Civil aircraft fault diagnosis method based on feature anchor point alignment in variable working condition scene

Abstract

The invention discloses a civil aircraft fault diagnosis method based on feature anchor point alignment in a variable working condition scene, and the method comprises the steps: calculating the maximum mean value difference of the edge distribution of labeled civil aircraft operation data as a source domain and unlabeled civil aircraft operation data as a target domain according to the output of a full connection layer in front of a bottleneck layer of a neural network; the edge distribution of the source domain and the target domain is aligned; secondly, cross-domain condition distribution alignment is assisted on the basis of feature anchor points, so that the phenomenon of fault feature confusion caused by unbalanced variable working condition fault data is avoided, and negative migration is eliminated; and finally, generating initial feature anchor points of different types of faults by using all fault data under all working conditions, iterating the distribution of the feature anchor points based on a statistical measurement method to enhance the discrimination, and assisting the network to perform cross-domain migration through the optimized and balanced feature anchor points. According to the method, the defect of non-frequent fault feature discrimination caused by fault data imbalance can be relieved, the discrimination of accidental faults and frequent faults is improved, and the flight safety and task reliability of a civil aircraft in the whole voyage are effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to a technology in the field of civil aviation flight control, specifically a civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions. Background Art

[0002] One of the challenges of civil aircraft fault diagnosis under equipment variable operating conditions is the problem of data set distribution offset. Since the working conditions of the aircraft vary greatly in different flight phases, the distribution of operating data changes accordingly, which brings difficulties to the state estimation and fault diagnosis of the entire flight. Secondly, the historical data of aircraft components are characterized by incomplete or unlabeled labels, which further increases the difficulty. In addition, due to the different probabilities of different faults, civil aircraft failures are unbalanced, which is reflected in the different proportions in civil aircraft operating data. Among them, occasional failures with a low proportion are often easily overlooked.

[0003] The core of most existing technologies is to optimize the distribution differences between different fields through distribution measurement methods, so as to obtain a certain cross-domain diagnostic capability. However, at the same time, they still cannot solve the risk of negative migration caused by unbalanced fault data, that is, due to the randomness of sampling, the information of occasional faults is missing, which makes the features of frequent faults and occasional faults in the global feature space mixed, thus affecting the diagnostic performance of the model. Summary of the invention

[0004] In order to solve the negative migration problem caused by the imbalance of fault data, the present invention proposes a civil aircraft fault diagnosis method based on feature anchor point alignment in variable working condition scenarios. Based on the transfer learning network based on feature mapping, the data under different working conditions are mapped into a well-designed feature space by aligning the cross-domain fault data features, realizing domain adaptation in variable working condition scenarios, and improving the influence of data distribution differences caused by variable working conditions of existing equipment and incomplete target data set labels on the performance of traditional diagnostic models. On this basis, a method based on feature anchor point alignment is adopted to alleviate the lack of discriminability of non-frequent fault features caused by the imbalance of fault data, improve the discriminability of occasional faults and frequent faults, and effectively ensure the flight safety and mission reliability of civil aircraft throughout the entire voyage.

[0005] The present invention is achieved through the following technical solutions:

[0006] The present invention relates to a civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating condition scenarios. According to the output of a fully connected layer before a bottleneck layer of a neural network, the maximum mean difference of the marginal distribution of labeled civil aircraft operation data as a source domain and the unlabeled civil aircraft operation data as a target domain is calculated to align the marginal distribution of the source domain with that of the target domain. Then, feature anchor point-assisted cross-domain conditional distribution alignment is adopted to avoid the phenomenon of fault feature confusion caused by imbalance of variable operating condition fault data, thereby eliminating negative migration. Finally, all fault data under all operating conditions are used to generate initial feature anchor points of different types of faults, and the distribution of feature anchor points is iterated based on a statistical measurement method to enhance its discriminability, and the network is assisted in cross-domain migration by optimizing the balanced feature anchor points.

[0007] The neural network includes: a feature extractor and a classifier, the classifier is composed of a bottleneck layer in a backbone network, and the rest of the backbone network is a feature extractor, wherein: the feature extractor performs feature extraction operations such as convolution and pooling according to the input civil aircraft operation data of the source domain and the target domain to obtain complex fault features, and the classifier performs dimensionality reduction processing based on feature information in the form of a bottleneck layer, maps the feature space calculated by the previous convolution, pooling and other layers to the sample label space, and obtains the fault classification result.

