Turbine blade load prediction model training method and training device and turbine blade load prediction method and prediction device

Abstract

The invention discloses a training method and device of a turbine blade load prediction model and a turbine blade load prediction method and device, and belongs to the technical field of turbine blade health monitoring. The training method comprises the steps that multiple sets of historical target operation parameters of a gas turbine are obtained, and simulation calculation is conducted to obtain a full-field load cloud picture; extracting the load of the stress concentration area and constructing a target output vector; utilizing a random Fourier feature mapping function to convert a first input vector formed by each group of historical target operation parameters into a low-dimensional random feature vector; and under a linear ridge regression framework, taking the low-dimensional random feature vector as input, taking the target output vector as a supervision signal to carry out joint training, and updating internal parameters of the random Fourier feature mapping function and a weight vector of a linear ridge regression model. According to the invention, the problem that the real-time life evaluation cannot be carried out based on the load because the accurate and rapid prediction of the load cannot be realized in the prior art is solved.

Description

Technical Field

[0001] This invention relates to a training method and device for a turbine blade load prediction model, and a turbine blade load prediction method and device, belonging to the field of turbine blade health monitoring technology. Background Technology

[0002] As a core power unit in power generation and aerospace, the operational reliability and lifespan of gas turbines directly affect the safety and economy of the entire system. Turbine blades, as key hot-end components of gas turbines, operate under extreme conditions of high temperature, high pressure, and high speed, enduring the combined effects of complex aerodynamic, thermal, and centrifugal loads. Under these conditions, blade materials face two main failure mechanisms: first, creep deformation occurs under sustained high temperatures; second, cyclical thermo-mechanical stresses during unit start-up, shutdown, and load changes lead to fatigue damage. Both creep and fatigue failure can cause blade fracture and even catastrophic accidents involving the entire system, resulting in incalculable economic losses and significant safety risks. Therefore, achieving accurate monitoring and prediction of key parameters such as stress, strain, and temperature of turbine blades is essential for ensuring the safe, reliable, and long-term operation of gas turbines, and it remains a long-standing technical challenge in this field.

[0003] In existing technologies, machine learning algorithms have become a common method for condition monitoring. However, these methods are mostly applied to fault diagnosis and classification based on signals such as vibration and acoustics, for example, identifying whether blades have pre-existing defects such as cracks, foreign object marks, or assembly deviations. This technical approach is essentially a "post-event" monitoring mode that identifies faults that have already occurred. Its core deficiency lies in its failure to address the fundamental problem of blade health management: existing technologies cannot achieve accurate and rapid prediction of loads, resulting in the inability to perform real-time life assessments based on loads. This leads to significant deficiencies in predictive maintenance and proactive safety control, failing to meet the current higher demands for precise condition management and intelligent control of gas turbines. Summary of the Invention

[0004] The purpose of this invention is to provide a training method and device for a turbine blade load prediction model, as well as a load prediction method and device. By jointly training with stochastic Fourier feature mapping and linear ridge regression, this invention addresses the problem that existing technologies cannot achieve accurate and rapid load prediction, thus preventing real-time life assessment based on load.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0006] In a first aspect, the present invention provides a training method for a turbine blade load prediction model, comprising:

[0007] Multiple sets of historical target operating parameters of the gas turbine are obtained, and simulation calculations are performed on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters;

[0008] Extract the loads in the stress concentration region of the full-field load cloud map corresponding to each set of historical target operating parameters, and construct the target output vector corresponding to each set of historical target operating parameters;

[0009] The first input vector, composed of each set of historical target running parameters, is converted into a low-dimensional random feature vector using a predefined random Fourier feature mapping function.

[0010] Within the linear ridge regression framework, the low-dimensional random feature vectors corresponding to each set of historical target running parameters are used as inputs, and the target output vectors corresponding to each set of historical target running parameters are used as supervision signals to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model, and update the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model.

