Hydroelectric generating set vibration trend prediction method based on LMD-TCN-ECA
The vibration signal of the hydroelectric unit is decomposed and feature extraction through the LMD-TCN-ECA method, which solves the problem of low accuracy when processing nonlinear vibration signals by the existing method, and accurately predicts the vibration trend of the hydroelectric unit and ensures the stable operation of the equipment.
Patent Information
- Application Number
- CN202510236141.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-06-20
AI Technical Summary
The existing vibration prediction methods for hydroelectric units are difficult to extract effective features when processing nonlinear vibration signals, resulting in low prediction accuracy.
The LMD-TCN-ECA-based method is used to decompose the vibration signal into sub-components and residual components through local mean decomposition (LMD), and time series features and weighted reconstruction are extracted using the time convolution network (TCN) and the high-efficiency channel attention module (ECA) to achieve accurate prediction of vibration trends.
This method can effectively retain the degradation characteristics in the original signal, improve the accuracy of vibration trend prediction, and ensure the stable operation and state maintenance of the hydroelectric unit.
Smart Images
Figure CN120180025A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of hydro-generator units, and particularly relates to a vibration trend prediction method for hydro-generator units based on LMD-TCN-ECA. Background Art
[0002] As the core equipment for converting water energy into electrical energy, the safe and stable operation of hydro-generator units directly affects the stability and safety of the equipment. However, during the operation of the units, they are affected by factors such as hydraulics, mechanics, and electromagnetics, which cause abnormal vibrations of the units. Abnormal vibrations of the units often affect the operation safety of the equipment and may even lead to equipment damage or accidents. Therefore, a vibration trend prediction method is needed to predict the vibrations of the units in advance, ensure the maximization of the operation efficiency of hydro-generator units, and avoid performance degradation caused by equipment problems.
[0003] Common vibration prediction methods for hydro-generator units mainly include prediction methods based on physical models and prediction methods based on data-driven. Prediction methods based on physical models mainly rely on the understanding of the physical characteristics and working principles of the units, and predict the vibration trend by establishing a mathematical model; prediction methods based on data-driven mainly rely on the vibration data collected from actual operation, and conduct predictions through data analysis and machine learning techniques.
[0004] Prediction methods based on physical models mainly establish a detailed mathematical model according to information such as the physical structure, material properties, and working state of the units. Establishing and calculating the physical model may be relatively complex, requiring high professional knowledge and computing resources. The accuracy of the model also depends on the correct understanding and precise modeling of all relevant factors of the system. Problems such as its complexity, real-time performance, and dependence on parameter accuracy limit its wide use in practical applications.
[0005] Existing mainstream data-driven prediction methods are mainly divided into traditional machine learning and deep learning methods. Traditional machine learning methods need to rely on expert experience for feature selection and extraction, and cannot handle high-dimensional data and data with non-linear relationships well. Deep learning methods have powerful automatic feature learning and complex pattern processing capabilities in vibration trend prediction, and can extract useful information from a large amount of data to complete relatively complex prediction tasks. However, in the actual working environment, hydro-generator units are affected by various factors, and the vibration signals of the units show non-linear characteristics. It is particularly important to extract effective features from the existing non-linear vibration signals to complete the vibration trend prediction task of the units. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to provide a vibration trend prediction method for hydropower units based on LMD-TCN-ECA. This method can decompose the original monitoring signals of hydropower units, maximize the retention of degradation features in the original signals, and effectively predict the changes in the vibration trends of the units, which has certain reference value for the stable operation and condition maintenance of the units.
[0007] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A vibration trend prediction method for hydropower units based on LMD-TCN-ECA, comprising the following steps: Step1. Data preprocessing: Collect the vibration signals of hydropower units monitored by sensors, perform local mean decomposition (LMD) on the data. According to the complexity and variation law of the signal itself, a complex multi-component signal is gradually decomposed into the sum of several sub-components and a residual component through multiple loop iterations. The decomposed sub-sequences and the residual component are used as the input quantities of the prediction model, and then the sequence is normalized by the maximum and minimum values, and the data is scaled to between [0, 1]. Step2. Vibration trend prediction of hydropower units: Establish a TCN-ECA prediction model, use the temporal convolutional network (TCN) to extract the time-series features in the components, and then use the efficient channel attention module (ECA) to reconstruct the features with weights to obtain the prediction indicators of each component. Then, the prediction indicators are respectively inverse-normalized and weighted and superimposed to obtain the final vibration trend prediction result. Step3. Model performance comparison: Input the actual monitoring data into the model for verification, compare the prediction results with the actual values, calculate the evaluation index values, and select other different models for comparison to prove the effectiveness of the proposed model.
