Data-model dual-drive intershaft bearing double-rotor bolt connection diagnosis method

Abstract

The invention belongs to the technical field of aero-engine structure health monitoring and fault diagnosis, and particularly relates to a data-model dual-drive intermediate bearing double-rotor bolt connection diagnosis method, which comprises the steps of simulation data generation through a physical model, data normalization and data set construction, model pre-training, transfer learning and diagnosis. By combining data driving and model driving methods, the problem of lack of marked fault data is solved, the rotor bolt connection state can be accurately diagnosed under the condition of no pre-marked fault data, and the problem that an aero-engine rotor system is difficult to diagnose due to lack of marked fault data is effectively solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft engine structural health monitoring and fault diagnosis, and in particular relates to a data-model dual-driven intermediate bearing double-rotor bolt connection diagnosis method. Background Art

[0002] In recent years, with the rapid development of big data and artificial intelligence technologies, data-driven fault diagnosis methods have gradually become a research hotspot. In mechanical system fault diagnosis, the common practice is to collect a large number of actual fault cases or generate labeled data through pre-fabricated fault tests to train the model and enable it to identify faults. Deep learning-based algorithms, in particular, often assume the availability of sufficient labeled data. However, in real-world applications, most data is unlabeled for the following reasons: 1) Relatively more healthy data but less fault data: Most mechanical systems operate in a normal healthy state, with a low frequency of faults. This results in a large amount of collected data being mostly normal operation data, but lacking sufficient fault data.

[0003] 2) High cost of label acquisition: Especially in complex mechanical systems, obtaining labels corresponding to fault data requires a lot of resources and costs.

[0004] To address these issues, few-shot learning methods have been introduced. These methods generate a small amount of labeled data by prefabricating faults, and then use few-shot learning algorithms to extract features from this data. However, for aircraft engine rotor systems, labeled data is extremely scarce, and there may even be no relevant data at all. Furthermore, prefabricating faults in aircraft engines is prohibitively expensive and poses safety risks, making it impractical. Therefore, few-shot methods still have significant limitations for this type of problem.

[0005] This type of problem, lacking labeled fault data, falls under the category of zero-shot learning. Unlike traditional supervised learning, zero-shot learning aims to classify data without labels. This paper aims to design a data-model-driven zero-shot diagnosis method. By leveraging a physical model to provide the characteristics of structural vibration response, combined with a data-driven algorithm, it enables diagnosis of rotor bolt connection conditions without any real labeled fault data. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method to solve the problem of bolt connection status diagnosis of aircraft engine rotor system in the absence of labeled fault data.

[0007] The present invention is implemented by providing a data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method, comprising the following steps: Step S0, generating simulation data using the physical model: establishing a finite element model of a dual-rotor with intermediate bearings with multiple possible bolt loosening states and normal states, and generating labeled simulation data through simulation. To enhance the generalization ability of the data, a domain randomization method is used to randomize key parameters during the simulation process to simulate the dispersion and uncertainty of the actual dual-rotor with intermediate bearings; Step S1, data normalization and dataset construction: Normalize the frequency components of the simulation data to eliminate the absolute magnitude effects caused by different excitation amplitudes or imbalances, obtain a simulation dataset, divide the simulation dataset into a training set and a validation set, and simultaneously, use the same normalization parameters on unlabeled real test data to construct a real dataset; Step S2, model pre-training: Use the training set to perform preliminary training on the neural network model. During the preliminary training process, the CenterLoss method is used to further optimize the feature extractor so that it can better cluster similar samples and separate features of different categories. The trained neural network model is verified using the validation set. Step S3, transfer learning and diagnosis: the labeled simulation dataset is used as the source domain, and the unlabeled real dataset is used as the target domain. By introducing adversarial loss and local domain alignment strategy, the feature distribution of the source domain and the target domain is adjusted so that the neural network model can adapt to the unlabeled real dataset of the target domain. Finally, the optimal parameters are obtained, and the neural network model after transfer learning is obtained. The neural network model after transfer learning is used to accurately diagnose the bolt connection status of the real intermediate bearing dual rotor.

