A method for fault diagnosis of aircraft engine bearings based on domain generalization

Abstract

The present invention discloses a method for fault diagnosis of aircraft engine bearings based on domain generalization, comprising: obtaining bearing data to be diagnosed; inputting the bearing data to be diagnosed into a fault diagnosis model to obtain a fault diagnosis result; wherein the fault diagnosis model comprises a feature extraction module and a condition comparison feature module, the feature extraction module and the condition comparison module are trained respectively using a training set, the feature extraction module and the condition comparison module are updated in combination with a corresponding loss function, and the fault diagnosis model is obtained, and the training set comprises bearing working data. The present invention utilizes domain generalization technology to perform fault diagnosis on aircraft engine bearings, and does not need to separately collect samples for each working condition of the aircraft engine and perform labeling operations, thus saving a lot of manpower and material resources.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and particularly relates to a fault diagnosis method for aero-engine bearings based on domain generalization. Background Art

[0002] As a core component of an aircraft, the stable operation of an aero-engine is crucial for ensuring flight safety. During the operation of the engine, the rolling bearing is a key component that bears a huge load. Its working environment is harsh and it is extremely prone to failure, which may lead to serious flight accidents. Therefore, in-depth research on the fault diagnosis method of aero-engine bearings helps to timely detect the characteristics of early bearing faults, thereby effectively preventing accidents. Therefore, it is necessary to provide a fault diagnosis method for aero-engine bearings to solve the above technical problems;

[0003] The following defects exist in the prior art: 1. In the prior art, the basic assumption that the test data and the training data are independently and identically distributed is followed. Once this assumption is broken, the diagnostic accuracy of the existing methods will drop significantly; 2. The existing methods all require collecting a sufficient number of healthy state labels. However, in the actual working environment of an aero-engine, it is impossible to collect enough data and mark enough labels for each task and each working condition; 3. There are domain adaptation-based methods to alleviate the above problems, but the domain adaptation methods can only solve the problem from one source domain to one target domain, while the working conditions of aero-engines vary greatly; 4. Applying domain adaptation technology can only perform offline fault diagnosis and cannot perform online fault diagnosis on the current operating data; 5. The existing domain generalization technologies do not extract domain discriminant information well, resulting in the generalization performance not meeting the requirements; 6. The existing domain generalization technologies extract domain-invariant features based on the method of measuring the difference in feature distributions, and this method sometimes extracts information irrelevant to the task and ignores the differences between different classes, resulting in poor quality of domain-invariant features and affecting domain generalization. Therefore, there is an urgent need for a fault diagnosis method for aero-engine bearings based on domain generalization. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a fault diagnosis method for aero-engine bearings based on domain generalization, which can use domain generalization technology to diagnose the faults of aero-engine bearings, without the need to separately collect samples and perform tagging operations for each working condition of the aero-engine, saving a large amount of manpower and material resources.

[0005] To achieve the above object, the present invention provides a fault diagnosis method for aero-engine bearings based on domain generalization, including:

[0006] Obtain the bearing data to be diagnosed;

[0007] Input the bearing data to be diagnosed into the fault diagnosis model to obtain the fault diagnosis result;

[0008] Among them, the fault diagnosis model includes a feature extraction module and a conditional contrast feature module. The feature extraction module and the conditional contrast module are respectively trained using the training set, and combined with the corresponding loss function, the feature extraction module and the conditional contrast module are updated to obtain the fault diagnosis model. The training set includes bearing working data.

[0009] Optionally, the bearing working data includes: bearing speed, bearing load, and acceleration signal.

[0010] Optionally, before training the fault diagnosis model using the training set, it further includes: annotating the training set to obtain the annotation categories;

[0011] Among them, the annotation categories are: fault type and fault size.

[0012] Optionally, the feature extraction module includes: a first feature extraction layer and a first classifier;

[0013] The first feature extraction layer is used to obtain multi-source domain features using a feature extractor;

[0014] The first classifier is used to adjust the multi-source domain features using an improved cross-entropy loss function to obtain the cross-entropy loss.

[0015] Optionally, the improved cross-entropy loss function is:

[0016]

[0017] Among them, L CCE is the total cross-entropy loss in cross-domain learning, is the j-th sample obtained from the i-th source domain, is its corresponding label, is the output after the observation of the sample passes through the feature extraction layer and the classification layer, L 2 is the cross-entropy loss function, M is the number of different source domains, N i is the number of samples in the i-th source domain.

