Aero-engine low-pressure turbine inducer fracture fault early warning method and system

Abstract

The invention discloses an aero-engine low-pressure turbine inducer fracture fault early warning method and system, and belongs to the technical field of aero-engine component state monitoring and fault diagnosis. The method comprises the following steps: S1, acquiring original vibration data of an aero-engine in an operation process by using a vibration sensor on the aero-engine; s2, preprocessing the original vibration data to obtain a preprocessed data set; s3, constructing a Transform auto-encoder model consisting of an input mapping layer, a Transform encoding-decoding module and an output reconstruction layer, wherein the Transform auto-encoder model is used for carrying out early warning on the fracture fault of the low-pressure turbine inducer of the aero-engine; s4, carrying out training, verification and fine tuning on the Transform auto-encoder model; and S5, defining an early warning threshold value by using a Transform auto-encoder model, and carrying out early warning on the fracture fault of the low-pressure turbine inducer of the aero-engine. According to the invention, high-precision early recognition of faults of the low-pressure turbine inducer can be realized, and a visual interaction interface is designed to enhance practicability; the problem of scarcity of fault samples in actual engineering is solved, and the applicability is high.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine component condition monitoring and fault diagnosis technology, and relates to an early warning method and system for low-pressure turbine guide vane rupture fault in aero-engines. Background Technology

[0002] As the core power plant of an aircraft, the aero-engine is one of the key systems determining flight safety, mission success rate, and combat effectiveness. Its operating environment is extremely harsh, requiring long-term stable operation under high temperature, high pressure, high speed, and complex aerodynamic loads. Therefore, it places extremely high demands on the structural integrity and dynamic performance of each component. The low-pressure turbine guide vane, as a crucial supporting component connecting the low-pressure turbine and the fan rotor, not only guides airflow and balances axial forces but also directly participates in constructing the vibration transmission path of the entire engine. This component is typically located in a high-cycle fatigue stress field and is susceptible to alternating loads, thermal shock, fretting wear, and material aging during long-term service, gradually leading to progressive failures such as microcracks, loose connections, and even structural fractures. In fact, once the low-pressure turbine guide vane of an aero-engine ruptures, it will cause an imbalance in the mass distribution of the engine's rear section, triggering a significant rotational imbalance response, manifested as violent fluctuations in vibration signals. If such vibration anomalies are not effectively identified and intervened in the early stages, they may further induce a chain of failures such as bearing damage, seal failure, and even rotor rubbing, potentially leading to in-flight engine shutdown or non-containment failure, threatening flight safety. Therefore, establishing a highly sensitive and robust fault monitoring mechanism to achieve real-time assessment and early warning of the low-pressure turbine guide vane status has become a key research focus and urgent engineering need in the field of aero-engine health management.

[0003] In recent years, several sudden malfunctions during ground testing have further highlighted the inadequacy of existing monitoring methods. Take, for example, the serious malfunction of the X-type engine during a prolonged performance evaluation test: During a test, the engine smoothly increased thrust from idle to a high-power ramp-up state and operated stably for approximately two minutes when the test personnel heard an abnormal noise in the control room. Monitoring video showed a sudden open flame in the left-hand side of the outer bypass duct rear casing. An emergency shutdown and fire extinguishing were immediately initiated. A visual inspection of the engine after extinguishing the fire revealed that the turbine low-pressure guide outer casing and the corresponding outer bypass duct rear casing on the left-hand side of the engine were punctured. The circumference of the puncture in the outer bypass duct rear casing was approximately 45 cm, and the width was approximately 8-10 cm. Debris from the low-pressure turbine guide vane and blades was found in the test workshop. Approximately half of the low-pressure turbine guide vane fragment was lodged at the puncture in the outer bypass duct rear casing, and nearby external piping was damaged and leaking.

[0004] In-depth analysis revealed that the root cause of the failure was the lack of axial tightness between the low-pressure turbine guide vane and the low-pressure turbine disk. During long-term engine operation, the connecting pin experienced continuous wear due to a design flaw in the circumferential limiting structure, leading to a vicious cycle of wear between the pin, the U-shaped groove on the disk, and the guide vane lug. Furthermore, unclear pin assembly processes and significant positional deviations in the U-shaped groove further accelerated the degradation of the mating interface. With accumulated operating time, the positioning function between the guide vane and the disk gradually failed, causing the guide vane to skew and rub against the turbine support structure, triggering fatigue crack propagation. Ultimately, under high-stress conditions, the guide vane ruptured and flew out, penetrating the low-pressure turbine outer casing and the outer bypass casing. This failure not only caused severe engine damage but also directly led to the interruption of the year's critical long-term evaluation test. This incident exposed significant shortcomings in the current fault monitoring system's ability to identify early signs, perceive dynamic evolution, and provide early warnings of critical component conditions.

[0005] Currently, in practical engineering applications, fault detection of critical components of aero-engines mainly relies on fixed threshold discrimination methods based on experience-set key parameters. This method sets fixed or segmented thresholds for key sensor parameters (such as vibration amplitude, temperature, and pressure), triggering an alarm when the monitored value exceeds a preset threshold. While this method is simple to implement and logically clear, it reveals several limitations in practical applications: First, fixed thresholds are difficult to adapt to changes in operating conditions such as different throttle states, flight altitudes, and atmospheric conditions during the engine's full envelope operation, easily leading to false alarms or missed alarms; second, early faults often manifest as weak signal characteristic shifts or spectral structure changes, while traditional threshold methods only focus on amplitude abrupt changes, making them insensitive to slowly evolving faults and unable to provide early warnings; third, complex coupling relationships exist between multi-channel vibration parameters, and threshold judgments for some single sensor parameters ignore system-level dynamic correlation characteristics, resulting in diagnostic results lacking comprehensiveness and interpretability.

[0006] To overcome the above shortcomings, data-driven intelligent fault diagnosis technology has gradually become a research hotspot in recent years. Machine learning-based state monitoring methods learn normal behavior patterns from historical operating data and achieve anomaly identification without the need for precise physical modeling, showing good application prospects. For example, the literature [1] uses Long Short-Term Memory Network (LSTM) to sequentially model the sensor time series, effectively captures time-dependent features through its internal memory mechanism, realizes fault prediction of engine rotor system, and achieves good diagnostic results. Reference [2] = {WuJ, Kong L, Kang S, et al. Aircraft engine fault diagnosis model based on 1DCNN-BiLSTM with CBAM[J]. Sensors, 2024, 24(3): 780} constructs a fault diagnosis model that combines a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM). By utilizing the local receptive field and weight sharing mechanism of the convolutional network, it exhibits excellent feature extraction capabilities when processing two-dimensional features such as the time-frequency diagram and envelope spectrum of vibration signals, and significantly improves the fault classification performance based on image-based signal representation. Reference [3] = {Chen J, Yan Z, Lin C, et al. Aero-engine high speed bearing fault diagnosis for data imbalance: A sample enhanced diagnostic method based on pre-training WGAN-GP[J]. Measurement, 2023, 213: 112709} addresses the problem of scarce and unbalanced vibration signal samples for high-speed bearings of aero-engines. It introduces the pre-training and generative adversarial enhancement (WGAN-GP) method to expand the data samples, thereby improving the fault identification accuracy and generalization ability of the model under conditions of few samples.

