Aero-engine fault diagnosis method based on conditional physical perception diffusion model

Abstract

The invention discloses an aero-engine fault diagnosis method based on a conditional physical perception diffusion model, and the method is characterized in that the method comprises the steps: obtaining a working condition monitoring parameter and a performance monitoring parameter in the operation process of an aero-engine; based on the thermodynamic system characteristics of the aero-engine, constructing a physical constraint system; a conditional physical perception diffusion model is constructed, noise addition and physical violation term injection are carried out in a physical perception forward process, and noise removal and physical violation term correction are respectively carried out through a noise prediction branch and a physical correction branch of a double-branch network by taking a fault type label as a condition in a reverse process. The double-branch network completes pre-training based on a physical constraint system; training the conditional physical perception diffusion model; based on the working condition monitoring parameters and the performance monitoring parameters, physical perception features are extracted through middle layer representation of the condition physical perception diffusion model, and fault diagnosis of the aero-engine is achieved through a fault diagnosis classifier.

Description

Technical Field

[0001] This invention relates to the field of aero-engine health monitoring and fault diagnosis technology, specifically to an aero-engine fault diagnosis method based on a conditional physical perception diffusion model. Background Technology

[0002] As a core component of an aircraft's power system, the performance and reliability of an aero-engine directly determine the overall performance and flight safety of the aircraft. To improve flight safety and optimize maintenance costs, accurate health assessments and fault diagnosis of engines are crucial. With the rapid development of big data and artificial intelligence technologies, data-driven methods are showing a strong application trend in the field of aero-engine diagnostics. Real-time monitoring and intelligent analysis of engine operating data through machine learning, deep learning, and other methods enable early warning of faults, accurate diagnosis, and predictive maintenance. This intelligent diagnostic technology based on massive amounts of operating data not only significantly improves engine reliability and safety and reduces maintenance costs, but also provides important support for optimized engine design and full lifecycle management.

[0003] However, acquiring aero-engine data faces severe challenges in actual operational scenarios, mainly in the following aspects: 1. Difficulty in Data Acquisition. High testing costs, stringent safety regulations, and complex and variable flight environments make obtaining sufficient high-quality training data extremely difficult. The scarcity of fault data is particularly pronounced; due to the high reliability design and rigorous maintenance procedures of aero engines, real-world fault cases are relatively rare, leading to a widespread problem of small sample sizes for diagnostic systems.

[0004] 2. Impact of Small Sample Size. This imbalanced data and scarce sample size severely restrict the training effectiveness of machine learning models, making diagnostic methods prone to overfitting. They also exhibit insufficient generalization ability when faced with new fault modes or unseen operating conditions, resulting in a significant drop in diagnostic accuracy. The small sample size also increases the model's sensitivity to noise, making it difficult to effectively distinguish between normal fluctuations and true fault signals, thus affecting the reliability and practicality of intelligent diagnostic systems for aero-engines in real-world engineering applications.

[0005] To address the problem of small sample sizes, existing technologies mainly include the following categories of methods: (1) Traditional data augmentation methods. Methods such as SMOTE (Synthetic Minority Oversampling Technique) effectively alleviate the class imbalance problem by generating synthetic samples through linear interpolation between minority class samples. Statistical models such as GMM (Gaussian Mixture Model) are based on probability distribution assumptions and achieve data augmentation through parameter estimation and random sampling. These methods are computationally simple and easy to implement, and are widely used in the field of traditional machine learning. However, these methods are often based on simplified distribution assumptions, making it difficult to capture complex data structures and nonlinear relationships, and limiting the realism and diversity of the generated samples.

[0006] (2) Deep Generative Model Methods. With the rapid development of deep learning technology, neural network-based generative models have demonstrated powerful data modeling capabilities. Variational Autoencoders (VAEs) learn the latent representation of data through an encoder-decoder architecture; Generative Adversarial Networks (GANs) utilize adversarial training mechanisms to achieve high-quality sample generation; Flow Models accurately model data distribution by constructing reversible neural network transformations; Conditional Generative Adversarial Networks (cGANs) methods control the generation process by introducing conditional information to achieve the embedding of specific working conditions to address the problem of insufficient samples under multiple working conditions.

[0007] However, existing methods have the following key technical shortcomings: First, traditional generative models neglect physical constraints. Existing research shows that traditional generative models primarily learn data distribution from a statistical perspective, ignoring the physical laws governing engineering systems. Aero-engines, as complex thermodynamic systems, are subject to strict physical constraints on their operational data, including thermodynamic laws, fluid mechanics principles, the law of conservation of energy, and material mechanical properties. For example, when applying GANs to generate data, it has been found that while the generated samples may have similar statistical characteristics, they violate the second law of thermodynamics, especially in the initial stages of training the generative adversarial network, where the generated data often deviates significantly from physical constraints. The compressor characteristic curves generated by VAEs exhibit physically impossible pressure-flow ratio combinations. This lack of physical constraints in traditional generative models results in statistically realistic but physically unreasonable generated data, making them unsuitable for applications in the highly safety-critical field of aero-engines.

[0008] Second, the physical embedding methods are unreasonable. Existing research has attempted to combine physical knowledge with generative models, such as Physics-Informed Neural Networks (PINNs) by adding physical equation constraints to the loss function, PGNNs by guiding network architecture design through prior knowledge, and variational autoencoders based on physical regularization. However, multiple empirical studies have revealed inherent flaws in these methods: conflicts between posterior constraints and data fitting terms lead to training instability; the weights of physical regularization terms are difficult to adjust—too strong leads to underfitting, too weak leads to constraint failure; and static constraints cannot adapt to the dynamic evolution of the generative process. These studies indicate that static, posterior-based physical constraint embedding methods have fundamental flaws: physical knowledge and data learning are antagonistic rather than collaborative. This passive constraint mechanism often results in both failure to guarantee physical consistency and compromised generative quality.

