Denoising optimization method based on wavelet basis and dynamic adaptive threshold

Abstract

The invention discloses a de-noising optimization method based on a wavelet basis and a dynamic adaptive threshold, and the method comprises the following steps: S1, obtaining a DAS-Morlet wavelet basis function based on a DAS signal in combination with a Morlet wavelet basis; s2, after a dynamic threshold decision model is constructed based on channel statistics, dynamic threshold processing is carried out on the DAS-Morlet wavelet signals; and S3, carrying out multi-scale fusion on the DAS-Morlet wavelet signal after threshold processing by adopting a progressive residual model, and outputting a reconstructed DAS signal. According to the method, the Morlet wavelet basis combined with the DAS features is adopted to form the DAS-Morlet wavelet signal, the DAS-Morlet wavelet signal is subjected to dynamic threshold processing based on the decision model of the dynamic threshold of the channel statistics, and meanwhile, the progressive residual model is introduced, so that the accuracy of the DAS signal after reconstruction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal denoising, and particularly relates to a denoising optimization method based on a wavelet base and a dynamic adaptive threshold. BACKGROUND

[0002] Pipeline transportation, as an efficient long-distance fluid medium transportation method, has gradually established as an indispensable transportation means in modern industrial production of many countries due to its multiple advantages such as economy, convenience, safety and reliability; tailings, as an inevitable product in the separation process of a concentrator, need to be converted into slurry through specific processing and then transported to a tailings pond for stacking and filling by a transportation means; compared with traditional road, railway and waterway transportation modes, pipeline transportation becomes an ideal choice for tailings slurry transportation of a concentrator due to its unique adaptability; therefore, it is necessary and important to monitor the transportation safety of a tailings pipeline.

[0003] Traditional methods such as mean filtering cannot achieve good results, and the deep learning method often has unexplainability, it is difficult to understand the high-dimensional and high-depth model, and these models need a large amount of data to realize; and the model with prior knowledge often has better results, so it has become a popular research direction at present; but if it is used in the field of processing and analysis of DAS signals of tailings pipelines, there are still the following problems: 1, the kernel of wavelet is used as feature extraction, there is no targeted analysis and modification for specific conditions, it is necessary to select a suitable wavelet base function, because some wavelet base functions such as Laplace wavelet base have exponential components, which will cause great difference in wavelet value distribution; 2, the hard threshold is discontinuous, and the setting of the soft threshold often leads to excessive smoothing of the signal, especially in high-frequency signals, details will be lost, such characteristics are difficult to separate noise components when dealing with complex noise and non-stationary signals; 3, the real-time acquisition rate of DAS distributed optical fiber acoustic vibration system reaches 100Mbps, and thousands of acoustic signals can be obtained in a single measurement, forming a high-density data matrix, which will form a huge amount of data, so it needs to be processed specially to speed up the processing ability of the signal processing algorithm.

[0004] In summary, in the current DAS signal processing, although the wavelet analysis as a feature extraction method has certain advantages, there are some problems, first, the selection of wavelet kernel needs to be analyzed and modified according to the specific situation; second, in the threshold processing, the discontinuity of hard threshold and the soft threshold may cause the signal to be excessively smoothed, especially when processing high frequency signals, details are easy to lose, which makes it difficult to effectively separate noise components in complex noise and non-stationary signal processing; in addition, in practical application, DAS faces such a large amount of data, which needs to be processed specifically to improve the denoising ability of the signal processing algorithm; therefore, in order to solve the above problems, a denoising optimization method based on wavelet base and dynamic adaptive threshold is proposed. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies in the prior art described above, and to provide a denoising optimization method based on wavelet base and dynamic adaptive threshold, which adopts Morlet wavelet base combined with DAS characteristics to form DAS-Morlet wavelet signal, and uses a dynamic threshold decision model based on channel statistics to process the DAS-Morlet wavelet signal, and introduces a progressive residual model to improve the accuracy of the reconstructed DAS signal.

[0006] To achieve the above purpose, the technical scheme adopted by the present application is: a denoising optimization method based on wavelet base and dynamic adaptive threshold, comprising the following steps: step S1: based on DAS signal combined with Morlet wavelet base, obtaining DAS-Morlet wavelet base function; step S2: constructing a dynamic threshold decision model based on channel statistics, and performing dynamic threshold processing on the DAS-Morlet wavelet signal; step S3: using a progressive residual model to perform multi-scale fusion on the DAS-Morlet wavelet signal after threshold processing, and outputting the reconstructed DAS signal.

