A flood discharge gate vibration signal enhancement method based on physical information and deep learning

Abstract

This invention discloses a method for enhancing the vibration signal of a flood discharge gate based on physical information and deep learning. First, the prototype vibration response signal of the flood discharge gate is collected as the original noisy signal. Then, a physical information enhancement method combining SSA and SVMD is used to process the original noisy signal, obtaining a clean signal with a high signal-to-noise ratio. Next, an intelligent signal enhancement model based on CBAM-CDAE is constructed, using the original noisy signal as input and the clean signal as a label to train the intelligent signal enhancement model, resulting in a trained signal enhancement model. Finally, the vibration response signal of the flood discharge gate to be enhanced is input into the trained signal enhancement model, outputting a vibration response signal with enhanced signal-to-noise ratio. This invention's method generates a clean signal using a combined physical information enhancement method to train the CBAM-CDAE model, solving the label missing problem while achieving rapid and effective enhancement of weak vibration signals from flood discharge gates, thus providing a feasible solution for online real-time processing of flood discharge gate vibration signals.

Description

Technical Field

[0001] This invention belongs to the field of signal-to-noise ratio enhancement technology for flood discharge gate vibration response signals, specifically relating to a method for enhancing flood discharge gate vibration signals based on physical information and deep learning. Background Technology

[0002] Vibration safety monitoring of spillway structures is a crucial means of ensuring the safe operation of floodgates. However, due to the large size and high stiffness of floodgates, their vibration response amplitude remains relatively weak even under spillway excitation. Furthermore, environmental interference often exists in the vibration response signal, easily submerging structural vibration characteristics in noise, which significantly hinders the extraction of vibration information from the spillway structure. Therefore, effective signal enhancement processing methods are needed to preserve structural characteristic information while suppressing the impact of environmental interference on effective information extraction. Traditional methods such as CEEMDAN and VMD suffer from problems such as long iterative computation time, insufficient real-time processing performance, and difficulty in selecting key parameters. Therefore, there is an urgent need to develop signal enhancement methods that both incorporate physical information and can be processed rapidly. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent signal enhancement method based on Singular Spectrum Analysis (SSA), Adaptive Successive Variational Mode Decomposition (SVMD), and Attention Mechanism-Convolutional Denoising Autoencoder (CBAM-CDAE). Addressing the problems of complex operation and low efficiency of traditional methods, this invention generates a high signal-to-noise ratio (SNR) training set using a physical information method combined with SSA-SVMD as the clean signal for training, solving the problem of deep learning models being unable to train due to the lack of clean signal labels in practical engineering applications. Simultaneously, by constructing an intelligent weak signal enhancement model based on CBAM-CDAE, rapid SNR enhancement of vibration signals is achieved. By combining the rapid processing capabilities of deep learning with the effectiveness of the SSA-SVMD physical information method in suppressing environmental interference, an intelligent deep learning signal enhancement model is constructed, thereby improving both weak signal enhancement processing capabilities and processing efficiency. This provides the possibility for online real-time processing of vibration signals from flood discharge gates.

[0004] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution.

[0005] A method for enhancing the vibration signal of a floodgate based on physical information and deep learning includes the following steps: Step S1: Collect the prototype vibration response signal of the floodgate as the original noisy signal; Step S2: The original noisy signal is processed using a physical information enhancement method based on Singular Spectral Analysis (SSA) and Adaptive Successive Variational Mode Decomposition (SVMD) to obtain a clean signal with a high signal-to-noise ratio, and a deep learning training sample set is constructed accordingly. The combined physical information enhancement method uses SSA to preprocess the original noisy signal to filter out high-frequency white noise, and then uses SVMD to adaptively decompose the processed signal. The Newton-Raphson optimization algorithm (NRBO) is introduced during the decomposition process. The envelope entropy is used as the fitness function to evaluate the quality of SVMD decomposition. Through NRBO iterative optimization, the effective modal component reconstructed signal is selected, the interference of low-frequency noise is filtered out, and a clean signal with high signal-to-noise ratio is obtained. Step S3: Construct an intelligent signal enhancement model based on attention mechanism-convolutional denoising autoencoder CBAM-CDAE. Use the original noisy signal as input and the clean signal as label to train the intelligent signal enhancement model and obtain the trained signal enhancement model. The intelligent signal enhancement model uses a Convolutional Block Attention Module (CBAM) that is deeply integrated with the encoding and decoding structure of the CDAE through a two-layer attention mechanism. CBAM dynamically adjusts the weights of each feature channel through the channel attention module to enhance redundant information related to noise. At the same time, it uses a spatial attention module to focus on key regions of the signal to enhance the discriminative learning of local features, thereby improving the denoising performance and signal fidelity of the model. Step S4: Input the vibration response signal of the floodgate to be enhanced into the trained signal enhancement model, and output the vibration response signal with enhanced signal-to-noise ratio.

[0006] Specifically, in step S2, the original noisy signal is processed to obtain a clean signal with a high signal-to-noise ratio. The process is as follows: Step S21: Use Singular Spectrum Analysis (SSA) to preprocess the original noisy signal. Reconstruct the signal by trajectory matrix embedding, singular value decomposition, grouping, and diagonal averaging to filter out high-frequency white noise. Step S22: Introduce the Newton-Raphson optimization algorithm (NRBO) and optimize the quadratic penalty factor parameter of SVMD using the minimization of envelope entropy as the fitness function. ; Step S23: Adaptively decompose the reconstructed signal after SSA preprocessing using SVMD with optimized quadratic penalty factor parameters to obtain a series of intrinsic mode components (IMFs). Step S24: Calculate the Pearson correlation coefficient between each IMF component and the original noisy signal, and screen the effective IMF components according to the preset threshold for reconstruction to filter out low-frequency environmental interference and finally obtain a pure signal that reflects the true vibration characteristics of the structure.

[0007] Furthermore, the reconstruction process of the original noisy signal in step S21 is as follows: Step S211: Trajectory matrix embedding; The original noisy signal is denoted as a signal of length... The original time series Choose an appropriate window size The original time series Mapped to A length of The vectors form the trajectory matrix. : (1); In the above formula, ; Step S212, Singular value decomposition; Define the covariance matrix For the covariance matrix Singular values ​​are obtained by performing eigenvalue decomposition. and eigenvectors ; In singular value analysis: (2); (3); In the above formula, Let L be the singular component matrix; The basic matrix; It is the k-th singular value; The left matrix, The right matrix, and All are orthogonal matrices; Combining formulas (2) and (3), we get: (4); In the above formula, Right matrix transpose; Step S213: Grouping; Calculate the variance contribution rate for each singular value. , Let i be the i-th singular value, and select the top singular values ​​whose cumulative contribution rate exceeds the threshold. matrix components , trajectory matrix Divided into effective signal matrix and noise interference matrix ; Step S214: Diagonal average; Effective signal matrix Converted into a reconstructed signal sequence with the same length as the measured vibration response sequence. : (5); (6); In the above formula, For matrix elements; This is the nth sampling point; is the number of columns in the trajectory matrix.

