An aero-engine fault diagnosis method, device, equipment and medium

Abstract

An aero-engine fault diagnosis method, device, equipment and medium; belong to the field of avionics. Adopting a self-supervised learning framework, making full use of a large amount of aero-engine unlabeled vibration data for pre-training, without relying on field experts for large-scale data labeling, effectively solving the problem of high cost of label acquisition and scarcity of fault samples in industrial scenarios; through frequency domain mask enhancement to construct diversified positive and negative samples, make up for the insufficient model training problem caused by insufficient fault samples, greatly reduce the dependence threshold of labeled data; fusion of contrast learning and frequency domain reconstruction collaborative learning logic, effectively solve the defects of existing methods that contrast learning alone is prone to pseudo-contrast and weak discriminability of reconstruction learning features, realize the dual goals of comprehensive feature capture and accurate distinction, significantly improve the recognition accuracy of similar fault types.

Description

Technical Field

[0001] This invention belongs to the field of avionics, specifically relating to a method, system, equipment, and medium for diagnosing aircraft engine faults. Background Technology

[0002] With the continuous advancement of aviation technology and the increasing complexity of flight missions, the operational status of aero engines, as the core power system of aircraft, directly affects flight safety and reliability. Therefore, efficient and accurate fault diagnosis of aero engines has become a crucial aspect of aviation operation and maintenance support. Especially during flight, aero engines operate under high loads and complex conditions for extended periods, making them susceptible to various factors such as structural fatigue, component wear, and aerodynamic anomalies, which can induce various malfunctions.

[0003] Aero-engine failures have several notable characteristics: First, they are diverse in type, including various structural or functional problems such as bearing damage, blade breakage, combustion anomalies, and rotor imbalance. Second, they exhibit drastic changes in operating conditions, with engines displaying significantly different dynamic responses and noise levels at different flight stages (such as takeoff, climb, cruise, and descent), leading to complex and varied failure manifestations. Third, they are subject to strong interference in sensor signals, as engine vibration data are often affected by aerodynamic noise, structural resonance, and electromagnetic interference, resulting in a large number of non-fault-related components mixed in with the signals. Fourth, they exhibit broad spectral characteristics, with fault signals distributed across different frequency bands, including both low-frequency structural vibration characteristics and high-frequency impact-related fault information, posing challenges to traditional feature extraction and pattern recognition methods.

[0004] Conventional fault diagnosis methods typically assume that a large amount of monitoring data has been collected under normal and various fault conditions of mechanical equipment. Supervised machine learning models are used to extract correlation information from data label mappings, achieving good diagnostic results. However, in real-world industrial scenarios, mechanical equipment monitoring data suffers from core problems such as unlabeled data and a small number of fault samples. On the one hand, labeling requires domain expert knowledge and significant manpower; due to limitations in conditions or costs, monitoring data is generally unlabeled. On the other hand, equipment has long normal operating cycles, faults develop and spread rapidly, and are quickly shut down for repair after they occur, resulting in a very high proportion of normal samples and a very small number of fault samples in the monitoring data.

[0005] To address the core issues of unlabeled data and limited fault samples, self-supervised learning has been increasingly adopted in fault diagnosis due to its advantages of requiring no manual annotation and leveraging massive amounts of unlabeled data for feature self-learning. This technology was initially explored extensively in computer vision and natural language processing, forming two core paradigms: contrastive learning (learning discriminative features by constructing similarity or dissimilarity relationships between samples) and reconstruction learning (learning representational features by restoring damaged data). In recent years, some studies have attempted to apply it to aero-engine fault diagnosis, but significant shortcomings remain. Existing methods often employ either contrastive learning or reconstruction learning alone. While contrastive learning can enhance the discriminative power of features, it is highly dependent on the logic of sample pair construction, making it prone to spurious comparisons in scenarios with scarce fault samples, where similar sample pairs may contain unidentified fault differences. Reconstruction learning, while preserving the global representation of the data, has weak discriminative power in the learned features, making it difficult to distinguish similar fault types. Neither approach can simultaneously ensure both the completeness of feature representation and the accuracy of discrimination. Summary of the Invention

