Aero-engine test failure cross-domain migration diagnosis method based on continuous diffusion
By constructing a fractional network model of the source and target domains and using the fractional gradient field divergence degree discrimination in the continuous diffusion process, the problem of unlabeled fault signal distribution offset in aero-engine testing was solved, improving diagnostic accuracy and reliability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST FORESTRY UNIV
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies suffer from low diagnostic accuracy due to the distribution deviation of unlabeled fault signals under different operating conditions during aero-engine testing. Traditional methods lack robustness and cannot meet the requirements for high-reliability diagnosis.
A diagnostic method based on continuous diffusion score modeling and cross-domain divergence discrimination mechanism is adopted. By constructing a source domain fault condition score network and a target domain global score network, the divergence of the score gradient field in the continuous diffusion process is used to achieve label-free fault identification under unknown working conditions.
It significantly improves the accuracy of fault diagnosis and the reliability of classification decisions under varying operating conditions and strong noise environments, and achieves cross-domain feature distribution alignment under unknown operating conditions in the target domain.
Smart Images
Figure CN122132919A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aero-engine condition monitoring and intelligent fault diagnosis technology, specifically relating to an aero-engine test fault diagnosis method based on a continuous diffusion model and cross-domain transfer learning. Background Technology
[0002] As the most critical and complex core component of aviation equipment systems, the operational status of aero-engines directly affects flight safety and equipment reliability. During engine development, certification, and service, extensive ground testing is required to monitor and evaluate engine performance parameters and structural health. The large-scale vibration signals generated during testing contain rich information about the equipment's operational status, serving as a crucial data foundation for fault diagnosis and health management. However, due to the complex structure, extreme operating environment, and wide range of operating conditions of aero-engines, their signals exhibit strong nonlinearity, strong non-stationarity, and strong noise interference characteristics. This results in traditional diagnostic methods based on rule thresholds or shallow feature engineering lacking robustness in complex environments and failing to meet the requirements for high-reliability diagnostics.
[0003] Fault diagnosis methods based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely applied in the fields of rotating machinery and aero-engines. These methods typically rely on supervised learning training with large amounts of labeled data, achieving fault mode recognition through automatic feature extraction and classifier modeling. However, in real-world engineering scenarios, acquiring test data with complete fault labeling information is extremely costly, and the data distribution varies significantly across different engine models, test benches, and operating conditions, resulting in widespread data distribution bias and relatively insufficient model generalization and cross-environment adaptability. Transfer learning and domain adaptation methods can, to some extent, reduce the distribution difference between the source and target domains through feature alignment, distribution matching, or adversarial training, thereby improving the model's diagnostic capabilities under unknown operating conditions. However, existing transfer learning methods are mostly based on discriminative model frameworks, primarily achieving cross-domain adaptation through minimizing statistical distribution distance or feature space alignment. However, discriminative models are prone to getting trapped in local feature mapping under varying operating conditions, making it difficult to capture the global probability evolution of signals. Therefore, there is an urgent need for a diagnostic method that can characterize the signal distribution evolution across operating conditions in a continuous probability space and achieve label-free fault identification. Based on this, this paper proposes a cross-domain migration diagnosis method for aero-engine test faults based on continuous diffusion fractional modeling and cross-domain divergence discrimination mechanism. This method constructs a dual-path collaborative modeling system of a source domain fault condition fractional network and a target domain global fractional network, and utilizes the divergence of the fractional gradient field during the continuous diffusion process to achieve label-free fault identification under unknown operating conditions. It demonstrates strong engineering applicability and robustness in practical applications of aero-engine test fault diagnosis. Summary of the Invention
[0004] The purpose of this invention is to solve the problem of low diagnostic accuracy caused by the distribution deviation of unlabeled fault signals under different operating conditions during aero-engine testing. Therefore, a cross-domain migration diagnostic method based on continuous diffusion fraction modeling and cross-domain divergence discrimination mechanism is proposed.
[0005] The aforementioned objectives are primarily achieved through the following technical solutions:
[0006] S1. Preparation and time-frequency feature mapping of source and target domain signal data;
[0007] (1) The source domain data is a set of historical test signal samples collected under the given test conditions and for which complete fault labeling information has been obtained. The source domain data is used to support the matching and discrimination of the source domain fault condition fractional network model; the target domain data comes from the actual operating or test environment that differs from the source domain, and includes unlabeled signal data without fault category annotations. and the signal to be measured This is used to achieve cross-domain feature distribution alignment and target domain fault state identification;
[0008] (2) Both the source domain and target domain data need to be converted into two-dimensional time-frequency spectra by continuous wavelet transform or short-time Fourier transform during subsequent training as input to the U-Net structure.
