Digital twinning-based aero-engine bearing fault diagnosis system and diagnosis method

Abstract

The invention discloses an aero-engine bearing fault diagnosis system and method based on digital twinning. The aero-engine bearing fault diagnosis system comprises a physical sensing layer, a data processing layer, an intelligent diagnosis layer and an application interaction layer. The physical sensing layer collects original vibration signals and working condition parameters of the bearing through a multi-class sensor network; the data processing layer adopts a variational mode decomposition algorithm to extract an intrinsic mode function component from the original vibration signal; the intelligent diagnosis layer extracts multi-scale features through a convolutional neural network layer by layer, and weights the features in combination with a double attention mechanism to focus key information. The diagnosis method outputs a result through signal acquisition, VMD processing, intelligent feature extraction and weighting and fault classification. According to the method, the advantages of signal decomposition of variational mode decomposition, deep feature learning of the convolutional neural network and attention focusing of a double attention mechanism are combined, and efficient and accurate diagnosis of the aero-engine bearing fault is realized.

Description

Technical Field

[0001] This invention pertains to the field of aero-engine condition monitoring and health management technology, specifically relating to an aero-engine bearing fault diagnosis system and method based on digital twins. Background Technology

[0002] As the "heart" of an aircraft, the safety and reliability of an aero-engine are crucial to flight safety. Bearings, as key rotating components of the engine, operate under harsh environments of high temperature, high speed, and heavy load, making them typical wear parts. Failures can range from affecting engine performance to potentially causing catastrophic accidents, resulting in huge economic losses and casualties. Traditional bearing maintenance strategies rely mainly on periodic inspections or reactive maintenance. Periodic inspections, based on fixed time cycles, may lead to unnecessary replacements of bearings in good condition, wasting resources. Reactive maintenance, performed after a failure, cannot avoid unplanned downtime and poses serious safety hazards. Therefore, developing condition-based maintenance (CBM) strategies to achieve real-time monitoring and accurate fault diagnosis of bearing operating conditions has become an urgent need in the field of aero-engine health management. In recent years, with the development of sensor technology and artificial intelligence, data-driven fault diagnosis methods have received widespread attention, especially those based on convolutional neural networks (CNNs). Deep learning technology, with its powerful feature self-learning capabilities, has achieved remarkable results in processing complex vibration signals. However, it still faces many challenges in the practical application of aero-engine bearing fault diagnosis: First, signal noise interference: early-stage bearing fault characteristic signals are usually very weak and easily submerged by strong background noise and vibration signals from other components during engine operation, making feature extraction difficult. Second, complex fault modes: bearings may exhibit multiple modes such as inner ring, outer ring, rolling element, or combined faults. Features of different fault modes may overlap in the time and frequency domains, making accurate identification difficult. Third, insufficient feature sensitivity: traditional deep learning diagnostic models may treat all features equally when extracting features, failing to focus on the most critical feature information for fault classification, affecting the accuracy and robustness of diagnosis. Therefore, there is an urgent need to develop intelligent diagnostic systems and methods that can effectively suppress noise and automatically extract highly discriminative fault features to improve the automation and intelligence level of aero-engine bearing fault diagnosis. Summary of the Invention

[0003] The purpose of this invention is to provide a fault diagnosis system and method for aero-engine bearings based on digital twins. This invention combines the signal decomposition of variational mode decomposition, deep feature learning of convolutional neural networks, and the attention-focusing advantages of a dual attention mechanism, achieving efficient and accurate diagnosis of aero-engine bearing faults.