[0007] It is worth noting that although the aforementioned deep learning methods have demonstrated certain performance in specific tasks, they still face common challenges in early fault warning applications for critical components of aero-engines. For example, they are not sensitive to early fault characteristics and have weak recognition capabilities. Furthermore, they require a large number of high-quality labels, making them difficult to implement in real-world engineering scenarios where fault samples are scarce.

[0008] To effectively prevent such aero-engine failures, it is urgent to develop a novel fault detection method adapted to real-world engineering conditions. Addressing the common problems of scarce fault samples and missing labels in practical engineering, this paper proposes a novel fault detection method for low-pressure turbine guide vanes in aero-engines based on a Transformer autoencoder. This method employs an unsupervised learning strategy, requiring only multi-channel vibration data under normal operating conditions to construct a baseline model, eliminating the need for fault samples and manual annotation, thus significantly enhancing its engineering applicability. Leveraging the powerful sequence modeling and high-precision reconstruction capabilities of the Transformer autoencoder, the method can effectively capture early, weak anomaly signals, achieving accurate detection of the deviation between the current operating state and the normal mode, thereby providing anomaly warnings. Simultaneously, this method exhibits good adaptability to operating conditions and cross-sample generalization performance, supports interactive visualization analysis, clearly presents the fault evolution process, and improves the interpretability of monitoring results. Overall, this method is expected to significantly improve the accuracy, robustness, and intelligence of aero-engine fault diagnosis, enhance operational practicality, and provide strong protection for flight safety. Summary of the Invention

[0009] This invention aims to overcome the shortcomings of existing technologies and provides a method for early warning of low-pressure turbine guide vane rupture faults in aero-engines based on a Transformer autoencoder and a design of a visual interactive system. It constructs a vibration signal feature sequence, trains a deep neural network model in an unsupervised manner, utilizes reconstruction errors to achieve high-precision early identification of low-pressure turbine guide vane faults, and designs a visual interactive interface to enhance practicality.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for detecting and warning of low-pressure turbine guide vane rupture fault in an aero-engine, the method comprising the following steps:

[0012] Step S1: Collect raw vibration data of the engine during operation using vibration sensors on the aircraft engine; specifically:

[0013] During the operation of an aero-engine, a piezoelectric axial vibration acceleration sensor installed near the engine's front bearing is used to collect the total front axle vibration, and a piezoelectric crystal vertical vibration acceleration sensor installed at a key support structure at the rear of the engine is used to collect the total rear vertical vibration. This ensures that the dynamic response of the aero-engine's vibration can be captured, obtaining raw vibration data. The raw vibration data is stored in time-series format, covering multiple operating conditions from engine start-up, idle, climb to maximum thrust and shutdown, forming the dataset for subsequent analysis.

[0014] Step S2: Preprocess the raw vibration data obtained in Step S1, including signal denoising and extraction, data standardization, and sliding time window division, to obtain the preprocessed dataset; specifically:

[0015] Step S2.1: First, filter technology is used to remove high-frequency noise and electromagnetic interference from the original vibration data, retaining the effective vibration components. Linear interpolation is then performed on the filtered vibration data to fill in the missing data points. The linear interpolation operation is shown in the following formula.

[0016] (1)

[0017] In equation (1), Indicates the sampling time of the missing data points; Indicates the numerical value to be filled; Indicates the sampling time of the point preceding the missing value; This indicates the value of the point preceding the missing value. Indicates the sampling time of the point following the missing value; This indicates the value of the point following the missing value.

[0018] Step S2.2: Then, extract the fundamental and harmonic vibration components related to the front axle vibration and rear sag vibration from the data processed in step S2.1 to generate the required... Each of the obtained feature parameters is standardized as shown in the following formula to obtain... A continuous time series;

[0019] (2)

[0020] In equation (2), the subscript Indicates the index of the value in the feature parameter; This indicates that the first characteristic parameter One value; This represents the mean of the feature parameters in this dimension; This represents the standard deviation of the feature parameter in this dimension; This represents a standardized continuous time series;

[0021] Step S2.3: Finally, select the time window length and sliding step size for the standardized continuous time series to perform sliding slices, generating preprocessed slices that conform to the input dimensions of the Transformer autoencoder model. Dimensional standardized dataset.

[0022] Step S3: Construct a Transformer autoencoder model consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, to provide early warning of low-pressure turbine guide vane rupture faults in aero-engines; specifically:

[0023] Construct a Transformer autoencoder model to learn the normal operation of the low-pressure turbine guide vane of an aero-engine. The time evolution law of vibration parameters is studied, and high-precision signal reconstruction is achieved. The Transformer autoencoder model consists of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, constructed in the following order:

[0024] Step S3.1: Construct the input mapping layer;

[0025] The input mapping layer is a fully connected network, and its input is the preprocessed data generated in step 2. 3D normalized dataset, used to preprocess the data Dimensional standardized data mapping to In a higher-dimensional space, we obtain High-dimensional variables in a high-dimensional space; its core function can be summarized as performing the function mapping shown in the following formula;

[0026] (3)

[0027] In equation (3), Indicates input Dimensional standardized data; Represents the learnable weight matrix; Represents the bias vector; Indicates mapping to High-dimensional variables in a high-dimensional space;

[0028] Step S3.2: Construct the Transformer encoding-decoding module;

[0029] The Transformer encoder-decoder module is an encoder-decoder architecture model consisting of multiple Transformer encoder layers and multiple Transformer decoder layers. It is used to receive high-dimensional variables generated by the input mapping layer and reconstruct the sequence.

[0030] First, a position encoding parameter is introduced and superimposed on the input high-dimensional variable to preserve the temporal information of the time series. The position encoding is shown in the following formula.

[0031] (4)

[0032] (5)

[0033] In equations (4) and (5), Indicates the position index in the time series; Indicates a dimension index; Indicates the total number of dimensions for positional encoding; Represents the position encoding matrix; This indicates that the position index in the position encoding matrix is And the dimension index is The corresponding scalar value; This indicates that the position index in the position encoding matrix is And the dimension index is The corresponding scalar value;

[0034] Then, using multiple Transformer encoder layers, the high-dimensional variables in the high-dimensional space after stacking position encoding are transformed into a low-dimensional representation and input into a latent space. Specifically, each Transformer encoder layer first assigns a key, value and query vector to each input, then performs a linear transformation and inputs it into a self-attention structure. Finally, the features obtained from the self-attention structure are concatenated and subjected to a linear transformation to obtain the low-dimensional representation in the latent space.

