A cross-domain fault diagnosis method for aircraft engines based on dynamic adversarial residual network

Abstract

The present invention provides an aircraft engine cross-domain fault diagnosis method based on a dynamic adversarial residual network. The method proposed by the present invention aligns the historical number distribution and the new unlabeled data distribution through dynamic adversarial learning, and introduces a residual module in the source domain label classifier to better adapt the classifier to the target domain data. The present invention overcomes the dependence of traditional data-driven methods on a large amount of identically distributed labeled data, makes better use of historical data, and can be applied to the problem of unsupervised cross-domain fault diagnosis of aircraft engines. It reduces the dependence on data annotation, is more in line with actual work scenarios, improves the ability to apply historical accumulated data, and helps data-driven methods to be better implemented in actual work environments.

Description

Technical Field

[0001] The invention belongs to the field of neural network computing, and in particular to an aero-engine cross-domain fault diagnosis method based on a dynamic adversarial residual network. Background Art

[0002] At present, the methods for diagnosing faults in the air circuit of aircraft engines mainly include two categories: model-based and data-driven methods. The traditional model-based method requires the establishment of an accurate mathematical model based on the operating conditions of the monitored engine and the knowledge of the mechanism of the equipment. However, the engine operation process contains complex aerodynamic thermodynamic processes. Due to the influence of factors such as noise in the actual operating environment, there are many uncertainties in the modeling process, and it is often difficult to obtain an accurate mathematical model. The data-driven method is completely based on engine monitoring data. Through data analysis methods such as SVM and deep neural networks, the potential laws of the measurable parameters of aircraft engines and the fault states of air circuit components are mined to achieve intelligent fault diagnosis. The data-driven method based on deep learning is currently a hot research topic in the field of rotating machinery fault diagnosis. Compared with traditional machine learning methods, deep learning methods can extract deeper feature representations of data and weaken the influence of noise in data. It is expected to build an end-to-end diagnostic model that can directly associate the original monitoring data with the health status of the equipment. In general, the method for diagnosing faults in the air circuit of aircraft engines based on machine learning methods uses the historical operation data of the engine as training data to train the model to identify the unlabeled data in the engine under the current operating state. This method assumes that the training data and test data come from the same probability distribution, and requires sufficient data to train a reliable diagnostic model. However, in actual aircraft engine gas path fault diagnosis, the following challenges are often faced:

[0003] 1) Due to changes in engine operating conditions, performance degradation, and differences in machine assembly before and after maintenance, the data originally used to build the diagnostic model often comes from a different distribution than the new data actually detected. Studies have shown that directly applying the old model will lead to a decrease in diagnostic performance due to differences in data distribution.

[0004] 2) During the operation of aircraft engines, it is not possible to frequently stop the equipment to check the health status, and aircraft engine fault data is difficult to obtain at a low cost. Therefore, it is impractical to collect a large amount of labeled sensor measurement data from the actual monitored engines to retrain the diagnostic model.

[0005] The above two problems limit the successful application of data-driven intelligent fault diagnosis methods in some unlabeled data situations. Fortunately, the emergence of transfer learning makes it possible to solve the above problems. The main idea of ​​transfer learning is to apply the knowledge learned in the old field to the new field by exploring the similarities between the source domain and the target domain in data, tasks or models. For the problem of air path fault diagnosis of aircraft engines, we can use transfer learning methods to extract domain-invariant features from the data collected under the historical working conditions of the engine to establish a fault diagnosis model that can be used for the current working state of the aircraft engine.

[0006] At present, DANN is the most representative deep learning-based transfer learning method. It draws on the idea of ​​GAN, an adversarial neural network, and introduces a domain classifier based on the feature extractor and label classifier to learn transferable features through adversarial learning. Specifically, the domain classifier is used to distinguish the source domain from the target domain, while the feature extractor attempts to learn the domain-invariant features of the two domains, and the label classifier uses the learned domain-invariant features to predict labels for samples in the two domains. Most of the effective deep learning-based transfer learning methods are improved from DANN. Although these methods have achieved good results in the field of fault diagnosis, they have two limitations:

[0007] 1) These methods all assume that the source and target domains share the same fault label classifier. However, in practical applications, there may be some small differences between the fault label classifiers of the two domains.

