A method for fault diagnosis of aero-engine gas path considering out-of-distribution faults

Abstract

The present invention provides an aircraft engine gas path fault diagnosis method that considers out-of-distribution faults, belonging to the field of aircraft engine fault diagnosis. The method comprises: collecting and preprocessing data; establishing an MLP feature extraction model Net1 and training the model to obtain a trained MLP feature extraction model Net1; training an out-of-distribution fault diagnosis model Net2 based on the MLP feature extraction model Net1; retaining the output of the second hidden layer of the MLP feature extraction model Net1 to obtain a projection of the real data in the feature space, and then fitting the projection of each type of data using a Gaussian distribution; sampling the fitted data distribution to obtain synthetic out-of-distribution data; estimating the synthetic out-of-distribution data and in-distribution data using an energy function, determining an uncertainty threshold, and obtaining a trained out-of-distribution fault diagnosis model Net2. The present invention can accurately determine whether an aircraft engine gas path fault is an in-distribution fault, avoiding the identification of unknown faults as known faults.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft engine fault diagnosis, and in particular relates to an aircraft engine gas path fault diagnosis method taking into account out-of-distribution faults. Background Art

[0002] Aircraft engines operate in harsh environments, often under variable and high-load conditions. Failure modes are diverse, resulting in maintenance costs accounting for over 40% of the total engine maintenance cost. Among all aircraft engine failure modes, air path system failures account for over 90%, and their repair costs reach 60% of the total engine maintenance cost. Air path system failure modes generally include erosion, fouling, corrosion, inter-lobe wear, and foreign object damage. Accurate air path system fault diagnosis can quickly locate faults, effectively reduce maintenance costs, and avoid significant economic losses and safety accidents.

[0003] Current data-driven methods for detecting air path faults in aircraft engines generally assume that the training and test data sets are independent and identically distributed. However, when this assumption fails, that is, when faced with a fault type not found in the training set, the neural network will produce a deterministic result, identifying it as a fault type in the training set. Hendrycks et al. published the paper "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks" at the 2017 International Conference on Learning Representations (ICLR). They proposed the concept of out-of-distribution (OOD) fault detection, which detects objects of unknown categories while maintaining the original task. In recent years, OOD detection has garnered widespread attention and has been extensively studied in fields such as computer vision, natural language processing, and fault diagnosis, aiming to improve the robustness of models to unknown samples based on the original deep learning task. Supervised OOD detection utilizes labeled ID (in-distribution) and OOD samples, treating OOD detection as a supervised classification problem. Wu et al. published a paper titled "Detecting Unexpected Faults of High-Speed ​​Train Bogie Based on Bayesian Deep Learning" in the journal IEEE Transactions on Vehicular Technology. They proposed a Bayesian unexpected fault detection method that only requires a small number of unexpected fault diagnosis samples to indicate whether an anomaly belongs to a known category. This method is applied to monitoring concurrent and out-of-order (OOD) faults in high-speed train bogies. However, since real OOD samples are difficult to obtain, generating OOD samples for training is an intuitive approach.In 2019, Verneker et al. published the paper "Out-of-Distribution Detection in Classifiers via Generation" at the Conference and Workshop on Neural Information Processing Systems (NeurIPS). The paper used a conditional variational autoencoder to generate two types of OOD samples: those outside the sample manifold and those on the sample manifold, effectively covering the entire distribution boundary. In 2021, Gao et al. published the paper "Uncertainty Enhanced Attention for OOD Detection" at the International Joint Conference on Neural Networks (IJCNN). The paper used a generative adversarial network to generate OOD samples and introduced an attention mechanism to improve detection accuracy. Although supervised methods have high accuracy, they face the challenge of obtaining high-quality labeled OOD samples. Summary of the Invention

[0004] This paper proposes a method for diagnosing aeroengine gas path faults that considers out-of-distribution faults. The method first extracts features, then fits the feature distribution. The method integrates the feature extraction method with Gaussian distribution fitting, and uses an MLP feature extraction network to obtain the projection of in-distribution data in feature space. The Gaussian distribution is then used to fit the feature distributions of various fault types. Data at the edge of the distribution is then sampled to generate synthetic out-of-distribution data for training uncertainty thresholds to identify ID and OOD samples. This method addresses the difficulty in obtaining high-quality labeled OOD samples in existing technologies.