[0004] The technical solution of this invention: a fault diagnosis system for aero-engine bearings based on digital twins, comprising a physical perception layer, a data processing layer, an intelligent diagnosis layer, and an application interaction layer; The physical sensing layer is equipped with a network of multiple sensors at the key bearing of the aero-engine to collect original vibration signals and operating parameters. The physical sensing layer also includes a three-dimensional virtual diagnostic model of the key bearing of the aero-engine, which is used to realize real-time mapping and synchronization of the spatial state of the physical bearing and the virtual model. The data collected by the multiple sensor networks is synchronously transmitted to the diagnostic model to support signal analysis in the twin scenario. The data processing layer includes a variational mode decomposition module, which uses a variational mode decomposition algorithm to adaptively reduce noise and extract features from the original vibration signal to obtain intrinsic mode function components. The intelligent diagnostic layer includes a convolutional neural network module and a dual attention mechanism module. The convolutional neural network module is used to extract multi-scale features hierarchically from intrinsic mode function components. The dual attention mechanism module uses channel attention and spatial attention to weight the multi-scale features to achieve key feature focusing. The application interaction layer includes a client-server architecture and a visualization module. The visualization module is used for visualizing fault results, providing health warnings, and supporting maintenance decisions.

[0005] In the aforementioned digital twin-based aero-engine bearing fault diagnosis system, the multi-sensor network includes vibration acceleration sensors and speed sensors.

[0006] In the aforementioned aero-engine bearing fault diagnosis system based on digital twins, the convolutional neural network consists of at least six one-dimensional convolutional layers, a linear rectifier unit activation function, a max pooling layer, and a random deactivation layer; the dual attention mechanism is deployed after the convolutional neural network, and the dual attention mechanism includes a channel attention mechanism and a spatial attention mechanism set in parallel.

[0007] In the aforementioned digital twin-based aero-engine bearing fault diagnosis system, the client-server architecture includes a client interface and a server backend; the client interface is equipped with a real-time vibration waveform display window, a diagnostic result output window, a fault analysis report panel, and a historical data backtracking query function.

[0008] The aforementioned diagnostic method for the aero-engine bearing fault diagnosis system based on digital twins specifically includes the following steps: Step S1: Collect the original vibration signals and operating parameters of key bearing parts of the aero-engine through a multi-type sensor network; Step S2: Adaptively decompose and denoise the acquired original vibration signal using the variational mode decomposition module to obtain the intrinsic mode function components; Step S3: Construct and train a diagnostic model that integrates convolutional neural networks and dual attention mechanisms. Input the intrinsic mode function components into the diagnostic model, extract multi-scale features through convolutional neural networks, and perform attention weighting on key features in the extracted multi-scale features through dual attention mechanisms to obtain weighted features. Step S4: Classify faults based on weighted features, and finally output the fault diagnosis results through the application interaction layer.

[0009] In the aforementioned diagnostic method, in step S2, the variational mode decomposition algorithm decomposes the original vibration signal into... The constrained variational problem is expressed as: quasi-orthogonal eigenmode function components. ; ; in, Represents the set of optimization variables. Indicates the first One intrinsic mode function; For the first The center frequency of each mode This represents the time partial derivative operator. Represents the Dirac function, The imaginary unit, This represents the convolution operation. Represents a time variable. This represents the original vibration signal.

[0010] In the aforementioned diagnostic method, step S3, the process of extracting multi-scale features through a convolutional neural network, involves inputting intrinsic mode function components as two-dimensional features into the convolutional neural network, extracting local abstract features of the original vibration signal through multi-layer convolution and pooling operations, and converting the extracted local abstract features into feature maps. The specific operation is as follows: ; in, This represents the convolution operation. For convolution kernel weights, This is the bias of the convolution kernel. This is the ReLU activation function.

[0011] In the aforementioned diagnostic method, the attention weighting includes channel attention weighting and spatial attention weighting; the channel attention weighting learns the weights of each feature channel in the feature map through a compressed activation network. This enhances the response to fault-sensitive channels; the spatial attention weighting generates an attention weight map focused on the time-domain interval corresponding to the fault pulse by calculating the correlation between spatial locations of feature maps. .

[0012] In the aforementioned diagnostic method, the channel attention weighting process first involves performing global average pooling on the feature map to obtain channel statistics. Then, weights are generated through two fully connected layers and a sigmoid function, with the specific formula as follows: ; in, It is the ReLU activation function. and For learning parameters, For channel statistics, This is the scaling ratio.

