Hydroelectric generating set vibration trend prediction method and system based on collaborative noise reduction and feature fusion

Abstract

The invention belongs to the technical field of equipment state monitoring and fault prediction, and discloses a hydroelectric generating set vibration trend prediction method and system based on collaborative noise reduction and feature fusion, and the method comprises the steps: firstly carrying out the two-stage noise reduction and trend reinforcement of an original vibration signal through the collaborative effect of wavelet threshold denoising and singular spectrum analysis; then decomposing the processed signal into a plurality of intrinsic mode function components by using fully adaptive noise set empirical mode decomposition; independently predicting each IMF component by adopting a gating circulation unit trained by an Adam optimizer and a learning rate attenuation strategy; and finally, summarizing prediction results of all components to obtain a final vibration trend. According to the method, the problems of large noise interference, insufficient feature extraction and poor long-term prediction stability in the prior art are effectively solved, the precision and robustness of hydroelectric generating set vibration trend prediction are remarkably improved, and a reliable basis is provided for predictive maintenance of a set.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring and fault prediction technology, specifically to a method and system for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion, and particularly to a hybrid prediction method and system combining collaborative noise reduction, signal decomposition and deep learning, for achieving high-precision and long-term stable prediction of the vibration trend of hydropower units. Background Technology

[0002] As a critical peak-shaving and frequency-regulating device in the power grid, the safe and stable operation of hydropower units is of paramount importance. Unit vibration signals are an important indicator reflecting their operating status; however, these signals typically exhibit strong nonlinearity and non-stationarity, and are easily affected by high-frequency noise interference in the field, making direct trend prediction difficult and inaccurate.

[0003] In existing technologies, single signal processing methods or prediction models have significant limitations. For example, while Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) can effectively decompose non-stationary signals, it is sensitive to noise in the original signal, easily generating spurious mode components and leading to feature extraction bias. Deep learning models such as Gated Recurrent Units (GRUs) perform well in time series prediction, but are prone to gradient decay and error accumulation in long-term predictions. Furthermore, traditional wavelet threshold denoising (WTD) methods rely on experience in selecting wavelet bases and the number of decomposition layers, resulting in poor adaptability; while Singular Spectrum Analysis (SSA) can effectively extract global trends, its ability to capture local transient features is limited.

[0004] Therefore, existing technologies lack a comprehensive solution that can effectively and collaboratively process noise, integrate multi-scale features, and optimize model training to overcome the above-mentioned deficiencies and achieve high-precision prediction of the vibration trend of hydropower units. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method and system for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion, aiming to improve prediction accuracy and long-term stability.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion is provided, the method comprising the following steps: S1 Signal Preprocessing and Collaborative Denoising: The original vibration time series signal of the hydropower unit was obtained by performing wavelet threshold denoising and singular spectrum analysis in sequence to obtain the trend sequence after collaborative denoising. S2 Multiscale Feature Decomposition: Perform fully adaptive noise set empirical mode decomposition on the trend sequence obtained in step S1 after collaborative noise reduction to obtain several intrinsic mode function components. S3 Component Prediction and Model Optimization: For each intrinsic mode function component obtained in step S2, an independent gated recurrent unit prediction model is established and trained and predicted. The training of the gated recurrent unit prediction model adopts the Adam optimizer and combined with the learning rate decay strategy. S4 Result Reconstruction and Output: The prediction results of all intrinsic mode function components obtained in step S3 are superimposed to reconstruct the final vibration trend prediction value.

[0007] As a preferred option, the specific steps of step S1 are as follows: acquire historical vibration time series data of the hydropower unit. First, the original vibration time series signal of the acquired hydropower unit is subjected to wavelet threshold denoising to perform preliminary noise reduction on the original vibration signal. Then, the processed signal is subjected to singular spectrum analysis to extract its global trend component in order to further suppress residual high-frequency noise and strengthen the core trend features, so as to obtain the trend sequence after collaborative noise reduction.

