Graph self-coding digital twinborn body construction method for aero-engine vibration overrun fault diagnosis

Abstract

The invention provides a graph self-encoding digital twin construction method for aero-engine vibration overrun fault diagnosis, and belongs to the field of aero-engine digital engineering maintenance and guarantee. The construction method is divided into three parts, namely an engine sensor network state correction diagram, a self-encoding model and a feedback decision, and comprises the following steps: acquiring an engine sensor signal, and selecting proper sample characteristics of data; performing denoising, normalization and resampling preprocessing on each piece of state data, constructing a time window, and dividing a training test set; constructing a graph self-coding digital twinborn body; and performing fault prediction by using the constructed digital twinborn body. According to the method, the vibration overrun fault of the engine can be efficiently identified, the normal working condition and the abnormal state can be accurately distinguished, and the accuracy and reliability of fault identification are remarkably improved, so that timely support is provided for preventive maintenance and operation decision of the engine, and a solid technical guarantee is provided for health management of the aero-engine.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of digital engineering maintenance and support of aero-engines, and relates to a graph auto-encoding digital twin construction method for aero-engine vibration overrun fault diagnosis. BACKGROUND

[0002] An aero-engine is known as the "heart" of an aircraft, providing core power for flight, and its safety and reliability are directly related to flight safety and mission success. Modern aero-engines have complex structures, integrating a compressor, a combustion chamber, a turbine, a control system and other key components, and work in extremely harsh environments, often under extreme conditions such as high temperature, high pressure and high speed, and long-term thermal stress and mechanical vibration. In such an environment, the engine often experiences blade cracks, rotor imbalance, bearing wear, structural loosening and other faults, which can cause abnormal vibration. Once the vibration exceeds the safety threshold, it may cause component failure, performance degradation, and even serious consequences such as in-flight shutdown. Therefore, establishing an efficient and accurate vibration state monitoring and fault diagnosis mechanism is crucial for ensuring flight safety, improving engine availability and reducing operating and maintenance costs.

[0003] Among them, vibration overrun, as one of the most common alarm events of aero-engines, has become an important factor affecting flight punctuality and engine in-service life. Vibration overrun usually manifests as the vibration amplitude of the engine rotor system exceeding the design-allowed threshold, and its causes are diverse, including but not limited to: mass imbalance caused by compressor or turbine blade fracture or fouling, support stiffness reduction caused by bearing wear or lubrication failure, rotor misalignment, structural resonance, casing deformation, etc. Taking a real test process of a certain type of aero-engine vibration overrun event as an example: during the test process, when the fan speed dropped to 80.3% (compressor speed 85.7%), the rear vertical vibration component suddenly rose to 51.7 mm / s, and the test operator was forced to pull back the throttle to slow speed at 48 minutes and 16 seconds, and shut down at 48 minutes and 46 seconds. This vibration overrun directly caused the test to be interrupted, the engine had to be disassembled for troubleshooting, which seriously affected the test progress and caused the external delivery task to be delayed.

[0004] After in-depth analysis, the root cause of the failure is that the assembly process of the locking washer does not follow the design requirements, and there are non-compliant assembly situations: the design drawing does not clearly make standard samples, resulting in a lack of quantifiable comparison benchmarks during assembly, and the locking washer cannot be accurately positioned; at the same time, after assembly, effective inspection records are not conducted, resulting in assembly defects not being discovered before shipment. After the engine is put into use, the locking washer gradually fails under high-speed working conditions, cannot effectively fix the special nut, causes the nut to loosen and causes insufficient compression force, and further causes the relative movement between the No. 3 fulcrum bearing inner steel sleeve and the shaft neck, ultimately leading to the whole machine vibration exceeding the limit. This case shows that the mode of relying on manual assembly and post-detection cannot effectively prevent and control the progressive vibration failure caused by mechanical connection loosening, and at the same time, it also exposes the key weak links in the current failure prevention and state monitoring system.

[0005] Currently, aeroengine fault diagnosis methods are mainly divided into model-driven and data-driven categories. Model-driven methods establish mathematical models (e.g., aerodynamic-thermal models, rotor dynamics models) based on the physical mechanisms of the engine, and achieve state estimation and residual analysis through state observers or filtering algorithms (e.g., Kalman filter, sliding mode observer). This type of method has clear physical meaning and strong interpretability, but its performance is highly dependent on accurate modeling. In the face of complex nonlinear dynamics, parameter uncertainty, and unmodeled dynamics, model-driven methods have poor adaptability, high modeling cost, and long modeling cycle. Another challenge is that vibration signals have strong nonlinearity, multi-source coupling, and strong dependence on working conditions. The abnormal features of faults are usually weak in time-frequency changes in the early stages, easily masked by noise or normal variable working condition dynamics, which makes traditional detection methods based on fixed thresholds or spectral analysis have high false negative rate, early warning lag, and frequent false positives. For example, existing systems often rely on single-point peak values or root mean square (RMS) values compared with empirical thresholds, which makes it difficult to distinguish between real faults and normal transient processes (e.g., acceleration, surge margin fluctuations), and cannot identify multi-sensor collaborative abnormal patterns. In addition, the vibration baseline of the engine varies significantly in different flight stages (e.g., takeoff, cruise, landing), and static thresholds are difficult to adapt to dynamic changes in the operating environment, which further limits the detection accuracy. With the continuous improvement of sensor technology and data acquisition capability, data-driven methods have gradually become a research hotspot. This type of method achieves anomaly detection and fault identification by mining statistical rules or time series features in historical operation data. Although data-driven methods do not require accurate modeling and have strong adaptability to complex systems, traditional models still have certain limitations in handling long sequence dependencies, multi-variable coupling, and early weak fault features. To address this issue, existing literature [1] = { Wang CX, Wang CF, Ding, BQ. Time-Frequency Postprocessing Method Based on Generalized S-Transform and Its Application to Aeroengine Rotor System Fault Diagnosis [J]. IEEE Transactions on Instrumentation and Measurement, 2024, Volume 73, 3001811} proposes a graph neural network method based on multi-scale graph wavelet convolution for multi-sensor data and time-varying speed conditions to enhance the cross-channel vibration feature capture ability and diagnostic interpretability, providing an effective means for high-precision, early fault warning.In addition, the literature [2] = {Jia SX, Li YB; Sun DY. An Interpretable Multiscale Graph Wavelet Neural Network For Aeroengine Fault Diagnosis Under Time-Varying Speeds [J]. Mechanical Systems And Signal Processing, 2025, Volume 240, 113308} fuses multi-sensor information and improves the interpretability of diagnosis, and proposes a new multi-scale graph neural network method, which provides an effective means for early and reliable warning of vibration over-limit faults. At the same time, the literature [3] = {Liu YP, Jiang HK, Liu CQ. Data-augmented wavelet capsule generative adversarial network for rolling bearing fault diagnosis [J]. Knowledge-Based Systems, 2022, Volume 252, 109439} focuses on the problem of data imbalance, and proposes a data-augmented wavelet capsule generative adversarial network. By introducing wavelet transform, the translation invariance of convolutional neural network is maintained, and capsule network is used to improve signal information analysis, which improves the diagnosis performance of unbalanced data in rolling bearing fault diagnosis.

