Turbine rotor-blade multi-modal vibration analysis and fault identification method and system

Abstract

This invention discloses a method and system for multimodal vibration analysis and fault identification of a steam turbine rotor-blade, belonging to the field of steam turbine monitoring technology. It solves the problem that existing methods are unable to clearly locate and separate fault features through simple spectrum analysis. The method includes noise suppression preprocessing of multimodal vibration signals, modal feature extraction of multimodal preprocessed information based on a convolutional neural network model, and identification and analysis of high-dimensional feature state vectors based on a fault identification model. This invention can provide accurate physical guidance for multimodal monitoring and feature extraction of vibration signals through a coupled three-dimensional rotor-blade model. By using frequency band and adaptive threshold noise suppression preprocessing in the multimodal vibration signals, and by using a convolutional neural network model guided by the coupled three-dimensional model to extract modal features from the multimodal preprocessed information, early weak fault features can be clearly located and separated.

Description

Technical Field

[0001] This invention belongs to the field of steam turbine monitoring technology, specifically relating to a method and system for multimodal vibration analysis and fault identification of steam turbine rotor-blade. Background Technology

[0002] Steam turbine generator sets are core power equipment in process industries such as power, petrochemical, and aviation. The operating status of their rotor-blade system directly determines the economy, safety, and availability of the entire unit. The steam turbine rotor-blade system is a key component for energy conversion. During operation, it is subjected to extreme conditions of high temperature, high pressure, and high speed for a long time. It is highly susceptible to various vibration faults induced by factors such as design defects, manufacturing errors, material fatigue, and operational instability. If such faults are not identified and diagnosed in a timely and accurate manner, they may lead to a decrease in unit efficiency and unplanned shutdowns, or even catastrophic accidents such as blade breakage, rotor imbalance, or even complete unit damage, resulting in huge economic losses and safety risks.

[0003] Currently, the industry commonly uses accelerometers or velocity sensors installed at critical locations such as bearing housings to collect vibration signals from turbine units. Existing methods mainly rely on the observation and interpretation of signals in the time or frequency domains (such as spectrum analysis). However, the vibration problem of the turbine rotor-blade system is highly complex. The impact or modulation signals generated by early weak faults are easily drowned out by background noise and vibrations from other components. In addition, a fault in one component often triggers abnormal responses in multiple modes, causing the fault characteristics to appear in the spectrum as complex sidebands or cross-modulation of frequency components within a wide frequency band. Existing methods cannot clearly locate and separate fault characteristics through simple spectrum analysis. To address these issues, we propose a method and system for multimodal vibration analysis and fault identification of turbine rotor-blade systems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and system for multimodal vibration analysis and fault identification of turbine rotor-blade. This solves the problem that the impact or modulation signals generated by early weak faults are easily submerged by background noise and vibrations of other components, causing the fault characteristics to appear in the spectrum as complex sidebands or cross-modulation of frequency components within a wide frequency band. Existing methods cannot clearly locate and separate fault characteristics through simple spectrum analysis.

[0005] This invention is implemented as follows: a method for multimodal vibration analysis and fault identification of a steam turbine rotor-blade, the method comprising: Based on the basic information of the turbine rotor-blade, a three-dimensional model of rotor-blade coupling considering dynamic coupling effects is constructed. The multi-modal monitoring points in the turbine rotor-blade architecture are determined through the three-dimensional model of rotor-blade coupling, and the multi-modal monitoring terminal is set at the multi-modal monitoring point. Based on the real-time acquisition of multimodal vibration signals of the turbine rotor-blade system by the multimodal monitoring terminal, noise suppression preprocessing of the multimodal vibration signals is performed to obtain noise-suppressed multimodal preprocessing information; Load multimodal preprocessed information, extract modal features from the multimodal preprocessed information based on a convolutional neural network model guided by a coupled 3D model, and output a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system; High-dimensional feature state vectors are obtained, and the high-dimensional feature state vectors are identified and analyzed based on a pre-built fault identification model to obtain the fault identification results of the turbine rotor-blade system, and the fault identification results are presented in a visual manner.

[0006] Preferably, the method for constructing a rotor-blade coupled three-dimensional model considering dynamic coupling effects based on the basic information of the turbine rotor-blade includes: Collect basic information on the turbine rotor and blades, and construct a three-dimensional finite element model of the turbine rotor and blade system based on the basic information on the turbine rotor and blades. The fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of the turbine rotor-blade system to the main interface modes and constraint modes, resulting in a reduced-order three-dimensional finite element model. The constraint modes include rotor flexibility mode, disk flexibility mode, blade flapping mode, and oscillation mode. The reduced-order three-dimensional finite element model is loaded, and a nonlinear behavior connector is introduced into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect. The model simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector. Fault samples are generated under typical fault conditions using a generative adversarial network. The rotor-blade coupled 3D model is iteratively trained using the fault samples to obtain a rotor-blade coupled 3D model that can cover typical faults.

[0007] Preferably, the preprocessing method for noise suppression of multimodal vibration signals includes: Multimodal vibration signals are acquired, and outlier and missing values ​​are processed to obtain cleaned multimodal vibration signals. After loading the cleaned multimodal vibration signal, perform spectral analysis on the multimodal vibration signal, output the signal spectral analysis results, identify the noise segment and the main frequency segment in the multimodal vibration signal, and divide the multimodal vibration signal into long and short windows based on the signal spectral analysis results to obtain at least one set of time sequence signal windows; The multimodal vibration signal within the time-series signal window is acquired, and the average density of the noise power spectrum within the time-series window is determined based on the Periodogram spectrum estimation method. It is then determined whether the average density of the noise power spectrum within the time-series window exceeds the preset power spectrum density threshold. If the average density of the noise power spectrum within the time window does not exceed the preset power spectrum density threshold, a second-order Butterworth filter is used to filter the multimodal vibration signal within the time window. During filtering, the multimodal preprocessing information output after filtering is represented as follows:

[0008]

[0009] in, This represents the multimodal preprocessing information output after filtering. This represents the input of multimodal vibration signals within a time-series signal window. This indicates the filtering process performed by the Butterworth filter. These are the acquisition frequency and cutoff frequency of the multimodal vibration signal, respectively. Let the filter order be . This is the transfer function of the Butterworth filter; If the average density of the noise power spectrum within the time window exceeds the preset power spectrum density threshold, a sixth-order Butterworth filter is used to filter the multimodal vibration signal within the time window. The system acquires the filtered multimodal preprocessing information, normalizes the multimodal preprocessing information, and outputs the normalized multimodal preprocessing information.

