A Ф-OTDR signal denoising method, system and storage medium based on improved VMD algorithm

Abstract

The invention discloses a Ф-OTDR signal denoising method, system and storage medium based on an improved VMD algorithm, and relates to the field of signal processing. The method comprises: optimizing the penalty factor and the number of decompositions of the VMD algorithm by using a ZOA algorithm to obtain an optimal penalty factor and the number of decompositions; inputting the acquired original Ф-OTDR signal into the VMD algorithm running with the optimal penalty factor and the number of decompositions to obtain multiple IMF components; using a deep learning model to determine the first IMF component that needs to participate in signal reconstruction in the IMF component, and the remaining IMF components to be processed; using an improved wavelet threshold algorithm to denoise the IMF component to be processed to obtain a denoised IMF component, and using a deep learning model to determine the second IMF component suitable for signal reconstruction in the denoised IMF component; reconstructing the first IMF component and the second IMF component to obtain a denoised signal. The present application can suppress noise more effectively, retain more signal details, and thus improve the clarity and analyzability of the signal.

Description

Technical Field

[0001] The invention relates to the technical field of signal processing, and in particular to a Φ-OTDR signal denoising method, system and storage medium based on an improved VMD algorithm. Background Art

[0002] In the Ф-OTDR system, the presence of noise poses a serious challenge to the accurate interpretation of the signal. After drilling, it is necessary to conduct preliminary measurements of various parameters of the mine to determine whether it meets the expected production requirements. Therefore, this requires the Ф-OTDR system to have the ability to deploy and analyze in unfamiliar environments.

[0003] Due to differences in environmental noise and other factors, the specific distribution of noise cannot be obtained when the Ф-OTDR system is first deployed. The estimated prior distribution often has a large gap with the actual distribution. Therefore, denoising in this case requires an algorithm that does not rely on prior knowledge of the noise distribution, namely the VMD algorithm. Variational mode decomposition (VMD) is based on completely non-recursive decomposition, which can decompose the noisy original signal into a series of IMF components and calculate the center frequency of each IMF component, which can effectively solve the phenomenon of modal aliasing in EMD.

[0004] However, each modal component decomposed by VMD contains both signal and noise. If any component is removed at random, the accuracy of the reconstructed signal will be affected. In addition, how to optimize the penalty factor and the number of decompositions of VMD is crucial to the accuracy of the results. Summary of the invention

[0005] The object of the present invention is to provide a Φ-OTDR signal denoising method, system and storage medium based on an improved VMD algorithm, so as to improve the denoising effect of the Φ-OTDR signal.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present application discloses a Φ-OTDR signal denoising method based on an improved VMD algorithm, comprising:

[0008] The ZOA algorithm is used to optimize the penalty factor and decomposition number of the VMD algorithm to obtain the optimal penalty factor and decomposition number;

[0009] The acquired original Ф-OTDR signal is input into the VMD algorithm running with the optimal penalty factor and decomposition number to obtain multiple IMF components;

[0010] A deep learning model is used to determine the first IMF component among the IMF components that needs to participate in signal reconstruction, and the remaining IMF components are to-be-processed;

[0011] An improved wavelet threshold algorithm is used to perform denoising on the IMF component to be processed to obtain a denoised IMF component, and a deep learning model is used to determine a second IMF component in the denoised IMF component that is suitable for signal reconstruction;

[0012] Reconstructing the first IMF component and the second IMF component to obtain a denoised signal;

[0013] The framework of the deep learning model includes a multi-scale CNN, a first feature extraction unit, a second feature extraction unit, and a fusion module, an adaptive gated LSTM, and an output decision module connected in sequence. The multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are arranged in parallel, and the outputs of the multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are connected to the fusion module.

[0014] Furthermore, the ZOA algorithm is used to optimize the penalty factor and the number of decompositions of the VMD algorithm, and the optimal penalty factor and the number of decompositions are obtained:

[0015] Step 11: Set the penalty factor and the decomposition number boundary, determine the optimization space, and randomly initialize the population in the optimization space:

[0016] ;

[0017] in, is an individual of the population, each Represents a set of penalty factors and the number of decompositions of VMD, is the lower boundary of the optimization space, is the upper boundary of the optimization space, r is a random number in [0,1];

[0018] Step 12: Apply the VMD algorithm to each individual in the population to process the original Ф-OTDR signal to obtain the reconstructed Ф-OTDR signal, calculate the reconstruction error according to the original Ф-OTDR signal and the reconstructed Ф-OTDR signal, and select the individual with the smallest reconstruction error as the current optimal individual;

[0019] Step 13: The current optimal individual is updated iteratively in the foraging phase and the defense phase of the ZOA algorithm to obtain the optimal individual; the optimal individual is the optimal penalty factor and the number of decompositions.

[0020] Furthermore, the improved wavelet threshold algorithm is a wavelet threshold algorithm using an improved threshold function, and the expression of the improved threshold function is:

[0021] ;

[0022] in, is the wavelet coefficient before processing, is the processed wavelet coefficient, is the threshold value set, is a sign function, is the maximum decomposition scale, , K is the length of the wavelet coefficients, , , N is the length of the signal; and is the adjustment factor used to adjust, The speed of change, It is used to adjust the degree of approximation of the function to the asymptote. The value range is , To adjust the threshold The vertical coordinate value of the cutoff point is in the range of .

