Turbine valve actuator fault diagnosis method and system
By collecting and analyzing the working parameters of steam turbine valve actuators, and using advanced models for prediction and classification, the problem of inaccurate fault diagnosis in existing technologies has been solved, enabling online fault diagnosis and safety improvement of actuators.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN TURBINE
- Filing Date
- 2024-12-20
- Publication Date
- 2026-05-15
AI Technical Summary
Existing fault diagnosis methods for steam turbine valve actuators are insufficient for accurately identifying the cause of faults, and cannot achieve condition-based maintenance or even fault prediction.
By collecting the current working parameters of the actuator, using the actuator motion state prediction model and state classification model, combined with the Informer model, Transformer model or LSTM model to predict the working parameters, constructing a one-hot encoded dataset, training a multilayer perceptron (MLP) model, random forest model or XGBoost model to perform state classification, and realizing online fault diagnosis.
It enables accurate judgment of the motion status of the actuator and fault classification, ensuring the safe operation of the unit, early detection of potential system safety hazards, and improving the safety of steam valve operation.
Smart Images
Figure CN119756842B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for identifying steam turbine faults. Background Technology
[0002] Steam valves are key components of steam turbines. Their opening and closing controls the supply and cutoff of steam to the turbine, thus regulating its power output. Each steam valve has an actuator that controls its operation. Using high-pressure fire-resistant oil as the medium, the actuator receives commands from the DEH control system and adjusts the opening of the turbine's inlet valves. The actuator's operating status directly affects the turbine's operation.
[0003] Current fault diagnosis methods for turbine valves and actuators in power plants mainly combine subjective diagnosis with model-based diagnosis. This relies on on-site inspections and a limited number of system measurement points for fault diagnosis. This approach requires a high level of expertise from personnel, and due to the hidden and diverse nature of hydraulic system faults, existing diagnostic methods struggle to accurately pinpoint the root cause of faults. Furthermore, they cannot achieve condition-based maintenance or even fault prediction. Summary of the Invention
[0004] The purpose of this invention is to overcome the problems of existing fault diagnosis methods for steam turbine valve actuators, which are difficult to accurately find the cause of the fault and cannot achieve condition-based maintenance or even fault prediction. This invention provides a fault diagnosis method and system for steam turbine valve actuators.
[0005] This invention provides a method for diagnosing faults in a steam turbine valve actuator, comprising the following steps:
[0006] Collect the current operating parameters of the actuator;
[0007] The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke;
[0008] Input the current working parameters into the actuator motion state prediction model to predict and output the actuator motion state in the future time period T1;
[0009] The actuator's motion status includes normal actuator operation, actuator spring failure, and other actuator abnormalities;
[0010] After obtaining three or more consecutive results indicating actuator spring failure or other actuator abnormalities, the actuator is determined to be faulty.
[0011] Furthermore, historical data of working parameters for time period T2 were collected to construct the first dataset;
[0012] Collect historical data of working parameters and corresponding actuator motion states for time period T3 to construct an initial second dataset;
[0013] The working parameter prediction model was trained using the first dataset; and the working parameter prediction data for time period T3 was obtained by using the trained working parameter prediction model.
[0014] By using time alignment, the historical data of the working parameters for time period T3 in the initial second dataset are replaced with the predicted data of the working parameters for time period T3, while keeping the corresponding actuator motion state unchanged; thus, the reconstructed second dataset is obtained.
[0015] The state classification model was trained using the reconstructed second dataset; the motion state prediction model of the actuator was then trained.
[0016] Furthermore, the process of constructing the initial second dataset also includes:
[0017] One-hot encoding is performed on different motion states of the actuator.
[0018] Furthermore, the working parameter prediction model includes the Informer model, the Transformer model, or the LSTM model.
[0019] Furthermore, state classification models include multilayer perceptron (MLP) models, random forest models, or XGBoost models.
