Intelligent estimation method for vibration state of EH oil pipe of steam turbine generator unit
By constructing a time-series neural network model with a Transformer-LSTM architecture and combining primary frequency regulation commands with vibration characteristics, the problem of early warning lag in traditional monitoring methods is solved. This enables accurate prediction and early warning of the vibration status of the EH oil pipe of the steam turbine generator set, improving prediction accuracy and operation and maintenance efficiency.
Patent Information
- Application Number
- CN202511832162.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-06
AI Technical Summary
Traditional monitoring methods suffer from early warning lag, failing to predict in a timely manner faults such as fatigue cracks and loose joints in the EH oil pipe of steam turbine generator sets. Existing intelligent models cannot effectively capture long-term time-series dependencies and local dynamic features, resulting in large vibration prediction errors and insufficient robustness. Furthermore, they do not fully integrate primary frequency regulation command parameters with vibration time-series characteristics.
A method for intelligently estimating the vibration state of the EH oil pipe of a steam turbine generator set is designed. By acquiring the primary frequency regulation command parameters and vibration dynamic features, a multi-dimensional joint temporal feature matrix is constructed. A temporal neural network model with a Transformer-LSTM architecture is adopted, and a two-stage framework of offline training and online estimation is combined to achieve accurate prediction of the vibration state and construct a hierarchical early warning system.
It achieves accurate prediction of vibration status, with prediction accuracy improved to 97.3% and response latency of only 35ms, providing an intervention time window of no less than 5 seconds, effectively avoiding failures, reducing operation and maintenance costs, and promoting the transformation of operation and maintenance mode from post-maintenance to preventive maintenance.
Smart Images
Figure CN121611518A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of safety monitoring of steam turbine generator sets, and in particular to an intelligent estimation method for the vibration state of the EH oil pipe of a steam turbine generator set. Background Technology
[0002] Traditional monitoring methods rely on mechanistic models and preset fixed thresholds, and only trigger alarms when vibration parameters exceed the thresholds. This results in a serious problem of early warning lag, which cannot provide maintenance personnel with enough time to intervene and makes it difficult to avoid failures such as pipeline fatigue cracks, loose joints, or even rupture and leakage.
[0003] Existing single intelligent models (such as RNN, SVR, BP neural network, etc.) have limitations: RNN has difficulty capturing long-range temporal dependencies, SVR has low prediction accuracy and long response delay, and a single LSTM or Transformer model cannot simultaneously take into account the accurate modeling of local dynamic features and global correlation information, resulting in large vibration prediction errors and insufficient robustness.
[0004] Traditional methods do not fully integrate primary frequency modulation command parameters with vibration timing characteristics, and cannot establish a dynamic correlation between "command and response", making it difficult to accurately predict vibration response under scenarios such as sudden load changes and frequency modulation direction changes. Summary of the Invention
[0005] The purpose of this invention is to solve the above-mentioned problems by designing an intelligent estimation method for the vibration state of the EH oil pipe of a steam turbine generator set.
[0006] The technical solution of the present invention to achieve the above objectives is a method for intelligent estimation of the vibration state of the EH oil pipe of a steam turbine generator set, comprising the following steps:
[0007] Obtain the frequency modulation command parameters, which include the load command change amount ∆P, the command change rate dP / dt, and the frequency modulation direction D. dir ;
[0008] Extract the vibration dynamic characteristics of the EH tubing, the vibration dynamic characteristics including the vibration frequency spectrum energy E high、 Root mean square value of amplitude and vibration acceleration A peak ;
[0009] The primary frequency modulation command parameters and the vibration dynamic features are aligned by timestamps to construct a multi-dimensional joint temporal feature matrix;
[0010] The multidimensional joint temporal feature matrix is input into a pre-trained temporal neural network model, which outputs an estimated value of the vibration state. The temporal neural network model is a Transformer-LSTM architecture.
