An aeroengine fault diagnosis method based on domain distribution prediction adversarial transfer

Abstract

The present invention relates to the field of aircraft engine gas path fault diagnosis, and proposes an aircraft engine fault diagnosis method based on domain distribution prediction adversarial migration. The fault diagnosis method can achieve transfer learning in continuous domains and multi-dimensional domains by constructing an adversarial migration model based on domain distribution prediction to characterize the domain information of aircraft engine working data, and solves the problem of decreased fault diagnosis accuracy due to lack of effective update data during the operation of aircraft engines due to performance degradation and flight state changes. At the same time, the diagnostic method has good robustness in the face of different scenarios, providing a new perspective for the diagnosis of air path faults in aircraft engines under variable states, and helping to realize the judgment of the health status of the engine during its entire operating life cycle.

Description

Technical Field

[0001] The invention relates to the field of aviation engine gas path fault diagnosis, and in particular to an aviation engine fault diagnosis method based on domain distribution prediction countermeasure migration. Background Art

[0002] When aircraft engines are affected by fouling, erosion, corrosion, etc. during actual use, components often deviate from the standard state, causing the engine gas path performance to degrade. If not handled promptly and appropriately, it will increase maintenance costs and even cause dangerous situations. Accurate and reliable diagnosis can help researchers quickly locate the engine fault location and provide a reference for subsequent operation and maintenance.

[0003] With the construction of industrial digital systems and the development of database technology, data-driven intelligent algorithms with strong nonlinear mapping capabilities are gradually being applied to engine fault diagnosis. In the prior art, when the database contains enough healthy and various fault data samples, fault diagnosis methods based on intelligent algorithms such as support vector machines (SVM), Markov models or artificial neural networks (ANN) have achieved good applications in engine diagnosis problems. The data of the above diagnostic methods are all collected from engines in the same working state, but the collected operating data of the engine in actual service generally do not obey the same distribution. The distribution difference of this data is reflected in two aspects: First, the difference in component performance levels, especially the continuous degradation of performance during use will cause continuous changes in data distribution. In this way, the data collected at the current moment cannot be used for future fault diagnosis, and then after the performance declines to a certain extent, the engine with only performance degradation may be misdiagnosed as a failure of some components. Second, because the flight conditions of the engine are constantly changing, which in turn affects the engine control law and causes performance changes, the model trained in a limited state cannot be applied to the diagnosis of multiple flight state points.

[0004] In order to solve the problem that the trained diagnostic model cannot be applied to new engines or new operating conditions, existing methods usually adjust the model in real time based on the newly added data. However, it is also greatly restricted in actual usage scenarios. During the diagnosis process, it may not be possible to collect enough data to meet the training needs, especially the amount of data in the engine failure state is less than that in the healthy state, which is likely to cause class imbalance problems. In addition, it is difficult to collect data over the entire life cycle of the engine, and the newly added data is often lacking in annotations or even unlabeled, making it impossible to retrain the diagnostic model. Therefore, in aircraft engine fault diagnosis, there is often a contradiction between the demand for high generalization capabilities of diagnostic methods and the limited available data. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes an aircraft engine fault diagnosis method based on domain distribution prediction adversarial migration. The fault diagnosis method solves the problems of continuous domain deviation and decreased accuracy of the fault diagnosis method caused by performance degradation or flight state changes by constructing an adversarial migration model based on domain distribution prediction.

[0006] The technical solution of the present invention is as follows:

[0007] An aircraft engine fault diagnosis method based on domain distribution prediction adversarial migration includes the following steps:

[0008] Step 1: Obtain parameter combinations of multiple engines at different operating times through sensor measurements, and use them to construct a training set;

[0009] Step 2: Establish an adversarial transfer model based on domain distribution prediction;

[0010] Step 3: Use the training set obtained in step 1 to train the adversarial transfer model;

[0011] Step 4: Apply the trained adversarial migration model to fault diagnosis of engine operation data.

