EMD (empirical mode decomposition) and LSTM (long short term memory) model coupled water turbine top cover fracture risk prediction method and system

Abstract

The invention relates to the field of hydroelectric generating set detection, and discloses a water turbine top cover fracture risk prediction method and system coupled with EMD and LSTM models, vibration and strain states of a water turbine are monitored in real time by arranging a vibration measurement sensor and a strain sensor in a dangerous area of a top cover, a database is constructed, time-frequency domain decomposition is performed, and signal features are extracted. And then, constructing an LSTM model by using the standardized data, and optimizing hyper-parameters of the LSTM model through Bayesian optimization. A rolling prediction method is adopted to predict future vibration and strain data, and a relative root-mean-square error combined with expert scoring is designed to perform early warning. According to the method, higher monitoring timeliness and accuracy can be realized, potential risks can be fed back in time, the processing capability of complex signals is improved, the false alarm rate is reduced, and the safety of the hydropower station is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of hydropower unit detection, and in particular to a method and system for predicting the risk of rupture of a turbine top cover by coupling an EMD and LSTM model. Background Art

[0002] The burst of the turbine cover is one of the important sources of accidents that cause flooding in the powerhouse of a hydropower station. After the turbine cover bursts, pressurized water quickly flows into the pit and flows out into the powerhouse through the turbine, causing serious losses, which can lead to equipment failures or power grid fluctuations. The consequences of a turbine cover burst are irreversible. Therefore, in order to improve the safety of hydropower stations, it is very important to conduct real-time detection and risk prediction of the integrity of the turbine cover.

[0003] However, for the detection and inspection of turbine top covers, traditional manual detection methods or expert database analysis methods require a large amount of manpower, have poor predictive power, and cannot provide timely feedback of danger signals. In addition, turbine top cover signals are multi-band non-stable signals, and the accuracy and reliability of conventional artificial intelligence prediction methods are low. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method and system for predicting the risk of turbine top cover rupture by coupling EMD and LSTM models.

[0005] The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM models includes the following steps:

[0007] S1, collects vibration and strain test data of different positions of the top cover under normal operating conditions of the turbine and builds a database;

[0008] Specifically, vibration sensors and strain sensors are arranged in dangerous areas of the top cover, and vibration and strain tests of the top cover are carried out under the full operating conditions of the turbine. Vibration and strain test data at different positions under normal operating conditions of the turbine are obtained, and a database is constructed; the output signals of the vibration sensors and strain sensors can not only be derived from historical operating data and experimental data, but can also be combined with data collected in real time by an online monitoring system to provide more comprehensive and real-time monitoring information.

[0009] In step S1, the vibration sensors and strain sensors are arranged in dangerous areas determined by finite element analysis, including the connection parts of the top cover, the areas with concentrated stress, and the key nodes where crack propagation may occur.

[0010] S2, decompose the vibration and strain signals of all measuring points in the time-frequency domain through empirical mode decomposition (EMD) to extract the signals of vibration modes and strain change modes of different modes;

[0011] The IMF components after EMD processing are screened and selected according to different frequency characteristics, retaining the modal components that are important for rupture risk prediction and discarding noise and signals in irrelevant frequency bands.

[0012] S3, standardize the decomposed vibration and strain modal signals, and build LSTM modeling based on the standardized sub-datasets;

[0013] The specific method of signal standardization is:

[0014]

[0015] in is the standardized input variable, u p is the vibration or strain signal after EMD modal decomposition, u p,ave is the mean value of the corresponding signal, and σ is the standard deviation of the corresponding signal.

[0016] The LSTM model includes an input layer, an LSTM layer, a fully connected layer, and an output layer; the training set data passes through the input layer, the LSTM layer, the fully connected layer, and the output layer in sequence.

[0017] The input layer of the LSTM model is designed according to the number of features and time series length of different modal signals to ensure that the model can fully capture the time series characteristics of the signal.