[0009] Third, the limitations of the uniform constraint handling strategy. Further analysis reveals that the aforementioned physical embedding methods all employ a uniform constraint handling strategy, such as applying the same energy conservation constraint weights to all fault types and using fixed physical regularization coefficients for different component failures. However, research on aero-engine failure mechanisms shows significant differences in the physical behavior of different fault types. For example, compressor stall primarily violates pressure monotonicity constraints, while combustion chamber coking mainly manifests as temperature-flow coupling imbalance; the physical characteristics of turbine blade cracks are completely different from those of bearing wear. The uniform, static physical constraint handling strategy of traditional methods ignores the significant differences in the violation patterns of physical constraints among different fault types and different engine components. It cannot dynamically adjust the importance weights of each physical constraint according to the fault type and component characteristics, resulting in fault samples lacking fault-specific physical behavior characteristics and failing to accurately reflect the complex physical mechanisms under real fault conditions.

[0010] Fourth, limitations of diffusion models. In recent years, diffusion models have demonstrated excellent generation quality and stability in fields such as image generation due to their denoising mechanisms. However, their direct application to aero-engine data generation still has significant limitations. First, traditional diffusion models are mainly based on data-driven probabilistic modeling and lack deep integration of domain-specific physical knowledge. Second, the standard Gaussian noise diffusion process cannot reflect the physical characteristics of aero-engine fault evolution. For example, fault characteristics such as abnormal temperature rise, pressure mutation, and vibration frequency shift often have specific physical patterns rather than simple random noise. Furthermore, the iterative denoising process of diffusion models is computationally expensive and lacks interpretability, making it difficult to trace the physical rationality of the generated samples.

[0011] In summary, to address the aforementioned technical shortcomings, it is necessary to develop an intelligent generation model that can deeply integrate physical knowledge, ensuring both the statistical authenticity and physical rationality of the generated data, while dynamically adjusting the physical constraint weights according to different fault types, in order to solve the small sample problem in the diagnosis of aero-engine gas path faults. Summary of the Invention

[0012] This invention addresses the shortcomings of existing technologies by proposing a fault diagnosis method for aero-engines based on a conditional physical perception diffusion model. It deeply embeds the physical laws of aero-engines into the generation process of the diffusion model and designs an adversarial physical perception learning mechanism and an adaptive constraint weight adjustment strategy. This fundamentally solves the key technical problems of traditional generative models, such as ignoring physical constraints, unreasonable physical embedding methods, and a single constraint processing strategy.

[0013] According to a first aspect of the embodiments of this application, a method for diagnosing aero-engine faults based on a conditional physics-aware diffusion model is provided, comprising: Step S1: Obtain the operating status monitoring parameters and performance monitoring parameters of the aero-engine during operation, and standardize the data; Step S2: Based on the thermodynamic characteristics of the aero-engine system, construct the physical constraint system of the aero-engine; Step S3: Construct a conditional physical perception diffusion model, wherein in the forward process of the physical perception of the conditional physical perception diffusion model, noise is added and physical violation terms are injected. In the backward process, the fault type label is used as a condition, and noise removal and physical violation term correction are performed through the noise prediction branch and physical correction branch of the dual-branch network, respectively. The dual-branch network is pre-trained based on the physical constraint system. Step S4: The conditional physical perception diffusion model is trained using an adversarial physical perception learning approach, wherein a time-dependent scheduling mechanism is used to gradually increase the weight of physical constraints. Step S5: Based on the working condition monitoring parameters and performance monitoring parameters, extract physical perception features using the intermediate layer representation of the conditional physical perception diffusion model, and use a pre-trained fault diagnosis classifier to realize the fault diagnosis of the aero-engine.

[0014] Furthermore, in step S2, the physical constraint system of the aero-engine includes: (2.1) Temperature monotonicity constraint: , in For the sigmoid function, For the first There are N temperature measurement points; (2.2) Pressure-temperature coupling constraint: , Where P and Temp represent pressure and temperature, respectively; (2.3) Energy conservation constraint: , in , "in" and "out" represent input and output, respectively. Indicates the first One sample, Indicates the number of input samples. Indicates the number of output samples. This represents the set of indices for the input samples. This represents the set of indices for the output samples. Used to prevent the denominator from being zero.

[0015] Furthermore, in step S3, the forward process of physical perception in the conditional physical perception diffusion model includes: Standard diffusion noise addition: ,in Standard Gaussian noise, This is the cumulative variance coefficient. In the first Noise addition factor in each diffusion step; Physical violation injection: ,in: For time-dependent physical violation strength scheduling functions, The initial time dependence coefficient, The total number of diffusion steps, The decay exponent, For structured physical violation vectors, including temperature inversion perturbations Pressure-temperature decoupling disturbance and energy imbalance disturbance , The monotonicity is disrupted by localized temperature reversals. By enhancing the inverse correlation between pressure and temperature, the coupling relationship is disrupted. The law of conservation of energy is violated by amplifying the energy difference between input and output.

[0016] Furthermore, in step S3, the reverse process of the conditional physical perception diffusion model is as follows: In each denoising step, denoising is performed simultaneously using the noise prediction branch and the physical correction branch of the dual-branch network: , in For noise terms, The variance coefficient, For the physical violation vector, the noise prediction branch Physical correction branch They are used to predict standard Gaussian noise, respectively. and physical violation vector c represents the fault type condition information, t represents the propagation time step, and θ represents the network parameters.