[0008] 2) These methods either only consider the marginal distribution adaptation of the data, or only consider the conditional distribution adaptation of the data, or set the marginal distribution and conditional distribution to the same weight when considering the joint distribution adaptation of the data. However, in practical applications, the marginal distribution and conditional distribution between data in different fields often contribute differently to domain adaptation. Summary of the invention

[0009] In order to solve the problems and limitations of current data-driven methods in the field of cross-domain fault diagnosis of aircraft engines, the present invention proposes an aircraft engine cross-domain gas path fault diagnosis method based on dynamic adversarial residual network (DARAN) from the perspective of transfer learning.

[0010] The dynamic adversarial residual adaption network DARAN (dynamic adversarial residual adaption network) includes a feature extractor, a global domain discriminator, a local domain discriminator, a source domain health status identifier and a target domain health status identifier. It inherits all the advantages of the DANN network, can learn domain-invariant features in end-to-end adversarial training, and achieves network optimization through a general stochastic gradient descent method. However, compared with the DANN network, it has made two improvements, which further improves the cross-domain diagnostic capability of the model. First, it is assumed that the source domain fault label classifier is different from the target domain label classifier, and the difference is a small perturbation function. Secondly, it is assumed that the marginal distribution and conditional distribution of aircraft engine gas path fault data are not equally important in the domain adaptation process, and can be quantitatively evaluated by the dynamic adversarial factor.

[0011] The present invention proposes an aero-engine cross-domain fault diagnosis method based on a dynamic adversarial residual network, comprising the following steps:

[0012] Step 1: normalize the collected source domain and target domain data to serve as training set data; the source domain and target domain data are aircraft engine state parameters;

[0013] Step 2: construct a dynamic adversarial residual network;

[0014] The dynamic adversarial residual network includes a feature extractor G f , source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator

[0015] Feature Extractor G f Extract the sample data features and use the feature extractor G f The extracted features are input into the source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator Source domain health status identifier G s and the target domain health status identifier G t They are used to classify the health status of source domain data and target domain data respectively;

[0016] Global domain discriminator G d and local domain discriminator Used to distinguish the domain origin of data, that is, whether it belongs to the target domain or the source domain.

[0017] In step 3, the classification losses of the source domain health identifier, target domain health identifier, global domain discriminator, and local domain discriminator are calculated respectively.

[0018] Step 4: Input the training set data into the dynamic adversarial residual network, perform back propagation through the stochastic gradient descent method, and train the network until convergence.

[0019] Furthermore, the aircraft engine state parameters include: high pressure shaft speed N h , low pressure shaft speed N l , compressor inlet temperature T 22 , compressor inlet pressure P 22 , compressor outlet temperature T3, compressor outlet pressure P3, high pressure turbine outlet temperature T 42 , high pressure turbine outlet pressure P 42 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 .

[0020] Furthermore, in step 1, the source domain and target domain data have the same feature space and label space, that is, But the probability distribution of source domain data P s (x s ,y s ) and the target domain data probability P t (x t ,y t ) distribution is different, that is, P s (x s ,y s )≠P t (x t ,y t ).

[0021] Furthermore, the source domain health status identifier G s and the target domain health status identifier G t The last layer activation function uses the softmax function;

[0022] Source domain health status identifier G s and the target domain health status identifier G t The difference is a small perturbation function, namely P s (y s |x s )=P t (y t |x t )+∈;

[0023] where ∈ is a small perturbation function that can be learned by the residual network.

[0024] Furthermore, the source domain health status identifier and the target domain health status identifier are connected through a residual network, which is expressed as

[0025] f S (fcs;θ r )=fcs+Δf(fcs;θ r )

[0026] Among them, fcs is the source domain health status identifier G s and the target domain health status identifier G t Share some of the extracted source domain data features; θ r is the parameter of the residual network; f S (fcs;θ r ) is used to calculate the output value of the source domain health status identifier before the softmax activation function; Δf(fcs; θ r ) satisfies the constraint condition |Δf(fcs;θ r )|<<|fcs|≈|f S (fcs;θ r )| is a small perturbation function.