[0009] S2. Train the conditional score network model for each fault category in the source domain using forward continuous diffusion noise-added data;
[0010] A forward continuous diffusion noise addition process is performed on the source domain fault category data using a predefined diffusion operator based on stochastic differential equations. The expression for this process is:
[0011] (1)
[0012] in, The drift coefficient, The diffusion coefficient is... This represents the standard Wiener process; by solving this equation, the original signal distribution can be smoothly transformed into a known Gaussian distribution.
[0013] (2) Establish a conditional fractional network model with a U-Net-type symmetric architecture containing skip connections, integrate residual convolutional blocks and dilated convolutions in the encoding path, and use them to extract long-range temporal dependence features of engine signals across scales; at the same time, embed a self-attention mechanism in the deep feature extraction of the network to capture the non-local coupling relationship in the time-frequency manifold of the signal.
[0014] (3) Learn the fractional gradient fields of each fault category at different diffusion time steps, and construct a set of source domain fault conditional fractional network models covering N known fault modes. The training objective is to minimize the following weighted score matching loss function:
[0015] (2)
[0016] This loss function forces the network to learn that the signal recovers to a specific fault mode under different noise levels. The optimal path.
[0017] S3. Train a global score network model for the target domain using unlabeled test signal data from the target domain;
[0018] Construct a global score network model for the target domain using the same network architecture as in step S2. Using unlabeled test signal data in the target domain Unsupervised training is performed to enable the model to learn the unconditional score distribution of the target domain in order to capture the background noise distribution manifold of a specific engine model under the current test environment.
[0019] S4. Cross-domain perturbation mapping and dual-path fractional gradient modeling of target domain signals;
[0020] right diffusion time step is The noise was obtained ;
[0021] (2) Parallel input target domain global score network model and source domain fault category conditional score network model;
[0022] (3) Calculate the environmental gradient of the target domain ;
[0023] (4) Calculate the theoretical gradient for each fault category. ,in Traversal Source domain faults; calculate cross-domain divergence. Select Minimum corresponding category As the final diagnostic result.
[0024] S5. Fault identification and decision output based on fractional gradient field divergence;
[0025] (1) Calculate the environmental fractional gradient field based on the consistency constraint criterion of fractional gradient field. Fractional gradient fields for each fault condition The vector divergence residuals under continuous perturbation scales are used to construct a cross-domain divergence metric function through time-dimensional weighted integrals. Cross-domain divergence The specific calculation expression is as follows:
[0026] (3)
[0027] in, For the global fractional gradient field of the target domain, For the source domain conditional fractional gradient field, This is a preset time weighting function used to adjust the weight of gradient bias on the decision result under different noise levels;
[0028] (2) Determine the fault category index that best matches the current signal probability evolution trajectory based on the criterion of minimizing the divergence of the full diffusion path. It outputs the corresponding fault diagnosis results and their confidence evaluation, realizing cross-domain transfer reasoning diagnosis under unknown working conditions in the target domain. The confidence evaluation is obtained by negative exponential normalization mapping of the cross-domain divergence degree corresponding to each fault category, and its calculation form is as follows:
[0029] (4)
[0030] in, Indicates the signal to be measured relative to the first Cross-domain divergence of fault condition distribution Represents the manifold of the signal to be measured and the source domain. Cumulative gradient divergence between fault-related prior manifolds ; This represents the total number of known fault categories in the source domain. The optimal diagnostic category is determined by comparing the confidence levels of each category, and the highest confidence level is used as the diagnostic reliability evaluation index. Invention Effects
[0031] This invention provides a cross-domain migration diagnostic method for aero-engine test faults based on continuous diffusion. The algorithm first trains a conditional fractional network model based on stochastic differential equations using historical test signals from the source domain to learn the fractional gradient fields of different fault categories at continuous diffusion time steps. Then, unsupervised training is performed using unlabeled signals from the target domain to construct a global fractional network model for the target domain, capturing the background noise distribution manifold of a specific engine model under the current test environment. For unknown fault signals in the target domain, a pre-defined diffusion operator is used to perform multi-timescale random noise addition to generate a noisy perturbation sequence, and a dual-path fractional gradient modeling method is used to estimate the fractional gradient fields of the environmental background distribution and the fault condition distribution, respectively. To address the diagnostic failure problem caused by inconsistent cross-domain feature distributions, a time-dimensional weighted integral algorithm is proposed to construct a cross-domain divergence metric function, and a negative exponential normalization mapping is used to calculate the diagnostic confidence. Experiments show that this method can effectively extract the long-range time-series dependency features and time-frequency manifold coupling relationship of engine signals, achieve cross-domain feature distribution alignment under unknown operating conditions in the target domain, and significantly improve the accuracy of fault diagnosis and the reliability of classification decision under variable operating conditions and strong noise environment. Attached Figure Description