[0008] The fully connected layer before the bottleneck layer of the neural network is composed of fully connected neurons, and a complex nonlinear function fitting structure is constructed using an activation function. The fully connected layer integrates the local fault features extracted by each convolutional layer in a feature weighted manner to obtain complete global fault feature information.

[0009] The neural network is trained in the following way:

[0010] Step 1: Cross-domain edge distribution alignment. Specifically, after the fault data of the source domain and the target domain are input into the feature extractor respectively, the cross-domain edge distribution alignment loss is calculated using the maximum mean difference, and then the data differences in different flight stages are eliminated.

[0011] Step 2: Cross-domain conditional distribution alignment is assisted by feature anchors. Specifically, after calculating the conditional distribution alignment loss using feature anchors and label information distribution, the prediction result of the classifier is used as the pseudo label to perform conditional distribution alignment of the source domain and the target domain.

[0012] Step 3: Generation and optimization of feature anchors. Specifically, after equally memorizing the feature distribution of all fault categories in the unbalanced data during the iteration process of different batches, the feature anchors are iteratively optimized through statistical measurements to form balanced feature anchors, providing balanced fault information for the next batch of learning to assist in the conditional distribution alignment of cross-domain data features. Technical Effects

[0013] The present invention utilizes unbalanced training data to continuously maintain balanced feature anchor points during the training phase; in the model iteration phase, the feature anchor points are statistically optimized for discriminability, and good and balanced fault features are provided in the next round of iteration; compared with the prior art, the present invention can effectively prevent the network from generating negative migration during the training process, and significantly improve the accuracy of civil aircraft fault diagnosis under unbalanced variable operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the present invention;

[0015] Figure 2 It is a schematic diagram of the fault diagnosis problem under variable working condition scenario;

[0016] Figure 3 A schematic diagram of the data imbalance problem existing in the existing methods when processing fault data of different working conditions;

[0017] Figure 4 Schematic diagram of the generation and optimization of feature anchors in the model training process and their application in cross-domain fault diagnosis;

[0018] Figure 5 Select the structure and parameters of the backbone network;

[0019] Figure 6 The distribution and visualization results of the embodiments in the feature space. DETAILED DESCRIPTION

[0020] This embodiment is aimed at Figure 2 The problem of fault diagnosis model failure caused by the distribution difference between the source domain and the target domain in the variable working condition scenario is solved by adopting a civil aircraft fault diagnosis method based on feature anchor point alignment in the variable working condition scenario, such as Figure 1 As shown, the specific civil aircraft fault diagnosis method includes the following steps:

[0021] Step 1) Align the edge distribution of the source domain and the target domain to prevent the failure of the civil aircraft fault diagnosis model due to distribution differences. Specifically, according to the output of the fully connected layer before the bottleneck layer of the neural network, calculate the maximum mean difference (MMD) of the edge distribution of the labeled civil aircraft operation data as the source domain and the unlabeled civil aircraft operation data as the target domain. s ),G(x t )), where: G(.) is the feature extractor, and are source domain samples and their corresponding labels respectively.

[0022] The maximum mean difference (MMD) refers to the maximum mean difference between two probability distributions p and q in the reconstructed Hilbert space (RKHS). Among them: kernel function is a mapping from original data to RKHS, is the mathematical expectation of the probability distribution p, is the mathematical expectation of the probability distribution q, To reconstruct the Hilbert space, all variables are in the reconstructed Hilbert space;

[0023] In this embodiment, the kernel function is a Gaussian kernel function: Where: x i ,x j is the civil aircraft failure data, σ is the variance of Gaussian distribution, and exp is the natural exponential function;

[0024] Step 2) Target Figure 3 The traditional distribution alignment method is prone to negative transfer. Based on the alignment of source domain and target domain data in step 1, feature anchor-assisted cross-domain conditional distribution alignment is adopted to further avoid the phenomenon of fault feature confusion caused by unbalanced variable condition fault data, thereby eliminating negative transfer. Figure 4 Shown: Calculate the maximum mean difference between the initial feature anchor and the target domain data after pseudo label marking Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, is a pseudo label, i is a pseudo label category, x t The target domain data.

[0025] The initial feature anchor point refers to: Where: G(·) is the feature extractor, x s is the source domain data, y is the true label, and c is the true label category. The initial anchor point means the mean of the source domain feature data.

[0026] The negative transfer refers to the problem that due to insufficient or missing sampling in neural network training, the characteristics of occasional faults are confused with the characteristics of frequent faults, which ultimately leads to a decline in the performance of the civil aircraft variable operating condition fault diagnosis model.