[0005] In order to solve the above technical problems, the specific technical solutions of the present invention are as follows:

[0006] A method for diagnosing air path faults in an aircraft engine taking into account out-of-distribution faults, the method comprising the following steps:

[0007] Step S1: using a plurality of different types of sensors to collect engine operating data as sample data.

[0008] Step S2: pre-processing the sample data collected in step S1 to obtain pre-processed sample data.

[0009] Step S3: For the sample data preprocessed in step S2, the fault type corresponding to each sample data is used as a label; one type of fault type sample in all the sample data is excluded and used as the subsequent true out-of-distribution samples, and the remaining data including normal samples and other fault samples constitute the in-distribution samples; the in-distribution samples are divided into training set, validation set and test set in a ratio of 7:2:1, and the excluded fault samples are retained as the true out-of-distribution data for model evaluation.

[0010] Step S41: Build an MLP feature extraction model Net1, which consists of an input layer, a first hidden layer, a second hidden layer, an output layer, and a classifier.

[0011] Step S42: constructing the loss function of the MLP feature extraction model Net1 using the mean square error.

[0012] Step S43: Train the MLP feature extraction model Net1, calculate the gradient of the loss function with respect to the model parameters to guide the update of the weights, and obtain the trained MLP feature extraction model Net1.

[0013] Step S44: Use the trained MLP feature extraction model Net1 to extract features from the samples in the distribution, export the output of the second hidden layer, obtain the projection of the real data in the feature space, and use Gaussian distribution to fit the projection of each type of data.

[0014] Step S45: sampling the distribution edge data of the feature space based on the fitted Gaussian distribution to obtain synthetic out-of-distribution samples.

[0015] Step S46: Evaluate the synthesized out-of-distribution samples and in-distribution samples through the energy function, construct the energy uncertainty loss function, obtain the optimal training parameters through training, calculate the uncertainty, and obtain a suitable uncertainty threshold so that the in-distribution samples have negative energy and the synthesized out-of-distribution samples have positive energy, and obtain the trained out-of-distribution fault diagnosis model Net2.

[0016] The energy function is as follows:

[0017]

[0018] Where E(x;θ) is the energy function; x is the input vector; θ is the trainable parameter; K represents the total number of fault categories; f k (x;θ) is the score of the kth class, calculated as follows:

[0019]

[0020] in, is the network weight of the output layer, and T represents transpose.

[0021] Construct the energy uncertainty loss function, which is expressed as follows:

[0022]

[0023] in, represents the energy uncertainty loss function; represents the mathematical expectation operation of a sample v sampled from the in-distribution sample set V; E(v;θ) represents the estimated value of the energy function for the sample v; φ(·) is a nonlinear MLP function used to distinguish in-distribution samples from synthetic out-of-distribution samples; It represents the mathematical expectation operation on the sample x sampled from the sample set D outside the synthetic distribution.

[0024] Through the energy uncertainty loss function Get the optimal training parameter θ * , according to the obtained optimal training parameters θ * , using the in-distribution samples and the synthetic out-of-distribution samples to calculate the uncertainty, the uncertainty threshold δ is calculated as follows:

[0025]

[0026] Among them, x * represents the input containing samples in the distribution and samples out of the synthetic distribution; G(x * ) represents x * The uncertainty value of E(x * θ * ) represents the energy function; the input parameter of the energy function is x * and the optimal training parameters θ * The uncertainty threshold makes the samples within the distribution have negative energy and the samples outside the synthetic distribution have positive energy. The uncertainty threshold δ is adjusted according to the recognition of the samples within the distribution and the samples outside the synthetic distribution. The uncertainty threshold is calculated for the samples within the distribution and the samples outside the synthetic distribution respectively, and then an uncertainty threshold δ that makes the out-of-distribution fault diagnosis accuracy the highest is selected. The training is then completed, and the trained out-of-distribution fault diagnosis model Net2 is obtained.

[0027] Step S5: Use the test set and real out-of-distribution data to test the trained out-of-distribution fault diagnosis model Net2 to obtain the identification results of out-of-distribution faults or in-distribution faults.

[0028] Furthermore, the step S2 includes:

[0029] Step S21: clear incomplete data and remove outliers to obtain cleaned sample data;

[0030] Step S22: normalize the cleaned sample data. The normalization method is as follows:

[0031]

[0032] Among them, x is the sample data after cleaning; x' is the normalized data; x min is the minimum value of the cleaned sample data; x max is the maximum value of the cleaned sample data.