[0012] Preferably, the adversarial transfer model consists of a feature extractor, a domain classifier and a label classifier, wherein the feature extractor is used to extract feature data from the source domain and the target domain, the domain classifier is used to determine the domain source of the feature data, and the label classifier is used to receive the feature data of the source domain and perform task classification.

[0013] Preferably, the domain classifier is a distribution prediction model for the domain source of feature data, and the domain source is judged by fitting actual continuous domain index values.

[0014] Preferably, the parameter combination includes performance parameters, state labels, and domain indexes; the input of the adversarial migration model is the performance parameters and the domain index, and the output is the state label.

[0015] Preferably, the parameter combination is derived from a continuous domain distribution or a multi-dimensional continuous domain distribution.

[0016] Preferably, the last layer of the domain classifier is a linear layer, which outputs a regression prediction value for the domain, and uses Gaussian distribution to describe the domain distribution prediction output of the domain classifier.

[0017] Preferably, for a parameter combination derived from a continuous domain distribution, the probability density function of the Gaussian distribution is:

[0018]

[0019] where μ and σ 2Represents the mean and variance of the predicted distribution of domain index u.

[0020] For parameter combinations derived from multidimensional continuous domain distribution, a hybrid dense network MDN is used to describe the domain. The Gaussian distribution is represented as a linear combination of multiple kernel functions, and its probability density function is:

[0021]

[0022]

[0023] Where u represents the domain index, σ represents the prediction standard deviation of the domain classifier, μ represents the prediction mean of the domain classifier, K represents the number of kernel functions, and α k represents the weight coefficient, represents the kth kernel function of domain index u, θ k Represents the parameters of the kth kernel function.

[0024] Preferably, for a parameter combination derived from a continuous domain distribution, the domain classification loss function L of the domain classifier is adv as follows:

[0025]

[0026] Where m represents the size of the training sample batch, u i Represents the domain index value of the i-th sample.

[0027] For a parameter combination derived from a multi-dimensional continuous domain distribution, the domain classification loss function L of the domain classifier adv as follows:

[0028]

[0029]

[0030] Where L represents the total likelihood function of the domain classifier, logσ 2 represents the prediction variance of the domain classifier, μ represents the prediction mean of the domain classifier, and L i represents the loss function for domain index prediction of a single data sample, σ k represents the standard deviation of the kth kernel function, μ k represents the mean of the kth kernel function.

[0031] Preferably, the total loss function L of the adversarial migration model is adv (θ f ,θ y ,θ d )as follows:

[0032] L adv (θ f ,θy ,θ d )=L y (θ f ,θ y )+γL mmd (θ f )+λL adv (θ f ,θ d )

[0033] Among them, L y (θ f ,θ y ) represents the label classification loss function of the label classifier, L mmd (θ f ) represents the joint distribution difference loss function of the feature layer, L adv (θ f ,θ d ) represents the domain classification loss function of the domain classifier, γ and λ are weight parameters, θ f ,θ d ,θ y denote the model parameters of the feature extractor, domain classifier, and label classifier, respectively.

[0034] Preferably, the joint distribution difference loss function L of the feature layer is mmd (θ f ) is expressed as:

[0035]

[0036] Among them, G f (·) represents the feature extraction function, n s ,n t Represents the number of samples in the source domain and the target domain respectively, x s ,x t Represent the input parameters of samples in the source domain and target domain respectively.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention proposes an aero-engine fault diagnosis method based on domain distribution prediction and adversarial migration, which replaces the discretization processing mode of domain information in the traditional transfer learning algorithm, and adopts a distribution model to characterize the domain information of the aero-engine working data, which can realize the transfer learning of continuous domain and multi-dimensional domain, and solves the problem that the aero-engine lacks effective update data due to performance degradation and flight state changes during operation, resulting in the decrease of fault diagnosis accuracy. At the same time, the diagnostic method has good robustness in the face of different scenarios, provides a new perspective for the diagnosis of gas path faults under the variable state of the aero-engine, and helps to realize the judgment of the healthy state of the engine under the full operating life cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. The features and advantages of the present invention can be more clearly understood by referring to the drawings. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 It is a flow chart of the aircraft engine fault diagnosis method based on domain distribution prediction countermeasure migration proposed by the present invention.