[0011] The jointly trained stochastic Fourier feature mapping function and linear ridge regression model are used as the turbine blade load prediction model.

[0012] Furthermore, the method for determining the target operating parameters includes:

[0013] Obtain multiple sets of historical raw operating parameters of the gas turbine;

[0014] Simulation calculations were performed on each set of historical original operating parameters to obtain the full-field load cloud map corresponding to each set of historical original operating parameters.

[0015] Extract the loads in the stress concentration region of the full-field load cloud map corresponding to each set of historical original operating parameters, and construct the original output vector corresponding to each set of historical original operating parameters;

[0016] Analyze the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters.

[0017] Use the original operating parameters whose correlation is greater than a preset threshold as the target operating parameters.

[0018] In a second aspect, the present invention provides a training apparatus for a turbine blade load prediction model, used to implement the training method for the turbine blade load prediction model described in the first aspect, comprising:

[0019] The full-field load cloud map determination module is used to acquire multiple sets of historical target operating parameters of the gas turbine, and to perform simulation calculations on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters.

[0020] The target output vector construction module is used to extract the loads in the stress concentration area of ​​the full-field load cloud map corresponding to each set of historical target operating parameters, and to construct the target output vector corresponding to each set of historical target operating parameters.

[0021] The random Fourier feature mapping module is used to convert the first input vector composed of each set of historical target running parameters into a low-dimensional random feature vector using a predefined random Fourier feature mapping function.

[0022] The model training module is used to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model under the linear ridge regression framework. It takes the low-dimensional random feature vector corresponding to each set of historical target running parameters as input and the target output vector corresponding to each set of historical target running parameters as supervision signal. It also updates the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model.

[0023] The turbine blade load prediction model determination module is used to use the jointly trained stochastic Fourier feature mapping function and linear ridge regression model as the turbine blade load prediction model.

[0024] Furthermore, it also includes: an operating parameter selection module, used to acquire multiple sets of historical original operating parameters of the gas turbine; to perform simulation calculations on each set of historical original operating parameters to obtain the full-field load cloud map corresponding to each set of historical original operating parameters; to extract the loads in the stress concentration areas of the full-field load cloud map corresponding to each set of historical original operating parameters, and to construct the original output vector corresponding to each set of historical original operating parameters; to calculate the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters; and to use the original operating parameters with a correlation greater than a preset threshold as the target operating parameters.

[0025] Thirdly, the present invention provides a method for predicting turbine blade loads, comprising:

[0026] Real-time acquisition of the current target operating parameters of the gas turbine;

[0027] Construct the current parameter vector based on the current target operating parameters;

[0028] The current parameter vector is input into a pre-trained turbine blade load prediction model, and the current parameter vector is converted into a low-dimensional random feature vector using the random Fourier feature mapping function.

[0029] The current load on the turbine blade is obtained by performing an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model.

[0030] The turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model, and the turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model described in the first aspect.

[0031] Furthermore, it also includes: determining the cumulative damage of the turbine blade based on the current load of the turbine blade, the SN curve of the turbine blade material, and the fatigue and creep damage model;

[0032] The remaining service life of the turbine blade is estimated based on the cumulative damage to the turbine blade.

[0033] Furthermore, it also includes: determining whether the turbine blade is in an off-design condition based on the current load of the turbine blade, and issuing a warning when the turbine blade is in an off-design condition.

[0034] Fourthly, the present invention provides a turbine blade load prediction device for implementing the turbine blade load prediction method described in the third aspect, comprising:

[0035] The operating parameter acquisition module is used to acquire the current target operating parameters of the gas turbine in real time.

[0036] The load prediction module is used to construct a current parameter vector based on the current target operating parameters; input the current parameter vector into a pre-trained turbine blade load prediction model, and use the stochastic Fourier feature mapping function to convert the current parameter vector into a low-dimensional random feature vector; perform an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model to obtain the current load of the turbine blade; wherein, the turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model, and the turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model described in the first aspect.

