Digital twinning fault diagnosis method for gas path of aero-engine under limited unbalanced data

Abstract

The invention discloses an aero-engine gas path digital twinning fault diagnosis method under limited unbalanced data. The method comprises the steps of obtaining a mechanism model of an aero-engine gas path; finite unbalanced data of aeroengine gas path faults under different working conditions are simulated and obtained through the mechanism model of the aeroengine gas path; constructing a data driving model of limited unbalanced data; performing model training on the data-driven model according to the data training set to obtain a trained data-driven model; based on the constructed model loss function, according to the data verification set, judging whether the trained data driving model is trained to be converged or not, so as to obtain a digital twin fault diagnosis model; and inputting the data test set into the digital twinning fault diagnosis model to obtain an aero-engine gas path fault diagnosis result. According to the invention, the problem of poor fault diagnosis precision of the gas path system under limited unbalanced data in the existing fault detection method for the gas path system of the aero-engine is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft engine automatic control, and in particular to an aircraft engine gas path digital twin fault diagnosis method under limited unbalanced data. Background Art

[0002] Aircraft engine fault detection is an important means to ensure flight safety. In recent years, intelligent fault diagnosis technology represented by deep learning has rapidly emerged and has attracted much attention due to its rapid diagnosis and high accuracy. Deep learning models use different algorithms to extract features that reflect the health status of the machine from the data, and then distinguish these features to achieve fault diagnosis.

[0003] Traditional deep learning models such as recursive neural networks (RNN), autoencoders (AE), residual neural networks (ResNet) and other algorithms have been widely studied and applied to the detection of faults in the gas path system of aircraft engines to improve the accuracy of gas path system fault diagnosis and ensure flight safety. However, in the actual training process, traditional deep learning models need to rely on a large amount of data to extract highly distinguishable fault-related features. Although this improves the accuracy of diagnosis, it is often challenging to obtain the required large-scale data sets. Although there are some open source data sets for reference, given that new models of aircraft engines are launched every year, it is impossible to collect data by conducting various artificial fault experiments just because a new model is launched.

[0004] In addition, most current studies are based on the assumption that data is identically distributed, but in practice, the data collected cannot have exactly the same amount of data and be in the same distribution. This idealized premise limits the actual application scenarios of the algorithm. This has led to the fact that most of the existing models based on identical data distribution are not very effective when applied to actual scenarios. Low fault diagnosis accuracy can easily lead to uncertainty in flight safety. On the other hand, the actual collected data sets are usually unbalanced, which leads to the existence of two phenomena in the data sets: feature shift and label shift. However, most studies on the phenomenon of unbalanced data shift focus on bearing problems, but relatively little attention is paid to the gas path system. In aircraft engine maintenance, the fault diagnosis of the gas path system is more critical than the bearing fault because it directly affects the core functions of the engine, such as intake, compression, combustion, and exhaust. Abnormalities in these functions may seriously affect engine performance and even threaten flight safety. In addition, in the case of limited data, there is still a problem that there are few studies on the simultaneous existence of feature shift and label shift in unbalanced data stored in images. Summary of the invention

[0005] The present invention provides an aero-engine gas path digital twin fault diagnosis method under limited unbalanced data to overcome the above-mentioned technical problems.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] A digital twin fault diagnosis method for aero-engine gas path under limited unbalanced data specifically includes the following steps:

[0008] S1: obtaining an aircraft engine gas path system model, and obtaining an aircraft engine gas path mechanism model based on the aircraft engine gas path system model;

[0009] S2: Through the mechanism model of the aircraft engine gas path, simulate and obtain the limited imbalance data of the aircraft engine gas path failure under different working conditions;

[0010] And divide the limited unbalanced data into data training set, data verification set and data test set;

[0011] And the data training set includes a source domain sample set and a target domain sample set;

[0012] S3: Construct a data-driven model for digital twin fault diagnosis of aircraft engine gas path based on limited imbalance data;

[0013] S4: Perform model training on the data-driven model according to the data training set to obtain the trained data-driven model;

[0014] S5: Based on the constructed model loss function, determine whether the trained data-driven model has been trained to convergence according to the data validation set;

[0015] If the trained data-driven model converges, the trained data-driven model is used as the digital twin fault diagnosis model; otherwise, the weight parameters in the data-driven model are adaptively adjusted based on the back propagation method, and step S4 is repeated;

[0016] S6: Input the data test set into the digital twin fault diagnosis model to obtain the aircraft engine gas path fault diagnosis results.

