A combustion engine vibration monitoring and analysis system

Abstract

This invention discloses a gas turbine vibration monitoring and analysis system, specifically relating to the field of intelligent monitoring and analysis. It includes a data synchronization acquisition module, a dynamic feature extraction module, a coupling characteristic modeling module, a time-frequency fusion enhancement module, a fault intelligent identification module, an energy flow tracing module, a state assessment and prediction module, and a three-dimensional visualization module. The data synchronization acquisition module employs a dual-layer synchronization architecture, deploying multiple monitoring point sensors to complete data preprocessing and synchronization error quantification. The time-frequency fusion enhancement module performs wavelet packet multi-scale decomposition and combines it with Hilbert-Huang transform to achieve Gaussian weighted fusion and feature enhancement. Through this dual-layer synchronization architecture, this invention achieves multi-dimensional coverage of vibration data for the entire gas turbine system, providing a highly consistent and reliable data foundation for subsequent fault diagnosis and significantly improving the accuracy of identifying complex coupled faults.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and analysis technology, and more specifically, to a gas turbine vibration monitoring and analysis system. Background Technology

[0002] As a core power source in fields such as power generation and petrochemicals, the operational stability of gas turbines directly affects the safety and efficiency of the entire industrial system. Currently, gas turbine vibration monitoring systems based on sensor technology and computer data processing technology are widely used. These systems collect vibration physical quantities by deploying various sensors at key parts of the gas turbine, and then monitor the turbine's operating status through digital signal processing and data analysis. With the development of multi-source data fusion technology, deep learning technology, and digital twin technology, gas turbine vibration monitoring systems have gradually acquired functions such as spectrum analysis, trend analysis, and simple fault diagnosis, enabling them to identify some common vibration fault types.

[0003] Existing gas turbine vibration monitoring systems have improved the reliability of gas turbine operation to a certain extent, enabling timely detection of obvious vibration anomalies and issuing early warnings. However, such systems still have many shortcomings: First, existing systems mostly use single-dimensional vibration parameter analysis methods, failing to achieve deep integration of rotor-bearing system dynamic response data and structure-foundation coupling transmission data, and lacking high-precision synchronous control of the data acquisition process, resulting in low accuracy in identifying complex vibration faults; Second, existing systems have limited time-frequency analysis capabilities for vibration signals, mostly using a single Fourier transform method, which is difficult to effectively capture transient vibration characteristics during gas turbine start-up, shutdown, and load changes, and has a short warning time for early faults such as oil film whirl and rotor thermal bending; Third, existing systems lack the ability to quantitatively analyze and track vibration energy transmission paths, only able to identify fault types, but unable to determine the root cause and propagation law of faults, making it difficult to guide targeted equipment maintenance; Fourth, the intelligent analysis models of existing systems mostly use single-structure neural networks, failing to fully utilize the complementarity of multi-source heterogeneous data, and lacking incremental learning capabilities, making it difficult to adapt to changes in gas turbine vibration characteristics under different operating conditions; Fifth, the visualization of existing systems mostly uses two-dimensional charts, which are difficult to intuitively present the vibration distribution and fault development process of the entire gas turbine system, resulting in poor human-machine interaction.

[0004] To address the shortcomings of the existing technologies, this invention proposes a gas turbine vibration monitoring and analysis system. This system is based on high-precision time synchronization technology, multi-scale time-frequency analysis technology, attention mechanism deep learning technology, and digital twin technology to construct a coupled analysis framework for the rotor-bearing system and the structure-foundation system. This enables accurate identification, early warning, and root cause localization of gas turbine vibration faults, effectively improving the safety and reliability of gas turbine operation. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a gas turbine vibration monitoring and analysis system, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a gas turbine vibration monitoring and analysis system, comprising: Data synchronization acquisition module: adopts a two-layer synchronization architecture, deploys multiple monitoring point sensors, and completes data preprocessing and synchronization error quantification; Dynamic feature extraction module: Achieves sensitive detection of faults at different stages through a three-level extraction strategy; Coupling characteristic modeling module: Employs a hybrid modeling approach to quantify the three-parameter coupling transfer characteristics; Time-frequency fusion enhancement module: Performs wavelet packet multi-scale decomposition, combined with Hilbert-Huang transform, to achieve Gaussian weighted fusion and feature enhancement; Fault Intelligent Recognition Module: Constructs a deep learning model and uses transfer learning and incremental learning to complete fault mode recognition and classification; Energy flow tracing module: Establish a five-node energy model, obtain energy transfer coefficients, solve energy flow equations, and locate the root cause of the fault; Status assessment and prediction module: The weights are determined by the analytic hierarchy process (AHP), and fuzzy comprehensive evaluation is performed to achieve LSTM trend prediction and health analysis. 3D visualization module: Modeling is performed using glTF format to complete real-time data mapping, enabling visualization of vibration status and interactive display of faults.

[0007] Preferably, the dual-layer synchronization architecture combines Precise Time Protocol (PTPv2) with GPS / BeiDou dual-mode timing. A master clock unit is deployed in the gas turbine control room, and UTC time is acquired through a GPS / BeiDou dual-mode receiver, achieving a synchronization accuracy better than 100 nanoseconds. A unified second pulse and IRIG-B time code reference are output to construct a system-wide time reference. For high-speed vibration acceleration sensors with a 50kHz sampling rate and eddy current sensors with a 20kHz sampling rate, a hardware synchronization acquisition unit is used to trigger sampling synchronously via hard-wired connections, achieving strict phase synchronization within microseconds. For low-frequency, slowly varying signals such as foundation settlement LVDT with a 1kHz sampling rate, timestamp alignment is performed using the Industrial Ethernet PTPv2 protocol. The multiple monitoring point sensors include 12 monitoring point sensors. One eddy current sensor (20kHz sampling rate) is installed in each of the X and Y vertical directions of each sliding bearing section to collect rotor radial vibration displacement. One piezoelectric acceleration sensor (50kHz sampling rate) is installed on each bearing housing to collect casing vibration acceleration. One strain gauge torque sensor (10 kHz sampling rate) is installed on the coupling to collect the torque transmitted by the shaft system. Six high-precision LVDT displacement sensors (1 kHz sampling rate) are installed at the four corners and the center of the gas turbine foundation to collect foundation settlement displacement. One accelerometer (20 kHz sampling rate) is installed 1 meter from the connection surface of the pipeline to collect pipeline vibration acceleration. Data preprocessing involves passing the collected raw data through a second-order Butterworth hardware filter circuit to remove power frequency interference and high-frequency noise. The low-pass filter cutoff frequency is set to 10 times the rated power frequency of the gas turbine. Then, software baseline correction is performed using a fifth-order polynomial fitting trend term removal method. The trend term of the fitted signal is subtracted from the original signal. Finally, data at different sampling rates are aligned by timestamps, converted to standard HDF5 format, and stored in a distributed time-series database. Low sampling rate data (1 kHz) is used for long-term trend analysis, and high sampling rate data (20 kHz and above) is used for transient feature extraction. Synchronization error quantification involves obtaining the data synchronization error index. The calculation formula is: ,in, This represents the total number of sensor nodes. For the first Local time of each sensor node The standard time of the master clock unit, and when If the time exceeds 5 microseconds, the module will automatically trigger the resynchronization process.

[0008] Preferably, the three-level extraction strategy is used to extract basic features, advanced features, and fault-sensitive features. Basic features correspond to obvious fault stages and can directly reflect the abnormal state of the system; advanced features correspond to early fault stages and can capture weak signs of faults; fault-sensitive features correspond to fault severity stages and can quantify the development level of the fault. The basic features are extracted as follows: a Hanning window is applied to the preprocessed vibration signal using a Fast Fourier Transform, with a window length of 1024 points and an overlap rate of 50%, to obtain the signal's spectral distribution. The power frequency amplitude, second harmonic amplitude, and half-harmonic amplitude of the vibration signal are then extracted. The basic characteristic parameters of frequency amplitude and total effective value are used to initially determine whether there is an anomaly in the system. The advanced characteristics include oil film whirl-oil film oscillation characteristics, rotor thermal bending characteristics, and shaft misalignment characteristics. The extraction method of oil film whirl-oil film oscillation characteristics is as follows: during the gas turbine speed-up process, vibration signals at the bearing are continuously collected in steps of 1 revolution per minute. The signal at each speed point is subjected to adaptive bandpass filtering of 0.4 to 0.5 times the power frequency. The effective value of the filtered signal is calculated to obtain the speed-half-frequency amplitude curve. After cubic spline interpolation of the curve, the first derivative is obtained to obtain the growth rate of the oil film whirl half-frequency amplitude. Simultaneously, the difference between the maximum and minimum values ​​of the phase difference of the X and Y directions (0.4 to 0.5 times the power frequency components) of the same bearing section within 10 seconds is calculated to obtain the oil film whirl phase difference fluctuation. ,in The angular velocity of the gas turbine rotor (unit: rad / s) indicates that this parameter is a function of rotor speed. The rotor thermal bending feature extraction method involves simultaneously acquiring the rotor vibration power frequency phase and the highest rotor surface temperature, accurately recording the times when the temperature peak and vibration peak occur, and calculating the mechanical angle through which the rotor rotates between the two times to obtain the rotor thermal bending hysteresis angle. The axial vibration displacement at the same moment is measured by two eddy current sensors installed on the front and rear ends of the rotor thrust disk, and the absolute difference between the two measurements is calculated to obtain the thermal bending axial vibration difference. The method for extracting shaft misalignment features is as follows: extract the amplitudes of the twice-power-frequency components in the X and Y directions of the same shaft section on both sides of the coupling, and calculate the absolute difference between the amplitudes in the two directions of the same section to obtain the misalignment twice-frequency amplitude difference. The torque transmitted by the shaft system is continuously measured using a strain gauge torque sensor at a sampling rate of 10 kHz. The difference between the maximum and minimum torque values ​​within 10 seconds is used to calculate the torque fluctuation of the coupling. The method for extracting the fault-sensitive features is as follows: after normalizing and dimensionlessly processing the high-level features to their maximum values, they are fused to obtain the oil film instability feature index. Rotor thermal bending characteristic index Shaft misalignment characteristic index Among these, the development of oil film instability faults is manifested in two aspects: firstly, the amplitude of oil film whirl increases rapidly with rotational speed (from...). The reflection is that the phase of the eddy current fluctuates drastically (due to...). (Reflection) A single parameter is insufficient to comprehensively assess the risk of oil film instability. Therefore, two dimensionless parameters are fused to obtain the oil film instability characteristic index. : ,in This represents the growth rate of the half-frequency amplitude of the oil film eddy after dimensionless scaling. This represents the dimensionless fluctuation of the oil film eddy phase difference. This serves as a reference maximum value for the half-frequency amplitude growth rate of oil film eddy. This is the reference maximum value for the oil film eddy phase difference fluctuation. Let be the angular velocity of the gas turbine rotor. The value range is from 0 to 1. The higher the value, the higher the risk of oil film instability: when When <0.3, the oil film is stable; when 0.3≤ When <0.7, there is a risk of early oil film eddy; when When the temperature is ≥0.7, oil film oscillation is about to occur, and the machine must be stopped immediately for inspection; the core characteristic of rotor thermal bending fault is that the vibration peak lags behind the temperature peak (due to...). (Reflection), while axial vibration shows significant differences (due to) (Reflection), by fusing the two dimensionless parameters, the rotor thermal bending characteristic index is obtained. : ,in The dimensionless rotor thermal bending hysteresis angle. The dimensionless thermal bending axial vibration difference. This is the reference maximum value for the rotor thermal bending hysteresis angle. This is the reference maximum value for the difference in axial vibration during thermal bending. The larger the value, the more severe the rotor thermal bending: when When 0.3 < 0.3, the rotor shows no obvious thermal bending; when 0.3 ≤ When <0.7, slight thermal bending exists; when When the value is ≥0.7, severe thermal bending occurs, requiring shutdown and cooling. The core characteristic of shaft misalignment faults is a significant difference in the amplitude of the second harmonic vibration (due to...). (Reflection), and at the same time, the torque transmitted by the coupling fluctuates drastically (due to) (Reflection), by fusing the two dimensionless parameters, the shaft misalignment characteristic index is obtained. : ,in This represents the dimensionless difference in the second harmonic amplitude of the misalignment. This represents the dimensionless torque fluctuation of the coupling. To be the reference maximum value for the misalignment of the second harmonic amplitude difference, This is the reference maximum value for the torque fluctuation of the coupling. The larger the value, the more severe the misalignment of the shaft system: when When 0.3 < 0.3, the shaft system is well aligned; when 0.3 ≤ When <0.7, there is a slight misalignment; when When the value is ≥0.7, the misalignment is severe and realignment is required.

