Turbine key parameter inversion model training method and device, inversion method and maintenance method

Abstract

The invention discloses an inversion method for key parameters of a steam turbine. The inversion method comprises the steps that first monitoring parameters of the steam turbine are acquired; performing feature extraction on the first monitoring parameter to obtain a first feature parameter; and constructing an adjacent matrix A, and training a mechanism enhanced graph neural network according to the first feature parameter and the adjacent matrix to obtain an inversion model. According to the method, the mechanism enhanced graph neural network technology is creatively applied, multi-source sensor data and a physical mechanism model are deeply fused, and soft measurement of parameters such as oil film characteristics and shaft seal dynamic gaps is achieved. The invention aims to fill up the capability weakness of the TSI system in the aspect of key internal state parameter monitoring, and provides a direct and reliable quantitative basis for refined operation optimization, condition-based maintenance decision and early fault early warning of the steam turbine, so that the safety, economy and intelligent management level of a unit are comprehensively improved.

Description

Technical Field

[0001] This application relates to the field of steam turbine condition monitoring and fault diagnosis technology, specifically to a method, device, inversion method, and maintenance method for training an inversion model of key parameters of a steam turbine. Background Technology

[0002] As the core equipment of modern energy and power systems, the safety, stability, and economy of large steam turbine units are of paramount importance. Within the complex mechanical structure of a steam turbine, a series of key internal parameters, such as the oil film thickness, oil film pressure distribution, oil film stiffness and damping coefficient of each sliding bearing, as well as the dynamic operating clearance of each stage of shaft seals (steam seals and oil seals), and the top shaft oil film height under turning gear conditions, directly determine the unit's vibration characteristics, operating efficiency, component wear rate, and overall lifespan.

[0003] However, due to the extreme working environment of high temperature, high pressure, and high-speed rotation inside the steam turbine, these key parameters are often difficult to measure directly, continuously, and accurately online using conventional sensors. Currently, engineering practice mainly relies on the following indirect or offline methods to obtain these parameters: First, theoretical estimation or empirical formula deduction based on design values. This method ignores the complexity and variability of actual operating conditions, and the results deviate significantly from the true values. Second, it relies on cold-state measurements during shutdown and maintenance, such as using feeler gauges to measure static clearances. The data obtained by this method is severely lagging and cannot reflect the dynamic changes such as thermal expansion and force deformation of the unit under operating conditions. Often, by the time problems are discovered, irreversible performance degradation or safety hazards have already occurred. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, this application provides a method, apparatus, inversion method and maintenance method for training a key parameter inversion model of a steam turbine, to solve at least one defect in the prior art.

[0005] This application provides a method for training a key parameter inversion model for a steam turbine, the method comprising: Obtain the first monitoring parameters of the steam turbine; The first feature parameter is obtained by extracting features from the first monitoring parameter; Construct an adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) consists of the key components of the steam turbine, and the edge set E consists of the physical connections or mutual influence relationships between the key components; An inversion model is obtained by enhancing the graph neural network based on the first feature parameters and the adjacency matrix training mechanism.

[0006] In one embodiment of this application, the first feature parameter includes: Time-domain characteristics, including axis trajectory parameters, vibration amplitude, and DC bias; Frequency domain characteristics include the frequency component amplitude and phase, and the full-spectrum energy distribution extracted by performing a fast Fourier transform on the monitoring parameters. Time-frequency domain features, including time-frequency features obtained through wavelet transform or short-time Fourier transform; Correlation features include the correlation between axial displacement variation patterns and thermal expansion.

[0007] In one embodiment of this application, the step of obtaining the inversion model based on the first feature parameters and the adjacency matrix training mechanism of the enhanced graph neural network includes: A graph structure is defined using the adjacency matrix A. A graph neural network is selected as the basic architecture, and a constraint term based on a physical mechanism model is added to the loss function of the graph neural network to construct a mechanism-enhanced graph neural network model. The mechanism-enhanced graph neural network model is trained using the first feature parameters and the corresponding label data to obtain a trained inversion model.

[0008] In one embodiment of this application, constructing the adjacency matrix A includes: The key components of the steam turbine system are defined as graph nodes; Determine the relationships between the nodes in the graph, including the first association with direct physical connection, the second association without direct physical connection but with mutual influence, and the third association without direct physical connection and without mutual influence. Assign weight values ​​to the first association, the second association, and the third association, respectively; Adjacency moments are constructed using the weight values ​​corresponding to the first association, the second association, and the third association as matrix elements.

[0009] In one embodiment of this application, the mechanism-enhanced graph neural network includes: Graph convolutional layers are used to perform spatial feature aggregation operations on nodes in the graph structure to obtain preliminary features of each node and its adjacent nodes. The graph attention layer is used to assign differentiated weights to different neighbor nodes through an attention mechanism, generating weighted node features. The temporal modeling layer is used to process the time series data of the weighted node features through a temporal neural network based on the weighted node features, and to capture the dynamic evolution of parameters in the time dimension. The physical constraint layer is used to calculate the constraint terms of the physical mechanism model based on the output of the temporal modeling layer during model training, and to add the constraint terms to the loss function.

[0010] In one embodiment of this application, the constraint includes at least one of the following: Fluid dynamic constraints based on Reynolds equations; Mechanical equilibrium constraints based on rotor dynamics equations; Thermal expansion constraints based on thermodynamic equations.

[0011] In one embodiment of this application, the first monitoring parameter includes: Based on the detailed physical model of the steam turbine constructed through finite element analysis, parametric simulation was performed to generate simulation data covering different operating conditions; Historical operating data of steam turbines under different operating conditions; The measured data obtained by fixed-point calibration under preset working conditions; In the process of training the mechanism-enhanced graph neural network, a phased training strategy is adopted. First, simulation data is used for pre-training, and then historical running data and measured data are used for fine-tuning.

[0012] This application provides a training device for a key parameter inversion model of a steam turbine, the training device comprising: The data acquisition module is used to acquire the monitoring parameters of the steam turbine; The feature extraction module is used to extract features from the monitoring parameters to obtain a feature parameter set; The matrix construction module is used to construct the adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) consists of the key components of the steam turbine, and the edge set E consists of the physical connections or mutual influence relationships between the key components; The model training module is used to enhance the graph neural network based on the feature set and the adjacency matrix training mechanism to obtain the inversion model.

