Aero-engine fault diagnosis method and device based on tensor decomposition and medium

Abstract

The invention relates to an aero-engine fault diagnosis method and device based on tensor decomposition and a medium, and the method comprises the steps: collecting the gas path parameter time sequence data of a plurality of parts of an aero-engine, and carrying out the time alignment and working condition label labeling; for the multi-parameter data of each component, a kernel principal component analysis and automatic encoder fused feature extraction method is adopted, and parameters of different components are unified into fixed dimension features; the feature matrixes obtained after feature extraction of all the components are combined, and a time-component-feature third-order tensor is constructed; training a high-order singular value decomposition model based on the engine health monitoring data; fault diagnosis is carried out by calculating the reconstruction error of the test data and the monitoring data; and fault positioning is realized by combining core tensor difference analysis. According to the method, the problem of feature fusion of engine multi-source heterogeneous data under variable working conditions is solved, and the engine fault detection sensitivity and positioning precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aero-engine health management, in particular to an aero-engine fault diagnosis method, device and medium based on tensor decomposition. BACKGROUND

[0002] The engine is the core power device of the aircraft, and its reliability directly affects the flight safety of the aircraft. Many internal parts of the engine often operate in extremely harsh environments such as high speed, high temperature and strong vibration. In particular, when the performance of the key components of the engine degrades to a certain extent, various faults are prone to occur. By timely fault diagnosis and positioning of the engine, the safety of the engine can be improved.

[0003] The current engine fault diagnosis method has the following problems: 1) difficulty in multi-source heterogeneous data fusion: the types and quantities of monitoring parameters of different components of the aero-engine (compressor, combustion chamber, gas turbine, power turbine, etc.) are different, forming heterogeneous data sources, which are difficult to process uniformly; 2) insufficient dynamic adaptability under variable conditions: the aero-engine may experience slow vehicle, cruise, maximum power and other conditions in a task, and the working conditions are frequently switched. The behavior patterns of monitoring parameters under different working conditions are significantly different, and the change law is complex. The existing method usually assumes a steady state condition, and it is difficult to handle dynamic condition switching; 3) low detection rate of weak faults: early fault signals of the engine are weak and are easily masked by condition changes and sensor noise. The capture rate and monitoring accuracy of the existing monitoring and diagnosis system for these early fault signs are low.

[0004] Some existing engine fault diagnosis methods also have certain limitations. Among them, the method based on physical model establishes a physical model such as compressor characteristic diagram and combustion chamber thermodynamic model to diagnose faults and evaluate performance of the engine. Although this method has theoretical rigor, it is complex to model and difficult to adapt to actual degradation; the data-driven method is some common deep learning method. Although it avoids physical modeling, it requires a large amount of labeled data, which is difficult to obtain in actual scenarios, and the black box characteristics result in poor interpretability of the diagnosis results.

[0005] Therefore, there is an urgent need for an aero-engine diagnosis method that can process dimensional differences, adapt to condition changes and accurately locate faults under a unified framework to meet the health management needs of modern aero-engines with high reliability and high safety. SUMMARY

[0006] The technical problem to be solved by the present application is to provide an aero-engine fault diagnosis method, device and medium based on tensor decomposition, which can meet the health management needs of modern aero-engines with high reliability and high safety.

[0007] The technical solution adopted by this invention to solve its technical problem is: to provide a method for diagnosing aero-engine faults based on tensor decomposition, comprising the following steps:

[0008] Collect time-series data of gas path parameters of multiple key components of the aero-engine under test during a complete operating cycle, and label the operating condition type at each time point according to the operating condition classification standard of the aero-engine under test.

[0009] The time-series data of the collected gas path parameters are aligned by timestamps, and the operating condition type of each timestamp is labeled according to the operating condition classification standard of the aero-engine under test. The time-series data of the gas path parameters are adaptively normalized according to the operating condition.

[0010] Kernel principal component analysis was performed to reduce the dimensionality of the time-series data of the normalized gas path parameters, and feature extraction was performed by stacked noise reduction autoencoder to obtain feature vectors of a preset dimension.

[0011] A time-feature-component test tensor X is constructed based on the number of time points, the number of key components of the aero-engine, and feature vectors of a preset dimension. test ;

[0012] Test tensor X test Project onto the health subspace and compute the test tensor X. test Global reconstruction error RE between the health baseline tensor X and the health baseline tensor X global Component-level reconstruction error RE c The difference between the core tensor and ΔG;

[0013] When the reconstruction error RE global If the set threshold is exceeded, the aircraft engine under test is determined to have malfunctioned;

[0014] Based on component-level reconstruction error RE c The faulty component of the aero-engine under test is located by the difference between the core tensor ΔG and the fault type is identified by analyzing the energy distribution of the feature dimension.

