physical lifetime prediction model based on the encoding results, and outputs the basic lifetime vector under multiple damage states.

[0037] Furthermore, the GPR residual correction module receives the lifetime vector generated by the multi-physics model parallel computing module and fuses it with visual recognition features and operating parameters to form the GPR input feature vector. In this module, GPR corrects the deviation of the physical model results according to the residual learning strategy and outputs the predicted lifetime value and its confidence interval.

[0038] Furthermore, the results output and visualization module presents the final results to the user, including the predicted lifespan, damage type encoding, various damage characteristic parameters, and confidence interval ranges.

[0039] The beneficial effects of this invention are:

[0040] (1) Achieve objective and automated life assessment: Replace manual judgment with image recognition results to effectively overcome the subjectivity and uncertainty of traditional experience methods and achieve standardization and automation of the assessment process.

[0041] (2) Significantly improve prediction accuracy: By integrating visual damage characteristics, dynamic working conditions and material properties, and through a dual mechanism of physical modeling and GPR calibration, the prediction results are comprehensive and reliable.

[0042] (3) Breaking through the bottleneck of small sample modeling: The innovative introduction of the GPR algorithm can still achieve stable and effective prediction results even with fewer than 15 samples.

[0043] (4) Provide confidence intervals to enhance decision support capabilities: The prediction results are not limited to lifetime point estimation, but also output confidence ranges, which helps in risk control and maintenance strategy formulation.

[0044] (5) The model system is open and has good scalability: It adopts a modular design and can integrate new damage mechanism models or replace the GPR algorithm at any time, and has good engineering iteration capabilities. Attached Figure Description

[0045] Figure 1 is a flowchart of the overall lifetime prediction process.

[0046] Figure 2 is a comparison diagram of the GPR correction of the present invention and the conventional correction. Detailed Implementation

[0047] The present invention will be specifically described below through embodiments. It should be noted that the following embodiments are only used to further illustrate the present invention, but are not limited thereto, unless otherwise stated.

[0048] The specific embodiments of the present invention are described in detail below with reference to the technical solutions:

[0049] This invention provides a method for predicting the remaining life of aero-engine gears based on multi-model fusion and Gaussian process regression, aiming to achieve high-precision and interpretable prediction of gear life under small sample conditions. The method is implemented jointly using MATLAB code modules and a visual recognition module. Input parameters include damage information extracted from gear images, operating parameters, and inherent gear characteristics. Output is the predicted remaining life value and confidence interval in hours.

[0050] This method mainly includes the following steps:

[0051] (1) Multi-source information collection and fusion

[0052] A. Comprehensive attribute parameters of the gear system are collected in all aspects, totaling 13 parameters, covering operating parameters (such as torque, speed, load variation characteristics, etc.), geometric parameters (such as center distance, module, tooth width, number of teeth, etc.), and material physical properties (such as hardness, elastic modulus, fatigue limit, etc.). These parameters are acquired through various means, including industrial sensors, digital design drawings, and experimental measurements, to ensure the comprehensiveness and accuracy of the input data. These attribute parameters are not only used for subsequent life calculations of the physical model, but also provide necessary physical background information for the machine learning model, ensuring the physical constraints and generalization ability of the prediction model from the source.

[0053] B. Employ a deep learning-based visual recognition module (such as YOLOv11) to automatically detect and recognize gear surfaces. This module directly extracts structured damage information from high-resolution images, including but not limited to damage types (dents, abrasions, pitting, spalling) and key feature parameters (damage area percentage, depth range, location distribution, etc.), ensuring accurate and repeatable quantitative descriptions of the damage. This module maintains high recognition accuracy even under complex backgrounds, varying lighting conditions, and partial occlusion, and transforms the original image information into numerical features that can be directly used by the computational model.

[0054] C. In the data fusion stage, this invention effectively fuses the collected operating condition parameters, geometric attribute parameters, and damage features obtained through visual recognition to form a unified comprehensive input feature set. Through preprocessing steps such as feature standardization, normalization, and encoding, it ensures that the multi-source information maintains dimensional consistency and numerical stability in subsequent model calls and fusion calculations. This fusion mechanism not only guarantees the complementarity of data from different sources but also establishes an efficient data channel for information interaction between multi-physics models and machine learning models.