[0006] It should be noted that although the above deep learning methods have shown certain effects in some specific tasks, in the actual application of early fault warning of key components of aero-engines, they still face some common challenges. For example, these methods have low sensitivity to early fault features and insufficient recognition ability. At the same time, the training of deep learning models often relies on a large amount of high-quality labeled data, which is difficult to obtain in actual engineering due to the scarcity of fault samples. Therefore, there is an urgent need for an intelligent detection method that can adaptively learn normal vibration patterns, accurately identify abnormal deviations, and have strong time series modeling capability, in order to realize early and reliable warning of vibration over-limit faults. At the same time, it is also necessary to build a digital twin system integrating state perception, intelligent diagnosis and process tracing, so as to realize full-process monitoring and early warning of key component assembly quality and running state. SUMMARY

[0007] The present application aims at the above-mentioned problems, and proposes an interpretable graph auto-encoding digital twin for constructing aero-engine vibration overrun fault diagnosis based on spatio-temporal graph learning and mechanism prior knowledge embedding fusion. The designed auto-encoding digital twin can determine whether the aero-engine has a vibration overrun fault according to the sensor data of the aero-engine. This method can efficiently identify engine vibration overrun faults, accurately distinguish normal working conditions from abnormal states, significantly improve the accuracy and reliability of fault identification, and thus provide timely support for preventive maintenance and operation decision of the engine, and provide a solid technical support for aero-engine health management.

[0008] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0009] A graph auto-encoding digital twin construction method for aero-engine vibration overrun fault diagnosis, the construction method is divided into three parts of engine sensor network state correction graph, auto-encoding model and feedback decision, including the following steps: obtaining engine sensor signal, selecting appropriate sample features of data; denoising, normalizing and resampling preprocessing of each state data, constructing time window, dividing training and test set; constructing graph auto-encoding digital twin; using the constructed digital twin for fault prediction. Specifically as follows:

[0010] Step S1: using the multi-source sensor network on the aero-engine, collecting the running state data of the aero-engine under various running states, obtaining the original data, and performing data sample feature selection to form an original data set with clear labels; specifically:

[0011] Step S1.1: relying on the multi-source sensor network arranged on the aero-engine, synchronously collecting the running state data of the engine under different working conditions as original data. The multi-source sensor network includes vibration acceleration sensor, speed sensor, temperature sensor and pressure sensor. The different working conditions include ground slow running, climbing, cruising and deceleration parking. The collected running state data covers the vibration characteristics, speed change, thermal parameters and aerodynamic performance indicators of key components, forming a high-dimensional, multi-channel time series data set. The time series data set contains multiple sensor signals such as vibration, speed, temperature and pressure;

[0012] The time series data set formed above is represented by a two-dimensional data sequence , where T is the number of time sampling points, N is the number of sensor channels, i.e. the dimension of the characteristic variable, each column corresponds to a physical characteristic variable, and each row represents the synchronous observation value of each variable at a certain time. This data structure can effectively reflect the multivariate coupling characteristics of the aero-engine in complex dynamic processes;

[0013] Step S1.2: In order to construct high-quality training sample set and test sample set, the collected raw data needs to be preliminarily labeled. The application adopts an improved constant threshold method based on engineering experience and historical troubleshooting data to set a safety threshold for the vibration principal component, and the safety threshold of the vibration principal component is 30. The vibration principal component is the effective value of the vibration velocity or acceleration in the vertical and horizontal directions.

[0014] When the amplitude of the vibration principal component continuously exceeds the preset safety threshold in a certain period of time and is accompanied by a sudden increase in vibration during the speed drop process, it is determined as a "vibration overrun fault" sample; the rest of the data that meets the normal operation specification is marked as a "normal working condition" sample.

[0015] Through the above method, the preliminary separation of fault data and normal data is realized, and the original data set with clear labels is formed, providing a reliable data basis for subsequent feature engineering and model training.

[0016] Step S2: The original data set with clear labels collected in step S1 is preprocessed to make it suitable for the form of graph auto-encoder digital twin architecture, and through data cleaning, time alignment and standardization processing, combined with sliding window slicing, a multi-channel time series sample is constructed to ensure the reliability and consistency of the input data, so that the graph auto-encoder can accurately learn the data features of the engine in normal state and realize effective anomaly detection; Specifically:

[0017] Step S2.1: For the vibration overrun fault generated during the operation of the aero-engine, a plurality of sub-data sets are established using the original data set with clear labels collected in step S1. For the missing values, jump points and abnormal values that obviously deviate from the physical law existing in the data set, the median filter in the sliding window is used, as shown in formula (1).

[0018]

[0019] Wherein, is the original signal, is the sliding step, is the filtered signal.

[0020] Step S2.2: The filtered data is identified and repaired by a linear interpolation method, as shown in formula (2).

[0021]

[0022] Wherein, is the filtered invalid data point, and are the front and rear valid points, is the repaired data. This method ensures the reliability and physical consistency of the data.

[0023] Step S2.3: Since there may be a sampling time offset problem between the vibration signal and other sensor signals in the original data set, linear interpolation alignment is needed based on the unified timestamp to ensure the synchronization of each sensor data signal at the same time step, and to avoid model misjudgment due to timing misalignment;

[0024] Step S2.4: In view of the large difference in the dimensions of different sensor data and the uneven distribution of numerical ranges in the original data set, Z-score standardization method is used for dimensionless processing to make all input data distributed in a similar interval, so as to improve the convergence speed and stability of model training. As shown in equation (3), the similar interval has a mean of 0 and a variance of 1.