[0010] Preferably, the method for modal feature extraction of multimodal preprocessed information includes: A pre-built convolutional neural network model guided by a coupled 3D model is constructed. Fault samples are generated through a generative adversarial network. The convolutional neural network model guided by the coupled 3D model is trained using the fault samples, and a converged convolutional neural network model is output. Loading multimodal preprocessing information, the Generative Adversarial Network (GAN) is driven by the multimodal preprocessing information to generate adversarial derivative physical information. The rotor-blade coupled three-dimensional model obtains the adversarial derivative physical information, and the simulation output based on the adversarial derivative physical information includes stress cloud diagrams and modal vibration mode frames. Simulated physical information and multimodal preprocessing information are acquired. Based on the first and second feature extraction branches in the convolutional neural network model, features are extracted from the simulated physical information and multimodal preprocessing information to obtain simulated physical features and multimodal vibration features. Simulated physical features and multimodal vibration features are acquired, and the simulated physical features and multimodal vibration features are interactively fused based on the cross-attention mechanism to obtain a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system.

[0011] Preferably, when the pre-constructed convolutional neural network model guided by the coupled 3D model is built, the convolutional neural network architecture is used as the basic architecture, and a cross-attention mechanism is introduced into the convolutional neural network architecture. The convolutional neural network architecture uses the cross-attention mechanism to interactively fuse simulated physical features and multimodal vibration features. The convolutional neural network architecture includes three convolutional layers, a max pooling layer, and an average pooling layer. The convolutional neural network model also includes a parallel branch block, which includes a first feature extraction branch and a second feature extraction branch. The first feature extraction branch includes a ResNet50 layer, a batch normalization layer, and a max pooling layer. The second feature extraction branch includes a ResNet layer, and the second feature extraction branch also includes two convolutional layers, a max pooling layer, and a ReLU activation function. The loss function of the convolutional neural network model is expressed as:

[0012]

[0013] in, These are the second feature extraction branch, the first feature extraction branch, and the loss function of the convolutional neural network architecture, respectively. These are the second feature extraction branch, the first feature extraction branch, and the weight coefficients of the convolutional neural network architecture, respectively.

[0014] Preferably, the fault identification model includes an input layer, a support vector machine (SVM) model, and an output layer. An autoencoder and a weight calibration layer are positioned between the input layer and the SVM model. The autoencoder is used to acquire high-dimensional feature state vectors and performs autoencoding processing on these vectors to obtain at least one set of multi-dimensional autoencoded features. The autoencoder is connected to the weight calibration layer, which is also connected to the SVM model. The weight calibration layer includes a graph neural network, a global average pooling layer, and two fully connected layers. The weight calibration layer learns the nonlinear cross-relationships of the autoencoded features in the graph neural network through the fully connected layers and the sigmoid activation function, and generates relevant weights for the autoencoded features. The graph neural network is used to acquire the autoencoded features and construct a multi-dimensional physical map of the turbine rotor-blade system based on the relevant weights of the autoencoded features. The SVM model is used to acquire the turbine rotor-blade physical map. The multi-dimensional physical map of the turbine rotor-blade system maps the autoencoded features of the multi-dimensional physical map to a linearly separable space. Based on the cosine similarity algorithm, the similarity distance between the autoencoded features and the decision boundary points is determined. The fault participation factors associated with multi-dimensional fault classification in the autoencoded features are determined through a decision function. These fault participation factors are input into a bidirectional long short-term memory network based on an attention mechanism. The bidirectional long short-term memory network uses a fault case library of typical faults as prior knowledge and outputs fault mode probabilities. Using the fault participation factors and fault mode probabilities as input, the remaining lifetime distribution of the turbine rotor-blade system is obtained through Monte Carlo rolling prediction. The fault mode probabilities and remaining lifetime distribution are obtained, and the fault identification results of the turbine rotor-blade system are generated based on these probabilities and remaining lifetime distribution, and then visualized.

[0015] Preferably, the method for identifying and analyzing high-dimensional feature state vectors based on a pre-built fault identification model includes: A high-dimensional feature state vector is obtained, and the high-dimensional feature state vector is processed by autoencoding to obtain at least one set of multi-dimensional autoencoded features. The weight labeling layer learns the nonlinear cross-relationship of the autoencoded features in the graph neural network through a fully connected layer and a nonlinear activation function Sigmoid function, and generates the relevant weights of the autoencoded features. Obtain autoencoded features and construct a multi-dimensional physical map of turbine rotor-blade based on the relevant weights of the autoencoded features; Obtain the multi-dimensional physical map of the turbine rotor-blade and map the autoencoded features of the multi-dimensional physical map of the turbine rotor-blade to a linearly separable space; The similarity distance between the self-encoded features and the decision boundary points is determined based on the cosine similarity algorithm. The similarity distance between the self-encoded features and the decision boundary points is obtained, and the fault participation factors associated with multi-dimensional fault classification in the self-encoded features are determined through the decision function. The fault-related factors are input into a bidirectional long short-term memory network based on an attention mechanism. The bidirectional long short-term memory network uses a fault case library of typical faults as prior knowledge and outputs the probability of fault modes. The formula for calculating the failure mode probability is as follows:

[0016]

[0017]

[0018] in, Indicates the probability of the failure mode. The linear transformation vector of the fault participation factors. The first factor representing the fault participation factor A probability distribution vector of elements, These are the fault participation factor weight matrix, fault participation factor input vector, and bias vector, respectively.

[0019]

[0020]

[0021] in, Let represent the hidden state vectors of the forward LSTM (Long Short-Term Memory) network at time t and the backward LSTM network at time t, respectively. The input representation of the fault participation factor at the current time t; Using fault participation factors and fault mode probabilities as inputs, the remaining lifetime distribution of the turbine rotor-blade system is obtained through Monte Carlo rolling prediction. The failure mode probability and remaining lifetime distribution are obtained. Based on the failure mode probability and remaining lifetime distribution, the failure identification results of the turbine rotor-blade system are generated and visualized.