[0055] (2) Damage type encoding and model calling mechanism

[0056] A. This invention designs a four-bit binary encoding scheme to uniquely identify gear damage types. Each bit corresponds to a specific damage type, specifically: indentation corresponds to code "0001", abrasion corresponds to code "0010", pitting corresponds to code "0100", and spalling corresponds to code "1000". When a certain type of damage exists, the corresponding position is "1", otherwise it is "0", thereby realizing a precise digital description of the damage type.

[0057] B. In the scenario of composite damage identification, a bitwise OR operation is used to generate a combined code. For example, when pressure marks and pitting exist at the same time, the code is "0101", which can unambiguously represent the coexistence of multiple damages. The advantage of this coding mechanism is its scalability and computational simplicity, which allows the system to quickly adapt to different combinations of damage quantities and types.

[0058] C. Based on the above coding mechanism, the system can automatically match and call the corresponding single damage calculation module in the physical lifetime prediction model library. In the case of multiple damages, it can call multiple related models in parallel to form a combination scheme for basic lifetime calculation, thereby accurately reflecting the coupling effect of different damage combinations on lifetime at the physical level.

[0059] (3) Multiphysics model-based lifetime prediction

[0060] This invention constructs a collaborative lifetime prediction framework for complex damage, characterized in that,

[0061] This invention constructs a collaborative lifetime prediction framework for complex damage. Within this framework, for four typical damage types in gear systems—denting, scuffing, pitting, and spalling—corresponding publicly available physical lifetime prediction models are selected to independently calculate the baseline lifetime value under each single damage condition, denoted as:

[0062]

[0063] These models, based on principles such as gear meshing mechanics, contact fatigue theory, and material damage evolution, can provide quantitative basis for predicting the theoretical life of each type of damage.

[0064] B. Design a four-bit binary damage encoding mechanism to accurately characterize the composite state of multiple damages. The encoding example is "0101" which corresponds to the simultaneous existence of pressure damage and pitting.

[0065] C. Combining the above four-bit binary encoding mechanism, the system can accurately characterize the composite state of multiple damages and realize the parallel invocation of multiple physical models under the drive of encoding, thereby generating a vector-form output containing the basic lifetime corresponding to each damage. This vector not only reflects the independent effects of different damages, but also implies the possible interaction trends between damages.

[0066] D. Compared with existing technologies that only support single damage prediction, this invention is the first to input the lifetime vector output by the multi-physics model as a physical feature under the composite damage state into the machine learning model, thereby significantly enhancing the model's generalization ability and prediction stability under multiple damage conditions while ensuring theoretical interpretability.

[0067] (4) GPR correction mechanism to address the problems of ideal model bias and sample scarcity.

[0068] This invention addresses the dual challenges of idealized assumptions in physical models for composite damage life prediction and the limited sample size in engineering practice. It proposes a residual correction method based on Gaussian process regression (GPR). In terms of input feature design, the system first uses the life vector obtained from the multi-physics model's basic life prediction as the primary physical feature. This vector is then fused with damage features acquired by the visual recognition module (including damage area percentage, damage depth range, and binary encoding of damage type) and collected operating parameters (torque, speed, material strength, etc.) to form a structured high-dimensional input feature vector. This vector simultaneously retains the theoretical prediction information from the physical model and the actual damage state characteristics observed in the field, thus providing sufficient correlation variables for GPR at the information level.

[0069] Regarding the learning strategy, the GPR model of this invention does not directly predict the entire lifetime value, but focuses on learning the residual distribution between the lifetime predicted by the physical model and the actual lifetime. This residual learning strategy can significantly reduce the modeling complexity of GPR in high-dimensional space, enabling it to maintain stable generalization ability even under small sample conditions, and effectively suppressing the risk of overfitting due to data scarcity. Through residual prediction, GPR can accurately capture the impact of factors that are not considered or are difficult to quantify in the physical model (such as complex load perturbations, microscopic defects in materials, etc.) on lifetime, achieving targeted correction of the physical model results.

[0070] The final lifetime prediction result is obtained by weighted summation of the minimum lifetime value output by the physical model and the GPR residual correction, and the mathematical expression is:

[0071]

[0072] in, , where is the weighting coefficient, used to balance the dominant role of the physical model in lifetime prediction with the improvement in prediction accuracy brought about by GPR correction. When At higher levels, the results depend more on the theoretical physics foundation; when At lower levels, the data-driven correction ratio increases to adapt to the actual conditions of complex operating situations.