[0027] The pseudo-label is the prediction result of the classifier on the target domain data. Among them: F(.) is the classifier, G(·) is the feature extractor, x t The target domain data.

[0028] The supervised learning training objective of the classifier F(.) is Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, x sis the source domain data, y is the true label, and c is the true label category; by optimizing the loss function to learn the source domain annotated data knowledge, a high confidence prediction result is given to the target domain data.

[0029] Step 3) Iterate and optimize feature anchors: After generating initial feature anchors of different types of faults using all fault data under all working conditions, iterate the distribution of feature anchors based on statistical measurement methods to enhance their discriminability, and then use the optimized and balanced feature anchors to assist the network in cross-domain migration, including:

[0030] 3.1 The training batch generated by random sampling of the unbalanced source domain fault training set under variable working conditions is expressed as in: is the training set that follows the distribution of source domain data, M is the number of batches, and B m is the total data structure of the mth batch, x s and y are the samples in the batch and their corresponding fault labels, with a total of c types of fault states (including healthy states, occasional faults, and frequent faults);

[0031] 3.2 The fault dataset of the target domain is represented as in: For a training set that follows the target domain distribution, the fault labels are unknown.

[0032] 3.3 During the training process of the mth batch, the feature anchors are updated by combining the data features in this batch with the feature anchors in the previous batch. Specifically, the feature anchors generated by the n fault data features with label c in the mth batch are Where: G(·) is the feature extractor, x s is the source domain data, y is the true label, c is the true label category, The feature anchor generated for the fault data feature with label c in the m-1th batch; when Or when m=1, the initial feature anchor point is

[0033] 3.4 After one data iteration (epoch), the final feature anchor will equally integrate all fault information in the training data, that is, the set of feature anchors that perceive global information and are balanced is

[0034] 3.5 Calculate the maximum mean difference between feature anchors of different labels based on the generated feature anchors: Calculate the loss function The feature anchor points for discriminative enhancement can be obtained through gradient descent Finally, we get a set of feature anchor points with enhanced discriminability. in: The feature anchor points generated for the n fault data features with labels i, j in the mth batch, is the gradient operator, It is the feature anchor point obtained after one data iteration. This step is not performed through a neural network, and the time cost is almost negligible.

[0035] 3.6 Align cross-domain samples with the same category labels in the feature space and separate cross-domain samples with different category labels: Although the training set data is unbalanced variable operating condition data, the generated feature anchor points are balanced, which will provide balanced fault information for the next round of batch learning to assist in the conditional distribution alignment of cross-domain data features, that is, optimize Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, x s is the source domain data, y is the true label, x t is the target domain data, and F(.) is the classifier.

[0036] After specific experiments, the planetary gearbox dataset of Southeast University, the bearing dataset of Case Western Reserve University, and the fault dataset of the overall drive generator of civil aircraft were used. The input vibration signal was uniformly sampled into 1×1024 in the Pytorch environment, and the Adam optimizer was selected for parameter optimization. In order to reduce the impact of randomness, the result was taken as the average of five repeated experiments. The backbone network model structure is a convolutional neural network, such as Figure 5 shown.

[0037] The overall experimental process includes hyperparameter optimization, fault diagnosis of planetary gearbox data from Southeast University, fault diagnosis of bearing data from Case Western Reserve University, and fault diagnosis of overall drive generator data from civil aircraft, as follows:

[0038] The Southeast University planetary gearbox data set contains 9 fault operation data of the planetary gearbox under two speed load ratios of 20HZ-0V and 30HZ-2V, and the specific content is shown in Table 1.

[0039] Table 1: Description of gearbox fault types in the Southeast University planetary gearbox dataset

[0040] The cross-domain task division method of the Southeast University planetary gearbox dataset is as follows: the "domains" representing different working conditions are set according to the bearing load and speed and are represented by capital letters, as shown in Table 3-4. In the experiment, two cross-domain diagnosis tasks can be constructed based on different working conditions, and the arrow "→" indicates domain migration. For example, "A→B" means that the fault data under working condition A in the cross-domain diagnosis task is labeled source domain data, while the fault data under working condition B is unlabeled target domain data. The cross-domain fault diagnosis algorithm needs to learn domain common knowledge through samples in the source domain and the target domain, and finally verify the accuracy of the diagnosis on the target domain. As shown in Table 2.