[0013] In the aforementioned diagnostic method, the specific formula for spatial attention weighting is as follows: ; in, and It is a linear transformation characteristic.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention enhances the accuracy and anti-interference capability of aero-engine bearing fault diagnosis through the effective preprocessing of a variational mode decomposition module, the deep feature extraction of a convolutional neural network module, and the precise focusing of a dual attention mechanism module. These three elements work synergistically to effectively capture early, subtle fault characteristics, achieving efficient and accurate fault identification. By introducing a variational mode decomposition algorithm, the original vibration signal is decomposed into more stable sub-signals, effectively overcoming the limitations of traditional fault diagnosis methods in high-noise and variable operating conditions, and significantly improving the signal-to-noise ratio of feature extraction. This invention utilizes a multi-layer convolutional neural network module to automatically learn fault features, avoiding the reliance on expert knowledge and manual feature extraction in traditional methods, and enabling the capture of deeper and more abstract fault patterns. The dual attention mechanism module allows the diagnostic model to adaptively focus on the most discriminative feature channels and signal time points, effectively enhancing the sensitivity to early, subtle faults and improving diagnostic accuracy. Attached Figure Description

[0015] Figure 1 This is a system diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a flowchart of the variational mode decomposition process of the present invention; Figure 4This is an example of the original signal diagram of the bearing outer ring under different fault conditions collected by NI-1 in embodiment of the present invention; Figure 5 This is an example of the original signal diagram of the bearing inner ring under different fault conditions collected by NI-1 in embodiment of the present invention; Figure 6 This is an example of the original signal diagram of the bearing rolling elements under different fault conditions collected by NI-1 in this invention. Figure 7 This is a combined raw signal diagram of different faults under different operating conditions collected by NI-1 in embodiment of the present invention; Figure 8 The variational mode decomposition diagram and corresponding spectrum diagram of the original diagnostic signal diagram are shown in the example. Figure 9 This is a schematic diagram of the overall structure of the dual attention mechanism of the present invention; Figure 10 Visual confusion matrix diagram for VMD-CNN-DANE fault classification in the example; Figure 11 This is a schematic diagram of the visual interface; Figure 12 The loss function and accuracy changes during the training process of VMD-CNN-DANET are illustrated in the example. Figure 13 The loss function and accuracy changes during the training process of VMD-CNN-DANE are illustrated in the example. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0017] Example: A digital twin-based fault diagnosis system for aero-engine bearings, such as... Figure 1 As shown, the core lies in building a closed-loop technology system of "physical sensing, signal processing, intelligent diagnosis and interactive feedback", which includes a physical sensing layer, a data processing layer, an intelligent diagnosis layer and an application interaction layer.

[0018] The physical sensing layer deploys a network of multiple sensors at key bearing locations of the aero-engine (including rotating components such as bearings and accessory gearboxes) to collect raw vibration signals and operating parameters. This network includes vibration acceleration sensors and rotational speed sensors, focusing on collecting high-frequency vibration acceleration signals, supplemented by operating parameters such as rotational speed and temperature, to comprehensively and in real-time capture the vibration state of the bearing during operation. The physical sensing layer also includes a three-dimensional virtual diagnostic model of the key aero-engine bearing, used to achieve real-time mapping and synchronization of the spatial state between the physical bearing and the virtual model. Data collected by the multiple sensor networks is synchronously transmitted to the diagnostic model, supporting signal analysis in a twin scenario.

[0019] The data processing layer includes a variational mode decomposition module, which uses a variational mode decomposition algorithm to adaptively reduce noise and extract features from the original vibration signal to obtain intrinsic mode function components.