[0017] Furthermore, the data-driven model constructed in S3 includes a one-dimensional convolutional neural network module, a domain distribution adaptation module, and a reweighting / regularization module connected in sequence;

[0018] The one-dimensional convolutional neural network module is used to extract the gas path fault characteristics of aircraft engines and make initial predictions of fault categories based on the gas path fault characteristics;

[0019] The domain distribution adaptation module is used to align the domain distribution of the output data of the one-dimensional convolutional neural network module, and optimize the prediction of the fault category of the gas path fault characteristics after the domain distribution alignment;

[0020] The reweighting / regularization module is used to perform class-balanced label distributed perception on the gas path fault features and the corresponding optimized predicted fault labels output by the domain distribution adaptation module to obtain the predicted fault category.

[0021] Furthermore, the S1 specifically includes the following steps:

[0022] S11: Get the aero-engine gas path system model, which is expressed as

[0023] y t =g(u 0:t ,t;θ)

[0024] Where: g(u 0:t ,t; θ) represents the air path system model of the aircraft engine; u 0:t represents the control input of the gas path system model; t represents the time parameter; θ represents the model design parameter of the gas path system model; y t Represents the output of the gas path system model;

[0025] S12: Based on the aircraft engine gas path system model, the mechanism model of the aircraft engine gas path is obtained, and its expression is:

[0026] z t =g(u 0:t ,t;θ)+ε t

[0027] Where: z t Represents the output of the mechanical model of the aircraft engine gas path; ε t Represents the model error parameter that satisfies the normal distribution.

[0028] Furthermore, the one-dimensional convolutional neural network module in S3 includes an input layer, a plurality of sequentially connected feature extraction network layers, and an output classification layer, wherein each of the feature extraction network layers has a different convolution kernel to extract gas path fault features of aircraft engines of different scales;

[0029] The input layer is used to input the sample data of the source domain sample set and the target domain sample set into the feature extraction network layer respectively;

[0030] And the feature extraction network layer includes a convolution layer, a batch normalization layer, an activation function layer, a pooling layer and a random drop layer;

[0031] The convolution layer is used to extract the fault data features of the aircraft engine gas path;

[0032] The batch normalization layer is used to normalize the output of the convolutional layer;

[0033] The activation function layer is used to perform a ReLU activation operation on the output of the batch normalization layer;

[0034] The pooling layer is used to perform a pooling operation on the output of the activation function layer;

[0035] The random discarding layer is used to perform a random discarding operation on the output data of the pooling layer according to a preset discarding probability to obtain the gas path fault characteristics of the aircraft engine;

[0036] The output classification layer is used to make an initial prediction of the fault category according to the gas path fault characteristics, and obtain a source domain sample feature set and a target domain sample feature set.

[0037] Furthermore, the expression for domain distribution alignment of the output data of the one-dimensional convolutional neural network module by the domain distribution adaptation module in S3 is:

[0038]

[0039] Where: d LMMD Represents the difference between the source domain sample feature set and the target domain sample feature set; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying D s With D t represent the source domain sample set and the target domain sample set respectively; and The kernel function representing the sample feature map in the source domain sample set; The kernel function representing the sample feature map in the target domain sample set; The kernel function representing the sample feature mapping between the source domain sample set and the target domain sample set; It represents the value obtained after the source domain or target domain sample feature passes through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain sample set; n t Represents the total number of samples in the target domain sample set.

[0040] Furthermore, the construction method of the reweighting / regularization module in S3 specifically includes:

[0041] S31: define the optimized prediction example output by the domain distribution adaptive module as Θ, and Θ = (x, y), where x represents the gas path fault feature output by the domain distribution adaptive module, y represents the optimized prediction fault category label, and y∈{1, ...C};

[0042] Define the output logarithm of a one-dimensional convolutional neural network module as Z = [z1,…,z C ], where z i =log e Xout , X out Represents the output of a one-dimensional convolutional neural network module;

[0043] S32: Define the reweighting factor α for each type of predicted fault category sample cm , whose expression is

[0044]

[0045] Where: β m represents the optimal parameters of the data-driven model found by the grid search method for each training; n c Represents the total number of samples in the data training set;