[0017] Furthermore, the dual-branch network is a pre-trained neural network, and the total loss function during pre-training is: , hyperparameters and For loss weights; Denoising loss ; Physical correction loss ; Weighted physical constraint loss ,in , , These are the temperature monotonicity constraint, pressure-temperature coupling constraint, and energy conservation constraint in the physical constraint system, respectively.

[0018] Further, in step S4, the total training loss function of the conditional physical perception diffusion model is: , in, and The denoising loss and physical correction loss are for the two-branch network. For loss weights; The adaptive physical constraint loss is obtained through a hierarchical condition injection mechanism and an adaptive constraint weight learning mechanism based on cross-attention. The hierarchical condition injection mechanism includes time step embedding and fault category label embedding. The time step embedding uses sinusoidal position encoding to map discrete time steps to a continuous vector space. The fault category labels are implemented using the embedding matrix of the science department. The embedded time steps and fault category labels are concatenated along the feature dimension to obtain the condition vector. The adaptive constraint weight learning specifically involves using the condition vector as the query and the physical constraints of the aero-engine as the key and value, and adaptively assigning attention weights through an attention mechanism. Thus obtain ,in For the i-th physical constraint loss, These are the corresponding attention weights; Physical constraint weights A time-dependent cosine annealing scheduling strategy is used to dynamically adjust the weights of physical constraints. and These are the maximum and minimum values ​​of the predetermined physical constraint weights, respectively.

[0019] Furthermore, in step S5, the fault diagnosis classifier adopts a multi-layer fully connected neural network structure, and its pre-training uses a multi-task learning architecture to simultaneously optimize the classification loss and physical consistency loss. The loss function is: , in, The standard cross-entropy loss is used for fault type identification. Physical consistency loss is calculated based on the predicted category. This is the input to the fault diagnosis classifier. For the predicted fault diagnosis results, This is the true value for fault diagnosis.

[0020] According to a second aspect of the embodiments of this application, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the method described in the first aspect.

[0021] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0022] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0023] The technical solutions provided by the embodiments of this application may include the following beneficial effects: Physical consistency guarantee of the generative model: By mathematizing and deeply integrating the physical laws of aero-engines, such as temperature monotonicity, pressure-temperature coupling, and energy conservation, into the diffusion model, the problem that traditional generative models cannot guarantee the physical rationality of generated data is fundamentally solved. The generated samples satisfy both statistical distribution and physical constraints.

[0024] Active physical embedding mechanism: An adversarial physical perception learning method is designed. By introducing structured physical violation noise during the diffusion process and designing a dual-branch prediction network, combined with a time-dependent scheduling mechanism, the conflict between physical knowledge and data-driven learning is resolved. This enables the active embedding and progressive recovery of physical constraints, avoiding the training instability and constraint failure problems of posterior constraint methods.

[0025] Fault-specific modeling capability: By dynamically learning the importance weights of each physical constraint under different fault types through an attention mechanism, the limitations of the traditional method of uniform constraint processing are solved. This enables the model to adaptively adjust physical constraints for different faults (such as compressor stall, turbine blade cracks, combustion chamber coking, etc.), and the generated samples more accurately reflect the fault-specific physical behavior characteristics.

[0026] Improved diagnostic performance with few samples: Under conditions of very few samples (only 50 per class), the diagnostic method based on the Conditional Physics-Aware Diffusion Model (CPADM) achieved an accuracy of 93%, which is a significant improvement compared to traditional methods (SVM 68%, Random Forest 57%), a 20 percentage point improvement compared to deep generative methods (GAN 73%), and an 18 percentage point improvement compared to meta-learning methods (Meta-Learning 75%). This fully demonstrates the effectiveness of physical embedding for learning with few samples.

[0027] Superior generation quality: The physical deviation of the generated data reaches 0.342 (significantly lower than cGAN's 0.587 and VAE's 0.523), the Wasserstein distance is 9818 (close to the optimal value), and the correlation preservation reaches 0.9548, which are comprehensively better than existing generative models, proving that CPADM effectively reduces physical constraint violations while maintaining the statistical properties of the data.

[0028] Engineering application value: It provides an interpretable and physically consistent data augmentation method, and the physical rationality of the generated samples is traceable, which meets the stringent safety and reliability requirements in the field of aero-engines. It also provides a reference solution for the small sample learning problem of other complex engineering systems.

[0029] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0031] Figure 1 This is a flowchart of the aero-engine fault diagnosis method based on the conditional physical sensing diffusion model of the present invention.

[0032] Figure 2 This is a schematic diagram of the physical constraints of an aero-engine.

[0033] Figure 3 This is a schematic diagram of the forward process of physical perception in the conditional physical perception diffusion model.

[0034] Figure 4 Framework diagram for fault diagnosis applications.

[0035] Figures 5-10 The charts show a comparison of the generated data quality for CPADM, SMOTE, RealNVP, cGAN, GMM, and cPGAN, including t-SNE visualization of samples generated by different methods and a comparison of physical constraint violations.

[0036] Figures 11-16 The fault diagnosis confusion matrix diagram shows the performance comparison of CPADM with SVM, Random Forest, GAN, MetaLearning, and XGBOOST on different evaluation metrics.

[0037] Figure 17 This is a block diagram of the aero-engine fault diagnosis device based on the conditional physical sensing diffusion model of the present invention.

[0038] Figure 18 This is a schematic diagram of an electronic device. Detailed Implementation

[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0040] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0041] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0042] Reference Figures 1 to 12 The present invention will be further described below.

