Turbonator rotor fault on-line diagnosis method based on multi-modal vibration characteristics

Abstract

The invention relates to the technical field of steam turbine generator state monitoring and fault diagnosis, in particular to a steam turbine generator rotor fault on-line diagnosis method based on multi-modal vibration characteristics, which comprises the following steps of: acquiring vibration signals of rotor bearing seat measuring points, marking fault types, constructing a supervised data set, and then, constructing a supervised data set; the method comprises the following steps: firstly, carrying out noise reduction preprocessing on a signal by adopting adaptive wavelet transform to obtain a preprocessed pure vibration signal, secondly, carrying out joint optimization on key parameters of variational mode decomposition through a hybrid optimization strategy to carry out mode decomposition on the pure vibration signal, screening out an effective mode component, and carrying out noise reduction on the effective mode component; the method comprises the steps of extracting time domain and frequency domain features to construct a multi-modal feature set, and finally performing training and diagnosis by using a double-layer bidirectional long-short-term memory network model to obtain a rotor fault diagnosis model, thereby realizing accurate capture of fault features, effectively improving diagnosis accuracy, and providing support for preventing fault expansion and guaranteeing safe operation of a unit.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steam turbine generator state monitoring and fault diagnosis, in particular to a steam turbine generator rotor fault online diagnosis method based on multi-modal vibration characteristics. BACKGROUND

[0002] As the core power generation equipment of the power system, the rotor operating state of the steam turbine generator directly determines the safety and stability of the unit and the power supply reliability. Under the long-term high-speed, high-temperature, and alternating load working conditions, the rotor is prone to typical faults such as unbalance, inter-turn short circuit, bearing wear, shaft bending, and steam-induced vibration due to mechanical wear, electromagnetic abnormalities, thermal deformation, etc. If not diagnosed and treated in time, it may lead to fault expansion, unplanned shutdown, and even equipment damage and safety accidents, causing significant economic losses. Therefore, accurate and real-time diagnosis of rotor faults is a key requirement to ensure the safe and efficient operation of power plants.

[0003] At present, steam turbine generator fault diagnosis is mostly based on vibration signal analysis, but the existing technology still has significant limitations and cannot meet the requirements of high precision and high real-time in the industrial field. For example, the patent CN119723158A "Fault diagnosis method based on vibration of stator end of steam turbine generator" converts the vibration signal into a two-dimensional image through Markov transition field and Gram angle field, and uses a VGG network with global attention mechanism to extract texture features to realize diagnosis. However, this method has a single feature extraction dimension and only focuses on spatial texture distribution, failing to fully exploit the time-domain features of the vibration signal. At the same time, it uses empirical mode decomposition (EMD) for noise reduction, which is prone to modal aliasing, affecting feature purity and diagnosis accuracy. For another example, the patent CN110702394A "Steam turbine generator set vibration fault diagnosis method based on vibration change characteristics" uses regularized features such as vibration amplitude fluctuation, jump time, and repeatability, and combines artificial threshold for fault judgment. This method is simple to implement, but it only relies on shallow statistical features and cannot capture the dynamic relationships between features, especially in the early stages of failure, when the feature signal amplitude is weak and easily overwhelmed by noise, often resulting in missed and false judgments.

[0004] In summary, the existing technology still has deficiencies in key aspects such as feature extraction comprehensiveness and signal preprocessing accuracy, and it is difficult to effectively deal with the non-stationarity and complexity of steam turbine generator vibration signals. Therefore, there is an urgent need for a fault diagnosis method that can integrate multi-dimensional feature extraction and intelligent optimization modeling mechanisms to address the characteristics of rotor vibration signals, improve the accuracy and reliability of diagnosis, and achieve efficient and intelligent identification of steam turbine generator rotor faults.

[0005] To solve the above-mentioned defects, a technical solution is provided. SUMMARY

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides an online fault diagnosis method for steam turbine generator rotors based on multimodal vibration characteristics.