[0023] Furthermore, it also includes using an improved simulated annealing algorithm to determine the threshold of the improved threshold function, including:

[0024] Step 41: Set the initial temperature , let the temperature T = , randomly generate an initial solution As a candidate threshold for improving the threshold function, and calculating the initial solution The corresponding objective function value E( );

[0025] Step 42, let T = KT, where K is the temperature drop rate, ranging from 0 to 1;

[0026] Step 43: Current solution Apply random perturbations, the neighborhood distribution of the perturbations is normal distribution, the variance is the difference between the current solution and the optimal solution, and in the current solution A new solution is generated in the neighborhood , and calculate the new solution The corresponding objective function value E( ) , calculate the difference between the new solution objective function value and the initial solution objective function value: ( )=E( )- E( ) ;

[0027] Step 44: If the objective function value of the new solution is better, accept the new solution as the current solution, otherwise Decide whether to accept the inferior new solution;

[0028] Step 45: If no new solution is accepted for N consecutive times, increase the variance to the historical maximum value to increase the exploration intensity, and set T = (1 + N / 20) T, and return to step 43, where N is a constant less than 10;

[0029] Step 46, at temperature T, repeat step 43 and step 44 for L times;

[0030] Step 47: When the temperature drops to the preset termination temperature, the algorithm ends and the last accepted solution is output as the threshold of the improved threshold function; otherwise, the algorithm returns to step 42.

[0031] Furthermore, the expression of the DTW algorithm is:

[0032] ;

[0033] in, is the distance matrix Line The cumulative distance when the column is are the values ​​of two time series, is the distance between the i-th point of the q sequence and the j-th point of the c sequence, To find the minimum cumulative distance path to the current position (i, j) from the cumulative distance matrix.

[0034] Further, the using a deep learning model to determine the first IMF component among the IMF components that needs to participate in signal reconstruction includes:

[0035] Multi-scale CNN extracts the features of each IMF component in the time domain and obtains the multi-scale time domain feature representation of the IMF component;

[0036] The first feature extraction unit performs frequency domain feature extraction on each IMF component to obtain a frequency domain feature representation of the IMF component;

[0037] The second feature extraction unit divides each IMF component according to the ERB frequency band, and calculates the MFCC and the first-order derivative and the second-order derivative of each frequency band after the division to obtain the MFCC features of the IMF component;

[0038] The fusion module fuses the multi-scale time domain feature representation of the IMF component, the frequency domain feature representation of the IMF component and the MFCC feature of the IMF component to obtain a fusion output;

[0039] The adaptive gated LSTM integrates the fusion output and outputs the fusion feature vector;

[0040] The output decision module receives the fused feature vector and outputs a judgment result to obtain the first IMF component that needs to participate in signal reconstruction.

[0041] Furthermore, the fusion module includes a merging layer.

[0042] Furthermore, the output decision module includes a fully connected layer.

[0043] Further, the multi-scale CNN includes a feature capture module, a first maximum pooling layer, a convolutional layer, a second maximum pooling layer, a global average pooling layer, a first fully connected layer, a random dropout layer, and a second fully connected layer, which are arranged in sequence;

[0044] The feature capture module includes a plurality of parallel convolutional layers and a merging layer connected after the plurality of parallel convolutional layers; wherein the convolution kernel size of each convolutional layer in the plurality of parallel convolutional layers is different.

[0045] In a second aspect, the present application discloses a Ф-OTDR signal denoising system based on an improved VMD algorithm, comprising:

[0046] A memory for storing instructions;

[0047] A processor is used to operate according to the instructions to execute the steps of the method as described in any one of the first aspects.

[0048] In a third aspect, the present application discloses a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements the denoising method as described in any one of the first aspects.

[0049] According to the above technical solution, the embodiments of the present invention have at least the following effects:

[0050] The present application firstly uses the ZOA algorithm to dynamically adjust the penalty factor and the number of decompositions of VMD, and can automatically find the optimal parameter combination, thereby improving the adaptability and efficiency of the algorithm, ensuring that the decomposition result is more consistent with the real structure of the original signal, and separating the multi-scale characteristics of the signal by obtaining multiple IMF components through VMD running with optimal parameters; applying the improved wavelet threshold algorithm to the IMF components not participating in the reconstruction for denoising can more accurately remove the noise, while protecting the signal details and improving the denoising effect; using a deep learning model to determine the first IMF component in the IMF component that needs to participate in the signal reconstruction and the second IMF component in the denoised IMF component that is suitable for signal reconstruction, the deep learning model can automatically learn the signal characteristics, intelligently screen out useful IMF components, avoid the subjectivity and errors caused by manual settings, and improve the processing efficiency; the denoised signal is obtained by reconstructing the first IMF component and the second IMF component, and the reconstructed signal is closer to the real signal, ensuring the high quality of the final signal, improving the signal resolution and reliability, and achieving a better denoising effect;