[0015] The step involves performing kernel principal component analysis to reduce the dimensionality of the time-series data of the normalized gas path parameters, and then extracting features using a stacked noise-reducing autoencoder to obtain a feature vector of a preset dimension. Specifically, this includes:

[0016] Combine all the gas path parameters of each key component into matrix X;

[0017] Kernel principal component analysis was used to extract m principal components from the matrix X to obtain the feature F. kpca In the kernel principal component analysis, the Gaussian kernel function is used to handle nonlinear relationships;

[0018] The extracted features F kpcaA stacked denoising autoencoder is used for processing. Robust feature representations are learned through an encoder-decoder structure. Noise injection and L1 regularization are employed to obtain the bottleneck layer features F of the encoder. sdae And dynamically determine the feature dimension d of each key component based on information entropy. c ;

[0019] The feature dimensions of each key component are globally unified to obtain feature vectors with preset dimensions.

[0020] The health baseline tensor X is obtained in the following way:

[0021] Collect health time-series data of gas path parameters of multiple key components of the aero-engine during a complete operating cycle when the engine is in a healthy state, and label the operating condition type at each time point according to the operating condition classification standard of the aero-engine.

[0022] The collected health time series data of the gas path parameters are aligned by timestamp, and the operating condition type of each timestamp is labeled according to the operating condition classification standard of the aero-engine. The health time series data of the gas path parameters are adaptively normalized according to the operating condition.

[0023] Kernel principal component analysis was performed to reduce the dimensionality of the normalized gas path parameters health time series data, and feature extraction was performed by stacked noise reduction autoencoder to obtain a health feature vector of preset dimension.

[0024] A time-feature-component health benchmark tensor is constructed based on the number of time points, the number of key components of the aero-engine, and health feature vectors of preset dimensions.

[0025] The test tensor Project onto the health subspace and compute the test tensor. With health baseline tensor Global reconstruction error RE between global Component-level reconstruction error RE c Differences between core tensors Specifically, it includes:

[0026] Based on health baseline tensor A baseline model is established using higher-order singular value decomposition to obtain the core tensor. And factor matrix U time U comp U feat ;

[0027] Calculate the test tensor X test Projecting onto the healthy subspace yields the test core tensor. The calculation method is as follows:

[0028] Wherein, x1, x2 and x3 represent tensors respectively The product of the three factor matrices in the corresponding modal;

[0029] Based on the test core tensor Reconstruct the test tensor, and the calculation method is: Wherein, The reconstructed test tensor;

[0030] According to the core tensor Test tensor And the reconstructed test tensor Calculate the global reconstruction error RE global , component-level reconstruction error RE c And core tensor difference

[0031] The health benchmark tensor is based on The benchmark model is established by high-order singular value decomposition, and the core tensor is obtained And factor matrix U time , U comp , U feat , specifically including:

[0032] Obtain the health benchmark tensor The expansion matrix of each mode, and singular value decomposition is performed on each expansion matrix to obtain the singular value vector of each expansion matrix;

[0033] According to the singular value vector of each expansion matrix, the cumulative variance contribution rate of singular value is calculated, and the singular value serial number corresponding to the variance contribution rate not less than the preset value is taken as the rank R i of each expansion matrix;

[0034] Select the first R i singular vectors of each expansion matrix to form factor matrix U time , U comp , U feat ;

[0035] Based on the factor matrix U time , U comp , U feat And the health benchmark tensor Calculate the core tensor The calculation method is:

[0036] The calculation method of the global reconstruction error RE global Is: The calculation method of the component-level reconstruction error RE c Is: The calculation method of the core tensor difference Is: where || || denotes the Frobenius norm, F is the Frobenius norm, is the test tensor is the tensor slice of the cth key component, is the reconstructed test tensor is the tensor slice of the cth key component.

[0037] said when the reconstruction error RE global exceeds a set threshold, the threshold is represented as: τ g = μ RE + 3σ RE , wherein τ g is the threshold, μ RE and σ RE are the mean and variance of the reconstruction error set of the healthy baseline tensor .