[0027] Define f T is the function of the output layer of the target domain health recognition before the softmax activation function, that is, G t (G f (x)) = softmax(f T (G f (x);θ y )), then fcs=f T (G f (x s );θ y ), where θ y G is the source domain health status identifier s and the target domain health identifier G t Parameters of the shared part of the recognizer.

[0028] Furthermore, the source domain health status identifier G in step 3 s and the target domain health status identifier G t Cross entropy is used as the classification loss function, and its calculation expressions are:

[0029]

[0030]

[0031] in, G is the source domain health status identifier s Classification loss, G is the target domain health status identifier t Classification loss, n s Indicates the number of source domain samples, n t Indicates the number of samples in the target domain.

[0032] Furthermore, the global domain discriminator G d and local domain discriminator Both use binary cross entropy as their classification loss function, and their calculation expressions are:

[0033]

[0034]

[0035] Among them, L g is the global domain discrimination loss, L l is the local domain identification loss, * represents the dot product operation, represents the probability that the i-th sample in the source domain is predicted to be of category c, is the probability that the jth sample in the target domain is predicted as category c. Furthermore, assuming that the distance between the source domain and the target domain is determined by the marginal distribution distance and the conditional distribution distance, and the distances between the two distributions are not equally important, the distance between the two distributions is dynamically adjusted through the dynamic adversarial factor μ∈[0,1], expressed as:

[0036] D(P s (x s ,y s ),P t (x t ,y t ))≈(1-μ)D(P s (x s ),P t (x t ))+μD(P s (y s |x s ),P t (y t |x t )) Specifically, during the network training process, the global domain identification loss is used to represent the marginal distribution distance, and the local domain identification loss is used to represent the conditional distribution distance. Therefore, the final domain identification loss can be expressed as:

[0037] L d =(1-μ)L g +μL l

[0038] The dynamic adversarial factor μ is calculated by global and local Dynamically and quantitatively adjust the weights of marginal distribution and global distribution, and the calculation expression is:

[0039]

[0040] in, For the global It can be calculated as For local Calculated as Furthermore, in step 3, the loss function is combined to obtain the final optimized loss function:

[0041]

[0042] Among them, θ f is the feature extractor parameter, θ g is the global domain discriminator parameter, is the local domain discriminator parameter, β and λ are the trade-off hyperparameters. The ultimate training goal is to find the optimal hyperparameters and It can be expressed as:

[0043]

[0044]

[0045] Further, input to the global domain discriminator G d and local domain discriminator The features in , first pass through the gradient reversal layer GRL to ensure adversarial training in end-to-end learning. The gradient reversal layer is defined as a pseudo function R λ , and assuming that its input is X, its forward calculation and back-propagation gradient calculation can be expressed as:

[0046] R λ (X) = X

[0047]

[0048] Where I is the identity matrix.

[0049] Beneficial effects:

[0050] DARAN uses a dynamic adversarial learning method to dynamically and quantitatively adjust the weights of the marginal distribution and the conditional distribution in the domain adaptation process through dynamic adversarial factors, so that the model learns domain-invariant features. At the same time, the shared classifier assumption is relaxed, and the residual network learns the slight differences between the health status identifiers in the source domain and the target domain, making the network more consistent with the target domain data structure, effectively solving the problem of cross-domain fault diagnosis of aircraft engines. It is applied to the field of aircraft engine fault diagnosis to improve the application ability of historical data and make data-driven fault diagnosis methods better adapt to real-world working scenarios. Cross-domain fault diagnosis experiments on a certain type of engine simulation fault data show that compared with traditional data-driven methods and many classic domain adaptation methods, DARAN has excellent diagnostic accuracy on target domain data. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is the overall framework diagram of the DARAN model;

[0052] Figure 2 This is the specific architecture diagram of the DARAN feature extractor;

[0053] Figure 3 This is a structural diagram of a twin-shaft turbofan engine;

[0054] Figure 4 This is a comparison chart of the accuracy of different methods in the cross-domain fault diagnosis task of aircraft engines;

[0055] Figure 5 It is a graph showing the impact of different noise levels in the data set on the diagnostic accuracy of the model;

[0056] Figure 6 This is the effect of the value of the hyperparameter β on the diagnostic accuracy of DARAN;