[0037] Furthermore, it also includes: a cumulative damage determination module, used to determine the cumulative damage of the turbine blade based on the current load of the turbine blade, the SN curve of the turbine blade material, and the fatigue and creep damage model;

[0038] The life assessment module is used to estimate the remaining service life of the turbine blade based on the cumulative damage to the turbine blade.

[0039] Furthermore, it also includes an anomaly warning module, used to determine whether the turbine blade is in an undesigned operating condition based on the current load of the turbine blade, and to issue a warning when the turbine blade is in an undesigned operating condition.

[0040] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0041] This invention, through simulation calculations of historical target operating parameters, can accurately capture complex load changes, ensuring high fidelity in the training data source. By using stochastic Fourier eigenmaps, the original nonlinear kernel ridge regression problem, solved in an implicit high-dimensional space, is transformed into a linear problem solved in an explicit low-dimensional space. Leveraging the powerful nonlinear modeling capabilities of kernel methods while possessing the prediction speed of linear models, the trained turbine blade load prediction model exhibits both real-time prediction and high prediction accuracy. By predicting turbine blade loads in real time using this model, the turbine blade's lifespan can be predicted, helping operators avoid or reduce operation under conditions that cause excessive damage to components, thereby extending component lifespan and improving the overall operational safety and economic efficiency of the gas turbine. This invention achieves true predictive maintenance by assessing the damage accumulation process through real-time load prediction. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a training method for a turbine blade load prediction model provided in an embodiment of the present invention.

[0043] Figure 2 A schematic flowchart of a turbine blade load prediction method provided in an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the structure of a training device for a turbine blade load prediction model provided in an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the structure of a turbine blade load prediction device provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0048] Example 1

[0049] like Figure 1 As shown in the figure, this embodiment introduces a training method for a turbine blade load prediction model, including:

[0050] Step 1: Obtain multiple sets of historical target operating parameters of the gas turbine, and perform simulation calculations on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters.

[0051] This embodiment utilizes the operating parameters of a gas turbine, taking these parameters as input data. By simulating and calculating the operating parameters, the corresponding loads are obtained. The input data and the corresponding loads are then used as training data, resulting in high-quality, high-fidelity training data. This data serves as the basis for subsequent machine learning models to learn physical laws.

[0052] Specifically, a precise three-dimensional geometric model of the gas turbine blades can be established, and a systematic experimental design scheme can be set. This scheme aims to systematically change a set of key gas turbine operating parameters to cover all operating conditions that it may encounter in actual operation, including normal, full load, and some extreme conditions. There are many types of gas turbine operating parameters; the operating parameters with a high correlation to the load can be selected from multiple parameters as target operating parameters, which can then be used as effective predictive factors.

[0053] Optionally, multiple sets of historical raw operating parameters of the gas turbine are obtained. These raw operating parameters can be all operating parameters of the gas turbine or operating parameters determined empirically. Simulation calculations are performed on each set of historical raw operating parameters to obtain a full-field load cloud map corresponding to the historical raw operating parameters. Loads in key regions are selected from the full-field load cloud map corresponding to the historical raw operating parameters to construct the original output vector. The correlation between the vector constructed from the historical raw operating parameters and the original output vector is statistically analyzed. Raw operating parameters with a correlation greater than a preset threshold are selected as target operating parameters. Target operating parameters may include: gas turbine inlet temperature, natural gas flow rate, variable guide vane (IGV) angle, rotational speed, airflow rate, unit power generation, and gas turbine exhaust temperature.

[0054] In this embodiment, the method for determining the target operating parameters includes:

[0055] Step 1.1: Obtain multiple sets of historical raw operating parameters of the gas turbine.

[0056] Step 1.2: Perform simulation calculations on each set of historical original operating parameters to obtain the full-field load cloud map corresponding to each set of historical original operating parameters.