[0038] said according to the component-level reconstruction error RE c and the core tensor difference locating the faulty component of the to-be-tested aero-engine, and identifying the fault type by analyzing the feature dimension energy distribution, specifically comprising:

[0039] According to the component-level reconstruction error RE c and the core tensor difference , the comprehensive anomaly score Score c of each key component is calculated, if the comprehensive anomaly score Score c of a certain key component is greater than the confidence threshold τ conf , the key component is determined as a faulty component, wherein the calculation method of the comprehensive anomaly score Score c is: α, β, γ are weight coefficients, is the Frobenius norm of the core tensor difference on the key component c, Consistency c is the anomaly temporal consistency of the component c;

[0040] Based on the core tensor difference , the anomaly energy distribution E f on the feature dimension is calculated, the top 10% features of the energy distribution are selected as the fault feature set F key , then the contribution degree ParaScore p of each original parameter is calculated, the parameter contribution degree ParaScore p is matched with the known fault feature library to determine the fault type, wherein the calculation method of the anomaly energy distribution E f is: r t denotes the mode index in the time dimension, R t denotes the rank of the unfolding matrix of the tensor in the time dimension, r c denotes the mode index in the component dimension, R c denotes the rank of the unfolding matrix of the tensor in the component dimension, f denotes the mode index in the feature dimension, i.e., the column index of the feature factor matrix U feat The calculation method of the parameter contribution degree ParaScore p is as follows: U feat [p,f] denotes the contribution degree of the original gas path parameter p to the fth feature mode.

[0041] The technical scheme adopted by the present application to solve its technical problems is to provide an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned aero-engine fault diagnosis method based on tensor decomposition when executing the computer program.

[0042] The technical scheme adopted by the present application to solve its technical problems is to provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the above-mentioned aero-engine fault diagnosis method based on tensor decomposition when executed by a processor.

[0043] Advantages

[0044] Compared with the prior art, the present application has the following advantages and positive effects: according to the characteristics of aero-engine operation data, the present application constructs a time-component-feature three-dimensional tensor model, and combines high-order singular value decomposition and working condition perception mechanism, effectively overcoming the technical bottlenecks of traditional methods in multi-source heterogeneous data fusion, variable working condition adaptability and weak fault recognition. The present application can automatically analyze the coupling relationship between the core tensor difference and the feature factor matrix, realize fault component positioning and fault type recognition, and significantly improve the decoupling ability and state recognition depth of the multi-dimensional fault mode of the aero-engine. This method not only preserves the physical interpretability, but also breaks through the limitations of dimension difference and working condition dynamic switching on the diagnosis accuracy, and also provides a method for fault diagnosis of other variable working condition equipment. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is a flowchart of the aero-engine fault diagnosis method based on tensor decomposition of the first embodiment of the present application;

[0046] Figure 2 is a compressor parameter normalization curve in the embodiment of the present application;

[0047] Figure 3 is a reconfiguration error analysis diagram in an embodiment of the present application;

[0048] Figure 4 is a fault location and characteristic energy analysis diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described below with reference to specific embodiments. It should be understood that these embodiments are intended to illustrate the present application and are not intended to limit the scope of the present application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the present application after reading the content taught in the present application, and these equivalent forms also fall within the scope of the appended claims of the present application.

[0050] A first embodiment of the present application relates to a method for diagnosing faults of an aero-engine based on tensor decomposition, as shown in the following steps: Figure 1

[0051] Step 1, collecting time series data of air path parameters (such as temperature, pressure, flow rate, rotation speed, vibration, etc.) of multiple key components (such as a compressor, a combustion chamber, a gas turbine, and a power turbine) of a to-be-tested aero-engine in a complete running cycle, and labeling a working condition type (such as slow taxi, cruising, maximum continuous, takeoff, etc.) of each time point according to a working condition division standard of the to-be-tested aero-engine.

[0052] Step 2, aligning the collected time series data of the air path parameters according to a time stamp, labeling a working condition type of each time stamp according to the working condition division standard of the to-be-tested aero-engine, and performing adaptive normalization on the time series data of the air path parameters according to the working condition.

[0053] Step 3, performing kernel principal component analysis dimension reduction on the normalized time series data of the air path parameters, and performing feature extraction through a stacked denoising autoencoder to obtain a feature vector of a preset dimension. This step specifically includes:

[0054] Firstly, all air path parameters of each key component are combined into a matrix X ∈ R T×p , where T is the number of time points, and p is the number of parameters of the key component;

[0055] Next, m principal components are extracted from the matrix X using kernel principal component analysis to obtain a feature F kpca ∈ R T×m , wherein a Gaussian kernel function is used to handle nonlinear relationships during kernel principal component analysis;

[0056] Then, the extracted feature F kpca ∈ R T×m ​The stacked denoising autoencoder is used for processing, a robust feature representation is learned through an encoder-decoder structure, noise injection and L1 regularization are used to obtain the bottleneck layer features of the encoder And dynamically determine the feature dimension d of each key component according to the information entropy c ;

[0057] Finally, the feature dimensions of all key components are globally unified to obtain the feature vector d of the preset dimension global , wherein

[0058] Step 4, constructing a time-feature-component test tensor X according to the number of time points, the number of key components of the aero-engine, and the feature vector of the preset dimension test .