[0020] This invention also provides a fault diagnosis system for steam turbine valve actuators, including a working parameter acquisition module and an actuator motion state prediction model:
[0021] The working parameter acquisition module is used to collect the current working parameters of the actuator;
[0022] The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke;
[0023] The actuator motion state prediction model includes a working parameter prediction model, a state classification model, and a continuous error counter;
[0024] The working parameter prediction model is used to predict and output the working parameters in the future time period T1, taking the current working parameters as input.
[0025] A state classification model is used to predict and output the motion state of the actuator within a future time period T1, taking the working parameters within the future time period T1 as input.
[0026] The actuator's motion status includes normal actuator operation, actuator spring failure, and other actuator abnormalities;
[0027] A continuous error counter is used to determine an actuator malfunction after obtaining three or more results indicating actuator spring failure or other actuator abnormalities.
[0028] Furthermore, the motion state prediction model of the actuator is trained using the training module:
[0029] The training module for the actuator motion state prediction model includes:
[0030] The dataset construction unit is used to collect historical data of working parameters for time period T2 to construct the first dataset; and to collect historical data of working parameters for time period T3 and the corresponding motion state of the actuator to construct the initial second dataset.
[0031] The working parameter prediction model training and prediction unit is used to train the working parameter prediction model using the first dataset; and to predict the working parameters for time period T3 using the trained working parameter prediction model to obtain the working parameter prediction data for time period T3.
[0032] The dataset construction unit is also used to replace the predicted working parameters for time period T3 in the initial second dataset with the predicted working parameters for time period T3 by means of time alignment, while keeping the corresponding actuator motion state unchanged; thus obtaining the reconstructed second dataset.
[0033] The state classification model training unit is used to train the state classification model using the reconstructed second dataset.
[0034] Furthermore, the dataset construction unit also includes a one-hot encoding unit;
[0035] One-hot encoding units are used to perform one-hot encoding on different motion states of the actuator during the construction of the initial second dataset.
[0036] Furthermore, the working parameter prediction model includes the Informer model, the Transformer model, or the LSTM model.
[0037] Furthermore, state classification models include multilayer perceptron (MLP) models, random forest models, or XGBoost models.
[0038] The beneficial effects of this invention are:
[0039] The present invention relates to a method and system for diagnosing faults in steam turbine valve actuators.
[0040] 1. The system can assess the motion status of the actuators online and classify faults to ensure the safe operation of the unit;
[0041] 2. This invention can predict the working parameter status of the actuator and detect potential system safety hazards in advance.
[0042] Achieving more accurate fault diagnosis of steam turbine valves and actuators can effectively improve the safety of steam valve operation, thereby better ensuring the safe operation of coal-fired power units. Attached Figure Description
[0043] Figure 1 This is a schematic diagram of the structure of the turbine valve actuator fault diagnosis method and system of the present invention.
[0044] Figure 2 This is a flowchart of the turbine valve actuator fault diagnosis method of the present invention;
[0045] Figure 3 This is a schematic diagram of the fault system of the steam turbine valve actuator of the present invention;
[0046] Figure 4 This is a schematic diagram of the structure of the actuator motion state prediction model in the turbine valve actuator fault diagnosis system of the present invention;
[0047] Figure 5 A comparison chart of actuator safety oil temperature changes predicted by the Informer model and collected locally over the next 24 hours;
[0048] Figure 6 Heatmap of the confusion matrix for the existing Informer model;
[0049] Figure 7 This is a confusion matrix heatmap of the actuator motion state prediction model composed of the Informer model, MLP model, and continuous error counter in this invention. Detailed Implementation
[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention. Specific Implementation Method 1
[0054] The method for diagnosing faults in turbine valve actuators according to this embodiment includes the following steps:
[0055] Collect the current operating parameters of the actuator;
[0056] The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke;
[0057] Input the current working parameters into the actuator motion state prediction model to predict and output the actuator motion state in the future time period T1;
[0058] The actuator's motion status includes normal actuator operation, actuator spring failure, and other actuator abnormalities;
[0059] After obtaining three or more consecutive results indicating actuator spring failure or other actuator abnormalities, the actuator is determined to be faulty.