[0057] This embodiment performs a complete coupled simulation calculation for each set of historical target operating parameters. Optionally, CFD (Computational Fluid Dynamics) methods can be used to calculate the aerodynamic pressure and temperature distribution on the turbine blade surface; the aerodynamic pressure and temperature distributions are then applied to an FEM (Finite Element Method) model, and combined with centrifugal force loads, to calculate the full-field load contour map inside the turbine blade. The full-field load contour map includes stress, strain, temperature, pressure, etc.

[0058] In this embodiment, simulation calculations are performed on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters, including:

[0059] Based on the historical target operating parameters for each set, the aerodynamic pressure and temperature distribution on the turbine blade surface are calculated using computational fluid dynamics (CFD) methods.

[0060] Aerodynamic pressure and temperature distributions are applied to the finite element method (FEM) model, and the full-field load cloud map inside the turbine blade is calculated based on the centrifugal force load, which serves as the full-field load cloud map corresponding to each set of historical target operating parameters.

[0061] Step 1.3: Extract the loads in the stress concentration areas of the full-field load cloud map corresponding to each set of historical original operating parameters, and construct the original output vector corresponding to each set of historical original operating parameters.

[0062] Step 1.4: Calculate the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters.

[0063] Step 1.5: Use the original operating parameters with a correlation greater than the preset threshold as the target operating parameters.

[0064] Step 2: Extract the loads in the stress concentration areas of the full-field load cloud map corresponding to each set of historical target operating parameters, and construct the target output vector corresponding to each set of historical target operating parameters.

[0065] In this embodiment, the stress concentration area is the area on the turbine blade that is prone to creep or fatigue damage.

[0066] Based on engineering experience and design principles, this embodiment identifies critical regions from the overall load cloud map. These critical regions are areas on the turbine blade prone to creep or fatigue damage. Critical regions are typically points of maximum stress, highest temperature, or high stress gradients, such as the leading edge, trailing edge, and tenon transition zone of the turbine blade. The loads in these critical regions are then constructed as target output vectors, which are the tagged data corresponding to historical target operating parameters.

[0067] Step 3: Use a predefined random Fourier feature mapping function to convert the first input vector, which consists of the historical target running parameters, into a low-dimensional random feature vector.

[0068] This embodiment uses kernel ridge regression for modeling.

[0069] Kernel tricks allow a model to learn a linear function in a high-dimensional or even infinite-dimensional feature space without explicitly calculating the coordinates of data points in that space. This is equivalent to learning a complex nonlinear function in the original input space. Using kernel functions such as the Gaussian radial basis function (RBF) kernel, highly complex nonlinear relationships between operating parameters and blade loads can be effectively captured.

[0070] Ridge regression: By adding an L2 norm regularization term to the loss function, it penalizes excessive growth of model coefficients. This regularization mechanism effectively prevents the model from overfitting on the training data, thereby improving the model's ability to generalize to unseen new data, which is crucial for handling potentially noisy real-world industrial data.

[0071] However, while the kernel ridge regression algorithm is powerful, its computational complexity for both training and prediction is highly dependent on the number of training samples N, as it requires calculating and storing an N×N Gram matrix. Specifically, the training complexity is O(N³) and the prediction complexity is O(N). When the training data is large, such as in this embodiment where thousands of simulation points are needed to ensure accuracy, this computational bottleneck prevents standard kernel ridge regression from meeting the requirements for real-time prediction.

[0072] To address this problem, this embodiment introduces stochastic Fourier features. The theoretical basis of this method is Buchner's theorem, which states that any continuous, shift-invariant kernel function, such as the commonly used Gaussian kernel, can be expressed as the expectation of its Fourier transform. The stochastic Fourier features method utilizes the Monte Carlo approximation to realize this theorem. This embodiment leverages the powerful nonlinear modeling capabilities of kernel methods while maintaining the prediction speed of linear models, thus resolving the contradiction between real-time performance and accuracy.