[0035] Finally, a multi-layer Transformer decoder layer is used, which receives the output of the Transformer encoder layer as the input to its cross-attention structure, and receives the same input sequence as the encoder as the input to its self-attention structure, and performs sequence reconstruction. Specifically, each Transformer decoder layer first assigns a key, value and query vector to each input, then performs a linear transformation and inputs it to the self-attention structure to model the internal dependencies of the input sequence, then fuses the output of the Transformer encoder layer through the cross-attention structure, and finally concatenates the features obtained from the cross-attention structure and performs a linear transformation to obtain the reconstructed high-dimensional output.

[0036] Furthermore, the core function of the Transformer encoder layer and the Transformer decoder layer is a multi-head attention mechanism, which performs attention calculation as shown in the following formula;

[0037] (6)

[0038] (7)

[0039] In equations (6) and (7), the subscripts Represents the attention head index; Indicates the first The output of each attention head; Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the first A learnable query projection matrix for each attention head; Indicates the first Learnable key projection matrix for each attention head; Indicates the first The learnable projection matrix of each attention head; This represents the query dimension for each header; For the number of attention heads; To output the projection matrix; This indicates a vector concatenation operation;

[0040] Step S3.3: Construct the output reconstruction layer;

[0041] The output reconstruction layer is a fully connected network used to receive the high-dimensional output of the Transformer encoder-decoder module and map the high-dimensional output to... The reconstructed output of the multidimensional signal ultimately achieves end-to-end reconstruction of the initial multidimensional input signal. Its core function can be described as executing the function mapping shown in the following formula.

[0042] (8)

[0043] In equation (8), This represents the high-dimensional representation of the decoder output; Represents the weight matrix of the reconstruction layer; Indicates the refactoring of bias terms; This represents the reconstructed output with the same dimension as the original input;

[0044] Step S4: Based on the preprocessed dataset from Step S2, train, validate, and fine-tune the Transformer autoencoder model constructed in Step S3; specifically:

[0045] The training objective of the Transformer autoencoder model for early warning of low-pressure turbine guide vane rupture in aero-engines is to minimize the preprocessed step. Dimensional Standardized Data and The mean square reconstruction error between the dimensional reconstruction outputs is expressed in the following form;

[0046] (9)

[0047] In equation (9), This indicates the batch size, which is the number of samples input into the neural network in a single run. This represents the time step for each sample; Indicates the index of the time step; Indicates the index of the sample; Indicates the first The sample at the th The actual input value at each time step; This represents the reconstructed value corresponding to the model; This represents the overall mean square error loss;

[0048] The specific process is as follows: First, select data from the preprocessed 14-dimensional standardized data obtained in step S2, specifically the data of the low-pressure turbine guide vane of the aero-engine during normal operation, and divide it into a training dataset and a validation dataset. Based on the size of the training dataset, select hyperparameters for training the Transformer autoencoder model, including the training batch size and learning rate. Second, optimize the parameters of the Transformer autoencoder model using an optimizer based on the training dataset. Third, validate the Transformer autoencoder model using the validation dataset, fine-tune the model hyperparameters based on the model's performance on the validation dataset, and retrain, setting a target error value. Finally, repeat the above training, validation, and fine-tuning process until the reconstruction error of the Transformer autoencoder model on the validation dataset is lower than the set target error value. The Transformer autoencoder training is then complete, and the optimal model parameters are saved. The target error value ranges from 0.005 to 0.01.

[0049] Step S5: Based on the Transformer autoencoder model trained in Step S4 for early warning of low-pressure turbine guide vane rupture in aero-engines, define the early warning threshold and issue an early warning; specifically:

[0050] The trained Transformer autoencoder model is used to define the warning threshold, and then used for early warning of low-pressure turbine guide vane rupture faults in aero-engines, specifically including:

[0051] Step S5.1: First, select a sufficient number of preprocessed data of the low-pressure turbine guide vane of the aero-engine during normal operation, load the optimal model parameters and input the data into the Transformer autoencoder model to obtain the corresponding reconstruction output;

[0052] Step S5.2: Next, calculate the reconstruction error value for each time window, and use the maximum reconstruction error value in all time windows within a single operation of the aero-engine as the low-pressure turbine guide vane anomaly score for this operation, thereby obtaining a set of low-pressure turbine guide vane anomaly scores corresponding to the normal operation data of the aero-engine low-pressure turbine guide vane.

[0053] Step S5.3: Then, calculate a warning threshold based on the obtained set. The warning threshold is defined as shown in the following formula;

[0054] (10)

[0055] In equation (10), This represents the maximum value of the low-pressure turbine guide vane anomaly score in the normal operation data of the aero-engine low-pressure turbine guide vane; This is a threshold coefficient, ranging from 1 to 1.05, used to compensate for model uncertainties; This indicates the set warning threshold;

[0056] Step S5.4: Finally, load the optimal model parameters, input the target data to be detected into the Transformer autoencoder model, and obtain the low-pressure turbine guide vane anomaly score of the target data. If rating Greater than the set warning threshold If the target data corresponds to an experimental low-pressure turbine guide vane of an aero-engine, it is determined that there is an abnormal risk; if the score is... Not greater than the set warning threshold If the target data is obtained, it is determined that the low-pressure turbine guide vane of the aero-engine in the experiment is operating normally.

[0057] A fault detection and early warning system for low-pressure turbine guide vane rupture in aero-engines is disclosed, comprising the following modules: a data acquisition and preprocessing module, a Transformer autoencoder model construction and training module, a reconstruction detection and anomaly scoring module, a threshold management and early warning decision module, and a visualization and interaction module; specifically:

[0058] The data acquisition and preprocessing module is primarily used to acquire and preprocess raw vibration data of the aero-engine from its operating environment. The module comprises two functional units: data acquisition and preprocessing. The data acquisition unit uses a piezoelectric axial vibration acceleration sensor installed near the engine's front bearing and a piezoelectric crystal vertical vibration acceleration sensor installed at the engine's rear critical support structure to collect the total vibration of the front and rear axles, saving it as time-series raw vibration data. The preprocessing unit performs filtering, linear interpolation, standardization, and sliding slicing on the raw vibration data to generate a preprocessed standardized dataset.

[0059] The Transformer autoencoder model construction and training module is integrated with the data acquisition and preprocessing module, encompassing three functional units: model construction, training optimization, and model storage. The model construction unit builds a Transformer autoencoder model based on a preset architecture, consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer. The training optimization unit uses a training set and a validation set formed after preprocessing normally operating vibration data to train, validate, and fine-tune the model's hyperparameters, aiming to minimize reconstruction error. The model storage unit saves the optimal model parameters after training for subsequent modules to access.