[0008] In the above Step1, the steps of performing local mean decomposition (LMD) on the vibration signals of hydropower units are as follows: Step1.1. According to the original signal, calculate the local mean function between two adjacent extreme points and the local envelope function between the extreme points. Step1.2. Smooth the local mean function and the local envelope function by the moving average method to obtain the local mean function and the local envelope function , and calculate the zero-mean function and perform demodulation to obtain the demodulation function. Step1.3. Obtain the local envelope estimation function of the demodulation function, and judge whether the local envelope estimation function is 1. When the condition is satisfied, obtain the pure frequency modulation signal ; if the condition is not satisfied, then the demodulation function Iteratively calculate by repeating the above steps for the original signal until the calculation results in a pure frequency-modulated signal; obtain the desired envelope signal function by multiplying all local envelope estimation functions, and obtain the original signal by multiplying the envelope signal and the pure frequency-modulated signal to obtain the first PF component of the original signal Step1.4. Extract from the original signal to obtain the remaining component . When the remaining component does not satisfy the condition of being a constant or strictly monotonic, regard as the original signal and repeat the above process k times until becomes a constant or a monotonic function. At this time, the original signal is decomposed into k PF components and 1 remaining component .
[0009] The specific process of the above Step1.1 is as follows: Find all extreme points in the original signal. The local average function between two adjacent extreme points is calculated as: ; The local envelope function between extreme points is calculated as: ; In the formula, is the i-th extreme point in the signal; is the (i + 1)-th extreme point in the signal.
[0010] The zero-mean function in the above Step1.2 is: ; Demodulate the zero-mean function to obtain the demodulation function as: .
[0011] The envelope signal function in the above Step1.3 is: ; The first PF component: .
[0012] The original signal in the above Step1.4 is: .
[0013] In the above Step1, the formula for normalizing the maximum and minimum values of the decomposed signal is as follows: ; Among them, and are the data before and after normalization respectively, and are the minimum and maximum values of the data respectively.
[0014] In the above Step2, the specific process of using the Temporal Convolutional Network (TCN) to extract the time series features in the component is as follows: The TCN network contains multiple TCN residual blocks. In the residual block, the data is calculated through one-dimensional dilated causal convolution. For the input of one-dimensional sequence, the receptive field can be expanded by the filter coefficient k and the dilation coefficient d of the convolutional kernel. Then the dilated convolution operation is: ; In the formula, represents the dilated convolution operation, represents the sequence data, represents the filter function, n is the length of the input sequence data, and i is the i-th data point in the input sequence data.
[0015] In the above Step2, the process of using the Efficient Channel Attention (ECA) module to perform weighted reconstruction on the features is as follows: First, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels. The calculation formula of the kernel size is as follows: ; In the formula, represents the odd number closest to t, C is the dimension size, b and are constants, and their values are 1 and 2 respectively; After obtaining the kernel size k, the one-dimensional convolution is applied to the input features to learn the importance of each channel relative to other channels. The process can be expressed by the following formula: ; In the formula, represents the one-dimensional convolution operation with kernel size k, x represents the input, and y represents the output.
[0016] In the above Step2, the steps for denormalizing the prediction results of each component are as follows: ; Among them, and represent the data before and after denormalization respectively.
[0017] A vibration trend prediction method for hydropower units based on LMD-TCN-ECA is provided by the present invention. Aiming at the problem that traditional methods cannot effectively extract useful features from the non-linear vibration signals of hydropower units, the vibration data is first decomposed by LMD to retain the degradation features in the data, and then the TCN is used to deeply mine the time series features in the decomposed sequence. The ECA is selectively introduced and the output weights of different channels are adjusted to enhance the network's attention to important features, and finally the accurate prediction of the vibration trend of the unit is realized to ensure the safe and stable operation of the unit.
[0018] It has the following beneficial effects: a. The present invention introduces the LMD method, which can decompose the complex non-linear vibration signals of hydropower units into the sum of several components, extract the characteristic information in different frequency bands of the original signal, and the TCN-ECA model can predict the decomposed signal, fully extract the important time series features in the decomposed sequence, and improve the prediction accuracy.