[0013] This application provides a method for inverting key parameters of a steam turbine, the method comprising: Obtain the second monitoring parameters of the steam turbine; The second monitoring parameter is subjected to feature extraction to obtain the second feature parameter; Based on the second feature parameter, the inversion result is obtained through the inversion model.

[0014] This application provides a maintenance method for a steam turbine, the maintenance method comprising: Obtain the third monitoring parameter of the steam turbine; The third monitoring parameter is subjected to feature extraction to obtain the third feature parameter; Based on the third feature parameter, the inversion result is obtained through the inversion model. Identify abnormal parameters based on the inversion parameters and preset identification rules; The anomaly source is determined based on the aforementioned anomaly parameters and the adjacency matrix A; For the source of the anomaly, an operation and maintenance plan is generated to maintain the steam turbine based on the operation and maintenance plan.

[0015] The beneficial effects of this application are: This application discloses a method for training a key parameter inversion model for a steam turbine, comprising: acquiring first monitoring parameters of the steam turbine; extracting features from the first monitoring parameters to obtain first feature parameters; constructing an adjacency matrix A, wherein the adjacency matrix A is a mathematical representation of the edge set E in a graph structure G=(V,E); wherein the graph node set V in the graph structure G=(V,E) consists of key components of the steam turbine, and the edge set E consists of physical connections or mutual influence relationships between the key components; and training a mechanism-enhanced graph neural network based on the first feature parameters and the adjacency matrix to obtain an inversion model. This application innovatively applies mechanism-enhanced graph neural network technology to deeply integrate multi-source sensor data and physical mechanism models, achieving "soft measurement" of parameters such as oil film characteristics and shaft seal dynamic clearance. This application aims to fill the gap in the TSI system's ability to monitor key internal state parameters, providing direct and reliable quantitative basis for refined operation optimization, condition-based maintenance decision-making, and early fault warning of the steam turbine, thereby comprehensively improving the safety, economy, and intelligent management level of the unit. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 This is a flowchart illustrating a method for training a key parameter inversion model for a steam turbine according to an embodiment of this application; Figure 2 This is a flowchart illustrating the training of an inversion model according to an embodiment of this application; Figure 3 This is a flowchart illustrating the construction of an adjacency matrix according to an embodiment of this application; Figure 4This is a schematic block diagram of a key parameter inversion model training device for a steam turbine according to an embodiment of this application; Figure 5 This is a flowchart of a method for inverting key parameters of a steam turbine according to an embodiment of this application; Figure 6 This is a trend graph of an embodiment of this application; Figure 7 This is a flowchart illustrating a steam turbine maintenance method according to an embodiment of this application; Figure 8 A schematic diagram of a computer system suitable for implementing the memory of the embodiments of this application is shown. Detailed Implementation

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for training a key parameter inversion model for a steam turbine, according to an embodiment of this application. Figure 1 In China, the training methods for the key parameter inversion model of steam turbines include: Step S110: Obtain the first monitoring parameters of the steam turbine; Specifically, multiple eddy current sensors are deployed at various bearings, shaft seals, and other critical locations on the turbine to collect raw signals such as shaft vibration and axial displacement at high frequencies. For example, a pair of eddy current sensors (in the X and Y directions) are placed at each of the turbine's sliding bearings (#1 to #6) to measure rotor vibration. Near the thrust bearing, two eddy current sensors are deployed to monitor axial displacement. Additionally, eddy current sensors are also deployed near the last stage blades of the low-pressure cylinder to monitor blade tip clearance. All eddy current sensors continuously acquire data at a sampling frequency of 4096 Hz. Simultaneously, a data interface with the power plant's DCS system is established via the OPC protocol, acquiring DCS operating parameters related to the unit's status at a frequency of 1 second. These parameters mainly include: speed, load, unit active power, main steam pressure, main steam temperature, extraction steam pressure and temperature at each stage, condenser vacuum, and bearing inlet and return oil temperatures and pressures.

[0022] All acquired eddy current signals and DCS operating parameters are sent to the server. The server runs an NTP (Network Time Protocol) client to synchronize with the power plant's master clock server, resolving the time alignment issue of multi-source data and ensuring that all data has a unified, high-precision timestamp.

[0023] Meanwhile, the high-frequency eddy current data is segmented by second, and the DCS operating data corresponding to that second is cleaned: outlier detection and removal are performed on all data, such as removing invalid data points caused by sensor failure or communication interruption; bandpass filtering is performed on the signal to remove high-frequency noise and DC drift interference.

[0024] Step S120: Extract features from the first monitoring parameter to obtain the first feature parameter; The first feature parameter includes: time-domain features, frequency-domain features, time-frequency-domain features, and correlation features; Time-domain characteristics, including axis trajectory parameters, vibration amplitude, and DC bias; Shaft center trajectory parameters: The shaft center trajectory is plotted using vibration signals in the X and Y directions, and its major axis (L), minor axis (S), tilt angle (θ), and center of mass coordinates (x, y) are calculated. c ,y c Specifically, time-domain feature extraction can calculate peak-to-peak values ​​from 4096 data points in the X and Y directions. Using the X and Y signals, an axisymmetric trajectory is plotted, and through least-squares ellipse fitting, 12 geometric parameters, including the major axis, minor axis, inclination angle, and centroid coordinates, are calculated.

[0025] , where σ x ,σ y σ represents the standard deviation in the X and Y directions. xy For covariance.

[0026] Vibration amplitude (peak-to-peak value) App=max ( x ) min ( x ), and DC bias .

[0027] Frequency domain characteristics include the frequency component amplitude and phase, and the full-spectrum energy distribution extracted by performing a fast Fourier transform on the monitoring parameters. Specifically, an FFT is performed on the data segment. Assuming the rotational speed is 3000 RPM (50Hz), the amplitude and phase of the subharmonic components in the range of 50Hz (1X), 100Hz (2X), 150Hz (3X), 200Hz (4X), and 15-25Hz (0.3~0.5X) are extracted from the spectrum, forming a total of 16 frequency domain features.

[0028] Time-frequency domain features, including time-frequency features obtained through wavelet transform or short-time Fourier transform; Specifically, for transient processes such as start-up and shutdown, wavelet transform or short-time Fourier transform (STFT) is used to capture the dynamic frequency characteristics of the signal; for example, if the speed or load change rate is detected to exceed the threshold, STFT analysis is initiated on the data of that period to generate the time spectrum and extract the trajectory of the power frequency peak changing with time as a feature.