[0025]

[0026] wherein, is the observation value of different sensor data in the original data set at time , is the result of Z-score standardization on the feature , is the mean value of the feature , N is the sample number, is the standard deviation of the feature , is the unified timestamp corresponding to the i-th time sampling point;

[0027] Step S2.5: Then, according to the dynamic characteristics of the typical operation process of the engine, a reasonable sliding window length is set, and the continuous time series data is divided into subsequence samples with overlapping relationship. Each subsequence sample contains multi-channel sensor multi-step historical information as the input sequence of the graph autoencoder, so as to effectively capture the dynamic evolution trend and potential fault mode of the vibration signal. The typical operation process is acceleration, steady state, deceleration;

[0028] Step S2.6: Finally, the constructed subsequence samples are divided into training set, validation set and test set, wherein the training set and the validation set only use the test run data confirmed to be fault-free, so as to ensure that the autoencoder learns the "normal" operation mode of the engine; the test set contains multiple groups of normal and abnormal working conditions, which are used to evaluate the abnormal detection ability and generalization performance of the model;

[0029] Step S3: constructing a graph auto-encoding digital twin, which is composed of an aero-engine sensor network state correction graph, a deep learning model of an auto-encoder, and a feedback decision control module. First, the correlation between the sensor data in the training set is fused with the prior knowledge of the aero-engine to construct a sensor network state correction graph; then, the multi-dimensional time series sensor data optimized by the state correction graph are input into the deep learning model based on the graph auto-encoder; finally, the feedback decision control module is introduced as the closed-loop interface between the digital twin and the physical system, and the establishment of the overall digital twin architecture is completed; the specific construction process is as follows:

[0030] Step S3.1: fusing the correlation between the prior knowledge of the aero-engine and the sensor data to construct an engine sensor network state correction graph; specifically:

[0031] Step S3.1.1: first, in the vibration overrun fault diagnosis of an aero-engine, the reasonable selection of sensors determines the sensing ability of the model to key fault features. In order to ensure that the constructed model has good physical interpretability and diagnostic pertinence, the present application performs mechanism-driven feature selection on the sensor data in the training set based on the prior knowledge of the structure topology, power transmission path and typical fault propagation mechanism of the aero-engine. Specifically, for the typical mechanical looseness class fault corresponding to vibration overrun, the excitation source mainly embodies the local eccentricity and looseness of the rotor system, and the vibration energy is transmitted along the shaft structure and is jointly affected by the throttle command and flight conditions. Therefore, the present application preferentially selects key variables directly related to vibration response or having a modulating effect on fault evolution as deep learning network nodes. The key variables include: vibration acceleration signals at the fan case and high-pressure compressor mounting points, low-pressure rotor speed and high-pressure rotor speed , throttle lever angle PLA, and flight altitude or inlet total pressure;

[0032] Step S3.1.2: secondly, the spatio-temporal information of the deep learning network graph after mechanism-driven feature selection only contains prior knowledge, however, the mutual influence between the connected nodes in the network graph is still unknown, therefore, the weight calculation between different nodes can describe the correlation between different sensor data. In order to capture the mutual relationship between different nodes in the network graph, the present application selects the Pearson correlation coefficient method suitable for information transmission to perform data-driven correlation analysis on the sensor data in the training set to obtain the node connection relationship of the deep learning network graph. As shown in equation (4);

[0033]

[0034] wherein, and are two vectors, is a vector and the number of samples, denotes the average value of the vector;

[0035] Step S3.1.3: Finally, the node selection of the deep learning network graph guided by prior knowledge is combined with the data-driven edge connection to generate a sensor network state correction graph that integrates mechanism cognition and measured characteristics. This method not only preserves the intrinsic causal logic of the system, but also captures the collaborative change pattern in actual operation, effectively improving the accuracy and interpretability of the autoencoder model in fault diagnosis;

[0036] Step S3.2: Input the multi-dimensional time series sensor data optimized by the state correction graph into the deep learning model of the autoencoder;

[0037] To effectively capture the long-range dependence and cross-channel dynamic association in the multi-sensor time series of an aero-engine, the present application uses a graph structure-based autoencoder to perform nonlinear feature learning and low-dimensional representation modeling on the multi-dimensional time series sensor data optimized by the state correction graph. The autoencoder consists of an encoder and a decoder, both of which use a multi-head self-attention mechanism (MHSA) as the core module, and can adaptively model the time series evolution pattern and potential coupling relationship between sensor variables. Specifically:

[0038] Step S3.2.1: First, map the multi-dimensional time series sensor data to an embedded vector through linear projection, as shown in equation (5);

[0039]

[0040] wherein, denotes the input embedded vector, is a learnable embedded weight matrix, is a position encoding used to preserve the order information of the time series, which is usually constructed using sine / cosine functions, as shown in equations (6) and (7);

[0041]

[0042]

[0043] wherein, is the embedding dimension of the model;

[0044] Step S3.2.2: Then, in the encoder, the input embedded sequence is input through N layers of multi-head self-attention and feedforward network modules to extract high-level abstract features of the input sequence. The output of a single-layer encoder is denoted as Z, as shown in equation (8);

[0045]

[0046] wherein, is an encoder feedforward network, is multi-head self-attention;

[0047] The multi-head self-attention is defined as shown in equations (9) and (10);

[0048]

[0049] wherein h is the number of attention heads, is an output projection matrix, is the Query of the i-th head, is the Key of the i-th head, is the Value of the i-th head, is a learnable projection, is the dimension of Query / Key for each attention head, is the time step, and T is the total length of time;

[0050] The feedforward network is generally composed of two linear mappings and a nonlinear activation function, as shown in equation (11);

[0051]

[0052] wherein, is the output of, is the first layer linear weight, is the first layer bias, is the first layer linear weight, is the first layer bias, is the hidden layer dimension of the feedforward network, and the size is 4 ;

[0053] Step S3.2.3: Secondly, the decoder receives the latent representation Z output by the encoder, and gradually reconstructs the original input sequence through a similar attention mechanism , as shown in equation (12);

[0054]

[0055] wherein, is the reconstructed data, is the encoder, is the weight coefficient;

[0056] The Decoder is composed of N layers of stacked mask self-attention, cross-attention, and feedforward network, and the single-layer structure can be represented as shown in equation (13);

[0057]

[0058] wherein, is the start embedding of the input sequence; is the multi-head self-attention module with temporal masking, CA is cross-attention, is the decoder feed-forward network;

[0059] The multi-head self-attention module with temporal masking is shown in equations (14) and (15);

[0060]

[0061] wherein, M is an upper triangular masking matrix, used to mask future time information;

[0062] The cross-attention module is shown in equation (16);

[0063]

[0064] wherein, Q is the decoder input, Z is the encoder output; the Key and Value of the cross-attention are obtained by linear mapping from the encoder output Z;

[0065] The feed-forward network is shown in equation (17);

[0066]

[0067] wherein, M is the output of the cross-attention module, is the first layer linear weight, is the first layer bias, is the first layer linear weight, is the first layer bias;

[0068] Step S3.2.4: Finally, the reconstructed representation obtained by stacking the above N layers of the decoder is recovered to the original feature space through linear mapping, as shown in equation (18);

[0069]

[0070] wherein, is the model reconstruction result, is the hidden representation output by the Nth layer of the decoder, is the output projection matrix of the decoder, used to map the hidden features back to the original sensor feature dimension The model is trained by minimizing the reconstruction error;

[0071] The loss function is defined as mean square error (MSE), as shown in equation (19);

[0072] ​

[0073] wherein, is real data, is reconstructed data, and LOSS is a loss function;

[0074] Through the graph autoencoder, the system can learn the normal operation mode of the sensor data in an unsupervised or semi-supervised mode, extract a potential feature representation Z sensitive to the anomaly, and provide a high-discriminative feature input for subsequent anomaly detection and state evaluation.