[0041] Table 2: Division method of the Southeast University planetary gearbox dataset domain

[0042] The Case Western Reserve University bearing data set records vibration data of 9 fault states under 4 working conditions (the speed load ratios are 1797rpm-0HP, 1772rpm-1HP, 1750rpm-2HP, and 1730rpm-3HP, respectively), as shown in Table 3.

[0043] Table 3: Description of shaft fault types in Case Western Reserve University bearing dataset

[0044] The method for dividing the domain of the Southeast University planetary gearbox dataset is shown in Table 4.

[0045] Table 4: Division method of the Southeast University planetary gearbox dataset domain

[0046] The civil aircraft integral drive generator fault data set refers to: an unbalanced variable operating condition fault data set constructed based on the real IDG operating data provided by Juneyao Airlines. This data set collects the oil temperature data of the integral drive engines of multiple different types of passenger aircraft from healthy operation to failure, and continued operation after the failure is repaired. The IDG full life cycle data is labeled according to the time nodes of failure and repair: 1) Data 10 days before the occurrence of the frequency deviation failure is at risk of failure (nearly broken); 2) Data when the frequency deviation failure occurs is a failure (broken); 3) Data after replacing or repairing the generator is in a healthy state (health); 4) Data when a low oil pressure alarm occurs is in a low oil pressure state (lowpressure). The specific content of the divided IDG fault data set is shown in Table 3. Since the power requirements of the aircraft's onboard electrical equipment loads will vary greatly with the flight phase, and the flight speed, temperature, pressure and other environments in different flight phases will also affect their working environment, the present invention briefly divides the civil aircraft IDG data of the entire flight segment into three working phases with large differences in power requirements and environments according to the flight phase: 1) take-off phase; 2) cruising phase; 3) landing phase, in which the climb, descent, approach and other phases are divided into adjacent flight phases in the same proportion. The data set description is shown in Table 5.

[0047] Table 5: Description of the data set of real civil aircraft integral drive generator operation

[0048] The division method of the data set domain of the civil aircraft integral drive generator fault data set is shown in Table 6.

[0049] Table 6: Division method of the domain of the civil aircraft integral drive generator operation data set

[0050] The hyperparameters include batch size, learning rate, epoch number and loss weight of the migration module. The optimal values ​​of each hyperparameter selected according to the experiment are shown in Table 7.

[0051] Table 7: Hyperparameter settings for deep learning networks

[0052] The experiment for determining the optimal batch size is as follows: on the planetary gearbox data set of Southeast University, multiple algorithms with a batch size range of 16-512 are compared, including: Multiple Kernel Maximum Mean Discrepancies (MK-MMD), Joint Maximum Mean Discrepancies (JMMD), Domain Adversarial Neural Network (DANN), Correlation Alignment (CORrelation Alignment, CORAL), Conditional Adversarial Domain Adaptation (CDAN) and Convolutional Neural Network (CNN), among which CNN is the benchmark algorithm adopted by the present invention and the results are shown in Table 8.

[0053] Table 8: Comparison of diagnostic accuracy of algorithms under different batch sizes

[0054] The experiment for determining the optimal training rounds is: on the Southeast University planetary gearbox dataset, a variety of algorithms with a training round range of 75-600 are compared, including: multi-core maximum mean difference method, joint maximum mean difference method, domain adversarial neural network correlation alignment method, conditional domain adversarial neural network and convolutional neural network. The results are shown in Table 9.

[0055] Table 9: Comparison of diagnostic accuracy of algorithms under different training rounds

[0056] The experiment for determining the optimal learning rate is: on the Southeast University planetary gearbox data set, a variety of algorithms with a learning rate range of 0.0001-0.1 are compared, including: multi-core maximum mean difference method, joint maximum mean difference method, domain adversarial neural network correlation alignment method, conditional domain adversarial neural network and convolutional neural network. The results are shown in Table 10.

[0057] Table 10: Comparison of diagnostic accuracy of algorithms under different learning rates

[0058] The experiment for determining the optimal migration loss weight is: on the Southeast University planetary gearbox data set, a variety of algorithms with a migration loss weight range of 0.1-1.0 are compared, including: multi-core maximum mean difference method, joint maximum mean difference method, domain adversarial neural network correlation alignment method, conditional domain adversarial neural network and convolutional neural network. The results are shown in Table 11.