[0051] In the framework of the deep learning model of the present application, the multi-scale CNN can extract features in the time domain for each IMF component, the first feature extraction unit and the second feature extraction unit can obtain frequency domain feature representation and MFCC features, and the adaptive gated LSTM can integrate the captured long-term dependencies and dynamic features for the output processing of the fusion module, and output a fused feature vector containing all IMF information. The local and global features of the signal are captured by the multi-scale CNN, and the long-term dependencies of the sequence data are retained and learned by the adaptive gated LSTM, so as to achieve accurate evaluation of the IMF and intelligent screening, provide strong support for subsequent signal reconstruction decisions, ensure the quality and accuracy of the reconstructed signal, and thus improve the efficiency of the overall signal processing and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the denoising method of the present invention;

[0053] Figure 2 A flowchart of an improved simulated annealing algorithm in the present invention;

[0054] Figure 3 This is a decomposition effect diagram of the VMD algorithm in the present invention processing the Ф-OTDR signal;

[0055] Figure 4 This is a comparison diagram of the denoised signal and the original signal by the method of the present invention;

[0056] Figure 5 This is a schematic diagram of the framework of the deep learning model of the present invention;

[0057] Figure 6 This is a connection diagram of each layer of the deep learning model framework of the present invention. DETAILED DESCRIPTION

[0058] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0059] The denoising method of the present application imports the noisy signal as input data into the variational mode decomposition (VMD) algorithm. The zebra intelligent optimization algorithm (ZOA) is applied to optimize the parameter selection of VMD to ensure the effectiveness and accuracy of the decomposition process. The noisy signal is decomposed by the VMD algorithm to obtain multiple intrinsic mode components (IMFs). The deep learning model output is used to determine which IMFs need to participate in signal reconstruction. The IMFs that do not need to participate in the reconstruction are subjected to improved wavelet threshold denoising to reduce the noise component, and the denoised IMFs are processed using a deep learning model to screen the signals that require parameter reconstruction again. Finally, these signals are reconstructed to obtain the denoising results, and the algorithm is executed. The final reconstructed signal and related evaluation indicators, such as signal-to-noise ratio, are output.

[0060] Example 1

[0061] like Figure 1 As shown, the present application discloses a Φ-OTDR signal denoising method based on an improved VMD algorithm, including step 1, using the ZOA algorithm to optimize the penalty factor and decomposition number of the VMD algorithm to obtain the optimal penalty factor and decomposition number. Step 2, input the acquired original Φ-OTDR signal into the VMD algorithm running with the optimal penalty factor and decomposition number to obtain multiple IMF components; step 3, using a deep learning model to determine the first IMF component that needs to participate in signal reconstruction in the IMF component, and the remaining is the IMF component to be processed; step 4, using an improved wavelet threshold algorithm to denoise the IMF component to be processed, obtain a denoised IMF component, and use a deep learning model to determine the second IMF component suitable for signal reconstruction in the denoised IMF component; step 5, reconstruct the first IMF component and the second IMF component to obtain a denoised signal. Among them, the framework of the deep learning model includes a multi-scale CNN, a first feature extraction unit, a second feature extraction unit and a fusion module, an adaptive gated LSTM and an output decision module connected in sequence. The multi-scale CNN, the first feature extraction unit and the second feature extraction unit are arranged in parallel, and the outputs of the multi-scale CNN, the first feature extraction unit and the second feature extraction unit are connected to the fusion module.

[0062] The present application firstly uses the ZOA algorithm to dynamically adjust the penalty factor and the number of decompositions of VMD, so as to automatically find the optimal parameter combination, improve the adaptability and efficiency of the algorithm, ensure that the decomposition result is more consistent with the real structure of the original signal, and separate the multi-scale characteristics of the signal by obtaining multiple IMF components through VMD running with optimal parameters; the improved wavelet threshold algorithm is applied to the IMF components not participating in the reconstruction for denoising, so as to more accurately remove the noise, protect the signal details and improve the denoising effect; the deep learning model is used to determine the first IMF component in the IMF component that needs to participate in the signal reconstruction and the second IMF component in the denoised IMF component that is suitable for the signal reconstruction, the deep learning model can automatically learn the signal characteristics, intelligently screen out the useful IMF components, avoid the subjectivity and errors caused by manual setting, and improve the processing efficiency; the denoised signal is obtained by reconstructing the first IMF component and the second IMF component, and the reconstructed signal is closer to the real signal, so as to ensure the high quality of the final signal, improve the resolvability and reliability of the signal, and achieve better denoising effect.

[0063] The present application is further described below through detailed steps.

[0064] Step 1: Use the ZOA algorithm to optimize the penalty factor and decomposition number of the VMD algorithm to obtain the optimal penalty factor and decomposition number, including:

[0065] Step 11: Set the penalty factor and the decomposition number boundary, determine the optimization space, and randomly initialize the population in the optimization space:

[0066] ;

[0067] in, is an individual of the population, each Represents a set of penalty factors and the number of decompositions of VMD, is the lower boundary of the optimization space, is the upper boundary of the optimization space, and r is a random number in [0,1].