[0073] Step 4: Under the linear ridge regression framework, the low-dimensional random feature vector corresponding to each set of historical target running parameters is used as input, and the target output vector corresponding to each set of historical target running parameters is used as supervision signal to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model, and update the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model.

[0074] This implementation utilizes stochastic Fourier feature mapping to transform kernel ridge regression into standard linear ridge regression. This significantly reduces computational complexity and improves model training speed when processing large datasets, such as thousands of training data sets, during training.

[0075] Step 5: Use the jointly trained stochastic Fourier feature mapping function and linear ridge regression model as the turbine blade load prediction model.

[0076] This embodiment tests the turbine blade load prediction model using test data. The test results are shown in Table 1 below.

[0077] Table 1 Test Results

[0078] Evaluation indicators Mean Absolute Error Squared error MSE Root Mean Square Error (RMSE) Average percentage error R2 Stress prediction results 19.04 680.09 26.08 3.49 0.92 Temperature prediction results 9.01 111.86 10.58 1.41 0.99

[0079] R² is used to measure how well the regression model fits the data. Its value is generally between 0 and 1. The closer it is to 1, the stronger the model's ability to explain the data variation.

[0080] When the test results meet the requirements, the trained turbine blade load prediction model can be deployed to a real industrial environment to perform online monitoring tasks. The turbine blade load prediction model in this application essentially creates a computationally efficient, real-time surrogate model for a high-fidelity physical simulation model. This surrogate model can accurately reproduce the complex physical behavior of the physical entity on a timescale synchronized with the physical entity, which is precisely the core goal pursued in the fields of modern industrial IoT and predictive health management.

[0081] The turbine blade load prediction model training method in this embodiment utilizes a kernel ridge regression model accelerated by stochastic Fourier features, and uses offline FEM / CFD simulation data as the training basis to quickly and accurately train the turbine blade load prediction model. Then, using the turbine blade load prediction model, the full-field equivalent stress, strain, and temperature loads of the turbine blades are predicted online from the real-time operating parameters of the gas turbine.

[0082] Example 2

[0083] Based on the same inventive concept as Embodiment 1, this embodiment introduces a training device 300 for a turbine blade load prediction model, used to implement the training method for the turbine blade load prediction model described in Embodiment 1, such as... Figure 3 As shown, it includes:

[0084] The full-field load cloud map determination module 302 is used to acquire multiple sets of historical target operating parameters of the gas turbine, and to perform simulation calculations on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters.

[0085] The target output vector construction module 304 is used to extract the loads in the stress concentration area of ​​the full-field load cloud map corresponding to each set of historical target operating parameters, and to construct the target output vector corresponding to each set of historical target operating parameters.

[0086] The random Fourier feature mapping module 306 is used to convert the first input vector composed of each set of historical target running parameters into a low-dimensional random feature vector using a predefined random Fourier feature mapping function.

[0087] The model training module 308 is used to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model under the linear ridge regression framework by taking the low-dimensional random feature vector corresponding to each set of historical target running parameters as input and the target output vector corresponding to each set of historical target running parameters as supervision signal, and updating the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model.

[0088] The turbine blade load prediction model determination module 310 is used to use the jointly trained stochastic Fourier feature mapping function and linear ridge regression model as the turbine blade load prediction model.

[0089] In some embodiments, the system further includes an operating parameter selection module, used to acquire multiple sets of historical original operating parameters of the gas turbine; perform simulation calculations on each set of historical original operating parameters to obtain a full-field load cloud map corresponding to each set of historical original operating parameters; extract the loads in the stress concentration areas of the full-field load cloud map corresponding to each set of historical original operating parameters, and construct the original output vector corresponding to each set of historical original operating parameters; calculate the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters; and use the original operating parameters with a correlation greater than a preset threshold as target operating parameters.

[0090] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0091] Example 3

[0092] Based on the same inventive concept as other embodiments, this embodiment introduces a method for predicting turbine blade loads, such as... Figure 2 As shown, it includes:

[0093] Step 1: Obtain the current target operating parameters of the gas turbine in real time.