[0008] Furthermore, in step S22, the quadratic penalty factor parameter of SVMD is optimized. The process is as follows: Step S221: Population initialization; Preset secondary penalty factor parameters The range of values ​​is ,exist Internal random generation The candidate solutions are Values ​​constitute the initial population. The first in the population candidate solutions The position is represented as: (7); In the above formula, Indicates the first The first candidate solution Position in each dimension , , The dimension of the population; express Uniformly distributed random numbers between; Step S222: Calculate fitness; With candidate solutions As a penalty factor for SVMD, the reconstructed signal sequence The decomposition yields a set of modal components; Calculate the envelope entropy of each modal component. The average value is then taken as the fitness value of the current individual. : (8); (9); (10); In the above formula, This represents the length of the original time series, i.e., the total number of sampling points. For the index of the sampling point, ; For the first Normalized envelope amplitude of each sampling point; Represents the natural logarithm; The instantaneous amplitude envelope is obtained after the signal undergoes the Hilbert transform; This represents the sum of the envelope amplitudes of all sampling points, after normalization. ; This represents the total number of modal components. Record the globally optimal individual (Minimum fitness) and worst individual (Maximum fitness); Step S223, NRSR search; NRSR derives the position update formula by approximating the first and second derivatives, and introduces adaptive parameters during the iteration process. To enhance the search capability of the NRSR optimization process, the improved NRSR search process is expressed as follows: (11); (12); In the above formula, , They represent the first , There are 10 candidate solutions; The objective function value; , They represent respectively to Find the first and second derivatives; express Uniformly distributed random numbers between; Indicates the current iteration number; Indicates the maximum number of iterations; Step S224, Trap Avoidance Operator (TAO); TAO avoids getting trapped in local optima by randomly selecting parameters and updating population positions, further improving the global search capability of NRSR. The update process of the trap avoidance operator TAO is represented as follows: (13); In the above formula, and These represent the updated position and the original position of an individual in the population, respectively. express Uniformly distributed random numbers between; and They are and Random numbers between; Indicates the worst position; Indicates the optimal position; For random disturbance terms, ; Step S225: Iterative optimization; For candidate solutions Calculate candidate positions according to step S223 Generate the updated version according to step S224 Combined with step S222, calculate , , The fitness of the solution is used to select the solution with the best fitness as the optimal solution. ,like Exceeding the preset range Then take and The nearest boundary value; When the maximum number of iterations is reached If the fitness change is less than a preset threshold, then the iteration stops and the global optimal solution is output. As a secondary penalty factor for SVMD.

[0009] Furthermore, in step S23, the SVMD with optimized quadratic penalty factor parameters is used to adaptively decompose the reconstructed signal after SSA preprocessing. The process is as follows: Assume the input signal is : (14); In the above formula, For the first First-order modal signal; Represents the residual signal, from the 1st order to the 2nd order. The sum of the first-order modal signals constitutes the composition; For the first First-order modal signal; This is an unprocessed signal; To ensure that the above assumptions are met, the following constraints need to be set; Modal compactness constraints: ensuring modal compactness Closely around its center frequency , is represented as: ; In the above formula, Indicates the first The center frequency of the first mode; Represents the convolution operator; Indicates time The partial derivative; Represents the Dirac function; The imaginary unit; is the base of the natural logarithm; Spectral overlap minimization constraint: in modal With the extracted Under the condition of minimum spectral overlap of the first mode, a suitable filter is selected, and its response frequency is [value missing]. Represented as: (15); In the above formula, It is a secondary penalty factor; Angular frequency; The established constraint for minimizing spectral overlap is expressed as: (16); In the above formula, This is the impulse response of the filter; Signal fidelity constraints: for signals from the first order to... Similar processing is applied to the first mode, setting... For the first The impulse response and frequency response of the first-order mode filter are obtained. : (17); The established signal fidelity constraints are: (18); Signal complete reconstruction constraint: Ensures that the signal can be completely reconstructed, expressed as: (19); Based on the above constraints, establish a mathematical model for constraint minimization: (20); In the above formula, To balance , and Parameters; This represents the IMF of the Lth modal component to be extracted. This is the accumulated signal of the extracted modes; This is the original observation signal; By minimizing the total cost function The modal components of each order are solved sequentially.

[0010] Furthermore, the formulas for calculating the Pearson correlation coefficient and threshold in step S24 are expressed as follows: (twenty one); (twenty two); In the above formula, The Pearson correlation coefficient; The covariance between each IMF component and the noise signal; The standard deviation of each IMF component; The standard deviation of the noisy signal; The threshold for the correlation coefficient; The maximum correlation coefficient; When the correlation coefficient R between the IMF and the original signal is greater than the threshold μ, it is retained, and the retained IMF components are superimposed to obtain the enhanced signal after physical information processing.

[0011] Specifically, the intelligent signal enhancement model constructed in step S3, based on the attention mechanism-convolutional denoising autoencoder CBAM-CDAE, adopts an encoder-decoder symmetrical structure, and embeds CBAM attention modules after the key layers of the encoder and decoder; the training process of the intelligent signal enhancement model is as follows: Step S31: Data preparation; Using the original noisy signal from step S1 as input and the clean signal obtained in step S2 as the training label of the model, the original noisy signal and the clean signal are divided into training set, validation set and test set in a ratio of 8:1:1. Step S32: Signal segmentation; Intelligent signal enhancement models require a fixed-length input; therefore, the long signal needs to be segmented into multiple short segments. Assume the length of each segment is... The output length after n convolutional and pooling layers is... Satisfy the following formula: (twenty three); In the above formula, , These are the filter sizes for the convolutional layer and the pooling layer, respectively. , These are the step sizes corresponding to these layers; Indicates the amount of zero fill used; Since the sizes before and after convolution are the same, the filter size of the merged layer is... and step length If they are equal, then formula (23) simplifies to: (twenty four); Step S33: Model training; The model training adopts a supervised learning approach, taking noisy signal segments as input and corresponding clean signal segments as target outputs. After each convolutional layer, a modified linear unit ReLU non-linear activation function is used, and a sigmoid function is used in the output layer. The mean squared error (MSE) is used as the loss function during model training. Let the length of the input signal be... The number of training samples is expressed as Then the MSE loss function Represented as: (25); In the above formula, For learnable parametric geometry; The loss function is at the sample level. The input signal to the model is the noisy signal segment; This is the output signal of the model, i.e., the signal after noise reduction; Training stops when the validation loss changes less than a preset threshold for 10 consecutive times. The model parameters with the minimum validation loss are saved to obtain the trained signal enhancement model.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. The method of this invention generates a high signal-to-noise ratio training set as a clean signal in training by using the physical information method of SSA-SVMD joint, which solves the problem that deep learning models cannot be trained due to the lack of clean signal labels in practical engineering applications.