[0033] Furthermore, step S41 includes the following steps:

[0034] The samples in the distribution are flattened to obtain the input vector x, which is transmitted to the first hidden layer through the input layer. The input-output conversion relationship of the first hidden layer is as follows:

[0035]

[0036] in, represents the input of the jth neuron in the first hidden layer; x i Represents the i-th component of the input layer input vector; is the network weight of the first hidden layer, which represents the connection weight between the j-th neuron of the first hidden layer and the i-th component of the input vector of the input layer; I represents the dimension of the input layer; represents the output of the jth neuron in the first hidden layer; ReLU represents the ReLU activation function.

[0037] The output of the first hidden layer is transmitted to the second hidden layer, and the input-output conversion relationship of the second hidden layer is as follows:

[0038]

[0039] in, represents the input of the jth neuron in the second hidden layer; is the network weight of the second hidden layer, which represents the connection weight between the jth neuron in the second hidden layer and the ith neuron in the first hidden layer; M represents the number of neurons in the first hidden layer; represents the output of the second hidden layer.

[0040] The output of the second hidden layer is transmitted to the output layer. The input-output conversion relationship of the output layer is as follows:

[0041]

[0042] Among them, y k is the output of the kth neuron in the output layer; J represents the number of neurons in the second hidden layer;

[0043] The output of the output layer is transmitted to the classifier for normalization. The normalization of the classifier is achieved through the Softmax function, which is defined as follows:

[0044]

[0045] Among them, Softmax(y k ) is the value of the k-th neuron output in the output layer after normalization by the classifier; K is the number of nodes in the output layer, which also represents the total number of output categories.

[0046] Furthermore, in step S42, the loss function of the MLP feature extraction model Net1 is calculated using mean square error, which is expressed as follows:

[0047]

[0048] Among them, L represents the Net1 loss function of the MLP feature extraction model, K represents the total number of output categories; t k Indicates the value of the kth position after the true category label t is converted into a vector.

[0049] Furthermore, in step S44,

[0050] The distribution of samples within the distribution in the low-dimensional feature space is fitted with a Gaussian distribution, and the formula is as follows:

[0051] p θ (h(x)|t k =k) ​​= N(μ k ,∑)

[0052] Among them, p θ (h(x)|t k =k) ​​represents the distribution of samples in the low-dimensional feature space; h(x)∈R m is the representation of the input vector x in the low-dimensional feature space, that is, the output of the second hidden layer, m represents the dimension of the feature space; t k represents the true category label of the sample; μ k is the Gaussian mean of fault categories k∈{1,2,…,K}, where K represents the total number of fault categories; ∑ is the bound covariance matrix; N(μ k k,∑) represents μ k The mean and ∑ are Gaussian distributions with covariance.

[0053] Furthermore, in step S45, synthetic out-of-distribution samples are sampled from the low-likelihood region of the Gaussian distribution; the synthetic out-of-distribution samples are located in the feature boundary region between the in-distribution data and the true out-of-distribution data, and are used to construct a more compact decision boundary; the sampling formula is as follows:

[0054]

[0055] Among them, V k represents the synthetic out-of-distribution sample sampled from the Gaussian distribution of category k; m represents the dimension of the feature space; and denote the empirical class mean and covariance of the training samples respectively; v k represents the outlier sample sampled from the kth class; ε is a small constant.

[0056] Compared with the prior art, the beneficial technical effect of the present invention is that the present invention can generate out-of-distribution samples based on in-distribution samples through a feature extraction network, and obtain an out-of-distribution fault diagnosis network, which can accurately determine whether the aircraft engine air path fault is an in-distribution fault, and avoid identifying unknown faults as known faults, thereby causing untimely maintenance and damage to the engine health. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] 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 any creative work.

[0058] Figure 1 This is a flow chart of an aircraft engine gas path fault diagnosis method considering out-of-distribution faults provided by an embodiment of the present invention.

[0059] Figure 2 It is a schematic diagram of the fusion of the two methods of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0061] The present invention proposes a method for diagnosing air path faults of aircraft engines taking into account external faults. Figure 1 As shown, the method includes the following steps:

[0062] Step S1: Data Collection: During multiple flight phases or software simulations of an aircraft engine, multiple different types of sensors are used to collect engine operating data as sample data at a sampling frequency within a specified range. The operating data includes various cross-sectional state monitoring parameters such as temperature, pressure, and speed. Specifically, the data includes: high-pressure rotor speed, low-pressure rotor speed, high-pressure compressor pressure, low-pressure compressor pressure, high-pressure compressor temperature, low-pressure compressor temperature, high-pressure turbine temperature, low-pressure turbine temperature, high-pressure turbine pressure, and low-pressure turbine pressure. The collected sample data is divided into normal data samples and fault data samples, and the fault type is also marked. Each sample consists of operating data for a period of time.