[0094] This embodiment allows the trained turbine blade load prediction model to be deployed on an edge computing device or a monitoring server connected to the gas turbine control system. The system is configured to receive data streams of the required target operating parameters, such as inlet air temperature and fuel flow rate, in real time and continuously via standard industrial protocols such as OPC-UA.

[0095] Step 2: Construct the current parameter vector based on the current target running parameters.

[0096] Step 3: Input the current parameter vector into the pre-trained turbine blade load prediction model, and use the random Fourier feature mapping function to convert the current parameter vector into a low-dimensional random feature vector.

[0097] Step 4: Perform an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model to obtain the current load of the turbine blade.

[0098] In this embodiment, the turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model. The turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model described in Embodiment 1.

[0099] In some embodiments, the method further includes determining the cumulative damage of the turbine blade based on its current load, the SN curve of the turbine blade material, and fatigue and creep damage models, and estimating the remaining service life of the turbine blade based on the cumulative damage. This embodiment can determine the cumulative damage of the turbine blade based on its current load, the SN curve of the material, and fatigue and creep damage models such as Miner's rule and the Larson-Miller parameter method, and estimate the remaining service life of the turbine blade based on the cumulative damage. Real-time load sensing capabilities can help operators avoid or reduce operation under conditions that cause excessive damage to components, thereby extending component life and improving the overall operational safety and economic efficiency of the gas turbine.

[0100] In some embodiments, the method further includes determining whether the turbine blade is under non-design conditions based on its current load, and issuing an early warning in such cases. The determination can also be based on the current load on the turbine blade itself; for example, it can identify non-design conditions that cause abnormal stress on the turbine blade and issue an early warning in such cases. This allows maintenance decisions to shift from a traditional, passive, fixed-time-interval-based model to a proactive, more scientific, and economical model based on actual health status, thereby optimizing maintenance costs and reducing unplanned downtime.

[0101] For example, the power plant's DCS (Distributed Control System) transmits the real-time operating parameters of the gas turbine (IGV angle, fuel flow, speed, power generation, exhaust temperature, etc.) to a monitoring server deployed on-site via an industrial network once per second.

[0102] A turbine blade load prediction model is deployed on the monitoring server. Upon receiving each new set of data, the model immediately performs calculations and outputs predicted values ​​for key points on the leading edge of the first-stage turbine blade, such as the maximum stress point, within milliseconds. The prediction results are displayed in real-time as trend curves on the operator monitoring interface in the power plant's control room, allowing operators to intuitively see the direct impact of current operations on the blade load. Simultaneously, all predicted data, along with corresponding operating parameters, are synchronously recorded in a historical database. The load data in the historical database is continuously accessed by a parallel fatigue life calculation module. This module updates the blade's consumed life fraction in real-time based on the material's SN curve and creep data. Maintenance planning engineers can use this system to clearly see the blade's lifespan consumption rate and, based on actual wear and tear rather than fixed operating hours, accurately plan the next major overhaul.

[0103] Example 4

[0104] Based on the same inventive concept as other embodiments, this embodiment introduces a turbine blade load prediction device 400, used to implement the turbine blade load prediction method described in Embodiment 3, such as... Figure 4 As shown, it includes:

[0105] The operating parameter acquisition module 402 is used to acquire the current target operating parameters of the gas turbine in real time;

[0106] The load prediction module 404 is used to construct a current parameter vector based on the current target operating parameters; input the current parameter vector into a pre-trained turbine blade load prediction model, and use the stochastic Fourier feature mapping function to convert the current parameter vector into a low-dimensional random feature vector; perform an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model to obtain the current load of the turbine blade; wherein, the turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model, and the turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model described in Example 2.

[0107] In some embodiments, the system further includes a cumulative damage determination module for determining the cumulative damage of the turbine blade based on the current load of the turbine blade, the SN curve of the turbine blade material, and a fatigue and creep damage model.