[0013] 2. The method of the present invention constructs an intelligent weak signal enhancement model based on CBAM-CDAE, inherits the enhancement advantages of traditional physical information enhancement methods, and relies on its fast processing capability to greatly improve the enhancement efficiency, thereby realizing rapid signal-to-noise ratio enhancement processing of vibration signals.

[0014] 3. The method of this invention combines the fast processing capability of deep learning with the effectiveness of the SSA-SVMD physical information method in suppressing environmental interference to construct an intelligent deep learning signal enhancement model, thereby improving the processing efficiency while enhancing the weak signal enhancement capability.

[0015] 4. The method of this invention addresses the problems of complex operation and low efficiency of traditional methods. By leveraging the powerful feature extraction advantages of deep learning, it can quickly and effectively enhance the weak vibration response signal of large-volume, high-stiffness floodgates, thereby providing a feasible solution for online real-time processing of floodgate vibration signals. Attached Figure Description

[0016] To provide a more intuitive understanding of the technical implementation of this invention, the accompanying drawings involved in the embodiments of this invention are briefly described below. These drawings are used to assist in illustrating the implementation methods and are not intended to limit the invention. Those skilled in the art can make derivative designs based on the drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for enhancing the vibration signal of a floodgate based on physical information and deep learning, according to the present invention. Figure 2 The flowchart shows the physical information enhancement method based on the combination of Singular Spectral Analysis (SSA) and Adaptive Successive Variational Mode Decomposition (SVMD) of the present invention. Figure 3 This is a flowchart illustrating the SVMD parameter optimization process in an embodiment of the present invention. Figure 4 This is a diagram of the CBAM-CDAE network structure in an embodiment of the present invention; Figure 5 This is a time-domain diagram of the simulated clean and noisy attenuated signals in an embodiment of the present invention; Figure 6 This is a time-domain comparison diagram of the simulated signal and the clean signal in an embodiment of the present invention; Figure 7 This is a schematic diagram of the loss of the CBAM-CDAE model in the simulation example of this invention; Figure 8 This is a time-domain diagram of the enhanced results in an embodiment of the present invention; Figure 9 This is a spectrum of the suppression results in an embodiment of the present invention; Figure 10 This is a comparison chart of SNR and RMSE of different methods under different noise levels in the embodiments of the present invention; Figure 11 This is a diagram showing the dam discharge and on-site testing during the implementation of this invention; Figure 12 This is a time-domain image of measurement point D9 before and after enhancement according to the present invention; Figure 13 This is a frequency domain diagram of measurement point D9 before and after enhancement in this invention; Figure 14 This is a loss diagram of the CBAM-CDAE model in an engineering example of the present invention; Figure 15 This is a time-domain plot of the D9 test set before and after enhancement according to the present invention; Figure 16 This is a frequency domain diagram before and after enhancement of the D9 test set in this invention. Detailed Implementation

[0018] To facilitate understanding and implementation of the present invention by those skilled in the art, the various steps of the method proposed in this invention are described in detail below. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various modifications or alterations to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0019] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a method for enhancing the vibration signal of a floodgate based on physical information and deep learning, including the following steps: Step S1: Collect the prototype vibration response signal of the floodgate as the original noisy signal; Step S2: The original noisy signal is processed using a physical information enhancement method based on Singular Spectral Analysis (SSA) and Adaptive Successive Variational Mode Decomposition (SVMD) to obtain a clean signal with a high signal-to-noise ratio, and a deep learning training sample set is constructed accordingly. The combined physical information enhancement method uses SSA to preprocess the original noisy signal to filter out high-frequency white noise, and then uses SVMD to adaptively decompose the processed signal. The Newton-Raphson optimization algorithm (NRBO) is introduced during the decomposition process. The envelope entropy is used as the fitness function to evaluate the quality of SVMD decomposition. Through NRBO iterative optimization, the effective modal component reconstructed signal is selected, the interference of low-frequency noise is filtered out, and a clean signal with high signal-to-noise ratio is obtained. Step S3: Construct an intelligent signal enhancement model based on attention mechanism-convolutional denoising autoencoder CBAM-CDAE. Use the original noisy signal as input and the clean signal as label to train the intelligent signal enhancement model and obtain the trained signal enhancement model. The intelligent signal enhancement model uses a Convolutional Block Attention Module (CBAM) that is deeply integrated with the encoding and decoding structure of the CDAE through a two-layer attention mechanism. CBAM dynamically adjusts the weights of each feature channel through the channel attention module to enhance redundant information related to noise. At the same time, it uses a spatial attention module to focus on key regions of the signal to enhance the discriminative learning of local features, thereby improving the denoising performance and signal fidelity of the model. Step S4: Input the vibration response signal of the floodgate to be enhanced into the trained signal enhancement model, and output the vibration response signal with enhanced signal-to-noise ratio.

[0020] Specifically, such as Figure 2 As shown, in step S2, the original noisy signal is processed to obtain a clean signal with a high signal-to-noise ratio. The process is as follows: Step S21: Use Singular Spectrum Analysis (SSA) to preprocess the original noisy signal. Reconstruct the signal by trajectory matrix embedding, singular value decomposition, grouping, and diagonal averaging to filter out high-frequency white noise. Step S22: Introduce the Newton-Raphson optimization algorithm (NRBO) and optimize the quadratic penalty factor parameter of SVMD using the minimization of envelope entropy as the fitness function. ; Step S23: Adaptively decompose the reconstructed signal after SSA preprocessing using SVMD with optimized quadratic penalty factor parameters to obtain a series of intrinsic mode components (IMFs). Step S24: Calculate the Pearson correlation coefficient between each IMF component and the original noisy signal, and screen the effective IMF components according to the preset threshold for reconstruction to filter out low-frequency environmental interference and finally obtain a pure signal that reflects the true vibration characteristics of the structure.