[0006] To address the core issues of unlabeled aero-engine monitoring data, scarce fault samples, and the difficulty of balancing feature representation completeness and discrimination accuracy in existing self-supervised methods under complex operating conditions, resulting in low diagnostic accuracy and insufficient generalization, this invention provides an aero-engine fault diagnosis method, device, equipment, and medium, belonging to the field of avionics. It employs a self-supervised learning framework, fully utilizing massive amounts of unlabeled vibration data from aero-engines for pre-training, eliminating the need for large-scale data annotation by domain experts, effectively solving the problems of high label acquisition costs and scarce fault samples in industrial scenarios. By constructing diverse positive and negative samples through frequency domain masking enhancement, it compensates for insufficient model training caused by insufficient fault samples, significantly reducing the reliance on labeled data. The integration of contrastive learning and frequency domain reconstruction collaborative learning logic effectively solves the shortcomings of existing methods that rely solely on contrastive learning, which is prone to pseudo-contrasts, and that reconstruction learning suffers from weak feature discrimination. This achieves the dual goals of comprehensive feature capture and accurate differentiation, significantly improving the identification accuracy of similar fault types.

[0007] A method for diagnosing aircraft engine faults, such as Figure 1 As shown, it includes the following steps:

[0008] S1. Perform data preprocessing on the collected aero-engine vibration time sequence signal;

[0009] S2. Data enhancement is performed on the preprocessed vibration time series signal using frequency domain masking to obtain enhanced samples; the enhanced samples include positive sample pairs and negative sample pairs.

[0010] S3. Input the enhanced sample into the encoder module for feature extraction and output the deep global representation vector; input the deep global representation vector into the projection head module for nonlinear mapping to obtain the discriminant adaptation embedding vector;

[0011] S4. Input the deep global representation vector and the discriminant adaptation embedding vector into the joint loss module to train and optimize the encoder module, thereby pre-training the encoder module and obtaining the trained encoder module.

[0012] S5. Construct a fault diagnosis model and use the fault diagnosis model to diagnose aero-engine faults.

[0013] Furthermore, in step S1, the method for preprocessing the collected aero-engine vibration time-series signal is as follows:

[0014] First, the continuously acquired vibration time-series signals By performing sliding segmentation with a fixed window length L, a set of time series segments is obtained. Time sequence fragment Where T represents the number of time frames, R is the set of real numbers, and N is the total number of time segments; the overlap rate is set to 20%~50% to preserve the transient characteristics of the transition phase of the operating condition;

[0015] Then, for time segment x i Perform normalization independently to obtain normalized time series segments. :

[0016] ;

[0017] In the formula: and and are the mean and standard deviation of the i-th segment, respectively; This is a numerically stable term to prevent division by zero errors.

[0018] Furthermore, in step S2, the process of data enhancement of the preprocessed vibration time-series signal using frequency domain masking is as follows:

[0019] Step 2.1, process the normalized time series segments. Perform short-time Fourier transform According to this, Converting a time-domain signal into a frequency-domain complex spectrum ;

[0020] Step 2.2, from the frequency domain complex spectrum Extracting spectral representation features, including amplitude spectrum. and phase spectrum ;

[0021] Step 2.3, only for amplitude spectrum Logarithmic compression is performed to obtain the compressed amplitude spectrum. To enhance high-frequency components;

[0022] Step 2.4: Based on the engine vibration characteristics, the compressed amplitude spectrum... Divided into K key frequency band sets Key frequency band set It covers three fault-sensitive regions: rotor fundamental frequency, gear meshing frequency, and bearing fault characteristic frequency; each frequency band B k Width W k Dynamically adjust based on prior knowledge to satisfy... f represents the frequency band; the frequency band includes a narrow band in the high-frequency region or a wide band in the low-frequency region; the high-frequency band is divided into narrow bands (to adapt to the energy concentration characteristics of high-frequency fault features such as bearing failure impact and gear meshing), and the low-frequency band is divided into wide bands (to match the wide-band distribution law of low-frequency vibration features such as rotor imbalance and structural resonance).

[0023] Step 2.5, then with probability Randomly select M frequency bands for masking;

[0024] Step 2.6: Compare the masked amplitude spectrum with the frequency domain complex spectrum. phase spectrum By fusing the data, the complex frequency spectrum is reconstructed to obtain the reconstructed complex frequency spectrum.

[0025] Step 2.7, Enhance sample generation;

[0026] To train a fault feature representation with strong discriminative power, based on the reconstructed frequency domain complex spectrum in step 2.6, the signal is converted into a time domain vibration signal by inverse short-time Fourier transform, and positive and negative sample pairs are constructed for comparative learning.