[0046] The marginal loss regularization is performed on each type of predicted fault category, and its expression is:

[0047]

[0048] Where: z i Represents the output logarithm of the neural network of the i-th sample; Δ i Represents the minimum distance from the i-th sample to the decision boundary of this type of sample; Represents the neural network output value of the cth sample; Represents the neural network output value of the i-th sample;

[0049] S33: constructing a model of class-balanced label distributed perception margin according to step S32, i.e., a reweighting / regularization module;

[0050] The reweighting / regularization module is used to perform distributed perception of class-balanced labels on the gas path fault features output by the domain distribution adaptation module and the corresponding optimized predicted fault labels to obtain the predicted fault category; the expression of the reweighting / regularization module is:

[0051]

[0052] Furthermore, the model loss function constructed in S5 is expressed as

[0053]

[0054] Where: Loss all Represents the model loss function; L CBDM The output of the reweighting / regularization module; represents the intermediate variable parameter of the model, ξ represents the constant parameter; t m represents the maximum number of training times; t represents the current number of training times.

[0055] Beneficial effect: The present invention provides a digital twin fault diagnosis method for aero-engine gas path under limited unbalanced data, which realizes fault diagnosis by combining a mechanism model with a data-driven model. That is, a mechanism model is constructed based on the aero-engine to simulate and obtain limited unbalanced data of aero-engine gas path faults under different working conditions, and the features of the limited unbalanced data about fault information are extracted through a one-dimensional convolutional neural network, and the feature distribution is aligned using domain adaptation technology to achieve optimized prediction of the fault type. In order to solve the problem of category imbalance, a reweighted / regularized module is used to perform distributed perception of class-balanced labels to obtain predicted fault categories, thereby enhancing the generalization ability of minority classes. The present invention accurately simulates the operating status and fault mode of the aero-engine by integrating data-driven and mechanism models, thereby achieving more efficient fault diagnosis and predictive maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0057] Figure 1 It is a flow chart of the digital twin fault diagnosis method of aero-engine gas path under limited unbalanced data of the present invention;

[0058] Figure 2 This is a model diagram of the mechanism modeling of the aero-engine gas path system in this embodiment;

[0059] Figure 3 is a schematic diagram of distribution alignment in this embodiment;

[0060] Figure 4 is a schematic diagram of a reweighting / regularization module in this embodiment;

[0061] Figure 5 This is a comparison chart of the F1 score results of the four-category fault diagnosis in this embodiment;

[0062] Figure 6 This is a comparison chart of the F1 score results of the six-category fault diagnosis in this embodiment;

[0063] Figure 7 This is a comparison diagram of the confusion matrix results of the four-category fault diagnosis in this embodiment;

[0064] Figure 8 This is a comparison chart of the confusion matrix results of the six-category fault diagnosis in this embodiment. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] This embodiment provides a method for diagnosing faults in the digital twin of an aircraft engine gas path under limited unbalanced data. Figure 1 As shown, the specific steps include:

[0067] S1: Obtain an aircraft engine gas path system model, and obtain an aircraft engine gas path mechanism model based on the aircraft engine gas path system model, such as Figure 2 As shown, the specific steps include:

[0068] S11: Get the aero-engine gas path system model, which is expressed as

[0069] y t =g(u 0:t ,t;θ)

[0070] Where: g(u 0:t ,t; θ) represents the gas path system model of the aircraft engine, such as the JT9D aircraft engine; u 0:t represents the control input of the gas path system model; t represents the time parameter; θ represents the model design parameter of the gas path system model; y t Represents the output of the gas path system model;

[0071] S12: Based on the aircraft engine gas path system model, the mechanism model of the aircraft engine gas path is obtained, and its expression is:

[0072] z t =g(u 0:t ,t;θ)+ε t

[0073] Where: z t Represents the output of the mechanical model of the aircraft engine gas path; ε t represents the model error parameter that satisfies the normal distribution, that is, the noise and

[0074] S2: Through the mechanism model of the aircraft engine gas path, simulate and obtain the limited unbalanced data of the aircraft engine gas path failure under different working conditions; limited unbalanced data: limited data refers to the unbalanced data with limited data volume and under different working conditions. The limited unbalanced data of the aircraft engine gas path failure is an existing professional term, which will not be explained in detail here;

[0075] The limited imbalanced data is divided into a data training set, a data verification set and a data test set, and the data training set includes a source domain sample set and a target domain sample set;