[0009] Preferably, the hybrid modeling method combines finite element modeling with experimental modal analysis to complete geometric modeling and mesh generation, experimental modal testing, and finite element model correction. Geometric modeling and mesh generation: First, a 3D geometric model of the gas turbine casing, foundation, vibration isolators, and piping is created using SolidWorks, then imported into ANSYS for mesh generation. The casing uses Shell181 elements, the foundation uses Solid186 elements, the vibration isolators use Combin14 spring-damping elements, and the piping uses Beam188 elements. Mesh generation combines free meshing and mapped meshing, with mesh refinement applied to key components such as bearing housings and flanges. The total number of meshes is controlled to within 500,000. Experimental modal testing: A multi-point excitation single-point response method is used. Experimental modal analysis was performed using a pulsed hammer to apply excitation in three orthogonal directions to the housing bearing housing. Twenty accelerometers were evenly distributed on the housing surface as response points to collect excitation and response signals. The experimental modal parameters (natural frequencies, damping ratios, and mode shapes) of the housing and foundation were estimated using the frequency response function. The finite element model was corrected using a combination of sensitivity analysis and a genetic algorithm. First, sensitivity analysis identified the parameters with the greatest impact on the modal parameters, including the housing elastic modulus, isolator stiffness, and foundation damping. Then, with the objective function of minimizing the error between the experimental and calculated modal frequencies, the genetic algorithm optimized these parameters, ensuring that the error between the corrected model's calculated results and the experimental results was less than 3%. The three-parameter coupling transfer characteristics included the housing-rotor modal coupling coefficient. Transmissivity of the foundation vibration isolation system Pipe-cylinder connection stiffness Casing-rotor modal coupling coefficient The calculation formula is: ,in, The casing modal damping ratio, The frequency difference between the casing and the rotor, measured in Hertz. The power frequency of the gas turbine rotor is expressed in Hertz (Hz). The value ranges from 0 to 1, with a larger value indicating a higher degree of modal coupling; Transitivity of the foundation vibration isolation system: ,in, This represents the vibration acceleration amplitude at the upper end of the vibration isolator, expressed in meters per second squared. This represents the vibration acceleration amplitude at the lower end of the vibration isolator, expressed in meters per second squared. The difference in vertical vibration acceleration between the vibration isolator and the vertical vibration isolator is expressed in meters per second squared. A value less than 1 indicates effective vibration isolation; the smaller the value, the better the vibration isolation effect. >1 indicates that the vibration is amplified, and the vibration isolator is at risk of failure; Pipe-cylinder connection stiffness: ,in, The dynamic stress of the connecting bolts is expressed in Pascals. The cross-sectional area of ​​the connecting bolts is expressed in square meters. This represents the relative displacement between the pipe and cylinder connection surfaces, in meters, and the module further calculates the stiffness attenuation rate. Used to assess the looseness of connecting bolts: ,in Design stiffness for pipe-cylinder connection bolts, The value ranges from 0 to 1, with a larger value indicating a more severe loosening of the bolt. When the percentage is less than 10%, the bolted connection is in good condition; when the percentage is less than or equal to 10%, the bolted connection is in good condition. If the percentage is less than 20%, the bolts are slightly loose and need to be checked during the next shutdown. When the rate is ≥20%, there is a serious risk of bolt loosening, and the module will immediately issue a warning message to prompt maintenance personnel to handle the situation promptly.

[0010] Preferably, the wavelet packet multi-scale decomposition uses the db4 wavelet to perform three-level wavelet packet decomposition on the preprocessed vibration signal, obtaining eight sub-band signals with different frequency scales. The frequency range of each sub-band signal is determined according to the rated speed of the gas turbine. For example, when the rated speed is 3000 r / min, the power frequency is 50 Hz, and the sub-band frequency ranges are 0-6.25 Hz, 6.25-12.5 Hz, ..., 37.5-50 Hz, respectively. The Hilbert-Huang transform is performed on each sub-band signal to perform empirical mode decomposition, obtaining several intrinsic mode functions. The correlation coefficient method is used to remove pseudo intrinsic mode function components. Components with a correlation coefficient less than 0.1 are pseudo components. The Hilbert transform is performed on the retained intrinsic mode function components to obtain the instantaneous frequency and instantaneous amplitude of each component, thereby constructing the Hilbert time-frequency distribution of each sub-band signal. The Gaussian weighted fusion is performed by assigning a Gaussian weight function to each sub-band based on its center frequency and frequency bandwidth, highlighting the characteristics near the center frequency and improving the focus of the time-frequency distribution. The fused time-frequency distribution... ,in, For the first The time-frequency distribution of each subband signal, expressed in squared amplitude per hertz. For the first The center frequency of the sub-band signal, measured in Hertz (Hz). For the first The frequency bandwidth of each sub-band signal is measured in Hertz, and the fused time-frequency distribution is converted into a 224×224 grayscale image, which is then input as a time-frequency feature map to the fault intelligent identification module. The feature enhancement uses the 3σ criterion to process the fused time-frequency distribution and calculates the mean of the background noise of the time-frequency distribution. with standard deviation It will be below the threshold Pixels are set to zero, while feature regions above the threshold are retained.

[0011] Preferably, the deep learning model adopts the MC-CNN network architecture, which includes three input channels: the first channel (rotor feature channel) inputs the dynamic feature vector of the rotor-bearing system (including the oil film instability feature index, rotor thermal bending feature index, shaft misalignment feature index, and basic features, totaling a 12-dimensional vector); the second channel (structural feature channel) inputs the structure-foundation coupling transfer characteristic vector (including the housing-rotor modal coupling coefficient, foundation vibration isolation system transfer rate, pipe-cylinder connection stiffness, and stiffness attenuation rate, totaling a 4-dimensional vector); and the third channel (time-frequency feature channel) inputs the multi-scale time-frequency fusion... The combined feature map (224×224 grayscale image) is generated. Each input channel corresponds to an independent convolutional branch. The rotor feature branch and structural feature branch employ a 1D convolutional neural network, each containing two 1D convolutional layers, one max-pooling layer, and one batch normalization layer. The time-frequency feature branch employs a 2D convolutional neural network, containing three 2D convolutional layers, two max-pooling layers, and one batch normalization layer. The local feature maps extracted from these three branches are input to the channel attention module (SE module). The SE module learns the importance weights of different channel features through global average pooling and two fully connected layers, thus affecting the feature map. Weighted fusion is performed, and the fused feature map is input into two fully connected layers. Finally, a softmax classifier outputs the identification results of 12 types of faults (including oil film eddy, oil film oscillation, rotor thermal bending, shaft misalignment, casing resonance, vibration isolator failure, bolt loosening, etc.). The transfer learning is performed by pre-training the model using a publicly available gas turbine vibration fault dataset. The dataset contains 10 common faults with a total of 10,000 samples. The optimizer used is the Adam optimizer, with an initial learning rate of 0.001, which decays to 0.1 every 10 epochs. The batch size is 32. The training rounds consist of 100 rounds, and the model is then fine-tuned using historical operating data of the target gas turbine. The incremental learning is achieved using an elastic weight consolidation algorithm. When a new fault mode is identified, the module automatically adds the features of that fault mode to the fault knowledge base and retains the weights that are important to the old fault modes through the elastic weight consolidation algorithm to prevent catastrophic forgetting. The model is incrementally updated every quarter. The fault mode recognition and classification process the multi-source heterogeneous features of the input through a trained MC-CNN model and outputs the identification results of fault type and fault location.

[0012] Preferably, the five-node energy model divides the gas turbine system into five subsystems: rotor, bearings, casing, foundation, and piping. Each subsystem serves as an independent energy node. The vibration energy of each energy node is calculated from the vibration signals of all monitoring points at that node. For acceleration signals, the vibration energy is proportional to the square of the acceleration. The energy transfer coefficient is the coefficient for any two adjacent energy nodes. and The energy transfer coefficient between two nodes is calculated by measuring their vibration power. The energy transfer coefficient represents the energy transfer rate from an energy node per unit time. Passed to node Energy occupied by nodes The proportion of total energy is calculated using the following formula: ,in, For the node Passed to node The vibration power, measured in watts. For nodes The vibrational energy, measured in joules. The dimension is the reciprocal of per second; the solution to the energy flow equation involves establishing the energy flow balance equation of the gas turbine system to describe the energy transfer and dissipation relationships between energy nodes: ,in, For the first The energy dissipation rate of each energy node, expressed as the reciprocal of its value per second. The time unit is seconds. The fourth-order Runge-Kutta method is used to solve the differential equation with a time step of 0.01 seconds, yielding the variation of vibration energy at each energy node over time. To locate the root cause of the fault, the energy percentage of each node is first calculated; nodes with an energy percentage exceeding 30% are considered suspected fault nodes. Simultaneously, preliminary screening is performed using structure-foundation coupling parameters: if the casing-rotor modal coupling coefficient... If the value is >0.7, prioritize checking for casing-rotor coupling faults. If the transmissibility of the foundation vibration isolation system is... If the value is greater than 1.2, then the failure of the vibration isolator should be investigated first. If the stiffness attenuation rate is... If the failure rate is >20%, it is directly determined to be a fault caused by loose pipe connection bolts. Then, the energy transfer coefficient from the suspected fault node to other nodes is calculated, and the path with the highest transfer coefficient is the main transfer path. Finally, the root cause of the fault is determined by combining the fault type output by the multi-channel fault identification module.

[0013] Preferably, the weights are determined by the analytic hierarchy process (AHP) to establish a gas turbine system state evaluation index system containing 4 primary indicators and 12 secondary indicators. The primary indicators include rotor-bearing system health (weight 0.4), structure-foundation system health (weight 0.3), fault severity (weight 0.2), and energy flow anomaly (weight 0.1). The secondary indicators are extracted feature parameters. The AHP is used to determine the weights of each indicator, construct a judgment matrix, and calculate the weights of each indicator. The fuzzy comprehensive evaluation involves establishing an evaluation set. The trapezoidal membership function is used to determine the membership degree of each indicator to different evaluation levels, and a fuzzy evaluation matrix is ​​constructed. The comprehensive evaluation result of the system is obtained through fuzzy matrix operations; the LSTM trend prediction: an LSTM model is constructed to predict the future trend of the system's comprehensive health index. The model contains two LSTM layers, each with 64 hidden units, one Dropout layer with a dropout rate of 0.2, and one fully connected layer. A sliding window method is used to generate training samples, with a window size of 30 days and a step size of 1 day. The input is the health index of the past 30 days, and the output is the health index of the next 7 days. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 50 training epochs. When the health index is predicted to be lower than 0.6 within the next 7 days, the module automatically issues a warning and provides corresponding maintenance suggestions; the health analysis is used to obtain the system's comprehensive health index. ,in, For the first The weight of each primary indicator, For the first The evaluation value of each primary indicator, The value ranges from 0 to 1, and the closer the value is to 1, the better the system is running.

[0014] Preferably, the glTF format modeling involves establishing a high-precision three-dimensional geometric model based on the gas turbine's design drawings and actual dimensions, using the glTF 2.0 format. The data mapping involves mapping the collected real-time data and analysis results from other modules onto the three-dimensional geometric model in real-time via the MQTT 3.1.1 protocol, forming a digital twin of the gas turbine system, with a data update frequency of 1 Hz. The vibration state visualization uses color coding technology to represent the vibration intensity at different locations, with the vibration intensity changing from 0 to its maximum value corresponding to a color gradient from blue to red. Real-time values ​​and trends of the oil film instability characteristic index, rotor thermal bending characteristic index, and shaft misalignment characteristic index are displayed at key locations such as the rotor and bearings. The casing, foundation, and pipe connection locations are also displayed. The real-time values ​​of rotor modal coupling coefficient, base vibration isolation system transmissivity, and stiffness attenuation rate are displayed using dynamic arrows to represent the vibration energy transmission path. The arrow thickness is proportional to the energy transmission coefficient. Clicking on any monitoring point allows viewing the real-time vibration data, historical trend curves, and spectrum analysis results for that point. The fault interactive display: When a fault occurs, the digital twin automatically switches the view to the fault location, highlights and flashes it, and plays a fault animation to show the development process of the fault. At the same time, a fault details window pops up, showing the fault type, severity, possible causes, and maintenance suggestions. The module supports historical data playback, allowing users to select any historical time period and replay the vibration state changes of the internal combustion engine at speeds of 0.5x, 1x, 2x, and 4x.