[0108] In some embodiments, a life assessment module is also included for estimating the remaining service life of the turbine blade based on the cumulative damage to the turbine blade.

[0109] In some embodiments, an anomaly warning module is also included, which is used to determine whether the turbine blade is in an off-design condition based on the current load of the turbine blade, and to issue a warning when the turbine blade is in an off-design condition.

[0110] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 3, and will not be repeated here.

[0111] Example 5

[0112] Based on the same inventive concept as other embodiments, this embodiment describes an electronic device, such as... Figure 5 As shown, it includes: processor 502, communication interface 504, memory 506, and communication bus 508.

[0113] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508.

[0114] Communication interface 504 is used to communicate with other electronic devices or servers.

[0115] The processor 502 is used to execute program 510, specifically the relevant steps in the above method embodiments.

[0116] Specifically, program 510 may include program code that includes computer operation instructions.

[0117] Processor 502 may be a central processing unit, a specific integrated circuit, or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0118] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0119] Specifically, program 510 can be used to cause processor 502 to execute the steps in the methods of embodiments 1 or 3 described above.

[0120] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0121] In summary, this invention, through simulation calculations of historical target operating parameters, can accurately capture complex load changes, ensuring high fidelity in the training data source. By using stochastic Fourier eigenmaps, the original nonlinear kernel ridge regression problem, solved in an implicit high-dimensional space, is transformed into a linear problem solved in an explicit low-dimensional space. Utilizing the powerful nonlinear modeling capabilities of kernel methods while possessing the prediction speed of linear models, the trained turbine blade load prediction model exhibits both real-time prediction and high prediction accuracy. By predicting turbine blade loads in real time using this model, the turbine blade's lifespan can be predicted, helping operators avoid or reduce operation under conditions that cause excessive damage to components, thereby extending component lifespan and improving the overall operational safety and economic efficiency of the gas turbine. This invention achieves true predictive maintenance by assessing the damage accumulation process through real-time load prediction.

[0122] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A training method for a turbine blade load prediction model, characterized in that, include: Multiple sets of historical target operating parameters of the gas turbine are obtained, and simulation calculations are performed on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters; Extract the loads in the stress concentration region of the full-field load cloud map corresponding to each set of historical target operating parameters, and construct the target output vector corresponding to each set of historical target operating parameters; The first input vector, composed of each set of historical target running parameters, is converted into a low-dimensional random feature vector using a predefined random Fourier feature mapping function. Within the linear ridge regression framework, the low-dimensional random feature vectors corresponding to each set of historical target running parameters are used as inputs, and the target output vectors corresponding to each set of historical target running parameters are used as supervision signals to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model, and update the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model. The jointly trained stochastic Fourier feature mapping function and linear ridge regression model are used as the turbine blade load prediction model.

2. The training method for the turbine blade load prediction model according to claim 1, characterized in that, The method for determining the target operating parameters includes: Obtain multiple sets of historical raw operating parameters of the gas turbine; Simulation calculations were performed on each set of historical original operating parameters to obtain the full-field load cloud map corresponding to each set of historical original operating parameters. Extract the loads in the stress concentration region of the full-field load cloud map corresponding to each set of historical original operating parameters, and construct the original output vector corresponding to each set of historical original operating parameters; Analyze the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters. Use the original operating parameters whose correlation is greater than a preset threshold as the target operating parameters.