[0011] 2. The method according to claim 1, characterized in that the specific steps of extracting the vibration dynamic features include:
[0012] The original tubing pressure signal was subjected to a 3-level discrete wavelet transform using the db4 wavelet basis, and the spectral energy of the high-frequency coefficients D3 in the 3rd level was extracted as the vibration frequency spectral energy E. high ;
[0013] The root mean square value of the pressure fluctuation is calculated using a sliding window with a window length of 10 seconds and a step size of 5 seconds to characterize the root mean square value of the amplitude.
[0014] The vibration acceleration A is derived based on the relationship between vibration frequency and physical quantities. peak And set A peak > 0.5×10 5 This serves as a threshold for determining vibration events.
[0015] The temporal neural network model, namely the Transformer-LSTM architecture, includes:
[0016] The input embedding layer is used to map input features into dense vectors;
[0017] The position encoding module is used to inject position information into the dense vector;
[0018] An N-layer Transformer encoder stack, including a multi-head self-attention mechanism and a feedforward neural network, is used to extract global long-range dependency features;
[0019] Time dimension extension module;
[0020] Two cascaded LSTM layers are used to extract local temporal dynamic features through their gating mechanism;
[0021] The feature fusion module is used to concatenate the global long-range dependency features and the local temporal dynamic features along the channel dimension.
[0022] A linear output layer is used to output vibration state estimates.
[0023] The method is implemented based on a two-stage framework of offline training and online estimation:
[0024] Offline training phase: Using continuous historical vibration data and frequency modulation command parameters as training inputs, the temporal neural network model is trained using the Adam optimizer and a cosine annealing learning rate scheduling strategy, and the model hyperparameters are optimized using a grid search method.
[0025] Online estimation stage: The unit operation data after processing is received in real time, and the vibration state estimate is output using the trained time-series neural network model to achieve real-time monitoring.
[0026] It also includes a hierarchical early warning system built based on the vibration acceleration predictions output by the model:
[0027] When the predicted vibration acceleration value exceeds the set threshold, the system immediately marks the vibration event and generates an early warning signal;
[0028] Establish a three-tiered early warning mechanism of "prediction-assessment-response" to provide maintenance personnel with an intervention window of no less than 5 seconds to support the implementation of frequency modulation rate optimization and enhanced proactive intervention measures during inspections.
[0029] On the other hand, this invention provides a time-series neural network model for an intelligent estimation method of the vibration state of the EH oil pipe in a steam turbine generator set. The model is a Transformer-LSTM architecture, and its hyperparameters are configured as follows:
[0030] The LSTM has 2 layers and the hidden layer dimension is 64.
[0031] The Transformer encoder has 2 layers, the hidden layer dimension is 64, and the number of attention heads is 4;
[0032] The output window size is 5 seconds;
[0033] The training rounds are 1000, and the batch size is 2048.
[0034] On the other hand, the present invention also provides an intelligent estimation system for the vibration state of the EH oil pipe of a steam turbine generator set, comprising:
[0035] The data acquisition module is used to acquire tubing pressure signals and primary frequency modulation command parameters;
[0036] The feature extraction module is used to perform vibration dynamic feature extraction methods;
[0037] The processing module loads a time-series neural network model to execute the method described in claim 1 and output vibration state estimates.
[0038] The early warning module is used to implement a tiered early warning system, generating and outputting early warning signals.
[0039] This invention provides an intelligent estimation method for the vibration state of EH oil pipes in steam turbine generator sets. It constructs a "command-response" time-series prediction mechanism, aligning primary frequency regulation command parameters (load command change, command change rate, and frequency regulation direction) with oil pipe vibration time-series features (frequency, amplitude, and vibration acceleration) by timestamps to form a multi-dimensional joint time-series feature matrix. Employing a Transformer-LSTM architecture, it combines the advantages of Transformer's self-attention mechanism in capturing long-range dependencies with the ability of LSTM's gating mechanism to mine local time-series dynamic features, achieving comprehensive modeling of complex time-series correlations. A dual-stage framework of offline training and online estimation is designed, integrating data preprocessing, feature engineering, parameter optimization, and a hierarchical early warning system to form a full-process intelligent monitoring solution.