[0060] Specifically, in this embodiment, appropriate actuator sensors are arranged at suitable locations on the actuator to measure parameters such as the inlet oil temperature, return oil temperature, high-pressure main steam valve hydraulic actuator safety oil temperature, working chamber pressure, non-working chamber pressure, spring force, and actuator stroke.
[0061] The actuator motion state prediction model in this embodiment first transmits data collected by sensors to the operating parameter prediction model for training, predicts the operating parameters of the actuator, and verifies the accuracy of the model's predictions. Based on the dataset predicted by the operating parameter prediction model, the overall motion state of the turbine valve actuator is one-hot encoded, and then trained using an MLP to detect common actuator faults.
[0062] Finally, the model diagnostic results are output to the intelligent front end to judge and classify the actuator faults. Specific Implementation Method Two
[0064] This embodiment is a further explanation of embodiment one. In this embodiment, historical data of working parameters for time period T2 are collected to construct the first dataset.
[0065] Collect historical data of working parameters and corresponding actuator motion states for time period T3 to construct an initial second dataset;
[0066] The working parameter prediction model was trained using the first dataset; and the working parameter prediction data for time period T3 was obtained by using the trained working parameter prediction model.
[0067] By using time alignment, the historical data of the working parameters for time period T3 in the initial second dataset are replaced with the predicted data of the working parameters for time period T3, while keeping the corresponding actuator motion state unchanged; thus, the reconstructed second dataset is obtained.
[0068] The state classification model was trained using the reconstructed second dataset; the motion state prediction model of the actuator was then trained.
[0069] The other technical features of this embodiment are exactly the same as those of Embodiment 1.
[0070] Specifically, the detailed process is as follows:
[0071] S1. Collect historical data of working parameters for time period T2 to construct the first dataset:
[0072] S2. Train the working parameter prediction model using the first dataset:
[0073] A time series forecasting model is used for training. Future operational parameter predictions are obtained by using historical operational parameter data.
[0074] S3. Collect historical data of working parameters for time period T3 to construct the initial second dataset:
[0075] While collecting historical working parameter data within the T3 time period, the corresponding labels for that time period are also collected.
[0076] The labels include the normal status of the actuator, the status of the actuator spring failure, and other abnormal statuses of the actuator.
[0077] S4. Predict the working parameters for time period T3 using the working parameter prediction model:
[0078] Utilizing a pre-trained working parameter prediction model: The model uses a portion of the data from the first few days of time period T3 to predict the working parameter prediction data. Assuming that the data for time period T3 includes known historical data from the previous days (day 1 of T3, day 2 of T3, etc.), this data can be used as input to predict the next working parameter (day 10 of T3, day ... 11 sky……).
[0079] S5. Replace the historical working parameter data for time period T3 with the predicted data using time alignment:
[0080] Replace the working parameters in the initial second dataset (such as the actual data for time period T3) with the predicted data of the working parameters.
[0081] The label remains unchanged, still indicating the operating status of the actuator (e.g., normal, spring failure, other abnormalities).
[0082] S6. Obtain the reconstructed second dataset:
[0083] The second dataset is reconstructed by replacing the historical data of the working parameters for the same time period (T3).
[0084] S7. Use the reconstructed second dataset to train a state classification model: This model can determine the motion state (normal, abnormal, etc.) of the actuator based on the predicted data of the working parameters.
[0085] In the above method, the duration of T3 should be at least equal to, or longer than, T1 to ensure that the model has sufficient data to learn and predict the motion state of the actuator. The time relationship between T2 and T3 does not necessarily have to be T2 first, followed by T3. However, to ensure a temporal relationship between training and validation data and to ensure the effectiveness of model training, T2 should precede T3. Data prior to T3 (i.e., T2 data) can be used to predict the operating parameters of T3. Alternatively, using a portion of the data preceding T3 (i.e., historical data close to T3) yields better results.