[0068] Step 12: Apply the VMD algorithm to each individual in the population to process the original Ф-OTDR signal to obtain the reconstructed Ф-OTDR signal, calculate the reconstruction error based on the original Ф-OTDR signal and the reconstructed Ф-OTDR signal, and select the individual with the smallest reconstruction error as the current optimal individual.

[0069] When calculating the reconstruction error of the original Ф-OTDR signal and the reconstructed Ф-OTDR signal, one or more combinations of mean square error, root mean square error, mean absolute error, correlation coefficient and signal-to-noise ratio may be used.

[0070] In a specific embodiment, the reconstruction error is calculated in the form of a combination of root mean square error and signal-to-noise ratio. The root mean square error can provide an accurate numerical error measure between the reconstructed signal and the original signal, and the signal-to-noise ratio is directly related to the performance of signal detection, especially the evaluation of signal quality under a noisy background. It is suitable for the needs of weak signal detection, high requirements for noise suppression, and accuracy of vibration event positioning during Ф-OTDR signal detection. In summary, the combination of the two ensures signal accuracy while also focusing on signal quality and performance in a noisy environment, providing a basis for efficient and accurate fiber optic monitoring and fault location in practical applications.

[0071] Step 13: The current optimal individual is iteratively updated in the foraging phase and the defense phase of the ZOA algorithm to obtain the optimal individual, which is the optimal penalty factor and the number of decompositions.

[0072] This step is a common step of the ZOA algorithm. A simple example is given below:

[0073] In the foraging phase, the positions of individual zebras in the population are updated with the following strategy:

[0074] ;

[0075] ;

[0076] in, is the state of individual zebra foraging, I is the random value of the population change control parameter {1,2}, and it is updated by the specified number of iterations.

[0077] In the defense phase, the positions of individual zebras in the population are updated. The zebras will adopt different strategies when dealing with different predators, and their probabilities are 50% respectively. The update strategies in the two situations are:

[0078] ;

[0079] Where t is the number of iterations, T is the maximum number of iterations, and R is a constant of 0.01. is the probability of switching between the two strategies, which is a random number in [0,1], This is the state of a zebra being attacked.

[0080] When updating the individual zebra position, if the objective function of the zebra individual in the new position has a better value, the new position is accepted and updated using the following formula:

[0081] ;

[0082] Iterate until the termination condition is met or the number of iterations is reached, and output the optimal individual obtained by the iteration, that is, the optimal penalty factor and number of decompositions.

[0083] Step 2: Input the acquired original Ф-OTDR signal into the VMD algorithm running with the optimal penalty factor and decomposition number to obtain multiple IMF components.

[0084] Using the VMD algorithm to process the Ф-OTDR signal is a conventional process and will not be further described in this application. It is worth noting that the VMD algorithm used in this application is based on the optimal penalty factor and number of decompositions obtained by optimizing the ZOA algorithm, and has good performance in processing the Ф-OTDR signal. Figure 3 As shown in Figure 1, when the Ф-OTDR signal is deeply analyzed, nine intrinsic mode functions (IMF components) are obtained. Figure 3 From (1), (2), (3), (4), (5) and (6), we can see that the first six IMF components show significant transients and irregularities, which indicates that they mainly contain noise components. Figure 3 From (7) in , we can see that the seventh IMF component can clearly show the shape of the main components of the signal, but from the waveform diagram, we can see that it still has a lot of noise to a certain extent. Figure 3 From (8) and (9), we can see that the last two IMF components present a relatively smooth and regular waveform, which clearly represents the key information of the signal. This reflects the excellent performance of the VMD algorithm in signal decomposition, which can effectively separate the signal and noise, and provides a solid foundation for the fine analysis of the signal.

[0085] The VMD algorithm of this application successfully decomposes the original Ф-OTDR signal into a series of IMF components, among which the first six IMF components mainly carry noise components, while the last three IMF components are rich in the main information of the signal, and the seventh IMF component is still interfered by noise to a certain extent. This decomposition strategy not only effectively distinguishes between signals and noise, but also provides a powerful means for noise suppression and signal feature extraction.

[0086] Step 3: Use a deep learning model to determine the first IMF component among the IMF components that needs to participate in signal reconstruction, and the remaining IMF components are to be processed.

[0087] The framework of the deep learning model in this application is as follows Figure 5 As shown, it includes a multi-scale CNN, a first feature extraction unit, a second feature extraction unit, and a fusion module, an adaptive gated LSTM, and an output decision module connected in sequence. The multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are arranged in parallel, and the outputs of the multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are connected to the fusion module.

[0088] like Figure 6As shown, the fusion module includes a merging layer, and the output decision module includes a fully connected layer. The multi-scale CNN includes a feature capture module, a first maximum pooling layer, a convolution layer, a second maximum pooling layer, a global average pooling layer, a first fully connected layer, a random inactivation layer, and a second fully connected layer arranged in sequence; the feature capture module includes multiple parallel convolution layers and a merging layer connected after the multiple parallel convolution layers; wherein the convolution kernel size of each convolution layer in the multiple parallel convolution layers is different. The training process of the deep learning model is a conventional process, and this application will not be further described.