[0059] Step 5, projecting the test tensor X test into the health subspace, calculating the global reconstruction error RE test , the component-level reconstruction error RE global , and the core tensor difference AG between the test tensor X c and the health benchmark tensor X.

[0060] The health benchmark tensor X in this step is obtained as follows:

[0061] Collect the health time series data of the gas path parameters of multiple key components during the complete running period of the aero-engine health state, and label the working condition type of each time point according to the working condition division standard of the aero-engine;

[0062] Align the collected health time series data of the gas path parameters according to the time stamp, label the working condition type of each time stamp according to the working condition division standard of the aero-engine, and adaptively normalize the health time series data of the gas path parameters according to the working condition;

[0063] Perform kernel principal component analysis dimension reduction on the normalized health time series data of the gas path parameters, and perform feature extraction through a stacked denoising autoencoder to obtain a health feature vector of a preset dimension;

[0064] Constructing a time-feature-component health benchmark tensor X according to the number of time points, the number of key components of the aero-engine, and the health feature vector of the preset dimension , wherein T is the number of time points, corresponding to the running length of the aero-engine, C is the number of key components in the aero-engine, and d is the preset dimension after feature extraction.

[0065] The global reconstruction error RE global , the component-level reconstruction error RE c , and the core tensor difference AG in this step are obtained as follows: ​

[0066] First, based on the health baseline tensor A baseline model is established using higher-order singular value decomposition to obtain the core tensor. And factor matrix U time U comp U feat Specifically:

[0067] Obtain the health baseline tensor The expansion matrix X of each mode unfold,i And perform singular value decomposition on each expanded matrix, i.e. Obtain the singular value vectors of each expanded matrix. R i Let n be the rank of the expanded matrix, and n be the smaller of the number of rows and columns of the expanded matrix.

[0068] Calculate the cumulative variance contribution rate of the singular values ​​based on the singular value vectors of each expanded matrix, and take the singular value index corresponding to a variance contribution rate of not less than 95% as the rank R of each expanded matrix. i ;

[0069] Select the first R of each expanded matrix i The singular vectors constitute the factor matrix U. time U comp U feat ;

[0070] Based on factor matrix U time U comp U feat and health baseline tensor Compute the core tensor The calculation method is as follows: Where ×1, ×2, and ×3 represent tensors respectively. The product of the three factor matrices in the corresponding modes.

[0071] Next, the test tensor is calculated. Projecting onto the healthy subspace yields the test core tensor. The calculation method is as follows:

[0072] Then, based on the test core tensor The reconstructed test tensor is calculated as follows: in, For the reconstructed test tensor.

[0073] Finally, based on the core tensor Test Tensor and the reconstructed test tensor Calculate the global reconstruction error RE global, component-level reconstruction error RE c and core tensor difference

[0074] where the global reconstruction error RE global is calculated as:

[0075] the component-level reconstruction error RE c is calculated as:

[0076] the core tensor difference is calculated as:

[0077] where || || F is the Frobenius norm, is the test tensor the tensor slice of the cth key component, is the reconstructed test tensor the tensor slice of the cth key component.

[0078] Step 6, when the reconstruction error RE global exceeds a set threshold, it is determined that the aero-engine under test has a fault. The threshold in this step is represented as: τ g = μ RE + 3σ RE , where τ g is the threshold, μ RE and σ RE are the mean and variance of the reconstruction error set of the healthy baseline tensor .

[0079] Step 7, according to the component-level reconstruction error RE c and the core tensor difference , the faulty component of the aero-engine under test is located, and the fault type is identified by analyzing the characteristic dimension energy distribution. This step specifically includes:

[0080] (1) According to the component-level reconstruction error RE c and the core tensor difference , the comprehensive anomaly score Score c of each key component is calculated, if the comprehensive anomaly score Score c of a certain key component is greater than the confidence threshold τ conf , then the key component is determined as a faulty component.