[0040] Functions: Offline stage: Based on historical vibration data of the unit and frequency regulation command parameters, complete model training and parameter optimization, and learn the local dynamic characteristics and global correlation law of vibration signals; Online stage: Receive unit operation data in real time, quickly output vibration state estimates, and trigger graded warnings when the predicted acceleration exceeds the threshold, achieving a 5-second advance warning; Supports dynamic adjustment of the prediction window scale to adapt to different frequency regulation scenarios and is compatible with the actual operating threshold settings of power plants.
[0041] The prediction accuracy is significantly improved, reaching 97.3%, which is more than 9.5% higher than the accuracy of traditional models such as RNN and SVR. Error indicators such as RMSE are greatly reduced, and vibration state estimation is more accurate. It solves the problem of early warning lag in traditional methods, providing vibration event early warning 5 seconds in advance, providing maintenance personnel with a critical intervention time window, and effectively avoiding pipeline failures and unexpected unit shutdowns. It reduces power plant operation and maintenance costs, reduces the frequency of manual inspections and unplanned maintenance expenditures, and promotes the transformation of operation and maintenance mode from "reactive maintenance" to "preventive maintenance". It operates efficiently and stably, with a model response latency of only 35ms, low memory requirements, low computational burden, and is suitable for real-time processing of millions of time series data, with strong compatibility. Attached Figure Description
[0042] Figure 1 This is a schematic diagram of the intelligent estimation method for the vibration state of the EH oil pipe of a steam turbine generator set as described in this invention;
[0043] Figure 2 This is a model architecture diagram of the intelligent estimation method for the vibration state of the EH oil pipe of a steam turbine generator set as described in this invention;
[0044] Figure 3 This is a prediction result diagram of the intelligent estimation method for the vibration state of the EH oil pipe of a steam turbine generator set as described in this invention. Detailed Implementation
[0045] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1-3 As shown, a smart estimation method for the vibration state of the EH oil pipe of a steam turbine generator set includes the following steps:
[0046] Obtain the primary frequency regulation command parameters, including the load command change ∆P, the command change rate dP / dt, and the frequency regulation direction D. dir ;
[0047] The vibration dynamic characteristics of the EH tubing were extracted, including the vibration frequency spectrum energy E. high、 Root mean square value of amplitude and vibration acceleration A peak ;
[0048] Align the primary frequency modulation command parameters with the vibration dynamic characteristics according to the timestamp to construct a multi-dimensional joint time-series feature matrix;
[0049] The multidimensional joint temporal feature matrix is input into a pre-trained temporal neural network model, which outputs an estimated value of the vibration state. The temporal neural network model is a Transformer-LSTM architecture.
[0050] The specific steps for extracting vibration dynamic features include:
[0051] The original tubing pressure signal was subjected to a 3-level discrete wavelet transform using the db4 wavelet basis, and the spectral energy of the high-frequency coefficients D3 in the 3rd level was extracted as the oscillation frequency spectral energy E. high ;
[0052] The root mean square value of pressure fluctuation is calculated using a sliding window with a window length of 10 seconds and a step size of 5 seconds to characterize the root mean square value of amplitude.
[0053] The vibration acceleration A is derived based on the relationship between vibration frequency and physical quantities. peak And set A peak > 0.5×10 5 This serves as a threshold for determining vibration events.
[0054] Temporal neural network models, namely the Transformer-LSTM architecture, include:
[0055] The input embedding layer is used to map input features into dense vectors;
[0056] The position encoding module is used to inject position information into dense vectors;
[0057] An N-layer Transformer encoder stack, including a multi-head self-attention mechanism and a feedforward neural network, is used to extract global long-range dependency features;
[0058] Time dimension extension module;
[0059] Two cascaded LSTM layers are used to extract local temporal dynamic features through their gating mechanism;
[0060] The feature fusion module is used to stitch together global long-range dependency features and local temporal dynamic features along the channel dimension.
[0061] A linear output layer is used to output vibration state estimates.