[0060] The reconstruction detection and anomaly scoring module is integrated with the data acquisition and preprocessing module and the Transformer autoencoder model construction and training module, encompassing three functional units: model loading, reconstruction detection, and scoring calculation. The model loading unit is responsible for loading the optimal model parameters. The reconstruction detection unit inputs the preprocessed vibration data to be detected into the model to obtain the corresponding reconstruction output. The scoring calculation unit calculates the reconstruction error for each time window and uses the maximum reconstruction error value across all time windows within a single aero-engine operation as the "low-pressure turbine guide vane anomaly score" for this mission, and outputs this score.

[0061] The threshold management and early warning decision module is connected to the reconstruction detection and anomaly score module, and includes two functional units: threshold calculation and early warning decision. The threshold calculation unit calculates the early warning threshold according to formula (10) based on the anomaly score set generated by a large amount of normal operation data. The early warning decision unit compares the anomaly score generated by the reconstruction detection and anomaly score module with the early warning threshold. If the anomaly score is greater than the early warning threshold, it is determined to be an operational anomaly. If the anomaly score is not greater than the early warning threshold, it is determined to be a normal operation. Finally, it outputs a normal operation or an anomaly signal.

[0062] The visualization interaction module integrates the data acquisition and preprocessing module, the reconstruction detection and anomaly scoring module, and the threshold management and early warning judgment module, encompassing two functional units: interactive testing and visualization of detection results. The interactive testing unit provides a graphical user interface, allowing users to manually select and import locally stored vibration data files and trigger a complete testing process for the selected file with a single click. The detection result visualization unit displays the detection results, including: a comparison curve between the original vibration signal and the model reconstruction signal, the reconstruction error value for each time window and its comparison with the early warning threshold, and the final risk assessment conclusion. This module provides users with flexible and intuitive analysis and verification tools, realizing a fully interactive operation from data selection and intelligent testing to result presentation.

[0063] The beneficial effects of this invention are as follows:

[0064] (1) This invention solves the problem of scarce fault samples in actual engineering and significantly improves the applicability of the method. This invention adopts an unsupervised learning paradigm. The core Transformer autoencoder model only needs to be trained using a large amount of vibration data under normal conditions, without any fault samples or manual labeling. This is in line with the engineering reality that there are few fault cases of key components of aero-engines and it is difficult to obtain effective fault data. It greatly reduces the data threshold and cost of technology implementation, making high-precision early warning possible in actual operation and maintenance.

[0065] (2) This invention provides a supporting visual interactive system, which enhances the practicality and acceptability of the system, allows users to flexibly import historical data files for testing, and clearly presents the comparison between the original signal and the reconstructed signal, the evolution of the reconstruction error, and the warning threshold line. This design transforms the model's judgment process into visual evidence that can be intuitively examined and compared, which not only facilitates technicians in verifying diagnostic results and tracing the fault initiation process, but also greatly improves the interpretability of the warning conclusions and the user's trust, effectively supporting the closed loop from intelligent alarm to maintenance decision-making. Attached Figure Description

[0066] Figure 1 This diagram illustrates the implementation steps of the early warning method for low-pressure turbine guide vane rupture fault in aero-engines proposed in this invention.

[0067] Figure 2 This is a schematic diagram of the Transformer autoencoder model for early warning of low-pressure turbine guide vane rupture fault in aero-engines as described in this invention;

[0068] Figure 3 The reconstruction error value of the trained Transformer autoencoder model for each time window on a set of normal operating data of a low-pressure turbine guide vane of an aero-engine;

[0069] Figure 4 This is a comparison between the reconstructed signal and the original signal of the trained Transformer autoencoder model on a set of normal operating data of a low-pressure turbine guide vane of an aero-engine. Figure 4 (a) in the figure shows the reconstruction effect of the Vg1Totl parameter; Figure 4 (b) in the figure shows the reconstruction effect of the Vg2Totl parameter; Figure 4 (c) in the figure shows the reconstruction effect of the Vg1N1_MF1 parameter; Figure 4 (d) in the figure represents the reconstruction effect of the Vg1N1_MF2 parameter; Figure 4 (e) in the figure is the reconstruction effect diagram of the Vg1N1_MF3 parameter; Figure 4 (f) in the figure is the reconstruction effect diagram of the Vg1N2_MF1 parameter; Figure 4 In the diagram, (g) represents the reconstruction effect of the Vg1N2_MF2 parameter; Figure 4 (h) in the diagram represents the reconstruction effect of the Vg1N2_MF3 parameter; Figure 4 In the diagram, (i) represents the reconstruction effect of the Vg2N1_MF1 parameter; Figure 4 In the diagram, (j) represents the reconstruction effect of the Vg2N1_MF2 parameter; Figure 4 In the diagram, (k) represents the reconstruction effect of the Vg2N1_MF3 parameter; Figure 4 (l) in the figure represents the reconstruction effect of the Vg2N2_MF1 parameter; Figure 4 In the diagram, (m) represents the reconstruction effect of the Vg2N2_MF2 parameter; Figure 4 In the diagram, (n) represents the reconstruction effect of the Vg2N2_MF3 parameter;

[0070] Figure 5 The reconstruction error value of the trained Transformer autoencoder model for each time window on a certain set of low-pressure turbine guide vane fault data of an aero-engine;

[0071] Figure 6 This is a comparison between the reconstructed signal and the original signal of the trained Transformer autoencoder model on a certain set of low-pressure turbine guide vane fault data of an aero-engine. Figure 6 (a) in the figure shows the reconstruction effect of the Vg1Totl parameter; Figure 6 (b) in the figure shows the reconstruction effect of the Vg2Totl parameter; Figure 6 (c) in the figure shows the reconstruction effect of the Vg1N1_MF1 parameter; Figure 6 (d) in the figure represents the reconstruction effect of the Vg1N1_MF2 parameter; Figure 6 (e) in the figure is the reconstruction effect diagram of the Vg1N1_MF3 parameter; Figure 6(f) in the figure is the reconstruction effect diagram of the Vg1N2_MF1 parameter; Figure 6 In the diagram, (g) represents the reconstruction effect of the Vg1N2_MF2 parameter; Figure 6 (h) in the diagram represents the reconstruction effect of the Vg1N2_MF3 parameter; Figure 6 In the diagram, (i) represents the reconstruction effect of the Vg2N1_MF1 parameter; Figure 6 In the diagram, (j) represents the reconstruction effect of the Vg2N1_MF2 parameter; Figure 6 In the diagram, (k) represents the reconstruction effect of the Vg2N1_MF3 parameter; Figure 6 (l) in the figure represents the reconstruction effect of the Vg2N2_MF1 parameter; Figure 6 In the diagram, (m) represents the reconstruction effect of the Vg2N2_MF2 parameter; Figure 6 In the diagram, (n) represents the reconstruction effect of the Vg2N2_MF3 parameter;

[0072] Figure 7 The graph shows the evolution trend of the low-pressure turbine guide vane anomaly score as a function of the number of experiments, obtained by testing the trained Transformer autoencoder model on all 241 sets of data in the complete dataset.