[0008] Preferably, wavelet threshold denoising in step S1 specifically includes the following steps: S1.1 Selecting appropriate wavelet basis functions and decomposition levels: Using the db4 wavelet basis, the original vibration signal is decomposed into 5 levels of wavelets to obtain the high-frequency detail coefficients and low-frequency approximation coefficients of each level. S1.2 Thresholding: The threshold of high-frequency coefficients in each layer is calculated using a general thresholding rule, and a soft thresholding function is selected to quantize the detail coefficients and suppress noise components. S1.3 Signal Reconstruction: Perform inverse wavelet transform on the coefficients of each layer after thresholding to reconstruct the signal after preliminary noise reduction.

[0009] Preferably, the singular spectrum analysis processing in step S1 specifically includes the following steps: S1.4 Embedding: The signal after preliminary noise reduction is converted into a one-dimensional time series, and then the one-dimensional time series of length N is further processed... x 1, x 2, ..., x N The sequence is mapped to a trajectory matrix X, and the embedding window length L is set to 1 / 3 of the original sequence length. S1.5 Singular Value Decomposition: Perform singular value decomposition on the trajectory matrix and calculate the covariance matrix S=XX. T Eigenvalues ​​are obtained by performing eigenvalue decomposition.λ m and eigenvectors U m ; S1.6 Grouping and Reconstruction: Based on the contribution of eigenvalues, select the top... K The main components are grouped together, and the trend component reflecting the global trend of the signal is reconstructed by the diagonal averaging method. This further suppresses high-frequency noise and strengthens the core trend, resulting in a trend sequence after synergistic noise reduction.

[0010] Preferably, step S2 specifically includes the following steps: S2.1 Add adaptive noise: Add a white noise sequence modulated by adaptive white noise amplitude to the trend sequence to be decomposed; S2.2 Empirical Mode Decomposition: Perform empirical mode decomposition on the signal after adding noise to obtain a set of intrinsic mode functions; S2.3 Integrated averaging: Repeat S2.1 and S2.2 for a total of N times, adding a different white noise sequence each time, and integrating and averaging the eigenmode function components of the corresponding order obtained from each decomposition to obtain a final set of eigenmode function components and a residual component. S2.4 Output decomposition results: The final output intrinsic mode function components are arranged from high frequency to low frequency, representing the local characteristics of the signal at different time scales, and the residual components represent the overall trend or baseline of the signal.

[0011] Preferably, the noise figure in the adaptive white noise is set to 0.2, and the number of noise additions is set to 500.

[0012] As a preferred option, the specific steps of step S3 are as follows: for each intrinsic mode function component obtained in step S2, construct and train an independent gated recurrent unit prediction sub-model; during the training process, use the Adam optimizer and combine it with a learning rate decay strategy to optimize the parameters of the gated recurrent unit prediction sub-model to improve training efficiency and model performance.

[0013] Preferably, in the training step of the gated recurrent unit prediction sub-model, the input time step is set to 12, the maximum number of training iterations is set to 200, and the initial learning rate is set to 0.001. The specific process of the learning rate decay strategy is as follows: During model training, when the validation set loss no longer decreases within P consecutive training cycles, the current learning rate is multiplied by the decay factor α (0 < α < 1) to decay until the preset minimum learning rate is reached.

[0014] As a preferred option, step S4 specifically involves: superimposing and summing the prediction results of each gated cyclic unit prediction sub-model for the corresponding intrinsic mode function components to obtain the final vibration trend prediction value; and evaluating the accuracy of the final prediction result using root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination.

[0015] According to another aspect of the present invention, a hydropower unit vibration trend prediction system for implementing the above method is also provided, the system comprising: (1) Data acquisition and collaborative noise reduction module, used to acquire the original vibration signal and perform the wavelet threshold denoising and singular spectrum analysis to obtain the trend sequence after collaborative noise reduction; (2) Signal decomposition module, used to perform empirical mode decomposition and integrate and average the intrinsic mode function components of the corresponding order obtained by each decomposition to obtain several intrinsic mode function components; (3) Parallel prediction module, which contains multiple gated recurrent unit prediction sub-models, used for parallel training and prediction of each intrinsic mode function component; (4) Result synthesis and output module, which is used to superimpose and sum the prediction results of each intrinsic mode function component and output the final vibration trend prediction value; The parallel prediction module has a built-in Adam optimizer and learning rate decay controller, which are used to optimize parameters during model training.