[0018] Optionally, the conditional contrast module includes: a second feature extraction layer and a second classifier;

[0019] The second feature extraction layer is used to obtain sample features using a feature extractor;

[0020] The second classifier is used to adjust the sample features using a conditional contrast function to obtain the conditional contrast loss.

[0021] Optionally, the conditional contrast function is:

[0022]

[0023] where is the conditional contrast loss, is the specific loss for a single sample u, M is the number of different source domains, and N i is the number of samples in the i-th source domain.

[0024] Optionally, updating the feature extraction module and the conditional contrast module includes:

[0025] Based on the cross-entropy loss and the conditional contrast loss, and based on the SGD optimization algorithm, update the feature extractor and the classifier to obtain the updated feature extraction layer and classifier.

[0026] Optionally, the updated feature extraction layer and classifier are:

[0027]

[0028] where f θ is the feature extraction layer, ε is the learning rate, L cc is the conditional contrast loss, α is the weight, L CCE is the cross-entropy loss, g θ is the classifier, f θ ← is the updated feature extraction layer, g θ ← is the updated classifier.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] (1) When an aero-engine is operating, it has different speeds and different loads, and the data generated will have different degrees of domain shift phenomena. Traditional deep learning fault diagnosis algorithms are unable to perform fault diagnosis work in this situation. However, this method can perform high-accuracy fault diagnosis on the operation data of aero-engine bearings under multiple working conditions with only one training.

[0031] (2) Currently, there are domain adaptation-based methods that can alleviate part of the problem of missing data labels, but usually domain adaptation can only solve part of the problem of missing sample labels and cannot solve the problem of lack of training.

[0032] (3) The present invention can perform intelligent fault diagnosis in an online manner, that is, along with the data acquisition of the sensor, it can also reflect the health status of the aero-engine bearing in real time.

[0033] (4) The present invention utilizes the improved softmax loss and the health state condition comparison to achieve the optimal effect in the fault diagnosis under unknown working conditions.

[0034] (5) The fault diagnosis method of the present invention does not require manual participation, has high processing efficiency, reduces the data acquisition cost while improving the diagnosis accuracy, and greatly improves the safety and reliability of the engine bearing. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments and descriptions thereof of this application are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0036] Figure 1 is a flowchart of a fault diagnosis method for an aero-engine bearing based on domain generalization according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments may be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0038] It should be noted that the steps shown in the flowchart of the drawings may be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than here.

[0039] The present invention provides a fault diagnosis method for an aero-engine bearing based on domain generalization, as Figure 1 shown, specifically including:

[0040] Obtain the bearing data to be diagnosed;

[0041] Input the bearing data to be diagnosed into the fault diagnosis model to obtain the fault diagnosis result;

[0042] Among them, the fault diagnosis model includes a feature extraction module and a conditional comparison feature module. The feature extraction module and the conditional comparison module are respectively trained using the training set, and combined with the corresponding loss function, the feature extraction module and the conditional comparison module are updated to obtain the fault diagnosis model. The training set includes the bearing working data.

[0043] Specifically, the trained model can be applied to many new working conditions and can perform online identification. Traditional manual methods can only conduct regular inspections or intervene manually when obvious faults occur in aero-engine bearings, resulting in lag in fault diagnosis. Traditional deep learning methods cannot perform high-precision fault diagnosis when the working conditions change. The network trained based on domain generalization technology can detect operation signals in real time, greatly improving the working safety of aero-engines.

[0044] Furthermore, the bearing working data includes: bearing speed, bearing load, and acceleration signal.

[0045] Furthermore, before training the fault diagnosis model using the training set, it also includes: annotating the training set to obtain the annotation categories;

[0046] Among them, the annotation categories are: fault type and fault size.

[0047] Specifically, using domain generalization technology for fault diagnosis of aero-engine bearings does not require separately collecting samples and performing tagging operations for each working condition of the aero-engine, saving a large amount of manpower and material resources.

[0048] Furthermore, the feature extraction module includes: the first feature extraction layer and the first classifier;

[0049] The first feature extraction layer is used to obtain multi-source domain features using a feature extractor;

[0050] The first classifier is used to adjust the multi-source domain features using an improved cross-entropy loss function to obtain the cross-entropy loss.