[0063] Step S2: pre-processing the sample data collected in step S1 to obtain pre-processed sample data.

[0064] Step S21: clear incomplete data and remove outliers (data points that are significantly different from other data points in the data set or deviate from the normal pattern) to obtain cleaned sample data.

[0065] Step S22: normalize the cleaned sample data. The normalization method is as follows:

[0066]

[0067] Among them, x is the sample data after cleaning; x' is the normalized data; x min is the minimum value of the cleaned sample data; x max is the maximum value of the cleaned sample data.

[0068] Step S3: For the sample data preprocessed in step S2, the corresponding fault type (including normal state) of each sample data is used as a label. Samples of a specific fault type are excluded from all sample data and used as the subsequent true out-of-distribution samples. The remaining data, including normal samples and other fault samples, constitute the in-distribution samples. The in-distribution samples are divided into training set, validation set, and test set in a ratio of 7:2:1, and the excluded fault samples are retained as the true out-of-distribution data for model evaluation.

[0069] Step S4: Establish an MLP (Multilayer Perception) feature extraction model Net1, construct a loss function of the MLP feature extraction model Net1, and obtain a trained MLP feature extraction model Net1 through training; train an out-of-distribution fault diagnosis model Net2 based on the MLP feature extraction model Net1; retain the output of the second hidden layer of the MLP feature extraction model Net1 to obtain the projection of the real data in the feature space, and then use Gaussian distribution to fit the projection of each type of data; sample the fitted data distribution to obtain synthetic out-of-distribution data; use an energy function to estimate the synthetic out-of-distribution data and in-distribution data, determine the uncertainty threshold, so that the score of the synthetic out-of-distribution data is positive and the score of the in-distribution data is negative, and obtain a trained out-of-distribution fault diagnosis model Net2.

[0070] In more detail, Figure 1-Figure 2 As shown, step S4 includes the following sub-steps:

[0071] Step S41: Build an MLP feature extraction model Net1, which consists of an input layer, a first hidden layer, a second hidden layer, an output layer, and a classifier. The input of the MLP feature extraction model Net1 is a sample within the distribution, and the output is a predicted fault type label.

[0072] The samples in the distribution are flattened to obtain the input vector x, which is transmitted to the first hidden layer through the input layer. The input-output conversion relationship of the first hidden layer is as follows:

[0073]

[0074] in, represents the input of the jth neuron in the first hidden layer; x i Represents the i-th component of the input layer input vector; is the network weight of the first hidden layer, which represents the connection weight between the j-th neuron of the first hidden layer and the i-th component of the input vector of the input layer; I represents the dimension of the input layer; represents the output of the jth neuron in the first hidden layer; ReLU represents the ReLU activation function.

[0075] The output of the first hidden layer is transmitted to the second hidden layer, and the input-output conversion relationship of the second hidden layer is as follows:

[0076]

[0077] in, represents the input of the jth neuron in the second hidden layer; is the network weight of the second hidden layer, which represents the connection weight between the jth neuron in the second hidden layer and the ith neuron in the first hidden layer; M represents the number of neurons in the first hidden layer; represents the output of the jth neuron in the second hidden layer.

[0078] The output of the second hidden layer is transmitted to the output layer. The input-output conversion relationship of the output layer is as follows:

[0079]

[0080] in, is the network weight of the output layer, which represents the connection weight between the kth neuron in the output layer and the jth neuron in the second hidden layer, y k is the output of the kth neuron in the output layer; J is the number of neurons in the second hidden layer.

[0081] The output of the output layer is transmitted to the classifier for normalization. The normalization of the classifier is achieved through the Softmax function, which is defined as follows:

[0082]

[0083] Among them, Softmax(y k ) is the value of the k-th neuron output in the output layer after normalization by the classifier; K is the number of nodes in the output layer, which also represents the total number of output categories.

[0084] Step S42: constructing the loss function of the MLP feature extraction model Net1 using the mean square error.