[0057] Figure 7 It is the diagnostic accuracy graph corresponding to adjusting the dynamic confrontation factor μ from 0 to 1;

[0058] Figure 8 This is the DARAN training and testing framework diagram; DETAILED DESCRIPTION

[0059] The present invention provides a method for cross-domain fault diagnosis of aircraft engines based on a dynamic adversarial residual network. Figure 1 As shown, the following steps are included:

[0060] Step 1: normalize the collected source domain and target domain data to serve as training set data; the source domain and target domain data are aircraft engine state parameters;

[0061] Both the source domain data and the target domain data are aircraft engine state parameters, including: high pressure shaft speed N h, low pressure shaft speed N l , compressor inlet temperature T 22 , compressor inlet pressure P 22 , compressor outlet temperature T3, compressor outlet pressure P3, high pressure turbine outlet temperature T 42 , high pressure turbine outlet pressure P 42 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 .

[0062] The source domain and target domain data have the same feature space and label space, namely But the probability distribution of source domain data P s (x s ,y s ) and the target domain data probability P t (x t ,y t ) distribution is different, that is, P s (x s ,y s )≠P t (x t ,y t ).

[0063] Step 2: construct a dynamic adversarial residual network;

[0064] The dynamic adversarial residual network includes a feature extractor G f , source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator ;

[0065] like Figure 2 As shown, a one-dimensional AlexNet is used as the feature extractor G f Extract the sample data features and use the feature extractor G f The extracted features are input into the source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator Source domain health status identifier G s and the target domain health status identifier G t They are used to classify the health status of source domain data and target domain data respectively; the last layer activation function adopts the softmax function.

[0066] And, let the source domain health status identifier G s and the target domain health status identifier Gt The difference is a small perturbation function, namely P s (y s |x s )=P t (y t |x t )+∈, where ∈ is a small perturbation function that can be learned through the residual network. Therefore, the source domain health identifier and the target domain health identifier are connected through the residual network, which can be expressed as

[0067]

[0068] Among them, fcs is the source domain health status identifier G s and the target domain health status identifier G t Share some of the extracted source domain data features; θ r is the parameter of the residual network; f S (fcs;θ r ) is used to calculate the output value of the source domain health status identifier before the softmax activation function; Δf(fcs; θ r ) satisfies the constraint condition |Δf(fcs;θ r )|<<|fcs|≈|f S (fcs;θ r )|. Define f T is the function of the output layer of the target domain health recognition before the softmax activation function, that is, G t (G f (x)) = softmax(f T (G f (x);θ y )), then fcs=f T (G f (x s );θ y ), where θ y G is the source domain health status identifier s and the target domain health status identifier G t Parameters of the shared part of the recognizer.

[0069] Global domain discriminator G d and local domain discriminator It is used to distinguish the domain source of data, that is, whether it comes from the target domain or the source domain. The last layer uses the sigmoid activation function, and through dynamic adversarial training with the feature extractor, the feature extractor can learn domain-invariant features. Among them, the global domain discriminator is used to align the marginal distribution, and the local domain discriminator is used to align the conditional distribution.

[0070] In step 3, the classification losses of the source domain health identifier, the target domain health identifier, the global domain discriminator, and the local domain discriminator are calculated respectively, and the loss functions are combined to obtain the final optimized loss function.

[0071] Source domain health status identifier G s and the target domain health status identifier G t Cross entropy is used as the classification loss function, and its calculation expressions are:

[0072]

[0073] in, G is the source domain health status identifier s Classification loss, G is the target domain health status identifier t Classification loss, n s Indicates the number of source domain samples, n t Indicates the number of samples in the target domain.

[0074] Global domain discriminator G d and local domain discriminator Both use binary cross entropy as their classification loss function, and their calculation expressions are:

[0075]

[0076]

[0077] Among them, L g is the global domain discrimination loss, L l is the local domain identification loss, * represents the dot product operation, represents the probability that the i-th sample in the source domain is predicted to be of category c, is the probability that the jth sample in the target domain is predicted to be of category c.