[0018] S4, optimizes the hyperparameters of the constructed LSTM through Bayesian optimization;

[0019] Optimize LSTM hyperparameters through Bayesian optimization. The specific steps include:

[0020]

[0021] Among them, P(f) represents the prior probability distribution of parameter f, P(H i |f) represents the likelihood distribution of parameter f, P(f|H i ) represents the conditional probability distribution of parameter f, P(H i ) indicates H i The prior probability distribution of is used, and the validation set is used for Bayesian optimization. The optimization objective function is the minimum value of the mean square error MSE.

[0022] S5, using LSTM to predict the data for a period of time in the future through the rolling prediction method, to obtain the data of different modal components of vibration and strain for a period of time in the future; by performing denormalization and signal reconstruction on the predicted data, the vibration and strain prediction results at different positions in the future are obtained;

[0023] The rolling prediction method uses sliding window technology to dynamically adjust the size and position of the prediction window to adapt to changes in turbine operating conditions.

[0024] Furthermore, in step S5, the specific method of de-standardization is:

[0025]

[0026] S6, calculate the relative root mean square error rRMSE of vibration and strain signals at different positions based on the predicted values ​​and the test values;

[0027] The relative root mean square error is calculated as:

[0028]

[0029] Among them, u i，pre Indicates the signal prediction value obtained based on the LSTM model, u i，test Represents the signal test value, and N represents the number of signal points.

[0030] S7, when the vibration signal or strain signal rRMSE at a certain position i is higher than the threshold C i,1 When the sum of the vibration signal and the strain signal rRMSE at a certain position i is higher than the threshold C i,2 When the vibration signal or strain signal rRMSE at adjacent positions i and j is higher than the threshold C i,3 When the vibration signal and strain signal rRMSE at adjacent positions i and j are higher than the threshold C i,4 Early warning when necessary;

[0031] Threshold C i,1 , threshold C i,2 , threshold C i,3 , threshold C i,4 Can be obtained through expert scoring mechanism.

[0032] S8, outputs the trained LSTM warning model and deploys it through the dynamic link library so that the model can be called by other systems.

[0033] The turbine top cover rupture risk prediction system coupling EMD and LSTM models includes perception layer, transmission layer and application layer.

[0034] The perception layer includes wireless integrated vibration and strain monitoring units, which are responsible for real-time collection of vibration and strain data of the turbine cover. Each strain monitoring unit has an internal integrated data processing module that can perform preliminary filtering and denoising on the original signal to ensure the quality of data transmitted to subsequent layers.

[0035] The transport layer uses the LoRa wireless network to wirelessly transmit multiple sets of monitoring signals to the LoRa gateway node. The data is centrally processed by the LoRa gateway node and then transmitted to the main switch via wired, wireless or the Internet.

[0036] In the application layer, the main switch is responsible for transmitting data between the main server and the monitoring unit. The data received by the main server will be stored and further processed, and then transmitted to the processor. The processor implements the prediction of the risk of rupture of the turbine top cover by calling the data and models in the memory. Among them, the main server analyzes the received data in real time, identifies abnormal signals and marks them to ensure the accuracy of subsequent processing; the processor calls the coupled empirical mode decomposition EMD and LSTM models stored in the memory, performs complex calculations and predictions, and obtains the risk assessment of rupture of the turbine top cover. The processor calls the coupled empirical mode decomposition EMD and LSTM models stored in the memory, performs complex calculations and predictions, and obtains the risk assessment of rupture of the turbine top cover.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention can realize real-time monitoring of the turbine top cover under all working conditions by arranging sensors in the dangerous area of ​​the top cover. Compared with the traditional manual detection method, it has higher timeliness and accuracy, and can timely feedback potential risk signals; through finite element analysis, the location of the potential danger source can be determined in advance, and the possible risk area can be identified in the design stage, so that these areas can be focused on in the subsequent monitoring, which improves the pertinence and effectiveness of monitoring; at the same time, vibration and strain sensors are used, so that the system can monitor the status of the turbine top cover from multiple dimensions, more comprehensively capture the danger signals of different frequency bands, and improve the detection ability of potential fault sources. In addition, the empirical mode decomposition (EMD) technology is used to decompose the vibration and strain signals of the turbine top cover in the time-frequency domain, which can effectively extract the signal characteristics of different modes and improve the processing ability of complex multi-band unstable signals; combined with the long short-term memory recursive neural network (LSTM), deep learning technology is used for data modeling and prediction, which significantly improves the accuracy and reliability of the prediction. Compared with the traditional analysis method, it can better adapt to the dynamically changing working conditions. Finally, setting multiple threshold conditions for early warning can flexibly respond to different locations and situations, reducing false alarm rates and improving safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0040] Figure 1 Schematic diagram of the process of the present invention;