[0027] Step 2.7.1, process the normalized time series segments. By applying different random frequency domain masks, positive sample pairs are obtained;

[0028] Step 2.7.3: Randomly select sample x from vibration data of different operating conditions or different engines. j (j≠i);

[0029] Step 2.7.3, Sample x j After meeting with x i Consistent preprocessing, frequency domain enhancement, reconstruction, and inverse transform yield enhanced samples. Enhanced samples and any positive sample in the positive sample pair constitute a negative sample pair.

[0030] Furthermore, in step S3, the encoder module adopts a four-layer cascaded one-dimensional residual network (1D-ResNet) convolutional structure; the encoder module has the ability to extract features multiple times, which is used to extract high-dimensional discriminative features related to faults. It can simultaneously capture the local texture details and global structural information of the input spectrum and output a deep global feature vector as the input feature basis for the contrast loss module and the fault diagnosis model.

[0031] Each layer of the one-dimensional residual network contains two residual blocks (Res Blocks), such as... Figure 2 As shown;

[0032] Each residual block (Res Block) consists of convolutional layer 1 (Conv 1d), batch normalization (BatchNorm 1d), GELU activation function, convolutional layer 2 (Conv 1d), and batch normalization (BatchNorm 1d). For example... Figure 3 As shown.

[0033] Furthermore, in step S3, the projection head module includes a fully connected network layer; as shown below. Figure 4 As shown;

[0034] The projection head module performs a nonlinear mapping on the deep global feature vector output by the encoder module to obtain a discriminant-adaptive embedding vector. This discriminant-adaptive embedding vector adapts to the feature representation adaptation requirements of the joint loss module.

[0035] Through the nonlinear mapping operation of the projection head module, the deep global feature vector is converted into a discriminant-adaptive embedding vector. This discriminant-adaptive embedding vector can be directly used to calculate the sample similarity in the contrast loss module, and can also serve as the input feature of the frequency domain reconstruction loss module, realizing a reverse mapping to the spectral space. This effectively enhances the discriminative power of positive and negative sample pairs in the feature space, providing a structurally adapted feature foundation for frequency domain reconstruction, while also enabling the model to learn semantic feature representations with cross-condition universality and physical interpretability.

[0036] Furthermore, in step S4, the joint loss module includes a contrast loss module and a frequency domain reconstruction loss module.

[0037] The joint loss module achieves collaborative training of the encoder and projector modules by fusing contrastive loss and frequency domain reconstruction loss based on optimal transmission theory. The joint loss module accurately measures sample similarity in the feature space through optimal transmission distance, while maintaining the physical interpretability of features through spectral reconstruction. Figure 5 As shown.

[0038] By using a joint loss module training mechanism, the encoder module and the projection head module are optimized collaboratively, thereby significantly improving the model's ability to represent deep features of engine vibration data in the unsupervised stage. This ultimately ensures the accuracy and robustness of the aero-engine fault identification task and effectively solves the technical problems of insufficient fault feature extraction and difficulty in distinguishing similar faults under complex operating conditions.

[0039] Furthermore, the contrast loss module employs the optimal transport theory. As a classic framework for measuring the difference between two probability distributions, optimal transport theory calculates the minimum cost required to transport one feature distribution to another by finding the optimal transport plan. This accurately characterizes the global correlation and local differences of samples in the feature space, avoiding the limitations of traditional similarity measures that only focus on point-to-point local relationships. It is particularly suitable for scenarios with complex fault features and imbalanced sample distributions. Based on optimal transport theory, the contrast loss module uses the Sinkhorn distance in the regularized form of optimal transport theory to calculate the minimum transport cost between feature distributions, thereby measuring the similarity between positive sample pairs. The optimal transport contrast loss L... OT Calculation formula:

[0040] ;

[0041] in, This represents the optimal transfer plan between samples. Let n be the cost matrix, and n represent the number of samples after data augmentation. The coefficient of the entropy regularization term, For entropy regularization, Let (a, b) represent the joint distribution set satisfying the marginal constraints, where (a, b) are the marginal probability distributions of the two augmented samples in the positive sample pair, respectively, and both a and b are uniform distributions. The contrastive loss guides the positive sample pairs to align in the embedding space by minimizing the transmission cost between samples, effectively improving the alignment of semantic features in the embedding space.