[0043] A fault diagnosis method for aero-engines based on a conditional physics-aware diffusion model, such as Figure 1 As shown, it includes the following steps: Step S1: Obtain the operating status monitoring parameters and performance monitoring parameters of the aero-engine during operation, and standardize the data; To monitor and diagnose faults in the complex operation of aero-engines, it is necessary to collect relevant parameters during the operation of the aero-engine, including but not limited to operating condition monitoring parameters (flight altitude, Mach number, throttle lever angle), performance monitoring parameters (flight altitude, Mach number, throttle lever angle, fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, bypass duct total pressure, etc.), and to standardize the collected data.

[0044] Step S2: Based on the thermodynamic characteristics of the aero-engine system, construct the physical constraint system of the aero-engine; Systematic mathematical modeling of the constraints of complex thermodynamic systems in aero-engines is conducted to construct a multi-level, multi-scale physical constraint system, such as... Figure 2 As shown, it specifically includes three types of physical constraints: (1) Temperature Monotonicity Constraint: Based on the second law of thermodynamics, the temperature in the engine should increase along the airflow direction. The temperature sequence is defined as follows: ,in Represents the i-th temperature measurement point, such as Figure 2 As shown, the temperature sequence represents the temperature measurement points of various components along the engine's airflow path. Temperature monotonicity violation. Defined as:

[0045] in This is the sigmoid function, used to ensure the differentiability of the constraints. The corresponding partial differential equation is of the form: This enables the constraint extension from discrete measurement points to continuous fields.

[0046] (2) Pressure-temperature coupling constraint: According to the ideal gas law and the principle of isentropic processes, the pressure P and temperature in a thermodynamic system are coupled. The relationship should be positively correlated. Define the coupling violation degree. for:

[0047] The coupling strength is quantified by the correlation coefficient to ensure that the generated data satisfies the basic thermodynamic relationship.

[0048] (3) Energy conservation constraint: Based on the principle of enthalpy balance, the input and output energies of the system should remain in balance. Define the degree of energy violation. for:

[0049] in , "in" and "out" represent input and output, respectively, ensuring that the system's input and output energy remain balanced. Indicates the first One sample, Indicates the number of input samples. Indicates the number of output samples. This represents the set of indices for the input samples. This represents the set of indices for the output samples. This represents a small constant introduced to prevent the denominator from being zero.

[0050] Step S3: Construct a conditional physical perception diffusion model, wherein in the forward physical perception process of the conditional physical perception diffusion model, such as... Figure 3 As shown, noise is added and physical violation terms are injected. In the reverse process, the fault type label is used as a condition, and noise removal and physical violation term correction are performed through the noise prediction branch and physical correction branch of the dual-branch network, respectively. The dual-branch network is pre-trained based on the physical constraint system. Building upon the forward process of the standard diffusion model, this paper innovatively designs a physical perception forward process, introducing structured physical violation terms while adding noise during diffusion, thus constructing an adversarial learning mechanism. The forward process of the traditional diffusion model is as follows:

[0051] The physical sensing forward process of this invention includes two steps: (1) Standard diffusion noise addition: ,in Standard Gaussian noise, This is the cumulative variance coefficient. In the first Noise addition coefficients (variance scheduling parameters) in each diffusion step.

[0052] (2) Physical violation injection: ,in For the physical violation vector, This is a time-dependent injection intensity coefficient.

[0053] Physical violation vector The construction is based on the three types of physical constraints defined in step 1, and adopts a structured design:

[0054] in The monotonicity is disrupted by localized temperature reversals. By enhancing the inverse correlation between pressure and temperature, the coupling relationship is disrupted. The conservation law is violated by amplifying the energy difference between input and output. (Time-dependent coefficient) Designed as follows: ,in The initial time dependence coefficient, The total number of diffusion steps, The decay exponent ensures that a strong physical violation is applied in the early stages of diffusion, which gradually weakens over time.

[0055] When backsampling to generate new data using the conditional physical sensing diffusion model, a physical guidance mechanism is designed to ensure the physical rationality of the generated samples: (1) Two-branch joint prediction: In each denoising step, both the noise prediction branch and the physical correction branch are used simultaneously:

[0056] in For noise terms, The variance coefficient, This is the physical violation vector.

[0057] The dual-branch network simultaneously learns two tasks: noise prediction and physics correction, enabling the active embedding of physical knowledge. ① Noise prediction branch: Predicted standard Gaussian noise The denoising objective follows the traditional diffusion model.

[0058] ②Physical Correction Branch: Predicting physical violation vectors Learn to identify and correct violations of physical constraints.

[0059] Where c is the fault type label, t is the diffusion time step, and θ is the network parameter. The noise prediction branch is implemented based on the U-Net architecture, including an encoder, bottleneck layer, and decoder, and introduces residual connections and attention mechanisms at each scale, using residual connections and skip connections to maintain feature transfer; the physics correction branch adopts a multilayer perceptron structure, with the input being the intermediate features of the encoder and the output being the predicted value of the physics violation; the two branches share encoder parameters to achieve feature reuse and collaborative learning, ensuring a deep integration of physical knowledge and data-driven learning. The dual-branch network is first pre-trained under the physical constraint system to obtain preliminary physical perception capabilities; then, as the core component of the diffusion model, it is optimized synchronously with the diffusion process during the adversarial physical perception learning stage, achieving end-to-end joint training.

[0060] For the dual-branch network, a multi-objective joint training strategy is designed to collaboratively optimize three objectives: noise prediction, physical correction, and physical consistency through a composite loss function. Total loss function for a two-branch network:

[0061] in: Hyperparameters and The relative importance of physical constraints and physical corrections is controlled separately, and a time-dependent dynamic scheduling strategy is adopted: In the early stages of training, emphasis is placed on physical constraints, while in the later stages, the data fitting ability is gradually enhanced. The training step (equivalent to the "time step" index) is incremented from 0; This represents the maximum number of steps that can be scheduled. The shape parameter for attenuation (>0).