[0032] Figure 1 Flowchart of a method for diagnosing cross-domain migration of test faults in aero-engines based on continuous diffusion;
[0033] Figure 2 A schematic diagram of the continuous diffusion noise addition process based on SDE;
[0034] Figure 3 U-Net network architecture diagram;
[0035] Figure 4 Logic diagram of dual-path fractional gradient modeling and divergence calculation. Specific implementation methods Specific implementation method one:
[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0038] like Figure 1 The overall process shown includes the following steps:
[0039] S1. Preparation and time-frequency feature mapping of source and target domain signal data;
[0040] S2. Train the conditional score network model for each fault category in the source domain using forward continuous diffusion noise-added data;
[0041] S3. Train a global score network model for the target domain using unlabeled test signal data from the target domain;
[0042] S4. Cross-domain perturbation mapping and dual-path fractional gradient modeling of target domain signals;
[0043] S5. Fault identification and decision output based on fractional gradient field divergence;
[0044] In this embodiment of the invention, during the model training phase, a large number of historical test signals from the source domain are selected as feature samples. A set of source domain fault condition fractional network models covering multiple fault modes is obtained through repeated iterative training using a forward continuous diffusion noise-adding process based on stochastic differential equations. Then, unsupervised training is performed using unlabeled data from the target domain, and a target domain global fractional network model representing the current test environment background distribution manifold is constructed and learned using the same network architecture, thus completing the construction of the dual-path fractional network model. In the cross-domain transfer diagnosis phase, a pre-defined diffusion operator is used to forward stochastically add noise to the unknown fault signal to be tested in the target domain, generating a noisy perturbation sequence. The dual-path fractional network model is then used to estimate the environmental background fractional gradient field and the fractional gradient fields of each fault condition. Finally, a time-dimensional weighted integral algorithm is used to calculate the cross-domain divergence degree, and the final fault diagnosis result and confidence evaluation are output through negative exponential normalization mapping.
[0045] The embodiments of the present invention will be described in detail below:
[0046] This invention employs historical test signals and test signals of aero-engines with different operating conditions. The invention is used to achieve fault identification and decision output under unknown operating conditions in the target domain. The specific implementation is as follows.
[0047] S1. Preparation and time-frequency feature mapping of source and target domain signal data;
[0048] In this embodiment, the first step is to construct a cross-domain heterogeneous database for model training and validation, which is the foundation for achieving transfer diagnostics. The source domain data is collected from aero-engines operating under predetermined standard test conditions, and includes a set of historical test signal samples with complete fault category labels acquired at different speeds and loads. These signals encompass a variety of common aero-engine failure modes, such as bearing fatigue spalling, blade foreign object damage, and rotor imbalance, and are used to support the subsequent discrimination and matching of the source domain fault condition fractional network model. The target domain data originates from actual test environments or operating sites that differ from the source domain's operating conditions. Its complexity lies in the presence of a large amount of unlabeled signal data without category annotations in the target domain. And unknown fault signals to be tested, acquired through sensors or extracted from historical records. .
[0049] Because aero-engine vibration signals are highly non-stationary and noisy, this embodiment utilizes a pre-defined time-frequency conversion operator to uniformly map the original one-dimensional signal during subsequent training and inference processes. Specifically, the time-domain sequence is converted into a two-dimensional time-frequency spectrum that can simultaneously characterize the time-domain evolution and frequency-domain energy distribution through continuous wavelet transform or short-time Fourier transform. This not only preserves the local features of the signal but also provides a standardized graphical input format for the subsequent deep feature extraction of the U-Net network structure.
[0050] S2. Train the conditional score network model for each fault category in the source domain using forward continuous diffusion noise-added data;
[0051] The core of this step lies in learning the prior distribution of source domain failure modes through a diffusion model. During implementation, a conditional fractional network model based on a symmetric encoder-decoder structure is established. ,like Figure 2 As shown, the source domain signal is noise-adding process through forward continuous diffusion. and its fault labels Joint modeling training is performed. This forward continuous diffusion process is based on stochastic differential equations, and its mathematical expression is:
[0052] (1)
[0053] in The drift coefficient, The diffusion coefficient is... This represents the standard Wiener process. By solving this equation step by step, the original complex engine signal distribution can be smoothly transformed into a known Gaussian distribution. For example... Figure 3 As shown, to enhance feature extraction capabilities, the Conditional Score Network employs a U-Net-type symmetric architecture with skip connections. Residual convolutional blocks are integrated into the encoding path to prevent gradient vanishing, and dilated convolutions are introduced to extract long-range temporal dependencies across scales. Furthermore, a self-attention mechanism is embedded in the bottleneck layer of the deep network to effectively capture non-local coupling relationships in the signal's time-frequency manifold. During training, the model iterates continuously by minimizing the weighted score matching loss function.