[0089] Specifically, the multi-scale CNN is mainly used to extract multi-level features from IMF. It consists of three parallel 1D convolutional layers, each with different convolution kernel sizes (such as 3, 5, and 7) to capture the characteristics of the signal at different time scales. The convolution layer is usually followed by a ReLU activation function to increase the nonlinear expression ability of the network. The output of the three parallel convolutional layers is then accepted by a merging layer and pooled through the first maximum pooling layer to form a higher-dimensional feature vector that contains rich multi-scale information. This feature vector is then input into a new 1D convolutional layer with a 3*3 convolution kernel, pooled through the second maximum pooling layer, and then output to the global average pooling layer. Finally, it passes through the first fully connected layer and a random dropout layer with a probability of 0.5, and is output by the second fully connected layer into the next module.

[0090] Data preprocessing and feature extraction: Use multi-scale CNN to extract features in the time domain for each IMF component, and obtain multi-scale time domain feature representation, totaling 256 dimensions; Adaptive Gabor transform is used to extract features in the frequency domain for each IMF component, and obtain its frequency domain feature representation, totaling 256 dimensions; Each IMF component is divided into 34 according to the ERB frequency band, and then the MFCC and its first-order derivative and second-order derivative of each frequency band are calculated, totaling 102 dimensions. This step converts each IMF component into a set of 614-dimensional feature vectors.

[0091] Time series modeling and fusion: The fusion module fuses the multi-scale time domain feature representation of the IMF component, the frequency domain feature representation of the IMF component, and the MFCC feature of the IMF component to obtain a fusion output, which is used as the input of the adaptive gated LSTM in the middle of the neural network. Through the time series modeling capability of the LSTM, the features of different IMF components are further fused. The LSTM layer integrates the captured long-term dependencies and dynamic features and outputs a fusion feature vector containing all IMF information.

[0092] Between the multi-scale CNN and the adaptive gated LSTM, the network designs an intermediate fusion module to integrate the feature representations of the multi-scale CNN and the adaptive gated LSTM. This fusion module is mainly a merging operation, which directly splices the time domain feature output of the CNN and the frequency domain features extracted by the adaptive Gabor change, as well as the MFCC after the IMF is divided by the ERB band and its first-order derivative and second-order derivative. The features are input into the adaptive gated LSTM module. The adaptive gated LSTM is responsible for processing time series data, especially long-term dependencies in long sequence data. In this application, the LSTM layer is designed to be adaptively gated, which means that it can dynamically adjust its gating mechanism according to the characteristics of the input data to more effectively remember or forget information. This adaptability enables the LSTM to more accurately identify and retain important features in the IMF while filtering out irrelevant or redundant information.

[0093] The output decision module receives the fused feature vector and outputs the judgment result to obtain the first IMF component that needs to participate in signal reconstruction. Specifically, according to the network output result, it is judged whether each IMF is suitable for being directly used as the reconstructed signal. The network output result of 1 is directly used as the first IMF component that needs to participate in signal reconstruction. The network output result of 0 is used as the remaining IMF component to be processed and needs to be further excluded to determine whether it can be finally adopted.

[0094] The entire network starts with a multi-scale CNN, undergoes feature fusion, enters an adaptive gated LSTM, and finally reaches the output layer, forming an end-to-end deep learning architecture that can automatically learn and extract meaningful features from the original IMF, and then make intelligent decisions about signal reconstruction based on these features. Through careful design and training, this network can demonstrate excellent performance in complex signal processing tasks.

[0095] The form of multi-scale CNN+adaptive gated LSTM in this application is that it can integrate information in the time and frequency domains, capture local and global features of the signal through multi-scale CNN, and use adaptive gated LSTM to retain and learn the long-term dependencies of sequence data, thereby achieving accurate evaluation of IMF. Specifically, the multi-scale CNN module first performs multi-level feature extraction on the IMF to ensure that details and patterns at different time scales can be captured; then, the adaptive gated LSTM component further processes these features, uses its internal memory mechanism to analyze the dynamic changes of the sequence, identifies which IMFs carry key information, and can effectively contribute to signal reconstruction, while eliminating those noise or redundant components. The entire network finally achieves intelligent screening by fusing multimodal information, that is, features of different scales and the dynamic characteristics of time series, providing strong support for subsequent signal reconstruction decisions, ensuring the quality and accuracy of the reconstructed signal, and thus improving the overall signal processing and analysis efficiency.

[0096] Step 4: Use an improved wavelet threshold algorithm to denoise the IMF component to be processed, obtain a denoised IMF component, and use a deep learning model to determine a second IMF component suitable for signal reconstruction in the denoised IMF component. Step 4 is described in detail below:

[0097] The prior art describes the following: Wavelet transform has a strong data-free correlation, which can concentrate the energy of the signal in some large wavelet coefficients in the wavelet domain, while the energy of the noise is distributed in the entire wavelet domain. Therefore, after wavelet decomposition, the amplitude of the wavelet coefficient of the signal is greater than the amplitude of the noise coefficient. It can be considered that the wavelet coefficient with a relatively large amplitude is generally dominated by the signal, while the coefficient with a relatively small amplitude is largely noise. Therefore, the threshold method is used to reduce most of the noise coefficients to 0, that is, the coefficients smaller than the threshold are removed as interference noise, and then wavelet reconstruction is performed to achieve noise reduction.