[0086] The above process, by combining predicted operating parameters with historical labels, allows the model to better understand the relationship between operating parameters and equipment status. This helps improve the accuracy of operating status predictions, especially when actual data is incomplete. Predictions based on historical data can simulate future scenarios, reducing human intervention. Reducing reliance on data through operating parameter predictions decreases the need for large amounts of historical data, effectively supplementing predictive data, particularly in environments where obtaining actual data is difficult. Specific Implementation Method 3
[0088] This embodiment is a further explanation of embodiment two. In this embodiment, the process of constructing the initial second dataset also includes:
[0089] One-hot encoding is performed on different motion states of the actuator.
[0090] The other technical features of this embodiment are exactly the same as those of Embodiment 2. Specific Implementation Method Four
[0092] This embodiment is a further explanation of embodiment two or three. In this embodiment, the working parameter prediction model includes an Informer model, a Transformer model, or an LSTM model.
[0093] The other technical features of this embodiment are exactly the same as those of Embodiment Two or Three.
[0094] Specifically, the Informer model is preferred for predicting working parameters, as it is particularly suitable for predicting long-term series data, especially for trend analysis and future forecasting over long periods.
[0095] However, the Transformer model or LSTM model can be used as an alternative. Detailed Implementation Method Five
[0097] This embodiment is a further explanation of embodiment two or three. In this embodiment, the state classification model includes the multilayer perceptron (MLP) model, the random forest model, or the XGBoost model.
[0098] The other technical features of this embodiment are exactly the same as those of Embodiment Two or Three.
[0099] Specifically, the preferred state classification model is the MLP (Multilayer Perceptron) model, which is used for classification problems.
[0100] However, random forest or XGBoost models can be used as alternatives. Specific Implementation Method Six
[0102] The turbine valve actuator fault diagnosis system of this embodiment includes a working parameter acquisition module and an actuator motion state prediction model:
[0103] The working parameter acquisition module is used to collect the current working parameters of the actuator;
[0104] The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke;
[0105] The actuator motion state prediction model includes a working parameter prediction model, a state classification model, and a continuous error counter;
[0106] The working parameter prediction model is used to predict and output the working parameters in the future time period T1, taking the current working parameters as input.
[0107] A state classification model is used to predict and output the motion state of the actuator within a future time period T1, taking the working parameters within the future time period T1 as input.
[0108] The actuator's motion status includes normal actuator operation, actuator spring failure, and other actuator abnormalities;
[0109] A continuous error counter is used to determine an actuator malfunction after obtaining three or more results indicating actuator spring failure or other actuator abnormalities.
[0110] Specifically, such as Figure 1 , 3 As shown in Figure 4, in this embodiment, appropriate actuator sensors are arranged at suitable locations on the actuator. The inlet oil temperature sensor 1, return oil temperature sensor 2, and safety oil temperature sensor 3 are used to measure the inlet oil temperature, return oil temperature, and safety oil temperature of the high-pressure main steam valve hydraulic actuator, respectively. The working chamber pressure sensor 4 and the non-working chamber pressure sensor 5 are used to measure the working chamber and non-working chamber pressures of the actuator, respectively. The spring force of the actuator is measured by the force gauge 6, and the stroke and other parameters of the actuator are measured by the displacement sensor 7.
[0111] The collected actuator operating parameters are then fed into the actuator motion state prediction model for motion state prediction. First, the data collected by the sensors is transmitted to the operating parameter prediction model for training, predicting the actuator operating parameters and verifying the accuracy of the model's predictions. Based on the dataset predicted by the operating parameter prediction model, the overall motion state of the turbine valve actuator is one-hot encoded, and then the MLP is trained on common actuator faults.