[0081] where the comprehensive anomaly score Score c is calculated as:

[0082]

[0083] In the formula, α, β, γ are weighting coefficients. It is the Frobenius norm of the core tensor difference on the key component c, Consistency c The consistency of abnormal time for component c is calculated as Consistency. c = Number of points continuously exceeding limits / T, i.e., continuously exceeding the failure threshold τ of each critical component. c (Proportion of time points at the historical 95th percentile of each key component);

[0084] (2) Based on core tensor differences Calculate the anomalous energy distribution E along the feature dimension f The top 10% of the energy distribution features are selected as the fault feature set F. key Subsequently, through the eigenfactor matrix U feat Perform physical parameter backtracking to map feature space anomalies to the original physical parameters p, i.e., calculate the contribution score ParaScore of each original parameter. p ParaScore of parameter contribution p Match the fault type with a known fault feature library to determine the fault type.

[0085] Among them, the abnormal energy distribution E f The calculation method is as follows:

[0086]

[0087] In the formula, r t R represents the modal index in the time dimension. t r represents the rank of the tensor's expansion matrix over the time dimension. c R represents the modal index in the component dimension. c The rank of the tensor expansion matrix in the component dimension is represented by f, and the modality index in the feature dimension is represented by f, i.e., the eigenfactor matrix U. feat Column index;

[0088] ParaScore p The calculation method is as follows:

[0089]

[0090] In the formula, U feat [p,f] represents the contribution of the original gas path parameter p to the f-th characteristic mode.

[0091] The present invention will be further illustrated by a specific embodiment below.

[0092] This embodiment uses a turboshaft engine as an example, employing simulation data to model real-world operating scenarios and demonstrate the entire process from data acquisition to fault location and identification. The simulation data is generated based on typical aero-engine parameter models, covering multiple operating conditions and fault modes.

[0093] The settings in this embodiment are as follows:

[0094] Simulation data generation: This simulation generates operational data for four key components of a turboshaft engine (compressor, combustion chamber, gas turbine, and power turbine) at 100 time points (T=100). The parameters for each key component are as follows:

[0095] Compressor: 5 parameters (inlet temperature, exhaust temperature, pressure ratio, speed, vibration)

[0096] Combustion chamber: 4 parameters (fuel flow, temperature, pressure, vibration)

[0097] Gas turbine: 4 parameters (inlet temperature, outlet temperature, speed, vibration)

[0098] Power Turbocharger: 3 parameters (speed, torque, vibration)

[0099] The operating conditions are divided as follows: time points 1-20 are for slow operation, time points 21-60 are for cruise, time points 61-80 are for maximum continuous operation, and time points 81-100 are for takeoff. All parameters are sampled at a uniform frequency of 1Hz. If interpolation is required, linear interpolation is used. Adaptive normalization is performed according to the operating conditions, normalizing the parameter values ​​within each condition to the [0,1] interval.

[0100] In the health data, all parameters follow a standard engine model (such as a derived model based on the NASAC-MAPSS dataset). Test data injection fault: Starting at time point 50, the compressor pressure ratio continuously drops by 10% (simulating compressor blade corrosion fault), while other components remain normal. The fault persists until time point 100.

[0101] The key parameters are as follows:

[0102] The number of time points T = 100, the number of components C = 4, and the unified feature dimension d = 10 (determined dynamically by information entropy and then rounded).

[0103] The health dataset consists of health tensors at 100 time points.

[0104] The test dataset has the same structure, but includes compressor faults.

[0105] Using the tensor decomposition-based aero-engine fault diagnosis method of the first embodiment, and based on a health benchmark model constructed from a health dataset, compressor faults in the test dataset are identified and located, including the following steps:

[0106] Step S1, collect turboshaft engine operation simulation data, generate air path parameter time series data of each key component, and label the working condition type of each time point of the simulation data according to the preset working condition division standard of the engine.

[0107] Step S2, align all parameters in the simulation data according to the time stamp. Since the sampling frequencies of the simulation data parameters are consistent, no interpolation is required. For each component parameter X, adaptive normalization is performed according to the working condition using the following formula, that is:

[0108]

[0109] As shown in Figure 2 , the parameter curve of the compressor changes greatly between the healthy stage and the fault stage (pressure ratio continuously decreases after time point 50), and the working condition switching will cause parameter fluctuation, but the fault stage data (pressure ratio decreases) is independent of the working condition.