[0041] Figure 2 A training diagram of the LSTM artificial intelligence model of the present invention;

[0042] Figure 3 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0044] like Figure 1 As shown in FIG. 1 , the turbine top cover rupture risk prediction method of coupling EMD and LSTM models includes:

[0045] S1: In the dangerous area of ​​the turbine top cover, several key nodes were identified through finite element analysis, including the connection parts of the top cover, the areas with concentrated stress, and the key points where cracks may propagate. Eight vibration sensors (numbered V1 to V8) and eight strain sensors (numbered S1 to S8) were installed at these locations to achieve multi-dimensional real-time monitoring. When the turbine was operating under full working conditions, vibration and strain signals were continuously collected for 3 months, and the data was accumulated to build a database.

[0046] S2: Perform empirical mode decomposition (EMD) on the signals of V1 to V8 and S1 to S8, extract their intrinsic mode functions (IMFs), and decompose each signal into 6 IMFs and a residual term. According to the frequency characteristics and energy contribution of IMF, IMF1 to IMF4 are retained, and IMF5 to IMF6 and the residual term are discarded to retain the modal components that are critical to the prediction of rupture risk.

[0047] S3: Standardize the vibration and strain modal signals obtained in step S2. The specific standardization formula is as follows:

[0048]

[0049] in is the standardized input variable, u p is the vibration or strain signal after EMD modal decomposition, u p,ave is the mean value of the corresponding signal, and σ is the standard deviation of the corresponding signal.

[0050] According to the standardized sub-datasets, the LSTM model is constructed. The specific process is as follows Figure 2 shown.

[0051] S4: Use Bayesian optimization to optimize the hyperparameters of the LSTM model to minimize the mean squared error (MSE). Adjust the learning rate, number of LSTM units, batch size, and number of layers. Set the learning rate between 0.0001 and 0.01, the number of LSTM units between 100 and 300, the batch size between 64 and 512, and the number of layers between 1 and 4. Find the best hyperparameter combination by minimizing the MSE on the validation set.

[0052] S5: Use the optimized LSTM model to perform rolling forecasts to predict vibration and strain data of different modes within the next 10 minutes. Use the optimized LSTM model to perform rolling forecasts to predict vibration and strain data within the next two weeks. The specific method of denormalization is:

[0053]

[0054] Each modal component is synthesized into vibration and strain prediction results at different locations to form the vibration and strain status of each measuring point in the next two weeks.

[0055] S6: Calculate the relative root mean square error (rRMSE) between the predicted value and the actual test value:

[0056]

[0057] Among them, u i，pre Indicates the signal prediction value obtained based on the LSTM model, u i，test Represents the signal test value, and N represents the number of signal points.

[0058] The test data is transmitted to the LoRa gateway node through the LoRa wireless network, and then transmitted to the main switch through the wired network to achieve real-time data processing.

[0059] S7: Perform early warning judgment based on the calculated rRMSE value. When any of the following conditions is met, the system issues an early warning signal: When the rRMSE of the vibration signal or strain signal at a certain position i is higher than the threshold C i,1When the sum of the vibration signal and the strain signal rRMSE at a certain position i is higher than the threshold C i,2 When the vibration signal or strain signal rRMSE at adjacent positions i and j is higher than the threshold C i,3 When the vibration signal and strain signal rRMSE at adjacent positions i and j are higher than the threshold C i,4 Early warning at that time.

[0060] Determine the threshold C through the expert scoring mechanism i,1 , C i,2 , C i,3 , C i,4 To ensure the accuracy and reliability of early warning. When the early warning conditions are met, the system will send out an alarm through the console and mobile devices simultaneously to notify maintenance personnel to take timely measures.