[0112] Then, the obtained model diagnostic results are output to the intelligent front end to judge and classify the actuator faults. Detailed Implementation Method Seven
[0114] This embodiment is a further explanation of embodiment six. In this embodiment, the motion state prediction model of the actuator is trained by the training module:
[0115] The training module for the actuator motion state prediction model includes:
[0116] The dataset construction unit is used to collect historical data of working parameters for time period T2 to construct the first dataset; and to collect historical data of working parameters for time period T3 and the corresponding motion state of the actuator to construct the initial second dataset.
[0117] The working parameter prediction model training and prediction unit is used to train the working parameter prediction model using the first dataset; and to predict the working parameters for time period T3 using the trained working parameter prediction model to obtain the working parameter prediction data for time period T3.
[0118] The dataset construction unit is also used to replace the predicted working parameters for time period T3 in the initial second dataset with the predicted working parameters for time period T3 by means of time alignment, while keeping the corresponding actuator motion state unchanged; thus obtaining the reconstructed second dataset.
[0119] The state classification model training unit is used to train the state classification model using the reconstructed second dataset.
[0120] The other technical features of this embodiment are exactly the same as those of Embodiment Six.
[0121] Specifically, the working parameter prediction model is first trained. Taking the actuator oil pipeline temperature as an example, the high-pressure main steam valve hydraulic actuator safety oil temperature collected by the sensor is input into the working parameter prediction model to obtain the predicted value of the main steam valve safety oil temperature for the next 28 days. The predicted temperature is then compared with the actual temperature value collected by the local sensor, as shown in the comparison graph. Figure 5 As shown, the predicted temperature trend is basically consistent with the actual temperature trend. This demonstrates that the Informer model's prediction of future data for the actuator is quite accurate.
[0122] Based on the dataset predicted by the working parameter prediction model, the overall motion state of the turbine valve actuator is encoded using one-hot encoding. Taking spring failure, a common actuator fault, as an example, the model is trained. A multilayer perceptron (MIP) model is used to classify the overall motion state of the actuator into three states: normal operation, spring failure, and other abnormalities. After multiple training iterations, the accuracy rate on the training set is 99.38%, and the accuracy rate on the test set is 99.4%.
[0123] Considering the inherent errors in model predictions, and the potential for serious consequences from accidental erroneous predictions that the actuator's operating status is malfunctioning, the model was optimized by introducing a continuous error counter. This counter is set to indicate an actuator malfunction after three consecutive (or other predetermined) predictions of an abnormal operating status. The confusion matrix heatmaps of the prediction models before and after the improvement are shown below. Figure 6 and Figure 7 As shown, the improved model makes more accurate predictions.
[0124] Once the working parameter prediction model is established, the working status of the actuator can be determined online, according to... Figure 2 The workflow can diagnose the causes of actuator malfunctions.
[0125] It can also be input into the background along with other measuring points of the steam turbine EH system, and comprehensively analyzed through the big data analysis platform to further improve the system's fault diagnosis capabilities and has a certain degree of scalability. Detailed Implementation Method Eight
[0127] This embodiment is a further explanation of embodiment seven. In this embodiment, the dataset construction unit also includes a one-hot encoding unit.
[0128] One-hot encoding units are used to perform one-hot encoding on different motion states of the actuator during the construction of the initial second dataset.
[0129] The other technical features of this embodiment are exactly the same as those of Embodiment Seven. Detailed Implementation Method Nine
[0131] This embodiment is a further explanation of embodiment seven or eight. In this embodiment, the working parameter prediction model includes an Informer model, a Transformer model, or an LSTM model.
[0132] The other technical features of this embodiment are exactly the same as those of embodiment seven or eight. Detailed Implementation Method Ten
[0134] This embodiment is a further explanation of embodiment seven or eight. In this embodiment, the state classification model includes a multilayer perceptron (MLP) model, a random forest model, or an XGBoost model.
[0135] The other technical features of this embodiment are exactly the same as those of embodiment seven or eight.
[0136] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other embodiments.