[0062] Denoising loss This ensures that the model can accurately predict the added Gaussian noise.

[0063] Physical correction loss This ensures that the model can accurately identify and predict physical violations.

[0064] Weighted physical constraint loss The physical consistency of the data after denoising is constrained, and will be explained in detail below.

[0065] To address the differences in physical constraint violation patterns across different fault types, an adaptive attention mechanism is designed to dynamically learn constraint weights. (1) Fault type coding: Embedding conditions The fault feature space is mapped through a multilayer perceptron.

[0066] (2) Calculation of constraint weights: ,in Let be a learnable query matrix, where This is the normalized weight vector for the three types of physical constraints.

[0067] (3) Weighted physical consistency loss:

[0068] in , , The violation loss represents the penalty for each of the three types of physical constraints.

[0069] In this way, the model can automatically adjust the importance of physical constraints according to different fault types, achieving fault-specific physical modeling.

[0070] Step S4: The conditional physical perception diffusion model is trained using an adversarial physical perception learning approach, wherein a time-dependent scheduling mechanism is used to gradually increase the weight of physical constraints. Specifically, such as Figure 4 As shown, this step may include: (4.1) Hierarchical Conditional Injection Mechanism: The hierarchical conditional injection mechanism includes time step embedding and fault category label embedding. Time step embedding uses sinusoidal positional coding to map discrete time steps to a continuous vector space:

[0071]

[0072] in For time steps, For embedded dimensions, For dimensional indexing.

[0073] Fault category embedding is achieved through a learnable embedding matrix. accomplish:

[0074] in For fault category labels, and These are learnable transformation parameters.

[0075] Conditional fusion concatenates and integrates time and fault information along the feature dimension into a unified conditional vector. This vector serves two purposes in subsequent modules: (i) as the modulation signal for the Adaptive Normalization Layer (AdaGN), injected into the features of each layer of the dual-branch network; and (ii) as the query vector for the cross-attention module, used to dynamically learn the weights of physical constraints. In this way, the model can achieve dynamic adjustment of physical consistency in the time-conditional joint space.

[0076] (4.2) Adaptive constraint weight learning adopts a cross-attention mechanism: In adaptive constraint weight learning, the joint condition vector As a query, it enables dynamic physical constraint modeling based on fault type and time state. Physical constraints are represented as keys and values, and the attention mechanism is based on... Adaptively assign attention weights based on the correlation between the physical constraint features and the physical constraint features. The final adaptive physics constraint loss is: ,in For the i-th physical constraint loss, These are the corresponding attention weights.

[0077] (4.3) Time-dependent scheduling mechanism and overall loss function: To balance the denoising stability of the diffusion model in the early training stage with the physical consistency in the later stage, a time-dependent cosine annealing scheduling strategy is adopted to dynamically adjust the physical constraint weights.

[0078]

[0079] and These are the predetermined maximum and minimum physical constraint weights, respectively. During training, the weights gradually approach the maximum value from the minimum. Early stage ( big): The noise level is relatively low, so rough noise reduction is mainly performed; in the mid-to-late stages ( Small): The value is relatively large, so physical corrections are the primary focus. Therefore, the final loss function is:

[0080] Step S5: Based on the working condition monitoring parameters and performance monitoring parameters, extract physical perception features using the intermediate layer representation of the conditional physical perception diffusion model, and use a pre-trained fault diagnosis classifier to realize the fault diagnosis of the aero-engine. Specifically, fault diagnosis is performed using a trained CPADM model, including: (1) Physical perception feature extraction: from the intermediate layer representation of CPADM (in (At any given moment) feature vectors containing rich physical semantics are extracted. These features implicitly contain physical constraint information such as temperature monotonicity, pressure-temperature coupling, and energy conservation.

[0081] It should be noted that during the training process of CPADM, for small sample fault categories, especially under the extreme condition of fewer than 100 samples per fault category, CPADM can also be used to generate high-quality augmented samples to expand the training set and achieve better training results.

[0082] (2) End-to-end diagnosis: The extracted physical perception features are input into the fault diagnosis classifier (such as a fully connected neural network) to realize fault type identification.

[0083] A fault diagnosis classifier is constructed based on the extracted physical perception features, employing a multi-layer fully connected neural network structure: It includes three fully connected layers with dimensions [d_feature, 512, 256, n_classes], where d_feature is the dimension of the fused features and n_classes is the number of fault types. A BatchNorm normalization layer and a ReLU activation function are added after each fully connected layer. A Dropout layer with a dropout rate of 0.3 is added after the first and second layers to prevent overfitting. Finally, a softmax layer outputs the probability distribution of each fault type.

[0084] in These are learnable parameters.

[0085] The diagnostic process simultaneously calculates the physical violation of the sample:

[0086] in This is a module defined by physical constraints. The final diagnostic decision considers classification probability and physical consistency.

[0087] The fault diagnosis classifier employs a multi-task learning architecture, simultaneously optimizing both classification loss and physical consistency loss, with the following loss function:

[0088] in The standard cross-entropy loss is used for fault type identification. The physical consistency loss is calculated based on the predicted categories to ensure the physical plausibility of the diagnostic results. This is the input to the fault diagnosis classifier. For the predicted fault diagnosis results, This is the true value for fault diagnosis.