[0054] (2)
[0055] This loss function forces the network to learn the gradient fields of each fault category at different diffusion time steps, and finally builds a set of source domain fault conditional fractional network models covering N known fault modes.
[0056] S3. Train a global score network model for the target domain using unlabeled test signal data from the target domain;
[0057] This embodiment further constructs a global score network model for the target domain. This network employs the same deep symmetric architecture as step S2, but adopts a completely unsupervised training strategy. The model utilizes unlabeled test signal data collected from the target domain. The technical objective of training is not to identify specific fault categories, but rather to capture the overall background noise distribution manifold of a specific engine model under the current target domain test environment. In this way, the model can learn the probability density evolution law specific to the current operating condition, thereby providing an accurate environmental background reference benchmark for subsequent steps to isolate the influence of environmental noise and achieve a pure fault feature divergence measurement.
[0058] S4. Cross-domain perturbation mapping and dual-path fractional gradient modeling of target domain signals;
[0059] Acquiring the unknown fault signal under test in the target domain Subsequently, this step performs multi-timescale forward random noise addition on it using a preset diffusion operator, generating a series of evolving noisy perturbation sequences on the continuous-time probability manifold. This process essentially involves projecting the signal under test into a perturbation space defined by the diffusion process, in order to observe the manifold evolution characteristics of the signal at different signal-to-noise ratios. For example... Figure 4 As shown, the generated perturbation sequence is input in parallel into the global fractional network model of the target domain and the set of conditional fractional network models for each fault category in the source domain. Under a unified diffusion time parameter t, the fractional gradient fields characterizing the current environmental background distribution are estimated respectively. And the fractional gradient field characterizing the distribution of prior conditions for N types of source domain faults. Through this dual-path fractional gradient modeling process, the signal under test simultaneously obtains gradient evolution guidance of environmental distribution and various fault distributions at each time step after being disturbed, laying the foundation for subsequent divergence comparison.
[0060] S5. Fault identification and decision output based on fractional gradient field divergence;
[0061] The final fault diagnosis is based on the fractional gradient field consistency constraint criterion. This embodiment calculates the environmental fractional gradient field. Fractional gradient fields for each fault condition Vector divergence residuals under continuous perturbation scales, with a pre-defined time weighting function introduced. We construct a cross-domain divergence measure function by performing weighted integration. The specific calculation expression is as follows:
[0062] (3)
[0063] The weighting function dynamically adjusts the impact of gradient bias on decision-making under different noise levels, ensuring that the model can make reasonable diagnostic judgments in both strong and weak noise stages. Based on the criterion of minimizing the divergence of the full diffusion path, the following is selected: Category index corresponding to the minimum value This serves as the final fault diagnosis result.
[0064] Furthermore, to quantify the reliability of the diagnosis, the system performs negative exponential normalization mapping on the cross-domain divergence corresponding to each fault category to obtain a confidence evaluation index:
[0065] (4)
[0066] in, Indicates the signal to be measured relative to the first Cross-domain divergence of fault condition distribution; Represents the manifold of the signal to be measured and the source domain. Cumulative gradient divergence between fault-related prior manifolds ; This represents the total number of known fault categories in the source domain. By comparing the confidence levels of different categories, the optimal diagnostic category is determined, and the maximum confidence level is used as a quantitative evaluation of diagnostic reliability, thus effectively realizing cross-domain transfer reasoning diagnosis under unknown operating conditions in the target domain.
Claims
1. A method for diagnosing cross-domain migration of aero-engine test faults based on continuous diffusion, characterized in that, Includes the following steps: S1. Preparation and time-frequency feature mapping of source and target domain signal data; S2. Train the conditional score network model for each fault category in the source domain using forward continuous diffusion noise-added data; S3. Train a global score network model for the target domain using unlabeled test signal data from the target domain; S4. Cross-domain perturbation mapping and dual-path fractional gradient modeling of target domain signals; S5. Fault identification and decision output based on fractional gradient field divergence.