[0016] The data acquisition and collaborative noise reduction module, signal decomposition module, parallel prediction module, and result synthesis and output module are connected in sequence to realize fully automated processing from raw signal acquisition to trend prediction output.

[0017] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: (1) Collaborative noise reduction to improve signal-to-noise ratio: Through the cascaded application of WTD and SSA, a collaborative noise reduction mechanism of "local refinement + global extraction" is formed, which effectively suppresses high-frequency noise and random interference, provides a high-quality data foundation for subsequent decomposition and prediction, and reduces the risk of false modes generated by CEEMDAN decomposition. (2) Multi-scale feature fusion to enhance expressive power: By combining the global trend component extracted by SSA with the local high-frequency detail component obtained by CEEMDAN decomposition, a multi-scale feature fusion strategy was constructed, enabling the model to capture both the long-term evolution law and short-term transient features of vibration signals. (3) Model optimization to improve prediction performance: The Adam optimizer is used in conjunction with the learning rate decay strategy to train the GRU model, which effectively accelerates the model convergence, alleviates the gradient decay problem, and improves the long-term prediction stability and accuracy of the model. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0019] Figure 2 The image shows the original vibration signal in the example.

[0020] Figure 3 This is a graph showing the SSA trend extraction results in the example.

[0021] Figure 4 The diagram shows the CEEMDAN decomposition results in the example.

[0022] Figure 5 This is a comparison chart of the final predicted result and the actual value in the example. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Please see Figure 1 and Figure 2 The present invention provides a method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion, comprising the following steps: S1 Signal Preprocessing and Collaborative Denoising: The original vibration time series signal of the hydropower unit was obtained by performing wavelet threshold denoising and singular spectrum analysis in sequence to obtain the trend sequence after collaborative denoising. The specific steps of step S1 are as follows: First, the historical vibration time series data of the hydropower unit is obtained. The original vibration time series signal of the hydropower unit is processed by WTD. High-frequency noise is initially filtered out by multi-scale decomposition and threshold screening mechanism to achieve preliminary noise reduction of the original vibration signal. Then, the processed signal is processed by SSA to extract and reconstruct the trend of the denoised signal. Through trajectory matrix construction, singular value decomposition and principal component reconstruction, the global trend component with physical meaning is extracted to further suppress residual high-frequency noise and strengthen the core trend features, significantly improve the signal-to-noise ratio, and obtain the trend sequence after collaborative noise reduction.

[0025] In step S1, wavelet threshold denoising (WTD) is performed: First, the original vibration time series signal of the hydropower unit is obtained. x ( t ),like Figure 2 As shown. WTD is used for preliminary noise reduction of the signal. The specific steps are as follows: S1.1 Selecting appropriate wavelet basis functions and decomposition levels: The db4 wavelet is selected as the basis function, and the original vibration signal is decomposed into 5 levels of wavelet decomposition to obtain the high-frequency detail coefficients of each level. d j(k) and low-frequency approximation coefficient a j(k) , where j = 1, 2, ..., 5; S1.2 Thresholding: The threshold of each layer of high-frequency coefficients is calculated using a general thresholding rule, and a soft thresholding function is selected to quantize the detail coefficients and suppress noise components; that is, according to the preset thresholding criterion (general threshold), the detail coefficients of each layer are subjected to soft thresholding to retain the high-energy coefficients that represent the real signal and remove the low-energy coefficients dominated by noise. S1.3 Signal Reconstruction: High-frequency detail coefficients of each layer after thresholding d j(k) and low-frequency approximation coefficient a j(k) Perform inverse wavelet transform to reconstruct the initial denoised signal. x WTD ( t ); In step S1, singular spectral analysis (SSA) is used to process the initial denoised signal. x WTD ( t Singular spectrum analysis is performed to extract global trend components and further suppress residual high-frequency noise. The specific steps are as follows: S1.4 Embedding: The signal after preliminary noise reduction is converted into a one-dimensional time series, and then the one-dimensional time series of length N is further processed... x 1, x 2, ..., x N The sequence is mapped to a trajectory matrix X, and the embedding window length L is set to 1 / 3 of the original sequence length. That is, choose an appropriate window length. L The original time series [ x 1 ,x 2 ,…,x N The trajectory matrix X is obtained by performing a lag arrangement:

[0026] In normal circumstances, L is taken. <N / 2。

[0027] Let K=N If L+1, then the trajectory matrix X is an L×K matrix:

[0028] S1.5 Singular Value Decomposition: Perform singular value decomposition on the trajectory matrix, calculate the covariance matrix S, and perform eigenvalue decomposition to obtain the eigenvalues. λ m and eigenvectors U m ;

[0029] in, It is a diagonal matrix; The left matrix; It is a right matrix; Since direct decomposition of matrix X is difficult, we first calculate the covariance matrix:

[0030] Where S is the covariance matrix, Let X be the transpose of the trajectory matrix X.

[0031] Eigenvalues ​​are obtained by performing eigenvalue decomposition on matrix S. λ 1 ,λ 2 ,…,λ L The corresponding feature vector is [ ]. U 1 ,U 2 ,…, U L ]; S1.6 Grouping and Reconstruction: Based on the contribution of eigenvalues, select the top... K The main components are grouped together, and the trend component reflecting the global trend of the signal is reconstructed by the diagonal averaging method. This further suppresses high-frequency noise and strengthens the core trend, resulting in a trend sequence after synergistic noise reduction.

[0032] The L components obtained from singular value decomposition are divided into... c Each set consists of a non-overlapping subset, and each group represents a specific trend pattern or signal component, such as long-term trends, periodic fluctuations, random noise, etc.

[0033]

[0034] in, l For grouped indexes, it means the first... l Groups l =1,2,…, c , I l For the first l The set of component indices contained in each group. For the first l The reconstructed trajectory submatrix corresponding to each group; Calculate the projection factor:

[0035] in, yes This reflects the temporal evolution pattern in the original sequence. Time period weighting It is the first of the trajectory matrix i List; Reconstructing the signal using time orthogonal functions and time principal components:

[0036] in, Indicates the first k The reconstructed component is at the... i The value at each point in time. It is the weight of each time period. U k,j Indicates the first k The first time empirical orthogonal function j One element, It is a reconstructed sequence; S2 Multi-scale Feature Decomposition: The trend sequence obtained in step S1 after collaborative noise reduction is subjected to CEEMDAN decomposition. By adaptively adding Gaussian white noise and performing multiple ensemble averaging operations, the non-stationary signal is decomposed into a series of intrinsic mode function (IMF) components and residual components, ultimately obtaining several IMF components, thus realizing multi-scale representation of the signal in the time-frequency domain. The specific steps are as follows: S2.1 Add adaptive noise: Add a white noise sequence modulated by adaptive white noise amplitude to the trend sequence to be decomposed; S2.2 Empirical Mode Decomposition: Perform empirical mode decomposition on the signal after adding noise to obtain a set of intrinsic mode functions; S2.3 Integrated Averaging: Repeat S2.1 and S2.2 for a total of N times, adding a different white noise sequence each time, and integrating and averaging the IMF components of the corresponding order obtained from each decomposition to obtain a final set of IMF components and a residual component. S2.4 Output decomposition results: The final output IMF components are arranged from high frequency to low frequency, representing the local characteristics of the signal at different time scales, and the residual components represent the overall trend or baseline of the signal.