[0051] Specifically, the feature extraction layer uses a one-dimensional convolutional neural network to extract features of multi-source domains, encodes the aero-engine bearing data into feature representations, and the feature extractor is denoted as f θ . The classifier uses an improved cross-entropy loss function to readjust the extracted features to make the intra-class distance closer and the inter-class distance farther, obtaining discriminative features, converting the features into prediction scores, and selecting the highest score as the prediction label. The classifier is denoted as g θ , calculates the task-specific cross-entropy loss L between the prediction and the true label CCE . Using the conditional contrast loss L cc maximizes the intra-class similarity and inter-class separability between different domains. At the same time, optimize L CCE and L cc to train the feature extractor.

[0052] Among them, the one-dimensional convolutional neural network contains a total of five convolutional modules. The first four convolutional modules include a convolutional layer, a pooling layer, and a weight normalization layer. The last two layers are fully connected layers, which are denoted as L 1 ,L2 。

[0053] The improved cross-entropy loss function is NormFace, and its formula is shown in (1):

[0054]

[0055] where L 2 is the feature of the last fully-connected layer, s is the scaling factor, and cos(θ yi ) is the angle between the feature and its corresponding class. Compared with the traditional Softmax loss, this loss function can significantly reduce the intra-class distance and simultaneously increase the inter-class distance, resulting in an obvious margin between different health states, which is beneficial to the recognition of domain generalization. The final improved cross-entropy loss function is shown in formula (2):

[0056]

[0057] where L CCE is the cross-entropy loss, is the j-th sample obtained from the i-th source domain, is its corresponding label, is the output after the observation of the sample passes through the feature extraction layer and the classification layer, L 2 is the feature of the last fully-connected layer, M is the number of source domains, and N i is the number of samples in the i-th source domain.

[0058] Furthermore, the conditional contrast module includes: a second feature extraction layer and a second classifier;

[0059] The second feature extraction layer is used to obtain sample features by using a feature extractor;

[0060] The second classifier is used to adjust the sample features by using a conditional contrast function to obtain a conditional contrast loss.

[0061] Specifically, samples with the same health state label in different source domains are set as positive samples, and other samples are set as negative samples. The purpose of this model is to maximize the mutual information between positive samples and simultaneously minimize the similarity between negative samples.

[0062] The advantage of this setting is that it avoids the situation in the measurement of feature distribution differences where, even if the feature distribution differences are well aligned, the distances between samples belonging to different classes are too close, resulting in poor performance of the domain generalization network. The proposed method only pulls samples belonging to the same health state together, enabling the model to find class-conditional invariant representations in multiple source domains. Its loss function is defined as formula (3):

[0063]

[0064] where ∣pos(u)∣ is the number of positive samples, and σ(a,b) = (a T b / |a||b|) is the similarity function of the given vectors a and b, and τ is the temperature function used to scale the contrast result. where g θ ⊙f θ represents the combination of the feature extractor and the classifier. represents the contrast loss with respect to the anchor sample x u . Therefore, the contrast loss of all samples is calculated as follows:

[0065]

[0066] where is the conditional contrast loss. is the specific loss function for a single sample u, M is the number of source domains, and N i is the number of samples in the i-th source domain. By minimizing , the lower bound of the mutual information between positive samples can be maximized. The theoretical formula of the mutual information can be described by formula (5):

[0067]

[0068] where p(a,b) is the joint distribution of (a,b), and p(a)p(b) is the product of the marginal distributions. The specific derivation is as follows:

[0069]

[0070] Initialize with a normal distribution according to the Xavier initialization method to avoid the problem of gradient vanishing or gradient explosion during training, and at the same time enable the signal to be transmitted deeper in the neural network and still remain within a reasonable range after passing through multiple neurons. As shown in formula (7):

[0071]

[0072] Furthermore, the update of the feature extraction module and the conditional contrast module includes:

[0073] Based on the cross-entropy loss and the conditional contrast loss, and based on the SGD optimization algorithm, update the feature extractor and the classifier to obtain the updated feature extraction layer and classifier.

[0074] Specifically, as above, the total optimization objective of the model is to jointly optimize and to improve the generalization diagnosis performance of the model on the operation data under unknown working conditions. Therefore, the total optimization objective is as shown in formula (8):

[0075]

[0076] Furthermore, the updated feature extraction layer and classifier are as follows:

[0077]

[0078] Among them, f θ is the feature extraction layer, ε is the learning rate, L cc is the conditional contrast loss, α is the weight balancing factor, L CCE is the cross-entropy loss, g θ is the classifier, f θ ← is the updated feature extraction layer, g θ ← is the updated classifier.