[0076] S3: Building data-driven models with limited imbalanced data;

[0077] The constructed data-driven model includes a one-dimensional convolutional neural network module, a domain distribution adaptation module, and a reweighting / regularization module connected in sequence;

[0078] The one-dimensional convolutional neural network module is used to extract the gas path fault characteristics of aircraft engines and make initial predictions of fault categories based on the gas path fault characteristics;

[0079] The domain distribution adaptation module is used to align the domain distribution of the output data of the one-dimensional convolutional neural network module, and optimize the prediction of the fault category of the gas path fault characteristics after the domain distribution alignment;

[0080] The reweighting / regularization module is used to perform distributed perception of class-balanced labels on the gas path fault features and corresponding optimized predicted fault labels output by the domain distribution adaptation module to obtain predicted fault categories;

[0081] In a specific embodiment, the one-dimensional convolutional neural network module includes an input layer, a plurality of sequentially connected feature extraction network layers and an output classification layer, each of the feature extraction network layers has a different convolution kernel to extract gas path fault features of aircraft engines of different scales;

[0082] The input layer is used to input the sample data of the source domain sample set and the target domain sample set into the feature extraction network layer respectively;

[0083] And the feature extraction network layer includes a convolution layer, a batch normalization layer, an activation function layer, a pooling layer and a random drop layer;

[0084] The convolution layer is used to extract the fault data features of the aircraft engine gas path, and the expression is:

[0085] X out =X in *W+b

[0086] Among them, X in represents the input finite unbalanced data; W represents the convolution kernel; b represents the bias term; * represents the convolution operation;

[0087] The batch normalization layer is used to normalize the output of the convolutional layer;

[0088] The expression of batch normalization is:

[0089] Where X represents the input feature X out ; μ and σ 2 are the mean and variance of the data in the batch respectively; γ and β represent learnable parameters; ε is a very small number to prevent the denominator from being zero; batch normalization normalizes the input of each layer to make the distribution of input data more stable, which helps the training and generalization of the network;

[0090] The activation function layer is used to perform a ReLU activation operation on the output of the batch normalization layer;

[0091] ReLU activation operation: For any input value X, when X is greater than 0, the output is X itself; when X is less than or equal to 0, the output is 0;

[0092] The pooling layer is used to perform a pooling operation on the output of the activation function layer;

[0093] This embodiment can speed up the calculation and prevent overfitting by adding a pooling operation;

[0094] The random discarding layer is used to perform a random discarding operation on the output data of the pooling layer according to a preset discarding probability to obtain the gas path fault characteristics of the aircraft engine;

[0095] In this embodiment, a Dropout layer with a dropout rate of 50% is also added to the network;

[0096] Expression of Dropout layer: Y = M⊙X

[0097] Where M represents a binary (0 or 1) matrix of the same shape as X, where the elements in M ​​are randomly selected to be 0 or 1 according to the preset drop probability;

[0098] The output classification layer is used to make an initial prediction of the fault category according to the gas path fault characteristics, and obtain a source domain sample feature set and a target domain sample feature set;

[0099] In a specific embodiment, the domain distribution adaptation module in S3 performs domain distribution alignment on the output data of the one-dimensional convolutional neural network module, specifically:

[0100] Get the difference between the source domain and the target domain, the expression is

[0101]

[0102] Where: Represents the difference value between the source domain and the target domain; and Represent the feature maps of samples in the source domain and the target domain respectively;

[0103] For a sample x i , with probability vector And calculate the weight for:

[0104]

[0105] in, Indicates p i The cth probability vector in ; Note that for the labeled sample set D s The sample in p i can be regarded as a one-hot vector, and for the unlabeled sample set D t For samples in , the predicted distribution of the network output is used as the pseudo label;

[0106] According to the calculation weight The difference between the source domain and the target domain can be used to obtain the local maximum mean difference (LMMD) formula, that is, the expression for domain distribution alignment of the output data of the one-dimensional convolutional neural network module is:

[0107]

[0108] Where: d LMMD Represents the difference between the source domain sample feature set and the target domain sample feature set; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying D s With D t represent the source domain sample set and the target domain sample set respectively; and k(·,·) represents the kernel and The kernel function representing the sample feature map in the source domain sample set; The kernel function representing the sample feature map in the target domain sample set; The kernel function representing the sample feature mapping between the source domain sample set and the target domain sample set; It represents the value obtained after the source domain or target domain sample feature passes through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain sample set; n t Represents the total number of samples in the target domain sample set;