[0021] Singular Value Analysis (SSA) is a nonparametric time series analysis method based on matrix factorization. By constructing a trajectory matrix and combining it with matrix factorization techniques, it adaptively separates the time series into non-repeating signal components. By calculating the contribution rate, effective signal components are selected and superimposed, thereby achieving noise filtering. SSA mainly consists of four steps: trajectory matrix embedding, singular value decomposition, grouping, and diagonal averaging. The specific process is as follows: Step S211: Trajectory matrix embedding; The original noisy signal is denoted as a signal of length... The original time series Choose an appropriate window size The original time series Mapped to A length of The vectors form the trajectory matrix. : (1); In the above formula, ; Step S212, Singular value decomposition; Define the covariance matrix For the covariance matrix Singular values ​​are obtained by performing eigenvalue decomposition. and eigenvectors ; In singular value analysis: (2); (3); In the above formula, Let L be the singular component matrix; The basic matrix; It is the k-th singular value; The left matrix, The right matrix, and All are orthogonal matrices; Combining formulas (2) and (3), we get: (4); In the above formula, Right matrix transpose; Step S213: Grouping; Calculate the variance contribution rate for each singular value. , Let i be the i-th singular value, and select the top singular values ​​whose cumulative contribution rate exceeds the threshold. matrix components , trajectory matrix Divided into effective signal matrix and noise interference matrix ; Step S214: Diagonal average; Effective signal matrix Converted into a reconstructed signal sequence with the same length as the measured vibration response sequence. : (5); (6); In the above formula, For matrix elements; This is the nth sampling point; is the number of columns in the trajectory matrix.

[0022] Furthermore, such as Figure 3 As shown, the Newton-Raphson Optimization Algorithm (NRBO) is a novel metaheuristic algorithm. It uses the Newton-Raphson Search Rule (NRSR) and the Trap Avoidance Operator (TAO) to explore the search domain, thereby defining the search path, improving search capability and convergence speed, and effectively avoiding local optima. In step S22, the quadratic penalty factor parameter of SVMD is optimized. The process is as follows: Step S221: Population initialization; Preset secondary penalty factor parameters The range of values ​​is ,exist Internal random generation The candidate solutions are Values ​​constitute the initial population. The first in the population candidate solutions The position is represented as: (7); In the above formula, Indicates the first The first candidate solution Position in each dimension , , The dimension of the population; express Uniformly distributed random numbers between; Step S222: Calculate fitness; With candidate solutions As a penalty factor for SVMD, the reconstructed signal sequence The decomposition yields a set of modal components; Calculate the envelope entropy of each modal component. The average value is then taken as the fitness value of the current individual. : (8); (9); (10); In the above formula, This represents the length of the original time series, i.e., the total number of sampling points. For the index of the sampling point, ; For the first Normalized envelope amplitude of each sampling point; Represents the natural logarithm; The instantaneous amplitude envelope is obtained after the signal undergoes the Hilbert transform; This represents the sum of the envelope amplitudes of all sampling points, after normalization. ; This represents the total number of modal components. Record the globally optimal individual (Minimum fitness) and worst individual (Maximum fitness); Step S223, NRSR search; NRSR derives the position update formula by approximating the first and second derivatives, and introduces adaptive parameters during the iteration process. To enhance the search capability of the NRSR optimization process, the improved NRSR search process is expressed as follows: (11); (12); In the above formula, , They represent the first , There are 10 candidate solutions; The objective function value; , They represent respectively to Find the first and second derivatives; express Uniformly distributed random numbers between; Indicates the current iteration number; Indicates the maximum number of iterations; Step S224, Trap Avoidance Operator (TAO); TAO avoids getting trapped in local optima by randomly selecting parameters and updating population positions, further improving the global search capability of NRSR. The update process of the trap avoidance operator TAO is represented as follows: (13); In the above formula, and These represent the updated position and the original position of an individual in the population, respectively. express Uniformly distributed random numbers between; and They are and Random numbers between; Indicates the worst position; Indicates the optimal position; For random disturbance terms, ; Step S225: Iterative optimization; For candidate solutions Calculate candidate positions according to step S223 Generate the updated version according to step S224 Combined with step S222, calculate , , The fitness of the solution is used to select the solution with the best fitness as the optimal solution. ,like Exceeding the preset range Then take and The nearest boundary value; When the maximum number of iterations is reached If the fitness change is less than a preset threshold, then the iteration stops and the global optimal solution is output. As a secondary penalty factor for SVMD.

[0023] Furthermore, SVMD is a nonlinear signal processing method based on a variational optimization framework, aiming to adaptively decompose complex signals into a series of Intrinsic Mode Functions (IMFs), where each IMF represents an independent frequency component in the signal. In step S23, SVMD with optimized quadratic penalty factor parameters is used to adaptively decompose the reconstructed signal after SSA preprocessing, as follows: Assume the input signal is : (14); In the above formula, For the first First-order modal signal; Represents the residual signal, from the 1st order to the 2nd order. The sum of the first-order modal signals constitutes the composition; For the first First-order modal signal; This is an unprocessed signal; To ensure that the above assumptions are met, the following constraints need to be set; Modal compactness constraints: ensuring modal compactness Closely around its center frequency , is represented as: ; In the above formula, Indicates the first The center frequency of the first mode; Represents the convolution operator; Indicates time The partial derivative; Represents the Dirac function; The imaginary unit; is the base of the natural logarithm; Spectral overlap minimization constraint: in modal With the extracted Under the condition of minimum spectral overlap of the first mode, a suitable filter is selected, and its response frequency is [value missing]. Represented as: (15); In the above formula, It is a secondary penalty factor; Angular frequency; The established constraint for minimizing spectral overlap is expressed as: (16); In the above formula, This is the impulse response of the filter; Signal fidelity constraints: for signals from the first order to... Similar processing is applied to the first mode, setting... For the first The impulse response and frequency response of the first-order mode filter are obtained. : (17); The established signal fidelity constraints are: (18); Signal complete reconstruction constraint: Ensures that the signal can be completely reconstructed, expressed as: (19); Based on the above constraints, establish a mathematical model for constraint minimization: (20); In the above formula, To balance , and Parameters; This represents the Lth modal component (IMF) to be extracted. This is the accumulated signal of the extracted modes; This is the original observation signal; By minimizing the total cost function The modal components of each order are solved sequentially.

