Aero-engine gas path fault diagnosis method considering distributed external faults

Abstract

The invention provides an aero-engine gas path fault diagnosis method considering distributed external faults, and belongs to the field of aero-engine fault diagnosis. The method comprises the following steps: collecting data and preprocessing the data; an MLP feature extraction model Net1 is established, and the trained MLP feature extraction model Net1 is obtained through training; training an out-of-distribution fault diagnosis model Net2 on the basis of the MLP feature extraction model Net1; reserving the output of a second hidden layer of the MLP feature extraction model Net1 to obtain the projection of real data in a feature space, and fitting the projection of each type of data by adopting Gaussian distribution; sampling the fitted data distribution to obtain data outside the synthetic distribution; and estimating the synthesized out-of-distribution data and in-distribution data by adopting an energy function, determining an uncertainty threshold, and obtaining a trained out-of-distribution fault diagnosis model Net2. According to the method, whether the aero-engine gas path fault is the distributed internal fault can be accurately judged, and the unknown fault is prevented from being recognized as the known fault.

Description

Technical Field

[0001] The present invention belongs to the field of aeroengine fault diagnosis, and particularly relates to an aeroengine gas path fault diagnosis method considering out-of-distribution faults. Background Art

[0002] Aeroengines operate in a harsh environment, often under variable and high-load operating conditions, with diverse fault modes. As a result, their maintenance costs account for more than 40% of the total aircraft maintenance costs. Among all the fault modes of aeroengines, gas path system faults account for more than 90%, and their repair costs reach 60% of the total engine repair costs. The fault modes of the gas path system generally include erosion, fouling, corrosion, interblade wear, foreign object damage, etc. Accurate gas path system fault diagnosis of the engine can achieve rapid fault location, effectively reduce maintenance costs, and avoid significant economic losses and safety accidents at the same time.

[0003] Currently, data-driven methods for aeroengine gas path faults generally assume that the training set and test set data are independently and identically distributed. However, when this assumption does not hold, that is, when facing fault types not present in the training set, the neural network will give a definite result and identify it as a fault type in the training set. Hendrycks et al. published a paper titled "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks" at the 2017 International Conference on Learning Representations (ICLR), proposing the concept of out-of-distribution (OOD) fault detection, which detects objects of unknown classes while ensuring the original task. In recent years, OOD detection has received extensive attention from scholars and has been deeply studied in fields such as computer vision, natural language processing, and fault diagnosis, with the aim of improving the model's robustness to unknown samples based on the original deep learning task. Supervised OOD detection uses labeled in-distribution (ID) samples 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", proposing 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 class and applying it to the monitoring of concurrent faults and OOD faults of high-speed train bogies. However, since real OOD samples are difficult to obtain, generating OOD samples for training is an intuitive method.In 2019, Verneker et al. published a paper titled "Out-of-Distribution Detection in Classifiers via Generation" at the Conference and Workshop on Neural Information Processing Systems (NeurIPS). They used conditional variational autoencoders to generate two types of out-of-distribution (OOD) samples, namely those outside the sample manifold and on the sample manifold, effectively covering the entire distribution boundary. In 2021, Gao et al. published a paper titled "Uncertainty Enhanced Attention for OOD Detection" at the International Joint Conference on Neural Networks (IJCNN). They used generative adversarial networks to generate OOD samples and introduced an attention mechanism to improve detection accuracy. Although supervised methods have high accuracy, there is a problem of difficulty in obtaining high-quality labeled OOD samples. Summary of the Invention

[0004] The present invention proposes an aeroengine gas path fault diagnosis method considering out-of-distribution faults. First, feature extraction is performed, and then the feature distribution is fitted. The feature extraction method and Gaussian distribution fitting are fused. The projection of in-distribution data in the feature space is obtained through an MLP feature extraction network. Then, the feature distributions of various fault types are fitted using Gaussian distribution, and the data at the distribution edge is sampled to obtain synthetic out-of-distribution data for training the uncertainty threshold to identify in-distribution (ID) samples and OOD samples. This is to solve the problem of difficulty in obtaining high-quality labeled OOD samples in the prior art.