[0029] Correlation characteristics, including the correlation between axial displacement variation patterns and thermal expansion; specifically, the correlation characteristics between axial displacement variation patterns and thermal expansion include: 10-minute moving average (μ 10 The 1-minute standard deviation (σ1) reflects the trend of thermal expansion and serves as a slow variable of thermal expansion; the 1-minute standard deviation (σ1) reflects the thrust fluctuation and serves as a fast variable of thrust fluctuation; and the rate of change of axial displacement (Δd / Δt) reflects the change of steam pressure.

[0030] After the above processing, every second, the system generates a vector containing dozens of dimensional features for each measuring point. These vectors, together with the operating parameters for that second, constitute the input data of the model.

[0031] Step S130: Construct an adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) consists of the key components of the steam turbine, and the edge set E consists of the physical connections or mutual influence relationships between the key components. The physical topology of the steam turbine system is mathematically modeled. A steam turbine is not an isolated collection of components, but a complex system composed of numerous components interacting through mechanical connections, heat transfer, and fluid dynamics. The adjacency matrix A quantifies and embeds these inter-component relationships into the model, enabling it to possess system-level reasoning capabilities.

[0032] Step S140: Based on the first feature parameters and the adjacency matrix training mechanism of the enhanced graph neural network, an inversion model is obtained.

[0033] The inversion model can accurately invert key parameters that cannot be directly measured or are too costly to measure, such as internal thermal stress of the rotor, stress at the root of the moving blade, and bearing oil film pressure, by using features extracted from the data (first feature parameter) and the physical structure of the system (adjacency matrix A).

[0034] This application employs a mechanism-enhanced graph neural network as the core inversion model, embedding physical laws as prior knowledge into the model. On the GNN architecture, graph convolutional networks (GCN), graph attention networks (GAT), or more complex spatiotemporal graph neural networks (STGCN) can be selected according to task requirements to effectively capture the spatial dependencies and temporal dynamics between sensor data.

[0035] Please see Figure 2 , Figure 2 This is a flowchart illustrating the training of an inversion model according to an embodiment of this application. Figure 2 In this context, the step of obtaining the inversion model by training an enhanced graph neural network based on the first feature parameters and the adjacency matrix includes: Step S210: Define a graph structure using the adjacency matrix A, select a graph neural network as the basic architecture, and add a constraint term based on a physical mechanism model to the loss function of the graph neural network to construct a mechanism-enhanced graph neural network model. Step S220: Using the first feature parameters and the label data corresponding to the first feature parameters, train the mechanism-enhanced graph neural network model to obtain the trained inversion model.

[0036] Specifically, in the model building phase, the system defines the topological connections of each component of the steam turbine using an adjacency matrix A, and selects a graph convolutional network as the basic architecture. Based on the standard classification or regression loss, physical constraint terms based on the rotor vibration differential equation are added to jointly constitute a mechanism-enhanced loss function, thereby constructing a mechanism-enhanced graph neural network model. In the model training phase, the first feature parameters (such as the vibration spectrum characteristics of each bearing housing) are used as node feature inputs to the graph structure, and the corresponding label data are used as supervision signals to iteratively update the model parameters. After training, the resulting inversion model can accurately infer the state or potential faults of key components inside the steam turbine based on real-time monitoring feature parameters.

[0037] Please see Figure 3 , Figure 3 This is a flowchart illustrating the construction of an adjacency matrix according to an embodiment of this application. Figure 3 In the process, constructing the adjacency matrix A includes: Step S310: Define the key components of the steam turbine system as graph nodes; Specifically, key components of the steam turbine system are defined as graphical nodes. Key components may include sensor measuring points, bearings, rotor shaft sections, critical steam seals, etc. For example, graphical nodes include: (1) 12 sensor nodes: X and Y direction measuring points of bearings #1-#6.

[0038] (2) Six bearing nodes: the center of bearings #1-#6.

[0039] (3) Six rotor shaft segment nodes: representing the main shaft segments of the high, medium and low pressure cylinder rotors.

[0040] (4) Six key steam seal nodes: steam seals at key locations such as the main steam valve of the high-pressure cylinder, the intermediate-pressure cylinder, and the low-pressure cylinder.

[0041] (5) Two other nodes: thrust bearing node and coupling node.

[0042] Step S320: Determine the relationships between the graph nodes, including a first association with a direct physical connection, a second association without a direct physical connection but with mutual influence, and a third association without a direct physical connection and without mutual influence; and assign weight values ​​to the first association, the second association, and the third association respectively. Specifically, the connection relationships of the edges are determined based on the equipment structural parameters. The turbine structural drawings are analyzed to determine whether there are direct physical connections between the nodes. If node i and node j are physically directly connected (e.g., the sensor node of bearing #3 is connected to its corresponding bearing #3 node; the bearing #3 node is also connected to its left and right adjacent rotor shaft segment nodes), then in the adjacency matrix A...ij =A ji =1. If there is no direct connection, the initial value A is 1. ij =0. This forms the basic binary skeleton of the adjacency matrix. Next, based on dynamic analysis and expert knowledge, the edge weights are quantified, and weights are assigned to edges between nodes that do not have direct physical connections and do not influence each other. The assignment of edge weights can be achieved through dynamic analysis and expert knowledge.

[0043] Dynamics analysis is performed using a rotor dynamics simulation model to calculate the cross-influence coefficient. For example, a unit excitation is applied at bearing #j, and the vibration response amplitude R at bearing #i is measured. ij Then the weight of the edge can be set as A. ij =f(R ij ), where f is a normalization function that maps the response magnitude to the interval (0,1).

[0044] For influences that are difficult to quantify precisely through models (such as heat conduction paths), domain experts assign values ​​based on their experience.

[0045] Step S330: Construct an adjacency moment using the weight values ​​corresponding to the first association relationship, the second association relationship, and the third association relationship as matrix elements.

[0046] The weight values ​​corresponding to the first, second, and third associations are used as matrix elements to fill matrix A, forming the final adjacency matrix. The adjacency matrix usually needs to be normalized to facilitate subsequent training of the graph neural network.