[0015] The technical effects and advantages of this invention are as follows: 1. This invention achieves high-precision synchronous acquisition of multi-source heterogeneous data through a dual-layer synchronous architecture combined with differentiated deployment of various types of sensors at multiple key monitoring points of the gas turbine. Simultaneously, by extracting dynamic features of the rotor-bearing system in a hierarchical manner and modeling the coupling and transmission characteristics of the structure-foundation, a dual-dimensional full-system feature system of rotor and structure is constructed. This results in a standardized full-system vibration dataset with low time synchronization error, as well as multi-dimensional feature vectors covering the entire transmission link between the rotor body and the structure. This solves the problems of low accuracy in identifying complex vibration faults caused by existing systems using single-dimensional vibration parameter analysis, lack of high-precision data synchronization control, and lack of deep integration of rotor-bearing and structure-foundation data. It achieves multi-dimensional coverage of the gas turbine's full-system vibration data, providing a highly consistent and reliable data foundation for subsequent fault diagnosis and significantly improving the accuracy of identifying complex coupled faults. 2. This invention employs a multi-scale time-frequency analysis method combining db4 wavelet 3-level decomposition and Hilbert-Huang transform, along with a Gaussian kernel weighted time-frequency fusion algorithm and 3σ criterion adaptive feature enhancement processing. This method obtains transient vibration characteristics under gas turbine start-up and shutdown, and variable load non-stationary operating conditions, thereby improving the fault feature signal-to-noise ratio. It solves the problem in the background technology where existing systems mostly use a single Fourier transform, have limited time-frequency analysis capabilities, and cannot effectively capture transient vibration characteristics, resulting in short early fault warning times. This invention achieves sensitive capture of early faults such as oil film whirl and rotor thermal bending, shortens the warning time, and significantly reduces the risk of fault deterioration and unplanned shutdowns. 3. This invention constructs a vibration energy flow model of five nodes: rotor, bearing, housing, foundation, and pipeline. It combines the calculation of energy transfer coefficients with the fourth-order Runge-Kutta method to solve the energy flow balance equation. At the same time, it integrates the structure-foundation coupling parameters to form a four-step fault tracing process. This results in the complete transmission path of vibration energy from the fault source to each monitoring point and the accurate location of the fault root cause. This significantly improves the accuracy of fault location and solves the problems in the background technology where existing systems lack the ability to quantitatively analyze and track vibration energy transmission paths, can only identify fault types but cannot determine the fault root cause and propagation law, and are difficult to guide targeted equipment maintenance. This invention achieves accurate root-cause location of vibration faults, shortens fault investigation time, provides a scientific basis for targeted equipment maintenance, and significantly reduces operation and maintenance costs. 4. This invention constructs a multi-channel convolutional neural network architecture based on channel attention mechanism, and combines transfer learning pre-training with elastic weights to consolidate incremental learning, thereby obtaining a fault identification model with strong generalization ability and sustainable evolution. This solves the problem in the background technology that the intelligent analysis model has a single structure and lacks incremental learning ability, which makes it unable to adapt to changes in the vibration characteristics of the gas turbine under different operating conditions. It achieves the benefits of high fault identification accuracy and automatic identification of new fault modes. 5. This invention constructs a high-precision three-dimensional geometric model of the gas turbine using the glTF2.0 format, and combines real-time data mapping and color-coded vibration visualization using the MQTT protocol to obtain a digital twin that is synchronized with the physical entity in real time. This solves the problem in the background technology that visualization is mostly in two-dimensional charts, which makes it difficult to intuitively present the vibration distribution and fault development process of the entire system, resulting in poor human-computer interaction. It achieves the benefits of intuitive three-dimensional display of vibration status, automatic fault location highlighting, and significantly improved system usability and operation and maintenance efficiency. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0017] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] As attached Figure 1 The gas turbine vibration monitoring and analysis system shown includes a data synchronization acquisition module, a dynamic feature extraction module, a coupling characteristic modeling module, a time-frequency fusion enhancement module, a fault intelligent identification module, an energy flow tracing module, a state assessment and prediction module, and a 3D visualization module. Upon system startup, it first completes global time synchronization and sensor self-checks, then enters continuous operation. All modules form a closed-loop data flow in the order of data acquisition, preprocessing, feature extraction, coupling modeling, fault identification, fault tracing, state assessment, and visualization. The output of each module directly serves as the input of the next module. The system covers all operating conditions of the gas turbine, including start-up and shutdown, variable load, and steady-state operation. The data update frequency is divided into three levels: 1 Hz, 10 Hz, and 100 Hz, depending on the module function.

[0019] The data synchronization acquisition module adopts a two-layer synchronization architecture, deploys multiple monitoring point sensors, and completes data preprocessing and synchronization error quantification. It should be noted that the described two-layer synchronization architecture adopts the Institute of Electrical and Electronics Engineers 1588 Precision Time Protocol version 2 (IEEE 1588). The PTPv2 system, combined with GPS and BeiDou dual-mode timing, utilizes a master clock unit deployed in the gas turbine control room. It acquires Coordinated Universal Time (UTC) via a GPS / BeiDou dual-mode receiver, achieving synchronization accuracy better than 100 nanoseconds. It outputs a unified second pulse and the IRIG-B time code reference from the range instrument group, constructing a system-wide time reference. For high-speed vibration accelerometers with a 50kHz sampling rate and eddy current sensors with a 20kHz sampling rate, a hardware synchronization acquisition unit is used to trigger sampling via hard-wired synchronization, achieving strict phase synchronization within microseconds. For low-frequency, slowly varying signals such as those from linear variable differential transformers (LVDTs) with a 1kHz sampling rate, timestamp alignment is achieved via the IEEE 1588 PTPv2 industrial Ethernet protocol. The multiple monitoring point sensors include 12 sensors: one eddy current sensor (20kHz sampling rate) is installed in each of the X and Y vertical directions of each sliding bearing section to collect rotor radial vibration displacement; and one piezoelectric accelerometer (50kHz sampling rate) is installed on each bearing housing to collect casing vibration. Dynamic acceleration is measured by installing a strain gauge torque sensor (10 kHz sampling rate) on the coupling to collect the torque transmitted through the shaft system; six high-precision LVDT displacement sensors (1 kHz sampling rate) are installed at the four corners and the middle of the gas turbine foundation to collect foundation settlement displacement; and an acceleration sensor (20 kHz sampling rate) is installed 1 meter from the connection surface of the pipeline to collect pipeline vibration acceleration. Data preprocessing involves first passing the collected raw data through a second-order Butterworth hardware filter circuit to remove power frequency interference and high-frequency noise. The low-pass filter cutoff frequency is set to 10 times the rated power frequency of the gas turbine. Then, software baseline correction is performed using a fifth-order polynomial fitting trend term removal method. The trend term of the fitted signal is subtracted from the original signal. Finally, data at different sampling rates are aligned by timestamps and converted to the standard Hierarchical Data Format Version 5 (HDF5) format for storage in a distributed time-series database. Low sampling rate data (1 kHz) is used for long-term trend analysis, and high sampling rate data (20 kHz and above) is used for transient feature extraction. Synchronization error quantification involves obtaining the data synchronization error index. The calculation formula is: ,in, This represents the total number of sensor nodes. For the first Local time of each sensor node The standard time of the master clock unit, and when When the time exceeds 5 microseconds, the module automatically triggers a resynchronization process. Specifically, upon system startup, the master clock unit deployed in the gas turbine control room first acquires standard world time via a dual-mode receiver of the Global Navigation Satellite System and the BeiDou system, achieving a synchronization accuracy better than 100 nanoseconds. Subsequently, the master clock unit sends synchronization commands to all sensor nodes distributed throughout the gas turbine via industrial Ethernet. Each sensor node's built-in slave clock unit performs precise time synchronization with the master clock unit, with synchronization accuracy controlled within 1 microsecond. Sensors are deployed according to 12 pre-planned key monitoring points: one eddy current sensor is installed in each of the two perpendicular directions of each sliding bearing section; one piezoelectric accelerometer is installed on each bearing housing; one strain gauge torque sensor is installed on the coupling; six high-precision displacement sensors are installed at the four corners and the center of the gas turbine foundation; and one accelerometer is installed one meter from the connection surface of the pipeline. Different types of sensors employ different sampling frequencies to meet the acquisition requirements of different signals. The acquired raw data is first filtered by a second-order Butterworth hardware filter circuit built into the sensor to remove power frequency interference and high-frequency noise. The cutoff frequency of the low-pass filter is set to ten times the rated power frequency of the gas turbine. The filtered data is then transmitted to a preprocessing unit, which uses a fifth-order polynomial fitting method to remove trend terms from the signal and complete baseline correction. Finally, the preprocessing unit precisely aligns the data from different sampling frequencies according to timestamps, converts it into a standard hierarchical data format, and stores it in a distributed time-series database. Low-sampling-frequency data is used for long-term trend analysis, while high-sampling-frequency data is used for transient feature extraction. During system operation, the master clock unit continuously calculates the time synchronization error of all sensor nodes. When the synchronization error exceeds a set threshold, a resynchronization process is automatically triggered to ensure the time consistency of all data.