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

[0006] An aeroengine gas path fault diagnosis method considering out-of-distribution faults, the method comprising the following steps:

[0007] Step S1: Collect engine operation data as sample data using multiple different types of sensors;

[0008] Step S2: Preprocess the sample data collected in Step S1 to obtain preprocessed sample data;

[0009] Step S3: For the sample data preprocessed in step S2, the fault type (including normal state) corresponding to each sample data is used as a label; samples of a specific fault type are excluded from all sample data and used as 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 a training set, a validation set, and a 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 S4: Establish an MLP 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;

[0011] 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.

[0012] Furthermore, the step S2 includes:

[0013] Step S21: clear incomplete data and remove outliers to obtain cleaned sample data;

[0014] Step S22: normalize the cleaned sample data. The normalization method is as follows:

[0015]

[0016] in, is the cleaned sample data; is the normalized data; is the minimum value of the cleaned sample data; is the maximum value of the cleaned sample data.

[0017] Furthermore, step S4 includes the following steps:

[0018] 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;

[0019] Step S42: Construct the loss function of the MLP feature extraction model Net1 using the mean square error.

[0020] 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.

[0021] Step S44: Use the trained MLP feature extraction model Net1 to extract features from in-distribution samples, export the output of the second hidden layer to obtain the projection of the real data in the feature space, and fit the projection of each class of data using a Gaussian distribution.

[0022] Step S45: Sample the distribution boundary data in the feature space based on the fitted Gaussian distribution to obtain synthetic out-of-distribution samples.

[0023] Step S46: Evaluate the synthetic out-of-distribution samples and in-distribution samples through an energy function, construct an energy uncertainty loss function, obtain the optimal training parameters through training, calculate the uncertainty, obtain a suitable uncertainty threshold, make the in-distribution samples have negative energy, and the synthetic out-of-distribution samples have positive energy, and obtain the trained out-of-distribution fault diagnosis model Net2.

[0024] Further, step S41 includes the following steps:

[0025] Flatten the in-distribution samples to obtain an input vector , and the input vector is transmitted through the input layer to the first hidden layer. The input-output conversion relationship of the first hidden layer is as follows:

[0026]

[0027] where represents the input of the th neuron in the first hidden layer; represents the th component of the input vector of the input layer; is the network weight of the first hidden layer, representing the connection weight between the th neuron in the first hidden layer and the th component of the input vector of the input layer; represents the dimension of the input layer; represents the output of the th neuron in the first hidden layer; represents the ReLU activation function;

[0028] The output of the first hidden layer is transmitted to the second hidden layer. The input-output conversion relationship of the second hidden layer is as follows:

[0029]

[0030] Among them, represents the input of the th neuron in the second hidden layer; is the network weight of the second hidden layer, representing the connection weight between the th neuron in the second hidden layer and the th neuron in the first hidden layer; represents the number of neurons in the first hidden layer; represents the output of the second hidden layer;

[0031] The output of the second hidden layer is transmitted to the output layer, and the input-output conversion relationship of the output layer is as follows:

[0032]

[0033] Among them, is the network weight of the output layer, representing the connection weight between the th neuron in the output layer and the th neuron in the second hidden layer, is the output of the th neuron in the output layer; represents the number of neurons in the second hidden layer;

[0034] The output of the output layer is transmitted to the classifier for normalization. The normalization of the classifier is implemented through the Softmax function. The Softmax function is defined as follows:

[0035]

[0036] Among them, is the value of the output of the th neuron in the output layer after normalization by the classifier; is the number of output layer nodes, and also represents the total number of output categories.

[0037] Furthermore, in step S42,

[0038] The loss function of the MLP feature extraction model Net1 is calculated using the mean squared error and is expressed as follows:

[0039]

[0040] Among them, represents the loss function of the MLP feature extraction model Net1, represents the total number of output categories; represents the true category label after being converted into a vector, the The value of a position.