[0007] Preferably, the DAS-Morlet wavelet base function is represented by the following formula:

[0008]

[0009] In the formula: s is the original wavelet scale parameter; ζ is the scale smoothing factor; t is the time variable, representing the sampling time point of the input signal; C is the envelope gain, which is used to control the steepness of the Sigmoid function; f DAS represents the target sound wave main frequency; u represents the translation factor, which controls the position of the wavelet on the time axis.

[0010] Preferably, the channel statistics include μ c and σ c , the μ c is represented by the following formula:

[0011]

[0012] The σ c is represented by the following formula:

[0013]

[0014] In the formula: T represents time; μ c represents the output feature map F(c, t) of the channel c is averaged on the time dimension T; σ c represents the fluctuation degree of the channel c feature.

[0015] Preferably, the dynamic threshold decision model comprises: step S201: based on the statistical quantity μ c and σ c dynamically perceive the feature distribution of each channel; step S202: introduce a dual-channel attention mechanism to generate threshold parameters α c , η c ; step S203: gradient update adjustment is performed on the threshold parameters α c and η c ; step S203: based on the statistical quantity μ c and σ c of step S201 and the α c and η c of step S203, a dynamic threshold formula based on channel statistics is obtained.

[0016] Preferably, the dynamic threshold decision model is represented by the following formula:

[0017]

[0018] In the formula: F(c, t) represents the input signal of channel c at time position t, η c represents the threshold offset; α c represents the soft and hard degree.

[0019] Preferably, the gradual residual model comprises: step S301: input the feature map, and split the feature map into three sub-feature maps [X1, X2, X3] using a 1×1 convolution kernel; step S302: input the sub-feature map X1 into a 1×1 convolution kernel for processing, and output Y1; step S303: input the sub-feature map X2 and the output Y1 of step S302 into a 3×3 convolution kernel, and output Y2; step S304: input the sub-feature map X3, the output Y1 of step S302 and the output Y2 of step S303 into a 5×5 convolution kernel, and output Y3; step S305: input the output Y1 of step S302, the output Y2 of step S303 and the output Y3 of step S304 into a 1×1 convolution kernel for feature fusion, and output the reconstructed DAS signal.

[0020] Preferably, the number of 1*1 convolution kernels is multiple, and the multiple 1*1 convolution kernels respectively perform feature map splitting, convolution processing and feature fusion.

[0021] Compared with the prior art, the application has the following advantages:

[0022] 1. The application adopts the Morlet wavelet base combined with the DAS feature to form a DAS-Morlet wavelet signal, a dynamic threshold decision model based on channel statistics is used to perform dynamic threshold processing on the DAS-Morlet wavelet signal, and a progressive residual model is introduced to improve the accuracy of the reconstructed DAS signal.

[0023] 2. The progressive residual model adopted by the application reduces the convolution calculation amount through grouping convolution, the cooperation of 1*1 convolution kernels, 3*3 convolution kernels and 5*5 convolution kernels, and improves the data processing speed.

[0024] 3. The application encodes the physical characteristics such as the sound wave main frequency f DAS of the DAS signal, the noise frequency band f noise to the wavelet base initialization, and enhances the pertinence of noise separation.

[0025] 4. The application adopts an adaptive threshold mechanism to dynamically generate alpha c and eta c in the double-channel attention, and solves the problems of discontinuity of hard threshold and excessive smoothing of soft threshold.