[0020] The intelligent diagnostic layer comprises, in sequence, an input layer, a convolutional neural network (CNN) module, a dual attention mechanism (DANet) module, a flattening layer, and an output layer, as follows: The input layer receives the multi-channel intrinsic mode function (IMF) components after variational mode decomposition (VMD) and normalization processing in step S2; The Convolutional Neural Network (CNN) module, as the core feature extraction unit, is used to learn abstract fault features from the input signal hierarchically and automatically, from shallow to deep layers, and transform them into feature maps. In this embodiment, the module is a deep architecture composed of multiple one-dimensional convolutional layers (Conv1D), nonlinear activation layers (ReLU), and pooling layers stacked together. Through multi-layer convolutional operations, the network can automatically and hierarchically learn the feature representation of fault signals, from shallow local impulses and transient impacts to deep complex patterns such as periodic envelopes and resonant frequency bands. The CNN module is used to extract multi-scale features hierarchically from the intrinsic mode function components. The CNN consists of at least six one-dimensional convolutional layers, linear rectified unit activation functions, max pooling layers, and random deactivation layers. Specifically, the convolutional neural network (CNN) comprises six one-dimensional convolutional (Conv1D) layers, with a preferred kernel size of 3. The first five convolutional layers use 64 filters, increasing to 128 filters in the sixth layer. Max-pooling layers and dropout layers are added at appropriate locations within the network to enhance the generalization ability of the diagnostic model. All convolutional layers employ one-dimensional convolutional kernels of size 3 to capture local short-term correlation features in the signal. Rectified linear units (ReLU) are used as activation functions after all convolutional layers to introduce non-linear mapping capabilities into the diagnostic model. To prevent overfitting and improve generalization, the following regularization techniques are integrated into the CNN: firstly, one-dimensional max-pooling (MaxPooling1D) layers are inserted after some convolutional layers to reduce the dimensionality of the feature maps and enhance the invariance of features to small temporal drifts. Then, L2 weight regularization is introduced in the first convolutional layer to suppress the complexity of the diagnostic model by penalizing excessively large weight values. Finally, a random deactivation layer (Dropout) is introduced in the latter half of the convolutional neural network with a dropout rate of 0.2. During training, some neurons are randomly deactivated, forcing the network to learn more robust feature representations.

[0021] The feature maps extracted by the CNN module are input into the Dual Attention Mechanism (DANet) module to achieve dynamic weighting of key fault information. Based on the features extracted by the CNN module, the Dual Attention Mechanism (DANet) module is introduced to focus on and enhance key fault information. This module consists of parallel channel attention and spatial attention mechanisms, which weight multi-scale features to achieve key feature focusing. The dual attention mechanism is deployed after the convolutional neural network and includes parallel channel attention and spatial attention mechanisms. The DANet module reshapes the feature maps and performs matrix multiplication to calculate the correlation matrices between channels and spatial locations. After Softmax normalization, the attention weights are obtained and multiplied with the original feature maps to achieve adaptive feature weighting. Finally, the fusion layer performs a residual connection between the outputs of the two attention mechanisms and the feature maps to preserve original information and accelerate the convergence of the diagnostic model.

[0022] The flatten layer flattens the multidimensional feature map after it has been weighted by the attention module, and converts it into a one-dimensional feature vector.

[0023] The output layer is a fully connected layer that receives the one-dimensional feature vector and uses a Softmax classifier to calculate the probability that the input sample belongs to each preset fault category. Finally, the category with the highest probability value is used as the diagnostic output of the diagnostic model.

[0024] The application interaction layer includes a client-server architecture and a visualization module. The visualization module is used for visualizing fault results, providing health warnings, and supporting maintenance decisions. The client-server architecture includes a client interface and a server backend. The client interface features a real-time vibration waveform display window, a diagnostic result output window, a fault analysis report panel, and a historical data backtracking query function.

[0025] Diagnostic methods for aero-engine bearing fault diagnosis systems based on digital twins, such as Figure 2 As shown, it includes the following steps: Step S1: Collect raw vibration signals and operating parameters of key bearing components of the aero-engine through a multi-sensor network; Specifically, in this embodiment, the effective sample data utilized by the method of the present invention comes from an aero-engine bearing failure simulation test bench. This sample data aims to simulate the complex operating conditions of an aero-engine in actual operation. The sensor deployment and corresponding data recording method are as follows: Six vibration acceleration sensors are arranged at key measuring points on the test bench, collecting raw vibration signals from multiple dimensions such as axial, horizontal radial, and vertical radial directions to ensure complete capture of the dynamic characteristics of the bearing. The acquisition system records the signals from each sensor as multi-channel data, labeled A1 to A6 in the database. Each set of data [A1, A2, A3, A4, A5, A6] has a precise timestamp, fully characterizing the comprehensive vibration state of the bearing at a specific moment, forming a multi-dimensional time series sample for subsequent analysis.