[0041] Figure 2 It is an adversarial migration model based on continuous domain distribution prediction established by the present invention.

[0042] Figure 3 The present invention establishes an adversarial migration model based on multi-dimensional continuous domain distribution prediction.

[0043] Figure 4 The present invention is a process diagram for diagnosing aircraft engine faults. DETAILED DESCRIPTION

[0044] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0046] like Figure 1 In the specific application of the aviation engine fault diagnosis method based on domain distribution prediction adversarial migration proposed in the present invention, the accuracy of the adversarial migration algorithm based on continuous domain distribution prediction is significantly improved compared with the existing migration algorithm, and it shows a diagnostic accuracy of more than 98% over the entire life cycle of the engine. When expanded to a multi-dimensional continuous domain, the adversarial migration algorithm based on multi-dimensional continuous domain distribution prediction has a leading advantage in multi-state point data sets that are independent of the training data set. In addition to achieving good domain migration in the domain dimension of the operating cycle, transfer learning is also achieved in the three-dimensional domain after expansion of altitude and Mach number, and it has the ability to diagnose in different states within the flight envelope.

[0047] In the diagnostic method of the present invention, an adversarial transfer model based on domain distribution prediction is established, and the adversarial transfer model includes a feature extractor, a domain classifier and a label classifier.

[0048] In the migration task, given the source domain D s and the target domain D t , where P s (x,y)≠P t (x, y), the data distribution is similar but different. When training the adversarial transfer model, the training data passes through the feature extractor f = G f (x;θ f ) are mapped to the same Z-dimensional feature space, that is, f∈R Z The network is then divided into a label classifier G y (x;θ y ) and domain classifier G d (x;θ d ), where x represents the model input and θ represents the model parameters. The label classifier receives the feature data extracted from the source domain for task classification; the domain classifier is used to determine the domain source of the feature data.

[0049] The core of transfer learning is the evaluation function that measures the difference in data distribution. This function does not need to be displayed like the maximum mean difference, but is learned in the process of adversarial game, also known as domain discrimination loss. The loss function of the adversarial transfer model includes the label classification loss L on the source domain. y and domain classification loss L adv Among them, L y Only the label classification loss of the source domain is calculated, L ad Compute the domain classification loss for the source and target domains.

[0050] In order to adapt to the continuous changes in domain distribution, the present invention is different from the traditional adversarial migration model in that it further optimizes the structure of the domain classifier. The domain information of the calibration data source is changed from the traditional discrete division method to the continuous division method. The output value of the domain classifier is no longer a discrete domain index value, but a distribution prediction model for the data source domain information, so as to realize the judgment of the domain source by fitting the actual continuous domain index value. This improvement to the domain classifier effectively solves the problem of continuous changes in data with the domain. Compared with the traditional discrete domain division method, the accuracy has been significantly improved, and it can be effectively extended to multi-dimensional domain data. Since the domain distribution continues to shift, the domain of the feature data is represented by a continuous index value, and the last layer of the domain classifier is set to a linear layer to output the regression prediction value of the domain. However, the regression prediction of the domain value by the domain classifier only guarantees the alignment of the domain index mean. Therefore, by introducing a probability distribution model in the domain classifier, a higher-dimensional alignment of the domain value distribution is achieved. Such as Figure 2 shown.