[0020] The dynamic feature extraction module achieves sensitive detection of faults at different stages through a three-level extraction strategy. It should be noted that the three-level extraction strategy is used to extract basic features, advanced features, and fault-sensitive features. Basic features correspond to obvious fault stages and can directly reflect the abnormal state of the system; advanced features correspond to early fault stages and can capture weak signs of faults; fault-sensitive features correspond to fault severity stages and can quantify the development level of the fault. The basic features are extracted as follows: a Hanning window is applied to the preprocessed vibration signal using a Fast Fourier Transform, with a window length of 1024 points and an overlap rate of 50%, to obtain the signal's spectral distribution, and the power frequency amplitude and second harmonic amplitude of the vibration signal are extracted. The basic characteristic parameters of half-frequency amplitude and total effective value are used to initially determine whether there is an anomaly in the system. The advanced characteristics include oil film whirl-oil film oscillation characteristics, rotor thermal bending characteristics, and shaft misalignment characteristics. The extraction method of oil film whirl-oil film oscillation characteristics is as follows: during the gas turbine speed-up process, vibration signals at the bearing are continuously collected in steps of 1 revolution per minute. The signal at each speed point is subjected to adaptive bandpass filtering of 0.4 to 0.5 times the power frequency. The effective value of the filtered signal is calculated to obtain the speed-half-frequency amplitude curve. After cubic spline interpolation of the curve, the first derivative is obtained to obtain the growth rate of the oil film whirl half-frequency amplitude. Simultaneously, the difference between the maximum and minimum values ​​of the phase difference of the X and Y directions (0.4 to 0.5 times the power frequency components) of the same bearing section within 10 seconds is calculated to obtain the oil film whirl phase difference fluctuation. ,in The angular velocity of the gas turbine rotor (unit: rad / s) indicates that this parameter is a function of rotor speed. The rotor thermal bending feature extraction method involves simultaneously acquiring the rotor vibration power frequency phase and the highest rotor surface temperature, accurately recording the times when the temperature peak and vibration peak occur, and calculating the mechanical angle through which the rotor rotates between the two times to obtain the rotor thermal bending hysteresis angle. The axial vibration displacement at the same moment is measured by two eddy current sensors installed on the front and rear ends of the rotor thrust disk, and the absolute difference between the two measurements is calculated to obtain the thermal bending axial vibration difference. The method for extracting shaft misalignment features is as follows: extract the amplitudes of the twice-power-frequency components in the X and Y directions of the same shaft section on both sides of the coupling, and calculate the absolute difference between the amplitudes in the two directions of the same section to obtain the misalignment twice-frequency amplitude difference. The torque transmitted by the shaft system is continuously measured using a strain gauge torque sensor at a sampling rate of 10 kHz. The difference between the maximum and minimum torque values ​​within 10 seconds is used to calculate the torque fluctuation of the coupling. The method for extracting the fault-sensitive features is as follows: after normalizing and dimensionlessly processing the high-level features to their maximum values, they are fused to obtain the oil film instability feature index. Rotor thermal bending characteristic index Shaft misalignment characteristic index Among these, the development of oil film instability faults is manifested in two aspects: firstly, the amplitude of oil film whirl increases rapidly with rotational speed (from...). The reflection is that the phase of the eddy current fluctuates drastically (due to...). (Reflection) A single parameter is insufficient to comprehensively assess the risk of oil film instability. Therefore, two dimensionless parameters are fused to obtain the oil film instability characteristic index. : ,in This represents the growth rate of the half-frequency amplitude of the oil film eddy after dimensionless scaling. This represents the dimensionless fluctuation of the oil film eddy phase difference. This serves as a reference maximum value for the half-frequency amplitude growth rate of oil film eddy. This is the reference maximum value for the oil film eddy phase difference fluctuation. Let be the angular velocity of the gas turbine rotor. The value range is from 0 to 1. The higher the value, the higher the risk of oil film instability: when When <0.3, the oil film is stable; when 0.3≤ When <0.7, there is a risk of early oil film eddy; when When the temperature is ≥0.7, oil film oscillation is about to occur, and the machine must be stopped immediately for inspection; the core characteristic of rotor thermal bending fault is that the vibration peak lags behind the temperature peak (due to...). (Reflection), while axial vibration shows significant differences (due to) (Reflection), by fusing the two dimensionless parameters, the rotor thermal bending characteristic index is obtained. : ,in The dimensionless rotor thermal bending hysteresis angle. The dimensionless thermal bending axial vibration difference. This is the reference maximum value for the rotor thermal bending hysteresis angle. This is the reference maximum value for the difference in axial vibration during thermal bending. The larger the value, the more severe the rotor thermal bending: when When 0.3 < 0.3, the rotor shows no obvious thermal bending; when 0.3 ≤ When <0.7, slight thermal bending exists; when When the value is ≥0.7, severe thermal bending occurs, requiring shutdown and cooling. The core characteristic of shaft misalignment faults is a significant difference in the amplitude of the second harmonic vibration (due to...). (Reflection), and at the same time, the torque transmitted by the coupling fluctuates drastically (due to) (Reflection), by fusing the two dimensionless parameters, the shaft misalignment characteristic index is obtained. : ,in This represents the dimensionless difference in the second harmonic amplitude of the misalignment. This represents the dimensionless torque fluctuation of the coupling. To be the reference maximum value for the misalignment of the second harmonic amplitude difference, This is the reference maximum value for the torque fluctuation of the coupling. The larger the value, the more severe the misalignment of the shaft system: when When 0.3 < 0.3, the shaft system is well aligned; when 0.3 ≤ When <0.7, there is a slight misalignment; when When the value is ≥0.7, the misalignment is severe and realignment is required. Specifically, after preprocessing, the rotor feature extraction unit immediately performs hierarchical feature extraction on the data. This process continues to run under all operating conditions of the gas turbine. First, basic feature extraction is performed. The feature extraction unit performs a Fast Fourier Transform with a Hanning window on the preprocessed vibration signal, with the window length set to 1024 points and the overlap rate set to 50%, to obtain the signal's spectral distribution. Then, basic feature parameters such as the power frequency amplitude, second harmonic amplitude, half-frequency amplitude, and total effective value of the vibration signal are extracted to preliminarily determine whether there are any abnormalities in the system. Next, advanced feature extraction is performed, extracting corresponding features for three typical faults. For oil film whirl and oil film oscillation faults, during the gas turbine's acceleration process, the feature extraction unit continuously collects vibration signals at the bearing in steps of one revolution per minute. The signal at each speed point is subjected to adaptive bandpass filtering at 0.4 to 0.5 times the power frequency, and the effective value of the filtered signal is calculated to obtain the relationship curve between the speed and the half-frequency amplitude. After performing cubic spline interpolation on the curve, the first derivative is calculated to obtain the growth rate of the half-frequency amplitude of the oil film whirl. Simultaneously, the maximum and minimum values ​​of the phase difference between the two perpendicular directions of the same bearing cross-section at 0.4 to 0.5 times the power frequency components within ten seconds are calculated to obtain the oil film whirl phase difference fluctuation. For rotor thermal bending faults, the feature extraction unit synchronously acquires the rotor vibration power frequency phase and the highest rotor surface temperature, accurately records the times when the temperature peak and vibration peak occur, and calculates the mechanical angle rotated by the rotor between the two times to obtain the rotor thermal bending hysteresis angle. Simultaneously, two eddy current sensors installed on the front and rear ends of the rotor thrust plate measure the axial vibration displacement at the same time, and the absolute difference between the two measurements is calculated to obtain the thermal bending axial vibration difference. For shaft misalignment faults, the feature extraction unit extracts the amplitudes of the two perpendicular directions of the same shaft cross-section on both sides of the coupling, which are twice the power frequency components. It calculates the absolute difference between the amplitudes in the two directions of the same cross-section to obtain the misalignment twice-frequency amplitude difference. Simultaneously, a strain gauge torque sensor continuously measures the torque transmitted by the shaft at a sampling rate of 10 kHz, calculating the difference between the maximum and minimum torque values ​​within ten seconds to obtain the coupling torque fluctuation. Finally, fault-sensitive features are extracted. Since different advanced features have different physical dimensions and cannot be directly fused, the feature extraction unit first performs maximum value normalization and dimensionless processing on all advanced features. Each original parameter is divided by its corresponding fault threshold value provided by the gas turbine manufacturer to obtain dimensionless parameters. Subsequently, the dimensionless parameters are fused and calculated to obtain the oil film instability characteristic index, rotor thermal bending characteristic index, and shaft misalignment characteristic index. Each index ranges from zero to one; a larger value indicates a higher severity of the corresponding fault. After extraction, the feature extraction unit combines the three fault-sensitive feature indices with the basic features and advanced features to form a dynamic feature vector of the rotor-bearing system, which is then input into the subsequent multi-channel fault identification module and system status evaluation module.Among them, as shown in Table 1, the performance comparison of the fault sensitivity feature index of the rotor-bearing system is as follows: the traditional single parameter can only reflect a single feature of the fault (such as oil film instability only looking at the half-frequency amplitude), and the false negative rate for early faults exceeds 30%; the fusion feature index of the present invention integrates multi-dimensional information such as the amplitude, phase, and dynamic changes of the fault, and improves the average identification accuracy of the three types of core rotor faults by 24.6%.

[0021] Table 1 Feature Index Name Actual measured value during normal operation Early fault measured values Measured values ​​of severe faults Traditional single-parameter recognition accuracy The accuracy of the fusion index recognition in this invention Oil film instability characteristic index 0.12 0.45 0.78 72.3% 98.2% Rotor thermal bending characteristic index 0.08 0.32 0.67 65.8% 96.7% Shaft misalignment characteristic index 0.06 0.28 0.72 79.5% 97.5% The coupling characteristic modeling module employs a hybrid modeling method to quantify the three-parameter coupling transmission characteristics. It should be noted that the hybrid modeling method combines finite element modeling with experimental modal analysis to complete geometric modeling and mesh generation, experimental modal testing, and finite element model correction. The geometric modeling and mesh generation are as follows: First, Dassault Systèmes' 3D computer-aided design software (SolidWorks) is used to create 3D geometric models of the gas turbine casing, foundation, vibration isolators, and piping. These models are then imported into ANSYS finite element analysis software for mesh generation. The casing uses four-node shell elements (Shell181) from ANSYS, the foundation uses twenty-node hexahedral solid elements (Solid186) from ANSYS, the vibration isolators use spring-damped elements (Combin14) from ANSYS, and the piping uses two-node three-dimensional beam elements (Beam188) from ANSYS. The mesh generation combines free meshing and mapped meshing methods for bearing housings, etc. Mesh refinement is applied to key components such as flanges, with the total number of meshes controlled below 500,000. The experimental modal testing employs a multi-point excitation, single-point response method for modal analysis. A pulsed hammer is used to apply excitation in three orthogonal directions to the housing bearing housing. Twenty accelerometers are evenly distributed on the housing surface as response points to collect excitation and response signals. The experimental modal parameters (natural frequency, damping ratio, and mode shape) of the housing and foundation are estimated using the frequency response function. The finite element model correction uses a combination of sensitivity analysis and genetic algorithms. First, sensitivity analysis identifies the parameters with the greatest impact on modal parameters, including the housing elastic modulus, vibration isolator stiffness, and foundation damping. Then, with the objective function of minimizing the error between the experimental and calculated modal frequencies, the genetic algorithm optimizes these parameters, ensuring that the error between the corrected model's calculated results and experimental results is less than 3%. The three-parameter coupling transfer characteristics include the housing-rotor modal coupling coefficient. Transmissivity of the foundation vibration isolation system Pipe-cylinder connection stiffness Casing-rotor modal coupling coefficient The calculation formula is: ,in, The casing modal damping ratio, The frequency difference between the casing and the rotor, measured in Hertz. The power frequency of the gas turbine rotor is expressed in Hertz (Hz). The value ranges from 0 to 1, with a larger value indicating a higher degree of modal coupling; Transitivity of the foundation vibration isolation system: ,in, This represents the vibration acceleration amplitude at the upper end of the vibration isolator, expressed in meters per second squared. This represents the vibration acceleration amplitude at the lower end of the vibration isolator, expressed in meters per second squared. The difference in vertical vibration acceleration between the vibration isolator and the vertical vibration isolator is expressed in meters per second squared. A value less than 1 indicates effective vibration isolation; the smaller the value, the better the vibration isolation effect. >1 indicates that the vibration is amplified, and the vibration isolator is at risk of failure; Pipe-cylinder connection stiffness: ,in, The dynamic stress of the connecting bolts is expressed in Pascals. The cross-sectional area of ​​the connecting bolts is expressed in square meters. This represents the relative displacement between the pipe and cylinder connection surfaces, in meters, and the module further calculates the stiffness attenuation rate. Used to assess the looseness of connecting bolts: ,in Design stiffness for pipe-cylinder connection bolts, The value ranges from 0 to 1, with a larger value indicating a more severe loosening of the bolt. When the percentage is less than 10%, the bolted connection is in good condition; when the percentage is less than or equal to 10%, the bolted connection is in good condition. If the percentage is less than 20%, the bolts are slightly loose and need to be checked during the next shutdown. When the bolt loosening rate is ≥20%, there is a serious risk of bolt loosening, and the module immediately issues an early warning to prompt maintenance personnel to handle the situation promptly. Specifically, this module consists of two parts: offline modeling and real-time quantization. Offline modeling is performed during the initial system deployment and after each major equipment overhaul, jointly completed by the finite element modeling unit and the experimental modal testing unit. First, the modeling unit establishes a three-dimensional geometric model including the casing, foundation, vibration isolators, and piping based on the gas turbine's design drawings and actual dimensions, and imports it into finite element analysis software for mesh generation. Shell elements are used for the casing, solid elements for the foundation, spring-damped elements for the vibration isolators, and beam elements for the piping. Mesh refinement is applied to key components such as bearing housings and flanges, with the total number of meshes controlled within 500,000. Subsequently, the experimental modal testing unit uses a multi-point excitation single-point response method for experimental modal analysis. A pulsed hammer is used to apply excitation in three orthogonal directions to the casing and bearing housing, and twenty accelerometers are evenly distributed on the casing surface as response points. Excitation and response signals are collected, and the experimental modal parameters of the casing and foundation, including natural frequencies, damping ratios, and mode shapes, are estimated through the frequency response function. Finally, the modeling unit uses a combination of sensitivity analysis and genetic algorithms to correct the finite element model. First, sensitivity analysis identifies the parameters with the greatest impact on modal parameters. Then, using the minimization of the error between the experimental and calculated modal frequencies as the objective function, the genetic algorithm optimizes these parameters, ensuring that the error between the corrected model's calculated results and experimental results is less than 3%. The real-time quantization process is continuously executed during gas turbine operation and is completed by the coupling parameter calculation unit. Based on the corrected finite element model, the calculation unit uses real-time acquired vibration data to calculate the shell-rotor modal coupling coefficient, the transmissivity of the foundation vibration isolation system, and the pipe-cylinder connection stiffness. The pipe-cylinder connection stiffness is calculated using the dynamic stress of the connecting bolts, the bolt cross-sectional area, and the relative displacement of the pipe-cylinder connection surface. The calculation unit further calculates the pipe-cylinder connection stiffness attenuation rate, i.e., the percentage decrease in actual bolt stiffness relative to the design stiffness. The design stiffness is calculated by the equipment manufacturer based on the bolt material grade, nominal diameter, effective length, and installation preload, and is provided at the time of equipment delivery. The stiffness attenuation rate ranges from zero to one; a higher value indicates more severe bolt loosening. When the attenuation rate exceeds 20%, the calculation unit immediately issues a bolt loosening warning. After calculation, the coupling parameter calculation unit combines the four parameters into a structure-foundation coupling characteristic vector, which is then input into the subsequent multi-channel fault identification module, vibration energy flow tracking module, and system state assessment module.Table 2 shows a comparison of the performance of the quantitative parameters for structure-foundation coupling characteristics. The traditional vibration amplitude method cannot distinguish whether the vibration originates from the rotor body or is transmitted through structural coupling, and the accuracy of locating structural faults is generally below 70%. The coupling parameters of this invention directly quantify the transmission characteristics of vibration between different components, enabling precise location of the root cause of the fault. The average location accuracy of structural faults is improved by 33.6%, with the location accuracy of loose pipe bolts approaching 100%.