[0008] Preferably, in step S0, the dynamic equation corresponding to the intermediate bearing dual-rotor finite element model is:

[0009] In the above formula, is the generalized displacement vector of the intermediate bearing dual-rotor system, 、 、 are the mass, damping and gyroscopic torque matrices, respectively, is the stiffness matrix without considering the asymmetry of the connection structure stiffness, is a time-varying function of stiffness, ε is an asymmetric coefficient indicating the degree of asymmetry of stiffness caused by bolt loosening, is the time-varying term caused by stiffness asymmetry, is the unbalanced excitation vector containing two speed excitation frequencies.

[0010] Preferably, in step S0, the specific operation of the domain randomization method is: when each data sample is generated by simulating the intermediate bearing dual-rotor finite element model, the key parameters are randomized within a certain range, and the key parameters include support stiffness, damping coefficient, stiffness asymmetry coefficient, unbalanced excitation size and phase, and the rotational speed of the two rotors.

[0011] Preferably, in step S1, the specific operation of normalizing the frequency components of the simulation data is: first, perform fast Fourier transform on the simulation data, then calculate the sum A of the amplitudes of the two rotor speed frequency components in the intermediate bearing dual-rotor system, divide the simulation data by A, so that the sum of the rotor fundamental frequency amplitudes is 1, and eliminate the response inconsistency caused by the excitation intensity.

[0012] Preferably, in step S2, the specific process of using CenterLoss to perform preliminary training on the neural network model is as follows: define is the feature vector extracted by the neural network model for the i-th data sample, is the center vector of category c, where d is the dimension of the feature vector. Assuming there are C categories, the goal of Centerloss is to minimize the Euclidean distance between the data sample features and their category centers:

[0013] Where N is the number of data samples in the training set, is a data sample Tags, Represents the Euclidean distance. CenterLoss and cross entropy loss are used together in training. The expression of cross entropy is:

[0014] in Is the neural network model for the data sample Belong to category The predicted probability of , the total loss function is: .

[0015] Preferably, in step S3, the specific operation of transfer learning is: first, for the source domain , target domain Through feature extractor Extracting high-level features and , then, the high-level features are input into the classifier and domain discriminator In the equation, we get the weight vector and classification loss and domain classification loss , the objective function of partial transfer learning is:

[0016] in, It is a hyperparameter used to weigh the loss of the source domain label classifier and the partial adversarial domain discriminator in the optimization problem. Finally, the optimal parameter needs to be found. , and : .

[0017] Compared with the prior art, the advantages of the present invention are: 1) By combining physical modeling with artificial intelligence technology, the system overcomes the challenge of lacking labeled fault data. It can accurately diagnose the rotor bolt connection status without pre-labeled fault data, effectively resolving the difficulty in diagnosing aircraft engine rotor systems due to the lack of labeled fault data. 2) By randomizing the physical model parameters, a source domain dataset covering various failure modes and normal operating conditions was constructed. A transfer learning method was introduced, which enabled high diagnostic accuracy to be maintained when dealing with actual rotor systems with uncertain parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flow chart of the data-model dual-driven intermediate bearing dual-rotor bolt connection diagnostic method provided by the present invention; Figure 2 Schematic diagram of a dual-rotor system with an intermediate bearing according to an embodiment of the present invention; Figure 3 Schematic diagram of the overall framework of the data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] Figure 1 A flow chart of the method of the present invention is provided. Figure 2As shown, the intermediate bearing dual rotor system of the present invention consists of a low-pressure rotor 1 and a high-pressure rotor 2, and the high-pressure rotor 2 is supported on the low-pressure rotor 1 through an intermediate bearing 3. The rotational frequency of the low-pressure rotor is denoted as f1, and the rotational frequency of the high-pressure rotor is denoted as f2. In the complete machine state, three vibration measuring points are provided, namely, the air intake casing measuring point 4, the intermediate casing measuring point 5, and the turbine casing measuring point 6. The high-pressure rotor 2 is provided with five rows of bolt connections, which are denoted as J1, J2, J3, J4, and J5 from front to back. This embodiment mainly identifies the bolt connection status of the high-pressure rotor 2.