[0051] The domain classifier fits the predicted mean μ and predicted variance logσ of the continuous domain index u 2 , describing the confidence interval that the domain index value may fall into. Fitting logσ 2 The reason is that there is no need to add an activation function. After adopting this strategy, you only need to accumulate the probability distribution of all domains in the integral domain and solve the maximum log likelihood of the target data. Gaussian distribution is used to describe the domain distribution prediction output of the domain classifier, and its probability density function is as follows:

[0052]

[0053] Where u represents the domain index, σ represents the prediction standard deviation of the domain classifier, and μ represents the prediction mean of the domain classifier.

[0054] The maximum likelihood estimation is used to estimate the structural parameters of the domain classifier. The negative likelihood logarithm is selected as the loss function of the domain classifier. After taking the logarithm of the likelihood function, the problem of solving the distribution model parameters is transformed into the problem of finding the extreme value. For all samples, the total likelihood function is calculated:

[0055]

[0056] After taking the negative logarithm of the above formula, the cumulative multiplication is transformed into an accumulation problem that is easy to solve. Therefore, the classification loss function of the domain classifier is expressed as:

[0057]

[0058] Where L represents the total likelihood function of the domain classifier, logσ 2 represents the prediction variance of the domain classifier, μ represents the prediction mean of the domain classifier, m represents the size of the training sample batch, and u i Represents the domain index value of the i-th sample.

[0059] In addition to the domain discrimination loss, the joint distribution difference loss is introduced to evaluate the joint maximum mean difference at the feature layer. The introduction of this loss helps to reduce the difference in joint probability distribution between the source domain and the target domain and improve the convergence of the adversarial transfer model. The joint distribution difference loss function of the feature layer can be expressed as:

[0060]

[0061] The Gaussian kernel function is used to improve the measurement performance of feature distribution differences. The Gaussian kernel function is as follows:

[0062]

[0063] Among them, x1 and x2 represent two different parameters.

[0064] In summary, the total loss function L of the adversarial transfer model is CDDAN (θ f ,θ y ,θ d ) can be expressed as:

[0065] L CDDAN (θ f ,θ y ,θ d )=L y (θ f ,θ y )+γL mmd (θ f )+λL adv (θ f ,θ d )

[0066] Among them, L y (θ f ,θ y ) represents the label classification loss function of the label classifier, L mmd (θ f ) represents the joint distribution difference loss function of the feature layer, L adv (θ f ,θ d ) represents the domain classification loss function of the domain classifier, γ and λ are weight parameters, θ f ,θ d ,θ y denote the model parameters of the feature extractor, domain classifier, and label classifier, respectively.

[0067] By minimizing the total loss function, the optimal solution for the model parameters can be obtained.

[0068]

[0069] It should be noted that the loss L adv The weight parameter λ is a negative number, so as to ensure the confusion of the source domain and the target domain data set. The weight parameters γ and λ control the balance of the label classifier and the domain classifier, representing the influence weight of various losses. The Adam optimizer is used to implement the parameter update of the model because of its fast convergence speed and not easy to fall into the local optimum. The optimal parameters of the model are obtained by iteratively using the error back propagation training algorithm, where the learning rate η controls the speed of model parameter iteration.

[0070]

[0071]

[0072]

[0073] When faced with a dataset from a multi-dimensional domain distribution, a mixed dense network (MDN) is introduced to describe the domain in which the data is located. MDN has the strong fitting ability of neural networks, and at the same time introduces the evaluation of model uncertainty through Gaussian mixture distribution. Gaussian mixture distribution is an extension of a single probability density distribution model, which can smoothly approximate density distributions of any shape when the classification of the sample is known. The structure diagram of the adversarial transfer model based on multi-dimensional continuous domain distribution prediction is shown in the figure below. Figure 4 shown.

[0074] Gaussian mixture distribution can be expressed as a linear combination of multiple kernel functions, as shown in the formula.

[0075]

[0076]

[0077] K represents the number of kernel functions, α k represents the weight coefficient, represents the kth kernel function of domain index u, θ k Represents the parameters of the kth kernel function. When the number of mixed Gaussian distributions is appropriate, any data distribution can be approximately fitted in theory.