2. The method for cross-domain migration diagnosis of aero-engine test faults based on continuous diffusion as described in claim 1, characterized in that, The preparation and time-frequency feature mapping of source and target domain signal data in step S1 includes the following steps: S11. Source domain data is a set of historical test signal samples collected under predetermined test conditions, for which complete fault labeling information has been obtained. The source domain data is used to support the matching and discrimination of the source domain fault condition fractional network model; the target domain data comes from the actual operating or test environment that differs from the source domain, and includes unlabeled signal data without fault category annotations. and the signal to be measured This is used to achieve cross-domain feature distribution alignment and target domain fault state identification; S12. Both the source and target domain data need to be converted into two-dimensional time-frequency spectra using continuous wavelet transform or short-time Fourier transform during subsequent training to serve as inputs to the U-Net structure.
3. The method for cross-domain migration diagnosis of aero-engine test faults based on continuous diffusion as described in claim 1, characterized in that, Step S2, which involves training a conditional score network model for each fault category in the source domain using forward continuous diffusion with added noise, includes the following steps: S21. A forward continuous diffusion noise addition process is performed on the source domain fault category data using a preset diffusion operator defined based on stochastic differential equations. The expression is as follows: （1） in, The drift coefficient, The diffusion coefficient is... This represents the standard Wiener process; by solving this equation, the original signal distribution can be smoothly transformed into a known Gaussian distribution. S22. Establish a conditional fractional network model with a U-Net-type symmetric architecture containing skip connections. Integrate residual convolutional blocks and dilated convolutions in the encoding path to extract long-range temporal dependency features of engine signals across scales. At the same time, embed a self-attention mechanism in the deep feature extraction of the network to capture the non-local coupling relationship in the time-frequency manifold of the signal. S23. Learn the gradient fields of each fault category at different diffusion time steps, and construct a set of source domain fault conditional score network models covering N known fault modes. The training objective is to minimize the following weighted score matching loss function: （2） This loss function forces the network to learn that the signal recovers to a specific fault mode under different noise levels. The optimal path.
4. The method for cross-domain migration diagnosis of aero-engine test faults based on continuous diffusion as described in claim 1, characterized in that, Step S3, which involves training a global score network model for the target domain using unlabeled test signal data, includes the following steps: Construct a global score network model for the target domain using the same network architecture as in step S2. Using unlabeled test signal data in the target domain Unsupervised training is performed to enable the model to learn the unconditional score distribution of the target domain in order to capture the background noise distribution manifold of a specific engine model under the current test environment.
5. The method for cross-domain migration diagnosis of aero-engine test faults based on continuous diffusion as described in claim 1, characterized in that, Step S4, which describes the cross-domain perturbation mapping and dual-path fractional gradient modeling of the target domain signal, includes the following steps: S41, to diffusion time step is The noise was obtained ; S42, will Parallel input target domain global score network model and source domain fault category conditional score network model; S43. Calculate the environmental gradient of the target domain. ; S44. Calculate the theoretical gradient for each fault category. ,in Traversal Source domain faults; calculate cross-domain divergence. ; Select Minimum corresponding category As the final diagnostic result.
6. The method for cross-domain migration diagnosis of aero-engine test faults based on continuous diffusion as described in claim 1, characterized in that, The fault discrimination and decision output based on fractional gradient field divergence in step S5 includes the following steps: S51. Calculate the environmental fractional gradient field based on the fractional gradient field consistency constraint criterion. Fractional gradient fields for each fault condition The vector divergence residuals under continuous perturbation scales are used to construct a cross-domain divergence metric function through time-dimensional weighted integrals. Cross-domain divergence The specific calculation expression is as follows: （3） in, For the global fractional gradient field of the target domain, For the source domain conditional fractional gradient field, This is a preset time weighting function used to adjust the weight of gradient bias on the decision result under different noise levels; S52. Determine the fault category index that best matches the current signal probability evolution trajectory based on the criterion of minimizing the divergence of the full diffusion path. It outputs the corresponding fault diagnosis results and their confidence evaluation, realizing cross-domain transfer reasoning diagnosis under unknown working conditions in the target domain. The confidence evaluation is obtained by negative exponential normalization mapping of the cross-domain divergence degree corresponding to each fault category, and its calculation form is as follows: （4） in, Indicates the signal to be measured relative to the first Cross-domain divergence of fault condition distribution Represents the manifold of the signal to be measured and the source domain. Cumulative gradient divergence between fault-related prior manifolds ; This represents the total number of known fault categories in the source domain. The optimal diagnostic category is determined by comparing the confidence levels of each category, and the maximum confidence level is used as the diagnostic reliability evaluation index.