[0059] Table 11: Comparison of diagnostic accuracy of algorithms under different loss weights

[0060] Under the optimal hyperparameters, the present invention is compared with the multi-core maximum mean difference method, the joint maximum mean difference method, the domain adversarial neural network correlation alignment method, the conditional domain adversarial neural network and the convolutional neural network method on the planetary gearbox fault dataset of Southeast University and the bearing fault dataset of Case Western Reserve University, and is verified on the civil aircraft integral drive generator dataset composed of real civil aircraft operation data provided by Juneyao Airlines. It is proved that when dealing with unbalanced civil aircraft fault datasets, the performance of cross-domain fault feature extraction of the present invention is better than other popular deep transfer learning methods, and the fault diagnosis accuracy is higher than other popular deep transfer learning methods.

[0061] The comparative experiment of the Southeast University planetary gearbox fault data set under the optimal hyperparameters is shown in Table 12.

[0062] Table 12: Variable operating condition fault diagnosis results based on the Southeast University planetary gearbox fault data set

[0063] The comparative experiment of the Case Western Reserve University bearing fault dataset under the optimal hyperparameters is shown in Table 13.

[0064] Table 13: Variable-condition fault diagnosis results based on Case Western Reserve University bearing fault dataset

[0065] The feature extraction performance comparison experiment of each deep transfer learning method under the optimal hyperparameters is: select task C→A from the Case Western Reserve University bearing data set for visualization experiment. Randomly select 200 fault samples from the target domain of the task and extract their feature information in the iterative process, and use t-SNE to draw a two-dimensional feature representation of the feature information, such as Figure 6 As shown in the figure, each point represents a fault sample, and points of different colors represent different health states.

[0066] The verification experiment on the civil aircraft integral drive generator dataset under the optimal hyperparameters is shown in Table 14.

[0067] Table 14: Fault diagnosis results based on the operating data of the overall drive generator of a civil aircraft under variable operating conditions

[0068] Compared with the prior art, the present invention fully demonstrates the enhancement of the discriminability of fault diagnosis features and the improvement of fault diagnosis model performance in variable operating conditions by performing relevant algorithm performance comparison and visualization experiments in two public data sets. Finally, the feasibility of the present invention is verified based on real civil aircraft IDG data, and the experiment verifies that the present invention can estimate and diagnose the overall drive generator status of civil aircraft in different flight phases based on the IDG operation data of fault unbalanced coupling.

[0069] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions, characterized in that: According to the output of the fully connected layer before the bottleneck layer of the neural network, the maximum mean difference of the marginal distribution of the labeled civil aircraft operation data as the source domain and the unlabeled civil aircraft operation data as the target domain is calculated to align the marginal distribution of the source domain with that of the target domain. Then, the cross-domain conditional distribution alignment based on feature anchors is adopted to avoid the confusion of fault features caused by the imbalance of variable operating condition fault data, thereby eliminating negative migration. Finally, all fault data under all operating conditions are used to generate initial feature anchors of different types of faults, and the distribution of feature anchors is iterated based on statistical measurements to enhance its discriminability. The network is assisted in cross-domain migration through the optimized and balanced feature anchors.

2. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 1 is characterized in that: The neural network includes: a feature extractor and a classifier, the classifier is composed of a bottleneck layer in a backbone network, and the rest of the backbone network is a feature extractor, wherein: the feature extractor performs feature extraction based on the input source domain and target domain civil aircraft operation data to obtain complex fault features, and the classifier performs dimensionality reduction processing based on feature information in a bottleneck layer manner, maps the calculated feature space to the sample label space, and obtains a fault classification result.

3. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 1 is characterized in that: The fully connected layer before the bottleneck layer of the neural network is composed of fully connected neurons, and a complex nonlinear function fitting structure is constructed using an activation function. The fully connected layer integrates the local fault features extracted by each convolutional layer in a feature weighted manner to obtain complete global fault feature information.

4. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 1 is characterized in that: The neural network is trained in the following way: Step 1: Cross-domain edge distribution alignment. Specifically, after the fault data of the source domain and the target domain are input into the feature extractor respectively, the cross-domain edge distribution alignment loss is calculated using the maximum mean difference, and then the data differences in different flight stages are eliminated. Step 2: Cross-domain conditional distribution alignment is assisted by feature anchors. Specifically, after calculating the conditional distribution alignment loss using feature anchors and label information distribution, the prediction result of the classifier is used as a pseudo label to perform conditional distribution alignment between the source domain and the target domain. Step 3: Generation and optimization of feature anchors. Specifically, after equally memorizing the feature distribution of all fault categories in the unbalanced data during the iteration process of different batches, the feature anchors are iteratively optimized through statistical measurements to form balanced feature anchors, providing balanced fault information for the next batch of learning to assist in the conditional distribution alignment of cross-domain data features.