[0078] Here we assume that the distance between the source domain and the target domain is determined by the marginal distribution distance and the conditional distribution distance, and the distances of the two distributions are not equally important. The distance between the two distributions can be dynamically adjusted by the dynamic adversarial factor μ∈[0,1], which can be expressed as:

[0079]

[0080] Specifically, during the network training process, the global domain discrimination loss is used to represent the marginal distribution distance, and the local domain discrimination loss is used to represent the conditional distribution distance. Therefore, the final domain discrimination loss can be expressed as:

[0081] L d =(1-μ)L g +μL l(7)

[0082] The dynamic adversarial factor μ is calculated by global and local To dynamically and quantitatively adjust the weights of edge distribution and global distribution, the calculation expression is:

[0083]

[0084] in, For the global It can be calculated as For local It can be calculated as The final optimized loss function obtained by combining loss functions is:

[0085]

[0086] Among them, θ f is the feature extractor parameter, θ g is the global domain discriminator parameter, is the local domain discriminator parameter, β and λ are the trade-off hyperparameters. The ultimate training goal is to find the optimal hyperparameters and It can be expressed as:

[0087]

[0088] Step 4: Input the training set data into the dynamic adversarial residual network, perform back propagation through the stochastic gradient descent method, and train the network until convergence.

[0089] In order to train the network through an end-to-end method, the input is sent to the global domain discriminator G d and local domain discriminator The features in are first passed through a gradient reversal layer (GRL) to ensure adversarial training in end-to-end learning. The gradient reversal layer is defined as a pseudo function R λ , and assuming that its input is X, its forward calculation and back-propagation gradient calculation can be expressed as:

[0090] R λ (X) = X (12)

[0091]

[0092] Where I is the unit matrix. At this point, the network training can be achieved by the stochastic gradient descent (SGD) method. All the parameters of the network are expressed as, Θ is updated as follows:

[0093]

[0094] Among them, α is the learning rate.

[0095] like Figure 8 As shown in Figure 2, during testing, the network only includes two parts: the feature extractor and the target domain health status identifier.

[0096] This case studies a certain type of twin-shaft turbofan engine, the structure of which is as follows: Figure 3 As shown in the figure. Common aircraft engine gas path component failures mainly include fan (Fan) failure, compressor (HPC) failure, high pressure turbine (HPT) failure, and low pressure turbine (LPT) failure. In the simulation model, the component failure is simulated by adjusting the efficiency coefficient SE and flow coefficient SW of each component. We use the health parameter vector HP = [SE1, SW1, SE2, SW2, SE3, SW3, SE4, SW4] to represent the performance degradation of the engine, where "1" corresponds to the fan, "2" corresponds to the compressor, "3" corresponds to the high pressure turbine, and "4" corresponds to the low pressure turbine. During the experiment, we obtained three sets of data with different distributions by changing the engine health parameters, working points, and fuel supply rules. Specifically, the working points of the three sets of data are shown in Table 1. Among them, at each working point, we collected 60 samples in each health state, that is, each data set includes 360×5 samples, and a single sample is a multidimensional time series with a length of 60 and a dimension of 10. In order to make the data more realistic, we assume that the sensor data noise conforms to the normal distribution and add Gaussian noise with a mean of 0 and a variance of 1% to all data sets.

[0097] Table 1 Working point table corresponding to each data set

[0098]

[0099] In order to evaluate the effectiveness of the DARAN of the present invention, we selected 6 methods including CNN, RTN, DANN, MADA, JAN, and DAAN without domain adaptation as comparison algorithms. The network structure of the feature extractor used by all methods is as follows Figure 2 As shown in the figure, the network is trained by the Adam method. During the training process, the batch size of all methods is set to 512, the learning rate is set to 0.001, and the weight decay coefficient is set to 0.0001. For the DARAN of the present invention, in all experiments, β is fixed to 0.5, and the adjustment strategy of λ is as follows:

[0100]

[0101] Among them, q = p / 1500 and p is the corresponding number of training steps, ranging from 0 to 1500, that is, the training epoch is set to 1500. In order to avoid the randomness of the experimental results, each group of experiments was repeated 10 times and the average value was taken.