[0026] The application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a denoising flowchart of the application;

[0028] Figure 2 is a structure diagram of the progressive residual module of the application;

[0029] Figure 3 is a denoising effect comparison diagram of different algorithms;

[0030] Figure 4 is an SNR parameter comparison diagram of Figure 3 . DETAILED DESCRIPTION

[0031] The application discloses a denoising optimization method based on a wavelet base and a dynamic adaptive threshold, comprising the following steps:

[0032] Step S1: obtaining a DAS-Morlet wavelet base function based on the DAS signal combined with the Morlet wavelet base;

[0033] For example Figure 1As shown, the application is applicable to monitoring of pipeline systems such as tailings pipelines, and the DAS system is arranged on the outer wall of the pipeline to facilitate monitoring of the pipeline network system. Taking wavelet domain parameter self-adaptive calibration as the core, firstly, the multi-scale propagation characteristics of the DAS signal are coded to the wavelet base initialization process, so that the center frequency of the Morlet wavelet base matches the effective sound wave frequency of the DAS signal, a DAS-Morlet wavelet base function is formed, and the Morlet wavelet kernel can accurately capture the target frequency band, realize signal decomposition guided by physical priori, separate the noise dominant frequency band and the weak signal bearing frequency band, that is, the DAS signal is modulated by combining the Morlet wavelet base with the DAS characteristics to separate different noise signals; then, the DAS-Morlet wavelet base function is dynamically thresholded based on the dynamic threshold method of channel statistics, and the noise is suppressed; finally, the DAS-Morlet wavelet signal is multi-scale fused through an asymptotic residual model to accelerate signal processing and finally construct a complete denoising signal.

[0034] The Morlet wavelet base is used as the basic wavelet base for improvement, and the Morlet wavelet formula is as follows:

[0035]

[0036] In the formula, C is the envelope gain, and t is the time variable.

[0037] Suppose that the effective sound wave frequency in the DAS signal is f DAS , the DAS noise dominant frequency is f noise , the frequency parameter of the Morlet wavelet is modified so that the center frequency of the Morlet wavelet matches f DAS , the Morlet wavelet kernel can accurately capture the target frequency band, and then:

[0038]

[0039] Adjust the Morlet wavelet scale, then:

[0040] s eff = s - ζ (ζ ∈ [-0.5, 0.5]) (3)

[0041] In the formula, s eff is the adjusted effective scale; s is the original wavelet scale parameter; ζ is a scale smoothing factor for adjusting the wavelet scale; and the effective scale s eff is used to enhance the time-frequency focusing of the Morlet wavelet base, solve the problem of overlapping of noise and effective signal frequency bands, and facilitate suppression of high-frequency noise.

[0042] When f noise is greater than f DASAt this time, increase ζ to compress the wavelet scale corresponding to the high-frequency noise, and through the smooth adjustment of ζ, the time-frequency focusing property of the Morlet wavelet base is adapted to the sound wave and noise band separation characteristics of the DAS signal; and the Sigmoid activation function is applied to the exponential term to facilitate the limitation of the wave value distribution range and avoid gradient explosion, then ψ DAS (t) is transformed into:

[0043]

[0044] ψ DAS (t) is adjusted in scale, and then the DAS-Morlet wavelet base function ψ DAS (t) is expressed as:

[0045]

[0046] In the formula: s∈[1,N k ], N k indicates different scale and translation combinations corresponding to the channel number; u is a translation factor, limiting the position of the wavelet on the time axis, u∈[0,K-1], K represents the total number of time points of the signal under the condition of discrete sampling; C is an envelope gain, controlling the steepness of the Sigmoid function; f DAS is the target sound wave main frequency; t is a time variable, indicating the sampling time point of the input signal.

[0047] Step S2: constructing a dynamic threshold decision model based on channel statistics, and performing dynamic threshold processing on the DAS-Morlet wavelet signal;

[0048] The dynamic threshold decision model comprises:

[0049] Step S201: dynamically sensing the feature distribution of each channel based on statistics μ c and σ c

[0050] The DAS-Morlet wavelet base function is used for channel processing. The DAS-Morlet wavelet base is used as a wavelet base convolution kernel to perform one-dimensional convolution operation on each input channel to generate an output feature map F(c,t)∈R^(c×T), c is the channel number, and T is the time point. Then, the global statistics are calculated along the channel dimension to model the spatial correlation, and then:

[0051]

[0052] In the formula: T represents time; μ c ​denotes the output feature map F(c, t) of the channel c is averaged on the time dimension T to capture the global feature intensity of this channel; F(c, t) denotes the output value at time point t after the input signal is one-dimensionally convolved with the cth DAS-Morlet wavelet kernel.

[0053]

[0054] In the formula: σ c denotes the fluctuation degree of the channel c feature, reflecting the robustness of the feature in this channel, i.e. the noise sensitivity; the statistical quantity μ c and σ c Dynamic perception of feature distribution in different channels.