[0007] To achieve the above objectives, the present invention can be implemented through the following technical solutions:

[0008] This invention provides an online fault diagnosis method for steam turbine generator rotors based on multimodal vibration characteristics, comprising the following steps:

[0009] Step 1: Collect vibration signals from the measuring points of the turbine generator rotor bearing housing, and label the collected vibration signals based on preset typical fault types to form a supervised dataset;

[0010] Step 2: Perform wavelet transform noise reduction preprocessing on the collected vibration signal to obtain a clean vibration signal after preprocessing;

[0011] Step 3: A hybrid optimization strategy of arithmetic optimization algorithm and sparrow search algorithm is adopted to jointly optimize the key parameters of variational mode decomposition, and the optimized parameters are used to perform mode decomposition on the pure vibration signal to extract pure multimodal vibration features and screen out effective modal components.

[0012] Step 4: Extract time-domain and frequency-domain feature parameters from the effective modal components to construct an effective modal feature set;

[0013] Step 5: Construct a two-layer bidirectional long short-term memory network model and train it using an effective modal feature set to obtain a rotor fault diagnosis model;

[0014] Step 6: After processing the collected vibration signals of the turbine generator rotor according to steps 2 to 4, import them as input data into the trained rotor fault diagnosis model, output the probability of each typical fault type, and take the typical fault type with the highest probability as the final fault diagnosis result.

[0015] Furthermore, typical fault types may include rotor imbalance, inter-turn short circuit, bearing wear, shaft bending, and steam flow excitation.

[0016] Furthermore, the collected vibration signals undergo wavelet transform noise reduction preprocessing. The specific processing steps are as follows:

[0017] The vibration signal is decomposed into five levels using the selected db4 wavelet basis to obtain the wavelet coefficients corresponding to each level.

[0018] The wavelet coefficients corresponding to the first and second layers are labeled as high-frequency coefficients and filtered out using the threshold suppression method. The wavelet coefficients corresponding to the third to fifth layers are labeled as mid-to-low-frequency coefficients and retained. The retained mid-to-low-frequency coefficients are then used for signal reconstruction to obtain a pure vibration signal.

[0019] Furthermore, the specific process for jointly optimizing the key parameters of variational mode decomposition is as follows:

[0020] Based on the main frequency range of the rotor vibration signal, the range of the mode number k in the variational mode decomposition is set to 3 to 7, and the range of the penalty factor α is set to 800 to 2500. A hybrid optimization strategy of arithmetic optimization algorithm and sparrow search algorithm is adopted to jointly optimize the parameters of variational mode decomposition.

[0021] The overall parameters of the optimization algorithm are set as follows: population size is 20, maximum number of iterations is 50, and the entire optimization process is divided into two parts: global exploration stage and local refinement stage.

[0022] Global exploration phase: An arithmetic optimization algorithm is used for global search, and the search space is expanded by addition and subtraction operators;

[0023] The update rules for addition and subtraction operations are defined as follows:

[0024] Addition operation:

[0025] Subtraction operation: in, This is the optimal solution for the t-th generation. T=50 represents the total number of iterations. The numbers are uniformly distributed random numbers;

[0026] Local refinement stage: Switch to the sparrow search algorithm to fine-tune the parameters;

[0027] The formula for updating the discoverer's location is as follows: , It is a constant;

[0028] The follower update formula is: , , where is a random variable that follows a normal distribution;

[0029] When an individual's fitness is less than 0.8 times the population's average fitness, a warning mechanism is triggered: , , is a random perturbation factor;

[0030] The objective function is optimized to minimize the weighted envelope, and its formula is as follows: ;

[0031] Among them, envelope extraction: , Let the probability distribution of the modal envelope be the energy summary. , Let i be the energy of the i-th mode. Total energy;

[0032] Therefore, the optimal parameters are obtained through the above hybrid optimization algorithm.

[0033] Furthermore, the specific process for screening effective modal components is as follows:

[0034] Variational mode decomposition was performed on the pure vibration signal using optimal parameters to obtain five intrinsic mode functions (IMF1 to IMF5), whose calculation formulas are as follows: , where u k (t) is the kth mode obtained from the decomposition, ω k It is the center frequency, and δ(t) is the Dirac function;

[0035] Modal analysis results show that IMF1 to IMF2 are the dominant high-frequency noise components, and IMF3 to IMF5 are the dominant fault characteristic components.