[0073] Figure 8 for Figure 7 A magnified view of the central vertical axis values ​​ranging from 0 to 1, overlaid with a threshold reference line;

[0074] Figure 9 This is a schematic diagram of the early warning method and system structure for low-pressure turbine guide vane rupture faults in aero-engines proposed in this invention. Detailed Implementation

[0075] To provide a more detailed explanation of the technical solutions and advantages proposed in this invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. (Appendix) Figure 1 This embodiment demonstrates the design process of an early warning system and a visual interactive interface for low-pressure turbine guide vane rupture failure in aero-engines, including the following steps:

[0076] Step S1: Collect raw vibration data of the engine during operation using vibration sensors on the aircraft engine; specifically:

[0077] This invention is based on experimental data of a low-pressure turbine guide vane rupture failure of an aero-engine provided by Dalian University of Technology in conjunction with the Guiyang Engine Design and Research Institute of Aero Engine Corporation of China. For the failure case described in the background section, all test data of the aero-engine were collected. The test data includes total front axle vibration data collected by a piezoelectric axial vibration acceleration sensor installed near the engine's front bearing and total rear vertical vibration data collected by a piezoelectric crystal vertical vibration acceleration sensor installed at a key support structure at the rear of the engine, yielding raw vibration data. This raw vibration data is stored in time-series format, covering multiple operating conditions of the engine from start-up, idle, climb to maximum thrust, and shutdown, totaling 241 CSV files, forming the dataset for subsequent analysis.

[0078] Step S2: Preprocess the raw vibration data obtained in Step S1, including signal denoising and extraction, data standardization, and sliding time window division, to obtain the preprocessed dataset; specifically:

[0079] Step S2.1: First, a low-pass filter is used to remove high-frequency noise and electromagnetic interference from the original vibration signal, retaining the effective vibration components. Then, linear interpolation is performed on the filtered vibration data to fill in the missing data points.

[0080] Step S2.2: Then, extract the fundamental and harmonic vibration components related to the front axle vibration and rear sag vibration from the data processed in step S2.1, generate the required 14 characteristic parameters, and perform independent standardization processing on each of the 14 characteristic parameters. The meaning of the characteristic parameters is shown in Table 1.

[0081] Table 1: Parameter Meaning Table

[0082]

[0083] Step S2.3: Finally, set the time window size to 100 and the sliding step size to 1, slide the standardized temporal vibration parameter sequence into overlapping sub-sequence samples, and generate a preprocessed 14-dimensional standardized dataset that conforms to the input dimension of the Transformer autoencoder model.

[0084] Step S3: Construct a Transformer autoencoder model consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, to provide early warning of low-pressure turbine guide vane rupture faults in aero-engines; specifically:

[0085] A Transformer autoencoder model based on the PyTorch framework is constructed to learn the temporal evolution of 14-dimensional vibration parameters during normal operation of the low-pressure turbine guide vane of an aero-engine and achieve high-precision signal reconstruction. The Transformer autoencoder model consists of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer. The model architecture is shown in the attached figure. Figure 2 As shown, it is constructed in the following order:

[0086] Step S3.1: Construct the input mapping layer;

[0087] The input mapping layer is a fully connected network with 1 layer. Its input is the preprocessed 14-dimensional standardized dataset generated in step 2. It is used to map the preprocessed 14-dimensional standardized data to a 128-dimensional high-dimensional space to obtain high-dimensional variables in the 128-dimensional high-dimensional space.

[0088] Step S3.2: Construct the Transformer encoding-decoding module;

[0089] The Transformer encoder-decoder module is an encoder-decoder architecture model consisting of multiple Transformer encoder layers and multiple Transformer decoder layers. It is used to receive high-dimensional variables generated by the input mapping layer and reconstruct the sequence.

[0090] First, a position encoding parameter is introduced and superimposed on the input high-dimensional variable to preserve the temporal information of the time series;

[0091] Then, using two Transformer encoder layers, the high-dimensional variables in the high-dimensional space after position encoding are transformed into a low-dimensional representation and input into a latent space. Specifically, each Transformer encoder layer first assigns a key, value and query vector to each input, then performs a linear transformation and inputs it into a self-attention structure. Finally, the features obtained from the self-attention structure are concatenated and subjected to a linear transformation to obtain the low-dimensional representation in the latent space.

[0092] Finally, a two-layer Transformer decoder layer is used. It receives the output of the Transformer encoder layer as the input to its cross-attention structure, and receives the same input sequence as the encoder as the input to its self-attention structure. Sequence reconstruction is then performed. Specifically, each Transformer decoder layer first assigns a key, value, and query vector to each input, then performs a linear transformation and inputs it to the self-attention structure to model the internal dependencies of the input sequence. Then, the output of the Transformer encoder layer is fused through the cross-attention structure. Finally, the features obtained from the cross-attention structure are concatenated and subjected to a linear transformation to obtain the reconstructed high-dimensional output.

[0093] The number of attention heads in the attention structures of the Transformer encoder layer and the Transformer decoder layer is set to 4;

[0094] Step S3.3: Construct the output reconstruction layer;

[0095] The output reconstruction layer is a fully connected network with a network layer number of 1. It is used to receive the high-dimensional output of the Transformer encoder-decoder module, map the high-dimensional output into a 14-dimensional reconstructed output, and finally realize the end-to-end reconstruction of the initial multi-dimensional input signal.

[0096] Step S4: Based on the preprocessed dataset from Step S2, train, validate, and fine-tune the Transformer autoencoder model constructed in Step S3; specifically:

[0097] The training objective of the Transformer autoencoder model for early warning of low-pressure turbine guide vane rupture in aero-engines is to minimize the mean square reconstruction error between the 14-dimensional standardized data after step preprocessing and the 14-dimensional reconstruction output.

[0098] The specific process is as follows: First, select a set of data on the normal operation of the low-pressure turbine guide vane of an aero-engine from the preprocessed 14-dimensional standardized data obtained in step S2. Shuffle the data and divide it into a training dataset and a validation dataset in an 8:2 ratio. Select hyperparameters for the Transformer autoencoder model based on the size of the training dataset. Set the training batch size to 128 and the initial learning rate to 0.0001. Second, optimize the parameters of the Transformer autoencoder model using the Adam optimizer based on the training dataset. Introduce a learning rate scheduler; halve the learning rate when the validation loss does not decrease for 5 consecutive rounds. To prevent overfitting, set an early stopping mechanism; terminate training when the validation loss does not decrease for 10 consecutive rounds. Third, validate the Transformer autoencoder model using the validation dataset. Fine-tune the model hyperparameters based on the model's performance on the validation dataset and retrain, setting a target error value of 0.005. Finally, repeat the above training, validation, and fine-tuning process until the reconstruction error of the Transformer autoencoder model on the validation dataset is lower than the set target error value. The Transformer autoencoder training is complete, and the optimal model parameters are saved.