[0098] The denoising process of wavelet threshold includes wavelet transform decomposition, threshold processing and signal reconstruction. When the signal is decomposed by wavelet, the larger the number of decomposition layers, the more obvious the different characteristics of noise and signal are, and the more conducive to the separation of the two. However, the larger the number of decomposition layers, the greater the distortion of the reconstructed signal will be, which will have a poor effect on the effect of signal denoising to a certain extent. Therefore, how to choose the number of decomposition layers to solve the contradiction between the signal-noise separation effect and the distortion of the reconstructed signal becomes an important issue in wavelet threshold denoising.

[0099] There are two types of traditional threshold functions, namely hard threshold and soft threshold. However, the soft threshold function usually smoothes the wavelet coefficients to reduce noise. However, excessive smoothness will cause the loss of signal details, resulting in excessive denoising. In addition, there is always a constant deviation between the actual wavelet coefficients and the processed wavelet coefficients, which will reduce the quality of the reconstructed signal. The hard threshold function will set the wavelet coefficients less than the threshold to zero, resulting in the complete elimination of small amplitude signals in the signal, causing the loss of effective information. In addition, since the hard threshold function directly truncates the wavelet coefficients, oscillation will occur during the signal reconstruction process, further generating the pseudo-Gibbs phenomenon.

[0100] In order to overcome the shortcomings of the soft threshold function and the hard threshold function in the prior art, the present application proposes an improved threshold function, which can be flexibly adjusted by changing the parameter factor between the values ​​of the hard threshold and the soft threshold function. The expression of the improved threshold function is:

[0101] ;

[0102] in, is the wavelet coefficient before processing, is the processed wavelet coefficient, is the threshold value set, is a sign function, is the maximum decomposition scale, , K is the length of the wavelet coefficients, , , N is the length of the signal; and is the adjustment factor used to adjust, The speed of change, It is used to adjust the degree of approximation of the function to the asymptote. The value range is , To adjust the threshold The vertical coordinate value of the cutoff point is in the range of .

[0103] The improved threshold function proposed in this application can be adjusted by and Two adjustment factors can more conveniently control the threshold function, while eliminating the discontinuity of the threshold function, making the function approach the desired position more quickly. This design ensures the continuity of the threshold function and does not cause mutations, thereby avoiding the occurrence of pseudo-Gibbs phenomenon, so that the denoised signal can maintain the smoothness of the original signal. At the same time, this improvement can reduce the deviation caused by signal compression due to wavelet decomposition and reconstruction, thereby improving the approximation between the reconstructed signal and the real signal.

[0104] In the process of processing actual signals, signals are often distributed in the low-medium frequency band, and noise is distributed in the low-medium-high frequency band according to different types. Therefore, simply taking soft threshold denoising or hard threshold denoising cannot bring effective denoising. Fixed threshold and heuristic threshold denoising are more thorough and more effective in denoising, but they are also easy to mistake useful signals for noise and remove them. Therefore, by selecting the optimal threshold, more useful information can be retained while maintaining a good denoising effect.

[0105] Therefore, wavelet threshold denoising can be converted into an optimization process with signal-to-noise ratio as the target, and simulated annealing is used for optimization to solve the optimal threshold for denoising.

[0106] The simulated annealing algorithm is a heuristic optimization algorithm used to solve various combinatorial optimization and search problems. However, the traditional simulated annealing algorithm has problems such as losing the optimal solution and easily falling into the local optimal solution.

[0107] In order to solve the shortcomings of the above-mentioned traditional simulated annealing algorithm and further improve the search ability of the algorithm, this application improves the traditional simulated annealing algorithm in combination with the characteristics of the target wavelet threshold, such as Figure 2 As shown, the improved method specifically includes the following processes:

[0108] Step 41: Set the initial temperature , let the temperature T = , randomly generate an initial solution As a candidate threshold for improving the threshold function, and calculating the initial solution The corresponding objective function value E( ).

[0109] Step 42: Let T=KT, where K is the temperature drop rate, which ranges from 0 to 1.

[0110] Step 43: Current solution Apply random perturbations, the neighborhood distribution of the perturbations is normal distribution, the variance is the difference between the current solution and the optimal solution, and in the current solution A new solution is generated in the neighborhood , and calculate the new solution The corresponding objective function value E( ), calculate the difference between the new solution objective function value and the initial solution objective function value: ( )=E( )- E( ) .

[0111] Step 44: If the objective function value of the new solution is better, that is ; accept the new solution as the current solution, otherwise follow the probability Decide whether to accept the inferior new solution.

[0112] Step 45: If no new solution is accepted for N consecutive times, increase the variance to the historical maximum value to increase the exploration intensity, and set T=(1+N / 20)T, and return to step 43, where N is a constant less than 10.