[0036] Effective modal components were screened based on dual indicators, and the selected results were confirmed as effective modal components.

[0037] Furthermore, the specific process of constructing an effective modal feature set is as follows:

[0038] Extract temporal feature parameters from effective modal components. Temporal feature parameters may include kurtosis, peak factor, and impulse factor.

[0039] Perform a fast Fourier transform on the effective modal components to convert the time-domain signal into a frequency-domain signal, and extract its frequency-domain characteristic parameters, which may include characteristic harmonic amplitude, harmonic energy ratio, and frequency band energy ratio.

[0040] The extracted time-domain and frequency-domain feature parameters are standardized and normalized, and the standardized features are then fused to construct an effective modal feature set.

[0041] Furthermore, the structural design for constructing a two-layer bidirectional long short-term memory network model includes:

[0042] Input layer: Vectorizes multimodal features, transforming them into a sequence of input features that the network can recognize;

[0043] Bidirectional LSTM layer: It adopts a two-layer stacked structure, with several hidden units in each layer and the activation function is tanh. The forward LSTM is used to capture the future temporal correlation of features, and the backward LSTM is used to capture the historical temporal correlation of features. The forward and backward LSTMs are concatenated and the output dimension is G.

[0044] Dropout layer: A dropout mechanism is added after the bidirectional LSTM layer, with the dropout probability set to F;

[0045] Output layer: Through a fully connected layer and a Softmax activation function, the probability distribution vectors of each typical fault category are output;

[0046] Weight initialization: The weight parameters of the LSTM layer and the fully connected layer are initialized using the He normal distribution.

[0047] Furthermore, the specific process of constructing the rotor fault diagnosis model is as follows:

[0048] Effective modal feature components are extracted from the effective modal feature set as training samples, cross-entropy is used as the loss function, and the Adam optimizer is used to adaptively and iteratively update the model parameters.

[0049] A validation set is introduced during the training process. The learning rate and network parameters are dynamically adjusted based on the diagnostic accuracy feedback of the validation set until the model achieves optimal performance on the validation set. Training is then stopped, and the corresponding weight parameters are saved to form a rotor fault diagnosis model.

[0050] The technical solution provided by this invention has the following advantages compared with the known prior art:

[0051] 1. High preprocessing accuracy: This invention targets the non-stationary characteristics of turbine generator rotor vibration signals, selects wavelet bases that adaptively match the signal features for multi-scale decomposition, effectively suppresses noise while maintaining key fault features, achieves a significant improvement in signal-to-noise ratio, and provides high-quality input data for subsequent feature extraction and recognition.

[0052] 2. High accuracy of mode decomposition: By introducing the hybrid optimization variational mode decomposition (VMD) algorithm and combining it with a global optimization strategy, the problem of traditional single algorithms easily getting trapped in local optima is avoided. The mode aliasing phenomenon is significantly reduced, ensuring that the physical meaning of each mode component is clear and the purity is high, thereby achieving accurate separation of multi-scale features of vibration signals.

[0053] 3. Excellent feature extraction: Constructing a multimodal vibration feature system that takes into account time domain, frequency domain and time-frequency domain feature information can comprehensively characterize the "impact-period" composite characteristics of rotor faults, effectively avoid the problems of missed and false judgments caused by a single feature dimension, and significantly improve the robustness and accuracy of fault identification.

[0054] 4. Superior diagnostic performance: Through full-process algorithm optimization and feature fusion design, accurate identification and real-time diagnosis of turbine generator rotor faults are achieved, significantly improving fault detection sensitivity and model generalization ability, and providing reliable technical support for the safe and stable operation of power plant units. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the overall process of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1 As shown, an online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics includes the following steps:

[0059] Step 1: Collect vibration signals from the measuring points of the turbine generator rotor bearing housing, and label the collected vibration signals based on preset typical fault types to form a supervised dataset;

[0060] Typical fault types include, but are not limited to, rotor imbalance, inter-turn short circuit, bearing wear, shaft bending, and steam flow excitation.