[0022] On the other hand, the present invention also provides a multimodal vibration analysis and fault identification system for a steam turbine rotor-blade, the system comprising: The 3D model construction module constructs a rotor-blade coupled 3D model that considers dynamic coupling effects based on the basic information of the turbine rotor-blade. The multimodal monitoring points in the turbine rotor-blade architecture are determined through the rotor-blade coupled 3D model, and the multimodal monitoring terminal is set at the multimodal monitoring point. The noise suppression module collects multimodal vibration signals of the turbine rotor-blade system in real time based on the multimodal monitoring terminal, performs noise suppression preprocessing on the multimodal vibration signals, and obtains noise-suppressed multimodal preprocessing information; The modal feature extraction module is used to load multimodal preprocessed information, extract modal features from the multimodal preprocessed information based on a convolutional neural network model guided by a coupled 3D model, and output a high-dimensional feature state vector characterizing the health status of the turbine rotor-blade system. The fault identification module is used to obtain high-dimensional feature state vectors, identify and analyze the high-dimensional feature state vectors based on the pre-built fault identification model, obtain the fault identification results of the turbine rotor-blade system, and visualize the fault identification results.

[0023] Preferably, the three-dimensional model construction module includes: The model reduction element is used to collect basic information of the turbine rotor-blade system. Based on the basic information of the turbine rotor-blade system, a three-dimensional finite element model of the turbine rotor-blade system is constructed. The fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of the turbine rotor-blade system to the interface principal modes and constraint modes, resulting in the reduced-order three-dimensional finite element model. Among them, the constraint modes include rotor flexibility mode, disk flexibility mode, blade flapping mode, and oscillation mode. Model coupling element is used to load the reduced-order three-dimensional finite element model. Nonlinear behavior connectors are introduced into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect. The model output unit simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector. Under typical fault conditions, it generates fault samples through a generative adversarial network. The rotor-blade coupled 3D model is iteratively trained using the fault samples to obtain a rotor-blade coupled 3D model that can cover typical faults.

[0024] Compared with the prior art, the embodiments of this application have the following main advantages: In this embodiment of the invention, by constructing a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect, accurate physical guidance can be provided for multimodal monitoring and feature extraction of vibration signals. By using frequency segmentation and adaptive threshold noise suppression preprocessing for noise in multimodal vibration signals, the problem that weak fault features in the original signal are easily submerged by noise is solved. Furthermore, by using a convolutional neural network model guided by the coupled three-dimensional model to extract modal features from the multimodal preprocessed information, early weak fault features can be clearly located and separated, making the generated high-dimensional feature state vector more comprehensively represent the health status of the system.

[0025] In this embodiment of the invention, a three-dimensional rotor-blade coupled model considering dynamic coupling effects is constructed using basic information of the turbine rotor-blade system. While ensuring high model accuracy, computational efficiency is significantly improved through order reduction and optimization techniques. The trained model can cover typical fault modes of the turbine rotor-blade system, making it suitable for real-time monitoring and fault diagnosis. At the same time, the model can output stress cloud diagrams and modal vibration frames under fault conditions, which can serve as physical guidance for feature extraction by convolutional neural networks, avoiding the black box nature of purely data-driven methods and improving the interpretability of fault identification.

[0026] In this embodiment of the invention, during the preprocessing of noise suppression for multimodal vibration signals, outliers and missing values ​​can eliminate non-vibration-related interference data at the source, ensuring data purity. By using Fourier transform to output the signal spectrum, the main frequency band and noise band can be clearly identified. When dividing the long and short windows, the window length can be flexibly set according to the bandwidth and duration of the main frequency band. The long and short window division results in long windows and short windows, which improve the frequency resolution of the main frequency band and enhance the noise tracking capability of the noise band. This takes into account the processing requirements of fault characteristics (low-frequency slow change) and noise (high-frequency fast change). Finally, by using Butterworth filters of different orders for vibration signals with different noise intensities, precise matching of strong noise matching strong filtering and weak noise matching weak filtering is achieved.

[0027] In this embodiment of the invention, when extracting modal features from multimodal preprocessed information, the guidance of the coupling model enables the CNN (Convolutional Neural Networks) to initially possess the physical correlation between faults and features. This avoids the black-box nature of purely data-driven models and improves the interpretability of features. By associating measured signals with physical simulation information, the simulated physical information can supplement the physical features of the fault. Furthermore, the simulated physical information provides deep damage features that cannot be directly obtained from measured signals, thus compensating for the deficiency that measured signals only reflect phenomena and not the essence. The convolutional neural network model can perform parallel feature extraction on heterogeneous simulated physical information and multimodal preprocessed information, avoiding information loss caused by forcibly fusing heterogeneous data in a single branch. The first branch retains the spatial details of the simulated physics, and the second branch retains the temporal dynamics of the vibration signal, ensuring the modal specificity of the two types of features. Finally, the high-dimensional feature state vector characterizing the health status of the turbine rotor-blade system integrates multi-dimensional information such as physical damage location, damage degree, and vibration response trend, achieving an upgrade from single fault type identification to a panoramic description of the health status.

[0028] In this embodiment of the invention, the fault identification model is based on a support vector machine (SVM) model and also includes an autoencoder and a weight calibration layer. The autoencoder effectively removes redundant information, and the sparsity of the autoencoded features directly corresponds to the physical damage of the fault, thus laying the foundation for the subsequent construction of a multi-dimensional physical map of the turbine rotor-blade. The multi-dimensional physical map of the turbine rotor-blade can transform abstract feature associations into visualized physical paths, and the faulty component can be accurately located through the node with the highest weight in the map. The SVM model can map the autoencoded features of the physical map to a linearly separable space and handle non-linear separability problems through kernel functions, maximizing the inter-class distance between different fault modes. Furthermore, the similarity distance is converted into a fault participation factor through a decision function, thereby replacing the traditional hard classification label and providing two-dimensional results of fault type and confidence, which better meets the risk quantification requirements of engineering decision-making. Finally, the fault mode probability and remaining life distribution provided support maintenance personnel to make comprehensive decisions, optimize maintenance plans, and reduce maintenance costs. Attached Figure Description

[0029] Figure 1 A schematic diagram of the implementation process of the multimodal vibration analysis and fault identification method for turbine rotor-blade is shown.

[0030] Figure 2 A schematic diagram of the implementation process of constructing a rotor-blade coupled three-dimensional model considering dynamic coupling effects based on the basic information of the turbine rotor-blade is shown.

[0031] Figure 3 A schematic diagram of the preprocessing method for noise suppression of multimodal vibration signals is shown.

[0032] Figure 4 A schematic diagram of the implementation process of modal feature extraction method for multimodal preprocessing information is shown.