[0041] Furthermore, in step S44,

[0042] The distribution of in-distribution samples in the low-dimensional feature space can be fitted with a Gaussian distribution, and the formula is as follows:

[0043] Where, Represents the distribution of in-distribution samples in the low-dimensional feature space; Is the input vector In the representation of the low-dimensional feature space, that is, the output of the second hidden layer, Represents the dimension of the feature space; Represents the true class label of the sample; Is the fault class Of the Gaussian mean, Represents the total number of fault classes; Is the bound covariance matrix; Represents as The mean, For the Gaussian distribution with covariance.

[0044] Furthermore, in step S45,

[0045] Synthesize out-of-distribution samples by sampling from the low-likelihood region of the Gaussian distribution; the synthesized out-of-distribution samples are located in the feature boundary region between in-distribution data and true out-of-distribution data, and are used to construct a more compact decision boundary, thereby enhancing the model's ability to distinguish out-of-distribution faults. The sampling formula is as follows:

[0046] Where, Represents the synthesized out-of-distribution sample sampled from the Gaussian distribution of class ; Represents the dimension of the feature space; And Respectively represent the empirical class mean and covariance of the training samples; Represents the transpose; Represents the outlier sample sampled from the th class; these samples are located in the likelihood-based subset; Is a small constant value to ensure that the sampled outlier samples are close enough to the class boundary.

[0047] Furthermore, in step S46,

[0048] The energy function is as follows:

[0049] Where, Is the energy function; Is the input vector; is a trainable parameter; represents the total number of fault categories; is the score of the

[0050] Construct an energy uncertainty loss function, which is expressed as follows:

[0051] ℒ uncertainty =  v ∼ V [ − log 1 1 + exp ( − ϕ ( E ( v ; θ ) ) ) ] +  x ∼ D [ − log exp ( − ϕ ( E ( x ; θ ) ) ) 1 + exp ( − ϕ ( E ( x ; θ ) ) ) ] where, represents the energy uncertainty loss function; represents the mathematical expectation operation on the sample sampled from the set of in-distribution samples ; represents the estimated value of the energy function for the sample ; is a non-linear MLP function used to distinguish in-distribution samples and synthetic out-of-distribution samples; represents the mathematical expectation operation on the sample sampled from the set of synthetic out-of-distribution samples ;

[0052] Obtain the optimal training parameters through the energy uncertainty loss function , and according to the obtained optimal training parameters , calculate the uncertainty using in-distribution samples and synthetic out-of-distribution samples. The uncertainty threshold is calculated as follows:

[0053]

[0054] where, represents the input containing in-distribution samples and synthetic out-of-distribution samples; represents the uncertainty value of; is the uncertainty threshold used to divide in-distribution and synthetic out-of-distribution samples; represents the energy function; the input parameters of the energy function are and the optimal training parameters . The uncertainty threshold makes in-distribution samples have negative energy and synthetic out-of-distribution samples have positive energy. Adjust the uncertainty threshold according to the recognition situations of in-distribution samples and synthetic out-of-distribution samples; calculate the uncertainty thresholds for in-distribution samples and synthetic out-of-distribution samples respectively, and then select an uncertainty threshold that makes the out-of-distribution fault diagnosis accuracy the highest, then the training ends, and the trained out-of-distribution fault diagnosis model Net2 is obtained.

[0055] Compared with the prior art, the beneficial technical effects of the present invention are as follows: 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 judge whether an aero-engine gas path fault is an in-distribution fault, avoiding misidentifying an unknown fault as a known fault, thereby preventing untimely maintenance and damage to the engine's health condition. Description of the Drawings

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0057] Figure 1 is a flowchart of a method for diagnosing aero-engine gas path faults considering out-of-distribution faults provided by an embodiment of the present invention.