[0085] The loss function of the MLP feature extraction model Net1 is calculated using mean square error, which is expressed as follows:

[0086]

[0087] Among them, L represents the Net1 loss function of the MLP feature extraction model, K represents the total number of output categories; t k Indicates the value of the kth position after the true category label t is converted into a vector.

[0088] Step S43: Train the MLP feature extraction model Net1, calculate the gradient of the loss function with respect to the model parameters to guide the update of the weights, and obtain the trained MLP feature extraction model Net1.

[0089] Loss function on the output layer network weights The gradient of is:

[0090]

[0091] in, represents the Hadamard product.

[0092] The loss function affects the weights of the second hidden layer network The gradient of is:

[0093]

[0094] The loss function is the weight of the first hidden layer network The gradient of is:

[0095]

[0096] If the learning rate is η, then the first hidden layer network weights The parameter update formula is:

[0097]

[0098] in, Represents the updated first hidden layer network weights

[0099] Then the second hidden layer network weights The parameter update formula is:

[0100]

[0101] in, Represents the updated second hidden layer network weights.

[0102] For the output layer network weights The parameter update formula is:

[0103]

[0104] in, Represents the updated output layer network weights.

[0105] When training the MLP feature extraction model Net1, a set of model hyperparameters is predefined, including the number of neurons in each hidden layer, activation function, learning rate, and optimizer type. The input is passed layer by layer through a weighted summation of the network. A loss function is then calculated using the output of the MLP feature extraction model Net1 and the true class labels. Starting from the output layer, the gradient of the loss function with respect to the model parameters is calculated to guide the weight update. This process is repeated until the loss function no longer decreases. A grid search is used to obtain the model's network weights based on the training set. Hyperparameters are determined based on the model's validation set accuracy. Once the hyperparameters are determined, the training and validation sets are combined and retrained. The trained MLP feature extraction model Net1 is obtained when model performance stabilizes.

[0106] The generalization performance of the model is evaluated based on indicators such as the accuracy and recall rate of the model test set. The formulas for the accuracy P and recall rate Rc are as follows:

[0107]

[0108] Among them, TP is the number of positive samples accurately predicted as positive samples; FP is the number of negative samples incorrectly predicted as positive samples; FN is the number of positive samples incorrectly predicted as negative samples.

[0109] After the training is completed, the trained MLP feature extraction model Net1 is obtained. The subsequent steps continue to train on the basis of the trained MLP feature extraction model Net1 to obtain an out-of-distribution fault diagnosis model Net2.

[0110] Step S44: Use the trained MLP feature extraction model Net1 to extract features from the samples in the distribution, export the output of the second hidden layer, obtain the projection of the real data in the feature space, and use Gaussian distribution to fit the projection of each type of data.

[0111] The distribution of samples within the distribution in the low-dimensional feature space can be fitted with a Gaussian distribution, and the formula is as follows:

[0112] p θ (h(x)|t=k)=N(μ k ,∑)where pθ(h(x)|t=k) represents the distribution of samples in the low-dimensional feature space; h(x)∈R m It is the representation of the input vector x in the low-dimensional feature space, that is, the output of the second hidden layer, m represents the dimension of the feature space; t represents the true category label of the sample; μ k is the Gaussian mean of fault categories k∈{1,2,…,K}, where K represents the total number of fault categories; ∑ is the bound covariance matrix; N(μ k ,∑) represents μ k The mean and ∑ are Gaussian distributions with covariance.

[0113] To estimate the parameter μ of the Gaussian distribution k and ∑, calculate the training samples The empirical average and covariance N is the total number of samples. For the k-th sample, the following formula is satisfied:

[0114]

[0115] Among them, x a represents the ath input sample; y a represents the true label of the a-th sample; h(x a) represents the sample x a After the trained MLP feature extraction model Net1, the feature vector representation of the second hidden layer; N k is the number of samples in the kth class; T represents the transpose.

[0116] Step S45: sampling the distribution edge data of the feature space based on the fitted Gaussian distribution to obtain synthetic out-of-distribution samples.

[0117] Edge data can be regarded as out-of-distribution samples, which can help distinguish out-of-distribution samples from in-distribution samples and avoid security accidents caused by misjudgment in verification.