[0102] First, by using the three sets of collected data, six sets of cross-domain fault diagnosis experiments were designed. The experimental results are as follows: Figure 4 As shown in Table 2. In each task, the dataset before the arrow is the source domain, and the dataset after the arrow is the target domain. Figure 4 As can be seen from Table 2, the average diagnostic accuracy of DARAN is 97.74%, which is much higher than that of CNN without domain adaptation, proving its effectiveness in cross-domain fault diagnosis. Compared with RTN, the diagnostic accuracy of DARAN is improved by 2.65%, and compared with DANN, the average diagnostic accuracy is improved by 4.71%, indicating the excellent performance of DARAN in domain adaptation.

[0103] Table 2 Diagnosis accuracy of cross-domain fault diagnosis experiment

[0104]

[0105] Secondly, in the actual working environment, the collected sensor data often contains noise, and different noise levels will also affect the difference between the source domain and the target domain. By changing the noise level of the target domain, the good generalization performance of the present invention is further demonstrated. Specifically, in all experiments, the source domain data noise is set to Gaussian noise with a mean of 0 and a variance of 1%, and the noise of the target domain is selected from {0.5%, 1%, 1.5%, 2%, 2.5%, 3%, 3.5%, 4%, 4.5%, 5%}. Finally, the experimental results are as follows Figure 5 As shown. Figure 5 It can be seen that the performance of all domain adaptation methods at all noise levels is higher than that of CNN without domain adaptation, and as the noise level increases, the diagnostic performance of CNN without domain adaptation decreases more than the domain adaptation method. In addition, the performance of DARAN in the present invention is better than other domain adaptation methods at all noise levels, which proves the robustness of the present invention to noise.

[0106] In addition, in order to further explore the performance of the present invention in cross-domain fault diagnosis, the value of the adjustable hyperparameter β is changed to obtain the experimental results as follows: Figure 5 As shown, and the dynamic confrontation factor μ is set to a fixed value, we can obtain Figure 6 The experimental results are shown in Figure 2. Figure 5 As shown in the figure, for any β value in the range of 0 to 1, the diagnostic accuracy of DARAN of the present invention is higher than that of DANN, which proves the effectiveness of DARAN of the present invention. Figure 6 From the experimental results shown, it can be seen that different values ​​of μ have a certain impact on the diagnostic accuracy, and only at one value, the diagnostic accuracy is better than that of DARAN of the present invention, indicating that the present invention can effectively find the optimal μ value in an end-to-end training framework.

[0107] Finally, to further explore the role of the residual module, the cosine distance is used to quantitatively measure the gap between the source domain health status identifier and the target domain health status identifier. The calculation formula is as follows:

[0108]

[0109] Select tasks A→B and A→C as the research objects. In task A→B, we calculate d(f s (x;θ r ),Δf(x;θ r ))=0.9346,d(f s (x;θ r ),f t (x))=0.9789. In task A→C, we calculate d(f s (x;θ r ),Δf(x;θ r ))=0.9511,d(f s (x;θ r ),f t (x))=0.9896. The experimental results verify the rationality of the small perturbation assumption.

Claims

1. A method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network, characterized in that: The steps include: Step 1: normalize the collected source domain and target domain data to serve as training set data; the source domain and target domain data are aircraft engine state parameters; Step 2: construct a dynamic adversarial residual network; The dynamic adversarial residual network includes a feature extractor G f , source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator Feature Extractor G f Extract the sample data features and use the feature extractor G f The extracted features are input into the source domain health status identifier G s , target domain health status identifier G t , global domain discriminator G d and local domain discriminator In the source domain health status identifier G s and the target domain health status identifier G t They are used to classify the health status of source domain data and target domain data respectively; Global domain discriminator G d and local domain discriminator Used to distinguish the domain source of data, that is, whether it belongs to the target domain or the source domain; Step 3, calculate the classification losses of the source domain health identifier, the target domain health identifier, the global domain discriminator, and the local domain discriminator respectively; Step 4: Input the training set data into the dynamic adversarial residual network, perform back propagation through the stochastic gradient descent method, and train the network until convergence; Source domain health status identifier G s and the target domain health status identifier G t The last layer activation function uses the softmax function; Source domain health status identifier G s and the target domain health status identifier G t The difference is a small perturbation function, namely P s (y s |x s )=P t (y t |x t )+∈; Where ∈ is a small perturbation function that can be learned through the residual network; The source domain health identifier and the target domain health identifier are connected through a residual network, which is expressed as f S (fcs;θ r )=fcs+Δf(fcs;θ r ) Among them, fcs is the source domain health status identifier G s and the target domain health status identifier G t Share some of the extracted source domain data features; θ r is the parameter of the residual network; f S (fcs;θ r ) is used to calculate the output value of the source domain health status identifier before the softmax activation function; Δf(fcs; θ r ) satisfies the constraint condition |Δf(fcs;θ r )|<<|fcs|≈|f S (fcs;θ r )| small perturbation function; define f T is the function of the output layer of the target domain health recognition before the softmax activation function, that is, G t (G f (x)) = softmax(f T (G f (x);θ y )), then fcs=f T (G f (x s );θ y ), where θ y G is the source domain health status identifier s and the target domain health identifier G t Parameters of the shared part of the recognizer.