[0061] The annotation process is based on the vibration characteristics of various rotor faults. For example, the characteristic of rotor imbalance fault is the abnormal proportion of octave amplitude, and the characteristic of bearing wear fault is the significant enhancement of high-frequency impact signal. In this way, a supervised dataset covering normal operation and various typical fault states is established to provide basic data support for the training of subsequent fault diagnosis models.

[0062] Step 2: Perform wavelet transform noise reduction preprocessing on the collected vibration signal to obtain a clean vibration signal after preprocessing. The specific implementation process is as follows:

[0063] To effectively suppress high-frequency electromagnetic noise (frequency range of 10-12kHz, originating from power grid harmonic interference) and low-frequency sensor zero drift (frequency range of 0.1-1Hz, originating from sensor temperature drift) in vibration signals, wavelet transform method is used.

[0064] The vibration signal was decomposed into five layers using the selected db4 wavelet basis to obtain the wavelet coefficients corresponding to each layer. The frequency ranges of each layer are as follows: layer 1 (6-12kHz), layer 2 (3-6kHz), layer 3 (1.5-3kHz), layer 4 (0.75-1.5kHz), and layer 5 (0.375-0.75kHz).

[0065] Specifically, the wavelet coefficients corresponding to the first and second layers are labeled as high-frequency coefficients and filtered out using the threshold suppression method. The wavelet coefficients corresponding to the third to fifth layers are labeled as mid-to-low-frequency coefficients and retained. The retained mid-to-low-frequency coefficients are used for signal reconstruction to obtain a pure vibration signal after removing noise interference, providing high-quality input data for subsequent fault feature extraction and model recognition.

[0066] Step 3: A hybrid optimization strategy of arithmetic optimization and sparrow search algorithm is adopted to jointly optimize the key parameters of variational mode decomposition (VMD). The optimized parameters are then used to perform mode decomposition on the pure vibration signal to extract pure multimodal vibration features and screen out effective modal components. The specific implementation process is as follows:

[0067] Based on the main frequency range of the rotor vibration signal (0.375~3kHz), the range of the mode number k in variational mode decomposition (VMD) is set to 3~7, and the range of the penalty factor α is set to 800~2500. A hybrid optimization strategy of arithmetic optimization algorithm and sparrow search algorithm is adopted to jointly optimize the key parameters of variational mode decomposition (VMD). The key parameters include the mode number and the penalty factor.

[0068] The overall parameters of the optimization algorithm are set as follows: population size is 20, maximum number of iterations is 50, and the entire optimization process is divided into two parts: global exploration stage and local refinement stage.

[0069] Global exploration phase (first 15 rounds, accounting for 30% of the total iterations): Arithmetic optimization algorithm is used for global search, and the search space is expanded by addition and subtraction operators to achieve diversified distribution of parameters;

[0070] The update rules for addition and subtraction operations are defined as follows:

[0071] Addition operation:

[0072] Subtraction operation: in, This is the optimal solution for the t-th generation. (T=50 is the total number of iterations) (Uniformly distributed random numbers);

[0073] Local refinement stage (last 35 rounds, accounting for 70% of the total iterations): switch to the sparrow search algorithm, and finely tune the parameters through the collective intelligence mechanism of "discoverer-follower-watcher".

[0074] The formula for updating the discoverer's location is as follows: ( (where the constant is)

[0075] The follower update formula is: , ( (where is a random variable that follows a normal distribution).

[0076] When an individual's fitness is less than 0.8 times the population's average fitness, the vigilant mechanism is triggered. , ( (where is a random perturbation factor to avoid getting trapped in local optima).