[0118] Synthetic out-of-distribution samples are obtained by sampling from the low-likelihood region of the Gaussian distribution. These synthetic out-of-distribution samples are located in the characteristic boundary region between the in-distribution data and the true out-of-distribution data, and are used to build a more compact decision boundary, thereby improving the model's ability to distinguish out-of-distribution faults. The sampling formula is as follows:

[0119] Among them, V k represents the synthetic out-of-distribution sample sampled from the Gaussian distribution of category k; m represents the dimension of the feature space; T represents the transpose; v k represents the outlier samples sampled from the kth class, which are in the likelihood-based subset, v k The mean is The covariance matrix is The multivariate normal distribution of ; ε is a constant value small enough to ensure that the sampled outliers are close enough to the class boundary.

[0120] Step S46: Evaluate the synthesized out-of-distribution samples and in-distribution samples through the energy function, construct the energy uncertainty loss function, obtain the optimal training parameters through training, calculate the uncertainty, and obtain a suitable uncertainty threshold so that the in-distribution samples have negative energy and the synthesized out-of-distribution samples have positive energy, and obtain the trained out-of-distribution fault diagnosis model Net2.

[0121] The energy function is as follows:

[0122]

[0123] Where E(x;θ) is the energy function; x is the input vector; θ is the trainable parameter; K represents the total number of fault categories; f k (x;θ) is the score of the kth class, calculated as follows:

[0124]

[0125] Construct the energy uncertainty loss function, which is expressed as follows:

[0126]

[0127] in, represents the energy uncertainty loss function; represents the mathematical expectation operation of a sample v sampled from the in-distribution sample set V; E(v;θ) represents the estimated value of the energy function for the sample v; φ(·) is a nonlinear MLP function used to distinguish in-distribution samples from synthetic out-of-distribution samples; It represents the mathematical expectation operation on the sample x sampled from the sample set D outside the synthetic distribution.

[0128] Through the energy uncertainty loss function Get the optimal training parameter θ * , according to the obtained optimal training parameters θ * , uncertainty is calculated using in-distribution samples and synthetic out-of-distribution samples, and the uncertainty threshold δ is obtained as follows:

[0129]

[0130] Among them, x * represents the input containing samples in the distribution and samples out of the synthetic distribution; G(x * ) represents x * The uncertainty value of E(x * θ * ) represents the energy function; the input parameter of the energy function is x * and the optimal training parameters θ * The uncertainty threshold makes the in-distribution samples have negative energy and the synthetic out-distribution samples have positive energy. The uncertainty threshold δ is adjusted according to the recognition of the in-distribution samples and the synthetic out-distribution samples. The uncertainty threshold is calculated for the in-distribution samples and the synthetic out-distribution samples respectively, and then an uncertainty threshold δ that makes the out-distribution fault diagnosis accuracy the highest is selected. The training is completed and the trained out-distribution fault diagnosis model Net2 is obtained.

[0131] Step S5: Use the test set and real out-of-distribution data to test the trained out-of-distribution fault diagnosis model Net2 to obtain the identification results of out-of-distribution faults or in-distribution faults.

[0132] For better understanding and implementation, specific embodiments are given below using simulation data in combination with the accompanying drawings to illustrate in detail the method used in the present invention.

[0133] The fault types are shown in Table 1:

[0134] Table 1 Fault type table

[0135] Gas circuit failure mode describe <![CDATA[Low-pressure compressor erosion m1]]> Decreased low-pressure compressor efficiency <![CDATA[Low-pressure compressor fouling m2]]> The mass flow rate of the low-pressure compressor decreases <![CDATA[High-pressure compressor erosion m3]]> High-pressure compressor efficiency decreases <![CDATA[High-pressure compressor fouling m4]]> High pressure compressor mass flow rate decreases <![CDATA[High-pressure turbine erosion m5]]> Reduced high-pressure turbine efficiency <![CDATA[High-pressure turbine fouling m6]]> High pressure turbine mass flow rate decreases <![CDATA[Low-pressure turbine erosion m7]]> Reduced low-pressure turbine efficiency <![CDATA[Low-pressure turbine fouling m8]]> Low-pressure turbine mass flow rate decreases

[0136] First, the collected data is cleaned of outliers and normalized.