[0073] Regarding uncertainty estimation, this invention utilizes the inherent predictive variance output capability of the GPR (Statistical Learning Model) and integrates it with the discrete characteristics of the fundamental lifetime vector of the multi-physics model to construct a confidence interval for composite damage lifetime prediction. This confidence interval not only reflects the predictive uncertainty of the statistical learning model but also fully embodies the reliability differences of the physical model under different damage states, thus providing engineers with a reliable quantitative basis for lifetime prediction. This method can still achieve stable confidence interval estimation under small sample conditions, significantly improving the interpretability and engineering usability of the prediction results.

[0074] (5) Implementation and System Deployment

[0075] This invention constructs an end-to-end automated composite damage lifetime prediction system at the implementation level, achieving seamless integration of the entire process from data acquisition, feature extraction, physical modeling to data-driven correction. The system as a whole consists of an image acquisition and visual recognition module, a damage binary encoding module, a multi-physical model parallel computing module, a GPR residual correction module, and a result output and visualization module, and achieves high-speed communication and data sharing among the various parts through an efficient modular interface design.

[0076] In the image acquisition and visual recognition module, the system uses deep learning visual recognition algorithms (such as YOLOv11) to automatically detect and classify damage on the gear surface, outputting structured data containing damage type and key dimensional features (such as area ratio and depth range). The recognition results of this module are directly input into the damage binary encoding module, which generates the corresponding damage state code according to a predefined four-bit binary encoding scheme, and uses the code as a control signal to drive the multi-physics model parallel computation module. This module automatically calls the corresponding single-damage physical lifetime prediction model based on the encoding results, outputting the basic lifetime vector under multiple damage states.

[0077] Subsequently, the GPR residual correction module receives the lifetime vector generated by the multi-physics model parallel computing module and fuses it with visual recognition features and operating condition parameters to form the GPR input feature vector. In this module, GPR corrects the bias in the physical model results according to the residual learning strategy and simultaneously outputs the predicted lifetime value and its confidence interval. The final results are presented to the user through the results output and visualization module, including the predicted lifetime value, damage type encoding, various damage feature parameters, and confidence interval ranges, ensuring that engineers can intuitively and quickly obtain key information.

[0078] In terms of system implementation, this invention uses MATLAB as the core development environment, leveraging its high efficiency in matrix operations, data visualization, and machine learning modeling to achieve a full-process computation response time of less than 3 seconds, thus meeting the requirements of real-time engineering applications. The entire system is the first to organically integrate visual recognition, multi-damage state encoding, parallel computation of multiple physical models, and GPR correction based on physical residuals. This overcomes the limitations of previous methods in predicting the lifespan of composite damage for single damages and single models, providing a high-precision, interpretable, and real-time deployable solution for the health management of gear systems under complex working conditions.

[0079] The original gear life obtained by manual interpretation or empirical formulas is compared with the GPR-corrected life estimated using the method of this application, as follows:

[0080] Table 1 Comparison of original lifespan and predicted lifespan

[0081]

[0082] The table above shows that the results obtained by using the method of this application can verify the accuracy and consistency of the prediction method under different damage types and combinations, and evaluate its reliability and promotion value in practical engineering applications.

[0083] The above embodiments are only used to illustrate the present invention. Any equivalent transformations and improvements made on the basis of the technical solutions of the present invention should not be excluded from the protection scope of the present invention.