[0078] The parameters of the mixed Gaussian distribution model are also determined by maximum likelihood estimation to ensure that the probability of f(u) is maximized under the given input conditions. The weight coefficient, mean, and standard deviation are all output parameters of the domain classifier to be optimized. The weight coefficient can be obtained by normalization using the following softmax function.

[0079]

[0080] in Output u for the domain classifier k The part corresponding to the weight coefficient in . Correspondingly, the mean and standard deviation of each Gaussian distribution can also be expressed as:

[0081]

[0082]

[0083] Taking the negative logarithm of the probability density of the output target value, we get the classification loss function definition of the domain classifier:

[0084]

[0085]

[0086] The optimization strategy for model parameters during training is the same as that for the one-dimensional continuous domain model.

[0087] Example 1

[0088] like Figure 3 The method of the present invention for diagnosing aero-engine faults based on continuous domain distribution prediction and adversarial migration is applied to the fault diagnosis of the operating data of the aero-engine during the whole life cycle, including the following steps:

[0089] Step 1: In the study of engine fault diagnosis, the health status of the engine is usually reflected by the parameter changes measured by the sensor. Combined with the actual measurement situation, the appropriate measurement parameter combination is selected through the parameter optimization method. When used as the input of the diagnostic model, the recorded measurement parameters are normalized to scale to the same dimension.

[0090] Step 2, randomly select data samples at each operating moment from the operating data of multiple engines. The complete data sample includes performance parameter y, state label R and domain index u (that is, the number of operating cycles at that moment). The data input of the adversarial migration model based on domain distribution prediction is the number of samples × the processed performance parameter (y,u), and the output is the number of samples × the corresponding state label R. The single-dimensional continuous domain migration only contains the single-dimensional domain information of the number of operating cycles, and the multi-dimensional continuous domain migration contains the multi-dimensional domain information of the flight working status and the number of operating cycles. The details of the input and output parameters of the model are shown in Table 1, where T represents the number of operating cycles, H and Ma represent the altitude and Mach number in the flight state, respectively.

[0091] Table 1

[0092]

[0093] Step 3: filter data samples based on the characteristics of the engine's operating data. The engine's usage data has the following characteristics: engine failure sample data only accounts for a small part of the entire data set, historical data is mostly concentrated in the initial operation stage and lacks sufficient full life cycle operation data, and newly recorded data lacks status annotations. Figure 4 As shown in the figure, during training, the data in the source domain T0 represents the historical operation data of the engine, and the data in the target domain T 500 ,T 1000 The data in …represents the newly added data that lacks annotations for the engine. The number of samples in each state in the test data set is evenly distributed, covering more fault states, and is used to test the model effect.

[0094] Step 4: The number of hidden layers of the model is selected from {10, 100, 200, ..., 600}. Experiments show that when the number of layers is small, the model will be underfitting in the target domain. When the number of layers is greater than 300, the model performance gradually converges. Therefore, the number of hidden layers of the model can be set to 300. The regularization parameters used to adjust the loss weight include γ for the distribution difference loss and λ for the domain classifier loss. Both parameters are selected from {10 -3 ,10 -2 ,…,10 2 ,10 3 ,}, and the learning rate parameter η is selected from {10 -6 ,10 -2 ,…,10 2 ,10 3 ,}, the loss coefficients after optimization are γ=1, λ=1, and the learning rate η is set to 10 -4 For a fair comparison, all deep networks are built using Pytorch.

[0095] Through the method described in this embodiment, engine fault diagnosis under continuous degradation and multiple flight states is achieved. The model after domain generalization has a stronger anti-interference ability for extreme data sets and is easier to adapt to various complex environments in actual situations. It solves the problem of continuous changes in the distribution of engine data sets and lack of annotations, realizes unsupervised migration to continuous target domains, and reduces the demand for data sample size.