[0089] This embodiment uses an aero-engine air path diagnostic benchmark dataset for validation. This dataset is based on a real physical system and includes comprehensive uncertainties to realistically represent and simulate all major sources of uncertainty in air path diagnostics on a common simulation and testing platform. These uncertainties include sensor noise and bias, manufacturing-induced individual variability, modeling errors, performance degradation, and uncertainties caused by data collection discreteness. BD-AGD includes four noise types, with different engines containing different noise categories, totaling four types (no noise (Type 0), white Gaussian noise (Type 1), near-realistic noise (Type 2), and Type 1 + incremental drift (Type 3)). Each engine contains four types of air path components that may fail: fan, compressor, high-pressure turbine, and low-pressure turbine. Only one type of failure occurs per flight, and it occurs before takeoff. The severity of the failure does not change during flight, and all components of the same type have the same failure severity.

[0090] For the study of EngineA, the noise2 subset was selected, which adds noise and makes reconstruction and diagnosis more difficult. The training set and test set were constructed by random sampling, and the sample size is shown in Table 1.

[0091] Table 1. Dataset Description

[0092] The specific implementation steps are as follows: Step 1: Data Preprocessing and Physical Property Analysis (2) Normalization: Z-score standardization is performed on each sensor parameter separately, with the formula x'=(x-μ) / σ, where μ is the mean and σ is the standard deviation, to ensure that parameters of different dimensions are comparable.

[0093] (3) Physical feature recognition: Based on the engine working principle, identify sensor parameters related to the three types of physical constraints: - Temperature sequence: T12 (fan inlet temperature) → T24 (LPC outlet temperature) → T30 (HPC outlet temperature) → T50 (LPT outlet temperature) - Pressure-temperature pair: (Ps30, T30) indicates the pressure-temperature coupling at the HPC outlet. - Energy parameters: fuel flow rate Wf, total temperature Tt, mass flow rate ṁ, etc. (4) Calculation of physical constraint violation: Calculate the violation degree of the three types of physical constraints for normal samples and faulty samples in the training set respectively, and establish a baseline physical behavior model: - Normal sample: =0.012, =0.008, =0.015 - HPC failure: =0.089, =0.156, =0.023 (most severe pressure-temperature coupling violation) - HPT malfunction: =0.134, =0.067, =0.091 (Most severe violation of temperature monotonicity) - FAN malfunction: =0.045, =0.038, =0.112 (Most serious violation of energy conservation) The above statistical results verify the differences in physical constraint violation modes among different fault types, providing a basis for subsequent adaptive weight learning.

[0094] Step 2: CPADM Model Construction and Training (1) Design of the forward process of physical perception: Based on the standard diffusion noise, a structured physical violation term is introduced. By constructing directional violation modes at specific locations, the model learns the recovery process from a physical violation state to a physical consistency state. Specifically, the physical violation term design includes: Temperature violation: δtemp breaks monotonicity by randomly selecting adjacent temperature measurement points for exchange, and the violation intensity follows a uniform distribution U(0.1, 0.3); Pressure-temperature violation: δPT reduces the correlation coefficient by adding reverse perturbations to pressure and temperature respectively, and the perturbation amplitude is σP=0.2·std(P) and σT=0.2·std(T); Energy violation: δenergy amplifies the input-output energy difference, and the violation ratio is U(0.05, 0.15); Time dependence coefficient: λt =0.5·(t / T)^2, the physical violation is stronger in the early stage (t<300) and gradually decays in the later stage.

[0095] (2) Constructing a dual-branch prediction network: Design a neural network architecture that includes a noise prediction branch and a physical correction branch. The noise prediction branch is implemented based on the U-Net architecture, including an encoder, a bottleneck layer, and a decoder. Residual connections and skip connections are used to maintain feature transfer. The physical correction branch adopts a multilayer perceptron structure. The input is the intermediate features of the encoder, and the output is the predicted value of the physical violation. The two branches share encoder parameters to achieve feature reuse and collaborative learning. Specifically, the number of diffusion steps is T=1000, with linear noise scheduling and βt increasing linearly from 0.0001 to 0.02. The U-Net encoder has 4 downsampling blocks with [64, 128, 256, 512] channels. The U-Net decoder has 4 upsampling blocks with [512, 256, 128, 64] channels.

[0096] (3) Design a conditional injection mechanism: embed the fault type label as conditional information into the generation process of the diffusion model so that the model can learn the unique physical behavior patterns and data distribution characteristics under each fault type. Conditional embedding dimensions: fault type label is embedded as a 128-dimensional vector and time is embedded as a 128-dimensional vector.

[0097] (4) Implement adaptive constraint weight learning: The importance weight of each physical constraint under different fault types is dynamically learned through the attention mechanism, and differentiated modeling is performed for the physical characteristics of different faults. Attention mechanism: Multi-head self-attention is introduced in layers with 256 and 512 channels, with 8 heads.

[0098] Step 3: Data Generation and Quality Assessment (1) Sampling parameter settings: - Number of sampling steps: Using DDIM to accelerate sampling, the actual number of sampling steps is 50. - Gradient guidance: Gradient correction is introduced at t=500, with a step size η=0.01. - Physical correction strength: λ phy =0.5 - Number of samples generated: 200 samples are generated for each fault type, expanding the training set to 250 (can be arbitrarily specified) samples / classes.