[0109] In a specific embodiment, the method for constructing the reweighting / regularization module in S3 specifically includes:

[0110] S31: define the optimized prediction example output by the domain distribution adaptive module as Θ, and Θ = (x, y), where x represents the gas path fault feature output by the domain distribution adaptive module, y represents the optimized prediction fault category label, and y∈{1, ...C};

[0111] Define the output logarithm of a one-dimensional convolutional neural network module as Z = [z1,…,z C ], where z i =log e X out , X out Represents the output of a one-dimensional convolutional neural network module; for each class c weight Defined as:

[0112]

[0113] Among them, β∈[0,1] is a preset hyperparameter related to the whole data set. For the so-called class balance, this embodiment defines a reweighting factor α c ;

[0114]

[0115] In this embodiment, in order to maintain the weight α c To accommodate the consistency of different dataset sizes, rewrite it as

[0116]

[0117] in, represents the weight of the i-th sample. c The β in the equation is a hyperparameter that needs to be adjusted manually. In this embodiment, the grid search method is used to determine the β value as β in order to improve the accuracy of the model. m , to obtain the reweighted weight factor α cm ;

[0118] S32: Define the reweighting factor α for each type of predicted fault category sample cm , whose expression is

[0119]

[0120] Where: β m represents the optimal parameters of the data-driven model found by the grid search method for each training; n c Represents the total number of samples in the data training set;

[0121] The marginal loss regularization is performed on each type of predicted fault category, and its expression is:

[0122]

[0123] Where: z i Represents the output logarithm of the neural network of the i-th sample; Δ i Represents the minimum distance from the i-th sample to the decision boundary of this type of sample; Represents the neural network output value of the cth sample; Represents the neural network output value of the i-th sample;

[0124] S33: constructing a model of class-balanced label distributed perception margin according to step S32, i.e., a reweighting / regularization module;

[0125] And according to the reweighting / regularization module, the gas path fault characteristics output by the domain distribution adaptation module and the corresponding optimized predicted fault labels are subjected to distributed perception of class balanced labels to obtain the predicted fault category; the expression of the reweighting / regularization module is:

[0126]

[0127] like Figure 4 As shown, in order to improve the generalization ability of minority classes and reduce boundary fragility, this embodiment proposes a class-balanced label distributed perception margin CBDM module that combines reweighting and regularization, namely a reweighting / regularization module; in the CBDM module, each class of samples has its own reweighting weight factor, which is used to adjust the boundary, thereby improving the boundary fragility problem;

[0128] S4: Perform model training on the data-driven model according to the data training set to obtain the trained data-driven model;

[0129] S5: Based on the constructed model loss function, determine whether the trained data-driven model has been trained to convergence according to the data validation set;

[0130] The constructed model loss function is expressed as

[0131]

[0132] Where: Loss all Represents the model loss function; L CBDM The output of the reweighting / regularization module; represents the intermediate variable parameter of the model, ξ represents the constant parameter; t m represents the maximum number of training times; t represents the current number of training times.

[0133] If the trained data-driven model converges, the trained data-driven model is used as the digital twin fault diagnosis model; otherwise, the weight parameters in the data-driven model are adaptively adjusted based on the back propagation method, and step S4 is repeated;

[0134] S6: Input the data test set into the digital twin fault diagnosis model to obtain the aircraft engine gas path fault diagnosis results.

[0135] Beneficial effects of this embodiment: This embodiment first builds a mechanism model of an aircraft engine, then performs data modeling and fusion on one-dimensional convolutional neural networks, domain adaptation and reweighted regularization, and obtains a data-driven model to correct unbalanced data distribution and improve diagnostic accuracy. That is, the mechanism modeling of the aircraft engine gas path system is carried out; the limited data is extracted by a one-dimensional convolutional neural network, the distribution is aligned by domain adaptation technology, and the bias and boundary fragility of the model are corrected by cost-sensitive reweighting and regularization technology. Finally, they are combined to establish a data-driven model, and finally the integrated model is used to perform fault diagnosis and monitoring on the aircraft engine gas path system.