[0026] To comprehensively simulate the operating conditions of an aircraft engine under different flight phases and variations in speed and load, this embodiment sets up four typical operating conditions, with specific parameter configurations shown in Table 1: Table 1 Operating Condition Configuration

[0027] To construct a fault sample set for training and validating the diagnostic model, various typical health conditions were pre-fabricated on the test bearings of the experimental rig using electrical discharge machining (EDM). This embodiment includes five bearing health conditions, as shown in Table 2: Table 2 Bearing Health Status Configuration

[0028] Data was collected under the above operating conditions and health statuses, with each sample containing 1024 data points, forming the basic dataset required for subsequent fault diagnosis.

[0029] Step S2: Adaptively decompose and denoise the acquired original vibration signal using the variational mode decomposition module to obtain the intrinsic mode function components; Variational mode decomposition (VMD), as an advanced adaptive signal processing technique, can decompose complex original vibration signals into a series of intrinsic mode function components with practical physical meaning. It can effectively suppress interference caused by strong background noise and operating condition fluctuations, and can accurately separate signal components related to early weak faults and those with specific center frequencies, providing high-quality input for subsequent feature extraction.

[0030] In this step, the acquired raw vibration signal is standardized and filtered. Then, considering the characteristics of the raw vibration signal being non-stationary, nonlinear, and having a strong noise background, Variational Mode Decomposition (VMD) is introduced for adaptive signal denoising and feature extraction. The raw vibration signal is then decomposed into quasi-orthogonal intrinsic mode functions (IMFs) by solving a constrained variational problem. The constrained variational problem is expressed as: ; ; in, Represents the set of optimization variables. Indicates the first One intrinsic mode function; For the first The center frequency of each mode This represents the time partial derivative operator. Represents the Dirac function, The imaginary unit, This represents the convolution operation. Represents a time variable. This represents the original vibration signal. VMD decomposes the original signal into a series of IMFs with specific sparse properties by iteratively searching for the optimal solution. The obtained IMFs effectively separate the background noise and contain the fault impact component, and effectively suppress the mode mixing phenomenon.

[0031] Solving the constrained variational problem of VMD using the above iterative algorithm, the final decomposed modes can be compactly represented in matrix form as follows: ; In this matrix, each column represents an IMF component.

[0032] Subsequently, all IMF components obtained from the decomposition are subjected to Min-max normalization, the expression of which is: ; In the formula, Represents the normalized i-th IMF data, and These represent the minimum and maximum values ​​of the IMF component, respectively.

[0033] In this embodiment, the penalty parameters for VMD are set. =2000, number of decomposed modes =5, convergence tolerance, decomposition process as follows: Figure 3 As shown. (Attached) Figure 4-7This image shows the raw signal from 1024 sample points acquired by the gear sensor NI-1 under normal conditions. VMD decomposition was performed on this signal, with a mode number of 5. The resulting IMF signal plots are shown in the attached image. Figure 8 As shown.

[0034] The sample data of the original vibration signals collected by each accelerometer under different operating conditions were segmented into non-overlapping segments according to each sample data point. Then, the segmented sample data were grouped according to the health status of the gear. Each accelerometer, as a different dimension, was used to stitch together the sample data of the same health status under different operating conditions to obtain multi-dimensional fused data, which was then divided into training set, validation set and test set in a 4:1:1 ratio.

[0035] Step S3: Construct and train a diagnostic model that integrates convolutional neural networks and dual attention mechanisms. Input the intrinsic mode function components into the diagnostic model, extract multi-scale features through convolutional neural networks, and perform attention weighting on key features in the extracted multi-scale features through dual attention mechanisms to obtain weighted features. In this step, the diagnostic model is a deep neural network. The process of extracting multi-scale features by the convolutional neural network involves inputting the intrinsic mode function components as two-dimensional features into the convolutional neural network, extracting local abstract features of the original vibration signal through multi-layer convolution and pooling operations, and transforming the extracted local abstract features into feature maps. The specific operation is as follows: ; in, This represents the convolution operation. For convolution kernel weights, This is the bias of the convolution kernel. This is the ReLU activation function.