[0077] The objective function is optimized to minimize the weighted envelope, and its formula is as follows: ;

[0078] Among them, envelope extraction: ( (Probability distribution of modal envelope), characterizing the irregularity of modal signals, energy abstraction: ( Let i be the energy of the i-th mode. (total energy), characterizing the uniformity of energy distribution;

[0079] The optimization process aims to obtain mode decomposition results that are both pure and energy-concentrated, thereby effectively avoiding mode aliasing.

[0080] The optimal parameters obtained by the above hybrid optimization algorithm are: k=5, α=1800. Using these optimal parameters, variational mode decomposition (VMD) is performed on the pure vibration signal reconstructed in step two to obtain 5 intrinsic mode functions (IMF1~IMF5), and their calculation formulas are as follows: , where u k (t) is the kth mode obtained from the decomposition, ω k It is the center frequency, and δ(t) is the Dirac function;

[0081] Modal analysis results show that IMF1 to IMF2 are the dominant high-frequency noise components (kurtosis of 2.1 and 2.8 respectively, both less than 3.5), and IMF3 to IMF5 are the dominant fault characteristic components (corresponding to bearing wear impact characteristics, rotor imbalance primary harmonic frequency, and inter-turn short circuit secondary harmonic frequency signal respectively).

[0082] Subsequently, effective modal components were screened based on dual indicators:

[0083] 1. Kurtosis value > 3.5 to ensure impact-related fault information is included;

[0084] 2. The Pearson correlation coefficient with the original vibration signal is >0.6 to ensure a strong correlation with the rotor state;

[0085] The final screening results are as follows: IMF3 (kurtosis 4.5, correlation coefficient 0.75), IMF4 (kurtosis 3.9, correlation coefficient 0.68), and IMF5 (kurtosis 3.7, correlation coefficient 0.64) were identified as effective modal components, while IMF1 and IMF2 were removed due to noise dominance.

[0086] Through the above processing, high-precision extraction of key fault features from vibration signals is achieved, providing high-confidence input features for subsequent fault identification models.

[0087] Step 4: Extract time-domain and frequency-domain feature parameters from the effective modal components to construct an effective modal feature set. The specific implementation process is as follows:

[0088] Temporal feature parameters are extracted from the effective modal components. These temporal feature parameters include, but are not limited to, kurtosis: used to characterize the impulsiveness of the signal and reflect the significance of transient impact components in the vibration signal.

[0089] Peak factor: Used to reflect the ratio between the peak value and the effective value of a signal, in order to evaluate the peak change characteristics in a vibration signal;

[0090] Pulse factor: Used to highlight the characteristics of instantaneous impulse amplitude changes in a signal, in order to reveal the strength of the fault impulse behavior;

[0091] Perform a Fast Fourier Transform (FFT) on the effective modal components to convert the time-domain signal into a frequency-domain signal, and extract its frequency-domain feature parameters, including but not limited to:

[0092] Characteristic harmonic amplitude: Used to identify amplitude variations of harmonic components in fault signals.

[0093] Harmonic energy ratio: Used to measure the proportion of harmonic components in the system's energy, reflecting the periodicity and stability of the system's vibration.

[0094] Frequency band energy percentage; used to describe the energy distribution characteristics of different frequency bands in order to identify the energy concentration area corresponding to the fault type;

[0095] The extracted time-domain and frequency-domain feature parameters are standardized to eliminate scale differences between features of different dimensions. The standardized features are then fused to construct an effective modal feature set.

[0096] Step 5: Construct a two-layer bidirectional long short-term memory network model and train it using an effective modal feature set to obtain the rotor fault diagnosis model. The specific implementation process is as follows:

[0097] Construct a two-layer bidirectional long short-term memory (LSTM) network model. The model structure is designed as follows:

[0098] Input layer: Vectorizes multimodal features, transforming them into a sequence of input features that the network can recognize;

[0099] Bidirectional LSTM layer: A two-layer stacked structure is adopted, with each layer having a number of hidden units (up to 128 hidden units) and the activation function being tanh. The forward LSTM is used to capture the future temporal correlation of features (such as the amplitude decay trend after bearing wear impact), and the backward LSTM is used to capture the historical temporal correlation of features (such as the frequency fluctuation pattern before the impact). By splicing the forward and backward outputs, a global feature cognition of the fault dynamic process is achieved. The output dimension after splicing is G (G is 256).