[0019] b. The present invention can accurately predict the vibration trend of hydropower units, effectively prevent the excessive amplitude of the units, and provide a reference basis for the condition-based maintenance of the units. Description of the Drawings
[0020] The present invention will be further described below in conjunction with the drawings and embodiments: Figure 1 is the schematic diagram of the step flow of the present invention; Figure 2 is the LMD decomposition diagram; Figure 3 is the TCN structure diagram; Figure 4 is the ECA structure diagram; Figure 5 is the schematic diagram of the real-time prediction process of the present invention; Figure 6 is the schematic diagram of the prediction result of the upper guide X-direction swing of the present invention; Figure 7 is the schematic diagram of the prediction result of the upper guide Y-direction swing of the present invention; Figure 8 is the schematic diagram of the prediction result of the water guide X-direction swing of the present invention; Figure 9 is the schematic diagram of the prediction result of the water guide Y-direction swing of the present invention. Detailed Embodiments
[0021] The technical solution of the present invention will be described in detail below in conjunction with the drawings and embodiments.
[0022] Embodiment 1: Combined with Figures 1 to 6 , the implementation and effect verification of the present invention will be described as follows: As Figure 1 shown in the figure, a vibration trend prediction method for hydropower units based on LMD-TCN-ECA includes the following steps: Step1. Data preprocessing: Collect the vibration signals of hydropower units monitored by sensors, perform local mean decomposition (LMD) on the data. According to the complexity and variation law of the signal itself, a complex multi-component signal is gradually decomposed into the sum of several sub-components and a residual component through multiple loop iterations. The decomposed sub-sequences and residual components are used as the input variables of the prediction model, and then the sequence is normalized by the maximum and minimum values, scaling the data to the range of [0, 1]; Step2. Vibration trend prediction of hydropower units: Establish a TCN-ECA prediction model, use the temporal convolutional network (TCN) to extract the time series features in the components, and then use the efficient channel attention (ECA) module to reconstruct the features with weights to obtain the prediction indicators of each component. Then, the prediction indicators are denormalized respectively and weighted and superimposed to obtain the final vibration trend prediction result; Step3. Model performance comparison: Input the actual monitored data into the model for verification, compare the prediction results with the actual values, calculate the evaluation index values, and select other different models for comparison to prove the effectiveness of the proposed model.
[0023] In the above Step1, the steps for performing local mean decomposition (LMD) on the vibration signals of hydropower units are as follows: Step1.1. According to the original signal, calculate the local average function between two adjacent extreme points and the local envelope function between the extreme points; Step1.2. Smooth the local average function and the local envelope function by the moving average method to obtain the local average function and the local envelope function , and calculate the zero-mean function and perform demodulation to obtain the demodulation function; Step1.3. Obtain the local envelope estimation function of the demodulation function, and judge whether the local envelope estimation function is 1. When the condition is satisfied, obtain the pure frequency modulation signal ; if the condition is not satisfied, then use the demodulation function as the original signal and repeat the above steps for iterative calculation until the calculation is a pure frequency modulation signal; by multiplying all the local envelope estimation functions, obtain the required envelope signal function, and multiply the envelope signal and the pure frequency modulation signal to obtain the first PF component of the original signal ; Step1.4. Extract from the original signal to obtain , getting the remaining component . When the remaining component does not meet the condition of being a constant or strictly monotonic, consider as the original signal and repeat the above process k times until becomes a constant or a monotonic function. At this time, the original signal is decomposed into k PF components and 1 remaining component .
[0024] The specific process of the above Step1.1 is as follows: Find all the extreme points in the original signal. The local average function between two adjacent extreme points is calculated as: ; The local envelope function between the extreme points is calculated as: ; In the formula, is the i-th extreme point in the signal; is the (i + 1)-th extreme point in the signal.
[0025] The zero-mean function in the above Step1.2 is: ; Demodulate the zero-mean function to obtain the demodulation function as: .
[0026] The envelope signal function in the above Step1.3 is: ; The first PF component: .
[0027] The original signal in the above Step1.4 is: .
[0028] In the above Step1, the formula for normalizing the maximum and minimum values of the decomposed signal is as follows: ; Among them, and are the data before and after normalization respectively, and are the minimum and maximum values of the data respectively.