[0036] like Figure 9 As shown, the attention weighting includes channel attention weighting and spatial attention weighting; the channel attention weighting learns the weights of each feature channel in the feature map through a compressed activation network. This enhances the response to fault-sensitive channels; the spatial attention weighting generates an attention weight map focused on the time-domain interval corresponding to the fault pulse by calculating the correlation between spatial locations of feature maps. .

[0037] Specifically, the channel attention mechanism learns the interdependencies between different feature channels and adaptively assigns higher weights to feature channels containing more fault information. The channel attention mechanism first performs global average pooling on the feature maps to obtain channel statistics. Then, weights are generated through two fully connected layers and a sigmoid function, specifically: ; in, It is the ReLU activation function. and For learning parameters, For channel statistics, This is the scaling ratio.

[0038] Spatial attention mechanism establishes long-range dependencies between different spatial (temporal) locations in the feature map, enabling the diagnostic model to focus on the most identifiable parts of the signal, such as periodic impulse components. The specific formula for spatial attention weighting is as follows: ; in, and It is a linear transformation characteristic.

[0039] The training process of the diagnostic model involves inputting the decomposed and feature-fused intrinsic mode function components into the model for forward propagation, calculating the error between the predicted output and the true label, and then using a backpropagation algorithm to calculate the gradient and update the diagnostic model parameters. During the training phase, the diagnostic model uses mini-batch stochastic gradient descent (such as the Adam optimizer) to iteratively optimize the model parameters, with a learning rate set to 0.0001. The diagnostic model uses classification cross-entropy as the loss function to measure the difference between the predicted and true distributions. To balance the convergence efficiency and computational resource consumption of the diagnostic model, the batch size is set to 16, and the number of training epochs is set to 15.

[0040] Two types of callback mechanisms are introduced during training: the first callback function is used to save the weight parameters of the best diagnostic model in real time when the accuracy on the validation set improves; the second callback function is used when the loss function decreases at a rate less than 1×10 for three consecutive training epochs. -3 Training is automatically terminated when necessary to prevent the diagnostic model from overfitting.

[0041] The training set was used for diagnostic model learning, the validation set was used for hyperparameter tuning and early stopping strategies to prevent overfitting, and the test set was used to evaluate the final performance of the diagnostic model, saving the diagnostic model that performed best on the validation set. During the training process, the diagnostic model showed a rapid overall decline in performance. Results show that the model reached stable convergence after approximately 8 to 10 iterations, the loss function stabilized, and the validation set accuracy remained at a high level, indicating that the diagnostic model has good stability and generalization ability.

[0042] The deep convolutional neural network (CNN) module in this invention leverages its powerful hierarchical learning capabilities to automatically extract abstract fault features ranging from low-level texture to high-level semantics from the decomposed signal. Through multi-layer convolution and pooling operations, the diagnostic model can progressively capture the deep and non-linear features of the fault signal, avoiding the heavy reliance on manual feature extraction and expert knowledge in traditional methods. This invention introduces a dual attention (DANet) mechanism to intelligently reshape the channel and spatial (temporal) dimensions of the features extracted by the CNN, enabling the diagnostic model to focus on the most identifiable fault information while suppressing interference from irrelevant features. This significantly improves the diagnostic model's recognition accuracy and generalization ability under complex signals.

[0043] Step S4: Classify faults based on weighted features, and finally output the fault diagnosis results through the application interaction layer.

[0044] In this embodiment, the feature map of the weighted features is flattened into a one-dimensional vector, which is then input to the output layer and the Softmax classifier. The output is the probability distribution of fault categories (e.g., inner race fault, outer race fault, rolling element fault, compound fault, etc.). The actual fault labels are used as supervision, and the cross-entropy loss function is used as the optimization objective. The confusion matrix is ​​as follows: Figure 10 As shown.