[0075] Step 3.3: Design a feedback control decision mechanism oriented to the reconstruction error of the graph autoencoder;

[0076] On the basis of modeling the normal operation state of the engine by the graph autoencoder, the application constructs a feedback decision control mechanism based on the reconstruction error, for real-time monitoring of the engine health state and identification of potential faults. Specifically:

[0077] First, the autoencoder is trained only using the training set composed of fault-free test data in the training phase, so that it can accurately reconstruct the input signal conforming to the normal operation mode;

[0078] Second, in the test phase, input the test set containing multiple normal and abnormal working conditions into the trained model to obtain the corresponding reconstruction output;

[0079] Finally, calculate the maximum reconstruction error of each test set data according to the defined reconstruction error formula, as shown in formula (21);

[0080]

[0081] wherein, is real data, is reconstructed data, is the window length, and k is the number of the window, is the maximum reconstruction error in each window of each test set data;

[0082] At the same time, based on the reconstruction error distribution of the normal samples in the training set, the adaptive threshold is determined by statistical method , as shown in formula (22);

[0083]

[0084] wherein, is the average of the reconstruction error, is an adjustable coefficient, is the variance of the reconstruction error;

[0085] Accordingly, the binary decision rule is established as shown in formula (23);

[0086]

[0087] wherein, is the maximum reconstruction error, which ranges from 0 to 100;

[0088] When the abnormal state is determined according to the binary decision rule, the feedback control module is triggered to perform the following operations: outputting an alarm signal and recording the time, working condition and sensor channel significantly deviating from the abnormal occurrence;

[0089] Step S4: training the graph auto-encoding digital twin and verifying the effectiveness thereof;

[0090] Step S4.1: first, according to the size of the training set obtained in step S2.6, suitable hyperparameters are selected for the model, including the training batch size, the number of iterations, the learning rate and the initial weight;

[0091] Step S4.2: then, the training set is inputted, and the Adam optimizer based on the stochastic gradient descent principle is used to train and optimize the weights of the auto-encoding digital twin, so that the target LOSS in S3.2.4 is minimized;

[0092] Step S4.3: finally, after the model training is completed, the model is verified using the verification set obtained in step S2.6, the model parameters are fine-tuned and retrained according to the verification result, the optimal model parameters are obtained, and the model is saved;

[0093] Step S5: testing the auto-encoding digital twin and verifying the effectiveness thereof;

[0094] Step S5.1: first, the test set data obtained in step S2.6 is inputted into the graph auto-encoding digital twin trained and saved in step S4.3, to obtain the corresponding reconstruction result. Then, the real data and the reconstruction data are substituted into the maximum reconstruction error calculation formula in step S3.3 to obtain the reconstruction error of the test set;

[0095] Step S5.2: then, the threshold is set according to step S3.3;

[0096] Step S5.3: secondly, according to the maximum reconstruction error calculated in step S5.1 and the threshold set in step S5.2, the data state is determined according to the binary decision rule given in step S3.3;

[0097] Step S5.4: according to S5.3, the state of the test data is determined, and is compared with the real state of the data to analyze the false negative rate and the false positive rate of the method.

[0098] The beneficial effects of the present application are:

[0099] (1) The present application significantly improves the diagnosis accuracy and early warning ability of aero-engine vibration overrun faults by constructing a fusion mechanism knowledge and data-driven graph auto-encoding digital twin. Compared with traditional methods relying on fixed thresholds or shallow models, this architecture can learn the normal vibration patterns of the engine under different working conditions, effectively capture the spatio-temporal correlation and long-term sequential dependence features between multiple sensors, significantly reduce the false positive rate and false negative rate under variable working conditions, and achieve accurate identification of faults.

[0100] (2) The present application introduces a sensor network state correction graph based on prior structure and a reconstruction error driven feedback decision mechanism, enhancing the physical interpretability and engineering practicability of the model. The system not only can monitor the health status in real time and automatically trigger alarms, but also can locate abnormal channels and assist fault tracing, forming a "perception-diagnosis-decision" closed loop, providing intelligent support for engine test troubleshooting and field maintenance, and significantly improving operation efficiency and flight safety guarantee capability. BRIEF DESCRIPTION OF DRAWINGS

[0101] Figure 1 The structural diagram of the graph auto-encoding digital twin of the present application;

[0102] Figure 2 The flowchart of the self-encoding digital twin construction method for aero-engine vibration overrun fault diagnosis proposed by the present application;

[0103] Figure 3 The working condition snapshot of the aero-engine data used by the present application;

[0104] Figure 4 The reconstruction error graph obtained when the graph auto-encoding digital twin is tested on each time window of normal data;

[0105] Figure 5 The comparison graph of original and reconstructed signals obtained when the graph auto-encoding digital twin is tested on each time window of normal data; Figure 5 (a) in is the reconstruction effect of Vg1Totl parameter; Figure 5 (b) in is the reconstruction effect of Vg2Totl parameter; Figure 5 (c) in is the reconstruction effect of Vg1N1_MF1 parameter; Figure 5 (d) in is the reconstruction effect of Vg1N1_MF2 parameter; Figure 5 (e) in is the reconstruction effect of Vg1N1_MF3 parameter; Figure 5 (f) in is the reconstruction effect of Vg1N2_MF1 parameter; Figure 5 (g) in is the reconstruction effect of Vg1N2_MF2 parameter; Figure 5 (h) in is the reconstruction effect of Vg1N2_MF3 parameter; Figure 5(i) in FIG. 2 is the reconstruction effect of the Vg2N1_MF1 parameter; Figure 5 (j) in FIG. 2 is the reconstruction effect of the Vg2N1_MF2 parameter; Figure 5 (k) in FIG. 2 is the reconstruction effect of the Vg2N1_MF3 parameter; Figure 5 (l) in FIG. 2 is the reconstruction effect of the Vg2N2_MF1 parameter; Figure 5 (m) in FIG. 2 is the reconstruction effect of the Vg2N2_MF2 parameter; Figure 6 (n) in FIG. 2 is the reconstruction effect of the Vg2N2_MF3 parameter;

[0106] Figure 7 is the reconstruction error map of the graph auto-encoding digital twin when testing on each time window of the abnormal data;