[0047] This application abstracts the complex turbine rotor-bearing-seal system into a graphical model G=(V, E). Nodes V represent key entities in the system, such as sensor measurement points, bearing centers, rotor shaft centroids, and the locations of critical steam and oil seals. Each node carries its own attributes; for example, sensor nodes carry extracted features, and component nodes carry design parameters. Edges E represent physical connections, force transmission paths, or mutual influence relationships between nodes, such as the oil film force exerted by the bearing on the journal, the thrust of steam on the shaft seal, and the mechanical transmission between adjacent shaft segments. The adjacency matrix of the graph is constructed based on equipment structural drawings, rotor dynamics analysis reports, and expert knowledge, quantifying the interaction strength between components. This graphical modeling approach allows the model to understand data from a global perspective of the system topology, rather than treating sensors as isolated information sources.

[0048] In one embodiment, the mechanism-enhanced graph neural network includes: a graph convolutional layer, a graph attention layer, a temporal modeling layer, and a physical constraint layer; The graph convolutional layer is used to perform spatial feature aggregation operations on the nodes in the graph structure to obtain preliminary features of each node and its neighboring nodes. The graph convolutional layer implements the node spatial feature aggregation operation. Based on the topological connection relationship defined by the adjacency matrix A, it performs convolution operations on the features of each node and its first-order neighboring nodes in the graph structure, thereby obtaining preliminary node features that incorporate local structural information.

[0049] The graph attention layer is used to assign differentiated weights to different neighbor nodes through an attention mechanism, generating weighted node features. The graph attention layer is a feature weighting unit that introduces an attention mechanism. It calculates the attention coefficients between a node and its neighbor nodes, assigns differentiated aggregation weights to different neighbor nodes, and thus generates weighted node features that better reflect the strength of the association.

[0050] The temporal modeling layer is used to process the time-series data of the weighted node features through a temporal neural network based on the weighted node features, capturing the dynamic evolution of parameters in the time dimension. The temporal modeling layer is a recurrent neural network unit that processes time-series data. Its input is the weighted feature sequence of each node at different time steps, and its output is the temporal features that capture the dynamic evolution of parameters in the time dimension.

[0051] The physical constraint layer is used during model training to calculate constraint terms for the physical mechanism model based on the output of the temporal modeling layer, and adds these constraint terms to the loss function. The physical constraint layer is a dedicated layer that applies physical constraints during model training. This layer itself does not participate in forward propagation; instead, during training, it calculates the deviation between the output of the temporal modeling layer and the preset physical mechanism model (such as the turbine rotor dynamics equations or vibration transmission models), and adds this deviation as a physical constraint term to the overall loss function of the model, thereby guiding the model's training direction to conform to physical laws.

[0052] In one embodiment, the constraint includes at least one of the following: fluid dynamics constraint based on Reynolds equations, mechanical equilibrium constraint based on rotor dynamics equations, and thermal expansion constraint based on thermodynamic equations.

[0053] Fluid dynamic constraints are used to ensure that the inverted oil film pressure distribution conforms to the laws of fluid dynamics;

[0054] In the formula p For oil film pressure, h For oil film thickness, The dynamic viscosity of the lubricating oil. U The speed is the surface velocity of the rotor.

[0055] Mechanical equilibrium constraints are used to ensure that the inverted oil film force and shaft vibration characteristics conform to dynamic equilibrium; Rotor dynamic equations: In the formula M , C , K These are the mass, damping, and stiffness matrices, respectively. F oil For oil film force, F unbalance This is an unbalanced excitation force.

[0056] Thermodynamic constraints are used to ensure that the inverted shaft seal clearance conforms to the laws of thermal expansion. The constraints are based on a simplified formula for linear thermal expansion: In the formula This is the change in length (related to axial displacement). a The coefficient of thermal expansion of the material. L 0 represents the original length. This represents the change in temperature.

[0057] In one embodiment, the first monitoring parameter includes: simulation data, historical operating data, and measured data; The simulation data is generated through parametric simulation based on a detailed physical model of the steam turbine constructed using finite element analysis. This simulation data can cover various operating conditions (such as different speeds, loads, and temperatures). This simulation data can be generated on a large scale by adjusting model parameters, covering a variety of scenarios including extreme operating conditions and rare faults, providing rich prior knowledge for the model. The detailed structural parameters of the steam turbine include: ① Geometric parameters: rotor diameter, shaft length, bearing span, and design values ​​for steam seal clearance; ② Material properties: elastic modulus, Poisson's ratio, and density; ③ Mass distribution: concentrated mass of each shaft segment and moment of inertia; ④ Lubrication parameters: lubricating oil viscosity-temperature characteristics and design values ​​for oil supply pressure / temperature.

[0058] Specifically, a detailed three-dimensional finite element model of the turbine was established using ANSYS software. This model includes the geometry, material, and mass distribution of the rotor system, the oil film dynamics model of the bearings (solved based on the Reynolds equations), and the fluid-structure interaction model of the steam seals. Through parametric simulation of the model, the operating conditions of the unit under different loads (30%–100%), different lubricating oil supply temperatures (38–48°C), and different degrees of steam seal clearance wear (0%–50%) were simulated, generating a large simulation dataset containing hundreds of operating points. Each operating point includes corresponding TSI monitoring data (generated through simulation) and actual internal parameters (such as minimum oil film thickness and shaft seal clearance), serving as "true value" labels for model training.

[0059] Historical operating data refers to the historical data of a steam turbine under different operating conditions. Historical operating data is multi-channel monitoring data that has been collected and recorded by a sensor array over a long period of time during the actual operation of the steam turbine. It truly reflects the operating status and degradation process of the equipment in the real environment.

[0060] Actual data refers to the measured data obtained by fixed-point calibration under preset operating conditions. Measured data is the data obtained by fixed-point calibration of the steam turbine under preset specific operating conditions using high-precision measuring instruments. For example, vibration and alignment data obtained by offline diagnostic equipment during maintenance. This data has extremely high accuracy and authority and is often used as a benchmark for model verification and calibration.

[0061] In the process of training the mechanism-enhanced graph neural network, a phased training strategy is adopted. First, simulation data is used for pre-training, and then historical running data and measured data are used for fine-tuning.

[0062] First, the model is pre-trained using large-scale, multi-scenario simulation data. This allows the model to initially learn the spatial topological relationships of various turbine components and their basic dynamic response laws under a wide range of operating conditions, laying a physically reasonable foundation for initial parameters. Subsequently, a relatively limited but more realistic amount of historical operating data and measured data is used to fine-tune the pre-trained model. This stage enables the model to adapt to the characteristics of real systems (such as manufacturing tolerances, installation errors, and environmental noise), thereby refining and adapting the universal laws learned in the simulation environment to applications specific to actual turbine units.