5. The method for civil aircraft fault diagnosis based on feature anchor point alignment under variable operating conditions according to any one of claims 1 to 4, characterized in that: include: Step 1) Align the marginal distribution of the source domain with that of the target domain. Specifically, according to the output of the fully connected layer before the bottleneck layer of the neural network, calculate the maximum mean difference (MMD) of the marginal distribution of the labeled civil aircraft operation data as the source domain and the unlabeled civil aircraft operation data as the target domain. s ),G(x t )), where: G(.) is the feature extractor, and They are source domain samples and their corresponding labels respectively; Step 2) Based on the alignment of source domain and target domain data in step 1, feature anchor-assisted cross-domain conditional distribution alignment is used to further avoid the phenomenon of fault feature confusion caused by imbalanced fault data under variable working conditions, thereby eliminating negative migration. Specifically, the maximum mean difference between the initial feature anchor and the target domain data marked with pseudo labels is calculated. Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, is a pseudo label, i is a pseudo label category, x t is the target domain data; Step 3) Iterate and optimize feature anchors: After generating initial feature anchors of different types of faults using all fault data under all working conditions, iterate the distribution of feature anchors based on statistical measurement methods to enhance their discriminability, and then use the optimized and balanced feature anchors to assist the network in cross-domain migration, including: 3.1 The training batch generated by random sampling of the unbalanced source domain fault training set under variable working conditions is expressed as in: is the training set that follows the distribution of source domain data, M is the number of batches, and B m is the total data structure of the mth batch, x s and y are the samples in the batch and their corresponding fault labels, and c is the number of fault status categories; 3.2 The fault dataset of the target domain is represented as in: For a training set that follows the target domain distribution, the fault labels are unknown; 3.3 During the training process of the mth batch, the feature anchors are updated by combining the data features in this batch with the feature anchors in the previous batch. Specifically, the feature anchors generated by the n fault data features with label c in the mth batch are Where: G(·) is the feature extractor, x s is the source domain data, y is the true label, c is the true label category, The feature anchor generated for the fault data feature with label c in the m-1th batch; when Or when m=1, the initial feature anchor point is 3.4 After one data iteration (epoch), the final feature anchor will equally integrate all fault information in the training data, that is, the set of feature anchors that perceive global information and are balanced is 3.5 Calculate the maximum mean difference between feature anchors of different labels based on the generated feature anchors: Calculate the loss function The feature anchor points for discriminative enhancement can be obtained through gradient descent Finally, we get a set of feature anchor points with enhanced discriminability. in: The feature anchor points generated for the n fault data features with labels i, j in the mth batch, is the gradient operator, It is the feature anchor point obtained after one data iteration; 3.6 Align cross-domain samples with the same category labels in the feature space and separate cross-domain samples with different category labels: Optimize Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, x s is the source domain data, y is the true label, x t is the target domain data, and F(.) is the classifier.

6. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 5 is characterized in that: The maximum mean difference (MMD) refers to the maximum mean difference between two probability distributions p and q in the reconstructed Hilbert space (RKHS). Among them: kernel function is a mapping from original data to RKHS, is the mathematical expectation of the probability distribution p, is the mathematical expectation of the probability distribution q, To reconstruct the Hilbert space, all variables are in the reconstructed Hilbert space; The kernel function is a Gaussian kernel function: Where: x i ,x j is the civil aircraft failure data, σ is the variance of Gaussian distribution, and exp is the natural exponential function.

7. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 5 is characterized in that: The initial feature anchor point refers to: Where: G(·) is the feature extractor, x s is the source domain data, y is the true label, c is the true label category, and the initial anchor point means the mean of the source domain feature data.

8. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 5 is characterized in that: The negative transfer refers to the problem that due to insufficient or missing sampling in neural network training, the characteristics of occasional faults are confused with the characteristics of frequent faults, which ultimately leads to a decline in the performance of the civil aircraft variable operating condition fault diagnosis model.

9. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 5 is characterized in that: The pseudo-label is the prediction result of the classifier on the target domain data. Among them: F(.) is the classifier, G(·) is the feature extractor, x t The target domain data.

10. The civil aircraft fault diagnosis method based on feature anchor point alignment under variable operating conditions according to claim 5, characterized in that: The supervised learning training objective of the classifier F(.) is Among them: A i is the feature anchor of category i in the previous round, G(·) is the feature extractor, x s is the source domain data, y is the true label, and c is the true label category; by optimizing the loss function to learn the source domain annotated data knowledge, a high confidence prediction result is given to the target domain data.

Images