[0037] Specifically, to the original signal x ( t Add to ) i ( i =1,2, …,m Gaussian white noise that follows a standard normal distribution w i ( t ), and a new signal is obtained:

[0038] in, The original signal, x i ( t ) is the first i The signal after adding noise. w i ( t ) is the first i The Gaussian white noise was added later; right x i ( t Perform EMD decomposition and retain only the first decomposed component. The first IMF component of the original signal is obtained by taking the average value. I IMF1 :

[0039] in, For the first i sub-noise signal x i ( t The first IMF component obtained by performing EMD decomposition. m Add the number of times to the total noise; Corresponding residual components r 1( t )for:

[0040] in, r 1( t () represents the residual components after the first decomposition; To the residual component r 1( t Add to ) i ( i=1,2,…,m Gaussian white noise that follows a standard normal distribution w i ( t New components are obtained:

[0041] in, For the first i Secondary residual components r 1( t The signal obtained after adding noise; right Perform EMD decomposition and retain only the first modal component. Then the second IMF component of the original signal I IMF2 for:

[0042] Repeated decomposition j ( j =1,2,… ,n This process continues until the residual cannot be decomposed further, at which point the following steps can be taken: I IMF1 , I IMF2 , ..., I IMFn and the corresponding residual components. Finally, the original signal can be expressed as:

[0043] in, The first of the original signals j One IMF component; The final residual component represents the trend term or monotonic component in the signal that cannot be decomposed. j For the index of the number of decompositions, j =1,2,…, n; n This represents the final total number of IMF components.

[0044] In the CEEMDAN decomposition, the noise coefficient is set to 0.2, the number of noise additions is set to 500, and the maximum number of envelope iterations is 5000. S3 Component Prediction and Model Optimization: For each IMF component obtained in step S2, an independent GRU prediction model is established and trained and predicted. The training of the GRU prediction model adopts the Adam optimizer and combined with the learning rate decay strategy. By dynamically adjusting the learning rate and gradient update mechanism, the convergence speed and long-term prediction stability of the model are improved. In the training steps of the GRU prediction sub-model, the input time step is set to 12, the maximum number of training iterations is set to 200, and the initial learning rate is set to 0.001.

[0045] The specific process of the learning rate decay strategy is as follows: During model training, when the validation set loss no longer decreases within P consecutive training epochs, the current learning rate is multiplied by the decay factor. α (0< α <1) Decrease the learning rate until the preset minimum learning rate is reached.

[0046] S4 Result Reconstruction and Output: The prediction results of all IMF components obtained in step S3 are linearly superimposed to reconstruct the final vibration trend prediction value.

[0047] The specific steps of step S4 are as follows: the prediction results of each GRU prediction sub-model for the corresponding IMF component are superimposed and summed to obtain the final vibration trend prediction value; the accuracy of the final prediction result is evaluated using the root mean square error (RMSE), mean absolute deviation (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²).

[0048] Simultaneously, a hydropower unit vibration trend prediction system for implementing the method is provided, characterized in that it includes: (1) Data acquisition and collaborative noise reduction module, used to acquire the original vibration signal and perform the WTD and SSA processing to obtain the trend sequence after collaborative noise reduction; (2) Signal decomposition module, used to perform the CEEMDAN decomposition and integrate and average the IMF components of the corresponding order obtained from each decomposition to obtain several IMF components; (3) Parallel prediction module, which contains multiple GRU prediction units, used for parallel training and prediction of each IMF component; (4) Result synthesis and output module, which is used to sum the prediction results of each IMF component and output the final vibration trend prediction value.

[0049] The data acquisition and collaborative noise reduction module, signal decomposition module, parallel prediction module, and result synthesis and output module are connected in sequence to realize fully automated processing from raw signal acquisition to trend prediction output.

[0050] The parallel prediction module has a built-in Adam optimizer and learning rate decay controller, which are used to optimize parameters during model training.

[0051] Taking the monitoring data of the Y-axis swing of the upper guide bearing of Unit #2 of a pumped storage power station in China from July 24, 2008 to December 15, 2011 as an example.