[0042] Furthermore, the frequency domain reconstruction loss module directly maps the discrimination adaptation embedding vector output by the projection head module back to the frequency domain to obtain the reconstructed spectrum;

[0043] Normalized time segments were calculated using the L2 norm. The spectrum and x-based i The differences between the reconstructed spectra of the enhanced samples are minimized to maintain the physical interpretability of the features, thereby ensuring that the encoder module preserves the normalized temporal segments while extracting discriminative features. Frequency domain structure information; frequency domain reconstruction loss L recon Calculation formula:

[0044] ;

[0045] Where t represents the time frame at time t, and F represents the frequency dimension. It is a normalized time sequence segment The frequency domain representation, To enhance the spectrum of the reconstructed sample;

[0046] The joint loss function is composed of a weighted sum of contrastive loss and reconstruction loss, and the joint loss function L... total formula:

[0047] ;

[0048] in, and All can be adjusted hyperparameters, representing L respectively. OT and L recon The weights are used to balance L OT and L recon The training weights.

[0049] Furthermore, in step S5, the fault diagnosis model includes an encoder module and a classifier module;

[0050] The encoder module is the encoder module trained in step S4; the output of the encoder module is used as the input of the classifier module.

[0051] The classifier module is a multilayer perceptron; it includes an activation function layer and multiple fully connected layers. The classifier module ultimately outputs a fault category label. Figure 6 As shown;

[0052] Furthermore, in step S5, the process of diagnosing aero-engine faults using a fault diagnosis model is as follows:

[0053] Step 5.1: Utilize publicly available datasets or aero-engine vibration signal datasets obtained under actual operating conditions;

[0054] The dataset contains normal states under different operating conditions and different types of fault states; input samples for training the fault diagnosis model;

[0055] Step 5.2: Input the samples into the encoder module for feature extraction; Combine the labeled fault category information, use the cross-entropy loss function as a supervision signal to perform supervised fine-tuning training on the classifier module and encoder module, thereby optimizing the classification performance and diagnostic accuracy of the fault diagnosis model, and finally achieving accurate identification and classification of fault types.

[0056] The cross-entropy loss function L CE for:

[0057] ;

[0058] Where λ represents the total number of fault categories. This represents the probability that the model predicts the class to be of type θ. Labels representing the true category. By minimizing the cross-entropy loss function, the parameters of the classifier module are optimized, making the prediction results closer to the true fault category. During training, the loss value is passed to the classifier and encoder modules through backpropagation to achieve end-to-end parameter updates. The fault diagnosis model makes full use of the structural knowledge acquired in the pre-training stage, and has strong transfer and generalization capabilities, making it suitable for fault diagnosis tasks under various aero-engine operating conditions.

[0059] An electronic device, characterized in that it comprises: one or more processors; a memory; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the above-described aircraft engine fault diagnosis method.

[0060] A computer-readable storage medium, characterized in that the computer-readable storage medium stores program code, which can be called by a processor to execute the above-described aero-engine fault diagnosis method.

[0061] The present invention has the following beneficial effects:

[0062] Adapted to unlabeled and fault sample-scarce scenarios, significantly reducing labeling dependence: This invention adopts a self-supervised learning framework, which can make full use of massive unlabeled vibration data of aero-engines to complete pre-training without relying on domain experts for large-scale data labeling, effectively solving the core pain points of high label acquisition costs and scarce fault samples in industrial scenarios; at the same time, by constructing diverse positive and negative samples through frequency domain masking, it makes up for the problem of insufficient model training caused by insufficient fault samples, and significantly reduces the threshold of dependence on labeled data.

[0063] Breaking through the limitations of a single self-supervised paradigm and balancing feature representation completeness and discrimination accuracy: This invention innovatively integrates contrastive learning and frequency domain reconstruction collaborative learning logic. The contrastive learning module enhances the discriminativeness of fault features, while the frequency domain reconstruction module ensures the integrity of spectral structure information. This effectively solves the defects of existing methods that rely solely on contrastive learning, which is prone to false contrasts, and the reconstruction learning feature has weak discriminativeness. It achieves the dual goals of comprehensive feature capture and accurate differentiation, significantly improving the recognition accuracy of similar fault types. Attached Figure Description

[0064] Figure 1 This is a flowchart of the present invention;

[0065] Figure 2 This is a schematic diagram of the residual module, the basic encoder module of the present invention;

[0066] Figure 3 This is a schematic diagram of the 1D convolutional network module, the basic encoder module of the present invention;

[0067] Figure 4 This is a schematic diagram of the projection head of the present invention;

[0068] Figure 5 This invention relates to a pre-trained model based on joint loss;

[0069] Figure 6 This invention provides a fault diagnosis model for aero-engines.