[0063] The model training adopts a supervised learning approach, utilizing historical running data, simulation data, and a small amount of actual data to optimize model parameters through the backpropagation algorithm.

[0064] In one embodiment, the inversion model can be constructed using a spatiotemporal graph convolutional network (STGCN). This model consists of three stacked spatiotemporal convolutional blocks, each containing a graph convolutional layer (using the GAT mechanism), a temporal convolutional layer (using a gated TCN), and a layer normalization layer. The model input is the feature sequence and adjacency matrix from the past 60 seconds, and the output is the predicted value of the internal key parameters at the current time step.

[0065] Physical constraints are introduced into the loss function of the inversion model. L_total The definition is as follows: L_ total=L_data +α×L_reynolds+β×L_dynamics, α and β are weighting coefficients used to balance the importance of data-driven terms and physical constraint terms. Their values ​​can be determined experimentally or by using an adaptive weighting strategy.

[0066] in: L_dataIt is the supervised learning loss, which is the mean square error (MSE) between the model's predicted value and the simulation / calibration data label.

[0067] L_reynolds The physical residual is the Reynolds equation residual. It is calculated by substituting the oil film pressure distribution derived from the model (mapped from the oil film thickness via a small fully connected network) into the simplified two-dimensional incompressible Reynolds equation, and then calculating the imbalance on both sides of the equation as the physical residual. This residual term penalizes predictions that do not conform to the laws of hydrodynamics.

[0068] L_dynamics The rotor dynamics residuals are obtained by inverting the oil film force (stiffness and damping coefficient multiplied by displacement and velocity) from the model as the excitation and substituting it into the simplified rotor dynamics equations. M +C +Kx=F_oil+F_unbalance The calculated vibration response differs from the TSI-measured vibration. This residual term ensures that the inverted oil film characteristics can reasonably explain the observed vibration phenomena.

[0069] Of course, loss function L_total Thermodynamic constraints can also be considered, i.e. L_total = L_data + α × L_ reynolds + β × L_dynamics + β 1 ×L_thermodynamics .

[0070] The training of the inversion model includes a pre-training phase, a fine-tuning phase, and a calibration and validation phase. In the pre-training phase of the inversion model: the STGCN model is pre-trained on a large scale using simulation data generated from operating points. This phase aims to enable the model to learn the basic and universal physical mapping relationships of the steam turbine system.

[0071] In the fine-tuning phase: The pre-trained model is fine-tuned using the unit's historical operating data from the past year. Since the historical operating data does not have real internal parameter labels, the weights (α, β) of the physical residual terms are increased in this phase, mainly relying on physical constraints to guide the model to adapt to the characteristics of the real unit.

[0072] During the calibration and verification phase: During the most recent unit overhaul, a feeler gauge was used to accurately measure the steam seal clearance in a cold state. During unit startup, the jacking oil film height data was recorded under turning gear conditions. These sparse but high-quality measured data were used for final model verification and calibration. The verification results showed that the model's prediction error for the cold-state steam seal clearance was less than 5%, and its prediction error for the jacking oil film height was less than 8%.

[0073] In summary, the key parameter inversion model training method for steam turbines proposed in this application has the following significant advantages compared to existing technologies: 1. Achieved online "soft measurement" of key internal state parameters, filling a gap in industry monitoring: This application, for the first time, realizes real-time, online, and undisturbed monitoring of a series of key internal state parameters that were previously impossible to obtain online, such as turbine oil film characteristics (minimum / average oil film thickness, peak pressure), dynamic clearance of each stage of shaft seals, and top shaft oil film height. This allows operation and maintenance personnel to see the true operating status of the unit for the first time, solving the long-standing monitoring "black box" problem that has plagued the industry, and providing an unprecedented data foundation for the refined management of turbines.

[0074] 2. Significantly improved accuracy and reliability of parameter estimation: Compared with traditional methods based on theoretical estimation of design values ​​or offline empirical formula inference, this application achieves an order-of-magnitude improvement in parameter estimation accuracy by fusing multi-source real-time data and utilizing a mechanism-enhanced GNN model for deep mining. The embedding of physical mechanisms ensures that the inversion results are not only data-fitting but also physically reasonable and reliable, effectively avoiding the spurious correlations and insufficient generalization capabilities that may arise from purely data-driven models, thus greatly enhancing the reliability of monitoring results.

[0075] 3. It provides direct quantitative basis for the refined operation optimization of steam turbines: By monitoring the oil film status and shaft seal clearance in real time, operators can optimize control parameters such as lubricating oil temperature and pressure based on precise internal parameter feedback, or adjust the start-up and shutdown curves and load change rates of the unit, so that the unit always operates in the most efficient and safest state range, thereby effectively achieving energy saving and consumption reduction and improving the economic efficiency of the unit.

[0076] 4. It has driven a revolutionary shift in maintenance models towards "condition-based maintenance": This application makes condition-based maintenance (CBM) based on the actual health condition of the equipment possible. By conducting long-term monitoring and prediction of factors such as shaft seal clearance wear rate and oil film degradation trend, maintenance plans can be accurately formulated, which not only avoids unplanned downtime caused by sudden failures, but also maximizes the utilization of the remaining life of components, significantly reduces maintenance costs and downtime losses, and comprehensively improves the level of equipment asset management.

[0077] 5. Provides revolutionary early fault warning capabilities: Many major faults (such as oil film instability and dynamic-static rubbing) will show subtle but trending abnormal changes in related internal parameters before they occur. This system can capture these physical precursors, which are more fundamental and earlier than traditional vibration amplitude alarms, advancing the fault warning window by hours or even days, gaining valuable time for preventative handling, and thus greatly improving the operational safety of the unit.

[0078] Please see Figure 4 , Figure 4This is a schematic block diagram of a key parameter inversion model training device for a steam turbine according to an embodiment of this application. Figure 4 In this context, the training device for the key parameter inversion model of the steam turbine includes: The data acquisition module 410 is used to acquire the monitoring parameters of the steam turbine; Feature extraction module 420 is used to extract features from the monitoring parameters to obtain a feature parameter set; The matrix construction module 430 is used to construct an adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) is composed of the key components of the steam turbine, and the edge set E is composed of the physical connections or mutual influence relationships between the key components. The model training module 440 is used to enhance the graph neural network based on the feature set and the adjacency matrix training mechanism to obtain the inversion model.