[0029] In the above Step 2, the specific process of using the Temporal Convolutional Network (TCN) to extract the time series features in the components is as follows: The TCN network contains multiple TCN residual blocks. In the residual blocks, one-dimensional dilated causal convolution is used to calculate the data. For the input of a one-dimensional sequence, the convolutional kernel can expand the receptive field through the filter coefficient k and the dilation coefficient d. Then the dilated convolution operation is: ; In the formula, represents the dilated convolution operation, represents the sequence data, represents the filter function, n is the length of the input sequence data, and i is the i-th data point in the input sequence data.
[0030] In the above Step 2, the process of using the Efficient Channel Attention (ECA) module to perform weighted reconstruction on the features is as follows: First, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels. The calculation formula for the kernel size is as follows: ; In the formula, represents the odd number closest to t, C is the dimension size, and b and are constants, with values of 1 and 2 respectively; After obtaining the kernel size k, the one-dimensional convolution is applied to the input features to learn the importance of each channel relative to other channels. The process can be expressed by the following formula: ; In the formula, represents the one-dimensional convolution operation with a kernel size of k, x represents the input, and y represents the output.
[0031] In the above Step 2, the steps for anti-normalizing the prediction results of each component are as follows: ; Among them, and represent the data before and after anti-normalization respectively.
[0032] Embodiment 2: As shown in the appendix Figure 1 The present invention provides a vibration trend prediction method for a hydropower unit based on LMD-TCN-ECA, including the following steps: Collect the vibration signals of the upper guide of the hydropower unit monitored by the sensor, perform LMD decomposition on the data, gradually decompose the complex multi-component signal into several sub-components and a residual component through multiple loop iterations, normalize the maximum and minimum values of the above sub-components, and input them into the TCN-ECA model. After the prediction indexes of each component are denormalized, they are weighted and superimposed to obtain the final vibration trend prediction result. Compare the predicted value with the actual value, and use the vibration data of different parts of the unit for experiments to evaluate the prediction performance of the model.
[0033] Specifically, acceleration sensors are installed in the X and Y directions of the upper guide and water guide bearing housings to monitor the vibration signals of the upper guide bearing. Compared with other sensors, acceleration sensors have the characteristics of high precision, small size, long life, easy installation, and good stability.
[0034] Furthermore, perform LMD decomposition on the monitored data. LMD can decompose a complex multi-component signal into the sum of several sub-components according to the complexity and variation law of the signal itself, so that characteristic information can be extracted from different frequency bands of the original signal. The decomposed sub-signals are as shown in the appendix Figure 2 and the calculation process is as follows: a. Find all the extreme points in the original signal, and calculate the local average function between two adjacent extreme points as:
[0035] Calculate the local envelope function between the extreme points as:
[0036] In the formula, is the i-th extreme point in the signal; is the (i + 1)-th extreme point in the signal.
[0037] b. Smooth the local average function and the local envelope function by the moving average method to obtain the local average function and the local envelope function , and calculate the zero-mean function as:
[0038] Demodulate the zero-mean function to obtain the demodulation function as:
[0039] c. Find the local envelope estimation function of the demodulation function , and judge whether the local envelope estimation function is 1. When the condition is satisfied, obtain the pure frequency modulation signal ; if the condition is not satisfied, then the demodulation function Repeat the above steps for the original signal to perform iterative calculations until the calculation results in a pure frequency modulation signal. By multiplying all the local envelope estimation functions, the desired envelope signal function is obtained as:
[0040] Through the envelope signal and the pure frequency modulation signal multiply, the first PF component of the original signal can be obtained:
[0041] d. Extract from the original signal to obtain the remaining component . When the remaining component does not satisfy the condition of being a constant or strictly monotonic, regard as the original signal and repeat the above process k times until becomes a constant or a monotonic function. At this time, the original signal is decomposed into k PF components and 1 remaining component . The original signal is:
[0042] Furthermore, perform min-max normalization on the decomposed sub-components, compress the data to between [0,1], which can effectively eliminate the influence of abnormal data and enable comparison between data indicators. The steps of normalization are as follows:
[0043] where, and are the data before and after normalization respectively, and are the minimum and maximum values of the data respectively.