[0022] Table 2 Coupling parameter name Actual measured value during normal operation Warning threshold critical value Fault failure measured value Traditional vibration amplitude method positioning accuracy The accuracy of the coupling parameter positioning in this invention Housing-rotor modal coupling coefficient 0.21 0.70 0.89 58.2% 95.3% Transmissivity of foundation vibration isolation system 0.32 1.20 1.35 61.7% 94.8% Pipe-cylinder connection stiffness attenuation rate 3% 20% 26% 68.5% 99.1% The time-frequency fusion enhancement module performs wavelet packet multi-scale decomposition and combines it with Hilbert-Huang transform to achieve Gaussian weighted fusion and feature enhancement. It should be noted that the wavelet packet multi-scale decomposition uses the multi-Beich wavelet (db4) wavelet to perform three-level wavelet packet decomposition on the preprocessed vibration signal, obtaining eight sub-band signals with different frequency scales. The frequency range of each sub-band signal is determined according to the rated speed of the gas turbine. For example, when the rated speed is 3000 r / min, the power frequency is 50 Hz, and the sub-band frequency ranges are 0-6.25 Hz, 6.25-12.5 Hz, ..., 37.5-50 Hz, respectively. The Hilbert-Huang transform is performed on each sub-band signal using empirical mode. The intrinsic mode functions (IMFs) are decomposed to obtain several intrinsic mode functions (IMFs). Pseudo-IMF components are eliminated using the correlation coefficient method; components with a correlation coefficient less than 0.1 are considered pseudo-components. A Hilbert transform is then performed on the remaining IMFs to obtain the instantaneous frequency and amplitude of each component, thereby constructing the Hilbert time-frequency distribution for each sub-band signal. The Gaussian weighted fusion process assigns a Gaussian weight function to each sub-band based on its center frequency and bandwidth, highlighting features near the center frequency and improving the focus of the time-frequency distribution. The fused time-frequency distribution... ,in, For the first The time-frequency distribution of each subband signal, expressed in squared amplitude per hertz. For the first The center frequency of the sub-band signal, measured in Hertz (Hz). For the first The frequency bandwidth of each sub-band signal is measured in Hertz, and the fused time-frequency distribution is converted into a 224×224 grayscale image, which is then input as a time-frequency feature map to the fault intelligent identification module. The feature enhancement uses the 3σ criterion to process the fused time-frequency distribution and calculates the mean of the background noise of the time-frequency distribution. with standard deviation It will be below the threshold Pixels are set to zero, while feature regions above the threshold are retained. Specifically, the time-frequency analysis unit starts immediately after data preprocessing, focusing on analyzing non-stationary operating conditions such as gas turbine start-up and shutdown and load changes. It also continues to run under steady-state conditions to capture sudden transient vibration characteristics. First, the time-frequency analysis unit uses db4 wavelets to perform three-level wavelet packet decomposition on the preprocessed vibration signal, obtaining eight sub-band signals at different frequency scales. The frequency range of each sub-band signal is determined according to the rated speed of the gas turbine. Then, empirical mode decomposition is performed on each sub-band signal to obtain several intrinsic mode functions. The correlation coefficient method is used to remove pseudo-intrinsic mode function components, i.e., components with a correlation coefficient less than 0.1. Hilbert transform is performed on the retained components to obtain the instantaneous frequency and instantaneous amplitude of each component, thereby constructing the Hilbert time-frequency distribution of each sub-band signal. To address the resolution differences in time-frequency distributions across different sub-bands, the time-frequency analysis unit employs a Gaussian kernel-based time-frequency fusion algorithm. Based on the center frequency and bandwidth of each sub-band, a Gaussian weighting function is assigned to highlight features near the center frequency, improving the focus of the time-frequency distribution. The fused time-frequency distribution is obtained through weighted summation. Finally, the time-frequency analysis unit uses a three-standard-deviation criterion to enhance the features of the fused time-frequency distribution. The mean and standard deviation of the background noise in the time-frequency distribution are calculated, and pixels below a threshold are set to zero, while feature regions above the threshold are retained, effectively removing background noise and interference components. After processing, the time-frequency analysis unit converts the enhanced time-frequency distribution into a 224x224 grayscale image, which is then input as the time-frequency feature map to the subsequent multi-channel fault identification module.

[0023] The fault intelligent identification module: constructs a deep learning model and uses transfer learning and incremental learning to complete fault mode identification and classification; It should be noted that the deep learning model adopts a multi-channel convolutional neural network (MC-CNN) architecture, which includes three input channels: the first channel (rotor feature channel) inputs the dynamic feature vector of the rotor-bearing system (including the oil film instability feature index, rotor thermal bending feature index, shaft misalignment feature index, and basic features, totaling a 12-dimensional vector); the second channel (structural feature channel) inputs the structure-foundation coupling transfer characteristic vector (including the housing-rotor modal coupling coefficient, foundation vibration isolation system transfer rate, pipe-cylinder connection stiffness, and stiffness attenuation rate, totaling a 4-dimensional vector); and the third channel (time-frequency feature channel) inputs the multi-scale time-frequency fusion. Feature maps (224×224 grayscale images); each input channel corresponds to an independent convolutional branch. The rotor feature branch and structural feature branch use a one-dimensional (1D) convolutional neural network, each containing two 1D convolutional layers, one max pooling layer, and one batch normalization layer; the time-frequency feature branch uses a two-dimensional (2D) convolutional neural network, containing three 2D convolutional layers, two max pooling layers, and one batch normalization layer. The local feature maps extracted from the three branches are input to the channel attention module, i.e., the squeeze-excitation module (SE module). The SE module learns the importance weights of different channel features through global average pooling and two fully connected layers, and focuses on the importance of features in each channel. The feature maps are weighted and fused, and the fused feature maps are input into two fully connected layers. Finally, a softmax classifier is used to output the identification results of 12 types of faults (including oil film eddy, oil film oscillation, rotor thermal bending, shaft misalignment, casing resonance, vibration isolator failure, bolt loosening, etc.). The transfer learning is performed by pre-training the model using a publicly available gas turbine vibration fault dataset. The dataset contains 10 common faults and a total of 10,000 samples. The optimizer uses an adaptive moment estimation optimizer (Adam optimizer) with an initial learning rate of 0.001, which decays every 10 training epochs. The original model was 0.1, with a batch size of 32 and 100 training rounds. Then, the model was fine-tuned using historical operating data of the target gas turbine. The incremental learning employed an elastic weight consolidation algorithm. When a new fault mode was identified, the module automatically added the features of that fault mode to the fault knowledge base and retained the weights important to the old fault modes through the elastic weight consolidation algorithm to prevent catastrophic forgetting. The model underwent incremental updates quarterly. Fault mode recognition and classification involved processing the multi-source heterogeneous features of the input using a trained MC-CNN model, outputting the identification results of the fault type and fault location. Specifically, the deep learning inference unit continuously performed fault recognition after receiving the features output from the preceding three modules. The inference unit constructed a multi-channel convolutional neural network model based on a channel attention mechanism, containing three independent input channels that respectively received the rotor-bearing system dynamic feature vector, the structure-foundation coupling characteristic vector, and the multi-scale time-frequency fusion feature map.Each input channel corresponds to an independent convolutional branch. The rotor feature branch and structural feature branch employ a one-dimensional convolutional neural network, each containing two one-dimensional convolutional layers, one max-pooling layer, and one batch normalization layer. The time-frequency feature branch employs a two-dimensional convolutional neural network, containing three two-dimensional convolutional layers, two max-pooling layers, and one batch normalization layer. The local feature maps extracted from the three branches are input into the channel attention module. The attention module learns the importance weights of different channel features through global average pooling and two fully connected layers, performing weighted fusion of the feature maps to highlight features that contribute significantly to fault identification and suppress interference from irrelevant features. The fused feature maps are input into two fully connected layers, and finally, a softmax classifier outputs the identification results for twelve types of faults, including fault type, fault location, and fault confidence. Model training is completed during the initial system deployment. First, the model is pre-trained using a publicly available gas turbine vibration fault dataset. The optimizer used is the Adam optimizer, with an initial learning rate of 0.001, decaying to 0.1 every ten training epochs. The batch size is 32, and the training epochs are 100. Then, the model is fine-tuned using historical operating data of the target gas turbine to improve its recognition accuracy on specific gas turbines. During system operation, the model undergoes incremental updates quarterly, employing an elastic weight consolidation algorithm to retain weights important to older fault modes and prevent catastrophic amnesia. When a new fault mode is identified, the inference unit automatically adds its features to the fault knowledge base and updates the model parameters through incremental learning. After identification, the inference unit outputs the fault type, fault location, and fault confidence level to the subsequent vibration energy flow tracing module, system state assessment module, and visualization interaction module.

[0024] Furthermore, the multi-channel convolutional neural network fault recognition model based on the channel attention mechanism is described in detail below: Overall Model Architecture and Hierarchical Connections: This model is specifically designed for fault identification of multi-source heterogeneous vibration data from gas turbines. It adopts a three-input, single-output branch fusion architecture, consisting of three independent feature extraction branches, a channel attention fusion module, and a classification output module connected in series. The three feature extraction branches work in parallel, processing different types of input data respectively. The extracted local features are uniformly input into the channel attention fusion module for weighted fusion, and the fused global features are input into the classification output module to obtain the final fault identification result. First Branch (Rotor Dynamic Feature Extraction Branch): This branch uses a one-dimensional convolutional neural network structure, consisting of two one-dimensional convolutional layers, a max-pooling layer, and a batch normalization layer connected in series. The first convolutional layer is responsible for extracting local one-dimensional features from the input data, the second convolutional layer is responsible for secondary abstraction of the extracted features, the max-pooling layer is used to reduce the feature dimensionality and retain key features, and the batch normalization layer is used to accelerate model convergence and prevent overfitting. Second Branch (Structural Coupling Feature Extraction Branch): This branch also uses a one-dimensional convolutional neural network structure, with the same hierarchical composition as the first branch, consisting of two one-dimensional convolutional layers, a max-pooling layer, and a batch normalization layer connected in series. This branch optimizes the convolutional kernel size to better suit the characteristics of structural coupling parameters, enabling more effective extraction of anomalous features from the structural system. The third branch (time-frequency feature extraction branch) employs a two-dimensional convolutional neural network structure, consisting of three two-dimensional convolutional layers, two max-pooling layers, and a batch normalization layer connected sequentially. The first two convolutional layers extract local texture features from the time-frequency image, the third convolutional layer extracts global semantic features, the two max-pooling layers progressively reduce the feature map size, and the batch normalization layer stabilizes the training process. The channel attention fusion module consists of a global average pooling layer, two fully connected layers, and a multiplication unit. The feature maps output from the three branches are first fed into the global average pooling layer, compressing each feature map into a one-dimensional feature vector. Then, the first fully connected layer performs dimensionality reduction on the feature vector, and the second fully connected layer performs dimensionality increase on the reduced feature vector and outputs the weight coefficients for each feature channel. Finally, the weight coefficients are multiplied channel-by-channel with the original feature map to complete the weighted fusion of features. The classification output module consists of two fully connected layers and a softmax classification layer connected sequentially. The first fully connected layer is responsible for mapping the fused feature vectors to a high-dimensional feature space, the second fully connected layer is responsible for mapping the high-dimensional feature vectors to a fault category space, and the softmax classification layer is responsible for converting the output into a probability distribution for each fault category.

[0025] Model training steps: Model training consists of three stages: pre-training, fine-tuning, and incremental learning. The entire training process is completed on industrial edge computing devices, without relying on cloud computing resources. Stage 1: Pre-training with a public dataset Before the system's initial deployment, the model is pre-trained using a publicly available general dataset of gas turbine vibration faults. This dataset contains ten common gas turbine vibration fault samples, with one thousand labeled samples for each fault class. During training, the dataset is randomly divided into training and validation sets at a ratio of seven to three. The optimizer uses an adaptive moment estimation algorithm, with an initial learning rate set to 0.1%, which decays to one-tenth of its original value every ten training cycles. The batch size is set to thirty-two, and the total number of training rounds is one hundred. In each training round, the model's recognition accuracy is calculated on the validation set. When the validation set accuracy no longer improves after ten consecutive rounds, training is terminated early, and the optimal model parameters are saved. Stage 2: Fine-tuning with target gas turbine data After pre-training, the model is fine-tuned using historical operating data from the target gas turbine. The historical data for the target gas turbine includes the equipment's normal operation data and historical fault data from the past year, all of which are manually labeled. During fine-tuning, the parameters of the first two convolutional layers of the model are frozen, and only the parameters of subsequent convolutional layers, fully connected layers, and attention modules are trained. The optimizer still uses the adaptive moment estimation algorithm, with an initial learning rate set to 0.01%, a batch size of 16, and a total of 50 training rounds. After fine-tuning, the model can adapt to the specific vibration characteristics of the target gas turbine, and the recognition accuracy is significantly improved. The third stage: After the quarterly incremental learning update system is put into operation, the model is incrementally updated every quarter. When the model identifies an unknown fault mode with a confidence level below a set threshold, the fault sample is automatically stored in the fault knowledge base. At the end of each quarter, maintenance personnel manually label the unknown samples in the fault knowledge base, and then use the elastic weight consolidation algorithm to incrementally train the model. During training, the model retains the weight parameters that are important for old fault modes, while learning the features of new fault modes to prevent catastrophic forgetting. The batch size for incremental training is set to eight, and the total number of training rounds is twenty.