2. According to claim 1, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: Aircraft engine status parameters include: high pressure shaft speed N h , low pressure shaft speed N l , compressor inlet temperature T 22 , compressor inlet pressure P 22 , compressor outlet temperature T3, compressor outlet pressure P3, high pressure turbine outlet temperature T 42 , high pressure turbine outlet pressure P 42 , low pressure turbine outlet temperature T 46 , low pressure turbine outlet pressure P 46 .

3. According to claim 1, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: In step 1, the source domain and target domain data have the same feature space and label space, that is, But the probability distribution of source domain data P s (x s ,y s ) and the target domain data probability P t (x t ,y t ) distribution is different, that is, P s (x s ,y s )≠P t (x t ,y t ).

4. According to claim 1, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: The source domain health status identifier G in step 3 s and the target domain health identifier G t Cross entropy is used as the classification loss function, and its calculation expressions are: in, G is the source domain health status identifier s Classification loss, G is the target domain health status identifier t Classification loss, n s Indicates the number of source domain samples, n t represents the number of samples in the target domain, is the probability that the jth sample in the target domain is predicted to be of category c.

5. According to claim 4, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: Global domain discriminator G d and local domain discriminator Both use binary cross entropy as their classification loss function, and their calculation expressions are: Among them, L g is the global domain discrimination loss, L l is the local domain identification loss, * represents the dot product operation, represents the probability that the i-th sample in the source domain is predicted to be of category c, is the probability that the jth sample in the target domain is predicted to be of category c.

6. According to claim 5, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: Assume that the distance between the source domain and the target domain is determined by the marginal distribution distance and the conditional distribution distance, and the distances of the two distributions are not equally important. The distance between the two distributions is dynamically adjusted by the dynamic adversarial factor μ∈[0,1], which is expressed as: D(P s (x s ,y s ),P t (x t ,y t ))≈(1-μ)D(P s (x s ),P t (x t ))+μD(P s (y s |x s ),P t (y t |x t )) Specifically, during the network training process, the global domain discrimination loss is used to represent the marginal distribution distance, and the local domain discrimination loss is used to represent the conditional distribution distance. Therefore, the final domain discrimination loss can be expressed as: L d =(1-μ)L g +μL l The dynamic adversarial factor μ is calculated by global and local Dynamically and quantitatively adjust the weights of marginal distribution and global distribution, and the calculation expression is: in, For the global It can be calculated as For local Calculated as 7. According to claim 6, a method for cross-domain fault diagnosis of aircraft engines based on dynamic adversarial residual network is characterized in that: In step 4, the loss function is also combined to obtain the final optimized loss function: Among them, θ f is the feature extractor parameter, θ g is the global domain discriminator parameter, is the local domain discriminator parameter, β and λ are the trade-off hyperparameters; the ultimate training goal is to find the optimal hyperparameter and It can be expressed as:

8. The method for cross-domain fault diagnosis of an aircraft engine based on a dynamic adversarial residual network according to claim 7 is characterized in that: Input to the global domain discriminator G d and local domain discriminator The features in the image are first passed through the gradient reversal layer GRL to ensure adversarial training in end-to-end learning. The gradient reversal layer is defined as a pseudo function R λ , and assuming that its input is X, its forward calculation and back-propagation gradient calculation can be expressed as: R λ (X)=X Where I is the identity matrix.

Images