[0021] The present invention provides a data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method, the basic framework is as follows Figure 3 The specific steps are as follows: Step S0, generating simulation data through the physical model: establishing a finite element model of a dual-rotor with intermediate bearings with multiple possible bolt loosening states and normal states, and the corresponding dynamic equation is:

[0022] In the above formula, is the generalized displacement vector of the intermediate bearing dual-rotor system, 、 、 are the mass, damping and gyroscopic torque matrices, respectively, is the stiffness matrix without considering the asymmetry of the connection structure stiffness, is a time-varying function of stiffness, ε is an asymmetric coefficient indicating the degree of asymmetry of stiffness caused by bolt loosening, is the time-varying term caused by stiffness asymmetry, is the unbalanced excitation vector containing two speed excitation frequencies.

[0023] Through simulation, labeled simulation data is generated, covering a variety of possible bolt loosening conditions and normal working conditions to construct a wide range of source domain data sets. When each data sample is generated through finite element model simulation, the key parameters are randomized within a certain range. The key parameters include: support stiffness, damping coefficient, stiffness asymmetry coefficient, unbalance excitation magnitude and phase, and the rotational speed of the two rotors. The random range of support stiffness is ±1% of the original design stiffness, the random range of damping coefficient is 0.01~0.05, the random range of stiffness asymmetry coefficient is 0.5~0.8, the magnitude of unbalance is 0~100gmm, the phase is randomized within 0~2pi, and the rotational frequencies of the low-pressure rotor and high-pressure rotor are 75Hz and 86Hz respectively, and are randomized within the range of ±0.5%.

[0024] Step S1: Data Normalization and Dataset Construction: After the simulation data is generated, a fast Fourier transform (FFT) is performed on the data. The sum of the amplitudes of the two rotor speed frequency components in the dual-rotor system is calculated as A. The simulation data is divided by A to equalize the sum of the rotor fundamental frequency amplitudes to 1, eliminating response inconsistencies caused by excitation intensity. Bolt connection states are categorized into six categories: loosened bolts in the first row (J1), loosened bolts in the second row (J2), loosened bolts in the third row (J3), loosened bolts in the fourth row (J4), loosened bolts in the fifth row (J5), and no loose bolts (J0). Labels are represented using a 1×6 one-hot encoding. The normalized simulation dataset is then divided into training and validation sets. The same normalization parameters are applied to unlabeled real test data to construct a real dataset.

[0025] Step S2, model pre-training: Use the training set to perform preliminary training on the neural network model, so that the neural network model's feature extractor has basic feature extraction capabilities. During the pre-training process, the CenterLoss method is used to further optimize the feature extractor, enabling it to better cluster similar samples and separate features of different categories.

[0026] definition is the feature vector extracted by the neural network model for the i-th data sample, is the center vector of category c, where d is the dimension of the feature vector. Assuming there are C categories, the goal of Centerloss is to minimize the Euclidean distance between the data sample features and their category centers:

[0027] Where N is the number of data samples in the training set, is a data sample Tags, Represents the Euclidean distance. CenterLoss and cross entropy loss are used together in training. The expression of cross entropy is:

[0028] in Is the neural network model for the data sample Belong to category The predicted probability of , the total loss function is: .

[0029] Step S3, transfer learning and diagnosis: take the labeled simulation dataset as the source domain and the unlabeled real dataset as the target domain, and adjust the feature distribution of the source and target domains by introducing adversarial loss and local domain alignment strategy. First, for the source domain (Simulation data) Target domain (actual unlabeled data) through feature extractor Extracting high-level features and , then, the features are input to the classifier and domain discriminator In the equation, we get the weight vector and classification loss and domain classification loss The objective function of partial transfer learning is

[0030] in, It is a hyperparameter used to weigh the loss of the source domain label classifier and the partial adversarial domain discriminator in the optimization problem. Finally, the optimal parameter needs to be found. , and :

[0031] Finally, the above optimal parameters are substituted into the neural network model, and the neural network model is used to accurately diagnose the bolt connection status of the real intermediate bearing dual-rotor system.