[0024] Furthermore, the formulas for calculating the Pearson correlation coefficient and threshold in step S24 are expressed as follows: (twenty one); (twenty two); In the above formula, The Pearson correlation coefficient; The covariance between each IMF component and the noise signal; The standard deviation of each IMF component; The standard deviation of the noisy signal; The threshold for the correlation coefficient; The maximum correlation coefficient; When the correlation coefficient R between the IMF and the original signal is greater than the threshold μ, it is retained, and the retained IMF components are superimposed to obtain the enhanced signal after physical information processing.

[0025] Specifically, the intelligent signal enhancement model based on attention mechanism-convolutional denoising autoencoder CBAM-CDAE constructed in step S3 adopts an encoder-decoder symmetrical structure, and embeds CBAM attention modules after the key layers of the encoder and decoder.

[0026] like Figure 4 The diagram shown is a CBAM-CDAE network structure diagram in this embodiment: (1) The input signal is divided into segments in the encoder, each segment having a length of [length missing]. These segments are divided into 16 sizes. The process involves a one-dimensional convolutional filter with a stride of 1, producing a product of size . The output features are then downsampled through a 1D average pooling layer with a stride of 2 and a window size of 2 to obtain a size of [missing information]. The pooled output. Such convolutional and pooling layers constitute the basic structural unit of the encoder. This structural unit is repeated twice, with the number of one-dimensional convolutional filters increasing to 32 and 64 respectively in subsequent structures. After this series of processing, the final output of the encoder is compressed to a length of... In short, the encoder is responsible for extracting features from the input signal and encoding these signals into feature representations of latent dimensions.

[0027] (2) Introducing the CBAM attention mechanism into the convolutional denoising autoencoder (CDAE) can effectively improve the network's adaptive enhancement capability for noise features. CBAM dynamically adjusts the weights of each feature channel through the channel attention module to enhance redundant information related to noise, while using the spatial attention module to focus on key signal regions to enhance discriminative learning of local features. This dual attention mechanism is deeply integrated with the CDAE's encoding and decoding structure, enabling the bottleneck layer to obtain more representative compressed features and the decoder to achieve more refined signal reconstruction. Ultimately, it significantly improves denoising performance and signal fidelity while maintaining low computational overhead.

[0028] (3) The decoder is symmetrical to the encoder, and its function is to reconstruct the compressed features to correspond to the original signal segment. The operation of the one-dimensional convolutional filter is similar to that of the encoder, while the one-dimensional upsampling layer is used to enlarge the size of the input feature map. Therefore, the decoder can reconstruct the compressed features to correspond to the original signal segment. The compressed feature mapping back to size is The signal. The final output layer of the decoder uses a step size of 1 and a kernel size of A one-dimensional convolutional filter is used to output a denoised signal. In this process, the decoder works to reconstruct the signal from the compressed feature space.

[0029] The training process of the intelligent signal enhancement model is as follows: Step S31: Data preparation; Using the original noisy signal from step S1 as input and the clean signal obtained in step S2 as the training label of the model, the original noisy signal and the clean signal are divided into training set, validation set and test set in a ratio of 8:1:1. Step S32: Signal segmentation; Intelligent signal enhancement models require a fixed-length input; therefore, the long signal needs to be segmented into multiple short segments. Assume the length of each segment is... The output length after n convolutional and pooling layers is... Satisfy the following formula: (twenty three); In the above formula, , These are the filter sizes for the convolutional layer and the pooling layer, respectively. , These are the step sizes corresponding to these layers; Indicates the amount of zero fill used; Since the sizes before and after convolution are the same, the filter size of the merged layer is... and step length If they are equal, then formula (23) simplifies to: (twenty four); According to formula (24), the constant N of the input signal must be... The multiple of this. Based on this formula, a signal with a length of N=128 is selected to execute the CDAE algorithm. Table 1 below shows the detailed information of each layer of the CBAM-CDAE model in this embodiment.

[0030] Table 1. CBAM-CDAE model with input signal length N=128 ; Step S33: Model training; The performance of the CBAM-CDAE model depends on the appropriate selection of activation function, loss function, and optimization algorithm.

[0031] In this embodiment, supervised learning is used for model training. The noisy signal segment is taken as input, and the corresponding clean signal segment is taken as the target output. A ReLU (Modified Linear Unit) non-linear activation function is used after each convolutional layer, and a sigmoid function is used in the output layer. The mean squared error (MSE) is used as the loss function during model training. Let the length of the input signal be... The number of training samples is expressed as Then the MSE loss function Represented as: (25); In the above formula, For learnable parametric geometry; The loss function is at the sample level. The input signal to the model is the noisy signal segment; This is the output signal of the model, i.e., the signal after noise reduction; Training stops when the validation loss changes less than a preset threshold for 10 consecutive times. The model parameters with the minimum validation loss are saved to obtain the trained signal enhancement model.

[0032] In deep learning model training, the choice of optimization algorithm has a significant impact on its convergence speed and accuracy. Common optimizers include stochastic gradient descent (SGD), Momentum, Adagrad, RMSProp, and Adam. Among them, Adagrad, RMSProp, and Adam employ an adaptive learning rate strategy, meaning their learning rates are not fixed. In this embodiment, the CDAE model is trained and optimized using the Adam optimizer. The Adam optimizer effectively combines the features of Adagrad and RMSProp to dynamically adjust weights, thereby achieving better training results.

[0033] The following numerical simulation and engineering example analysis further illustrate the enhancement effect of the method of the present invention on the signal-to-noise ratio of the vibration response signal of the floodgate.

[0034] Model parameter settings In this embodiment, the SSA-SVMD and CDAE models were used. During the model computation, the experimental computer was equipped with a 12th Gen Intel(R) Core(TM) i5-12600KF 3.70 GHz processor, 32GB of memory, an NVIDIA GeForce RTX 3060 graphics processor, and a Win10 64-bit operating system. The window size in SSA was 256, the training learning rate for the CDAE model was set to 0.0001, the optimizer was Adam, and the loss function was mean squared error.

[0035] I. Numerical simulation analysis; 1) Numerical simulation signal construction; To verify the signal enhancement effect of the proposed method, an analog signal containing dense frequency components was designed in this embodiment. By superimposing white noise and low-frequency noise, a signal with a sampling frequency of 200Hz and a duration of 10s is generated. This is used to simulate the measured flow-induced vibration signal, and its specific functional form is as follows: ; In the above formula, P represents the noise level. The standard deviation of the pure signal; Indicates a pure signal White noise of the same length; The sampling frequency is set to 200Hz, the sampling time to 768s, and the noise level to P=50%. The signal is divided into 1200 signal sets, each 0.64s long, and these sets are further divided into training, validation, and test sets with a ratio of 8:1:1. Figure 5 The figure shows the time-domain plots of the obtained clean signal and the noisy signal.