[0107] Figure 7 is the original and reconstructed signal comparison map of the graph auto-encoding digital twin when testing on each time window of the abnormal data; Figure 7 (a) in FIG. 1 is the reconstruction effect of the Vg1Totl parameter; Figure 7 (b) in FIG. 1 is the reconstruction effect of the Vg2Totl parameter; Figure 7 (c) in FIG. 1 is the reconstruction effect of the Vg1N1_MF1 parameter; Figure 7 (d) in FIG. 1 is the reconstruction effect of the Vg1N1_MF2 parameter; Figure 7 (e) in FIG. 1 is the reconstruction effect of the Vg1N1_MF3 parameter; Figure 7 (f) in FIG. 1 is the reconstruction effect of the Vg1N2_MF1 parameter; Figure 7 (g) in FIG. 1 is the reconstruction effect of the Vg1N2_MF2 parameter; Figure 7 (h) in FIG. 1 is the reconstruction effect of the Vg1N2_MF3 parameter; Figure 7 (i) in FIG. 2 is the reconstruction effect of the Vg2N1_MF1 parameter; Figure 7 (j) in FIG. 2 is the reconstruction effect of the Vg2N1_MF2 parameter; Figure 7 (k) in FIG. 2 is the reconstruction effect of the Vg2N1_MF3 parameter; Figure 8 (l) in FIG. 2 is the reconstruction effect of the Vg2N2_MF1 parameter; Figures 1-2 (m) in FIG. 2 is the reconstruction effect of the Vg2N2_MF2 parameter; Figure 3 (n) in FIG. 2 is the reconstruction effect of the Vg2N2_MF3 parameter;

[0108] Figure 1 is the abnormality identification result of the graph auto-encoding digital twin on the 25 groups of experimental data. DETAILED DESCRIPTION

[0109] In order to make the technical solutions and advantages of the present application more detailed, the specific embodiments of the present application will be described in detail below in combination with the drawings and analysis of experimental data.

[0110] The embodiment relates to a graph auto-encoding digital twin design architecture for an aero-engine vibration overrun fault diagnosis and a construction method thereof, and a specific architecture is as shown in the accompanying drawings. Figure 4 The specific architecture comprises the following steps.

[0111] Step S1: using a multi-source sensor network on an aero-engine, collecting running state data of the aero-engine in multiple running states to obtain original data, and performing data sample feature selection to form an original data set with clear labels; specifically,

[0112] The present application is based on the vibration overrun event experimental data in a real test process of an aero-engine provided by Dalian University of Technology and China Aero-engine Design Institute of Guangyin, and the execution process of the method is introduced in detail.

[0113] Step S1.1: for the fault case described in the background art, all test data of the aero-engine are collected, including: fan case vertical and horizontal direction vibration acceleration sensors, low-pressure and high-pressure rotor speed sensors, bearing cavity temperature sensor (PT100) No. 3, compressor outlet pressure sensor, fuel flow meter and throttle lever angle (PLA) sensor. In a complete test cycle, the running state data of the engine under typical working conditions such as cold start, slow vehicle, climbing, cruising and deceleration parking are synchronously collected, a total of 25 csv file data, a single test lasts about 60 minutes, and a time sequence matrix is formed, as shown in the accompanying drawings. Figure 5

[0114] Step S1.2: the original data collected are preliminarily labeled, and an improved constant threshold method based on engineering experience and historical excluded fault data and Vg1Totl<30 or Vg2Totl<40 are used for sample classification. When the vibration main component amplitude continuously exceeds the preset safety threshold in a period of time and the vibration suddenly increases in the process of speed drop, the sample is determined as a "vibration overrun fault" sample; the data conforming to the normal operation specification are marked as "normal working condition" samples. Thus, the preliminary separation of fault data and normal data is realized, and an original data set with clear labels is formed; in the embodiment, the safety threshold of the vibration main component is 30.

[0115] ​Step S2: The raw data set collected in step S1 with clear labels is preprocessed to be suitable for the form of the graph auto-encoder digital twin architecture, and through data cleaning, time alignment and standardization processing, combined with sliding window slicing to construct multi-channel time sequence samples, to ensure that the input data is reliable and consistent, so that the graph auto-encoder can accurately learn the data characteristics of the normal state of the engine and realize effective anomaly detection; Specifically:

[0116] Step S2.1: For the vibration over-limit fault occurring during the operation of the aero-engine, a plurality of sub-data sets are established using the raw data set collected in step S1 with clear labels, and for the missing values, jump points and abnormal values obviously deviating from the physical law existing in the data set, the median filtering in the sliding window is adopted;

[0117] Step S2.2: The filtered data is identified and repaired for a small amount of missing or abnormal data points by using a linear interpolation method in combination;

[0118] Step S2.3: Since there may be sampling time offset problems between the vibration signal and other sensor signals in the raw data set, linear interpolation alignment is needed based on a unified timestamp, we adopt to uniformly resample all sensor data to every 0.1 second as a time point, and to align the time with the throttle lever angle signal as the reference, to solve the problem of different synchronization of multi-source signals;

[0119] Step S2.4: In view of the problem of large dimension difference and uneven numerical range distribution of different sensor data in the raw data set, Z-score standardization method is adopted for dimensionless processing, so that all input data are distributed in a similar interval, to improve the convergence speed and stability of model training;

[0120] Step S2.5: Then, according to the dynamic characteristics of the typical operation process of the engine, a reasonable sliding window length is set, and the continuous time sequence data is cut into subsequence samples with overlapping relationship. We adopt to cut the continuous data into 10-second windows with a sliding step of 1 second, to generate overlapping time sequence samples, each sample containing the evolution information of multiple variables in the dynamic process;

[0121] Step S2.6: Finally, the constructed subsequence samples are divided into training set, validation set and test set, we adopt to take three normal working condition labeled samples from the data of 25 groups of csv files as the training set and test set, and the rest as the test set;

[0122] Step S3: Constructing the graph auto-encoder digital twin, which consists of aero-engine sensor network state correction graph, deep learning model based on graph auto-encoder and feedback decision control module, as shown in the attached Figure 6The correlation between the sensor data in the training set and the prior knowledge of the aero-engine is fused to construct a sensor network state correction graph. Then, the multi-dimensional time series sensor data optimized by the state correction graph is input into a deep learning model based on a graph autoencoder. Finally, a feedback decision control module is introduced as a closed-loop interface between the digital twin and the physical system to complete the establishment of the overall digital twin architecture. The specific construction process is as follows:

[0123] Step S3.1.1: The correlation between the prior knowledge of the aero-engine and the sensor data is fused to construct an engine sensor network state correction graph. Specifically:

[0124] Based on the prior knowledge of the structural topology, power transmission path and typical fault propagation mechanism of the aero-engine, mechanism-driven feature selection is performed on the sensor data in the training set. In this embodiment, 14 core vibration monitoring parameters are selected as the key nodes of the sensor network. The specific parameters are shown in Table 1.

[0125] Table 1

[0126]

[0127] Step S3.1.2: The Pearson correlation coefficient method suitable for information transmission is selected as the data-driven correlation analysis to obtain the node connection relationship of the deep learning network graph.

[0128] Step S3.1.3: Finally, the node selection of the deep learning network graph guided by the prior knowledge is combined with the edge connection obtained by data-driven to generate a sensor network state correction graph that fuses mechanism cognition and measured characteristics.

[0129] Step S3.2: The multi-dimensional time series sensor data optimized by the state correction graph is input into a deep learning model based on a graph autoencoder.