[0136] This embodiment conducts a simulation verification study on a proposed aero-engine gas path digital twin fault diagnosis method under limited unbalanced data to verify its effectiveness:

[0137] The working condition of the gas path system when the aircraft engine is working is simulated by using the mechanism model. The data is collected to form a data set and stored in the form of images. The dimension of each sample is [1, 8, 500]. The data set contains four working states, namely healthy (0), Fan02 (1), Fan03 (2) and LPC02 (3). There are 240 healthy state data in the source domain data set and 36 other fault data. The imbalance ratio reaches 6.67:1, which is a serious imbalance state.

[0138] Each of the four states of the target domain dataset contains 400 data. In addition, the source domain dataset and the target domain dataset are divided into four working modes under each type of label, so that the formed dataset meets the setting of the problem in this embodiment. The source domain dataset and the target domain dataset images are respectively as follows: Figure 3 In the experiment, this embodiment divides the target domain, does not change the original distribution during the division process, uses 70% of the target domain data set as the training set, and uses 30% as the test set for the experiment; in the simulation test, the model parameters are as follows: Bateh size = 32, Max epoch = 80, Learning rate = 0.01, Weight decay = 1e-3, Label smoothing = 0.1, Trade off factor = 0.05.

[0139] This embodiment uses software to perform computer simulation, and the results are compared with the unbalanced data algorithm in existing research. Figures 5 to 8 shown. Figures 5 and 6 It is the F1 score of the diagnostic accuracy after model diagnosis. Figures 7 and 8 Confusion matrix diagram for model diagnosis. Figures 5 to 8 It can be seen that the established digital twin fault diagnosis model improves the influence of the limited unbalanced aircraft engine gas path system data under variable working conditions on the fault diagnosis accuracy. Therefore, this embodiment can conclude that the proposed aircraft engine gas path system fault diagnosis method under limited data is effective.

[0140] In summary, this embodiment realizes fault diagnosis by combining the mechanism model with the data-driven model. First, a mechanism model is constructed based on the JT9D aircraft engine, and then the convolutional neural network (CNN) and domain adaptation technology are used for feature extraction and distribution alignment. In order to solve the problem of class imbalance, the class-balanced label distributed perception margin (CBDM) method is adopted to enhance the generalization ability of minority classes. Finally, by integrating data-driven and mechanism models, the operating status and failure mode of the aircraft engine are accurately simulated, achieving more efficient fault diagnosis and predictive maintenance.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital twin fault diagnosis method for aero-engine gas path under limited unbalanced data, characterized in that: The specific steps include: S1: obtaining an aircraft engine gas path system model, and obtaining an aircraft engine gas path mechanism model based on the aircraft engine gas path system model; S2: Through the mechanism model of the aircraft engine gas path, simulate and obtain the limited imbalance data of the aircraft engine gas path failure under different working conditions; The limited imbalanced data is divided into a data training set, a data verification set and a data test set, and the data training set includes a source domain sample set and a target domain sample set; S3: Construct a data-driven model for digital twin fault diagnosis of aircraft engine gas path based on limited imbalance data; S4: Perform model training on the data-driven model according to the data training set to obtain the trained data-driven model; S5: Based on the constructed model loss function, determine whether the trained data-driven model has been trained to convergence according to the data validation set; If the trained data-driven model converges, the trained data-driven model is used as the digital twin fault diagnosis model; otherwise, the weight parameters in the data-driven model are adaptively adjusted based on the back propagation method, and step S4 is repeated; S6: Input the data test set into the digital twin fault diagnosis model to obtain the aircraft engine gas path fault diagnosis results.

2. According to the method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data in claim 1, it is characterized in that: The data-driven model constructed in S3 includes a one-dimensional convolutional neural network module, a domain distribution adaptation module, and a reweighting / regularization module connected in sequence; The one-dimensional convolutional neural network module is used to extract the gas path fault characteristics of aircraft engines and make initial predictions of fault categories based on the gas path fault characteristics; The domain distribution adaptation module is used to align the domain distribution of the output data of the one-dimensional convolutional neural network module, and optimize the prediction of the fault category of the gas path fault characteristics after the domain distribution alignment; The reweighting / regularization module is used to perform class-balanced label distributed perception on the gas path fault features and the corresponding optimized predicted fault labels output by the domain distribution adaptation module to obtain the predicted fault category.