[0079] It should be noted that the turbine key parameter inversion device and the turbine key parameter inversion method provided in the above embodiments belong to the same concept. The specific operation methods of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the turbine key parameter inversion device provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0080] Embodiments of this application also provide an inversion device for key parameters of a steam turbine, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by one or more processors, the memory implements the inversion method for key parameters of a steam turbine as described in the above embodiments.

[0081] Embodiments of this application also provide one or more machine-readable media having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the inversion method for key parameters of the steam turbine in the above embodiments.

[0082] Please see Figure 5 , Figure 5 This is a flowchart illustrating a method for inverting key parameters of a steam turbine according to an embodiment of this application. Figure 5 A method for inverting key parameters of a steam turbine includes: Step S510: Obtain the second monitoring parameters of the steam turbine; Specifically, vibration sensors, temperature sensors, and pressure sensors installed in key components such as the rotor, bearings, and cylinders are used to collect real-time monitoring parameters of the steam turbine to ensure the real-time nature and synchronization of the data. Step S520: Extract features from the second monitoring parameter to obtain the second feature parameter; Feature extraction is performed on the second monitoring parameter using a feature extraction algorithm consistent with the model training phase (such as time-frequency analysis, eigenvalue calculation, etc.). The original vibration waveform, temperature sequence, etc. are converted into second feature parameters that can effectively characterize the state of the component, such as vibration energy in a specific frequency band, axis trajectory characteristics, temperature gradient, etc.

[0083] Specifically, real-time data is continuously received and processed. To capture temporal dynamics, the model (such as STGCN in the example) requires an input containing information from a past period. The system maintains a fixed-length (e.g., 60 seconds, corresponding to 60 time steps) sliding window of feature vectors. At each new time point t, the latest feature vector is added to the window, while the oldest is removed.

[0084] Step S530: Based on the second feature parameters, obtain the inversion result through the inversion model.

[0085] Based on the second characteristic parameter, the inversion result is obtained through the inversion model, including: by receiving the second characteristic parameter as input, and using the nonlinear mapping and relational reasoning capabilities of the inversion model, the inversion result of key internal parameters that cannot be directly measured or are difficult to obtain in real time is automatically calculated, such as the bearing oil film thickness, oil film pressure distribution, oil film stiffness and damping coefficient, and dynamic operating clearance of each stage of shaft seal of the steam turbine.

[0086] Specifically, at time point t, the feature vector sequence constructed in step S520 (a tensor with dimensions [N, T, F], where N is the number of nodes, T is the number of time steps such as 60, and F is the feature dimension) and the pre-constructed adjacency matrix A are used together as the input to the model.

[0087] The tensor [N, T, F] is fed into a trained and deployed mechanism-enhanced graph neural network model (Model) for a forward computation to obtain the inversion result. Calculation formula: In the formula This is the predicted result vector output by the model at the current time t. Each element of this vector corresponds to an internal parameter that needs to be inverted. For including from t T+1 arrive t The input tensor is the feature sequence of N nodes at time t, and A is the adjacency matrix representing the system topology.

[0088] In one embodiment, confidence intervals for the inversion parameters are calculated in real time. These confidence intervals are used to quantify the uncertainty of the inversion results and provide support for operational decisions. Specifically, the Monte Carlo method can be used to estimate the confidence intervals, including: Gaussian noise (σ=0.01) with a preset distribution and intensity is added to the input feature parameters to generate multiple sets of perturbation samples, such as 100 sets. These perturbation samples are then input into the inversion model to obtain corresponding inversion results. Based on these inversion results, the statistical characteristics of the inversion results are calculated, including the mean (μ) and standard deviation (σ). The confidence interval μ±2σ (95% confidence level) of the inversion parameters is determined based on these statistical characteristics. If the confidence level of the inversion result is within the confidence interval, the inversion result is considered reliable.

[0089] In one embodiment, different dedicated models are used for different operating conditions (such as full-load stable operation, low-load operation, and start-up / shutdown processes) to improve inversion accuracy. This application adopts a "model routing" or "hybrid expert model" strategy, including: acquiring key operating parameters of the steam turbine in real time; identifying the current macroscopic operating state of the steam turbine based on the key operating parameters; selecting a target dedicated inversion model corresponding to the current macroscopic operating state from multiple pre-trained dedicated inversion models according to preset model switching rules; wherein the multiple pre-trained dedicated inversion models are trained for different typical operating conditions; and obtaining the inversion results of key parameters through the target dedicated inversion model based on the current monitoring parameters of the steam turbine.

[0090] Specifically, the macroscopic operating status of the unit is first determined in real time based on key operating parameters (such as speed and active power) collected by the DCS, and the data stream is then routed to the most suitable dedicated model based on the macroscopic operating status of the unit; as follows: When 5 rpm < engine speed < 2950 rpm, it is determined to be a start-stop condition, and the start-stop dedicated model is called; when engine speed < 5 rpm, it is determined to be a turning gear condition, and no analysis is performed in this case.

[0091] The start-stop dedicated model (Model_LowLoad) is trained using historical start-stop process data from the unit. It focuses more on analyzing time-frequency domain characteristics (such as STFT and wavelet analysis results) to accurately capture transient phenomena such as crossing critical speeds.

[0092] When the engine speed is greater than 2950 RPM and the load is less than 240 MW (40% of the rated load), the condition is determined to be a low-load condition. In this case, the low-load dedicated model is invoked. Low-load dedicated model: Fine-tuned using historical low-load operating data, it can better learn the patterns of insufficient thermal deformation and differences in system dynamic characteristics under low load compared to high load.

[0093] When the engine speed is greater than 2950 RPM and the load is greater than or equal to 240 MW, the condition is determined to be a high-load or full-load stable operating condition. In this case, the high-load stable operation model (Model_Stable) is invoked. High-load stable operation model: The main model of the system, which is trained using a large amount of stable operation data and has the highest accuracy in parameter inversion under steady-state conditions.

[0094] In this way, the system can ensure that the "domain expert" model is used for calculation under any operating condition, thereby improving the overall inversion accuracy and robustness.

[0095] In one embodiment, the inversion method further includes: visualizing the inversion results. The visualization methods include: A three-dimensional dynamic cloud map is used to display the three-dimensional distribution of oil film thickness and pressure inside the bearing; Trend curve (see appendix) Figure 6 This is used to display the changes in key shaft seal clearance and minimum oil film thickness over time, load, speed, etc. Users can select one or more internal parameters to view their trend curves over any past time period (such as 24 hours or 30 days) as a function of load, speed, and other operating conditions.