[0099] (2) Generate quality assessment indicators: The following five metrics are used to evaluate the quality of the generated data: a) Wasserstein distance (EMD): Measures the difference between the generated distribution and the true distribution; a smaller value is better. b) Kolmogorov-Smirnov test (KS): measures the consistency of the distribution; the smaller the value, the better. c) Correlation Preservation: Measures the preservation of correlation between features; a value closer to 1 is better. d) Physics Deviation: Measures the degree of violation of physical constraints; a smaller value is better. Step 4: Compare with baseline methods: The following six methods were selected as baselines: SMOTE (classical oversampling method); GMM (Gaussian mixture model); VAE (variable autoencoder); cGAN (conditional generative adversarial network); RealNVP (real-valued non-volume-preserving flow model); and cPGAN (conditional GAN ​​with physical regularization). Experimental results are shown in Table 2.

[0100] Table 2 Comparison of generated data quality

[0101] This application employs four complementary evaluation metrics to comprehensively assess the performance of each generation method: (1) Kolmogorov-Smirnov (KS) test statistic, used to quantify the difference between the generated distribution and the true distribution; (2) Wasserstein distance, which measures the optimal transmission cost between two probability distributions; (3) correlation preservation, which assesses the fidelity of the coupling relationship between sensors; and (4) physical deviation, which quantifies the degree to which the generated samples violate thermodynamic constraints.

[0102] The CPADM method demonstrated superior overall performance, achieving an optimal value of 0.1387 in the KS test, significantly lower than other methods, indicating the highest good fit between its generated distribution and the true distribution. Its Wasserstein distance reached 9818, second only to the GMM method, proving its excellent performance in probability distribution matching. The correlation preservation was 0.9548, successfully maintaining the coupling relationship between sensors. The physical deviation of 0.342 was globally optimal, on par with cPGAN. Figures 5-10 The t-SNE visualization reveals that the data points generated by CPADM exhibit ideal distribution characteristics: the data points are evenly and continuously distributed, without unnatural gaps or over-concentration; multiple reasonable clustering structures are formed, with clear boundaries and appropriate transition regions between clusters; the point cloud density shows a natural decrease from the cluster center to the edge, consistent with the distribution pattern of real data. These results fully demonstrate that CPADM successfully achieves a balance between statistical realism and physical plausibility by gradually applying physical constraints during the diffusion process.

[0103] Step 4: Fault Diagnosis Experiment (1) Feature extraction: Extract physical perception features from the trained CPADM model. Specifically, during the backsampling process, when t=T / 2=500, extract the feature vector (512-dimensional) of the bottleneck layer of U-Net.

[0104] (2) Classifier training: The fully connected neural network classifier was trained using the extracted 512-dimensional features. The architecture was: input layer (512-dimensional) → hidden layer 1 (256-dimensional, ReLU, Dropout 0.3) → hidden layer 2 (128-dimensional, ReLU, Dropout 0.2) → output layer (5-dimensional, Softmax). Training parameters: optimizer Adam, learning rate 1e-3, batch size 64, training for 100 epochs, and loss function is cross-entropy loss.

[0105] (3) Baseline method comparison: The following six methods were selected as diagnostic baselines: SVM (Support Vector Machine): a classic machine learning method with RBF kernel function; Random Forest: an ensemble learning method with 100 trees; XGBoost: a gradient boosting tree method with a tree depth of 6 and a learning rate of 0.1; Meta Learning: a few-shot learning method based on MAML; GAN+Classifier: a classifier trained after generating data using cGAN; SMOTE+Classifier: a classifier trained after augmenting data using SMOTE.

[0106] (4) Evaluation indicators: The diagnostic performance is evaluated using four indicators: accuracy, precision, recall and F1 score. All indicators are calculated using macro averaging.

[0107] Experimental results are as follows Figures 11-16 As shown in Table 2: Table 2 Comparison of Fault Diagnosis Results

[0108] from Figure 11 Figure 16Table 2 shows that the CPADM method achieved the best performance across all evaluation metrics, with an accuracy of 93.33%, representing an improvement of 18.23 percentage points compared to the second-place Meta Learning (75.10%) and 25 percentage points compared to the traditional machine learning method SVM (68.33%). This fully demonstrates: a) the effectiveness of physical embedding: compared to GAN+Classifier (73.33%), CPADM improved by 20 percentage points, indicating that physical constraints significantly improve the quality of generated data and diagnostic performance. b) the necessity of the conditional mechanism: CPADM can generate specific samples for different fault types, avoiding class confusion. c) the superiority of physical perception features: the extracted intermediate layer features incorporate physical semantics, making them more discriminative than the original sensor data.

[0109] Corresponding to the aforementioned embodiments of the aero-engine fault diagnosis method based on the conditional physical perception diffusion model, this application also provides embodiments of an aero-engine fault diagnosis device based on the conditional physical perception diffusion model.

[0110] Figure 17 This is a block diagram of an aero-engine fault diagnosis device based on a conditional physical sensing diffusion model, according to an exemplary embodiment. (Refer to...) Figure 17 The device may include: The data acquisition module 21 is used to acquire the working status monitoring parameters and performance monitoring parameters of the aero-engine during operation, and to standardize the data. The physical constraint system construction module 22 is used to construct the physical constraint system of the aero-engine based on the thermodynamic system characteristics of the aero-engine; Model building module 23 is used to build a conditional physical perception diffusion model. In the forward process of the physical perception of the conditional physical perception diffusion model, noise is added and physical violation terms are injected. In the backward process, the fault type label is used as a condition, and noise removal and physical violation term correction are performed through the noise prediction branch and physical correction branch of the dual-branch network, respectively. The dual-branch network is pre-trained based on the physical constraint system. The model training module 24 is used to train the conditional physical perception diffusion model using an adversarial physical perception learning approach, wherein a time-dependent scheduling mechanism is used to gradually increase the weight of physical constraints. The fault diagnosis module 25 is used to extract physical perception features based on working condition monitoring parameters and performance monitoring parameters, using the intermediate layer representation of the conditional physical perception diffusion model, and to realize the fault diagnosis of aero-engines using a pre-trained fault diagnosis classifier.