Claims
1. Turbine valve actuator fault diagnosis method, the steps are as follows: Collect the current operating parameters of the actuator; The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke; Input the current operating parameters into the actuator motion state prediction model and predict and output the future time period. T The motion state of the actuator within 1; The actuator's motion state includes normal actuator operation, actuator spring failure, and other actuator abnormalities; After obtaining three or more consecutive results indicating actuator spring failure or other actuator abnormalities, the actuator is determined to be faulty. Its features are, The process of training the motion state prediction model of the actuator includes: Collection period T The first dataset is constructed using historical data of the working parameters of 2. Collection period T 3. Construct the initial second dataset using historical data of working parameters and corresponding actuator motion states; The working parameter prediction model is trained using the first dataset; and the time period is predicted using the trained working parameter prediction model. T 3. Predicted working parameters; By aligning the time intervals in the initial and second datasets, the time intervals are concentrated. T 3. Replace historical data of working parameters with time periods. T The working parameter prediction data of 3 are obtained while keeping the corresponding actuator motion state unchanged; thus, the reconstructed second dataset is obtained. The state classification model was trained using the reconstructed second dataset; the motion state prediction model of the actuator was then trained.
2. The method for diagnosing faults in turbine valve actuators according to claim 1, characterized in that, The process of constructing the initial second dataset also includes: One-hot encoding is performed on different motion states of the actuator.
3. The method for diagnosing faults in turbine valve actuators according to claim 1 or 2, characterized in that, The working parameter prediction model includes the Informer model, Transformer model, or LSTM model.
4. The method for diagnosing faults in turbine valve actuators according to claim 1 or 2, characterized in that, The state classification model includes the Multilayer Perceptron (MLP) model, the Random Forest model, or the XGBoost model.
5. Turbine valve actuator fault diagnosis system, including a working parameter acquisition module and an actuator motion state prediction model: The working parameter acquisition module is used to collect the current working parameters of the actuator; The current operating parameters of the actuator include the current actuator inlet oil temperature, actuator return oil temperature, actuator working chamber pressure, actuator non-working chamber pressure, actuator spring force, and actuator stroke; The actuator motion state prediction model includes a working parameter prediction model, a state classification model, and a continuous error counter; The operating parameter prediction model is used to predict and output future time periods using the current operating parameters as input. T Operating parameters within 1; The state classification model is used to classify future time periods. T Using the working parameters within 1 as input, predict and output the future time period. T The motion state of the actuator within 1; The actuator's motion state includes normal actuator operation, actuator spring failure, and other actuator abnormalities; The continuous error counter is used to determine the actuator malfunction after obtaining three or more results of actuator spring failure or other actuator abnormalities. Its features are, The actuator motion state prediction model is trained using the following module: The training module for the actuator motion state prediction model includes: Dataset building blocks, used to collect data over time periods. T The first dataset is constructed using historical data of the working parameters; and the collection period is also included. T 3. Construct the initial second dataset using historical data of working parameters and corresponding actuator motion states; The work parameter prediction model training and prediction unit is used to train the work parameter prediction model using the first dataset; and to predict the time period using the trained work parameter prediction model. T 3. Predicted working parameters; The dataset building unit is also used to align time periods in the initial second dataset using a time alignment method. T 3. Work parameter prediction data replaced with time period. T The working parameter prediction data of 3 are obtained while keeping the corresponding actuator motion state unchanged; thus, the reconstructed second dataset is obtained. The state classification model training unit is used to train the state classification model using the reconstructed second dataset.
6. The turbine valve actuator fault diagnosis system according to claim 5, characterized in that, The dataset construction unit also includes a one-hot encoding unit; The one-hot encoding unit is used to perform one-hot encoding on different motion states of the actuator during the construction of the initial second dataset.
7. The turbine valve actuator fault diagnosis system according to claim 5 or 6, characterized in that, The working parameter prediction model includes the Informer model, Transformer model, or LSTM model.
8. The turbine valve actuator fault diagnosis system according to claim 5 or 6, characterized in that, The state classification model includes the Multilayer Perceptron (MLP) model, the Random Forest model, or the XGBoost model.