[0052] First, the original vibration signal undergoes two-stage noise reduction and trend enhancement through the synergistic effect of WTD and SSA. Then, CEEMDAN decomposition is used to decompose the processed signal into multiple IMF components. Next, a gated recurrent unit trained with an Adam optimizer and a learning rate decay strategy is used to independently predict each IMF component. Finally, the prediction results of all components are summarized to obtain the final vibration trend. The flowchart is as follows: Figure 1 As shown.

[0053] (1) Signal preprocessing and collaborative noise reduction.

[0054] Acquire raw vibration signals, such as Figure 2 As shown, WTD with 5-level decomposition using db4 wavelet is used for initial noise reduction of the signal. Subsequently, SSA processing is performed on the denoised signal, with the window length L set to 1 / 3 of the total signal length, to extract its global trend component, as shown. Figure 3 As shown.

[0055] (2) Multiscale feature decomposition.

[0056] The trend sequence extracted by SSA was decomposed using CEEMDAN, with a noise figure of 0.2, 500 noise additions, and a maximum of 5000 iterations. The signal was ultimately decomposed into 13 IMF components, such as... Figure 4 As shown.

[0057] (3) Component prediction and model optimization.

[0058] A GRU prediction sub-model is built for each IMF component. The model parameters are set as follows: input step size 12, maximum number of iterations 200, initial learning rate 0.001, using the Adam optimizer with learning rate decay enabled. Training and prediction are performed separately for each component.

[0059] (4) Result reconstruction and output.

[0060] The prediction results of all IMF components are summed to obtain the final predicted vibration trend value, as shown below. Figure 5 As shown in the figure. Calculations show that the prediction results have an RMSE of 3.4024, a MAE of 2.5351, a MAPE of 4.6517%, and an R² of 0.9523, indicating extremely high prediction accuracy.

[0061] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion, characterized in that, Includes the following steps: S1 Signal Preprocessing and Collaborative Denoising: The original vibration time series signal of the hydropower unit was obtained by performing wavelet threshold denoising and singular spectrum analysis in sequence to obtain the trend sequence after collaborative denoising. S2 Multiscale Feature Decomposition: Perform fully adaptive noise set empirical mode decomposition on the trend sequence obtained in step S1 after collaborative noise reduction to obtain several intrinsic mode function components. S3 Component Prediction and Model Optimization: For each intrinsic mode function component obtained in step S2, an independent gated recurrent unit prediction model is established and trained and predicted. The training of the gated recurrent unit prediction model adopts the Adam optimizer and combined with the learning rate decay strategy. S4 Result Reconstruction and Output: The prediction results of all intrinsic mode function components obtained in step S3 are superimposed to reconstruct the final vibration trend prediction value.

2. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 1, characterized in that, The specific steps of step S1 are as follows: First, the original vibration time series data of the hydropower unit is obtained. Then, the original vibration time series signal of the hydropower unit is initially denoised by wavelet threshold denoising. Then, the processed signal is subjected to singular spectrum analysis to extract its global trend component in order to further suppress residual high-frequency noise and strengthen the core trend features, so as to obtain the trend sequence after collaborative denoising.

3. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 2, characterized in that, Step S1, wavelet thresholding for denoising, specifically includes the following steps: S1.1 Selecting appropriate wavelet basis functions and decomposition levels: Using the db4 wavelet basis, the original vibration signal is decomposed into 5 levels of wavelets to obtain the high-frequency detail coefficients and low-frequency approximation coefficients of each level. S1.2 Thresholding: The threshold of high-frequency coefficients in each layer is calculated using a general thresholding rule, and a soft thresholding function is selected to quantize the detail coefficients and suppress noise components. S1.3 Signal Reconstruction: Perform inverse wavelet transform on the coefficients of each layer after thresholding to reconstruct the signal after preliminary noise reduction.

4. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 2, characterized in that, The singular spectrum analysis process in step S1 specifically includes the following steps: S1.4 Embedding: The signal after preliminary noise reduction is converted into a one-dimensional time series, and then the one-dimensional time series of length N is further processed... x 1, x 2, ..., x N The sequence is mapped to a trajectory matrix X, and the embedding window length L is set to 1 / 3 of the original sequence length. S1.5 Singular Value Decomposition: Perform singular value decomposition on the trajectory matrix and calculate the covariance matrix S=XX. T Eigenvalues ​​are obtained by performing eigenvalue decomposition. λ m and eigenvectors U m ; S1.6 Grouping and Reconstruction: Based on the contribution of eigenvalues, select the top... K The main components are grouped together, and the trend component reflecting the global trend of the signal is reconstructed by the diagonal averaging method. This further suppresses high-frequency noise and strengthens the core trend, resulting in a trend sequence after synergistic noise reduction.

5. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1 Add adaptive noise: Add a white noise sequence modulated by adaptive white noise amplitude to the trend sequence to be decomposed; S2.2 Empirical Mode Decomposition: Perform empirical mode decomposition on the signal after adding noise to obtain a set of intrinsic mode functions; S2.3 Integrated averaging: Repeat S2.1 and S2.2 for a total of N times, adding a different white noise sequence each time, and integrating and averaging the eigenmode function components of the corresponding order obtained from each decomposition to obtain a final set of eigenmode function components and a residual component. S2.4 Output decomposition results: The final output intrinsic mode function components are arranged from high frequency to low frequency, representing the local characteristics of the signal at different time scales, and the residual components represent the overall trend or baseline of the signal.

6. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 5, characterized in that, The noise figure in the adaptive white noise is set to 0.2, and the number of noise additions is set to 500.

7. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 1, characterized in that, The specific steps of step S3 are as follows: For each intrinsic mode function component obtained in step S2, construct and train an independent gated recurrent unit prediction sub-model; during the training process, use the Adam optimizer and combine it with the learning rate decay strategy to optimize the parameters of the gated recurrent unit prediction sub-model to improve training efficiency and model performance.

8. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 7, characterized in that, In the training step of the gated recurrent unit prediction sub-model, the input time step is set to 12, the maximum number of training iterations is set to 200, and the initial learning rate is set to 0.

001. The specific process of the learning rate decay strategy is as follows: During model training, when the validation set loss no longer decreases within P consecutive training cycles, the current learning rate is multiplied by the decay factor α to decay the learning rate, where 0 < α < 1, until the preset minimum learning rate is reached.

9. The method for predicting the vibration trend of hydropower units based on collaborative noise reduction and feature fusion as described in claim 1, characterized in that, The specific steps of step S4 are as follows: the prediction results of each gated cyclic unit prediction sub-model for the corresponding intrinsic mode function components are superimposed and summed to obtain the final vibration trend prediction value; the accuracy of the final prediction result is evaluated using root mean square error, mean absolute error, mean absolute percentage error and coefficient of determination.

10. A system for predicting the vibration trend of a hydroelectric generator unit for implementing the method of any one of claims 1 to 9, characterized in that, include: (1) Data acquisition and collaborative noise reduction module, used to acquire the original vibration signal and perform the wavelet threshold denoising and singular spectrum analysis to obtain the trend sequence after collaborative noise reduction; (2) Signal decomposition module, used to perform empirical mode decomposition and integrate and average the intrinsic mode function components of the corresponding order obtained by each decomposition to obtain several intrinsic mode function components; (3) Parallel prediction module, which contains multiple gated recurrent unit prediction sub-models, used for parallel training and prediction of each intrinsic mode function component; (4) Result synthesis and output module, which is used to superimpose and sum the prediction results of each intrinsic mode function component and output the final vibration trend prediction value; The data acquisition and collaborative noise reduction module, signal decomposition module, parallel prediction module, and result synthesis and output module are connected in sequence to realize fully automated processing from raw signal acquisition to trend prediction output; the parallel prediction module has a built-in Adam optimizer and learning rate decay controller to optimize parameters during model training.

Images