[0045] The trained and validated diagnostic model from step S4 is deployed in the aero-engine condition monitoring and fault diagnosis system. Test set samples are input into the trained diagnostic model for inference calculations. During the inference phase, the real-time acquired and preprocessed intrinsic mode function components are input into the optimal diagnostic components. Multi-scale feature extraction is achieved through a CNN module, and feature weighting and information enhancement are achieved using the channel attention and spatial attention mechanisms in the DANet structure. Real-time bearing fault diagnosis results are then output, and the visualization interface is shown below. Figure 11 As shown, this enables accurate classification of inner ring faults, outer ring faults, rolling element faults, and combined faults. For example... Figure 12 and Figure 13 As shown, the diagnostic model's final accuracy rate stabilized at over 98.5%, verifying the feasibility and engineering applicability of this method in complex environments with multiple operating conditions.

[0046] During actual engine operation, raw vibration signals from critical bearings are acquired in real time and fed into a deployed diagnostic model. The diagnostic model automatically and continuously executes the entire process of signal decomposition (step S2), feature extraction and attention weighting (step S3), and fault classification (step S4), ultimately outputting an accurate diagnostic result of the bearing's current health status (e.g., "normal," "inner race fault," and "rolling element fault"). This diagnostic result can be used to trigger alarms and generate equipment health reports, providing crucial decision-making support for implementing Condition-Based Maintenance (CBM) strategies, thereby ensuring the safe and economical operation of the aero-engine.

[0047] This invention enhances the accuracy and anti-interference capability of aero-engine bearing fault diagnosis through the effective preprocessing of a variational mode decomposition module, the deep feature extraction of a convolutional neural network module, and the precise focusing of a dual attention mechanism module. These three elements work synergistically to effectively capture early, subtle fault characteristics, achieving efficient and accurate fault identification. By introducing a variational mode decomposition algorithm, the original vibration signal is decomposed into more stable sub-signals, effectively overcoming the limitations of traditional fault diagnosis methods in high-noise and variable operating conditions, and significantly improving the signal-to-noise ratio of feature extraction. This invention utilizes a multi-layer convolutional neural network module to automatically learn fault features, avoiding the reliance on expert knowledge and manual feature extraction in traditional methods, and enabling the capture of deeper and more abstract fault patterns. The dual attention mechanism module allows the diagnostic model to adaptively focus on the most discriminative feature channels and signal time points, effectively enhancing the sensitivity to early, subtle faults and improving diagnostic accuracy.

[0048] Comparative Analysis: The diagnostic model was compared with models of various structures, including CNN, CNN-ECA, VMD-CNN, and VMD-CNN-ECA. The comparative experiments were conducted under four different operating conditions, and the results are shown in Table 3 below. Table 3 Comparison of different structural models under four different working conditions

[0049] As shown in Table 3, the diagnostic model exhibits the highest classification accuracy under all operating conditions. Specifically, under complex operating condition C, the validation set accuracy reaches 98.9%, an improvement of approximately 6.2% compared to the traditional CNN model (92.7%). This indicates that the combined application of VMD decomposition and dual attention mechanism employed in this invention significantly improves fault identification capability and model robustness.

[0050] In summary, compared to traditional shallow learning methods based on time-domain or frequency-domain features, the diagnostic model of this invention exhibits stronger anti-interference capabilities in noisy signal environments. Through adaptive signal decomposition using VMD, multi-scale convolutional feature extraction using CNN, and attention weighting mechanism using DANet, the model achieves high recognition rates even with small sample sizes, significantly improving the accuracy of condition monitoring and fault early warning capabilities for aero-engine bearings.

Claims

1. An aero-engine bearing fault diagnosis system based on digital twinning, characterized in that: The physical perception layer, the data processing layer, the intelligent diagnosis layer and the application interaction layer are included. The physical perception layer is provided with a multi-type sensor network for collecting original vibration signals and working condition parameters at a key bearing part of an aero-engine, and further includes a three-dimensional virtual diagnosis model of the key bearing of the aero-engine, for realizing real-time mapping and synchronization of spatial states of the physical bearing and the virtual model, and the data collected by the multi-type sensor network is synchronously transmitted to the diagnosis model to support signal analysis in a twin scene. The data processing layer is provided with a variational modal decomposition module, which adopts a variational modal decomposition algorithm to adaptively denoise and extract features from the original vibration signals to obtain an intrinsic modal function component. The intelligent diagnosis layer includes a convolutional neural network module and a double attention mechanism module, the convolutional neural network module is used to extract multi-scale features from the intrinsic modal function component, and the double attention mechanism module is used to focus on key features by weighting the multi-scale features through channel attention and spatial attention. The application interaction layer includes a client-server architecture and a visualization module, and the visualization module is used for visual display of fault results, health warning and maintenance decision support.