[0033] Figure 5 The diagram illustrates the implementation process of a high-dimensional feature state vector identification and analysis method based on a pre-built fault identification model.

[0034] Figure 6 A schematic diagram of the structure of a turbine rotor-blade multimodal vibration analysis and fault identification system is shown. Detailed Implementation

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0036] To address the issue that early, weak faults can easily mask impact or modulation signals with background noise and vibrations from other components, resulting in fault characteristics appearing as complex sidebands or cross-modulations of frequency components across a wide frequency band, existing methods struggle to clearly locate and separate fault characteristics through simple spectral analysis. Therefore, we propose a multimodal vibration analysis and fault identification method and system for turbine rotor-blade systems. In short, the method first constructs a coupled three-dimensional model of the turbine rotor-blade system, considering dynamic coupling effects, based on the basic information of the turbine rotor-blade system. Noise suppression preprocessing is then performed on the multimodal vibration signals. Next, a convolutional neural network model guided by the coupled three-dimensional model is used to extract modal features from the preprocessed multimodal information. Finally, a pre-constructed fault identification model is used to identify and analyze the high-dimensional feature state vectors, yielding the fault identification results for the turbine rotor-blade system. In this embodiment of the invention, by constructing a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect, accurate physical guidance can be provided for multimodal monitoring and feature extraction of vibration signals. By using frequency segmentation and adaptive threshold noise suppression preprocessing for noise in multimodal vibration signals, the problem that weak fault features in the original signal are easily submerged by noise is solved. Furthermore, by using a convolutional neural network model guided by the coupled three-dimensional model to extract modal features from the multimodal preprocessed information, early weak fault features can be clearly located and separated, making the generated high-dimensional feature state vector more comprehensively represent the health status of the system.

[0037] This invention provides a method for multimodal vibration analysis and fault identification of a steam turbine rotor-blade. Figure 1 The diagram illustrates the implementation flow of a multimodal vibration analysis and fault identification method for a steam turbine rotor-blade system. Specifically, the method includes: S10. Based on the basic information of the turbine rotor-blade, construct a three-dimensional model of rotor-blade coupling that considers dynamic coupling effects. Use the three-dimensional model of rotor-blade coupling to determine the multi-modal monitoring points in the turbine rotor-blade architecture and set the multi-modal monitoring terminal at the multi-modal monitoring point. S20: Based on the multimodal monitoring terminal, multimodal vibration signals of the turbine rotor-blade system are collected in real time. Noise suppression preprocessing of the multimodal vibration signals is performed to obtain noise-suppressed multimodal preprocessing information. It should be noted that multimodal vibration signals refer to the set of vibration signals generated by different components, different vibration modes, and different physical quantity carriers during the operation of the turbine rotor-blade system. Multimodal monitoring points can be set at rotor journal, disk edge, blade root, and blade tip. The multimodal monitoring terminal can be a vibration sensor, velocity sensor, strain sensor, or temperature sensor.

[0038] S30, load multimodal preprocessing information, extract modal features from multimodal preprocessing information based on a convolutional neural network model guided by a coupled three-dimensional model, and output a high-dimensional feature state vector characterizing the health status of the turbine rotor-blade system; S40: Obtain high-dimensional feature state vectors, identify and analyze the high-dimensional feature state vectors based on the pre-built fault identification model, obtain the fault identification results of the turbine rotor-blade system, and visualize the fault identification results.

[0039] In this embodiment of the invention, by constructing a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect, accurate physical guidance can be provided for multimodal monitoring and feature extraction of vibration signals. By using frequency segmentation and adaptive threshold noise suppression preprocessing for noise in multimodal vibration signals, the problem that weak fault features in the original signal are easily submerged by noise is solved. Furthermore, by using a convolutional neural network model guided by the coupled three-dimensional model to extract modal features from the multimodal preprocessed information, early weak fault features can be clearly located and separated, making the generated high-dimensional feature state vector more comprehensively represent the health status of the system.

[0040] This invention provides a method for constructing a rotor-blade coupled three-dimensional model considering dynamic coupling effects based on the basic information of the turbine rotor-blade. Figure 2 The diagram illustrates the implementation process of a method for constructing a rotor-blade coupled 3D model considering dynamic coupling effects based on basic turbine rotor-blade information. The method specifically includes: S101. Collect basic information on the turbine rotor-blade system and construct a three-dimensional finite element model of the turbine rotor-blade system based on the basic information. When constructing the three-dimensional finite element model of the turbine rotor-blade system, geometric parameters, material properties, assembly relationships, and boundary conditions are used as core inputs to construct a three-dimensional finite element model with all elements, ensuring that the three-dimensional finite element model has high geometric and physical fidelity. S102, the fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of the turbine rotor-blade system to the main interface modes and constraint modes, resulting in a reduced-order three-dimensional finite element model. The constraint modes include rotor flexibility mode, disk compliance mode, blade flapping mode, and shimmy mode. In this embodiment of the invention, the fixed interface modal method is used to reduce the order of the three-dimensional finite element model, thereby preserving the main form of the overall vibration of the system and preserving the local modes that are strongly correlated with the fault. On the one hand, this can reduce the amount of dynamic analysis calculation, and on the other hand, it can ensure that the reduced-order model focuses on the fault-sensitive characteristics. S103, Load the reduced-order three-dimensional finite element model, introduce a nonlinear behavior connector into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model considering the dynamic coupling effect. In this embodiment of the invention, by introducing a nonlinear behavior connector, a nonlinear mechanical model is embedded at the component interface of the reduced-order model, thereby replacing the traditional rigid connection assumption, and thus realistically simulating the dynamic coupling effect between components. The parameters of the nonlinear connector can be adjusted according to the operating conditions, ensuring the dynamic adaptability of the model to the changing operating conditions of the steam turbine. S104 simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector, and generates fault samples through a generative adversarial network under typical fault conditions. The rotor-blade coupled three-dimensional model is iteratively trained using the fault samples to obtain a rotor-blade coupled three-dimensional model that can cover typical faults. The typical fault conditions can be rotor imbalance, blade crack, wheel loosening, and bearing oil film oscillation.