[0110] Step S3, for the air path parameter data of each key component, sequentially perform kernel principal component analysis dimension reduction and stacked denoising autoencoder feature extraction, and output a unified dimension feature vector, wherein,

[0111] Kernel principal component analysis (KPCA): use Gaussian kernel function Kernel parameter σ = 1.0. The number of principal components m = 8 (95% variance is retained). The input matrix X of the compressor ∈R 100×5 , the KPCA output F kpca ∈R 100×8 .

[0112] Stacked denoising autoencoder (SDAE): encoder-decoder structure, 3-layer network (input layer-hidden layer-bottleneck layer), noise injection rate is set to 0.1, and L1 regularization coefficient is 0.01. The dimension of the bottleneck layer is determined by information entropy: calculate the entropy H = -∑p i log p i , take the maximum entropy corresponding to the dimension d c (the dimension d c = 10 in this embodiment), and output F sdae ∈R 100×10 .

[0113] Unified dimension: global d global = min(max(d c ), 12) = 10. All components output 10-dimensional feature vectors.

[0114] Step S4, combine the feature matrix to construct the tensor wherein,

[0115] Mode 1 (time): 100 time points;

[0116] Mode 2 (component): 4 components;

[0117] Mode 3 (feature): 10 dimensional feature.

[0118] Health data tensor Training, testing data tensor With fault, for fault diagnosis.

[0119] Step S5, based on engine health state operation data, using high-order singular value decomposition (HOSVD) to train the benchmark model.

[0120] S5-1, expand the tensor to three mode matrices, respectively get time mode expansion X unfold,1 ∈R 100×40 (, component mode expansion X unfold,2 ∈R 4×1000 , feature mode expansion X unfold,1 ∈R 10×400 For each expansion matrix, do SVD decomposition, for example Get singular value vector Λ1 = [σ1, σ2,...].

[0121] S5-2, calculate the cumulative variance contribution rate, take the rank of variance ≥95%, get time mode R t = 5, component mode R c = 3, feature mode R d = 4.

[0122] S5-3, construct factor matrix: U time ∈R 100×5 (The first 5 left singular vectors), U comp ∈R 4×3 , U feat ∈R 10×4 .

[0123] S5-4, calculate the core tensor:

[0124]

[0125] Step S6, construct the test tensor Based on the health benchmark tensor Fault detection.

[0126] S6-1, project the test tensor To the health subspace:

[0127]

[0128] S6-2, reconstruct the test tensor:

[0129] S6-3, calculate global reconstruction error, component-level reconstruction error and core tensor difference, respectively;

[0130] Global reconstruction error

[0131] Component-level reconstruction error: compressor RE1 = 0.25, combustion chamber RE2 = 0.08, gas turbine RE3 = 0.07, power turbine RE4 = 0.06;

[0132] Core tensor difference

[0133] S6-4, due to the health data RE global μ RE = 0.05 and σ RE = 0.02, the threshold τ g = μ RE + 3σ RE = 0.11. Therefore, RE global = 0.15 > τ g , it is determined that the engine is malfunctioning.

[0134] Figure 3 (a) in FIG. 6-3 shows the global reconstruction error over time, and the curve shows that the test data RE global significantly increases after t = 50 (blue curve), and continues to exceed the threshold τ g = 0.11 (red dashed line); Figure 3 (b) in FIG. 6-3 shows the component-level reconstruction error bar chart, and the compressor RE1 = 0.25 is much higher than other components (combustion chamber 0.08, etc.), indicating that the compressor is abnormal.

[0135] Step S7, according to the component-level reconstruction error RE c and the core tensor difference , the malfunctioning component is located, and the fault type is identified by analyzing the feature dimension energy distribution.

[0136] S7-1, perform fault component positioning, and calculate the comprehensive abnormality score of each component (weight coefficients are α = 0.5, β = 0.3, γ = 0.2):

[0137] Compressor: Consistency1 = 0.98 (where t = 50-100 is out of limit), Score1 = 0.5 x 0.25 + 0.3 x 0.18 + 0.2 x 0.98 = 0.375; other components: Score c < 0.1.

[0138] Due to the confidence threshold τ conf =0.7, because Score1 = 0.375 > 0.7 * 0.375 / max(Score) (the proportion after normalization is > 70%), the compressor is judged to be faulty.

[0139] S7-2, perform fault type identification, first calculate the energy distribution of the feature dimension. Then take the first 10% of the energy characteristics F key ={f|E f >Q 90 (E)}, in this embodiment f=3,7 corresponds to high energy; subsequently, using By tracing back the physical parameters, since the pressure ratio ParaScore = 0.85 and the others < 0.2 in the compressor parameters, it can be matched with the fault feature library (abnormal pressure ratio → decreased compressor efficiency), and diagnosed as compressor blade failure.