[0096] The 3D dashboard is used to quantitatively display the current values, safety thresholds, and alarm status of various key parameters. On the 3D model of the steam turbine, the minimum oil film thickness and maximum oil film pressure of each bearing, as well as the dynamic clearance of each key steam seal, are displayed in real time, and their health status is represented by colors (green-yellow-red).

[0097] This application deploys a pre-trained inversion model, after lightweight processing, to a high-performance computing platform or edge computing unit bypassing the TSI system. After deployment, the model receives online data streams in real time, performing parameter inversion calculations at the second or sub-second level. The inversion results are continuously verified for model accuracy by comparing them with detailed offline simulation results or sparse direct measurement data. To transform abstract numbers into intuitive insights, the system visually displays the inversion results on a 3D visualization platform, such as generating distribution cloud maps of oil film thickness and pressure, dynamic change curves of shaft seal clearance, and status dashboards for key parameters, providing clear and easily understandable decision support information for operation and maintenance personnel.

[0098] The following describes a specific embodiment of the method for inverting key parameters of a steam turbine. One day, while operating stably at a load of 550MW, the DCS system showed that the horizontal vibration value of bearing #2 slowly increased from 35μm to 48μm within 12 hours. Although it did not reach the alarm value, the trend was noteworthy.

[0099] 1. Data preprocessing and feature extraction: This system processes the rotor vibration data at bearing #2 in real time. Taking a data point one hour before the fault warning as an example, the processing process is as follows: (1) Collect the raw X and Y vibration data for 1 second (4096 points each). (2) After bandpass filtering and outlier removal, feature extraction is performed: ① Time domain features: The peak-to-peak value of X is calculated to be 47.5 μm, and that of Y is 38.2 μm. The major axis of the shaft center trajectory is 47.8 μm, the minor axis is 10.1 μm, and the tilt angle is 85 degrees. ② Frequency domain features: FFT analysis shows that the amplitude of the power frequency (1X) component is 45 μm, accounting for 92% of the total vibration energy, and there are no obvious high-order harmonics or subharmonics. The above features, together with other operating parameters (such as load 550MW, oil supply temperature of bearing #2 43℃, etc.), constitute the input vector at this moment.

[0100] 2. Intelligent Inversion and Result Display: The input vector is fed into the pre-trained STGCN model, and the inverted internal parameters of bearing #2 are immediately updated on the visualization interface. The operator retrieves the historical trend curve of the minimum oil film thickness (hmin) of bearing #2 (see attached). Figure 6 Data shows that, in sync with the slow increase in vibration, the inversion value of hmin has shown a continuous downward trend over the past 24 hours, from 210 μm 24 hours ago to the current 145 μm, which is close to the warning value of the design limit (see Table 1).

[0101] Table 1

[0102] Simultaneously, the peak oil film pressure retrieved by the system showed an abnormally high increase. Based on these anomalies in internal parameters, the system automatically generated a warning: "Bearing #2 may have poor lubrication or alignment issues; it is recommended to check the lubrication system or monitor the unit's thermal deformation." Operators immediately intensified monitoring of the return oil temperature of bearing #2 and inspected the lubrication filter, finding slight signs of blockage. After timely cleaning of the filter, the minimum oil film thickness retrieved for bearing #2 returned to normal levels, and the vibration stabilized, successfully preventing an unplanned shutdown that could have caused bearing wear.

[0103] As can be seen from the above embodiments, this application, through systematic data processing, innovative graph structure modeling, core mechanism-enhanced GNN model, and user-friendly visualization, has successfully transformed a complex industrial problem into an implementable, verifiable, and valuable engineering solution, taking a key step toward realizing intelligent operation and maintenance of steam turbines.

[0104] Please see Figure 7 , Figure 7 This is a flowchart illustrating a steam turbine maintenance method according to an embodiment of this application. Figure 6 A method for maintaining a steam turbine includes: Step S710: Obtain the third monitoring parameter of the steam turbine; Step S720: Extract features from the third monitoring parameter to obtain the third feature parameter; Step S730: Based on the third feature parameter, obtain the inversion result through the inversion model; Step S740: Identify abnormal parameters based on the inversion parameters and preset identification rules; Step S750: Determine the anomaly source based on the anomaly parameters and the adjacency matrix A; Step S760: For the source of the anomaly, generate an operation and maintenance plan to maintain the steam turbine based on the operation and maintenance plan.

[0105] Specifically, when the system performs diagnostic tasks, it first acquires the third monitoring parameter and extracts the third feature parameter, then uses an inversion model to obtain the inversion results of internal key parameters. Based on preset rules, such as comparing the inverted parameters with threshold ranges, historical baselines, or dynamic model predictions, it automatically identifies abnormal parameters that deviate from the normal range. Using an adjacency matrix A representing component relationships, and employing graph traversal algorithms or influence propagation models, it analyzes the propagation path of abnormal parameters in the graph structure to locate the most likely source component of the anomaly. Finally, based on the anomaly source characteristics (type, severity, and centrality in the graph structure) combined with a maintenance knowledge base, it automatically generates an operation and maintenance plan containing maintenance content, priorities, and time recommendations.

[0106] In one embodiment, the identification of abnormal parameters can be achieved in the following way: Construct a health status index system for steam turbines, including oil film thickness margin index, shaft seal clearance degradation index, and vibration stability index; ①Indicator Definition: Oil Film Thickness Margin Index: Shaft seal clearance degradation index: Vibration stability index: .

[0107] ②Overall score: , This represents the weighting coefficient, which is set by experts.

[0108] (3) When an abnormal trend occurs (the comprehensive score does not meet the preset rules), root cause analysis is performed using the inverted internal parameters to determine the type and location of the fault; the specific steps of root cause analysis are as follows: ① Trend comparison: Compare the inversion parameters (e.g.) h min ① Historical normal values ​​and current values; ② Rule matching: If h min A decrease accompanied by an increase in oil film pressure indicates poor lubrication or misalignment; if Sudden increase → abnormal wear or thermal deformation of shaft seal; ③ Fault location: locate the source of the anomaly by combining the node correlation in the adjacency matrix.