[0041] In this embodiment of the invention, a three-dimensional rotor-blade coupled model considering dynamic coupling effects is constructed using basic information of the turbine rotor-blade system. While ensuring high model accuracy, computational efficiency is significantly improved through order reduction and optimization techniques. The trained model can cover typical fault modes of the turbine rotor-blade system, making it suitable for real-time monitoring and fault diagnosis. At the same time, the model can output stress cloud diagrams and modal vibration frames under fault conditions, which can serve as physical guidance for feature extraction by convolutional neural networks, avoiding the black box nature of purely data-driven methods and improving the interpretability of fault identification.

[0042] This invention provides a preprocessing method for noise suppression of multimodal vibration signals. Figure 3 This diagram illustrates the implementation flow of a preprocessing method for noise suppression of multimodal vibration signals. The preprocessing method specifically includes: S201. Acquire multimodal vibration signals and process outliers and missing values ​​in the multimodal vibration signals to obtain cleaned multimodal vibration signals. Outlier processing can be done using box plots, while missing value processing can be done using spline interpolation. By eliminating outliers and missing values, interference data that is not essentially related to vibration can be removed at the source, ensuring the purity of the data.

[0043] S202, Load the cleaned multimodal vibration signal, perform spectrum analysis on the multimodal vibration signal, output the signal spectrum analysis results, identify the noise segment and the main frequency segment in the multimodal vibration signal, and divide the multimodal vibration signal into long and short windows based on the signal spectrum analysis results to obtain at least one set of time-series signal windows. Among them, the output signal spectrum through Fourier transform can clearly identify the main frequency segment and the noise segment. When dividing the long and short windows, the window length can be flexibly set according to the bandwidth and duration of the main frequency segment. The long and short window division results in long windows and short windows, so that the long window improves the frequency resolution of the main frequency segment, and the short window enhances the noise tracking capability of the noise segment, taking into account the processing needs of fault characteristics (low frequency slow change) and noise (high frequency fast change). S203: Acquire multimodal vibration signals within the time-series signal window, and determine the average density of the noise power spectrum within the time-series window based on the Periodogram spectrum estimation method. The determination of the average density of the noise power spectrum within the time-series window based on the Periodogram spectrum estimation method can transform the noise intensity from a subjective judgment to an objective numerical index, thereby providing a basis for adaptively selecting the filter order and achieving accurate matching of strong noise matching strong filtering and weak noise matching weak filtering. S204, determine whether the average density of the noise power spectrum within the time window exceeds the preset power spectrum density threshold. S205, if the average density of the noise power spectrum within the time window does not exceed the preset power spectrum density threshold, a second-order Butterworth filter is used to filter the multimodal vibration signal within the time signal window. The second-order filter is used to slightly suppress residual noise while ensuring that the main frequency band characteristics are not distorted. During filtering, the multimodal preprocessing information output after filtering is represented as follows:

[0044]

[0045] in, This represents the multimodal preprocessing information output after filtering. This represents the input of multimodal vibration signals within a time-series signal window. This indicates the filtering process performed by the Butterworth filter. These are the acquisition frequency and cutoff frequency of the multimodal vibration signal, respectively. Let the filter order be . This is the transfer function of the Butterworth filter; S206. If the average density of the noise power spectrum within the time window exceeds the preset power spectrum density threshold, a sixth-order Butterworth filter is used to filter the multimodal vibration signal within the time signal window. S207: Obtain the filtered multimodal preprocessing information, normalize the multimodal preprocessing information, and output the normalized multimodal preprocessing information.

[0046] In this embodiment of the invention, during the preprocessing of noise suppression for multimodal vibration signals, outliers and missing values ​​can eliminate non-vibration-related interference data at the source, ensuring data purity. By using Fourier transform to output the signal spectrum, the main frequency band and noise band can be clearly identified. When dividing the long and short windows, the window length can be flexibly set according to the bandwidth and duration of the main frequency band. The long and short window division results in long windows and short windows, which improve the frequency resolution of the main frequency band and enhance the noise tracking capability of the noise band. This takes into account the processing requirements of fault characteristics (low-frequency slow change) and noise (high-frequency fast change). Finally, by using Butterworth filters of different orders for vibration signals with different noise intensities, precise matching of strong noise matching strong filtering and weak noise matching weak filtering is achieved.

[0047] This invention provides a method for modal feature extraction from multimodal preprocessed information. Figure 4 This diagram illustrates the implementation flow of a method for modal feature extraction from multimodal preprocessed information. The method specifically includes: S301: A pre-built convolutional neural network model guided by a coupled 3D model is constructed. Fault samples are generated through a generative adversarial network. The fault samples are used to train the convolutional neural network model guided by the coupled 3D model, and a converged convolutional neural network model is output. The pre-constructed convolutional neural network model guided by the coupled 3D model is based on a convolutional neural network architecture, and a cross-attention mechanism is introduced into the convolutional neural network architecture. The convolutional neural network architecture uses the cross-attention mechanism to interactively fuse simulated physical features and multimodal vibration features. The convolutional neural network architecture includes three convolutional layers, a max pooling layer, and an average pooling layer. The convolutional neural network model also includes a parallel branch block, which includes a first feature extraction branch and a second feature extraction branch. The first feature extraction branch includes a ResNet50 layer, a batch normalization layer, and a max pooling layer. The second feature extraction branch includes a ResNet layer and also includes two convolutional layers, a max pooling layer, and a ReLU activation function. The loss function of the convolutional neural network model is expressed as:

[0048]

[0049] in, These are the second feature extraction branch, the first feature extraction branch, and the loss function of the convolutional neural network architecture, respectively. These are the second feature extraction branch, the first feature extraction branch, and the weight coefficients of the convolutional neural network architecture, respectively.

[0050] S302, load multimodal preprocessing information, generate adversarial network driven by multimodal preprocessing information to generate adversarial derivative physical information, rotor-blade coupled three-dimensional model to obtain adversarial derivative physical information, simulation output based on adversarial derivative physical information including stress cloud map and modal vibration frame; S303: Acquire simulated physical information and multimodal preprocessing information. Based on the parallel first and second feature extraction branches in the convolutional neural network model, feature extraction is performed on the simulated physical information and multimodal preprocessing information to obtain simulated physical features and multimodal vibration features. Among them, the residual connection of esNet50 solves the gradient vanishing problem of deep networks, enabling the first branch to extract more refined spatial features; the local perception characteristics of the convolutional layer enable the second branch to capture short-term abrupt changes in vibration signals. S304: Obtain simulated physical features and multimodal vibration features. Based on the cross-attention mechanism, the simulated physical features and multimodal vibration features are interactively fused to obtain a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system.