[0062] The method is implemented based on a two-stage framework of offline training and online estimation:
[0063] Offline training phase: Using continuous historical vibration data and frequency modulation command parameters as training inputs, the temporal neural network model is trained using the Adam optimizer and a cosine annealing learning rate scheduling strategy, and the model hyperparameters are optimized using the grid search method.
[0064] Online estimation stage: Real-time reception of processed unit operation data, output of vibration state estimate using trained time-series neural network model, and real-time monitoring.
[0065] It also includes a hierarchical early warning system built based on the vibration acceleration predictions output by the model:
[0066] When the predicted vibration acceleration exceeds the set threshold, the system immediately marks the vibration event and generates an early warning signal;
[0067] Establish a three-tiered early warning mechanism of "prediction-assessment-response" to provide maintenance personnel with an intervention window of no less than 5 seconds to support the implementation of frequency modulation rate optimization and enhanced proactive intervention measures during inspections.
[0068] A time-series neural network model for an intelligent estimation method of the vibration state of the EH oil pipe in a steam turbine generator set is presented. The model uses a Transformer-LSTM architecture, and its hyperparameters are configured as follows:
[0069] The LSTM has 2 layers and the hidden layer dimension is 64.
[0070] The Transformer encoder has 2 layers, the hidden layer dimension is 64, and the number of attention heads is 4;
[0071] The output window size is 5 seconds;
[0072] The training rounds are 1000, and the batch size is 2048.
[0073] A smart estimation system for the vibration state of EH oil pipes in a steam turbine generator set includes:
[0074] The data acquisition module is used to acquire tubing pressure signals and primary frequency modulation command parameters;
[0075] The feature extraction module is used to perform vibration dynamic feature extraction methods;
[0076] The processing module loads a time-series neural network model to execute the method of claim 1 and output vibration state estimates.
[0077] The early warning module is used to implement a tiered early warning system, generating and outputting early warning signals.
[0078] The key feature of this implementation scheme is that it incorporates the load command change ∆P, the command change rate dP / dt, and the frequency modulation direction D. dir The parameters of the first frequency modulation command, and the frequency spectrum energy E high , Root mean square value of amplitude, vibration acceleration A peak By aligning the vibration dynamic features with timestamps, a multidimensional joint temporal feature matrix is constructed to realize the mathematical representation of the correlation between commands and vibration dynamics.
[0079] A temporal neural network model with a Transformer-LSTM architecture includes an input embedding layer, a position encoding module, an N-layer Transformer encoder stack, a temporal dimension extension module, a two-layer cascaded LSTM layer, a feature fusion module, and a linear output layer. The Transformer encoder stack contains a multi-head self-attention mechanism and a feedforward neural network to extract global long-range dependent features. The LSTM layer extracts local temporal dynamic features through a forget gate, input gate, and output gate gating mechanism. The feature fusion module concatenates two types of features along the channel dimension to maximize the retention of complementary advantages.
[0080] The method for extracting vibration dynamic features involves performing a three-level discrete wavelet transform on the original tubing pressure signal using the db4 wavelet basis, and extracting the spectral energy E of the third-level high-frequency coefficients D3. high The root mean square value of pressure fluctuation is calculated using a sliding window with a 10-second window and a 5-second step size to characterize the amplitude; the vibration acceleration is derived based on the relationship between vibration frequency and physical quantities, and A is set. peak >0.5×10 5 The threshold for determining vibration events.
[0081] The framework employs a two-stage approach: offline training and online estimation. In the offline stage, continuous historical vibration data and frequency modulation command parameters are used as inputs. The Adam optimizer is combined with a cosine annealing learning rate scheduling strategy to train the model, and hyperparameters such as the number of LSTM layers and the number of Transformer attention heads are optimized using a grid search method. In the online stage, the processed data is received in real time, and the trained model is used to output vibration state estimates to achieve real-time monitoring.
[0082] The tiered early warning system, which provides a 5-second advance warning, constructs a three-level mechanism of "prediction-assessment-response" based on the vibration acceleration prediction value output by the model. When the prediction value exceeds the set threshold, the vibration event is immediately marked and an early warning signal is generated, providing maintenance personnel with an intervention time window of no less than 5 seconds, supporting the implementation of proactive intervention measures such as frequency modulation rate optimization and enhanced inspection.