[0137] On this basis, we build the MLP feature extraction model Net1, taking m8 as out-of-distribution faults and the others as in-distribution faults. The input is the in-distribution data and the fault type label corresponding to the in-distribution data, and the output is the predicted label. The MLP parameters are set as follows: the number of hidden layers is 2, the first hidden layer has 120 neurons, the second hidden layer has 84 neurons, the ReLU function is selected as the activation function, the optimization algorithm is Adam, the learning rate is 0.001, and the ReLU function is calculated as follows:

[0138] f(x)=max(0,x)

[0139] On this basis, Gaussian distribution fitting is performed according to step S42, and then sampling is performed, and the constant ε used for limitation is 10 -4 , faults m1 to m7 are regarded as in-distribution faults, m8 as out-distribution faults, and the uncertainty threshold δ is set to 5%. The specific diagnosis results are shown in Table 2. The model can effectively identify out-distribution faults, which helps to avoid identifying out-distribution faults as in-distribution faults, thereby avoiding taking wrong maintenance measures, delaying the maintenance window and causing serious safety accidents.

[0140] Table 2 Out-of-distribution fault diagnosis results

[0141] Fault type Accuracy (%) Distributed failure 100 Out-of-distribution fault 100

[0142] This paper uses an MLP feature extraction model to extract the projection of samples within a distribution into feature space. A Gaussian distribution is then fitted, and low-likelihood regions of the distribution are sampled. Finally, an energy function is used to calculate uncertainty to distinguish between ID and OOD samples. The model is trained using simulation data, using 10 measurable flight parameters as input. Results demonstrate that the proposed model is highly accurate and stable, effectively distinguishing between ID and OOD samples.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A method for diagnosing air path faults in an aircraft engine considering out-of-distribution faults, characterized in that: The method comprises the following steps: Step S1: using multiple sensors of different types to collect engine operating data as sample data; Step S2: preprocessing the sample data collected in step S1 to obtain preprocessed sample data; Step S3: For the sample data pre-processed in step S2, the fault type corresponding to each sample data is used as a label; Exclude one type of fault sample from all sample data and use it as the subsequent true out-of-distribution samples. The remaining data, including normal samples and other fault samples, constitute the in-distribution samples. Divide the in-distribution samples into training set, validation set, and test set in a ratio of 7:2:1, and retain the excluded fault samples as the true out-of-distribution data for model evaluation. Step S41: Building an MLP feature extraction model Net1, which consists of an input layer, a first hidden layer, a second hidden layer, an output layer, and a classifier; Step S42: constructing the loss function of the MLP feature extraction model Net1 using the mean square error; Step S43: training the MLP feature extraction model Net1, calculating the gradient of the loss function with respect to the model parameters to guide the update of the weights, and obtaining the trained MLP feature extraction model Net1; Step S44: Use the trained MLP feature extraction model Net1 to extract features from the samples in the distribution, export the output of the second hidden layer, obtain the projection of the real data in the feature space, and use Gaussian distribution to fit the projection of each type of data; Step S45: sampling the distribution edge data of the feature space based on the fitted Gaussian distribution to obtain synthetic distribution out-of-distribution samples; Step S46: Evaluate the synthesized out-of-distribution samples and in-distribution samples through the energy function, construct an energy uncertainty loss function, obtain the optimal training parameters through training, calculate the uncertainty, and obtain a suitable uncertainty threshold so that the in-distribution samples have negative energy and the synthesized out-of-distribution samples have positive energy, thereby obtaining the trained out-of-distribution fault diagnosis model Net2; The energy function is as follows: Where E(x;θ) is the energy function; x is the input vector; θ is the trainable parameter; K represents the total number of fault categories; f k (x;θ) is the score of the kth class, calculated as follows: in, is the network weight of the output layer, T represents transpose; Construct the energy uncertainty loss function, which is expressed as follows: in, represents the energy uncertainty loss function; represents the mathematical expectation operation of a sample v sampled from the in-distribution sample set V; E(v;θ) represents the estimated value of the energy function for the sample v; φ(·) is a nonlinear MLP function used to distinguish in-distribution samples from synthetic out-of-distribution samples; It represents the mathematical expectation operation of the sample x sampled from the sample set D outside the synthetic distribution; Through the energy uncertainty loss function Get the optimal training parameter θ * , according to the obtained optimal training parameters θ * , using the in-distribution samples and the synthetic out-of-distribution samples to calculate the uncertainty, the uncertainty threshold δ is calculated as follows: Among them, x * represents the input containing samples in the distribution and samples out of the synthetic distribution; G(x * ) represents x * The uncertainty value of E(x * θ * ) represents the energy function; the input parameter of the energy function is x * and the optimal training parameters θ * The uncertainty threshold makes the samples within the distribution have negative energy and the samples outside the synthetic distribution have positive energy. The uncertainty threshold δ is adjusted according to the recognition of the samples within the distribution and the synthetic samples outside the distribution. The uncertainty threshold is calculated for the samples within the distribution and the synthetic samples outside the distribution respectively, and then an uncertainty threshold δ that makes the out-of-distribution fault diagnosis accuracy the highest is selected. The training is then completed, and the trained out-of-distribution fault diagnosis model Net2 is obtained. Step S5: Use the test set and real out-of-distribution data to test the trained out-of-distribution fault diagnosis model Net2 to obtain the identification results of out-of-distribution faults or in-distribution faults.