[0113] Step 46: At temperature T, repeat step 43 and step 44 L times.

[0114] Step 47: When the temperature drops to the preset termination temperature, the algorithm ends and the last accepted solution is output as the threshold of the improved threshold function; otherwise, the algorithm returns to step 42.

[0115] By improving the simulated annealing algorithm, the algorithm's optimization ability is improved, reducing the situation where the search falls into the local optimal solution. By converting the threshold calculation into an optimization based on the signal-to-noise ratio, the threshold selection problem is solved, and by improving the optimization algorithm, its search ability is enhanced.

[0116] In step 4, the process of using a deep learning model to determine the second IMF component suitable for signal reconstruction in the denoised IMF component is the same as that in step 3. The denoised IMF component is input into the deep learning model. If the result output by the deep learning model is 1, the denoised IMF component is the second IMF component suitable for signal reconstruction. If the result output by the deep learning model is 0, the denoised IMF component is eliminated. This step can further screen the second IMF component suitable for signal reconstruction in the denoised IMF component, obtain more signal details, and improve the denoising effect.

[0117] Step 5: Reconstruct the first IMF component and the second IMF component to obtain a denoised signal.

[0118] The reconstruction process may be completed by directly summing the first IMF component and the second IMF component or by other means. This process is a conventional process and will not be further described in this application.

[0119] In summary, this application has the following advantages:

[0120] The Zebra intelligent optimization algorithm is used to optimize the parameter selection of VMD. The algorithm adopts parallel computing and local search strategies to improve the search efficiency. For VMD parameter optimization problems that require a lot of calculations, the optimization results can be obtained more quickly. The deep learning model is innovatively selected as the standard metric for selecting the reconstructed signal, avoiding the manual operation of manually selecting IMF in the original algorithm. The originally discarded IMF is further denoised, and the denoised signal is attempted to participate in the reconstruction, which improves the utilization of the signal and avoids the loss of effective information; the improved wavelet threshold denoising algorithm is used to reduce the algorithm's dependence on parameter selection.

[0121] The denoising method of this application is used to compare with the original signal, such as Figure 4 As shown, it can be clearly seen that the outline of the denoised signal obtained by the denoising method of the present application is clearer, and the denoised signal part is almost completely wrapped in the original signal. While retaining the effective signal to a great extent, the noise is suppressed.

[0122] The denoising method of this application is compared with other methods (such as wavelet threshold denoising, simulated annealing algorithm optimized VMD denoising, particle swarm algorithm optimized VMD denoising, etc.), and the specific comparison results are shown in Table 1.

[0123] Table 1. Signal-to-noise ratio of 5dB simulation signal after denoising

[0124]

[0125] In summary, compared with the traditional VMD denoising technology, the method of the present application has further improved the denoising effect, can more effectively suppress noise, retain more signal details, thereby improving the clarity and analyzability of the signal.

[0126] Example 2

[0127] In one embodiment of the present application, the present application also discloses a Ф-OTDR signal denoising system based on an improved VMD algorithm, comprising: a memory for storing instructions; a processor for executing the instructions so that the system performs the operation of the denoising method of any of the above embodiments.

[0128] The system may include, but is not limited to: at least one processor, at least one memory, and a bus connecting different system components (including the memory and the processor). The memory stores program code, which can be executed by the processor so that the processor executes the steps of various exemplary embodiments of the present invention described in the "Exemplary Method" section above. For example, the processor may execute the following steps: Figure 1 or Figure 2 Follow the steps shown in .

[0129] The memory may include readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory, and may further include read only memory (ROM).

[0130] The memory may also include a program / utility having a set (at least one) of program modules, such program modules including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0131] The bus may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0132] The system may also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the system, and / or any device that enables the system to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via input / output (I / O) interfaces. Furthermore, the system may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the system via a bus.

[0133] It should be understood that although not shown in the figures, other hardware and / or software modules may be used in conjunction with the system, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0134] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0135] Example 3

[0136] In an exemplary embodiment of the present disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the above method of the present specification is stored. In some possible implementations, various aspects of the present invention can also be implemented in the form of a program product, which includes a program code, and when the program product is run on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present invention described in the above "Exemplary Method" section of the present specification.

[0137] The program product for implementing the above method according to an embodiment of the present invention can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto, and in this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0138] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0139] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may output, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0140] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

[0141] Program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0142] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A Ф-OTDR signal denoising method based on an improved VMD algorithm, characterized in that: include: The penalty factor and decomposition number of the VMD algorithm are optimized using the ZOA algorithm to obtain the optimal penalty factor and decomposition number; The acquired original Ф-OTDR signal is input into the VMD algorithm running with the optimal penalty factor and decomposition number to obtain multiple IMF components; A deep learning model is used to determine the first IMF component among the IMF components that needs to participate in signal reconstruction, and the remaining IMF components are to-be-processed; An improved wavelet threshold algorithm is used to perform denoising on the IMF component to be processed to obtain a denoised IMF component, and a deep learning model is used to determine a second IMF component in the denoised IMF component that is suitable for signal reconstruction; Reconstructing the first IMF component and the second IMF component to obtain a denoised signal; The framework of the deep learning model includes a multi-scale CNN, a first feature extraction unit, a second feature extraction unit, and a fusion module, an adaptive gated LSTM, and an output decision module connected in sequence. The multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are arranged in parallel, and the outputs of the multi-scale CNN, the first feature extraction unit, and the second feature extraction unit are connected to the fusion module.

2. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 1, characterized in that: The ZOA algorithm is used to optimize the penalty factor and number of decompositions of the VMD algorithm. The optimal penalty factor and number of decompositions include: Step 11: Set the penalty factor and the decomposition number boundary, determine the optimization space, and randomly initialize the population in the optimization space: ； in, is an individual of the population, each Represents a set of penalty factors and the number of decompositions of VMD, is the lower boundary of the optimization space, is the upper boundary of the optimization space, r is a random number in [0,1]; Step 12: Apply the VMD algorithm to each individual in the population to process the original Ф-OTDR signal to obtain the reconstructed Ф-OTDR signal, calculate the reconstruction error according to the original Ф-OTDR signal and the reconstructed Ф-OTDR signal, and select the individual with the smallest reconstruction error as the current optimal individual; Step 13: The current optimal individual is updated iteratively in the foraging phase and the defense phase of the ZOA algorithm to obtain the optimal individual; the optimal individual is the optimal penalty factor and the number of decompositions.

3. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 1, characterized in that: The improved wavelet threshold algorithm is a wavelet threshold algorithm using an improved threshold function, and the expression of the improved threshold function is: ； in, is the wavelet coefficient before processing, is the processed wavelet coefficient, is the threshold value set, is a sign function, , is the maximum decomposition scale, , k is the length of the wavelet coefficients, , N is the length of the signal; and is the adjustment factor used to adjust The speed of change, It is used to adjust the degree of approximation of the function to the asymptote. The value range is , To adjust the threshold The vertical coordinate value of the cutoff point is in the range of .

4. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 3, characterized in that: It also includes using an improved simulated annealing algorithm to determine the threshold of the improved threshold function, including: Step 41: Set the initial temperature , let the temperature T = , randomly generate an initial solution As a candidate threshold for improving the threshold function, and calculating the initial solution The corresponding objective function value E( ); Step 42, let T = K1T, where K1 is the temperature drop rate, ranging from 0 to 1; Step 43: Current solution Apply random perturbations, the neighborhood distribution of the perturbations is normal distribution, the variance is the difference between the current solution and the optimal solution, and in the current solution A new solution is generated in the neighborhood , and calculate the new solution The corresponding objective function value E( ) , calculate the difference between the new solution objective function value and the initial solution objective function value: ( )=E( )- E( ) ; Step 44: If the objective function value of the new solution is better, accept the new solution as the current solution, otherwise Decide whether to accept the inferior new solution; Step 45: If no new solution is accepted for N1 consecutive times, increase the variance to the historical maximum value to increase the exploration intensity, and set T = (1 + N1 / 20)T, and return to step 43, where N1 is a constant less than 10; Step 46, at temperature T, repeat step 43 and step 44 L times; Step 47: When the temperature drops to the preset termination temperature, the algorithm ends and the last accepted solution is output as the threshold of the improved threshold function; otherwise, the algorithm returns to step 42.

5. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 1, characterized in that: The using of a deep learning model to determine the first IMF component among the IMF components that needs to participate in signal reconstruction comprises: Multi-scale CNN extracts the features of each IMF component in the time domain and obtains the multi-scale time domain feature representation of the IMF component; The first feature extraction unit performs frequency domain feature extraction on each IMF component to obtain a frequency domain feature representation of the IMF component; The second feature extraction unit divides each IMF component according to the ERB frequency band, and calculates the MFCC and the first-order derivative and the second-order derivative of each frequency band after the division to obtain the MFCC features of the IMF component; The fusion module fuses the multi-scale time domain feature representation of the IMF component, the frequency domain feature representation of the IMF component and the MFCC feature of the IMF component to obtain a fusion output; The adaptive gated LSTM integrates the fusion output and outputs the fusion feature vector; The output decision module receives the fused feature vector and outputs a judgment result to obtain the first IMF component that needs to participate in signal reconstruction.

6. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 5, characterized in that: The fusion module includes a merging layer.

7. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 5, characterized in that: The output decision module includes a fully connected layer.

8. The Φ-OTDR signal denoising method based on the improved VMD algorithm according to claim 1, characterized in that: The multi-scale CNN includes a feature capture module, a first maximum pooling layer, a convolution layer, a second maximum pooling layer, a global average pooling layer, a first fully connected layer, a random dropout layer, and a second fully connected layer, which are arranged in sequence; The feature capture module includes a plurality of parallel convolutional layers and a merging layer connected after the plurality of parallel convolutional layers; wherein the convolution kernel size of each convolutional layer in the plurality of parallel convolutional layers is different.

9. A Ф-OTDR signal denoising system based on an improved VMD algorithm, characterized in that: include: A memory for storing instructions; A processor, configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the denoising method according to any one of claims 1 to 8 is implemented.

Images