[0070] Figure 7 This is a schematic diagram of the computer device of the present invention. Detailed Implementation

[0071] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0072] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the features of the following embodiments and examples can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0073] This invention also provides an aero-engine fault diagnosis system based on contrastive learning and reconstruction learning, including a data augmentation module, a model fine-tuning module, and a fault diagnosis module.

[0074] This invention provides a method for diagnosing aero-engine faults. Through self-supervised pre-training, frequency domain enhancement, optimal transmission contrastive learning, and spectrum reconstruction, it effectively improves the modeling capability for complex operating condition data under unlabeled conditions. On one hand, it significantly reduces the reliance on large amounts of labeled fault data: during the pre-training stage, this method utilizes a frequency domain mask enhancement strategy to construct positive sample pairs, optimizes the structural representation in the feature space by combining optimal transmission distance, and then maintains the physical interpretability of the features through frequency domain reconstruction. Therefore, it can obtain discriminative embedded representations without relying on explicitly labeled anomalous samples. On the other hand, it effectively solves the problem of difficult engine fault feature extraction under multi-source interference and complex vibration backgrounds: this invention uses a feature extraction model trained on the frequency domain structure, which can accurately perceive weak anomalous patterns in engine vibration signals, exhibiting strong robustness and generalization ability. Furthermore, in the fine-tuning stage, it combines limited labeled samples to achieve high-precision fault identification.

[0075] In this embodiment, a computer device is provided, such as... Figure 7 As shown, it includes a memory 701, a processor 702, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned aero-engine fault diagnosis methods based on contrastive learning and reconstruction learning.

[0076] Specifically, the computer device can be a computer terminal, a server, or a similar computing device.

[0077] The present invention also provides a computer-readable storage medium storing a computer program that executes the above-described aero-engine fault diagnosis model and training method based on contrastive learning and reconstruction learning, in order to solve the problems of complex and variable engine operating conditions, unstable fault feature extraction, and insufficient model generalization ability in the prior art, thereby achieving efficient and accurate engine fault identification and health status assessment.

[0078] Specifically, computer-readable storage media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media does not include transient media, such as modulated data signals and carrier waves.