[0099] Step S5: Based on the Transformer autoencoder model trained in Step S4 for early warning of low-pressure turbine guide vane rupture in aero-engines, define the early warning threshold and issue an early warning; specifically:

[0100] The trained Transformer autoencoder model is used to define the warning threshold, and then used for early warning of low-pressure turbine guide vane rupture faults in aero-engines, specifically including:

[0101] Step S5.1: First, select 104 sets of preprocessed data of the low-pressure turbine guide vane of the aero-engine that have been identified by professionals during normal operation, load the optimal model parameters and input the data into the Transformer autoencoder model to obtain the corresponding reconstruction output;

[0102] Step S5.2: Next, calculate the reconstruction error value for each time window, and use the maximum reconstruction error value across all time windows within a single aero-engine operation as the low-pressure turbine guide vane anomaly score for this operation, thus obtaining a set of low-pressure turbine guide vane anomaly scores corresponding to these 104 sets of aero-engine low-pressure turbine guide vane normal operation data; Appendix Figure 3The figure shows the reconstruction error values ​​of the trained Transformer autoencoder model for each time window on a set of normal operation data of aero-engine low-pressure turbine guide vanes. The figure indicates that the model has good reconstruction capability for the normal operation data, with a peak reconstruction error (i.e., the low-pressure turbine guide vane anomaly score) of 0.0207 across all time windows. Figure 4 The figure shows the comparison between the reconstructed signal and the original signal of the trained Transformer autoencoder model on a set of normal operation data of a low-pressure turbine guide vane of an aero-engine. The figure shows that the reconstructed signal of the model basically overlaps with the original signal, and the reconstruction effect is good.

[0103] Step S5.3: Then, a warning threshold is calculated based on the obtained set. The maximum score of the abnormality of the low-pressure turbine guide vane in this set of data is 0.0846. The threshold coefficient is selected as 1.05. Therefore, the warning threshold is set to 0.0888 according to formula (10).

[0104] Step S5.4: Finally, load the optimal model parameters and input all 241 sets of data in this dataset into the trained Transformer autoencoder model to obtain the low-pressure turbine guide vane anomaly score for the 241 sets of data; Appendix Figure 5 The figure shows the reconstruction error values ​​of the trained Transformer autoencoder model for each time window on a set of low-pressure turbine guide vane fault data of an aero-engine. The figure indicates that the model has poor reconstruction ability for fault operation data, with a peak reconstruction error (i.e., the low-pressure turbine guide vane anomaly score) of 2.3056 across all time windows. Figure 6 The figure shows a comparison between the reconstructed signal and the original signal of the trained Transformer autoencoder model on a set of low-pressure turbine guide vane fault data of an aero-engine. The figure indicates that the reconstructed signal deviates significantly from the original signal, resulting in poor reconstruction performance. Figure 7 The graph shows the evolution trend of the low-pressure turbine guide vane anomaly score as a function of the number of experiments, obtained by testing the trained Transformer autoencoder model on all 241 sets of data in the complete dataset. The curves in the graph indicate that the model described in this method clearly has the ability to distinguish between normal operation data and fault data of the low-pressure turbine guide vane of an aero-engine. In the early experiments of normal operation of the aero-engine low-pressure turbine guide vane, the anomaly score is very low; in the later experiments of low-pressure turbine guide vane faults, the anomaly score is very high. (Attached) Figure 8 For the appendix Figure 7The figure shows a magnified view of the vertical axis values ​​in the range of 0 to 1, with an overlay of a warning threshold reference line of 0.0888. The curves in the figure indicate that the set warning threshold can separate the early normal operation data and the later failure data of the low-pressure turbine guide vane of the aero-engine. In the 222nd experiment, a high abnormality score of the low-pressure turbine guide vane was measured for the first time, with a value of 0.8732, which greatly exceeded the set threshold. The date corresponding to this experiment was 13 days earlier than the date when the failure actually occurred. It can be considered that the method proposed in this invention can capture the early failure 13 days in advance.

[0105] To verify the reliability and early warning capability of the low-pressure turbine guide vane rupture fault early warning system for aero-engines described in this invention, a comparative analysis of 241 sets of data was conducted using a widely accepted mainstream discrimination method in the industry. Experimental results show that this mainstream method first identified the abnormal signal in the 230th experiment, with the corresponding experiment date issuing the warning 7 days later than the method proposed in this invention. This result demonstrates that the method proposed in this invention significantly outperforms existing mainstream discrimination methods in terms of fault detection sensitivity, response timeliness, and early anomaly identification capability. It possesses higher engineering application value and reliability, and can effectively support predictive maintenance and flight safety assurance for key components of aero-engines.

[0106] This embodiment also provides an early warning system for low-pressure turbine guide vane rupture faults in aero-engines. This system can implement the aforementioned early warning method for low-pressure turbine guide vane rupture faults in aero-engines in modules, including a data acquisition and preprocessing module, a Transformer autoencoder model construction and training module, a reconstruction detection and anomaly scoring module, a threshold management and early warning decision module, and a visualization and interaction module; specifically:

[0107] Data acquisition and preprocessing module: such as Figure 9 The system acquires and preprocesses raw vibration data from the engine's operating environment through a data acquisition and preprocessing module, encompassing two functional units: data acquisition and preprocessing. Based on this specific implementation case, the data acquisition unit uses a piezoelectric axial vibration acceleration sensor installed near the engine's front bearing and a piezoelectric crystal vertical vibration acceleration sensor installed at the engine's rear critical support structure to collect the total vibration of the front and rear axles, saving it as time-series raw vibration data. The preprocessing unit filters, linearly interpolates, standardizes, and performs sliding slicing on the raw vibration data to generate a preprocessed standardized dataset.

[0108] Transformer autoencoder model construction and training module: such as Figure 9The Transformer autoencoder model building and training module is connected to the data acquisition and preprocessing module, encompassing three functional units: model building, training optimization, and model storage. Based on this specific implementation case, the model building unit constructs a Transformer autoencoder model consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, according to the model architecture described in step S3. The training optimization unit uses the training and validation sets formed after preprocessing normally operating vibration data to train, validate, and fine-tune the model with the goal of minimizing the reconstruction error, until the reconstruction error is below 0.005. The model storage unit is used to save the optimal model parameters after training for subsequent modules to access.