[0096] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0097] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0098] In the present invention, the terms "first", "second", "third", "fourth" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. The term "plurality" refers to two or more, unless otherwise clearly defined.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An aircraft engine fault diagnosis method based on domain distribution prediction adversarial migration, characterized in that: The following steps are involved: Step 1: Obtain parameter combinations of multiple engines at different operating times through sensor measurements, and use them to construct a training set; Step 2: Establish an adversarial transfer model based on domain distribution prediction; Step 3: Use the training set obtained in step 1 to train the adversarial transfer model; Step 4: applying the trained adversarial migration model to fault diagnosis of engine operation data; The adversarial transfer model is composed of a feature extractor, a domain classifier and a label classifier, wherein the feature extractor is used to extract feature data from the source domain and the target domain, the domain classifier is used to determine the domain source of the feature data, and the label classifier is used to receive the feature data of the source domain and perform task classification; The domain classifier is a distribution prediction model for the domain source of feature data, and realizes the judgment of the domain source by fitting the actual continuous domain index value; The parameter combination includes performance parameters, state labels, and domain indexes. The input of the adversarial migration model is the performance parameters and the domain index, and the output is the state label.

2. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 1 is characterized in that: The parameter combination is derived from a continuous domain distribution or a multi-dimensional continuous domain distribution.

3. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 2 is characterized in that: The last layer of the domain classifier is a linear layer, which outputs a regression prediction value for the domain, and uses Gaussian distribution to describe the domain distribution prediction output of the domain classifier.

4. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 3 is characterized in that: For parameter combinations derived from continuous domain distribution, the probability density function of the Gaussian distribution is: For parameter combinations derived from multidimensional continuous domain distribution, a hybrid dense network MDN is used to describe the domain. The Gaussian distribution is represented as a linear combination of multiple kernel functions, and its probability density function is: Where u represents the domain index, σ represents the prediction standard deviation of the domain classifier, μ represents the prediction mean of the domain classifier, K represents the number of kernel functions, and α k represents the weight coefficient, represents the kth kernel function of domain index u, θ k Represents the parameters of the kth kernel function.

5. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 4 is characterized in that: For parameter combinations derived from continuous domain distributions, the domain classification loss function L of the domain classifier adv as follows: Where m represents the size of the training sample batch, u i Represents the domain index value of the i-th sample; For a parameter combination derived from a multi-dimensional continuous domain distribution, the domain classification loss function L of the domain classifier adv as follows: Where L represents the total likelihood function of the domain classifier, logσ 2 represents the prediction variance of the domain classifier, μ represents the prediction mean of the domain classifier, and L i represents the loss function for domain index prediction of a single data sample, σ k represents the standard deviation of the kth kernel function, μ k represents the mean of the kth kernel function.

6. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 5, characterized in that: The total loss function L of the adversarial transfer model CDDAN (θ f ,θ y ,θ d )as follows: L CDDAN (i f ,i y ,i d )=L y (i f ,i y )+γL mmd (i f )+λL adv (i f ,i d ) Among them, L y (θ f ,θ y ) represents the label classification loss function of the label classifier, L mmd (θ f ) represents the joint distribution difference loss function of the feature layer, L adv (θ f ,θ d ) represents the domain classification loss function of the domain classifier, γ and λ are weight parameters, θ f ,θ d ,θ y denote the model parameters of the feature extractor, domain classifier, and label classifier, respectively.

7. The method for aircraft engine fault diagnosis based on domain distribution prediction and countermeasure migration according to claim 6, characterized in that: The joint distribution difference loss function L of the feature layer mmd (θ f ) is expressed as: Among them, G f (·) represents the feature extraction function, n s ,n t Represents the number of samples in the source domain and the target domain respectively, x s ,x t Represent the input parameters of samples in the source domain and target domain respectively.

Images