3. A training device for a turbine blade load prediction model, characterized in that, A training method for implementing the turbine blade load prediction model according to any one of claims 1-2 includes: The full-field load cloud map determination module is used to acquire multiple sets of historical target operating parameters of the gas turbine, and to perform simulation calculations on each set of historical target operating parameters to obtain the full-field load cloud map corresponding to each set of historical target operating parameters. The target output vector construction module is used to extract the loads in the stress concentration area of ​​the full-field load cloud map corresponding to each set of historical target operating parameters, and to construct the target output vector corresponding to each set of historical target operating parameters. The random Fourier feature mapping module is used to convert the first input vector composed of each set of historical target running parameters into a low-dimensional random feature vector using a predefined random Fourier feature mapping function. The model training module is used to jointly train the stochastic Fourier feature mapping function and the linear ridge regression model under the linear ridge regression framework. It takes the low-dimensional random feature vector corresponding to each set of historical target running parameters as input and the target output vector corresponding to each set of historical target running parameters as supervision signal. It also updates the internal parameters of the stochastic Fourier feature mapping function and the weight vector of the linear ridge regression model. The turbine blade load prediction model determination module is used to use the jointly trained stochastic Fourier feature mapping function and linear ridge regression model as the turbine blade load prediction model.

4. The training device for the turbine blade load prediction model according to claim 3, characterized in that, Also includes: The operating parameter selection module is used to obtain multiple sets of historical original operating parameters of the gas turbine; simulation calculations are performed on each set of historical original operating parameters to obtain the full-field load cloud map corresponding to each set of historical original operating parameters; Extract the loads in the stress concentration areas of the full-field load cloud map corresponding to each set of historical original operating parameters, and construct the original output vector corresponding to each set of historical original operating parameters; calculate the correlation between the second input vector formed by each set of historical original operating parameters and the original output vector corresponding to each set of historical original operating parameters; and take the original operating parameters with a correlation greater than a preset threshold as the target operating parameters.

5. A method for predicting turbine blade loads, characterized in that, include: Real-time acquisition of the current target operating parameters of the gas turbine; Construct the current parameter vector based on the current target operating parameters; The current parameter vector is input into a pre-trained turbine blade load prediction model, and the current parameter vector is converted into a low-dimensional random feature vector using the random Fourier feature mapping function. The current load on the turbine blade is obtained by performing an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model. The turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model, and the turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model according to claim 1 or 2.

6. The turbine blade load prediction method according to claim 5, characterized in that, Also includes: Based on the current load of the turbine blade, the SN curve of the turbine blade material, and the fatigue and creep damage model, the cumulative damage of the turbine blade is determined. The remaining service life of the turbine blade is estimated based on the cumulative damage to the turbine blade.

7. The turbine blade load prediction method according to claim 5, characterized in that, Also includes: Based on the current load of the turbine blade, determine whether the turbine blade is in an off-design condition, and issue an early warning if it is in an off-design condition.

8. A turbine blade load prediction device, characterized in that, The method for predicting turbine blade loads according to any one of claims 5-7 includes: The operating parameter acquisition module is used to acquire the current target operating parameters of the gas turbine in real time. The load prediction module is used to construct a current parameter vector based on the current target operating parameters; input the current parameter vector into a pre-trained turbine blade load prediction model, and use the stochastic Fourier feature mapping function to convert the current parameter vector into a low-dimensional random feature vector; perform an inner product operation between the low-dimensional random feature vector and the weight vector in the linear ridge regression model to obtain the current load of the turbine blade; wherein, the turbine blade load prediction model includes a stochastic Fourier feature mapping function and a linear ridge regression model, and the turbine blade load prediction model is trained based on the training method of the turbine blade load prediction model according to claim 1 or 2.

9. The turbine blade load prediction device according to claim 8, characterized in that, Also includes: The cumulative damage determination module is used to determine the cumulative damage of the turbine blade based on the current load of the turbine blade, the SN curve of the turbine blade material, and the fatigue and creep damage model. The life assessment module is used to estimate the remaining service life of the turbine blade based on the cumulative damage to the turbine blade.

10. The turbine blade load prediction device according to claim 8, characterized in that, Also includes: An anomaly warning module is used to determine whether the turbine blade is in an undesigned operating condition based on the current load of the turbine blade, and to issue a warning when the turbine blade is in an undesigned operating condition.

Images