2. The digital-twin-based aero-engine bearing fault diagnosis system according to claim 1, characterized in that: The multi-type sensor network includes a vibration acceleration sensor and a rotational speed sensor.

3. The digital-twin-based aero-engine bearing fault diagnosis system according to claim 1, characterized in that: The convolutional neural network is composed of at least six one-dimensional convolutional layers, linear rectifier unit activation functions, maximum pooling layers and random inactivation layers; the double attention mechanism is arranged after the convolutional neural network, and the double attention mechanism includes a channel attention mechanism and a spatial attention mechanism arranged in parallel.

4. The digital-twin-based aero-engine bearing fault diagnosis system according to claim 1, characterized in that: The client-server architecture includes a client interface and a server background; the client interface is provided with a real-time vibration waveform display window, a diagnosis result output window, a fault analysis report panel and a historical data backtracking query function.

5. The diagnosis method of the digital-twin-based aeroengine bearing fault diagnosis system according to any one of claims 1-4, characterized in that, The method specifically includes the following steps: Step S1: collecting original vibration signals and working condition parameters at a key bearing part of an aero-engine through a multi-type sensor network; Step S2: adaptively decomposing and denoising the collected original vibration signals through a variational modal decomposition module to obtain an intrinsic modal function component; Step S3: constructing and training a diagnosis model integrating a convolutional neural network and a double attention mechanism, inputting the intrinsic modal function component into the diagnosis model, extracting multi-scale features through the convolutional neural network, and weighting key features in the extracted multi-scale features through the double attention mechanism to obtain weighted features; Step S4: classifying faults based on the weighted features, and finally outputting fault diagnosis results through an application interaction layer.

6. The diagnostic method according to claim 5, characterized in that: In step S2, the variational modal decomposition algorithm decomposes the original vibration signal into K quasi-orthogonal proper orthogonal decomposition (POD) components by solving a constrained variational problem expressed as: min f (x) + R(x) subject to g(x) = 0 where f(x) is a cost function, R(x) is a regularization term, g(x) is a constraint function, and x is the unknown variable. In step S2, the variational modal decomposition algorithm decomposes the original vibration signal into K quasi-orthogonal proper orthogonal decomposition (POD) components by solving ； ； wherein, represents a set of optimization variables, represents the i-th eigenmode function; is the center frequency of the i-th mode, represents the time derivative operator, represents the Dirac function, is the imaginary unit, represents the convolution operation, represents the time variable, represents the original vibration signal.​​ 7. The diagnostic method according to claim 5, characterized in that: In step S3, the process of extracting multi-scale features by the convolutional neural network is that the intrinsic mode function components are input as two-dimensional features into the convolutional neural network, local abstract features of the original vibration signal are extracted through multi-layer convolution and pooling operation, and the extracted local abstract features are converted into a feature map The specific operation is as follows: ； wherein, denotes a convolution operation, is a convolution kernel weight, is a bias of the convolution kernel, is a ReLU activation function.

8. The diagnostic method according to claim 7, characterized in that: The attention weighting includes channel attention weighting and spatial attention weighting; the channel attention weighting learns the weight of each feature channel in the feature map through a compression excitation network, and enhances the response to a fault-sensitive channel ; and the spatial attention weighting generates an attention weight map focusing on a time domain interval corresponding to a fault pulse by calculating the correlation between spatial positions of the feature map .

9. The diagnostic method according to claim 8, characterized in that: The process of the channel attention weighting is to firstly perform global average pooling on the feature map to obtain channel statistics Then, weights are generated by two fully connected layers and a Sigmoid function, and the specific formula is: ； wherein, is a ReLU activation function, and is a learning parameter, is a channel statistic, is a scaling ratio.

10. The diagnostic method according to claim 8, characterized in that: The specific formula of the spatial attention weighting is as follows: ； wherein and are linear transformation characteristics.

Images