[0044] Furthermore, input the decomposed sub-signals into the TCN model respectively. By increasing the number of network layers, changing the dilation coefficient and the size of the filter to expand its receptive field, the model can extract historical information more flexibly. The TCN model structure is as shown in Appendix Figure 4 . The TCN network contains multiple TCN residual blocks. In the residual block, data is mainly calculated through one-dimensional dilated causal convolution. For the input of one-dimensional sequences, the convolutional kernel can expand the receptive field through the filter coefficient k and the dilation coefficient d. Then the dilated convolution operation is:
[0045] In the formula, Denotes the dilated convolution operation, Denotes the sequence data, Denotes the filter function, where n is the length of the input sequence data and i is the i-th data point in the input sequence data.
[0046] Furthermore, the time series features extracted by the TCN model are input into the ECA module, and the one-dimensional convolution is used to dynamically select the information of each channel, enhancing the network's attention to the features of important channels. The ECA structure is as shown in the appendix Figure 5 as follows, and its calculation process is as follows: First, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels. The calculation formula of the kernel size is as follows:
[0047] In the formula, denotes the odd number closest to t, C is the dimension size, and b and are constants with values of 1 and 2 respectively.
[0048] After obtaining the kernel size k, the one-dimensional convolution is applied to the input features to learn the importance of each channel relative to other channels. The process can be expressed by the following formula:
[0049] In the formula, denotes the one-dimensional convolution operation with kernel size k, x represents the input, and y represents the output.
[0050] Furthermore, the inverse normalization is performed on the prediction results of each component to better compare the predicted values with the actual values. The specific steps are as follows:
[0051] Among them, and represent the data before and after inverse normalization respectively.
[0052] Furthermore, the degradation indicators extracted by the model are stacked together to obtain the result. The calculation process is as follows:
[0053] Among them, is the prediction result, is the n-th degradation indicator.
[0054] Furthermore, the process of real-time prediction of the unit vibration signal is as shown in the appendix Figure 5As shown in the figure, the collected dataset is decomposed by LMD and normalized, and then divided into a training set and a test set. The training set is trained, and the optimal model parameters are saved. The test set is imported into the trained model for vibration trend prediction. Finally, the predicted values are compared with the actual values, and repeated experiments are carried out using the vibration data of different components of the unit to evaluate the prediction performance of the model.
[0055] Example verification and analysis: The effectiveness of the present invention is verified below through experimental analysis of the vibration data in the X and Y directions of the upper guide bearing of Unit 5 of a hydropower station in China. Excluding the data with zero vibration during the shutdown process, taking the peak-to-peak value of the swing of the upper guide bearing during the steady operation of the unit as the research object, one data point is collected every 1 hour, and a total of 5,500 sample points are collected. The first 5,000 sample points are selected for training, and the last 500 sample points are used for testing. When training, the root mean square function (MSE) is selected as the loss function, the Adam algorithm is used as the optimizer, the training batch size is 64, the learning rate is 0.001, and the number of epochs is 300. The final number of decomposed sub-components by LMD is 8, the number of TCN residual blocks is 4, the output dimension size of the dilated convolution is 8, and the dropout rate is 0.3. At the same time, the mean absolute error (MAE) and the root mean square error (RMSE) are used as indicators to evaluate the performance of the model. The smaller the values of MAE and RMSE, the better the prediction effect of the model. The calculation formulas are as follows:
[0056]
[0057] Among them, is the total number of samples, represents the actual remaining life percentage of the i-th sample, represents the predicted value of the remaining life of the i-th sample.
[0058] The predicted results of the vibration trend of the hydropower unit are as Figure 6 shown. It can be seen from the figure that the numerical values of the vibration trends in the X and Y directions of the upper guide and water guide of the unit obtained by model prediction are basically consistent with the actual values, indicating that the method of the present invention can effectively predict the vibration trend of the unit. In addition, three groups of models are selected for comparison to evaluate the model performance. After conducting experiments using the swing data in the X direction of the upper guide respectively, the prediction results are shown in Table 1. It can be seen from the table that the prediction accuracy of the present invention is better than that of other models, proving that the method of the present invention has certain advantages in the vibration trend prediction of hydropower units.