[0130] The graph structure-based autoencoder is used to perform nonlinear feature learning and low-dimensional representation modeling on the multi-dimensional time series sensor data optimized by the state correction graph. The autoencoder is composed of two parts: two layers of encoder (Encoder) and two layers of decoder (Decoder).

[0131] Step S3.2.1: First, the multi-dimensional time series sensor data is mapped to an embedding vector through linear projection, and the position encoding is added. The linear projection is a fully connected layer with 1 network layer, whose input is the processed 14-dimensional time series sensor data, used to map the preprocessed 14-dimensional standardized data to a high-dimensional space of 256 dimensions, and then add the sine position encoding.

[0132] Step S3.2.2: The autoencoder is composed of 2 layers of encoders (Encoder), which are stacked by 4 layers of multi-head self-attention modules and feedforward networks, each layer containing 8 attention heads, capable of capturing complex cross-time and cross-variable dependencies between sensor channels, the feedforward network is composed of two fully connected layers, the intermediate hidden dimension is expanded to 512, the activation function adopts ReLU to enhance the nonlinear expression ability, while the residual connection and layer normalization operation are reserved between layers to ensure the stability of training;

[0133] Step S3.2.3: The autoencoder is composed of 2 layers of decoders (Decoder), which are also composed of 2 layers of graph decoding layers stacked, each layer integrating self-attention mechanism, cross-attention mechanism and feedforward network. Among them, the self-attention is used to model the internal time sequence dependence of the output sequence, and the cross-attention is responsible for fusing the context feature representation of the encoder output, realizing the effective extraction and utilization of key information. The feedforward network structure is consistent with the encoder, containing two fully connected layers, the intermediate dimension is 512, cooperating with GeLU activation function and residual connection, ensuring the consistency and expression ability of the features in the decoding process;

[0134] Step S3.2.4: Finally, the output reconstruction layer is constructed, which is linearly mapped to the original feature space, and the reconstruction layer maps the 256-dimensional autoencoder output back to the 14-dimensional space with a fully connected network, and the number of network layers is set to 1;

[0135] Step S3.3: A feedback control decision mechanism is designed for the graph autoencoder reconstruction error;

[0136] On the basis of modeling the normal running state of the engine by the graph autoencoder, the application constructs a feedback decision control mechanism driven by reconstruction error, which is used for real-time monitoring of the engine health state and identifying potential faults;

[0137] Firstly, the autoencoder is trained only using the training set composed of fault-free test data in the training stage, so that it can accurately reconstruct the input signal conforming to the normal running mode;

[0138] Secondly, in the test stage, input the test set containing multiple groups of normal and abnormal working conditions into the trained model to obtain the corresponding reconstruction output;

[0139] Finally, the maximum reconstruction error of each test set data is calculated according to the defined reconstruction error formula;

[0140] The adaptive threshold is set by the mean plus twice the standard deviation. When the reconstruction error of the test sample exceeds the threshold, it is determined to be abnormal, and the system immediately triggers the feedback action: generates an alarm signal, records the time of abnormal occurrence, the current working condition and the abnormal sensor channel with the highest contribution, and can push the diagnosis result to the ground maintenance system or flight control system to realize early warning and closed-loop response.

[0141] Step S4: training the graph auto-encoding digital twin;

[0142] Step S4.1: First, according to the training set obtained in step S2.6, ensure that the data does not contain any abnormal disturbance or fault feature. According to the sample size of the training set and the hardware resource configuration, set the hyperparameters of the model training: set the training batch size to 128 to balance the training efficiency and gradient stability; set the initial learning rate to 0.0001 to ensure that the model can converge smoothly in the early stage of training;

[0143] Step S4.2: Second, the Adam optimizer is used to iteratively optimize the learnable parameters of the auto-encoding digital twin, and the loss function is the mean square error (MSE) between the input and the reconstruction output. During the training process, the reconstruction loss on the validation set is calculated at the end of each round, and the model parameters with the lowest loss are automatically saved to ensure that the optimal model is obtained. To improve learning efficiency, a learning rate scheduling strategy based on validation loss change is introduced: when the validation loss does not decrease for 5 consecutive training rounds, the learning rate is multiplied by 0.5 for decay. At the same time, to prevent model overfitting and save training resources, set the early stopping mechanism: if the validation loss does not improve for 10 rounds, terminate the training in advance to avoid invalid iterations. Finally, save the trained model structure and parameters;

[0144] Step S4.3: Finally, after the model training is completed, the validation set obtained in step S2.6 is used to verify the model, the model parameters are adjusted according to the verification result and retrained to obtain the optimal model parameters, and the model is saved.

[0145] Step S5: test the auto-encoding digital twin to verify its effectiveness;

[0146] Step S5.1: First, input the test set data obtained in step S2.6 into the graph auto-encoding digital twin trained and saved in step S4.3 to obtain the corresponding reconstruction result. Then, the real data and the reconstruction data are substituted into the maximum reconstruction error calculation formula in step S3.3 to obtain the reconstruction error of the test set. Figure 7 The reconstruction error value of the trained auto-encoding digital twin model on the normal operation data of the engine is shown in the figure, which indicates that the model has good reconstruction ability for normal operation data, and the peak reconstruction error in all time windows is 53.7768; Figure 8The comparison between the reconstructed signal and the original signal is shown, and the figure shows that the signal reconstructed by the model is basically consistent with the original signal, and the reconstruction effect is good; Appendix ​ The reconstruction error value of the trained self-encoding digital twin model on the engine fault operation data is shown, and the figure shows that the model has poor reconstruction ability for fault operation data, and the peak value of the reconstruction error in all time windows is 249; Appendix ​ The comparison between the reconstructed signal and the original signal is shown, and the figure shows that the signal reconstructed by the model is basically consistent with the original signal, and the reconstruction effect is good; Appendix

[0147] Step S5.2: Then, set the threshold according to step S3.3;

[0148] Set the threshold, ; ; That is , a binary decision rule is established;

[0149] Step S5.3: Secondly, according to the maximum reconstruction error calculated in step S5.1, and in combination with the threshold set in step S5.2, the data state is determined according to the binary decision rule given in step S3.3;

[0150] Step S5.4: According to S5.3, the state of the test data is judged, and compared with the true state of the data, the false alarm rate and the false alarm rate of the method are analyzed. This experiment takes out 25 groups of data of the same batch of aero-engine experiments, mixes normal data and abnormal data, uses the established digital twin to reconstruct, calculates the error before and after reconstruction, and according to the set threshold, distinguishes normal and abnormal data. Appendix ​ The distribution of 25 groups of aero-engine experimental data is shown, wherein the red points are fault data, the gray points are normal data, and the red dotted line is the threshold line.