Claims

1. An aero-engine gear residual life prediction method based on multi-model fusion and Gaussian process regression, characterized in that, The method is realized by combining a MATLAB code module and a visual recognition module, input parameters include damage information extracted from the gear image, working condition parameters and gear inherent characteristics, and the output is the remaining life prediction value and confidence interval in hours; The method comprises the following steps: S1. Multi-source information acquisition and fusion S1.1 Collecting comprehensive attribute parameters of the gear system in all directions through industrial sensors, digital design drawings and experimental measurement methods; S1.2 Using a visual recognition module based on deep learning to automatically detect and identify the gear surface, and directly extracting structured damage information from high-resolution images to convert the original image information into numerical features that can be directly used by the calculation model; S1.3 In the data fusion stage, the working condition parameters, geometric attribute parameters and damage features obtained by visual recognition are fused to form a unified comprehensive input feature set; S2. Damage type coding and model calling mechanism S2.1 A four-bit binary coding scheme is used to uniquely identify the gear damage type, each bit corresponds to a specific damage type; S2.2 In the case of composite damage identification, a bitwise "or" operation is used to generate a combined code; S2.3 Based on the above coding mechanism, the system can automatically match and call the corresponding single-damage calculation module in the physical life prediction model library, and in the case of multiple damages, multiple related models are called in parallel to form a combined solution for basic life calculation; S3. Multi-physical model basic life prediction S3.1 For the four typical damages of gear system, i.e. indentation, scratch, pitting and spalling, the corresponding public physical life prediction model is selected to independently calculate the basic life value under each single damage condition; S3.2 Combined with the four-bit binary coding mechanism in step S2, the composite state of multiple damages is represented, and the parallel calling of multiple physical models is realized under the coding drive to generate a vector form output containing the basic life of each damage ; S4. GPR correction mechanism to address ideal model deviation and sample scarcity problem S4.1 The life vector obtained by the multi-physical model basic life prediction is used as the main physical feature, and is fused with the damage features obtained by the visual recognition module and the working condition parameters collected to form a structured high-dimensional input feature vector; S4.2 In terms of learning strategy, the residual distribution between the physical model prediction life and the actual life is learned. Through residual prediction, GPR accurately captures the influence of factors that are not considered or difficult to quantify by the physical model on the life; S4.3 The final life prediction result is obtained by weighting and superimposing the minimum life value output by the physical model and the GPR residual correction amount; S4.4 In terms of uncertainty estimation, the inherent prediction variance output capability of GPR is used to fuse with the dispersion characteristics of the multi-physical model basic life vector to construct the confidence interval of the composite damage life prediction.

2. The method of claim 1, wherein, In step S1.1, the comprehensive attribute parameters include working condition parameters, geometric structure parameters and material physical properties, including torque, speed, load variation characteristics, center distance, modulus, tooth width, tooth number, hardness, elastic modulus and fatigue limit.

3. The method of claim 1, wherein, In step S1.2, the damage information includes damage type and key feature parameters, the damage type is indentation, scratch, pitting and spalling, and the key feature parameters are damage area ratio, depth range and position distribution.

4. The method of claim 1, wherein, In step S1.3, the data in the comprehensive input feature set undergoes feature standardization, normalization, and encoding preprocessing to ensure that the multi-source information maintains dimensional consistency and numerical stability in subsequent model calls and fusion calculations.

5. The method of claim 1, wherein, In step S4.3, the final lifetime prediction result is: ； wherein, is a weight coefficient, is a minimum life value output by the physical model, is a GPR residual correction amount.

6. A composite damage life prediction system characterized by, This system implements the method described in any one of claims 1-5; The system comprises an image acquisition and visual recognition module, a damage binary encoding module, a multi-physics model parallel computing module, a GPR residual correction module, and a result output and visualization module. It also achieves communication and data sharing among the various parts through a modular interface design.

7. The system of claim 6, wherein: In the image acquisition and visual recognition module, the system uses a deep learning visual recognition algorithm to automatically detect and classify damage on the gear surface, and outputs structured data containing damage type and key dimensional features. The recognition results are directly input into the damage binary encoding module, which generates the corresponding damage state code according to a predefined four-bit binary encoding scheme, and uses the code as a control signal to drive the multi-physics model parallel computing module. This module automatically calls the corresponding single-damage physical lifetime prediction model according to the encoding results, and outputs the basic lifetime vector under multiple damage states.

8. The system of claim 6, wherein: The GPR residual correction module receives the lifetime vector generated by the multi-physics model parallel computing module and fuses it with visual recognition features and operating parameters to form the GPR input feature vector. In this module, GPR corrects the bias of the physical model results according to the residual learning strategy and outputs the predicted lifetime value and its confidence interval.

9. The system of claim 6, wherein: The results output and visualization module presents the final results to the user, including life prediction values, damage type codes, various damage characteristic parameters, and confidence interval ranges.

Images