[0140] Figure 4 The diagram shown is a fault location and characteristic energy analysis diagram, where (a) is a core tensor difference heatmap. (a) The component slices are concentrated in the compressor (c=1) component, and the energy in this region is significant (in dark red); (b) is a component anomaly score histogram, showing that the compressor Score1=0.375, which is much higher than other components; (c) is the characteristic energy distribution, where the energy of feature f=3,7 is significantly higher than other features and is identified as a key feature; (d) is the parameter contribution map, in which the pressure ratio is dominant.

[0141] It is easy to see that this invention constructs a three-dimensional tensor model of time, components, and features based on the characteristics of aero-engine operating data, and integrates high-order singular value decomposition and operating condition perception mechanisms, effectively overcoming the technical bottlenecks of traditional methods in multi-source heterogeneous data fusion, adaptability to changing operating conditions, and identification of weak faults. This invention can automatically resolve the coupling relationship between core tensor differences and feature factor matrices, enabling fault component location and fault type identification, significantly improving the decoupling capability and depth of state awareness for multi-dimensional fault modes in aero-engines. While preserving physical interpretability, this method overcomes the limitations on diagnostic accuracy imposed by dimensional differences and dynamic switching of operating conditions, and also provides a new approach for fault diagnosis of other equipment operating under changing conditions.

[0142] A second embodiment of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described aero-engine fault diagnosis method based on tensor decomposition.

[0143] The third embodiment of the present application relates to a computer readable storage medium, which stores a computer program, and the computer program, when executed by a processor, implements the steps of the above-mentioned method for diagnosing aero-engine faults based on tensor decomposition.

[0144] Those skilled in the art will understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0145] The present application is described with reference to the flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0146] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction method, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0147] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that carries out the functions specified in one or more flows and / or blocks.

[0148] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for aero-engine fault diagnosis based on tensor decomposition, characterized in that, The method comprises the following steps: Collecting time series data of gas path parameters of multiple key components of the to-be-tested aero-engine in a complete operation cycle, and labeling the working condition type of each time point according to the working condition division standard of the to-be-tested aero-engine; Aligning the collected time series data of the gas path parameters according to the time stamp, labeling the working condition type of each time stamp according to the working condition division standard of the to-be-tested aero-engine, and adaptively normalizing the time series data of the gas path parameters according to the working condition; Performing kernel principal component analysis dimension reduction on the normalized time series data of the gas path parameters, and performing feature extraction through a stacked denoising autoencoder to obtain a feature vector of a preset dimension; According to the number of time points, the number of key components of the aero-engine, and the preset dimension characteristic vector, a time-feature-component test tensor X is constructed test ; Projecting the test tensor X test to the healthy subspace, computing the test tensor X test global reconstruction error RE between the test tensor X global , component-level reconstruction error RE c and core tensor difference AG; When the reconstruction error RE global exceeds a set threshold, it is determined that the aero-engine under test has a fault. According to component-level reconstruction error RE c and core tensor difference AG, the faulty component of the aero-engine under test is located, and the fault type is identified by analyzing the characteristic dimension energy distribution.

2. The tensor decomposition based aeroengine fault diagnosis method of claim 1, wherein, The kernel principal component analysis dimension reduction on the normalized time series data of the gas path parameters and the feature extraction through the stacked denoising autoencoder to obtain the feature vector of the preset dimension specifically comprise: Combining all the gas path parameters of each key component into a matrix X; extracting m principal components from the matrix X using kernel principal component analysis, resulting in features F kpca wherein the kernel principal component analysis uses a Gaussian kernel function to handle non-linear relationships; The extracted feature F kpca The extracted feature F is processed by using a stacked denoising autoencoder, a robust feature representation is learned by using an encoder-decoder structure, noise injection and L1 regularization are used, and the bottleneck layer feature F of the encoder is obtained sdae , and the feature dimension d of each key component is dynamically determined according to the information entropy c ; Uniformly the feature dimensions of all the key components to obtain the feature vector of the preset dimension.