[0109] Reconstruct the detection and anomaly severity scoring module: such as Figure 9 The reconstruction detection and anomaly scoring module is integrated with the data acquisition and preprocessing module and the Transformer autoencoder model building and training module, encompassing three functional units: model loading, reconstruction detection, and scoring calculation. Based on this specific implementation case, the model loading unit is responsible for loading the optimal model parameters stored in the Transformer autoencoder model building and training module; the reconstruction detection unit inputs the preprocessed vibration data to be detected into the model to obtain the corresponding reconstruction output; and the scoring calculation unit calculates the reconstruction error for each time window, using the maximum reconstruction error value across all time windows within a single aero-engine operation as the "low-pressure turbine guide vane anomaly score" for this mission, and outputs this score.

[0110] Threshold management and early warning decision module: such as Figure 9 The threshold management and early warning decision module is connected to the reconstruction detection and anomaly score module, which includes two functional units: threshold calculation and early warning decision. Based on this specific implementation case, the threshold calculation unit calculates the early warning threshold of 0.0888 according to formula (10) based on the set of anomaly scores generated by 104 sets of preprocessed data of the low-pressure turbine guide vane of the aero-engine during normal operation, which have been certified by professionals. The early warning decision unit compares the anomaly score generated by the reconstruction detection and anomaly score module with the early warning threshold. If the anomaly score is greater than the early warning threshold, it is determined to be an operational anomaly. If the anomaly score is not greater than the early warning threshold, it is determined to be a normal operation. Finally, it outputs a normal operation or anomaly signal.

[0111] Visual interactive modules: such as Figure 9This interactive visualization module integrates the data acquisition and preprocessing module, the reconstruction detection and anomaly scoring module, and the threshold management and early warning judgment module, encompassing two functional units: interactive testing and visualization of test results. Based on this specific implementation case, the interactive testing unit uses a graphical user interface built on the PyQt5 framework, allowing users to manually select and import locally stored vibration data files and trigger a complete testing process for the selected file with a single click. The test result visualization unit displays the test results, including: a comparison curve between the original vibration signal and the model reconstruction signal, the reconstruction error value for each time window and its comparison with the early warning threshold, and the final risk assessment conclusion. Staff can use the interactive visualization interface to test the selected data and generate test results, and assess whether maintenance of the low-pressure turbine guide vane is necessary based on the results.

[0112] The above embodiments are only used to illustrate specific implementations of the present invention and should not be construed as limiting the scope of protection of the present invention. It should be noted that those skilled in the art can make various modifications, equivalent substitutions, or appropriate variations to the embodiments within the scope of the technical solution of the present invention without departing from the basic principles and concept of the present invention, and all such modifications and variations should fall within the scope of protection of the present invention.

Claims

1. A method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine, characterized in that, The aforementioned early warning method includes the following steps: Step S1: Collect raw vibration data of the engine during operation using vibration sensors on the aircraft engine; Step S2: Preprocess the raw vibration data obtained in step S1, including signal denoising and extraction, data standardization and sliding time window division, to obtain the preprocessed dataset; Step S3: Construct a Transformer autoencoder model consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, to provide early warning of low-pressure turbine guide vane rupture faults in aero-engines; The Transformer autoencoder model, consisting of an input mapping layer, a Transformer encoder-decoder module, and an output reconstruction layer, is used to learn the normal operation of the low-pressure turbine guide vane of an aero-engine. The temporal evolution law of vibration parameters is studied, and high-precision signal reconstruction is achieved. Step S4: Based on the dataset preprocessed in Step S2, train, validate, and fine-tune the Transformer autoencoder model constructed in Step S3; Step S5: Define the warning threshold using the trained Transformer autoencoder model, and provide early warning of low-pressure turbine guide vane rupture fault in aero-engines based on the Transformer autoencoder model.

2. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 1, characterized in that, The specific steps of S1 are as follows: During the operation of an aero-engine, a piezoelectric axial vibration acceleration sensor installed near the front bearing of the engine is used to collect the total front shaft vibration, and a piezoelectric crystal vertical vibration acceleration sensor installed at the key support structure at the rear of the engine is used to collect the total rear vertical vibration, so as to ensure that the dynamic response of the aero-engine vibration can be captured and the raw vibration data can be obtained. The raw vibration data is stored in time series form, covering multiple operating stages of the engine from start-up, idle, climbing to maximum thrust and shutdown, forming the dataset for subsequent analysis.

3. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 2, characterized in that, The specific steps of S2 are as follows: Step S2.1: Use filtering methods to remove high-frequency noise and electromagnetic interference from the original vibration data, retain the effective vibration components, and perform linear interpolation on the filtered vibration data to fill in the missing data points. The linear interpolation operation is shown in the following formula; （1）； In equation (1), Indicates the sampling time of the missing data points; Indicates the numerical value to be filled; Indicates the sampling time of the point preceding the missing value; This indicates the value of the point preceding the missing value. Indicates the sampling time of the point following the missing value; This indicates the value of the point following the missing value. Step S2.2: Extract the fundamental and harmonic vibration components related to the front axle vibration and rear sag vibration from the data processed in step S2.1, and generate the required... Each of the obtained feature parameters is standardized as shown in the following formula to obtain... A continuous time series; （2）； In equation (2), the subscript Indicates the index of the value in the feature parameter; This indicates that the first characteristic parameter One value; This represents the mean of the feature parameters in this dimension; This represents the standard deviation of the feature parameter in this dimension; This represents a standardized continuous time series; Step S2.3: Select the time window length and sliding step size for the standardized continuous time series to perform sliding slices, generating preprocessed slices that conform to the input dimensions of the Transformer autoencoder model. Dimensional standardized dataset.

4. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 3, characterized in that, In step S3, the construction process of the Transformer autoencoder model is as follows: Step S3.1: Construct the input mapping layer; The input mapping layer is a fully connected network, and its input is the preprocessed data generated in step 2. 3D normalized dataset, used to preprocess the data Dimensional standardized data mapping to In a higher-dimensional space, we obtain High-dimensional variables in a high-dimensional space; Step S3.2: Construct the Transformer encoding-decoding module; The Transformer encoder-decoder module is an encoder-decoder architecture model consisting of multiple Transformer encoder layers and multiple Transformer decoder layers. It is used to receive high-dimensional variables generated by the input mapping layer and reconstruct the sequence. Step S3.2.1: First, introduce position encoding parameters and superimpose them on the input high-dimensional variables. The position encoding is shown in the following formula. （4）； （5）； In equations (4) and (5), Indicates the position index in the time series; Indicates a dimension index; Indicates the total number of dimensions for positional encoding; Represents the position encoding matrix; This indicates that the position index in the position encoding matrix is And the dimension index is The corresponding scalar value; This indicates that the position index in the position encoding matrix is And the dimension index is The corresponding scalar value; Step S3.2.2: Then, using a multi-layer Transformer encoder layer, the high-dimensional variables in the high-dimensional space after stacking position encoding are transformed into a low-dimensional representation and input into a latent space. Each Transformer encoder layer first assigns a key, value and query vector to each input, then performs a linear transformation and inputs it into a self-attention structure. Finally, the features obtained from the self-attention structure are concatenated and a linear transformation is performed to obtain the low-dimensional representation in the latent space. Step S3.2.3: Finally, a multi-layer Transformer decoder layer is used. It receives the output of the Transformer encoder layer as the input to its cross-attention structure, and simultaneously receives the same input sequence as the encoder as the input to its self-attention structure, and performs sequence reconstruction. Each Transformer decoder layer first assigns a key, value, and query vector to each input, then performs a linear transformation and inputs it to the self-attention structure to model the internal dependencies of the input sequence. Then, it fuses the output of the Transformer encoder layer through the cross-attention structure. Finally, it concatenates the features obtained from the cross-attention structure and performs a linear transformation to obtain the reconstructed high-dimensional output. Step S3.3: Construct the output reconstruction layer; The output reconstruction layer is a fully connected network used to receive the high-dimensional output of the Transformer encoder-decoder module and map the high-dimensional output to... The reconstructed output of the multidimensional signal ultimately achieves end-to-end reconstruction of the initial multidimensional input signal.

5. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 4, characterized in that, Step S3: In step S3.1, the core function of the input mapping layer is to perform the function mapping shown in the following formula; （3）； In equation (3), Indicates input Dimensional standardized data; Represents the learnable weight matrix; Represents the bias vector; Indicates mapping to High-dimensional variables in a high-dimensional space; In step S3.2, the core function of the Transformer encoder layer and the Transformer decoder layer is a multi-head attention mechanism, which performs the attention calculation as shown in the following formula; （6）； （7）； In equations (6) and (7), the subscripts Represents the attention head index; Indicates the first The output of each attention head; Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the first A learnable query projection matrix for each attention head; Indicates the first Learnable key projection matrix for each attention head; Indicates the first The learnable projection matrix of each attention head; This represents the query dimension for each header; For the number of attention heads; To output the projection matrix; This indicates a vector concatenation operation; In step S3.3, the core function of the output reconstruction layer is described as executing the function mapping shown in the following formula; （8）； In equation (8), This represents the high-dimensional representation of the decoder output; Represents the weight matrix of the reconstruction layer; Indicates the refactoring of bias terms; This represents the reconstructed output with the same dimension as the original input.

6. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 4, characterized in that, The specific steps of S4 are as follows: The training objective of the Transformer autoencoder model for early warning of low-pressure turbine guide vane rupture in aero-engines is to minimize the preprocessed step. Dimensional Standardized Data and The mean square reconstruction error between the dimensional reconstruction outputs is expressed in the following form; （9）； In equation (9), This indicates the batch size, which is the number of samples input into the neural network in a single run. This represents the time step for each sample; Indicates the index of the time step; Indicates the index of the sample; Indicates the first The sample at the th The actual input value at each time step; This represents the reconstructed value corresponding to the model; This represents the overall mean square error loss; First, from the preprocessed 14-dimensional standardized data obtained in step S2, data on the normal operation of the low-pressure turbine guide vane of the aero-engine is selected to divide the dataset into training and validation datasets. Hyperparameters, including batch size and learning rate, are selected for training the Transformer autoencoder model based on the size of the training dataset. Second, the parameters of the Transformer autoencoder model are optimized using an optimizer based on the training dataset. Third, the Transformer autoencoder model is validated using the validation dataset. The hyperparameters are fine-tuned based on the model's performance on the validation dataset, and the model is retrained with a target error value set. Finally, the above training, validation, and fine-tuning process is repeated until the reconstruction error of the Transformer autoencoder model on the validation dataset is lower than the set target error value. The Transformer autoencoder training is then complete, and the optimal model parameters are saved.

7. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 6, characterized in that, The target error value ranges from 0.005 to 0.

01.

8. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 6, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Filter the preprocessed data of the low-pressure turbine guide vane of the aero-engine during normal operation, load the optimal model parameters and input the data into the Transformer autoencoder model to obtain the corresponding reconstruction output; Step S5.2: Calculate the reconstruction error value for each time window, and use the maximum reconstruction error value in all time windows within a single operation of the aero-engine as the low-pressure turbine guide vane anomaly score for this operation, thereby obtaining a set of low-pressure turbine guide vane anomaly scores corresponding to the normal operation data of the aero-engine low-pressure turbine guide vane. Step S5.3: Calculate a warning threshold based on the obtained set. The warning threshold is defined as shown in the following formula; （10）； In equation (10), This represents the maximum value of the low-pressure turbine guide vane anomaly score in the normal operation data of the aero-engine low-pressure turbine guide vane; This is the threshold coefficient; This indicates the set warning threshold; Step S5.4: Load the optimal model parameters, input the target data to be detected into the Transformer autoencoder model, and obtain the low-pressure turbine guide vane anomaly score of the target data. If rating Greater than the set warning threshold If the target data corresponds to an experimental low-pressure turbine guide vane of an aero-engine, it is determined that there is an abnormal risk; if the score is... Not greater than the set warning threshold If the target data is obtained, it is determined that the low-pressure turbine guide vane of the aero-engine in the experiment is operating normally.

9. The method and system for early warning of low-pressure turbine guide vane rupture fault in an aero-engine according to claim 8, characterized in that, In step S5.3, The threshold coefficient ranges from 1 to 1.05 and is used to compensate for model uncertainties.

10. A fault detection and early warning system for low-pressure turbine guide vane rupture in an aero-engine, characterized in that, The method for early warning of low-pressure turbine guide vane rupture faults in aero-engines, as described in any one of claims 1-9, is implemented through an early warning system for low-pressure turbine guide vane rupture faults in aero-engines. This includes a data acquisition and preprocessing module, a Transformer autoencoder model construction and training module, a reconstruction detection and anomaly severity scoring module, a threshold management and early warning decision module, and a visualization and interaction module. Specifically: The data acquisition and preprocessing module is used to acquire raw vibration data of the aero-engine from the operating environment and perform preprocessing. The Transformer autoencoder model building and training module is connected to the data acquisition and preprocessing module, and covers three functional units: model building, training optimization and model storage. The reconstruction detection and anomaly scoring module is connected to the data acquisition and preprocessing module and the Transformer autoencoder model construction and training module, and covers three functional units: model loading, reconstruction detection and scoring calculation. The threshold management and early warning decision module is connected to the reconstruction detection and anomaly score module, and covers two functional units: threshold calculation and early warning decision. The visualization interaction module includes a data acquisition and preprocessing module, a reconstruction detection and anomaly scoring module, and a threshold management and early warning decision module, covering two functional units: interactive testing and visualization of detection results.

Images