[0079] Obviously, those skilled in the art should understand that the modules or steps of the above-described embodiments of the present invention can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of the present invention are not limited to any particular hardware and software combination.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing faults in an aircraft engine, characterized in that, Includes the following steps: S1. Perform data preprocessing on the collected aero-engine vibration time sequence signal; First, the continuously acquired vibration time-series signals By performing sliding segmentation with a fixed window length L, a set of time series segments is obtained. Time sequence fragment Where T represents the number of time frames, R is the set of real numbers, and N is the total number of time segments; the overlap rate is set to 20%~50% to preserve the transient characteristics of the transition phase of the operating condition; Then, for time segment x i Perform normalization independently to obtain normalized time series segments. : ； In the formula, and and are the mean and standard deviation of the i-th segment, respectively; This is a numerically stable term to prevent division by zero errors; S2. Data enhancement is performed on the preprocessed vibration time series signal using frequency domain masking to obtain enhanced samples; the enhanced samples include positive sample pairs and negative sample pairs. The process of data enhancement of the preprocessed vibration time series signal using frequency domain masking is as follows: Step 2.1, process the normalized time series segments. Perform short-time Fourier transform According to this, Converting a time-domain signal into a frequency-domain complex spectrum ; Step 2.2, from the frequency domain complex spectrum Extracting spectral representation features, including amplitude spectrum. and phase spectrum ; Step 2.3, only for amplitude spectrum Logarithmic compression is performed to obtain the compressed amplitude spectrum. To enhance high-frequency components; Step 2.4: Based on the engine vibration characteristics, the compressed amplitude spectrum... Divided into K key frequency band sets Key frequency band set It covers three fault-sensitive regions: rotor fundamental frequency, gear meshing frequency, and bearing fault characteristic frequency; each frequency band B k Width W k Dynamically adjust based on prior knowledge to satisfy... f represents the frequency band; Step 2.5, then with probability Randomly select M frequency bands for masking; Step 2.6: Compare the masked amplitude spectrum with the frequency domain complex spectrum. phase spectrum By fusing the data, the complex frequency spectrum is reconstructed to obtain the reconstructed complex frequency spectrum. Step 2.7, Enhance sample generation; Step 2.7.1, process the normalized time series segments. By applying different random frequency domain masks, positive sample pairs are obtained; Step 2.7.3: Randomly select sample x from vibration data of different operating conditions or different engines. j j≠i; Step 2.7.3, Sample x j After meeting with x i Consistent preprocessing, frequency domain enhancement, reconstruction, and inverse transform yield enhanced samples. Enhanced samples and any positive sample in the positive sample pair constitute a negative sample pair. S3. Input the enhanced sample into the encoder module for feature extraction and output the deep global representation vector; input the deep global representation vector into the projection head module for nonlinear mapping to obtain the discriminant adaptation embedding vector; The encoder module adopts a four-layer cascaded one-dimensional residual network convolutional structure; Each layer of the one-dimensional residual network contains two residual blocks (Res Blocks). Each residual block (Res Block) consists of convolutional layer 1, batch normalization, GELU activation function, convolutional layer 2, and batch normalization. The projection head module includes a fully connected network layer; The projection head module performs a nonlinear mapping on the deep global feature vector output by the encoder module to obtain the discriminant adaptation embedding vector; S4. Input the deep global representation vector and the discriminant adaptation embedding vector into the joint loss module to train and optimize the encoder module, thereby pre-training the encoder module and obtaining the trained encoder module. The joint loss module includes a contrast loss module and a frequency domain reconstruction loss module; The contrast loss module employs optimal transmission theory, specifically the Sinkhorn distance under the regularized form of optimal transmission theory, to calculate the minimum transmission cost between feature distributions. This cost is used to measure the similarity between positive sample pairs. The optimal transmission contrast loss L... OT Calculation formula: ； in, This represents the optimal transfer plan between samples. Let n be the cost matrix, and n represent the number of samples after data augmentation. The coefficient of the entropy regularization term, For entropy regularization, Let (a, b) represent the set of joint distributions that satisfy the marginal constraints, where (a, b) are the marginal probability distributions of the two augmented samples in the positive sample pair, and a and b are both uniform distributions. S5. Construct a fault diagnosis model and use the fault diagnosis model to diagnose aero-engine faults; The fault diagnosis model includes an encoder module and a classifier module; The encoder module is the encoder module trained in step S4; the output of the encoder module is used as the input of the classifier module. The classifier module is a multilayer perceptron, which includes an activation function layer and multiple fully connected layers. The classifier module finally outputs a fault category label. The process of diagnosing aero-engine faults using a fault diagnosis model: Step 5.1: Utilize publicly available datasets or aero-engine vibration signal datasets obtained under actual operating conditions; The dataset contains normal states under different operating conditions and different types of fault states; input samples for training the fault diagnosis model; Step 5.2: Input the samples into the encoder module for feature extraction; Combine the labeled fault category information, use the cross-entropy loss function as a supervision signal to perform supervised fine-tuning training on the classifier module and encoder module, thereby optimizing the classification performance and diagnostic accuracy of the fault diagnosis model, and finally achieving accurate identification and classification of fault types. The cross-entropy loss function L CE for: ； Where λ represents the total number of fault categories. This represents the probability that the model predicts the class to be of type θ. Labels that represent the true category.

2. The method for diagnosing aircraft engine faults according to claim 1, characterized in that, In step S4, the frequency domain reconstruction loss module directly maps the discrimination adaptation embedding vector output by the projection head module back to the frequency domain to obtain the reconstructed spectrum; Normalized time segments were calculated using the L2 norm. The spectrum and x-based i The differences between the reconstructed spectra of the enhanced samples are minimized to maintain the physical interpretability of the features, thereby ensuring that the encoder module preserves the normalized temporal segments while extracting discriminative features. Frequency domain structure information; frequency domain reconstruction loss L recon Calculation formula: ； Where t represents the time frame at time t, and F represents the frequency dimension. It is a normalized time sequence segment The frequency domain representation, To enhance the spectrum of the reconstructed sample; The joint loss function is composed of a weighted sum of contrastive loss and reconstruction loss. The joint loss function L... total formula: ； in, and All can be adjusted hyperparameters, representing L respectively. OT and L recon The weights are used to balance L OT and L recon The training weights.

3. An electronic device, characterized in that, include: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method as described in any one of claims 1 to 2.

4. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 2.

Images