[0109] (4) Generate optimized maintenance plans and operation suggestions based on the prediction results and root cause analysis to achieve condition-based maintenance.

[0110] Figure 8 A schematic diagram of a computer system suitable for implementing the memory of embodiments of this application is shown. It should be noted that... Figure 8 The computer system with the memory shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0111] like Figure 8 As shown, the computer system 800 includes a Central Processing Unit (CPU) 801, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on a program stored in Read-Only Memory (ROM) 802 or a program loaded from storage into Random Access Memory (RAM) 803. The RAM also stores various programs and data required for system operation. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An Input / Output (I / O) interface 805 is also connected to the bus 804.

[0112] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.

[0113] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing a computer program for performing an inversion method for key parameters of the steam turbine of the aforementioned embodiments. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium 811. When the computer program is executed by a central processing unit (CPU) 801, it performs various functions defined in the system of this application.

[0114] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM) 803, read-only memory (ROM) 802, erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0115] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block or combination of blocks in a block diagram or flowchart may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0116] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0117] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a computer's processor, causes the computer to perform the aforementioned method for inverting key parameters of a steam turbine. This computer-readable storage medium may be included in the memory described in the above embodiments, or it may exist independently and not incorporated into that memory.

[0118] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the inversion method for key parameters of a steam turbine provided in the various embodiments described above.

[0119] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for training a key parameter inversion model for a steam turbine, characterized in that, The inversion model training method includes: Obtain the first monitoring parameters of the steam turbine; The first feature parameter is obtained by extracting features from the first monitoring parameter; Construct an adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) consists of the key components of the steam turbine, and the edge set E consists of the physical connections or mutual influence relationships between the key components; An inversion model is obtained by enhancing the graph neural network based on the first feature parameters and the adjacency matrix training mechanism.

2. The method for training a key parameter inversion model of a steam turbine according to claim 1, characterized in that, The first feature parameter includes: Time-domain characteristics, including axis trajectory parameters, vibration amplitude, and DC bias; Frequency domain characteristics include the frequency component amplitude and phase, and the full-spectrum energy distribution extracted by performing a fast Fourier transform on the monitoring parameters. Time-frequency domain features, including time-frequency features obtained through wavelet transform or short-time Fourier transform; Correlation features include the correlation between axial displacement variation patterns and thermal expansion.

3. The method for training a key parameter inversion model of a steam turbine according to claim 1, characterized in that, The step of enhancing the graph neural network based on the first feature parameters and the adjacency matrix training mechanism to obtain the inversion model includes: A graph structure is defined using the adjacency matrix A. A graph neural network is selected as the basic architecture, and a constraint term based on a physical mechanism model is added to the loss function of the graph neural network to construct a mechanism-enhanced graph neural network model. The mechanism-enhanced graph neural network model is trained using the first feature parameters and the corresponding label data to obtain a trained inversion model.

4. The method for training a key parameter inversion model of a steam turbine according to claim 1, characterized in that, The construction of the adjacency matrix A includes: The key components of the steam turbine system are defined as graph nodes; Determine the relationships between the nodes in the graph, including the first association with direct physical connection, the second association without direct physical connection but with mutual influence, and the third association without direct physical connection and without mutual influence. Assign weight values ​​to the first association, the second association, and the third association, respectively; Adjacency moments are constructed using the weight values ​​corresponding to the first association, the second association, and the third association as matrix elements.

5. The method for training a key parameter inversion model of a steam turbine according to claim 1, characterized in that, The mechanism-enhanced graph neural network includes: Graph convolutional layers are used to perform spatial feature aggregation operations on nodes in the graph structure to obtain preliminary features of each node and its adjacent nodes. The graph attention layer is used to assign differentiated weights to different neighbor nodes through an attention mechanism, generating weighted node features. The temporal modeling layer is used to process the time series data of the weighted node features through a temporal neural network based on the weighted node features, and to capture the dynamic evolution of parameters in the time dimension. The physical constraint layer is used to calculate the constraint terms of the physical mechanism model based on the output of the temporal modeling layer during model training, and to add the constraint terms to the loss function.

6. The method for training a key parameter inversion model of a steam turbine according to claim 5, characterized in that, The constraints include at least one of the following: Fluid dynamic constraints based on Reynolds equations; Mechanical equilibrium constraints based on rotor dynamics equations; Thermal expansion constraints based on thermodynamic equations.

7. The method for training a key parameter inversion model of a steam turbine according to claim 1, characterized in that, The first monitoring parameter includes: Based on the detailed physical model of the steam turbine constructed through finite element analysis, parametric simulation was performed to generate simulation data covering different operating conditions; Historical operating data of steam turbines under different operating conditions; The measured data obtained by fixed-point calibration under preset working conditions; In the process of training the mechanism-enhanced graph neural network, a phased training strategy is adopted. First, simulation data is used for pre-training, and then historical running data and measured data are used for fine-tuning.

8. A training device for a key parameter inversion model of a steam turbine, characterized in that, The inversion model training device includes: The data acquisition module is used to acquire the monitoring parameters of the steam turbine; The feature extraction module is used to extract features from the monitoring parameters to obtain a feature parameter set; The matrix construction module is used to construct the adjacency matrix A, which is a mathematical representation of the edge set E in the graph structure G=(V,E); wherein, the graph node set V in the graph structure G=(V,E) consists of the key components of the steam turbine, and the edge set E consists of the physical connections or mutual influence relationships between the key components; The model training module is used to enhance the graph neural network based on the feature set and the adjacency matrix training mechanism to obtain the inversion model.

9. A method for inverting key parameters of a steam turbine, characterized in that, The inversion method includes: Obtain the second monitoring parameters of the steam turbine; The second monitoring parameter is subjected to feature extraction to obtain the second feature parameter; Based on the second characteristic parameter, the inversion result is obtained using the inversion model described in any one of claims 1-7.

10. A method for maintaining a steam turbine, characterized in that, The maintenance method includes: Obtain the third monitoring parameter of the steam turbine; The third monitoring parameter is subjected to feature extraction to obtain the third feature parameter; Based on the third characteristic parameter, the inversion result is obtained using the inversion model described in any one of claims 1-7; Identify abnormal parameters based on the inversion parameters and preset identification rules; The anomaly source is determined based on the aforementioned anomaly parameters and the adjacency matrix A; For the source of the anomaly, an operation and maintenance plan is generated to maintain the steam turbine based on the operation and maintenance plan.

Images