3. The method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data according to claim 2 is characterized in that: The S1 specifically includes the following steps: S11: Get the aero-engine gas path system model, which is expressed as y t =g(u 0:t ,t;θ) Where: g(u 0:t ,t; θ) represents the air path system model of the aircraft engine; u 0:t represents the control input of the gas path system model; t represents the time parameter; θ represents the model design parameter of the gas path system model; y t Represents the output of the gas path system model; S12: Based on the aircraft engine gas path system model, the mechanism model of the aircraft engine gas path is obtained, and its expression is: z t =g(u 0:t ,t;θ)+ε t Where: z t Represents the output of the mechanical model of the aircraft engine gas path; ε t Represents the model error parameter that satisfies the normal distribution.

4. The method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data according to claim 3 is characterized in that: The one-dimensional convolutional neural network module in S3 includes an input layer, a plurality of sequentially connected feature extraction network layers, and an output classification layer, wherein each feature extraction network layer has a different convolution kernel to extract gas path fault features of aircraft engines of different scales; The input layer is used to input the sample data of the source domain sample set and the target domain sample set into the feature extraction network layer respectively; And the feature extraction network layer includes a convolution layer, a batch normalization layer, an activation function layer, a pooling layer and a random drop layer; The convolution layer is used to extract the fault data features of the aircraft engine gas path; The batch normalization layer is used to normalize the output of the convolutional layer; The activation function layer is used to perform a ReLU activation operation on the output of the batch normalization layer; The pooling layer is used to perform a pooling operation on the output of the activation function layer; The random discarding layer is used to perform a random discarding operation on the output data of the pooling layer according to a preset discarding probability to obtain the gas path fault characteristics of the aircraft engine; The output classification layer is used to make an initial prediction of the fault category according to the gas path fault characteristics, and obtain a source domain sample feature set and a target domain sample feature set.

5. The method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data according to claim 4 is characterized in that: The expression for domain distribution alignment of the output data of the one-dimensional convolutional neural network module by the domain distribution adaptation module in S3 is: Where: d LMMD Represents the difference between the source domain sample feature set and the target domain sample feature set; and Representation sample With sample The probability of belonging to the air path fault category c of the aircraft engine and satisfying and D s With D t represent the source domain sample set and the target domain sample set respectively; and The kernel function representing the sample feature map in the source domain sample set; The kernel function representing the sample feature map in the target domain sample set; The kernel function representing the sample feature mapping between the source domain sample set and the target domain sample set; It represents the value obtained after the source domain or target domain sample feature passes through the activation function; C represents the number of fault categories; n s Represents the total number of samples in the source domain sample set; n t Represents the total number of samples in the target domain sample set.

6. The method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data according to claim 5 is characterized in that: The construction method of the reweighting / regularization module described in S3 specifically includes: S31: define the optimized prediction example output by the domain distribution adaptive module as Θ, and Θ = (x, y), where x represents the gas path fault feature output by the domain distribution adaptive module, y represents the optimized prediction fault category label, and y∈{1, ...C}; Define the output logarithm of a one-dimensional convolutional neural network module as Z = [z1,…,z C ], where z i =log e X out , X out Represents the output of a one-dimensional convolutional neural network module; S32: Define the reweighting factor α for each type of predicted fault category sample cm , whose expression is Where: β m represents the optimal parameters of the data-driven model found by the grid search method for each training; n c Represents the total number of samples in the data training set; The marginal loss regularization is performed on each type of predicted fault category, and its expression is: Where: z i Represents the output logarithm of the neural network of the i-th sample; Δ i Represents the minimum distance from the i-th sample to the decision boundary of this type of sample; Represents the neural network output value of the cth sample; Represents the neural network output value of the i-th sample; S33: constructing a model of class-balanced label distributed perception margin according to step S32, i.e., a reweighting / regularization module; And according to the reweighting / regularization module, the gas path fault characteristics output by the domain distribution adaptation module and the corresponding optimized predicted fault labels are subjected to distributed perception of class balanced labels to obtain the predicted fault category; the expression of the reweighting / regularization module is:

7. The method for digital twin fault diagnosis of aero-engine gas path under limited unbalanced data according to claim 6 is characterized in that: The model loss function constructed in S5 is expressed as Where: Loss all Represents the model loss function; L CBDM The output of the reweighting / regularization module; represents the intermediate variable parameter of the model, ξ represents the constant parameter; t m represents the maximum number of training times; t represents the current number of training times.

Images