[0151] The above examples are only used to illustrate the specific embodiments of the present application, and are not limited to the scope of the patent protection of the present application. Those skilled in the art can make various modifications, equivalent replacements or reasonable modifications to the above examples within the scope of the technical solutions of the present application without deviating from the basic principles and concepts of the present application, and these shall belong to the protection scope of the present application.

Claims

1. A method for constructing a graph auto-encoding digital twin for diagnosing an aero-engine vibration out-of-limit fault, characterized in that, The construction method comprises an engine sensor network state correction graph, a self-encoding model and a feedback decision, and comprises the following steps: Step S1: using a multi-source sensor network on an aero-engine, collecting running state data of the aero-engine under multiple running states to obtain original data, and performing data sample feature selection to form an original data set with clear labels; Step S2: preprocessing the original data set with clear labels collected in step S1 to process it into a form suitable for a graph self-encoding digital twin architecture, and through data cleaning, time alignment and standardization processing, combining with a sliding window slicing to construct a multi-channel time sequence sample, so that the graph self-encoder can accurately learn the data features of the normal state of the engine and realize effective anomaly detection; Step S3: constructing a graph self-encoding digital twin, which is composed of an aero-engine sensor network state correction graph, a self-encoder deep learning model and a feedback decision control module; first, the correlation between the sensor data in the training set is fused with the prior knowledge of the aero-engine to construct a sensor network state correction graph; then, the multi-dimensional time sequence sensor data optimized by the state correction graph are input into the deep learning model based on the graph self-encoder; finally, the feedback decision control module is introduced as a closed-loop interface between the digital twin and the physical system, and the establishment of the overall digital twin architecture is completed; Step S4: training the graph self-encoding digital twin; Step S5: testing the graph self-encoding digital twin, and using the constructed digital twin to perform fault prediction.

2. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 1, characterized in that, The step S1 is specifically: Step S1.1: relying on a multi-source sensor network arranged on an aero-engine, synchronously collecting running state data of the engine under different working conditions as original data; the multi-source sensor network comprises a vibration acceleration sensor, a speed sensor, a temperature sensor and a pressure sensor; the different working conditions comprise ground slow running, climbing, cruising and deceleration parking; the collected running state data covers vibration characteristics, speed changes, thermal parameters and aerodynamic performance indicators of key components, forming a high-dimensional, multi-channel time sequence data set; the time sequence data set contains multiple sensor signals; Step S1.2: to construct high-quality training sample sets and test sample sets, the collected original data is preliminarily labeled; an improved constant threshold method based on engineering experience and historical fault elimination data is used to set a safety threshold for the vibration principal component; When the amplitude of the vibration principal component continuously exceeds the preset safety threshold in a period and is accompanied by a sudden increase in vibration during the speed drop process, it is determined as a "vibration overrun fault" sample; the rest of the data that meets the normal operation specification is marked as a "normal working condition" sample; The preliminary separation of fault data and normal data is realized to form an original data set with clear labels.

3. The method for constructing a graph-encoded digital twin for diagnosing excessive vibration faults in aero-engines according to claim 2, characterized in that, In the step S1: In the step S1.1, the time series dataset is represented as a two-dimensional data sequence where T is the number of time sampling points, N is the number of sensor channels, i.e., the dimension of the feature variables, each column corresponds to a physical feature variable, and each row represents the synchronous observation values of the variables at a certain time. In the step S1.2, the vibration principal component is the effective value of the vertical and horizontal vibration velocity or acceleration; the safety threshold of the vibration principal component is 30.

4. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 2, characterized in that, The step S2 is specifically: Step S2.1: For the vibration over-limit fault generated in the operation of the aero-engine, a plurality of sub-datasets are established using the original dataset with clear labels collected in step S1, and for the missing values, jump points and abnormal values obviously deviating from the physical law existing in the dataset, a sliding window median filter is used to obtain the filtered data ; Step S2.2: filtering the data The identification and repair are performed by the method combined with linear interpolation to obtain the repaired and filled data. Step S2.3: Linear interpolation alignment based on uniform timestamp to ensure the synchronization of each sensor data signal at the same time step; Step S2.4: Z-score standardization method is used for dimensionless processing to make all input data distributed in similar intervals, improving the convergence speed and stability of model training; Step S2.5: According to the dynamic characteristics of the typical operation process of the engine, the length of the sliding window is set, and the continuous time series data is divided into subsequence samples with overlapping relationship; Each subsequence sample contains multiple channels of sensor multi-step historical information as the input sequence of the graph autoencoder, effectively capturing the dynamic evolution trend and potential fault mode of the vibration signal; Step S2.6: The constructed subsequence samples are divided into training set, validation set and test set, wherein the training set and the validation set only use the test data confirmed to be fault-free to ensure that the autoencoder learns the "normal" operation mode of the engine; The test set contains multiple normal and abnormal working conditions, which is used to evaluate the abnormal detection ability and generalization performance of the model.

5. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 4, characterized in that, In the step S2: The median filtering in the step S2.1 is shown in formula (1); (1)； wherein, is the original signal, is the sliding step size ， is the filtered signal; The identification and repair in the step S2.2 is shown in formula (2); (2)； wherein, is the filtered invalid data point, and are the preceding and succeeding valid points, is the repaired interpolated data; The Z-score standardization method for dimensionless processing in the step S2.4 is shown in formula (3); The similar interval is with mean value 0 and variance 1; (3)； wherein, is the observation value of the different sensor data in the original data set at time , is the result after Z-score standardization of the feature , is the mean value of the feature , and N is the sample quantity, is the standard deviation of the feature , is the unified timestamp corresponding to the i-th time sampling point; In the step S2.5, the typical operation process is acceleration, steady state, and deceleration.

6. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 4, characterized in that, The step S3 is specifically: Step S3.1: The correlation between prior knowledge of an aero-engine and sensor data is fused to construct an engine sensor network state correction graph; Step S3.2: In order to effectively capture the long-range dependence and cross-channel dynamic correlation in the multi-sensor time series of the aero-engine, a graph structure based autoencoder is used for nonlinear feature learning and low-dimensional representation modeling of the multi-dimensional time series sensor data optimized by the state correction graph, that is, the multi-dimensional time series sensor data optimized by the state correction graph is input into the deep learning model of the autoencoder; The autoencoder is composed of an encoder Encoder and a decoder Decoder, both of which take multi-head self-attention mechanism MHSA as the core module, which can adaptively model the time evolution pattern and potential coupling relationship between sensor variables; Step 3.3: On the basis of modeling the normal operation state of the engine by the graph autoencoder, a feedback decision control mechanism based on reconstruction error driving is constructed for real-time monitoring of the engine health state and identification of potential faults.