[0100] Dropout layer: A Dropout mechanism is added after the bidirectional LSTM layer, with the dropout probability set to F (F=0.2) to randomly deactivate some neurons, thereby suppressing the risk of overfitting and improving the model's generalization ability;

[0101] Fully connected layer and output layer: Set up a hidden fully connected layer (256 neurons, activation function is ReLU), and introduce the Softmax activation function in the output layer to output the probability distribution vector of each typical fault category;

[0102] Weight initialization: The weight parameters of the LSTM layer and the fully connected layer are initialized using the He normal distribution to prevent gradient vanishing or gradient explosion and ensure model convergence stability.

[0103] Secondly, the two-layer bidirectional long short-term memory network (LSTM) model is trained and optimized: effective modal feature components are extracted from the effective modal feature set as training samples, cross-entropy is used as the loss function, and the Adam optimizer is used to adaptively update the model parameters iteratively.

[0104] A validation set is introduced during the training process. The learning rate and network parameters are dynamically adjusted based on the diagnostic accuracy feedback of the validation set until the model performs optimally on the validation set. Training is then stopped and the corresponding weight parameters are saved to form the final rotor fault diagnosis model.

[0105] By constructing and training the bidirectional LSTM model, bidirectional feature learning of rotor vibration signals in the time series dimension is realized. This enables comprehensive analysis of the dynamic evolution before and after a fault occurs, significantly improving the accuracy of fault identification and early diagnosis capabilities.

[0106] Step 6: After processing the collected vibration signals of the turbine generator rotor according to steps 2 to 4, import them as input data into the trained rotor fault diagnosis model, output the probability of each typical fault type, and take the typical fault type with the highest probability as the final fault diagnosis result.

[0107] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement any of the methods described above;

[0108] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the methods described above.

[0109] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for online diagnosis of rotor faults in a steam turbine generator based on multimodal vibration characteristics, characterized in that, Includes the following steps: Step 1: Collect vibration signals from the measuring points of the turbine generator rotor bearing housing, and label the collected vibration signals based on preset typical fault types to form a supervised dataset; Step 2: Perform wavelet transform noise reduction preprocessing on the collected vibration signal to obtain a clean vibration signal after preprocessing; Step 3: A hybrid optimization strategy of arithmetic optimization algorithm and sparrow search algorithm is adopted to jointly optimize the key parameters of variational mode decomposition, and the optimized parameters are used to perform mode decomposition on the pure vibration signal to extract pure multimodal vibration features and screen out effective modal components. Step 4: Extract time-domain and frequency-domain feature parameters from the effective modal components to construct an effective modal feature set; Step 5: Construct a two-layer bidirectional long short-term memory network model and train it using an effective modal feature set to obtain a rotor fault diagnosis model; Step 6: After processing the collected vibration signals of the turbine generator rotor according to steps 2 to 4, import them as input data into the trained rotor fault diagnosis model, output the probability of each typical fault type, and take the typical fault type with the highest probability as the final fault diagnosis result.

2. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, Typical fault types may include rotor imbalance, inter-turn short circuit, bearing wear, shaft bending, and steam flow excitation.

3. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, The collected vibration signals were preprocessed using wavelet transform for noise reduction. The specific processing steps are as follows: The vibration signal is decomposed into five levels using the selected db4 wavelet basis to obtain the wavelet coefficients corresponding to each level. The wavelet coefficients corresponding to the first and second layers are labeled as high-frequency coefficients and filtered out using the threshold suppression method. The wavelet coefficients corresponding to the third to fifth layers are labeled as mid-to-low-frequency coefficients and retained. The retained mid-to-low-frequency coefficients are then used for signal reconstruction to obtain a pure vibration signal.

4. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, The specific process of jointly optimizing the key parameters of variational mode decomposition is as follows: Based on the main frequency range of the rotor vibration signal, the range of the mode number k in the variational mode decomposition is set to 3 to 7, and the range of the penalty factor α is set to 800 to 2500. A hybrid optimization strategy of arithmetic optimization algorithm and sparrow search algorithm is adopted to jointly optimize the parameters of variational mode decomposition. The overall parameters of the optimization algorithm are set as follows: population size is 20, maximum number of iterations is 50, and the entire optimization process is divided into two parts: global exploration stage and local refinement stage. Global exploration phase: An arithmetic optimization algorithm is used for global search, and the search space is expanded by addition and subtraction operators; The update rules for addition and subtraction operations are defined as follows: Addition operation: ; Subtraction operation: in, This is the optimal solution for the t-th generation. T=50 represents the total number of iterations. The numbers are uniformly distributed random numbers; Local refinement stage: Switch to the sparrow search algorithm to fine-tune the parameters; The formula for updating the discoverer's location is as follows: , It is a constant; The follower update formula is: , , where is a random variable that follows a normal distribution; When an individual's fitness is less than 0.8 times the population's average fitness, a warning mechanism is triggered: , , is a random perturbation factor; The objective function is optimized to minimize the weighted envelope, and its formula is as follows: ; Among them, envelope extraction: , Let the probability distribution of the modal envelope be the energy summary. , Let i be the energy of the i-th mode. Total energy; Therefore, the optimal parameters are obtained through the above hybrid optimization algorithm.

5. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 4, characterized in that, The specific process for screening effective modal components is as follows: Variational mode decomposition was performed on the pure vibration signal using optimal parameters to obtain five intrinsic mode functions (IMF1 to IMF5), whose calculation formulas are as follows: , where u k (t) is the kth mode obtained from the decomposition, ω k It is the center frequency, and δ(t) is the Dirac function; Modal analysis results show that IMF1 to IMF2 are the dominant high-frequency noise components, and IMF3 to IMF5 are the dominant fault characteristic components. Effective modal components were screened based on dual indicators, and the selected results were confirmed as effective modal components.

6. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, The specific process of constructing an effective modal feature set is as follows: Extract temporal feature parameters from effective modal components. Temporal feature parameters may include kurtosis, peak factor, and impulse factor. Perform a Fast Fourier Transform on the effective modal components to convert the time-domain signal into a frequency-domain signal, and extract its frequency-domain feature parameters. Frequency domain characteristic parameters may include characteristic octave amplitude, harmonic energy ratio, and frequency band energy proportion; The extracted time-domain and frequency-domain feature parameters are standardized and normalized, and the standardized features are then fused to construct an effective modal feature set.

7. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, The structural design for constructing a two-layer bidirectional long short-term memory network model includes: Input layer: Vectorizes multimodal features, transforming them into a sequence of input features that the network can recognize; Bidirectional LSTM layer: It adopts a two-layer stacked structure, with several hidden units in each layer and the activation function is tanh. The forward LSTM is used to capture the future temporal correlation of features, and the backward LSTM is used to capture the historical temporal correlation of features. The forward and backward LSTMs are concatenated and the output dimension is G. Dropout layer: A dropout mechanism is added after the bidirectional LSTM layer, with the dropout probability set to F; Output layer: Through a fully connected layer and a Softmax activation function, the probability distribution vectors of each typical fault category are output; Weight initialization: The weight parameters of the LSTM layer and the fully connected layer are initialized using the He normal distribution.

8. The online fault diagnosis method for steam turbine generator rotor based on multimodal vibration characteristics according to claim 1, characterized in that, The specific process of constructing a rotor fault diagnosis model is as follows: Effective modal feature components are extracted from the effective modal feature set as training samples, cross-entropy is used as the loss function, and the Adam optimizer is used to adaptively and iteratively update the model parameters. A validation set is introduced during the training process. The learning rate and network parameters are dynamically adjusted based on the diagnostic accuracy feedback of the validation set until the model achieves optimal performance on the validation set. Training is then stopped, and the corresponding weight parameters are saved to form a rotor fault diagnosis model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method steps of any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method steps of any one of claims 1 to 8.

Images