[0059] Table 1 Comparison of prediction results
Claims
1. A method for predicting vibration trend of hydropower units based on LMD-TCN-ECA, characterized in that: The following steps are involved: Step 1, data preprocessing: collect the vibration signals of the hydropower unit monitored by the sensor, perform local mean decomposition (LMD) on the data, and gradually decompose a complex multi-component signal into the sum of several sub-components and a residual component through multiple loop iterations according to the complexity and change law of the signal itself. The sub-sequences and residual components obtained after decomposition are used as the input of the prediction model, and then the sequence is normalized to the maximum and minimum values, and the data is scaled to [0,1]; Step 2, Hydropower unit vibration trend prediction: Establish a TCN-ECA prediction model, use the time convolution network TCN to extract the time series features in the components, and then use the efficient channel attention module ECA to weighted reconstruct the features to obtain the prediction indicators of each component, and then denormalize the prediction indicators respectively, and weighted superposition to obtain the final vibration trend prediction results; Step 3, model performance comparison: input the actual monitoring data into the model for verification, compare the predicted results with the actual values, calculate the evaluation index values, and select other different models for comparison to prove the effectiveness of the proposed model.
2. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 1 is characterized in that: In the aforementioned Step 1, the steps of performing local mean decomposition (LMD) on the vibration signal of the hydropower unit are as follows: Step 1.1, according to the original signal, calculate the local average function between two adjacent extreme points and the local envelope function between extreme points; Step 1.2: Smooth the local average function and the local envelope function by moving average method to obtain the local average function and the local envelope function , and calculate the zero mean function and demodulate it to obtain the demodulation function; Step 1.3, find the demodulation function The local envelope estimation function , judge whether the local envelope estimation function is 1. When the condition is met, a pure FM signal is obtained. ; If the condition is not met, the demodulation function Repeat the above steps as the original signal for iterative calculation until the calculation Until it is a pure FM signal; by multiplying all the local envelope estimation functions, the required envelope signal function is obtained, and the envelope signal and pure FM signal Multiply them to get the original signal The first PF component of Step 1.4, from the original signal Extract , and obtain the remaining component , when the residual component does not satisfy the condition of being constant or strictly monotonic, is regarded as the original signal, and the above process is repeated k times until becomes a constant or a monotonic function, then the original signal is decomposed into k PF components and 1 residual component .
3. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 2 is characterized in that: The specific process of Step 1.1 is as follows: Find all extreme points in the original signal and calculate the local average function between two adjacent extreme points: ； The local envelope function between extreme points is calculated as: ； In the formula, is the i-th extreme point in the signal; is the i+1th extreme point in the signal.
4. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 3 is characterized in that: The zero mean function in Step 1.2 is: ； Demodulate the zero mean function and get the demodulation function as follows: 。 5. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 4 is characterized in that: The envelope signal function in Step 1.3 is: ； The first PF component: 。 6. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 5 is characterized in that: The original signal in Step 1.4 for: 。 7. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 6 is characterized in that: In the above Step 1, the formula for normalizing the maximum and minimum values of the decomposed signal is as follows: ； in, and are the data before and after normalization, and are the minimum and maximum values of the data respectively.
8. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 7 is characterized in that: In the above Step 2, the specific process of using the time convolution network TCN to extract the time series features in the component is as follows: The TCN network contains multiple TCN residual blocks. The residual blocks calculate the data through one-dimensional dilated causal convolution. For the input of one-dimensional sequence, the convolution kernel can expand the receptive field through the filter coefficient k and the dilation coefficient d. The dilated convolution operation is: ； In the formula, represents the dilated convolution operation, represents sequence data, Represents the filter function, n is the length of the input sequence data, and i is the i-th data point in the input sequence data.
9. The method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 8 is characterized in that: In the above Step 2, the process of weighted reconstruction of features using the efficient channel attention module ECA is as follows: First, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels. The calculation formula of the kernel size is as follows: ； In the formula, represents the odd number closest to t, C is the dimension size, b and is a constant, with values of 1 and 2 respectively; After obtaining the kernel size k, a one-dimensional convolution is applied to the input features to learn the importance of each channel relative to other channels; the process can be expressed by the following formula: ； In the formula, represents a one-dimensional convolution operation with a kernel size of k, x represents the input, and y represents the output.
10. A method for predicting vibration trend of a hydropower unit based on LMD-TCN-ECA according to claim 9, characterized in that: In the aforementioned Step 2, the steps of denormalizing the prediction results of each component are as follows: ； in, and Represent the data before and after anti-normalization respectively.