7. The method for constructing a graph-encoded digital twin for diagnosing excessive vibration faults in aero-engines according to claim 6, characterized in that, In the step S3: The step S3.1 is specifically: Step S3.1.1: Mechanism-driven feature screening on sensor data in the training set; select key variables directly related to vibration response or having a modulating effect on fault evolution as deep learning network nodes; the key variables include: vibration acceleration signals at the fan case and high-pressure compressor mounting fulcrum, low-pressure rotor speed and high-pressure rotor speed , throttle lever angle PLA, and flight altitude or inlet total pressure Step S3.1.2: In order to capture the mutual relationship between different nodes in the network graph, the Pearson correlation coefficient method is selected to perform data-driven correlation analysis on the sensor data in the training set to obtain the node connection relationship of the deep learning network graph; Step S3.1.3: Combine the node selection of the deep learning network graph guided by prior knowledge with the edge connection obtained by data-driven to generate a sensor network state correction graph that combines mechanism cognition and measured characteristics; The step S3.2 is specifically: Step S3.2.1: Map the multi-dimensional time series sensor data to an embedding vector through linear projection, as shown in formula (5); (5)； wherein, represents an input embedding vector; is a learnable embedding weight matrix; is a position encoding used to preserve the sequential information of the time series, usually constructed with sine / cosine functions; Step S3.2.2: In the encoder, input the embedding sequence By the N-layer multi-head self-attention and the feed-forward network module, high-level abstract features of the input sequence are extracted; the output of the single-layer encoder is represented as Z, as shown in equation (8); (8)； wherein, is an encoder feed-forward network, is multi-head self-attention; Step S3.2.3: The decoder receives the latent representation Z output by the encoder and reconstructs the original input sequence step by step through a similar attention mechanism as shown in equation (12); (12)； wherein, is the reconstructed data; is the weight coefficient; is an encoder consisting of N layers of stacked mask self-attention, cross-attention, and feed-forward networks, whose single layer structure is represented as shown in equation (13); (13)； wherein, is a start embedding for the input sequence; is a multi-head self-attention module with temporal masking; CA is a cross-attention module; is a decoder feed-forward network; Step S3.2.4: Restore the reconstructed representation obtained by stacking the N-layer decoder to the original feature space through linear mapping, as shown in formula (18); (18)； wherein, is the model reconstruction result; is the hidden representation output by the Nth layer of the decoder; is the output projection matrix of the decoder for mapping the hidden features back to the original sensor feature dimension model is trained by minimizing the reconstruction error; The loss function is defined as mean square error (MSE), as shown in formula (19); (19)； wherein, is real data, is reconstructed data, LOSS is a loss function; Through the graph autoencoder, the normal operation mode of the sensor data can be learned in an unsupervised or semi-supervised mode, and the latent feature representation Z sensitive to anomalies is extracted, providing high-discriminative feature input for subsequent anomaly detection and state evaluation; Step 3.3: On the basis of modeling the normal operation state of the engine by the graph autoencoder, a feedback decision control mechanism driven by reconstruction error is constructed for real-time monitoring of the engine health state and identification of potential faults; specifically: First, the autoencoder is trained only using the training set composed of fault-free test data in the training stage, so that it can accurately reconstruct the input signal conforming to the normal operation mode; Second, in the test stage, input the test set containing multiple normal and abnormal working conditions into the trained model to obtain the corresponding reconstruction output; Finally, calculate the maximum reconstruction error of each test set data according to the defined reconstruction error formula, as shown in formula (21); (21)； wherein, is the real data, is the reconstructed data, is the window length, k is the number of the window, is the maximum reconstruction error in each window of the test set data. At the same time, based on the reconstruction error distribution of the normal samples in the training set, an adaptive threshold is determined using statistical methods, as shown in formula (22); (22)； wherein, is the average of the reconstruction errors, is an adjustable coefficient, is the variance of the reconstruction errors; Accordingly, the binary decision rule is established as shown in formula (23); (23)； wherein, is the maximum reconstruction error, which ranges between 0 and 100; When the abnormal state is judged according to the binary decision rule, the feedback control module is triggered, and the following operations are performed: output an alarm signal and record the time, working condition and significantly deviated sensor channel of the abnormal occurrence.

8. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 7, characterized in that, The step S3 in the method is specifically: In the step S3.2.1, the sine / cosine function is used to construct as shown in formulas (6) and (7); (6)； (7)； wherein, is the embedding dimension of the model; In the step S3.2.2, the multi-head self-attention are defined as shown in equations (9) and (10); (9)； (10)； where h is the number of attention heads, is the output projection matrix, is the Query of the i-th head, is the Key of the i-th head, is the Value of the i-th head, is the learnable projection, is the dimension of Query / Key for each attention head, t is the time step, and T is the total length of time. In the step S3.2.2, the encoder feedforward network is composed of two linear mappings and a nonlinear activation function, as shown in formula (11); (11)； wherein, is the output, is the first layer linear weight, is the first layer bias, is the first layer linear weight, is the first layer bias, is the hidden layer dimension of the feedforward network, size 4 ;​ In the step S3.2.3, the multi-head self-attention module with time masking is as shown in formulas (14) and (15); (14)； (15)； Wherein, M is an upper triangular masking matrix for masking future time information; In the step S3.2.3, the cross-attention module is as shown in formula (16); (16)； where Q is the decoder input and Z is the encoder output; the Key and Value of cross attention are obtained by linear mapping; In the step S3.2.3, the decoder feedforward network As shown in equation (17); (17)； wherein M is the output of the cross-attention module, is the first layer linear weight, is the first layer bias, is the first layer linear weight, is the first layer bias.

9. The method for constructing a graph-encoded digital twin for diagnosing vibration exceedance faults in aero-engines according to claim 8, characterized in that, The step S4 is specifically: Step S4.1: According to the size of the training set obtained in step S2.6, select appropriate hyperparameters for the model, including training batch size, iteration number, learning rate and initial weight; Step S4.2: input training set The training is performed using Adam optimizer based on the principle of stochastic gradient descent to train and optimize the weights of the auto-encoding digital twin, so that the target LOSS in S3.2.4 is minimized. Step S4.3: After the model training is completed, the validation set obtained in step S2.6 is used to verify the model, the model parameters are adjusted according to the verification result and retrained to obtain the optimal model parameters, and the model is saved.

10. The method for constructing a graph-encoded digital twin for diagnosing excessive vibration faults in aero-engines according to claim 9, characterized in that, The step S5 is specifically: Step S5.1: Input the test set data obtained in step S2.6 into the graph autoencoder digital twin saved in step S4.3 to obtain the corresponding reconstruction result; then, the real data and the reconstruction data are substituted into the maximum reconstruction error calculation formula in step S3.3 to obtain the reconstruction error of the test set; Step S5.2: Set the threshold according to step S3.3; Step S5.3: According to the maximum reconstruction error calculated in step S5.1, and combined with the threshold set in step S5.2, the data state is determined according to the binary decision rule given in step S3.3; Step S5.4: According to the judgment of step S5.3, the state of the test data is judged, and compared with the true state of the data, the false negative rate and the false positive rate of the method are analyzed.

Images