Claims

1. A data-model dual-driven diagnostic method for intermediate bearing dual-rotor bolt connections, characterized in that: The steps include: Step S0, generating simulation data using a physical model: establishing a finite element model of a dual-rotor with intermediate bearings with multiple possible bolt loosening states and normal states, generating labeled simulation data through simulation, and using a domain randomization method to randomize key parameters during the simulation process to simulate the dispersion and uncertainty of the actual dual-rotor with intermediate bearings; Step S1, data normalization and dataset construction: normalize the frequency components of the simulated data to obtain a simulated dataset, divide the simulated dataset into a training set and a validation set, and use the same normalization parameters for unlabeled real test data to construct a real dataset; Step S2, model pre-training: use the training set to perform preliminary training on the neural network model. During the preliminary training process, the CenterLoss method is used to further optimize the feature extractor, and the trained neural network model is verified using the validation set; Step S3, transfer learning and diagnosis: the labeled simulation dataset is used as the source domain, and the unlabeled real dataset is used as the target domain. By introducing adversarial loss and local domain alignment strategy, the feature distribution of the source domain and the target domain is adjusted so that the neural network model can adapt to the unlabeled real dataset of the target domain. Finally, the optimal parameters are obtained, and the neural network model after transfer learning is obtained. The neural network model after transfer learning is used to accurately diagnose the bolt connection status of the real intermediate bearing dual rotor.

2. The data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to claim 1 is characterized in that: In step S0, the dynamic equation corresponding to the intermediate bearing dual-rotor finite element model is: ； In the above formula, is the generalized displacement vector of the intermediate bearing dual-rotor system, 、 、 are the mass, damping and gyroscopic torque matrices, respectively, is the stiffness matrix without considering the asymmetry of the connection structure stiffness, is a time-varying function of stiffness, ε is an asymmetric coefficient indicating the degree of asymmetry of stiffness caused by bolt loosening, is the time-varying term caused by stiffness asymmetry, is the unbalanced excitation vector containing two speed excitation frequencies.

3. The data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to claim 1 is characterized in that: In step S0, the specific operation of the domain randomization method is: when each data sample is generated by simulating the intermediate bearing dual-rotor finite element model, key parameters are randomized within a certain range. The key parameters include support stiffness, damping coefficient, stiffness asymmetry coefficient, unbalanced excitation magnitude and phase, and the rotational speed of the two rotors.

4. The data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to claim 1, characterized in that: In step S1, the specific operation of normalizing the frequency components of the simulation data is as follows: first, the simulation data is subjected to a fast Fourier transform, and then the sum A of the amplitudes of the two rotor speed frequency components in the intermediate bearing dual-rotor system is calculated, and the simulation data is divided by A to make the sum of the rotor fundamental frequency amplitudes 1, thereby eliminating the response inconsistency caused by the excitation intensity.

5. The data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to claim 1, characterized in that: In step S2, the specific process of using CenterLoss to perform preliminary training on the neural network model is as follows: define is the feature vector extracted by the neural network model for the i-th data sample, is the center vector of category c, where d is the dimension of the feature vector. Assuming there are C categories, the goal of Centerloss is to minimize the Euclidean distance between the data sample features and their category centers: ； Where N is the number of data samples in the training set, is a data sample Tags, Represents the Euclidean distance. CenterLoss and cross entropy loss are used together in training. The expression of cross entropy is: ； in Is the neural network model for the data sample Belong to category The predicted probability of , the total loss function is: 。 6. The data-model dual-driven intermediate bearing dual-rotor bolt connection diagnosis method according to claim 1, characterized in that: In step S3, the specific operation of transfer learning is: first, for the source domain , target domain Through feature extractor Extracting high-level features and , then, the high-level features are input into the classifier and domain discriminator In the equation, we get the weight vector and classification loss and domain classification loss , the objective function of partial transfer learning is: ； in, It is a hyperparameter used to weigh the loss of the source domain label classifier and the partial adversarial domain discriminator in the optimization problem. Finally, the optimal parameter needs to be found. , and : 。

Images