2. The method for diagnosing aero-engine gas path faults considering out-of-distribution faults according to claim 1, characterized in that: The step S2 comprises: Step S21: clear incomplete data and remove outliers to obtain cleaned sample data; Step S22: normalize the cleaned sample data. The normalization method is as follows: Among them, x is the cleaned sample data; x′ is the normalized data; x min is the minimum value of the cleaned sample data; x max is the maximum value of the cleaned sample data.

3. The method for diagnosing air path faults of an aircraft engine considering out-of-distribution faults according to claim 2, characterized in that: Step S41 includes the following steps: The samples in the distribution are flattened to obtain the input vector x, which is transmitted to the first hidden layer through the input layer. The input-output conversion relationship of the first hidden layer is as follows: in, represents the input of the jth neuron in the first hidden layer; x i Represents the i-th component of the input layer input vector; is the network weight of the first hidden layer, which represents the connection weight between the j-th neuron of the first hidden layer and the i-th component of the input vector of the input layer; I represents the dimension of the input layer; represents the output of the jth neuron in the first hidden layer; ReLU represents the ReLU activation function; The output of the first hidden layer is transmitted to the second hidden layer, and the input-output conversion relationship of the second hidden layer is as follows: in, represents the input of the jth neuron in the second hidden layer; is the network weight of the second hidden layer, which represents the connection weight between the jth neuron in the second hidden layer and the ith neuron in the first hidden layer; M represents the number of neurons in the first hidden layer; represents the output of the second hidden layer; The output of the second hidden layer is transmitted to the output layer. The input-output conversion relationship of the output layer is as follows: Among them, y k is the output of the kth neuron in the output layer; J represents the number of neurons in the second hidden layer; The output of the output layer is transmitted to the classifier for normalization. The normalization of the classifier is achieved through the Softmax function, which is defined as follows: Among them, Softmax(y k ) is the value of the k-th neuron output in the output layer after normalization by the classifier; K is the number of nodes in the output layer, which also represents the total number of output categories.

4. The method for diagnosing air path faults of an aircraft engine considering out-of-distribution faults according to claim 3, characterized in that: In step S42, The loss function of the MLP feature extraction model Net1 is calculated using mean square error, which is expressed as follows: Among them, L represents the Net1 loss function of the MLP feature extraction model, K represents the total number of output categories; t k Indicates the value of the kth position after the true category label t is converted into a vector.

5. The method for diagnosing air path faults of an aircraft engine considering out-of-distribution faults according to claim 4, characterized in that: In step S44, The distribution of samples within the distribution in the low-dimensional feature space is fitted with a Gaussian distribution, and the formula is as follows: p θ (h(x)|t k =k)=N(μ k ,∑) Among them, p θ (h(x)|t k =k) ​​represents the distribution of samples in the low-dimensional feature space; h(x)∈R m is the representation of the input vector x in the low-dimensional feature space, that is, the output of the second hidden layer, m represents the dimension of the feature space; t k represents the true category label of the sample; μ k is the Gaussian mean of fault categories k∈{1,2,…,K}, where K represents the total number of fault categories; ∑ is the bound covariance matrix; N(μ k ,∑) represents μ k The mean and ∑ are Gaussian distributions with covariance.

6. The method for diagnosing air path faults of an aircraft engine considering out-of-distribution faults according to claim 5, characterized in that: In step S45, Synthetic out-of-distribution samples are obtained by sampling from the low-likelihood region of the Gaussian distribution. The synthetic out-of-distribution samples are located in the feature boundary region between the in-distribution data and the true out-of-distribution data, and are used to construct a more compact decision boundary. The sampling formula is as follows: Among them, V k represents the synthetic out-of-distribution sample sampled from the Gaussian distribution of category k; m represents the dimension of the feature space; and denote the empirical class mean and covariance of the training samples respectively; v k represents the outlier sample sampled from the kth class; ε is a small constant.

Images