[0026] Model Integration with Gas Turbine Vibration Monitoring Scenarios: This model is specifically designed for the multi-source data characteristics and fault diagnosis needs of gas turbine vibration monitoring. The input and output data have a clear intrinsic correlation with the gas turbine's operating state. Input Data Settings: The model's three input channels correspond to three core types of data for gas turbine vibration monitoring. All input data comes from the output of preceding modules. The first channel receives the rotor-bearing system dynamic feature vector, which contains basic, advanced, and fault-sensitive features reflecting the rotor system's operating state. The second channel receives the structure-foundation coupling characteristic vector, which contains key parameters reflecting the coupled vibration characteristics of the structural system. The third channel receives a multi-scale time-frequency fusion feature map, which contains time-frequency information reflecting the transient vibration characteristics of the gas turbine. These three types of data describe the gas turbine's vibration state from different dimensions, complementing each other and comprehensively characterizing the fault features. Output Data Settings: The model outputs the identification results for twelve types of gas turbine vibration faults, including fault type, fault location, and fault confidence. The fault types cover twelve of the most common vibration faults during gas turbine operation, including oil film whirl, oil film oscillation, rotor thermal bending, shaft misalignment, casing resonance, vibration isolator failure, and loose pipe connection bolts. Fault locations are pinpointed to specific components of the gas turbine, such as bearing number one, coupling number two, and pipe number three. Fault confidence represents the model's certainty regarding the identification result, ranging from zero to one. Output application: The fault identification results output by the model are directly input into the subsequent vibration energy flow transmission path tracing and source tracing module and the system state dynamic assessment and trend prediction module. In the fault source tracing module, the fault identification results are cross-validated with the energy flow analysis results to improve the accuracy of fault location. In the state assessment module, fault type and fault confidence, as core components of the fault severity evaluation index, are used to calculate the system's comprehensive health index. Simultaneously, the fault identification results are also input into a digital twin-driven 3D visualization interactive module, providing an intuitive display for maintenance personnel.

[0027] The energy flow tracing module: establishes a five-node energy model, obtains the energy transfer coefficient, solves the energy flow equation, and locates the root cause of the fault. It should be noted that the five-node energy model divides the gas turbine system into five subsystems: rotor, bearings, casing, foundation, and piping. Each subsystem is considered an independent energy node. The vibration energy of each energy node is calculated from the vibration signals of all monitoring points at that node. For acceleration signals, the vibration energy is proportional to the square of the acceleration. The energy transfer coefficient is calculated for any two adjacent energy nodes. and The energy transfer coefficient between two nodes is calculated by measuring their vibration power. The energy transfer coefficient represents the energy transfer rate from an energy node per unit time. Passed to node Energy occupied by nodes The proportion of total energy is calculated using the following formula: ,in, For the node Passed to node The vibration power, measured in watts. For nodes The vibrational energy, measured in joules. The dimension is the reciprocal of per second; the solution to the energy flow equation involves establishing the energy flow balance equation of the gas turbine system to describe the energy transfer and dissipation relationships between energy nodes: ,in, For the first The energy dissipation rate of each energy node, expressed as the reciprocal of its value per second. The time unit is seconds. The fourth-order Runge-Kutta method is used to solve the differential equation with a time step of 0.01 seconds, yielding the variation of vibration energy at each energy node over time. To locate the root cause of the fault, the energy percentage of each node is first calculated; nodes with an energy percentage exceeding 30% are considered suspected fault nodes. Simultaneously, preliminary screening is performed using structure-foundation coupling parameters: if the casing-rotor modal coupling coefficient... If the value is >0.7, prioritize checking for casing-rotor coupling faults. If the transmissibility of the foundation vibration isolation system is... If the value is greater than 1.2, then the failure of the vibration isolator should be investigated first. If the stiffness attenuation rate is... If the energy transfer rate is greater than 20%, it is directly identified as a loose pipe connection bolt fault. Then, the energy transfer coefficient from the suspected fault node to other nodes is calculated, and the path with the highest transfer coefficient is the main transfer path. Finally, the root cause of the fault is determined by combining the fault type output by the multi-channel fault identification module. Specifically, the energy flow analysis unit starts immediately after the fault identification module outputs a suspected fault to locate the root cause. First, the energy flow analysis unit divides the gas turbine system into five subsystems: rotor, bearings, casing, foundation, and piping, with each subsystem as an independent energy node. The vibration energy of each energy node is calculated from the vibration signals of all monitoring points of that node. For acceleration signals, the vibration energy is proportional to the square of the acceleration. Then, the energy flow analysis unit calculates the energy transfer coefficient between any two adjacent energy nodes, that is, the proportion of energy transferred from one node to another per unit time relative to the total energy of the previous node. Based on this, the energy flow balance equation of the gas turbine system is established to describe the energy transfer and dissipation relationship between each energy node. The fourth-order Runge-Kutta method is used to solve this differential equation with a time step of 0.01 seconds to obtain the variation law of vibration energy of each energy node over time. To improve the accuracy of fault location, the energy flow analysis unit employs a four-step source tracing process based on structure-foundation coupling parameters: First, the energy percentage of each node is calculated; nodes with an energy percentage exceeding 30% are considered suspicious fault nodes. Second, preliminary screening is performed using coupling parameters: if the casing-rotor modal coupling coefficient exceeds 0.7, casing-rotor coupling faults are prioritized; if the base vibration isolation system transmission rate exceeds 1.2, isolator failure is prioritized; if the pipe-cylinder connection stiffness attenuation rate exceeds 20%, pipe connection bolt loosening is directly identified. Third, the energy transfer coefficient from the suspicious fault node to other nodes is calculated; the path with the highest transmission coefficient is the main transmission path. Fourth, the root cause of the fault is finally determined by combining the fault type output by the multi-channel fault identification module. After source tracing is completed, the energy flow analysis unit outputs the fault root cause location, main energy transfer path, and energy percentage of each node to the subsequent system status assessment module and visualization interaction module.

[0028] The status assessment and prediction module: determines the weights through the analytic hierarchy process, performs fuzzy comprehensive evaluation, and realizes LSTM trend prediction and health analysis; It should be noted that the weighting of the analytic hierarchy process (AHP) is determined by establishing a gas turbine system state evaluation index system containing 4 primary indicators and 12 secondary indicators. The primary indicators include rotor-bearing system health (weight 0.4), structure-foundation system health (weight 0.3), fault severity (weight 0.2), and energy flow anomaly (weight 0.1). The secondary indicators are various extracted feature parameters. The AHP is used to determine the weight of each indicator, construct a judgment matrix, and calculate the weight of each indicator. The fuzzy comprehensive evaluation involves establishing an evaluation set. The trapezoidal membership function is used to determine the membership degree of each indicator to different evaluation levels, and a fuzzy evaluation matrix is ​​constructed. The comprehensive evaluation result of the system is obtained through fuzzy matrix operations; the Long Short-Term Memory (LSTM) network trend prediction: an LSTM model is constructed to predict the future trend of the system's comprehensive health index. The model contains two LSTM layers, each with 64 hidden units, one random deactivation layer (Dropout layer) with a dropout rate of 0.2, and one fully connected layer. A sliding window method is used to generate training samples, with a window size of 30 days and a step size of 1 day. The input is the health index of the past 30 days, and the output is the health index of the next 7 days. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 50 training epochs. When the health index is predicted to be lower than 0.6 within the next 7 days, the module automatically issues a warning and provides corresponding maintenance suggestions; the health analysis is used to obtain the system's comprehensive health index. ,in, For the first The weight of each primary indicator, For the first The evaluation value of each primary indicator, The value ranges from 0 to 1, with a value closer to 1 indicating a better system operating condition. Specifically, the state assessment unit and trend prediction unit continuously execute during system operation, updating the assessment results every second. First, the state assessment unit establishes a gas turbine system state evaluation index system containing four primary indicators and twelve secondary indicators. All secondary indicators are derived from the output of the preceding modules. The four primary indicators are rotor-bearing system health, structure-foundation system health, fault severity, and energy flow anomaly, with corresponding weights of 0.4, 0.3, 0.2, and 0.1, respectively. The state assessment unit uses the analytic hierarchy process (AHP) to determine the weights of each indicator, inviting ten gas turbine experts to perform pairwise comparisons and scores, constructing a judgment matrix, and calculating the weights. Subsequently, an evaluation set is established, containing four levels: excellent, good, qualified, and unqualified. A trapezoidal membership function is used to determine the membership degree of each indicator to different evaluation levels. A fuzzy evaluation matrix is ​​constructed, and the comprehensive evaluation result of the system is obtained through fuzzy matrix operations. The comprehensive health index of the system is calculated, with a value ranging from zero to one, where a value closer to one indicates a better system operating condition. The trend prediction unit constructs a Long Short-Term Memory (LSTM) network model to predict the future trend of the system's comprehensive health index. The model includes two LSTM network layers, one Dropout layer, and one fully connected layer. This module is divided into two independent sub-units, each adapted to different operation and maintenance scenarios. The two sub-units are interconnected but functionally independent: 1. Real-time health assessment sub-unit: refreshed every second, based on the fuzzy comprehensive evaluation method, integrating real-time feature parameters output from previous modules to calculate the gas turbine system's comprehensive health index, used for real-time fault warning and anomaly alarm; 2. Long-term trend prediction sub-unit: updated daily, based on the LSTM network model, using the daily comprehensive health index of the past 30 days as input to predict the daily health trend for the next 7 days, used for formulating equipment preventative maintenance plans. The trend prediction unit constructs the LSTM network model, using the sliding window method to generate training samples. The window size is 30 days, the step size is 1 day, the input is the daily comprehensive health index of the past 30 days, and the output is the predicted daily health index for the next 7 days. The model comprises two LSTM layers (64 hidden units per layer), one Dropout layer, and one fully connected layer. After training, it automatically updates the prediction results daily at midnight using the previous day's health data. When the health index is predicted to fall below 0.6 within the next seven days, a maintenance warning is triggered. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 50 training epochs. When the trend prediction unit predicts that the health index will fall below 0.6 within the next seven days, it automatically issues a warning and provides corresponding maintenance suggestions based on fault identification and tracing results. After evaluation and prediction, the status evaluation unit and trend prediction unit output the system's overall health index, the health trend for the next seven days, the warning information, and the maintenance suggestions to the visualization and interactive module.

[0029] In addition, the long short-term memory network health trend prediction model is explained as follows: Overall Model Architecture and Hierarchical Connections: This model is specifically designed for time-series prediction of the comprehensive health index of gas turbine systems. It adopts a sequence-to-sequence architecture, consisting of two Long Short-Term Memory (LSTM) network layers, a Dropout layer, and a fully connected layer connected in series. The first LSM layer extracts short-term features from the input time series, capturing the changing patterns of the health index over a short period. This layer contains 64 hidden units, each with three gating structures: an input gate, a forget gate, and an output gate, effectively controlling the flow and retention of information. The second LSM layer extracts long-term features from the input time series, capturing the changing trends of the health index over a longer period. This layer also contains 64 hidden units, with the output of the first LSM layer as its input, allowing for further abstraction of higher-level temporal features based on short-term features. The Dropout layer, placed after the two LSM layers, prevents overfitting. During training, this layer randomly discards a certain percentage of neuron outputs, preventing the model from over-relying on specific features and improving its generalization ability. Fully connected layer: As the output layer of the model, it is responsible for mapping the temporal features output by the second Long Short-Term Memory network layer to predicted health index values ​​for the next seven days. This layer contains seven neurons, each corresponding to a predicted health index value for the next day.

[0030] Model Training Steps: Model training is completed during the initial system deployment, and subsequently incrementally updated quarterly along with the fault identification model. Training Sample Generation: Training samples are generated using a sliding window method. Based on the historical comprehensive health index data of the gas turbine over the past three years, a window size of thirty days and a step size of one day are set. Each time the window slides, the health index of the thirty days within the window is used as the input sample, and the health index of the following seven consecutive days is used as the corresponding label sample. Approximately one thousand training samples are generated in this way. All samples are randomly divided into training and validation sets at an 8:2 ratio. Model Training Process: The optimizer uses the adaptive moment estimation algorithm, with an initial learning rate set to 0.1%, a batch size of 64, and a total of fifty training rounds. In each training round, the mean squared error loss function value is calculated on the validation set. When the validation set loss value no longer decreases for five consecutive rounds, training is terminated early, and the optimal model parameters are saved. Incremental Update Process: At the end of each quarter, the model is incrementally updated using the newly generated health index data from that quarter. During updates, new training samples are generated using the same sliding window method as during initial training. The model is then fine-tuned using these new samples for twenty training epochs. Incremental updates allow the model to continuously learn the latest patterns of health status changes, improving prediction accuracy.