[0058] Figure 2 is a schematic diagram of the fusion of two methods of the present invention. Detailed Embodiments

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0060] The present invention proposes a method for diagnosing aero-engine gas path faults considering out-of-distribution faults, as Figure 1 shown, the method includes the following steps:

[0061] Step S1: Data acquisition: In multiple flight stages of an aero-engine or in software simulation, multiple different types of sensors are used to collect engine operation data as sample data at a sampling frequency within a specified range. The operation data includes cross-section state monitoring parameters such as temperature, pressure, and rotational speed, specifically including: 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 their fault types are marked at the same time. Each sample consists of operation data for a period of time.

[0062] Step S2: Preprocess the sample data collected in step S1 to obtain preprocessed sample data.

[0063] Step S21: Clear incomplete data and eliminate outliers (data points in the dataset that are significantly different or deviate from the normal pattern compared to other data points) to obtain the cleaned sample data.

[0064] Step S22: Normalize the cleaned sample data. The normalization method is as follows:

[0065]

[0066] where, is the cleaned sample data; is the normalized data; is the minimum value of the cleaned sample data; is the maximum value of the cleaned sample data.

[0067] Step S3: For the sample data preprocessed in Step S2, use the corresponding fault types (including the normal state) of each sample data as labels. Exclude the samples of a specific fault type from all the sample data and use them as the subsequent out-of-distribution samples in reality. The data including normal samples and other fault samples form the in-distribution samples. Divide the in-distribution samples into a training set, a validation set, and a test set according to the ratio of 7:2:1, and retain the excluded fault samples as the real out-of-distribution data for model evaluation.

[0068] Step S4: Establish an MLP (Multilayer Perception) feature extraction model Net1, construct the loss function of the MLP feature extraction model Net1, and obtain the 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 the Gaussian distribution to fit the projection of each class 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 the in-distribution data, determine the uncertainty threshold, so that the synthetic out-of-distribution data gets a positive score and the in-distribution data gets a negative score, and obtain the trained out-of-distribution fault diagnosis model Net2.

[0069] More specifically, as Figures 1 - 2 shown, Step S4 includes the following sub-steps:

[0070] Step S41: Build the MLP feature extraction model Net1. The MLP feature extraction model Net1 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 the in-distribution samples, and the output is the predicted fault type label.

[0071] Flatten the in-distribution samples to obtain the input vector , and the input vector is transmitted to the first hidden layer through the input layer. The input-output conversion relationship of the first hidden layer is as follows:

[0072]

[0073] where, represents the input of the -th neuron in the first hidden layer; represents the -th component of the input vector of the input layer; is the network weight of the first hidden layer, representing the connection weight between the -th neuron in the first hidden layer and the -th component of the input vector of the input layer; represents the dimension of the input layer; represents the output of the -th neuron in the first hidden layer; represents the ReLU activation function.

[0074] The output of the first hidden layer is transmitted to the second hidden layer. The input-output conversion relationship of the second hidden layer is as follows:

[0075]

[0076] where, represents the input of the -th neuron in the second hidden layer; is the network weight of the second hidden layer, representing the connection weight between the -th neuron in the second hidden layer and the -th neuron in the first hidden layer; represents the number of neurons in the first hidden layer; represents the output of the -th neuron in the second hidden layer.

[0077] 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:

[0078]

[0079] where, is the network weight of the output layer, representing the The connection weight between the neuron and the neuron in the second hidden layer, is the output of the neuron in the output layer;

[0080] The output of the output layer is transmitted to the classifier for normalization, and the normalization of the classifier is achieved through the Softmax function. The Softmax function is defined as follows:

[0081]

[0082] where, is the value after normalization of the output of the neuron in the output layer through the classifier; is the number of nodes in the output layer, and also represents the total number of output categories.

[0083] Step S42: Construct the loss function of the MLP feature extraction model Net1 using the mean squared error.

[0084] The loss function of the MLP feature extraction model Net1 is calculated using the mean squared error and is expressed as follows:

[0085]

[0086] where, represents the loss function of the MLP feature extraction model Net1, represents the total number of output categories; represents the true class label after being converted into a vector, the value at the th position.

[0087] 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.