[0111] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0112] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0113] Accordingly, this application also provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the aircraft engine fault diagnosis method based on the conditional physical perception diffusion model as described above.

[0114] Accordingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for diagnosing aero-engine faults based on a conditional physics-aware diffusion model. Figure 18 The diagram shown is a hardware structure diagram of any data processing-capable device where a deep learning dataset access system is located, according to an embodiment of the present invention. Except for... Figure 18 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0115] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the aforementioned aero-engine fault diagnosis method based on a conditional physics-aware diffusion model. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0116] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A method for diagnosing aero-engine faults based on a conditional physics-aware diffusion model, characterized in that, include: Step S1: Obtain the operating status monitoring parameters and performance monitoring parameters of the aero-engine during operation, and standardize the data; Step S2: Based on the thermodynamic characteristics of the aero-engine system, construct the physical constraint system of the aero-engine; Step S3: Construct a conditional physical perception diffusion model, wherein in the forward process of the physical perception of the conditional physical perception diffusion model, noise is added and physical violation terms are injected. In the backward process, the fault type label is used as a condition, and noise removal and physical violation term correction are performed through the noise prediction branch and physical correction branch of the dual-branch network, respectively. The dual-branch network is pre-trained based on the physical constraint system. Step S4: The conditional physical perception diffusion model is trained using an adversarial physical perception learning approach, wherein a time-dependent scheduling mechanism is used to gradually increase the weight of physical constraints. Step S5: Based on the working condition monitoring parameters and performance monitoring parameters, extract physical perception features using the intermediate layer representation of the conditional physical perception diffusion model, and use a pre-trained fault diagnosis classifier to realize the fault diagnosis of the aero-engine.

2. The method according to claim 1, characterized in that, In step S2, the physical constraint system of the aero-engine includes: (2.1) Temperature monotonicity constraint: ， in For the sigmoid function, For the first There are N temperature measurement points; (2.2) Pressure-temperature coupling constraint: ， Where P and Temp represent pressure and temperature, respectively; (2.3) Energy conservation constraint: ， in , "in" and "out" represent input and output, respectively. Indicates the first One sample, Indicates the number of input samples. Indicates the number of output samples. This represents the set of indices for the input samples. This represents the set of indices for the output samples. Used to prevent the denominator from being zero.

3. The method according to claim 1, characterized in that, In step S3, the forward process of physical perception in the conditional physical perception diffusion model includes: Standard diffusion noise addition: ,in Standard Gaussian noise, This is the cumulative variance coefficient. In the first Noise addition factor in each diffusion step; Physical violation injection: ,in: For time-dependent physical violation strength scheduling functions, The initial time dependence coefficient, The total number of diffusion steps, The decay exponent, For structured physical violation vectors, including temperature inversion perturbations Pressure-temperature decoupling disturbance and energy imbalance disturbance , The monotonicity is disrupted by localized temperature reversals. By enhancing the inverse correlation between pressure and temperature, the coupling relationship is disrupted. The law of conservation of energy is violated by amplifying the energy difference between input and output.

4. The method according to claim 1, characterized in that, In step S3, the reverse process of the conditional physical perception diffusion model is as follows: In each denoising step, denoising is performed simultaneously using the noise prediction branch and the physical correction branch of the dual-branch network: ， in For noise terms, The variance coefficient, For the physical violation vector, the noise prediction branch Physical correction branch They are used to predict standard Gaussian noise, respectively. and physical violation vector c represents the fault type condition information, t represents the propagation time step, and θ represents the network parameters.

5. The method according to claim 1, characterized in that, The dual-branch network is a pre-trained neural network, and the total loss function during pre-training is: ， hyperparameters and For loss weights; Denoising loss ; Physical correction loss ; Weighted physical constraint loss ,in , , These are the temperature monotonicity constraint, pressure-temperature coupling constraint, and energy conservation constraint in the physical constraint system, respectively.

6. The method according to claim 1, characterized in that, In step S4, the total training loss function of the conditional physical perception diffusion model is: ， in, and The denoising loss and physical correction loss are for the two-branch network. For loss weights; The adaptive physical constraint loss is obtained through a hierarchical condition injection mechanism and an adaptive constraint weight learning mechanism based on cross-attention. The hierarchical condition injection mechanism includes time step embedding and fault category label embedding. The time step embedding uses sinusoidal position encoding to map discrete time steps to a continuous vector space. The fault category labels are implemented using the embedding matrix of the science department. The embedded time steps and fault category labels are concatenated along the feature dimension to obtain the condition vector. The adaptive constraint weight learning specifically involves using the condition vector as the query and the physical constraints of the aero-engine as the key and value, and adaptively assigning attention weights through an attention mechanism. Thus obtain ,in For the i-th physical constraint loss, These are the corresponding attention weights; Physical constraint weights A time-dependent cosine annealing scheduling strategy is used to dynamically adjust the weights of physical constraints. and These are the maximum and minimum values ​​of the predetermined physical constraint weights, respectively.

7. The method according to claim 1, characterized in that, In step S5, the fault diagnosis classifier adopts a multi-layer fully connected neural network structure. During pre-training, it employs a multi-task learning architecture, simultaneously optimizing the classification loss and physical consistency loss. The loss function is: ， in, The standard cross-entropy loss is used for fault type identification. Physical consistency loss is calculated based on the predicted category. This is the input to the fault diagnosis classifier. For the predicted fault diagnosis results, This is the true value for fault diagnosis.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the method as described in any one of claims 1-7.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

Images