[0079] Input the operation data under the new working conditions into the above trained model, set the model to the test mode, set the gradient update to 0, and judge the prediction result of the model.

[0080] The innovations obtained through this embodiment include:

[0081] (1) For the traditional manual discrimination method, a large amount of prior knowledge and human input are required, and the identification usually has a lag. Using the traditional deep learning method, it is necessary to collect the operation data under all working conditions and label them. Although prior knowledge and a large amount of human input are not required, the cost of collecting labels and operation data is still expensive. Using the domain generalization technology for fault diagnosis of aero-engine bearings does not require collecting samples and labeling them separately for each working condition of the aero-engine, saving a large amount of manpower and material resources.

[0082] (2) The trained model can be applied to many new working conditions and can perform online identification. The traditional manual method can only perform regular inspections or intervene manually when obvious faults occur in the aero-engine bearings, and its fault diagnosis has a lag. For the traditional deep learning method, when the working condition changes, high-precision fault diagnosis cannot be performed. The network trained based on the domain generalization technology can detect the operation signal in real time, greatly improving the working safety of the aero-engine.

[0083] (3) By using the improved softmax loss function, NormFace is used to adjust the classification features, ensuring that there is a certain margin between the classification boundaries, clearly promoting the compact representation of features, and being beneficial to the improvement of domain generalization performance.

[0084] (4) By introducing the healthy state conditional contrast loss to measure the similarity between classes and avoiding using the maximum mean difference, deep correlation contrast or adversarial methods to compare feature differences, it can ensure that the similar class features between different domains are closer, further improving the generalization performance of aero-engine bearing fault diagnosis.

[0085] The above are only the preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for fault diagnosis of aircraft engine bearings based on domain generalization, characterized in that: include: Obtain the bearing data to be diagnosed; Inputting the bearing data to be diagnosed into a fault diagnosis model to obtain a fault diagnosis result; Wherein, the fault diagnosis model includes a feature extraction module and a condition comparison feature module, the feature extraction module and the condition comparison module are trained respectively using a training set, and the feature extraction module and the condition comparison module are updated in combination with a corresponding loss function to obtain the fault diagnosis model, and the training set includes bearing working data; The feature extraction module includes: a first feature extraction layer and a first classifier; The first feature extraction layer is used to obtain multi-source domain features using a feature extractor; The first classifier is used to adjust the multi-source domain features by using an improved cross entropy loss function to obtain a cross entropy loss; The improved cross entropy loss function is: Among them, L CCE is the cross entropy loss function in cross-domain fault diagnosis, is the jth sample obtained from the i-th source domain, For its corresponding label, is the output of the sample observation after the feature extraction layer and the classification layer, L2 is the feature of the last fully connected layer, M is the number of different source domains, representing M source domains, N i is the number of samples in the i-th source domain; The condition comparison module includes: a second feature extraction layer and a second classifier; The second feature extraction layer is used to obtain sample features using a feature extractor; The second classifier is used to adjust the sample features by using a conditional contrast function to obtain a conditional contrast loss; The conditional comparison function is: in, is the conditional contrast loss, is the specific loss function for a single sample u, M is the number of source domains, N i is the number of samples in the i-th source domain; Updating the feature extraction module and the condition comparison module includes: Based on the cross entropy loss and conditional contrast loss and the SGD optimization algorithm, the feature extractor and classifier are updated to obtain the updated feature extraction layer and classifier; The updated feature extraction layer and classifier are: Among them, f θ is the special diagnosis extraction layer, ε is the learning rate, L cc is the conditional contrast loss, α is the weight balancing factor, and L CCE is the cross entropy loss, g θ is the classifier, f θ ← is the updated feature extraction layer, g θ ← is the updated classifier.

2. The method for diagnosing aircraft engine bearing faults based on domain generalization according to claim 1, characterized in that: The bearing operating data includes: bearing speed, bearing load and acceleration signal.

3. The method for diagnosing aircraft engine bearing faults based on domain generalization according to claim 2, characterized in that: Before using the training set to train the fault diagnosis model, the method further includes: labeling the training set to obtain a labeling category; The marking categories are: fault type and fault size.

Images