[0088] The gradient of the loss function with respect to the output layer network weight is:

[0089]

[0090] where, represents the Hadamard product.

[0091] The gradient of the loss function with respect to the second hidden layer network weight is:

[0092]

[0093] The gradient of the loss function with respect to the weights of the first hidden layer is as follows:

[0094]

[0095] If the learning rate is , then the parameter update formula for the weights of the first hidden layer is:

[0096]

[0097] where represents the updated weights of the first hidden layer

[0098] Then the parameter update formula for the weights of the second hidden layer is:

[0099]

[0100] where represents the updated weights of the second hidden layer.

[0101] For the weights of the output layer the parameter update formula is:

[0102]

[0103] where represents the updated weights of the output layer.

[0104] When training the MLP feature extraction model Net1, the set of model hyperparameters is pre - divided, including the number of neurons in each hidden layer, the activation function, the learning rate, and the type of optimizer. The input is passed layer - by - layer through the network weighted sum to the output, and then the loss function is calculated based on the output of the MLP feature extraction model Net1 and the true class labels. The gradient of the loss function with respect to the model parameters is calculated starting from the output layer to guide the weight update. This process is repeated until the loss function no longer decreases. Grid search is used to obtain the network weights of the model based on the training set, and the hyperparameters are determined according to the accuracy of the model validation set; after determining the hyperparameters, the training set and the validation set are merged again for training, and the trained MLP feature extraction model Net1 is obtained when the model performance no longer changes.

[0105] The generalization performance of the model is evaluated according to indicators such as the accuracy and recall rate of the model test set. The accuracy and recall rate formulas are as follows:

[0106]

[0107]

[0108] Among them, is the number of positive samples accurately predicted as positive samples; is the number of negative samples mispredicted as positive samples; is the number of positive samples mispredicted as negative samples.

[0109] After the training is completed, the trained MLP feature extraction model Net1 is obtained. In the following steps, an out-of-distribution fault diagnosis model Net2 is continuously trained based on the trained MLP feature extraction model Net1.

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

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

[0112] Among them, represents the distribution of in-distribution samples in the low-dimensional feature space; is the input vector in the representation of the low-dimensional feature space, that is, the output of the second hidden layer, represents the dimension of the feature space; represents the true class label of the sample; is the fault class of the Gaussian mean value, represents the total number of fault classes; is the bound covariance matrix; represents as the mean, as the covariance of the Gaussian distribution.

[0113] In order to estimate the parameters and of the Gaussian distribution, calculate the empirical class mean value and covariance of the training samples , is the total number of samples, for the th class of samples, the following formula is satisfied:

[0114]

[0115] Among them, represents the th input sample; represents the The true label of a sample; Denote the sample The feature vector representation of the sample after passing through the trained MLP feature extraction model Net1 in the second hidden layer; Is the Number of samples in the class; Denote the transpose.

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

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

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

[0119] Where, Denote the out-of-distribution synthetic samples sampled from the Gaussian distribution of class ; Denote the feature space dimension; Denote the transpose; Denote the outlier samples sampled from the class, and these samples are located in the likelihood-based subset, Subject to a multivariate normal distribution with mean and covariance matrix ; Is a sufficiently small constant value to ensure that the sampled outlier samples are close enough to the class boundary.

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

[0121] The energy function is as follows:

[0122] Where, Is the energy function; Is the input vector; Is the trainable parameter; Denote the total number of fault classes; Is the The score of the class is calculated as follows:

[0123] Construct an energy uncertainty loss function, expressed as follows:

[0124] ℒ uncertainty =  v ∼ V [ − log 1 1 + exp ( − ϕ ( E ( v ; θ ) ) ) ] +  x ∼ D [ − log exp ( − ϕ ( E ( x ; θ ) ) ) 1 + exp ( − ϕ ( E ( x ; θ ) ) ) ] where represents the energy uncertainty loss function; represents the mathematical expectation operation on the samples sampled from the in-distribution sample set ; represents the estimated value of the energy function for the sample ; is a non-linear MLP function used to distinguish in-distribution samples and synthetic out-of-distribution samples; represents the mathematical expectation operation on the samples sampled from the synthetic out-of-distribution sample set ;