[0083] The hyperparameter configuration scheme adapted for thermal power units is characterized by: 2 LSTM layers with a hidden layer dimension of 64, 2 Transformer encoder layers with a hidden layer dimension of 64, 4 attention heads, an output window size of 5 seconds, 1000 training epochs, and a batch size of 2048. This configuration ensures that the model improves prediction accuracy while reducing memory requirements and computational latency.
[0084] In this implementation scheme, based on the core logic of "data-driven - model building - real-time estimation - early warning," a dynamic correlation model between a primary frequency modulation command and the vibration response of the EH pipeline is established by integrating a temporal neural network of Transformer and LSTM, achieving accurate prediction and early warning of vibration status. Its core is to utilize Transformer to capture the long-range dependency between the frequency modulation command and the vibration sequence, and to use LSTM to mine the local temporal dynamic features of the vibration signal. A two-stage framework is used to complete offline model training and online real-time estimation, and finally, intervention signals are output through a tiered early warning system.
[0085] Detailed Workflow
[0086] Phase 1: Offline Training Process
[0087] Data Acquisition and Preprocessing: Operating data from the EH oil pipe vibration monitoring system of the 350MW steam turbine unit in the thermal power plant was collected to obtain valid time-series records, covering nine core variables including unit load, main steam pressure, and frequency regulation commands. Data quality control, normalization, and outlier removal were performed to ensure data reliability.
[0088] Feature construction: Extracting primary frequency modulation command features (load command change ∆P, command change rate dP / dt, frequency modulation direction D) dir ) and vibration dynamic characteristics (vibration frequency spectrum energy E) high , Root mean square value of amplitude, vibration acceleration A peak The data is aligned with timestamps to form a multidimensional joint temporal feature matrix, and vibration event labels are added. peak >0.5×10 5 y=1 when y=1, otherwise y=0.
[0089] Model Construction and Training: A Transformer-LSTM model was built, with 2 LSTM layers, 64 hidden layers, 2 Transformer encoder layers, 4 attention heads, and an output window size of 5 seconds. Mean Squared Error (MSE) was used as the loss function, and the Adam optimizer was combined with a cosine annealing learning rate scheduling strategy (initial learning rate 1e-3, minimum learning rate 1e-5, epochs 50). The model was trained for 1000 epochs on the training set to learn the local dynamics and global correlation patterns of vibration signals.
[0090] Figure 3 This is a graph showing the prediction results.
[0091] Model validation and optimization: Use the remaining data as a test set to validate the model performance, compare the predicted values with the actual values, optimize the hyperparameters using the grid search method, and ensure that the model R² reaches above 0.9733 and the response latency is controlled within 35ms.
[0092] Phase Two: Online Estimation and Early Warning Process
[0093] Real-time data acquisition and processing: Collect unit operation data online in real time, with a sampling period of 1 second, repeat the data preprocessing and feature construction steps of the offline stage, and generate multi-dimensional time series features that meet the model input requirements.
[0094] Real-time vibration state estimation: The processed features are input into the trained model and mapped into a dense vector through the input embedding layer. After position information is injected through position encoding, the global long-range dependent features are extracted by the Transformer encoder stack. At the same time, the sequence after time dimension expansion is input into the LSTM layer to mine local dynamic features. The two types of features are concatenated by the fusion module and output as vibration acceleration prediction values through the linear output layer.
[0095] Tiered Early Warning and Response: The early warning judgment module compares the predicted value with a preset threshold. If A... peak >0.5×10 5 If the event is marked as a vibration event (early warning status), an early warning signal is generated. The system triggers a three-level "prediction-assessment-response" mechanism based on the early warning status, prompting maintenance personnel to implement proactive intervention measures such as frequency modulation rate optimization and enhanced inspection of key nodes, thereby achieving an early warning 5 seconds in advance.