3. The tensor decomposition based aeroengine fault diagnosis method of claim 1, wherein, The health benchmark tensor X is obtained by: Collecting health time series data of gas path parameters of multiple key components of the aero-engine in a health state in a complete operation cycle, and labeling the working condition type of each time point according to the working condition division standard of the aero-engine; Aligning the collected health time series data of the gas path parameters according to the time stamp, labeling the working condition type of each time stamp according to the working condition division standard of the aero-engine, and adaptively normalizing the health time series data of the gas path parameters according to the working condition; Performing kernel principal component analysis dimension reduction on the normalized health time series data of the gas path parameters, and performing feature extraction through a stacked denoising autoencoder to obtain a health feature vector of a preset dimension; According to the number of time points, the number of key components of the aero-engine, and the health feature vector of the preset dimension, a health benchmark tensor of time-feature-component is constructed.

4. The tensor decomposition based aeroengine fault diagnosis method of claim 1, wherein, said test tensor projected onto the healthy subspace, computing a global reconstruction error RE between the test tensor and the healthy benchmark tensor global , a component-level reconstruction error RE c and a core tensor difference in particular comprises: Health benchmark tensor A benchmark model is established by high-order singular value decomposition to obtain a core tensor and a factor matrix U time , U comp , U feat ; Computing the test tensor Projection onto the healthy subspace, resulting in the test core tensor The computation is done as follows: where xi, x2, and x3represent tensors the product of the three factor matrices in the corresponding modality; based on the test core tensor reconstructing the test tensor, computed as: wherein, is the reconstructed test tensor; According to the core tensor Test tensor And the reconstructed test tensor Calculate the global reconstruction error RE global , component-level reconstruction error RE c And the core tensor difference 5. The tensor decomposition based aeroengine fault diagnosis method of claim 4, wherein, The health benchmark tensor A benchmark model is established by high-order singular value decomposition to obtain a core tensor And a factor matrix U time , U comp , U feat , specifically comprising: Obtaining a health benchmark tensor The expansion matrix of each mode is decomposed, and singular value decomposition is performed on each expansion matrix to obtain singular value vectors of each expansion matrix; The singular value cumulative variance contribution rate is calculated according to the singular value vector of each expansion matrix, and the singular value serial number corresponding to the variance contribution rate not less than a preset value is taken as the rank R of each expansion matrix i ; Select the first R i singular vectors of each unfolded matrix to form the factor matrix U time , U comp , U feat ; Based on the factor matrix U time , U comp , U feat and the health benchmark tensor The core tensor is calculated The calculation is as follows:

6. The tensor decomposition based aeroengine fault diagnosis method of claim 4, wherein, The global reconstruction error RE global is calculated as follows: The component-level reconstruction error RE c is calculated as follows: The core tensor difference is calculated as follows: where || || is the Frobenius norm, F is the Frobenius norm, is the test tensor is the tensor slice of the cth key component, is the reconstructed test tensor is the tensor slice of the cth key component.

7. The tensor decomposition based aeroengine fault diagnosis method of claim 1, wherein, The reconstruction error RE global When the reconstruction error RE g = μ RE + 3σ RE , where τ g is the threshold, μ RE and σ RE are the mean and variance of the reconstruction error set of the healthy baseline tensor , respectively.

8. The tensor decomposition based aeroengine fault diagnosis method of claim 1, wherein, The component-level reconstruction error RE c And core tensor difference Locating the faulty component of the to-be-tested aero-engine, and identifying the fault type by analyzing the characteristic dimension energy distribution, specifically comprising: Based on component-level reconstruction error RE c Differences with core tensors Calculate the overall anomaly score for each critical component. c If the overall abnormal score of a certain key component is Score c Greater than the confidence threshold τ conf If the critical component is identified as faulty, the overall anomaly score will be used to determine the faulty component. c The calculation method is as follows: a, b, g are weight coefficients, is the Frobenius norm of the core tensor difference on the critical component c, Consistency c is the abnormal temporal consistency of component c; Based on core tensor difference Compute abnormal energy distribution E on feature dimension f Select top 10% features as fault feature set F key Then compute contribution degree ParaScore of each original parameter p Match parameter contribution degree ParaScore p With known fault feature library to determine fault type, wherein, the computation method of abnormal energy distribution E f is as follows: r t represents the mode index on time dimension, R t represents the rank of the unfolding matrix of the tensor on time dimension, r c represents the mode index on component dimension, R c represents the rank of the unfolding matrix of the tensor on component dimension, f represents the mode index on feature dimension; the computation method of parameter contribution degree ParaScore p is as follows: U feat [p,f] represents the degree of contribution of the original gas path parameter p to the fth characteristic mode.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the aero-engine fault diagnosis method based on tensor decomposition as claimed in any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the aero-engine fault diagnosis method based on tensor decomposition as claimed in any one of claims 1-8.

Images