[0125] Obtain the optimal training parameters through the energy uncertainty loss function . According to the obtained optimal training parameters , calculate the uncertainty using in-distribution samples and synthetic out-of-distribution samples. The uncertainty threshold is obtained in the following way:

[0126]

[0127] where represents the input containing in-distribution samples and synthetic out-of-distribution samples; represents the uncertainty value of; is the uncertainty threshold used to divide in-distribution and synthetic out-of-distribution samples; represents the energy function; the input parameters of the energy function are and the optimal training parameters . The uncertainty threshold makes in-distribution samples have negative energy and synthetic out-of-distribution samples have positive energy. Adjust the uncertainty threshold according to the recognition situation of in-distribution samples and synthetic out-of-distribution samples; calculate the uncertainty threshold for in-distribution samples and synthetic out-of-distribution samples respectively, and then select an uncertainty threshold that maximizes the out-of-distribution fault diagnosis accuracy rate. Then the training ends, and the trained out-of-distribution fault diagnosis model Net2 is obtained.

[0128] Step S5: Use the test set and real out-of-distribution data to test through the trained out-of-distribution fault diagnosis model Net2 to obtain the recognition results of out-of-distribution faults or in-distribution faults.

[0129] For better understanding and implementation, the following uses simulation data combined with the attached drawings to give specific embodiments to illustrate in detail the method used in the present invention.

[0130] The fault types are shown in Table 1:

[0131] Table 1 Fault Type Table

[0132]

[0133] First, clean the collected data of outliers and normalize it.

[0134] On this basis, build the MLP feature extraction model Net1, and use As out-of-distribution faults, and the others are in-distribution faults. The input is in-distribution data and the corresponding fault type labels of in-distribution data, and the output is the predicted labels. 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 selects Adam, the learning rate is 0.001, and the ReLU function. The calculation method is as follows:

[0135]

[0136] On this basis, perform Gaussian distribution fitting according to step S42, and then perform sampling for the limiting constant , and regard the fault To the fault As in-distribution faults, As out-of-distribution faults, the uncertainty threshold Is set to 5%. The specific diagnostic results obtained are shown in Table 2. The model can effectively identify out-of-distribution faults, which helps to avoid misidentifying out-of-distribution faults as in-distribution faults, thereby avoiding taking wrong maintenance measures and delaying the maintenance window, resulting in serious safety accidents.

[0137] Table 2 Out-of-Distribution Fault Diagnosis Result Table

[0138]

[0139] The present invention uses an MLP feature extraction model to obtain the projection of in-distribution samples in the feature space, fits them using a Gaussian distribution, samples the low-likelihood regions of the distribution, and finally uses an energy function to distinguish ID samples and OOD samples by calculating uncertainty. With 10 measurable flight parameters as inputs, the model is trained using simulation data. The results show that the model has high accuracy and stability and can effectively distinguish ID samples and OOD samples.

[0140] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are all included in the protection scope of the present invention.

Claims

1. An aero-engine gas path fault diagnosis method 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 samples of a specific fault type from all sample data and use them as 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, validation, and test sets in a ratio of 7:2:1, and retain the excluded fault samples as true out-of-distribution data for model evaluation. Step S4: Establish an MLP 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; 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 aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 1, wherein, 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, is the cleaned sample data; is the normalized data; is the minimum value of the cleaned sample data; is the maximum value of the cleaned sample data.

3. The aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 1, characterized in that, Step S4 includes the following steps: 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 synthetic out-of-distribution samples and in-distribution samples through an 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 synthetic out-of-distribution samples have positive energy, thereby obtaining a trained out-of-distribution fault diagnosis model Net2.