[0051] In this embodiment of the invention, when extracting modal features from multimodal preprocessed information, the guidance of the coupling model enables the CNN to initially possess the physical correlation between faults and features, avoiding the black box nature of purely data-driven models and improving the interpretability of features. By associating measured signals with physical simulation information, the simulated physical information can supplement the physical features of the fault. Furthermore, the simulated physical information provides deep damage features that cannot be directly obtained from measured signals, compensating for the deficiency that measured signals only reflect phenomena and not the essence. The convolutional neural network model can perform parallel feature extraction on heterogeneous simulated physical information and multimodal preprocessed information, avoiding information loss caused by forcibly fusing heterogeneous data in a single branch. The first branch retains the spatial details of the simulated physics, and the second branch retains the temporal dynamics of the vibration signal, ensuring the modal specificity of the two types of features. Finally, the high-dimensional feature state vector characterizing the health status of the turbine rotor-blade system integrates multi-dimensional information such as physical damage location, damage degree, and vibration response trend, achieving an upgrade from single fault type identification to a panoramic description of the health status.

[0052] In this embodiment of the invention, the fault identification model includes an input layer, a support vector machine (SVM) model, and an output layer. An autoencoder and a weight calibration layer are positioned between the input layer and the SVM model. The autoencoder is used to acquire a high-dimensional feature state vector and performs autoencoding processing on the high-dimensional feature state vector to obtain at least one set of multi-dimensional autoencoded features. The autoencoder is connected to the weight calibration layer, which is connected to the SVM model. The weight calibration layer includes a graph neural network, a global average pooling layer, and two fully connected layers. The weight calibration layer learns the nonlinear cross-relationships of the autoencoded features in the graph neural network through the fully connected layers and the nonlinear activation function Sigmoid function, and generates relevant weights for the autoencoded features. The graph neural network is used to acquire the autoencoded features and construct a multi-dimensional physical map of the turbine rotor-blade based on the relevant weights of the autoencoded features. The SVM model is used to acquire the turbine rotor-blade physical map. A multi-dimensional physical map of the turbine rotor-blade system is generated. The autoencoded features of the multi-dimensional physical map are mapped to a linearly separable space. The similarity distance between the autoencoded features and the decision boundary points is determined based on the cosine similarity algorithm. The fault participation factors associated with multi-dimensional fault classification in the autoencoded features are determined through a decision function. The fault participation factors are input into a bidirectional long short-term memory network based on an attention mechanism. The bidirectional long short-term memory network uses a fault case library of typical faults as prior knowledge and outputs the fault mode probability. The remaining lifetime distribution of the turbine rotor-blade system is obtained through Monte Carlo rolling prediction using the fault participation factors and the fault mode probability as input. The fault mode probability and the remaining lifetime distribution are obtained. The fault identification results of the turbine rotor-blade system are generated based on the fault mode probability and the remaining lifetime distribution and are visualized.

[0053] This invention provides a method for identifying and analyzing high-dimensional feature state vectors based on a pre-built fault identification model. Figure 5 The diagram illustrates the implementation flow of a method for identifying and analyzing high-dimensional feature state vectors based on a pre-built fault identification model. Specifically, this method includes: S401, obtain a high-dimensional feature state vector, perform autoencoding processing on the high-dimensional feature state vector to obtain at least one set of multi-dimensional autoencoded features, and the weight labeling layer learns the nonlinear cross-relation of autoencoded features in the graph neural network through a fully connected layer and a nonlinear activation function Sigmoid function, and generates the relevant weights of the autoencoded features; S402, acquire autoencoded features, and construct a multi-dimensional physical map of turbine rotor-blade based on the relevant weights of the autoencoded features; S403, acquire the multi-dimensional physical map of the turbine rotor-blade, and map the autoencoded features of the multi-dimensional physical map of the turbine rotor-blade to a linearly separable space; S404 determines the similarity distance between autoencoded features and decision boundary points based on the cosine similarity algorithm, obtains the similarity distance between autoencoded features and decision boundary points, and determines the fault participation factors in the autoencoded features that are associated with multi-dimensional fault classification through the decision function.

[0054] S405, the fault-related factors are input into a Bi-LSTM network based on an attention mechanism. The Bi-LSTM network uses a fault case library of typical faults as prior knowledge and outputs the probability of fault modes. Bi-LSTM can effectively process time series data, capture the changing patterns of fault features over time, and improve the accuracy of fault mode recognition. The attention mechanism can dynamically focus on important time steps and features, improve the model's sensitivity to key information, and enhance its predictive ability. The formula for calculating the failure mode probability is as follows:

[0055]

[0056]

[0057] in, Indicates the probability of the failure mode. The linear transformation vector of the fault participation factors. The first factor representing the fault participation factor A probability distribution vector of elements, These are the fault participation factor weight matrix, fault participation factor input vector, and bias vector, respectively.

[0058]

[0059]

[0060] in, Let represent the hidden state vectors of the forward LSTM and the backward LSTM at time t, respectively. The input representation of the fault participation factor at the current time t; S406 uses fault participation factors and fault mode probabilities as inputs to obtain the remaining life distribution of the turbine rotor-blade system through Monte Carlo rolling prediction. S407: Obtain the failure mode probability and remaining lifetime distribution, generate the failure identification results of the turbine rotor-blade system based on the failure mode probability and remaining lifetime distribution, and visualize the failure identification results.

[0061] In this embodiment of the invention, the fault identification model is based on a support vector machine (SVM) model and also includes an autoencoder and a weight calibration layer. The autoencoder effectively removes redundant information, and the sparsity of the autoencoded features directly corresponds to the physical damage of the fault, thus laying the foundation for the subsequent construction of a multi-dimensional physical map of the turbine rotor-blade. The multi-dimensional physical map of the turbine rotor-blade can transform abstract feature associations into visualized physical paths, and the faulty component can be accurately located through the node with the highest weight in the map. The SVM model can map the autoencoded features of the physical map to a linearly separable space and handle non-linear separability problems through kernel functions, maximizing the inter-class distance between different fault modes. Furthermore, the similarity distance is converted into a fault participation factor through a decision function, thereby replacing the traditional hard classification label and providing two-dimensional results of fault type and confidence, which better meets the risk quantification requirements of engineering decision-making. Finally, the fault mode probability and remaining life distribution provided support maintenance personnel to make comprehensive decisions, optimize maintenance plans, and reduce maintenance costs.