[0036] like Figure 6 As shown, the validation set signal is taken and subjected to Fourier transform to obtain the time-frequency domain diagram of the validation set before and after adding noise.

[0037] The parameter thresholds for SVMD were calculated according to formula (22). To verify the rationality of the SSA-SVMD parameters and the CBAM-CDAE network framework parameter design, this embodiment adopted the controlled variable method and compared the effects of different window lengths in SSA, different number of layers in CBAM-CDAE, convolution kernel size, pooling strategy, and input segment length on the model enhancement performance and efficiency using a 50% noise signal system, in order to determine the optimal parameter configuration. The experimental results are shown in Tables 2-4 below.

[0038] Table 2. Pearson correlation coefficients of IMFs ; Table 3. Comparison of Enhancement Effects of SSA with Different Window Lengths ; Table 4. Comparison of CBAM-CDAE Network Framework Parameters ; Based on the analysis results in Tables 3 and 4, the following conclusions can be drawn: The enhancement effect is optimal when the SSA window length is 256. When the network has 3 layers, a kernel size of 25, uses average pooling, and the input segment length is 128, the model achieves optimal output SNR and RMSE, while maintaining high computational efficiency. Specifically, average pooling achieves higher SNR and lower RMSE compared to max pooling, indicating that its smoothing characteristics are more suitable for this enhancement task. In the layer comparison, although the 2-layer network structure is slightly faster, its enhancement performance decreases significantly; the 4-layer network does not bring performance improvement but instead increases computational cost. Increasing the kernel size to 36 brings performance close to the optimal configuration, but considering both model complexity and efficiency, a kernel size of 25 is more balanced. The comparison of different input lengths shows that a length of 128 achieves the best balance between performance and efficiency. Therefore, subsequent experiments will be conducted based on this set of parameters (3 layers, 25 convolutional kernels, average pooling, and 128 segment lengths). This configuration ensures both the enhancement effect and the efficiency of the model's operation.

[0039] Use training and validation sets to train the model, such as... Figure 7 As shown, the training set loss and validation set loss tend to stabilize and exhibit good fit as they decrease continuously. The enhanced signal is obtained by feeding the test set into the trained CBAM-CDAE model. The time-frequency plots of the clean signal and the enhanced signal are compared, and evaluation metrics are calculated to verify the effectiveness of the convolutional denoising autoencoder enhancement.

[0040] Depend on Figure 8 It can be seen that the enhanced signal is compared to Figure 6A significant portion of the redundant noise was filtered out, and the time history curve of the signal showed good fit with the clean signal. The test set signal was subjected to Fourier transform, and the spectrograms of the simulated signal before and after enhancement were plotted to further verify the enhancement effect.

[0041] from Figure 9 As can be seen, noise is almost completely filtered out, and the clean signal and the enhanced signal are highly consistent. To verify the performance of the convolutional enhancement stacked autoencoder, the enhancement effect under different noise levels was measured using root mean square error and signal-to-noise ratio as performance indicators.

[0042] To verify the enhancement effectiveness of CBAM-CDAE under different noise levels (10%~50%), this study performed enhancement processing on simulated signals with different noise levels for each group and compared its enhancement performance with that of CNN, LSTM, SSA-SVMD, and SSA-CEEMDAN. The relevant data are shown in Table 5 below. Furthermore, histograms were plotted to visually present the differences in enhancement effects of the five methods under different noise levels, allowing for a clearer evaluation of the performance of each method.

[0043] Table 5. Comparison of Evaluation Indicators of Enhancement Effect of Different Methods at Different Noise Levels ; From Table 5 and Figure 10 A comparison of the SSA-CEEMDAN and SSA-SVMD methods reveals that, with the same window length, SSA-CEEMDAN has a heavier computational burden and insufficient decomposition stability, primarily affected by parameters such as noise amplitude and the number of noise additions. In contrast, SVMD, through adaptive successive modal synchronization optimization, maintains computational efficiency comparable to VMD while reducing the preset modality number K, significantly lowering computational complexity while achieving better denoising accuracy. Therefore, this invention selects SSA-SVMD as the method for generating high-precision denoised reference signals, aiming to provide the most reliable and purest training dataset for subsequent CBAM-CDAE.

[0044] Furthermore, the deep learning model proposed in this invention demonstrates excellent performance in signal denoising tasks. Compared to classical network models, CBAM-CDAE exhibits faster denoising speed, higher SNR, and lower RMSE, indicating that CBAM-CDAE can achieve more efficient and higher-quality signal enhancement. Compared to traditional enhancement methods, the signal-to-noise ratio of the signal processed by CBAM-CDAE is improved by more than 16.55 dB, the RMSE is reduced by more than 83% compared to the original signal, and the processing time per cycle is only 0.12 seconds, significantly outperforming the other two traditional enhancement methods. These results validate the effectiveness of CBAM-CDAE in enhancement and its superior efficiency.

[0045] II. Engineering Case Analysis; A certain floodgate consists of 7 gates and 8 piers. To further study the structural vibration characteristics of this floodgate, prototype vibration tests were conducted on the piers. A total of 14 horizontal sensors were installed on piers #6, #7, and #8. Figure 11 As shown, the measuring points are arranged on the top of the gate pier. As shown in Table 6 below, the test condition is that gates H6 and H7 are open, the gate opening is 0.3m, the upstream water level is 28.22m, and the test uses a BY-S07 sensor with a sampling frequency of 200Hz.

[0046] Table 6. Vibration Test Conditions for Prototype Flood Discharge Gate ; During data collection, noise from water flow, the environment, and mechanical equipment is inevitable. To ensure that the signal can more accurately respond to the dynamic characteristics of the structure in subsequent experiments, the method mentioned in this invention is used to enhance the signal. Figure 12 As shown in (a), SSA and SVMD were used to enhance the signals at 14 measurement points of the prototype vibration test signal. The parameters were selected to match those of the simulation signal, resulting in time history diagrams. In this example, only measurement point D9 was selected to demonstrate the enhancement effect of SSA and SVMD. Figure 12 (b) and Figure 13 The image shows the time-frequency diagrams before and after D9 enhancement; Depend on Figure 12 (b) and Figure 13 As can be seen, the SSA-SVMD enhancement method retains most of the main frequency signal and its energy, while eliminating low-frequency noise and background white noise. The enhanced signal is approximated as the clean signal in the convolutional denoising autoencoder, thereby training the enhancement network.