[0096] Model iteration and update: Regularly incorporate online operating data into historical datasets, retrain the model, continuously optimize hyperparameters, and improve the model's adaptability and prediction accuracy under different operating conditions.
[0097] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.
Claims
1. A method for intelligent estimation of the vibration state of the EH oil pipe of a steam turbine generator unit, characterized in that, The method comprises the following steps: Obtaining a frequency modulation instruction parameter, the instruction parameter including a load instruction change ∆P, an instruction change rate dP / dt and a frequency modulation direction D dir ; extracting the vibration dynamic characteristics of the EH tubing, said vibration dynamic characteristics comprising a spectrum of vibration frequencies energy E high、 a root mean square value of the amplitude and a vibration acceleration A peak ; timestamping the primary frequency modulation instruction parameters and the vibration dynamic characteristics to construct a multi-dimensional joint time series feature matrix; inputting the multi-dimensional joint time series feature matrix into a pre-trained time series neural network model to output an estimated value of the vibration state, wherein the time series neural network model is a Transformer-LSTM fusion architecture.
2. The method of claim 1, wherein, The specific steps of extracting the vibration dynamic characteristics comprise: The original tubing pressure signal is subjected to three-layer discrete wavelet transform using a db4 wavelet base, and the spectral energy of the third-layer high-frequency coefficient D3 is extracted as the vibration frequency spectral energy E high ; calculating the root mean square value of the pressure fluctuation through a sliding window with a window length of 10 seconds and a step length of 5 seconds to represent the amplitude root mean square value; Based on the relationship between the vibration frequency and the physical quantity, the vibration acceleration A is derived peak , and A peak > 0.5 x 10 5 is set as the determination threshold of the vibration event.
3. The method of claim 1, wherein, the time series neural network model, i.e., the Transformer-LSTM architecture, comprises: an input embedding layer for mapping input features into dense vectors; a position encoding module for injecting position information into the dense vectors; an N-layer Transformer encoder stack containing a multi-head self-attention mechanism and a feedforward neural network for extracting global long-range dependency features; a time dimension expansion module; two cascaded LSTM layers for extracting local time series dynamic features through their gating mechanisms; a feature fusion module for concatenating the global long-range dependency features and the local time series dynamic features in the channel dimension; a linear output layer for outputting the vibration state estimated value.
4. The method of claim 1, wherein, The implementation of the method is based on a two-stage framework of offline training and online estimation: in the offline training stage, the time series neural network model is trained using the continuous historical vibration data and the frequency modulation instruction parameters as training inputs, an Adam optimizer, and a cosine annealing learning rate scheduling strategy, and the model hyperparameters are optimized through a grid search method; in the online estimation stage, the trained time series neural network model is used to output the vibration state estimated value based on the processed unit operation data, thereby realizing real-time monitoring.
5. The method of claim 1, wherein, a hierarchical early warning system based on the vibration acceleration prediction value output by the model is further included: when the vibration acceleration prediction value exceeds a set threshold, the system immediately labels a vibration event and generates an early warning signal; a "prediction-evaluation-response" three-level early warning mechanism is constructed to provide an intervention time window of no less than 5 seconds for the operation and maintenance personnel to support the implementation of frequency modulation rate optimization and reinforcement of active intervention measures.
6. A timing neural network model for implementing the method of any one of claims 1 to 5, characterized in that, The model is a Transformer-LSTM architecture, and the hyperparameter configuration is as follows: the number of LSTM layers is 2, and the hidden layer dimension is 64; the number of Transformer encoder layers is 2, the hidden layer dimension is 64, and the number of attention heads is 4; the output window size is 5 seconds; the training rounds are 1000, and the batch size is 2048.
7. A steam turbine generator unit EH oil pipe vibration state intelligent estimation system characterized by, The system comprises: a data acquisition module for acquiring the oil pipe pressure signal and the primary frequency modulation instruction parameters; a feature extraction module for executing the vibration dynamic characteristic extraction method; a processing module loaded with the time series neural network model for executing the method of claim 1 to output the vibration state estimated value; an early warning module for executing the hierarchical early warning system to generate and output the early warning signal.