4. The aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 3, characterized in that, Step S41 includes the following steps: Flatten the in-distribution samples to obtain an input vector , and the input vector is transmitted to the first hidden layer through the input layer. The input-output conversion relationship of the first hidden layer is as follows: Among them, represents the input of the th neuron in the first hidden layer; represents the th component of the input vector of the input layer; is the network weight of the first hidden layer, representing the connection weight between the th neuron in the first hidden layer and the th component of the input vector of the input layer; represents the dimension of the input layer; represents the output of the th neuron in the first hidden layer; 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: Among them, represents the input of the th neuron in the second hidden layer; is the network weight of the second hidden layer, representing the th neuron in the second hidden layer and the th neuron in the first hidden layer, and is the connection weight between them; 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, and the input-output conversion relationship of the output layer is as follows: Among them, is the network weight of the output layer, indicating the connection weight between the th neuron in the output layer and the th neuron in the second hidden layer; is the output of the th neuron in the output layer; 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 implemented through the Softmax function, and the Softmax function is defined as follows: Among them, is the value of the output of the -th neuron in the output layer after being normalized by the classifier; is the number of nodes in the output layer, and also represents the total number of output categories.

5. The aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 4, wherein, In step S42, The loss function of the MLP feature extraction model Net1 is calculated using the mean squared error and is expressed as follows: Among them, represents the loss function of the MLP feature extraction model Net1, represents the total number of output categories; represents the true category label after being converted into a vector, the value at the -th position.

6. The method for diagnosing a gas path fault of an aero-engine considering out-of-distribution faults according to claim 5, wherein In step S44, The distribution of in-distribution samples in the low-dimensional feature space is fitted with a Gaussian distribution, and the formula is as follows: Among them, represents the distribution of in-distribution samples in the low-dimensional feature space; is the input vector in the low-dimensional feature space, that is, the output of the second hidden layer, represents the dimension of the feature space; represents the true class label of the sample; is the fault class of the Gaussian mean, represents the total number of fault classes; is the bound covariance matrix; represents the mean, and is the Gaussian distribution with covariance.

7. The aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 6, characterized in that 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 in-distribution data and true out-of-distribution data and are used to construct a more compact decision boundary; the sampling formula is as follows: wherein, represents an out-of-sample of the synthetic distribution sampled from the Gaussian distribution of class ; represents the dimension of the feature space; and respectively represent the empirical class mean and covariance of the training samples; represents the transpose; represents an outlier sample sampled from the -th class; is a tiny constant.

8. The aero-engine gas path fault diagnosis method considering out-of-distribution faults according to claim 7, characterized in that In step S46, The energy function is as follows: in, is the energy function; is the input vector; is a trainable parameter; Indicates the total number of fault categories; It is The score of a class is calculated as follows: Construct an energy uncertainty loss function, which is expressed as follows: Among them, represents the energy uncertainty loss function; represents the mathematical expectation operation on the samples sampled from the in-distribution sample set ; represents the estimated value of the energy function for the sample ; is a non-linear MLP function used to distinguish in-distribution samples and synthetic out-of-distribution samples; represents the mathematical expectation operation on the samples sampled from the synthetic out-of-distribution sample set ; Through the energy uncertainty loss function Obtain the optimal training parameters , according to the obtained optimal training parameters , use in-distribution samples and synthetic out-of-distribution samples to calculate uncertainty, and the uncertainty threshold is calculated as follows: Among them, represents the input containing in-distribution samples and synthetic out-of-distribution samples; represents the uncertainty value of; is the uncertainty threshold, used to distinguish in-distribution samples from synthetic out-of-distribution samples; represents the energy function; the input parameters of the energy function are and the optimal training parameters ; the uncertainty threshold makes in-distribution samples have negative energy and synthetic out-of-distribution samples have positive energy. The uncertainty threshold is adjusted according to the recognition situations of in-distribution samples and synthetic out-of-distribution samples; the uncertainty thresholds are calculated separately for in-distribution samples and synthetic out-of-distribution samples, and then an uncertainty threshold that maximizes the out-of-distribution fault diagnosis accuracy is selected , then the training ends, and the trained out-of-distribution fault diagnosis model Net2 is obtained.

Images