[0047] The training method used by CBAM-CDAE involves dividing the signal into multiple signal sequences, each with a length of 128. Since the total length of a single measurement point signal is 60,000, 59,904 data points are extracted from each measurement point and divided into training, validation, and test sets in a 7:1:1 ratio, resulting in 13 training sets (46,592 data points each), 13 validation sets (6,656 data points each), and 13 test sets (6,656 data points each). Furthermore, the SSA-SVMD-enhanced data is correspondingly divided and then fed into CBAM-CDAE for model training. Figure 14 As shown, during training, both the training set loss and the validation set loss gradually stabilized while continuously decreasing, indicating a good model fit. Finally, the test set was fed into the trained model to obtain the signal after test set enhancement at each measurement point, as shown... Figure 15 and Figure 16The image shows the time-frequency diagram of the test set at measurement point D9.

[0048] Depend on Figure 15 It can be seen that the signal peak value is reduced after enhancement processing by the CBAM-CDAE model, indicating that noise has been removed and the signal curve is smoother. Combined with... Figure 16 It can be seen that the enhanced signal not only effectively filters out most of the background white noise and low-frequency noise, but also largely retains the key characteristics of the main frequency signal.

[0049] By comparison Figure 13 and Figure 16 It is evident that the enhancement performance of the CBAM-CDAE model does not surpass that of the traditional SSA-SVMD method. The essence lies in the fact that CBAM-CDAE does not directly achieve superior enhancement performance, but rather simulates and solidifies its enhancement mode by learning the "pseudo-pure" signal labels generated by traditional methods. The core advantage of this method is that it improves enhancement efficiency while maintaining the enhancement accuracy of traditional methods, achieving a leap from "effective enhancement" to "fast and effective enhancement."

[0050] To verify the efficiency of the method of the present invention in terms of enhancement, the time consumed by different enhancement methods was compared, as shown in Table 7 below: Table 7. Time (s) required for test set augmentation at various measurement points using different methods ; The data in Table 7 shows that the time required by the method of the present invention is much less than that of the traditional augmentation methods SSA-SVMD and SSA-CEEMDAN. This verifies that the method of the present invention, by integrating traditional augmentation methods and deep learning algorithms, achieves effective denoising while optimizing the efficiency of the augmentation process, demonstrating the comprehensive advantages of the method of the present invention in the field of augmentation.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for flood sluice vibration signal enhancement based on physical information and deep learning, characterized in that, Includes the following steps: Step S1: Collect the prototype vibration response signal of the floodgate as the original noisy signal; Step S2: The original noisy signal is processed using a physical information enhancement method based on Singular Spectral Analysis (SSA) and Adaptive Successive Variational Mode Decomposition (SVMD) to obtain a clean signal with a high signal-to-noise ratio, and a deep learning training sample set is constructed accordingly. The combined physical information enhancement method uses SSA to preprocess the original noisy signal to filter out high-frequency white noise, and then uses SVMD to adaptively decompose the processed signal. The Newton-Raphson optimization algorithm (NRBO) is introduced during the decomposition process. The envelope entropy is used as the fitness function to evaluate the quality of SVMD decomposition. Through NRBO iterative optimization, the effective modal component reconstructed signal is selected, the interference of low-frequency noise is filtered out, and a clean signal with high signal-to-noise ratio is obtained. Step S3: Construct an intelligent signal enhancement model based on attention mechanism-convolutional denoising autoencoder CBAM-CDAE. Use the original noisy signal as input and the clean signal as label to train the intelligent signal enhancement model and obtain the trained signal enhancement model. The intelligent signal enhancement model uses a Convolutional Block Attention Module (CBAM) that is deeply integrated with the encoding and decoding structure of the CDAE through a two-layer attention mechanism. CBAM dynamically adjusts the weights of each feature channel through the channel attention module to enhance redundant information related to noise. At the same time, it uses a spatial attention module to focus on key regions of the signal to enhance the discriminative learning of local features, thereby improving the denoising performance and signal fidelity of the model. Step S4: Input the vibration response signal of the floodgate to be enhanced into the trained signal enhancement model, and output the vibration response signal with enhanced signal-to-noise ratio.

2. The physical information and deep learning-based flood sluice vibration signal enhancement method according to claim 1, characterized in that, In step S2, the original noisy signal is processed to obtain a clean signal with a high signal-to-noise ratio. The process is as follows: Step S21: Use Singular Spectrum Analysis (SSA) to preprocess the original noisy signal. Reconstruct the signal by trajectory matrix embedding, singular value decomposition, grouping, and diagonal averaging to filter out high-frequency white noise. Step S22, introduce Newton-Raphson optimization algorithm NRBO to minimize the envelope entropy as fitness function, optimize the quadratic penalty factor parameter of SVMD ; Step S23: Adaptively decompose the reconstructed signal after SSA preprocessing using SVMD with optimized quadratic penalty factor parameters to obtain a series of intrinsic mode components (IMFs). Step S24: Calculate the Pearson correlation coefficient between each IMF component and the original noisy signal, and screen the effective IMF components according to the preset threshold for reconstruction to filter out low-frequency environmental interference and finally obtain a pure signal that reflects the true vibration characteristics of the structure.

3. The physical information and deep learning-based flood sluice vibration signal enhancement method according to claim 2, characterized in that, The reconstruction process of the original noisy signal in step S21 is as follows: Step S211: Trajectory matrix embedding; Let the original noisy signal be denoted as an original time series of length , select a proper window size , map the original time series to vectors of length , compose a trajectory matrix : （1）； In the above formulae, ; Step S212, Singular value decomposition; Definition of the covariance matrix Eigenvalue decomposition of the covariance matrix and eigenvectors ;​ In singular value analysis: （2）； （3）； In the above formula, is the Lth singular component matrix; is the basic matrix; is the kth singular value; is the left matrix, is the right matrix, and are orthogonal matrices; Combining equations (2) and (3), we get: （4）； In the above formulae, is the transpose of the right matrix ; Step S213: Grouping; calculating a variance contribution rate of each singular value , is the i-th singular value, and the first matrix components whose cumulative contribution rate exceeds a threshold value are selected , the trajectory matrix is divided into an effective signal matrix and a noise interference matrix ; Step S214: Diagonal average; Effective signal matrix Converted into a reconstructed signal sequence with the same length as the measured vibration response sequence. : （5）； （6）； In the above formula, For matrix elements; This is the nth sampling point; is the number of columns in the trajectory matrix.