[0061] S8: Integrate the DLL file into the existing monitoring system, data acquisition system and control system of the hydropower station to realize real-time data sharing and timely transmission of early warning information. Integrate the DLL file into the existing monitoring system, data acquisition system and control system of the hydropower station to realize real-time data sharing and timely transmission of early warning information. Deploy the system on the cloud platform to ensure application adaptability and stability in different scales and environments, and support remote monitoring and management.

[0062] On the other hand, this embodiment also provides a turbine top cover rupture risk prediction system that couples the EMD and LSTM models.

[0063] Reference Figure 3 , Figure 3 The hardware structure diagram of the turbine top cover rupture risk prediction system is shown in FIG. In the embodiment of the present application, the turbine top cover rupture risk prediction system may include a perception layer, a transmission layer and an application layer.

[0064] The perception layer includes wireless integrated vibration and strain monitoring units, which are responsible for collecting vibration and strain data of the turbine cover in real time. These strain monitoring units use highly sensitive sensors that can accurately capture tiny changes, thereby promptly detecting potential rupture risks. Each strain monitoring unit has an integrated data processing module that can perform preliminary filtering and denoising on the original signal to ensure the quality of data transmitted to subsequent layers.

[0065] The transmission layer uses the LoRa wireless network to wirelessly transmit multiple sets of monitoring signals to the LoRa gateway node. LoRa technology has the characteristics of long distance and low power consumption, and is suitable for use in complex environments such as hydropower stations. Data is centrally processed by the LoRa gateway node, and then transmitted to the main switch via wired, wireless or the Internet.

[0066] In the application layer, the main switch is responsible for transmitting data between the main server and the monitoring unit. The data received by the main server will be stored and further processed, and then transmitted to the processor. The processor implements the prediction of the risk of rupture of the turbine top cover by calling the data and models in the memory. Among them, the main server analyzes the received data in real time, identifies abnormal signals and marks them to ensure the accuracy of subsequent processing; the processor calls the coupled empirical mode decomposition (EMD) and LSTM models stored in the memory, performs complex calculations and predictions, and obtains the risk assessment of rupture of the turbine top cover. The processor calls the coupled empirical mode decomposition (EMD) and LSTM models stored in the memory, performs complex calculations and predictions, and obtains the risk assessment of rupture of the turbine top cover.

[0067] Although the present application has described the preferred embodiments in detail, those skilled in the art may modify and adjust these embodiments after understanding the basic innovative concepts. Therefore, the appended claims are intended to cover the preferred embodiments and all changes and modifications that meet the scope of implementation of the present application.

[0068] This article introduces in detail a method and system for predicting the risk of turbine top cover rupture by combining empirical mode decomposition and LSTM model. These embodiments are intended to help understand the method and core ideas of this application; for those skilled in the art, based on the concept of this application, the specific implementation methods and applications may be different. Therefore, the contents of this specification should not be regarded as limiting this application.

Claims

1. A method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM models, characterized in that: The following steps are involved: S1, collects vibration and strain test data of different positions of the top cover under normal operating conditions of the turbine and builds a database; S2, decompose the vibration and strain signals of all measuring points in the time-frequency domain through empirical mode decomposition (EMD) to extract the signals of vibration modes and strain change modes of different modes; S3, standardize the decomposed vibration and strain modal signals, and build LSTM modeling based on the standardized sub-datasets; S4, optimizes the hyperparameters of the constructed LSTM through Bayesian optimization; S5, using LSTM to predict the data for a period of time in the future through the rolling prediction method, to obtain the data of different modal components of vibration and strain for a period of time in the future; by performing denormalization and signal reconstruction on the predicted data, the vibration and strain prediction results at different positions in the future are obtained; S6, calculate the relative root mean square error rRMSE of vibration and strain signals at different positions based on the predicted values ​​and the test values; S7, when the vibration signal or strain signal rRMSE at a certain position i is higher than the threshold C i,1 When the sum of the vibration signal and the strain signal rRMSE at a certain position i is higher than the threshold C i,2 When the vibration signal or strain signal rRMSE at adjacent positions i and j is higher than the threshold C i,3 When the vibration signal and strain signal rRMSE at adjacent positions i and j are higher than the threshold C i,4 Early warning when necessary; S8, outputs the trained LSTM warning model and deploys it through the dynamic link library so that the model can be called by other systems.

2. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S1, vibration sensors and strain sensors are arranged in the dangerous area of ​​the top cover to perform vibration and strain tests on the top cover under the full operating state of the turbine, and obtain vibration and strain test data at different positions under normal operating conditions of the turbine, or build a database in combination with data collected in real time by an online monitoring system.

3. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In the step S1, the vibration sensors and strain sensors are arranged in the dangerous areas determined by finite element analysis, including the connection parts of the top cover, the areas with concentrated stress, and the key nodes where crack propagation may occur.

4. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S2, the IMF components after EMD processing are screened and selected according to different frequency characteristics, and the modal components that are important for rupture risk prediction are retained, while the noise and signals of irrelevant frequency bands are discarded.

5. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S3, the specific method of signal standardization is: In step S5, the specific method of de-standardization is: in is the standardized input variable, u p is the vibration or strain signal after EMD modal decomposition, u p,ave is the mean value of the corresponding signal, and σ is the standard deviation of the corresponding signal.

6. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S3, the LSTM model includes an input layer, an LSTM layer, a fully connected layer, and an output layer; the training set data passes through the input layer, the LSTM layer, the fully connected layer, and the output layer in sequence; The input layer of the LSTM model is designed according to the number of features and time series length of different modal signals to ensure that the model can fully capture the time series characteristics of the signal.

7. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S4, the LSTM hyperparameters are optimized by Bayesian optimization, and the specific steps include: Among them, P(f) represents the prior probability distribution of parameter f, P(H i |f) represents the likelihood distribution of parameter f, P(f|H i ) represents the conditional probability distribution of parameter f, P(H i ) indicates H i The prior probability distribution of is used, and the validation set is used for Bayesian optimization. The optimization objective function is the minimum value of the mean square error MSE.

8. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S5, the rolling prediction method adopts the sliding window technology to dynamically adjust the size and position of the prediction window to adapt to the changes in the operating conditions of the turbine.

9. The method for predicting the risk of turbine top cover rupture by coupling EMD and LSTM model according to claim 1, characterized in that: In step S6, the relative root mean square error is calculated as follows: Among them, u i，pre Indicates the signal prediction value obtained based on the LSTM model, u i，test Represents the signal test value, and N represents the number of signal points.

10. A turbine top cover rupture risk prediction system coupled with EMD and LSTM models, used to implement the turbine top cover rupture risk prediction method coupled with EMD and LSTM models as claimed in any one of claims 1 to 8, characterized in that: Includes perception layer, transport layer and application layer; The perception layer includes a wireless integrated vibration and strain monitoring unit, which is responsible for collecting vibration and strain data of the turbine top cover in real time; each strain monitoring unit has an internal integrated data processing module, which can perform preliminary filtering and denoising on the original signal to ensure the quality of data transmitted to subsequent layers; The transmission layer uses the LoRa wireless network to wirelessly transmit multiple sets of monitoring signals to the LoRa gateway node; the data is centrally processed by the LoRa gateway node, and then transmitted to the main switch via wired, wireless or the Internet; In the application layer, the main switch is responsible for transferring data between the main server and the monitoring unit; the data received by the main server will be stored and further processed before being sent to the processor; The processor predicts the risk of rupture of the turbine top cover by calling the data and models in the memory; the main server analyzes the received data in real time, identifies and marks abnormal signals to ensure the accuracy of subsequent processing; the processor calls the coupled empirical mode decomposition EMD and LSTM models stored in the memory, performs complex calculations and predictions, and obtains a rupture risk assessment for the turbine top cover; the processor calls the coupled empirical mode decomposition EMD and LSTM models stored in the memory, performs complex calculations and predictions, and obtains a rupture risk assessment for the turbine top cover.

Images