[0062] On the other hand, embodiments of the present invention also provide a multimodal vibration analysis and fault identification system for a steam turbine rotor-blade. Figure 6 The diagram shows a schematic of a multimodal vibration analysis and fault identification system for a steam turbine rotor-blade. This system specifically includes: The 3D model construction module 100 constructs a rotor-blade coupled 3D model considering dynamic coupling effects based on the basic information of the turbine rotor-blade. The multi-modal monitoring points in the turbine rotor-blade architecture are determined through the rotor-blade coupled 3D model, and the multi-modal monitoring terminal is set at the multi-modal monitoring point. The three-dimensional model construction module 100 includes: Model reduction element 110 is used to collect basic information of turbine rotor-blade. Based on the basic information of turbine rotor-blade, a three-dimensional finite element model of turbine rotor-blade system is constructed. The fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of turbine rotor-blade system to the main interface modes and constraint modes, and the reduced-order three-dimensional finite element model is obtained. Among them, the constraint modes include rotor flexibility mode, disk flexibility mode, blade flapping mode, and oscillation mode. Model coupling element 120 is used to load the reduced-order three-dimensional finite element model. Nonlinear behavior connectors are introduced into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect. Model output unit 130 simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector, and generates fault samples through a generative adversarial network under typical fault conditions. The rotor-blade coupled three-dimensional model is iteratively trained using the fault samples to obtain a rotor-blade coupled three-dimensional model that can cover typical faults.

[0063] The noise suppression module 200 collects multi-modal vibration signals of the turbine rotor-blade system in real time based on the multi-modal monitoring terminal, performs noise suppression preprocessing on the multi-modal vibration signals, and obtains noise-suppressed multi-modal preprocessing information. The modal feature extraction module 300 is used to load multimodal preprocessing information, extract modal features from the multimodal preprocessing information based on a convolutional neural network model guided by a coupled three-dimensional model, and output a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system. The fault identification module 400 is used to acquire high-dimensional feature state vectors, identify and analyze the high-dimensional feature state vectors based on the pre-built fault identification model, obtain the fault identification results of the turbine rotor-blade system, and visualize the fault identification results.

[0064] In summary, this invention provides a method and system for multimodal vibration analysis and fault identification of turbine rotor-blade. In the embodiments of this invention, by constructing a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect, accurate physical guidance can be provided for multimodal monitoring and feature extraction of vibration signals. Furthermore, by using frequency segmentation and adaptive threshold noise suppression preprocessing to suppress noise in multimodal vibration signals, the problem that weak fault features in the original signal are easily submerged by noise is solved. Moreover, by using a convolutional neural network model guided by the coupled three-dimensional model to extract modal features from the multimodal preprocessed information, early weak fault features can be clearly located and separated, making the generated high-dimensional feature state vector more comprehensively represent the health status of the system.

[0065] It should be noted that, for the sake of simplicity, the foregoing embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to the present invention. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0066] It should be understood that the disclosed apparatus can be implemented in other ways, given the several embodiments provided in this application. For example, the apparatus embodiments described above are merely illustrative; the division of units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or communication connections shown or discussed may be through some interfaces; the indirect coupling or communication connections between devices or units may be telecommunications or other forms.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for steam turbine rotor-blade multi-modal vibration analysis and fault identification, characterized in that, The method includes: Based on the basic information of the turbine rotor-blade, a three-dimensional model of rotor-blade coupling considering dynamic coupling effects is constructed. The multi-modal monitoring points in the turbine rotor-blade architecture are determined through the three-dimensional model of rotor-blade coupling, and the multi-modal monitoring terminal is set at the multi-modal monitoring point. Based on the real-time acquisition of multimodal vibration signals of the turbine rotor-blade system by the multimodal monitoring terminal, noise suppression preprocessing of the multimodal vibration signals is performed to obtain noise-suppressed multimodal preprocessing information; Load multimodal preprocessed information, extract modal features from the multimodal preprocessed information based on a convolutional neural network model guided by a coupled 3D model, and output a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system; High-dimensional feature state vectors are obtained, and the high-dimensional feature state vectors are identified and analyzed based on a pre-built fault identification model to obtain the fault identification results of the turbine rotor-blade system, and the fault identification results are presented in a visual manner.

2. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 1, wherein: The method for constructing a rotor-blade coupled three-dimensional model considering dynamic coupling effects based on the basic information of the turbine rotor-blade includes: Collect basic information on the turbine rotor and blades, and construct a three-dimensional finite element model of the turbine rotor and blade system based on the basic information on the turbine rotor and blades. The fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of the turbine rotor-blade system to the main interface modes and constraint modes, resulting in a reduced-order three-dimensional finite element model. The constraint modes include rotor flexibility mode, disk flexibility mode, blade flapping mode, and oscillation mode. The reduced-order three-dimensional finite element model is loaded, and a nonlinear behavior connector is introduced into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect. The model simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector. Fault samples are generated under typical fault conditions using a generative adversarial network. The rotor-blade coupled 3D model is iteratively trained using the fault samples to obtain a rotor-blade coupled 3D model that can cover typical faults.

3. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 1, wherein: The method for noise suppression preprocessing of multimodal vibration signals includes: Multimodal vibration signals are acquired, and outlier and missing values ​​are processed to obtain cleaned multimodal vibration signals. After loading the cleaned multimodal vibration signal, perform spectral analysis on the multimodal vibration signal, output the signal spectral analysis results, identify the noise segment and the main frequency segment in the multimodal vibration signal, and divide the multimodal vibration signal into long and short windows based on the signal spectral analysis results to obtain at least one set of time sequence signal windows; The multimodal vibration signal within the time-series signal window is acquired, and the average density of the noise power spectrum within the time-series window is determined based on the Periodogram spectrum estimation method. It is then determined whether the average density of the noise power spectrum within the time-series window exceeds the preset power spectrum density threshold. If the average density of the noise power spectrum within the time window does not exceed the preset power spectrum density threshold, a second-order Butterworth filter is used to filter the multimodal vibration signal within the time window. If the average density of the noise power spectrum within the time window exceeds the preset power spectrum density threshold, a sixth-order Butterworth filter is used to filter the multimodal vibration signal within the time window. The system acquires the filtered multimodal preprocessing information, normalizes the multimodal preprocessing information, and outputs the normalized multimodal preprocessing information.

4. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 3, wherein: Methods for modal feature extraction from multimodal preprocessed information include: A pre-built convolutional neural network model guided by a coupled 3D model is constructed. Fault samples are generated through a generative adversarial network. The convolutional neural network model guided by the coupled 3D model is trained using the fault samples, and a converged convolutional neural network model is output. Loading multimodal preprocessing information, the Generative Adversarial Network (GAN) is driven by the multimodal preprocessing information to generate adversarial derivative physical information. The rotor-blade coupled three-dimensional model obtains the adversarial derivative physical information, and the simulation output based on the adversarial derivative physical information includes stress cloud diagrams and modal vibration mode frames. Simulated physical information and multimodal preprocessing information are acquired. Based on the first and second feature extraction branches in the convolutional neural network model, features are extracted from the simulated physical information and multimodal preprocessing information to obtain simulated physical features and multimodal vibration features. Simulated physical features and multimodal vibration features are acquired, and the simulated physical features and multimodal vibration features are interactively fused based on the cross-attention mechanism to obtain a high-dimensional feature state vector characterizing the health state of the turbine rotor-blade system.

5. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 4, wherein: When constructing the pre-built convolutional neural network model guided by the coupled 3D model, the convolutional neural network architecture is used as the basic architecture, and a cross-attention mechanism is introduced into the convolutional neural network architecture. The convolutional neural network architecture uses the cross-attention mechanism to interactively fuse simulated physical features and multimodal vibration features. The convolutional neural network architecture includes three convolutional layers, a max pooling layer, and an average pooling layer. The convolutional neural network model also includes a parallel branch block, which includes a first feature extraction branch and a second feature extraction branch. The first feature extraction branch includes a ResNet50 layer, a batch normalization layer, and a max pooling layer. The second feature extraction branch includes a ResNet layer, and the second feature extraction branch also includes two convolutional layers, a max pooling layer, and a ReLU activation function.

6. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 1, wherein: The fault identification model includes an input layer, a support vector machine model, and an output layer. An autoencoder and a weight calibration layer are set between the input layer and the support vector machine model. The autoencoder is used to obtain a high-dimensional feature state vector and performs autoencoding processing on the high-dimensional feature state vector to obtain at least one set of multi-dimensional autoencoded features. The autoencoder is connected to the weight calibration layer, and the weight calibration layer is connected to the support vector machine model. The weight calibration layer includes a graph neural network, a global average pooling layer, and two fully connected layers.

7. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 6, wherein: The method for identifying and analyzing high-dimensional feature state vectors based on a pre-built fault identification model includes: A high-dimensional feature state vector is obtained, and the high-dimensional feature state vector is processed by autoencoding to obtain at least one set of multi-dimensional autoencoded features. The weight labeling layer learns the nonlinear cross-relationship of the autoencoded features in the graph neural network through a fully connected layer and a nonlinear activation function Sigmoid function, and generates the relevant weights of the autoencoded features. Obtain autoencoded features and construct a multi-dimensional physical map of turbine rotor-blade based on the relevant weights of the autoencoded features; Obtain the multi-dimensional physical map of the turbine rotor-blade and map the autoencoded features of the multi-dimensional physical map of the turbine rotor-blade to a linearly separable space; The similarity distance between the self-encoded features and the decision boundary points is determined based on the cosine similarity algorithm. The fault participation factors associated with multi-dimensional fault classification in the self-encoded features are determined through the decision function.

8. The gas turbine rotor-blade multi-modal vibration analysis and fault identification method of claim 7, wherein: The method for identifying and analyzing high-dimensional feature state vectors based on a pre-built fault identification model further includes: The fault-related factors are input into a bidirectional long short-term memory network based on an attention mechanism. The bidirectional long short-term memory network uses a fault case library of typical faults as prior knowledge and outputs the probability of fault modes. Using fault participation factors and fault mode probabilities as inputs, the remaining lifetime distribution of the turbine rotor-blade system is obtained through Monte Carlo rolling prediction. The failure mode probability and remaining lifetime distribution are obtained. Based on the failure mode probability and remaining lifetime distribution, the failure identification results of the turbine rotor-blade system are generated and visualized.

9. A multimodal vibration analysis and fault identification system for a steam turbine rotor-blade, used to implement the multimodal vibration analysis and fault identification method for a steam turbine rotor-blade as described in any one of claims 1-8, characterized in that: the system comprises: The 3D model construction module constructs a rotor-blade coupled 3D model that considers dynamic coupling effects based on the basic information of the turbine rotor-blade. The multi-modal monitoring points in the turbine rotor-blade architecture are determined through the rotor-blade coupled 3D model, and the multi-modal monitoring terminal is set at the multi-modal monitoring point. The noise suppression module collects multimodal vibration signals of the turbine rotor-blade system in real time based on the multimodal monitoring terminal, performs noise suppression preprocessing on the multimodal vibration signals, and obtains noise-suppressed multimodal preprocessing information; The modal feature extraction module is used to load multimodal preprocessed information, extract modal features from the multimodal preprocessed information based on a convolutional neural network model guided by a coupled 3D model, and output a high-dimensional feature state vector characterizing the health status of the turbine rotor-blade system. The fault identification module is used to obtain high-dimensional feature state vectors, identify and analyze the high-dimensional feature state vectors based on the pre-built fault identification model, obtain the fault identification results of the turbine rotor-blade system, and visualize the fault identification results.

10. The gas turbine rotor-blade multi-modal vibration analysis and fault identification system of claim 9, wherein: The 3D model construction module includes: The model reduction element is used to collect basic information of the turbine rotor-blade system. Based on the basic information of the turbine rotor-blade system, a three-dimensional finite element model of the turbine rotor-blade system is constructed. The fixed interface modal method is used to reduce the degrees of freedom of the three-dimensional finite element model of the turbine rotor-blade system to the interface principal modes and constraint modes, resulting in the reduced-order three-dimensional finite element model. Among them, the constraint modes include rotor flexibility mode, disk flexibility mode, blade flapping mode, and oscillation mode. Model coupling element is used to load the reduced-order three-dimensional finite element model. Nonlinear behavior connectors are introduced into the three-dimensional finite element model to form a rotor-blade coupled three-dimensional model that considers the dynamic coupling effect. The model output unit simulates typical faults in the turbine rotor-blade system based on a nonlinear behavior connector. Under typical fault conditions, it generates fault samples through a generative adversarial network. The rotor-blade coupled 3D model is iteratively trained using the fault samples to obtain a rotor-blade coupled 3D model that can cover typical faults.

Images