[0031] Model Integration with Gas Turbine Condition Assessment: This model is specifically designed for the preventative maintenance needs of gas turbines. It can predict future condition changes based on historical health data, providing a scientific basis for operation and maintenance decisions. Input Data Settings: The model's input is the sequence of the gas turbine system's comprehensive health index over the past thirty days. The comprehensive health index is calculated by the system condition dynamic assessment module, integrating information from four aspects: rotor-bearing system health, structure-foundation system health, fault severity, and energy flow anomaly, comprehensively reflecting the overall operating status of the gas turbine system. The input data time step is one day, with each time step corresponding to the comprehensive health index for that day. Output Data Settings: The model's output is the predicted sequence of the gas turbine system's comprehensive health index for the next seven days. Each output value corresponds to the predicted health index result for the next day, ranging from zero to one. The closer the value is to one, the better the system's operating status. Application of Output Results: The health trend prediction results output by the model are directly input into the digital twin-driven 3D visualization interactive module, displayed to operation and maintenance personnel in the form of trend curves. When the model predicts that the health index will fall below the set warning threshold within the next seven days, the system will automatically issue a warning message and, based on the fault identification and tracing results, provide corresponding maintenance suggestions. Maintenance personnel can use the warning information to plan maintenance in advance, avoiding unplanned downtime.

[0032] The three-dimensional visualization module uses a graphical language transmission format (glTF format) for modeling, completes real-time data mapping, and realizes visualization of vibration status and interactive display of faults.

[0033] It should be noted that the graphical language transmission format modeling involves establishing a high-precision three-dimensional geometric model based on the gas turbine's design drawings and actual dimensions. The model uses the graphical language transmission format version 2.0 (glTF2.0). The data mapping involves mapping the collected real-time data and analysis results from other modules onto the three-dimensional geometric model in real time via the Message Queue Telemetry Transmission Protocol version 3.1.1 (MQTT3.1.1) protocol, forming a digital twin of the gas turbine system. The data update frequency is 1 Hz. The vibration state visualization uses color coding technology to represent the vibration intensity of different parts. The vibration intensity changes from 0 to the maximum value, corresponding to a color gradient from blue to red. Real-time values ​​and trends of the oil film instability characteristic index, rotor thermal bending characteristic index, and shaft misalignment characteristic index are displayed at key parts such as the rotor and bearings. The system displays real-time values ​​of the casing-rotor modal coupling coefficient, foundation vibration isolation system transmissivity, and stiffness attenuation rate at the casing, foundation, and pipe connection points. Dynamic arrows represent the vibration energy transmission path, with arrow thickness proportional to the energy transmission coefficient. Clicking on any monitoring point allows viewing real-time vibration data, historical trend curves, and spectrum analysis results for that point. The fault interactive display automatically switches the view to the fault location when a fault occurs, highlighting and flashing the fault location, and playing a fault animation to show the fault's development process. Simultaneously, a fault details window pops up, displaying the fault type, severity, possible causes, and maintenance suggestions. The module supports historical data playback, allowing users to select any historical time period and replay the engine's vibration state changes at speeds of 0.5x, 1x, 2x, and 4x. Specifically, the visualization rendering unit runs continuously after system startup, updating the digital twin's state every second. First, the rendering unit establishes a high-precision 3D geometric model based on the engine's design drawings and actual dimensions. The model uses the glTF2.0 format to improve loading speed and rendering efficiency. The real-time data collected by the multi-source heterogeneous data synchronous acquisition module and the analysis results from other modules are mapped onto the 3D geometric model in real time via a message queue telemetry transmission protocol, forming a digital twin of the gas turbine system. The rendering unit achieves multi-dimensional data visualization: color coding technology is used to represent the vibration intensity of different parts, with the vibration intensity changing from minimum to maximum value corresponding to a color gradient from blue to red; real-time values ​​and trends of three fault sensitivity characteristic indices are displayed for key parts such as the rotor and bearings; real-time values ​​of three coupling parameters and stiffness attenuation rate are displayed for the casing, foundation, and pipe connection parts; dynamic arrows are used to represent the vibration energy transmission path, with arrow thickness proportional to the energy transfer coefficient. When a fault occurs, the digital twin automatically switches the view to the fault location, highlighting and flashing it, and plays a fault animation to show the fault development process. Simultaneously, a fault details window pops up, displaying the fault type, severity, possible causes, energy transmission path, and maintenance recommendations.The module also supports historical data playback, allowing users to select any historical time period and play back the vibration state changes of the internal combustion engine at speeds of 0.5x, 1x, 2x, and 4x, providing support for fault analysis and accident investigation.

[0034] In summary, as in Example 1: Early Warning of Oil Film Whirl During the startup and acceleration process of a gas turbine at a power plant, with a rated speed of 3,000 revolutions per minute, a high-precision synchronous acquisition and preprocessing module for multi-source heterogeneous data collected rotor radial vibration signals through an eddy current sensor deployed at bearing number one. All data were synchronized at the microsecond level using a precise time protocol. After hardware filtering and baseline correction, the preprocessing unit transmitted the data to the rotor-bearing system dynamic feature hierarchical extraction module. The feature extraction unit first extracted basic features and found a slight upward trend in the half-frequency amplitude. Subsequently, advanced feature extraction was performed, calculating the growth rate of the half-frequency amplitude of oil film eddy current and the phase difference fluctuation at a step size of one revolution per minute during the acceleration process. After dimensionless processing, the fused calculation yielded an oil film instability characteristic index of 0.42, which is within the early fault warning range. Simultaneously, the multi-scale vibration signal time-frequency fusion and feature enhancement module performed wavelet packet decomposition and Hilbert-Huang transform on the vibration signal to obtain the fused time-frequency distribution, clearly showing the energy concentration region at 0.48 times the power frequency. The attention-based multi-channel fault identification module receives the rotor feature vector and time-frequency feature map, analyzes them through a multi-channel convolutional neural network, and outputs the identification result of "early oil film whirl in bearing No. 1" with a fault confidence level of 92%. After the vibration energy flow transmission path tracking and source tracing module is activated, it is found that the energy proportion of the rotor node is 42%, which is a suspected fault node. The main energy transmission path is rotor → bearing No. 1 → casing. Combined with the casing-rotor modal coupling coefficient of 0.32, casing-rotor coupling faults are ruled out, and the root cause of the fault is finally determined to be oil film whirl in bearing No. 1. The system status dynamic assessment and trend prediction module calculates the system comprehensive health index to be 0.73, predicts that the health index will drop to 0.58 in the next two hours, issues a yellow warning, and suggests reducing the acceleration rate and closely monitoring vibration changes. The digital twin-driven 3D visualization interactive module automatically switches the perspective to the bearing No. 1 part, highlights it in yellow, and shows the development process of oil film whirl and maintenance suggestions. The maintenance personnel adjusted the speed-up strategy in a timely manner based on the early warning information, thus preventing oil film oscillation.

[0035] Example 2: Identification of Loose Pipe Connection Bolts During steady-state operation of a gas turbine in a petrochemical plant, a high-precision synchronous acquisition and preprocessing module for multi-source heterogeneous data collected pipeline vibration signals and relative displacement signals of the connection surface through acceleration sensors and eddy current sensors deployed at the connection between the pipeline and the cylinder. The coupling parameter calculation unit of the structure-foundation coupling transmission characteristic modeling and quantification module calculated the pipeline-cylinder connection stiffness in real time and found that the stiffness value was significantly lower than the design value. The calculation unit further calculated that the pipeline-cylinder connection stiffness attenuation rate was 26%, exceeding the 20% warning threshold, and immediately issued a bolt loosening warning. Simultaneously, the multi-channel fault identification module based on an attention mechanism received the structure-foundation coupling characteristic vector and output the identification result "loose bolt of pipeline No. 3," with a fault confidence level of 98%. After the vibration energy flow transmission path tracing and source tracing module was activated, it calculated that the energy proportion of the pipeline node was 38%, making it a suspected fault node, with the main energy transmission path being pipeline → cylinder → casing. Combined with the result that the stiffness attenuation rate exceeded 20%, the root cause of the fault was directly determined to be the loose bolt of pipeline No. 3. The system status dynamic assessment and trend prediction module calculated the system's overall health index to be 0.68, predicting that it would drop to 0.54 within the next 24 hours, issuing an orange warning and recommending that the bolts be tightened during the next planned shutdown. The digital twin-driven 3D visualization interaction module automatically switched the view to the connection point of pipe number 3, highlighting it in orange and displaying the location of the loose bolts and maintenance steps. Maintenance personnel tightened the bolts during the next shutdown, and after restarting, the stiffness decay rate recovered to below 5%, and the system operated normally.

[0036] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A gas turbine vibration monitoring and analysis system, characterized in that, include: Data synchronization acquisition module: adopts a two-layer synchronization architecture, deploys multiple monitoring point sensors, and completes data preprocessing and synchronization error quantification; Dynamic feature extraction module: Achieves sensitive detection of faults at different stages through a three-level extraction strategy; Coupling characteristic modeling module: Employs a hybrid modeling approach to quantify the three-parameter coupling transfer characteristics; Time-frequency fusion enhancement module: Performs wavelet packet multi-scale decomposition, combined with Hilbert-Huang transform, to achieve Gaussian weighted fusion and feature enhancement; Fault Intelligent Recognition Module: Constructs a deep learning model and uses transfer learning and incremental learning to complete fault mode recognition and classification; Energy flow tracing module: Establish a five-node energy model, obtain energy transfer coefficients, solve energy flow equations, and locate the root cause of the fault; Status assessment and prediction module: The weights are determined by the analytic hierarchy process (AHP), and fuzzy comprehensive evaluation is performed to achieve LSTM trend prediction and health analysis. 3D visualization module: Modeling is performed using glTF format to complete real-time data mapping, enabling visualization of vibration status and interactive display of faults.

2. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The dual-layer synchronization architecture combines a precise time protocol with GPS / BeiDou dual-mode time synchronization. A master clock unit is deployed in the gas turbine control room, and UTC time is acquired through a GPS / BeiDou dual-mode receiver. The multiple monitoring point sensors include 12 sensors: one eddy current sensor is installed in each of the X and Y vertical directions of each sliding bearing section to collect rotor radial vibration displacement; one piezoelectric accelerometer is installed on each bearing housing to collect casing vibration acceleration; one strain gauge torque sensor is installed on the coupling to collect shaft transmission torque; and six high-precision LVDT displacement sensors are installed at the four corners and the center of the gas turbine foundation to collect... For basic settlement displacement, an accelerometer is installed 1 meter from the pipe connection surface to collect pipe vibration acceleration. The data preprocessing involves first passing the collected raw data through a second-order Butterworth hardware filter circuit to remove power frequency interference and high-frequency noise. The cutoff frequency of the low-pass filter is set to 10 times the rated power frequency of the gas turbine. Then, software baseline correction is performed using a fifth-order polynomial fitting trend term removal method. The trend term of the fitted signal is subtracted from the original signal. Finally, data with different sampling rates are aligned by timestamps, converted to standard HDF5 format, and stored in a distributed time series database. Low sampling rate data is used for long-term trend analysis, and high sampling rate data is used for transient feature extraction. The synchronization error quantization is used to obtain the data synchronization error index. , and when If the time exceeds 5 microseconds, the module will automatically trigger the resynchronization process.

3. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The three-level extraction strategy is used to extract basic features, high-level features, and fault-sensitive features; The basic feature extraction method is as follows: A Fast Fourier Transform with a Hanning window is applied to the preprocessed vibration signal, with a window length of 1024 points and an overlap rate of 50%, to obtain the signal's spectral distribution. Basic feature parameters, including the power frequency amplitude, second harmonic amplitude, half-frequency amplitude, and total effective value of the vibration signal, are extracted for preliminary judgment of whether the system has any abnormalities. The advanced features include oil film whirl-oil film oscillation features, rotor thermal bending features, and shaft misalignment features. The oil film whirl-oil film oscillation feature is extracted as follows: During the gas turbine's acceleration process, vibration signals at the bearing are continuously acquired in steps of 1 revolution per minute. An adaptive bandpass filter of 0.4 to 0.5 times the power frequency is applied to the signal at each speed point. The effective value of the filtered signal is calculated to obtain the speed-half-frequency amplitude curve. The curve is then subjected to cubic spline interpolation, and the first derivative is calculated to obtain the oil film whirl half-frequency amplitude growth rate. Simultaneously, the difference between the maximum and minimum values ​​of the phase difference of the X and Y directions (0.4 to 0.5 times the power frequency components) of the same bearing section within 10 seconds is calculated to obtain the oil film whirl phase difference fluctuation. The method for extracting rotor thermal bending features is as follows: synchronously acquire the rotor vibration power frequency phase and the highest temperature on the rotor surface, accurately record the time when the temperature peak and vibration peak occur, and calculate the mechanical angle through which the rotor rotates between the two times to obtain the rotor thermal bending hysteresis angle. The axial vibration displacement at the same moment is measured by two eddy current sensors installed on the front and rear ends of the rotor thrust disk, and the absolute difference between the two measurements is calculated to obtain the thermal bending axial vibration difference. The method for extracting shaft misalignment features is as follows: extract the amplitudes of the twice-power-frequency components in the X and Y directions of the same shaft section on both sides of the coupling, and calculate the absolute difference between the amplitudes in the two directions of the same section to obtain the misalignment twice-frequency amplitude difference. The torque transmitted by the shaft system is continuously measured using a strain gauge torque sensor at a sampling rate of 10 kHz. The difference between the maximum and minimum torque values ​​within 10 seconds is used to calculate the torque fluctuation of the coupling. The method for extracting the fault-sensitive features is as follows: after normalizing and dimensionlessly processing the high-level features to their maximum values, they are fused to obtain the oil film instability feature index. Rotor thermal bending characteristic index Shaft misalignment characteristic index ;in, , This represents the growth rate of the half-frequency amplitude of the oil film eddy after dimensionless scaling. This represents the dimensionless fluctuation of the oil film eddy phase difference. Let be the angular velocity of the gas turbine rotor. The value range is from 0 to 1. The higher the value, the higher the risk of oil film instability: when When <0.3, the oil film is stable; when 0.3≤ When <0.7, there is a risk of early oil film eddy; when When the oil film oscillation is ≥0.7, the machine must be stopped immediately for inspection. , The dimensionless rotor thermal bending hysteresis angle. The dimensionless thermal bending axial vibration difference. The larger the value, the more severe the rotor thermal bending: when When the value is less than 0.3, the rotor shows no obvious thermal bending. When 0.3≤ When <0.7, slight thermal bending exists; when When the value is ≥0.7, severe thermal bending occurs, requiring shutdown for cooling. , This represents the dimensionless difference in the second harmonic amplitude of the misalignment. This represents the dimensionless torque fluctuation of the coupling. The larger the value, the more severe the misalignment of the shaft system: when When 0.3 < 0.3, the shaft system is well aligned; when 0.3 ≤ When <0.7, there is a slight misalignment; when When the value is ≥0.7, the misalignment is severe and realignment is required.

4. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The hybrid modeling method combines finite element modeling with experimental modal analysis to complete geometric modeling and mesh generation, experimental modal testing, and finite element model correction. Geometric modeling and mesh generation: First, a 3D geometric model of the gas turbine casing, foundation, vibration isolators, and piping is created using SolidWorks. This model is then imported into ANSYS for mesh generation. The casing uses Shell181 elements, the foundation uses Solid186 elements, the vibration isolators use Combin14 spring-damped elements, and the piping uses Beam188 elements. Mesh generation combines free meshing and mapped meshing, with mesh refinement applied to key areas such as bearing housings and flanges. The total number of mesh elements is controlled to within 500,000. The experimental model… Modal testing: A multi-point excitation, single-point response method was used for experimental modal analysis. A pulsed hammer applied excitation in three orthogonal directions to the housing bearing seat. Twenty accelerometers were evenly distributed on the housing surface as response points to collect excitation and response signals. The experimental modal parameters of the housing and foundation were estimated using the frequency response function. Finite element model correction: A combination of sensitivity analysis and genetic algorithm was used. First, sensitivity analysis identified the parameters with the greatest impact on modal parameters, including the housing elastic modulus, vibration isolator stiffness, and foundation damping. Then, with the objective function of minimizing the error between the experimental and calculated modal frequencies, the genetic algorithm optimized these parameters, ensuring that the error between the calculated and experimental results of the corrected model was less than 3%. The three-parameter coupling transmission characteristics include the casing-rotor modal coupling coefficient. Transmissivity of the foundation vibration isolation system Pipe-cylinder connection stiffness ,in, , The casing modal damping ratio, For the frequency difference between the casing and the rotor, The operating frequency of the gas turbine rotor. The value ranges from 0 to 1, and the larger the value, the higher the degree of modal coupling. , This represents the amplitude of the vibration acceleration at the upper end of the vibration isolator. This represents the amplitude of the vibration acceleration at the lower end of the vibration isolator. The difference in vibration acceleration between the upper and lower parts of the vibration isolator. A value less than 1 indicates effective vibration isolation; the smaller the value, the better the vibration isolation effect. >1 indicates that the vibration is amplified, and the vibration isolator is at risk of failure; Pipe-cylinder connection stiffness: , For the dynamic stress of the connecting bolts, The cross-sectional area of ​​the connecting bolt. The module calculates the relative displacement between the pipe and cylinder connection surfaces and further calculates the stiffness attenuation rate. Used to assess the looseness of connecting bolts: ,in Design stiffness for pipe-cylinder connection bolts, The value ranges from 0 to 1, with a larger value indicating a more severe loosening of the bolt. When the percentage is less than 10%, the bolted connection is in good condition; when the percentage is less than or equal to 10%, the bolted connection is in good condition. If the percentage is less than 20%, the bolts are slightly loose and need to be checked during the next shutdown. When the rate is ≥20%, there is a serious risk of bolt loosening, and the module will immediately issue a warning message to prompt maintenance personnel to handle the situation promptly.

5. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The wavelet packet multi-scale decomposition uses the db4 wavelet to perform three-level wavelet packet decomposition on the preprocessed vibration signal, obtaining eight sub-band signals at different frequency scales. The frequency range of each sub-band signal is determined according to the rated speed of the gas turbine. The Hilbert-Huang transform is performed on each sub-band signal using empirical mode decomposition to obtain several intrinsic mode functions (IMFs). The correlation coefficient method is used to eliminate pseudo-IMF components; components with a correlation coefficient less than 0.1 are considered pseudo-components. The remaining IMF components are then subjected to Hilbert transform to obtain the instantaneous frequency and instantaneous amplitude of each component, thereby constructing the Hilbert time-frequency distribution of each sub-band signal. The Gaussian weighted fusion assigns a Gaussian weight function to each sub-band based on its center frequency and bandwidth, highlighting features near the center frequency and improving the focus of the time-frequency distribution. The fused time-frequency distribution... , For the first The time-frequency distribution of each sub-band signal, For the first The center frequency of the sub-band signal For the first The frequency bandwidth of each sub-band signal is calculated, and the fused time-frequency distribution is converted into a 224×224 grayscale image, which is then input into the fault intelligent identification module as a time-frequency feature map. The feature enhancement employs the 3σ criterion to process the fused time-frequency distribution and calculates the mean of the background noise in the time-frequency distribution. with standard deviation It will be below the threshold Pixels are set to zero, while feature regions above the threshold are retained.

6. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The deep learning model employs an MC-CNN network architecture, comprising three input channels: the first channel inputs the dynamic feature vector of the rotor-bearing system, the second channel inputs the structure-foundation coupling and transmission characteristic vector, and the third channel inputs a multi-scale time-frequency fusion feature map. Each input channel corresponds to an independent convolutional branch. The rotor feature branch and the structure feature branch utilize a 1D convolutional neural network, each containing two 1D convolutional layers, one max-pooling layer, and one batch normalization layer. The time-frequency feature branch utilizes a 2D convolutional neural network, containing three 2D convolutional layers, two max-pooling layers, and one batch normalization layer. The local feature maps extracted from these three branches are input to the channel attention module. The SE module learns the importance weights of different channel features through global average pooling and two fully connected layers, performing weighted fusion of the feature maps. The fused feature map is then input to two fully connected layers, and finally, a softmax classifier outputs 12 types of fault identification. The transfer learning process involves pre-training the model using a publicly available gas turbine vibration fault dataset. This dataset contains 10,000 samples across 10 common fault categories. The Adam optimizer is used, with an initial learning rate of 0.001, decreasing to 0.1 every 10 epochs. The batch size is 32, and the training run consists of 100 epochs. The model is then fine-tuned using historical operating data from the target gas turbine. The incremental learning process employs an elastic weight consolidation algorithm. When a new fault mode is identified, the module automatically adds its features to the fault knowledge base and retains the weights important to older fault modes to prevent catastrophic forgetting. The model undergoes incremental updates quarterly. The fault mode recognition and classification process uses a pre-trained MC-CNN model to process the input multi-source heterogeneous features and outputs the identification results of fault type and fault location.

7. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The five-node energy model divides the gas turbine system into five subsystems: rotor, bearings, casing, foundation, and piping. Each subsystem is considered an independent energy node. The vibration energy of each energy node is calculated from the vibration signals of all monitoring points at that node. For acceleration signals, the vibration energy is proportional to the square of the acceleration. The energy transfer coefficient is calculated for any two adjacent energy nodes. and The energy transfer coefficient between two nodes is calculated by measuring their vibration power. The energy transfer coefficient represents the energy transfer rate from an energy node per unit time. Passed to node Energy occupied by nodes The proportion of total energy is calculated using the following formula: ,in, For the node Passed to node Vibration power, For nodes Vibrational energy, The dimension is the reciprocal of per second; the solution to the energy flow equation involves establishing the energy flow balance equation of the gas turbine system to describe the energy transfer and dissipation relationships between energy nodes: ,in, For the first The energy dissipation rate of each energy node, expressed as the reciprocal of its value per second. For time, the fourth-order Runge-Kutta method was used to solve the differential equation with a time step of 0.01 seconds, obtaining the variation of vibration energy of each energy node with time. To locate the root cause of the fault, the energy percentage of each node was first calculated; nodes with an energy percentage exceeding 30% were considered suspected fault nodes. Simultaneously, preliminary screening was performed using structure-foundation coupling parameters: if the casing-rotor modal coupling coefficient... If the value is >0.7, prioritize checking for casing-rotor coupling faults. If the transmissibility of the foundation vibration isolation system is... If the value is greater than 1.2, then the failure of the vibration isolator should be investigated first. If the stiffness attenuation rate is... If the failure rate is >20%, it is directly determined to be a fault caused by loose pipe connection bolts. Then, the energy transfer coefficient from the suspected fault node to other nodes is calculated, and the path with the highest transfer coefficient is the main transfer path. Finally, the root cause of the fault is determined by combining the fault type output by the multi-channel fault identification module.

8. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The analytic hierarchy process (AHP) is used to determine the weights and establish a gas turbine system state evaluation index system comprising 4 primary indicators and 12 secondary indicators. The primary indicators include rotor-bearing system health, structure-foundation system health, fault severity, and energy flow anomaly. The secondary indicators are extracted feature parameters. The AHP is used to determine the weights of each indicator, construct a judgment matrix, and calculate the weights of each indicator. The fuzzy comprehensive evaluation involves establishing an evaluation set. The trapezoidal membership function is used to determine the membership degree of each indicator to different evaluation levels, and a fuzzy evaluation matrix is ​​constructed. The comprehensive evaluation result of the system is obtained through fuzzy matrix operations. The LSTM trend prediction: an LSTM model is constructed to predict the future trend of the comprehensive health index of the system. The model contains two LSTM layers, each with 64 hidden units, one Dropout layer with a dropout rate of 0.2, and one fully connected layer. The sliding window method is used to generate training samples with a window size of 30 days and a step size of 1 day. The input is the health index of the past 30 days, and the output is the health index of the next 7 days. The model training uses the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 50 training rounds. When it is predicted that the health index will be lower than 0.6 in the next 7 days, the module automatically issues an early warning and provides corresponding maintenance suggestions. The health analysis is used to obtain the system's overall health index. , The value ranges from 0 to 1, and the closer the value is to 1, the better the system is running.

9. The gas turbine vibration monitoring and analysis system according to claim 1, characterized in that: The glTF format modeling involves establishing a high-precision three-dimensional geometric model based on the gas turbine's design drawings and actual dimensions. The model uses the glTF 2.0 format. The data mapping involves mapping the collected real-time data and analysis results from other modules onto the three-dimensional geometric model in real time via the MQTT 3.1.1 protocol, forming a digital twin of the gas turbine system. The data update frequency is 1 Hz. The vibration state visualization uses color coding technology to represent the vibration intensity of different parts, with the vibration intensity changing from 0 to the maximum value corresponding to a color gradient from blue to red. The fault interactive display automatically switches the view to the fault location when a fault occurs, highlighting and flashing the fault location, playing a fault animation to show the development process of the fault, and simultaneously popping up a fault details window to display the fault type, severity, possible causes, and maintenance suggestions.

Images