4. The method for enhancing the vibration signal of a floodgate based on physical information and deep learning according to claim 3, characterized in that, In step S22, the quadratic penalty factor parameter of SVMD is optimized. The process is as follows: Step S221: Population initialization; Preset secondary penalty factor parameters The range of values ​​is ,exist Internal random generation The candidate solutions are Values ​​constitute the initial population. The first in the population candidate solutions The position is represented as: （7）； In the above formula, Indicates the first The first candidate solution Position in each dimension , , The dimension of the population; express Uniformly distributed random numbers between; Step S222: Calculate fitness; With candidate solutions As a penalty factor for SVMD, the reconstructed signal sequence The decomposition yields a set of modal components; Calculate the envelope entropy of each modal component. The average value is then taken as the fitness value of the current individual. : （8）； （9）； （10）； In the above formula, This represents the length of the original time series, i.e., the total number of sampling points. For the index of the sampling point, ; For the first Normalized envelope amplitude of each sampling point; Represents the natural logarithm; The instantaneous amplitude envelope is obtained after the signal undergoes the Hilbert transform; This represents the sum of the envelope amplitudes of all sampling points, after normalization. ; This represents the total number of modal components. Record the globally optimal individual and the worst individual ; Step S223, NRSR search; NRSR derives the position update formula by approximating the first and second derivatives, and introduces adaptive parameters during the iteration process. To enhance the search capability of the NRSR optimization process, the improved NRSR search process is expressed as follows: （11）； （12）； In the above formula, , They represent the first , There are 10 candidate solutions; The objective function value; , They represent respectively to Find the first and second derivatives; express Uniformly distributed random numbers between; Indicates the current iteration number; Indicates the maximum number of iterations; Step S224, Trap Avoidance Operator (TAO); TAO avoids getting trapped in local optima by randomly selecting parameters and updating population positions, further improving the global search capability of NRSR. The update process of the trap avoidance operator TAO is represented as follows: （13）； In the above formula, and These represent the updated position and the original position of an individual in the population, respectively. express Uniformly distributed random numbers between; and They are and Random numbers between; Indicates the worst position; Indicates the optimal position; For random disturbance terms, ; Step S225: Iterative optimization; For candidate solutions Calculate candidate positions according to step S223 Generate the updated version according to step S224 Combined with step S222, calculate , , The fitness of the solution is used to select the solution with the best fitness as the optimal solution. ,like Exceeding the preset range Then take and The nearest boundary value; When the maximum number of iterations is reached If the fitness change is less than a preset threshold, then the iteration stops and the global optimal solution is output. As a secondary penalty factor for SVMD.

5. The method for enhancing the vibration signal of a floodgate based on physical information and deep learning according to claim 4, characterized in that, In step S23, the SVMD with optimized quadratic penalty factor parameters is used to adaptively decompose the reconstructed signal after SSA preprocessing. The process is as follows: Assume the input signal is : （14）； In the above formula, For the first First-order modal signal; Represents the residual signal, from the 1st order to the 2nd order. The sum of the first-order modal signals constitutes the composition; For the first First-order modal signal; This is an unprocessed signal; To ensure that the above assumptions are met, the following constraints need to be set; Modal compactness constraints: ensuring modal compactness Closely around its center frequency , represented as: ； In the above formula, Indicates the first The center frequency of the first mode; Represents the convolution operator; Indicates time The partial derivative; Represents the Dirac function; The imaginary unit; is the base of the natural logarithm; Spectral overlap minimization constraint: in modal With the extracted Under the condition of minimum spectral overlap of the first mode, a suitable filter is selected, and its response frequency is [value missing]. Represented as: （15）； In the above formula, For the second-order penalty factor parameter; Angular frequency; The established constraint for minimizing spectral overlap is expressed as: （16）； In the above formula, This is the impulse response of the filter; Signal fidelity constraints: for signals from the first order to... Similar processing is applied to the first mode, setting... For the first The impulse response and frequency response of the first-order mode filter are obtained. : （17）； The established signal fidelity constraints are: （18）； Signal complete reconstruction constraint: Ensures that the signal can be completely reconstructed, expressed as: （19）； Based on the above constraints, establish a mathematical model for constraint minimization: （20）； In the above formula, To balance , and The secondary penalty factor parameter; This represents the IMF of the Lth modal component to be extracted. This is the accumulated signal of the extracted modes; This is the original observation signal; By minimizing the total cost function The modal components of each order are solved sequentially.

6. The method for enhancing the vibration signal of a floodgate based on physical information and deep learning according to claim 5, characterized in that, The formulas for calculating the Pearson correlation coefficient and threshold in step S24 are expressed as follows: （21）； （22）； In the above formula, The Pearson correlation coefficient; The covariance between each IMF component and the noise signal; The standard deviation of each IMF component; The standard deviation of the noisy signal; The threshold for the correlation coefficient; The maximum correlation coefficient; When the correlation coefficient R between the IMF and the original signal is greater than the threshold μ, it is retained, and the retained IMF components are superimposed to obtain the enhanced signal after physical information processing.

7. The method for enhancing the vibration signal of a floodgate based on physical information and deep learning according to claim 6, characterized in that, The intelligent signal enhancement model constructed in step S3, based on the attention mechanism-convolutional denoising autoencoder CBAM-CDAE, adopts an encoder-decoder symmetrical structure, and embeds CBAM attention modules after the key layers of the encoder and decoder; the training process of the intelligent signal enhancement model is as follows: Step S31: Data preparation; Using the original noisy signal from step S1 as input and the clean signal obtained in step S2 as the training label of the model, the original noisy signal and the clean signal are divided into training set, validation set and test set in a ratio of 8:1:

1. Step S32: Signal segmentation; Intelligent signal enhancement models require a fixed-length input; therefore, the long signal needs to be segmented into multiple short segments. Assume the length of each segment is... The output length after n convolutional and pooling layers is... Satisfy the following formula: （23）； In the above formula, , These are the filter sizes for the convolutional layer and the pooling layer, respectively. , These are the step sizes corresponding to these layers; Indicates the amount of zero fill used; Since the sizes before and after convolution are the same, the filter size of the merged layer is... and step length If they are equal, then formula (23) simplifies to: （24）； Step S33: Model training; The model is trained using supervised learning, taking noisy signal segments as input and corresponding clean signal segments as the target output. A modified linear unit (ReLU) non-linear activation function is used after each convolutional layer, and a sigmoid function is used in the output layer. Mean squared error (MSE) is used as the loss function during model training. Training stops when the validation loss changes less than a preset threshold for 10 consecutive times. The model parameters with the minimum validation loss are saved to obtain the trained signal enhancement model.

Images