[0055] Step S202: Introducing a dual-channel attention mechanism to generate threshold parameters α c and η c ;

[0056] Threshold parameters α c and η c are adjusted by gradient update, so that the threshold function adaptively selects the soft and hard threshold ratio and the threshold size based on the statistical characteristics of the input signal, and balances the denoising strength and detail preservation.

[0057] Step S203: Gradient update adjustment is performed on threshold parameters α c and η c ;

[0058] The gradient of α c is represented by the following formula:

[0059]

[0060] Where: y c,t denotes the output signal after dynamic threshold processing, L is the loss function, and denotes the difference between the denoised output y c,t and the real clean signal ,

[0061]

[0062] The gradient of η c is represented by the following formula:

[0063]

[0064] Where:

[0065]

[0066] The change trend of dynamic parameter α c :

[0067]

[0068] The above formula shows:

[0069] When |F(c,t)| approaches η c , α c →0 (approaching hard threshold), reducing high frequency signal attenuation;

[0070] When |F(c,t)| >> η c , α c →1 (approaching soft threshold), suppressing strong noise.

[0071] Step S203: Based on the statistical quantity μ c , σ c of step S201 and the α c , η c of step S203, the dynamic threshold formula based on channel statistics is obtained.

[0072] The dynamic threshold decision model is represented by the following formula:

[0073]

[0074] In the formula, F(c,t) represents the input signal of channel c at time position t, which is the characteristic coefficient after wavelet transform, as the original signal to be denoised, the amplitude of F(c,t) determines whether threshold processing is needed; when the amplitude of F(c,t) is less than η c , the signal is directly set to zero, because this part of the signal is mainly composed of noise, and setting to zero is equivalent to directly removing the noise; when the amplitude of F(c,t) is not less than η c , the signal is processed, because the signal exceeding the threshold is mainly composed of useful signal, but there is still noise mixed in, directly retaining the original value will retain all the noise, and directly setting to zero will lose useful signal, so it is selected to subtract α c η c from |F(c,t)|; the sign function sgn(F(c,t)) is used to retain the sign of the input signal, ensuring that the signal after wavelet basis function denoising is consistent in phase with the original signal; η c represents the threshold offset, which determines whether the signal is truncated; α c represents the soft and hard balance weight control threshold processing, balancing the denoising strength and detail preservation.

[0075] The above formula (13) generates independent threshold parameters for each channel through η c and α c , which facilitates adaptation to the differences between high frequency (more noise) and low frequency (strong signal), α cThe dynamic adjustment of the soft and hard ratio balances the denoising intensity and detail preservation, and finally a dynamic threshold process based on channel statistics converts the fixed rule of traditional threshold denoising into a learnable dynamic process, so that the model can adapt to the signal characteristics of different channels and different time points.

[0076] Step S3: using a progressive residual model to perform multi-scale fusion on the DAS-Morlet wavelet signal after threshold processing, and outputting a reconstructed DAS signal.

[0077] The progressive residual model is used to realize the collaborative optimization of multi-scale feature fusion and calculation acceleration.

[0078] The progressive residual model comprises:

[0079] Step S301: inputting a feature map, and using a 1x1 convolution kernel to split the feature map into three sub-feature maps [X1, X2, X3];

[0080] Step S302: inputting the sub-feature map X1 into a 1x1 convolution kernel for processing, and outputting Y1;

[0081] Step S303: inputting the sub-feature map X2 and Y1 output by step S302 into a 3x3 convolution kernel, and outputting Y2;

[0082] Step S304: inputting the sub-feature map X3, Y1 output by step S302 and Y2 output by step S303 into a 5x5 convolution kernel, and outputting Y3;

[0083] Step S305: inputting Y1 output by step S302, Y2 output by step S303 and Y3 output by step S304 into a 1x1 convolution kernel for feature fusion, and outputting a reconstructed DAS signal.

[0084] The number of the 1x1 convolution kernels is multiple, and the multiple 1x1 convolution kernels respectively perform feature map splitting, convolution processing and feature fusion.

[0085] The number of the 1x1 convolution kernels in the application is three.

[0086] The 1x1 convolution kernel in the above step S301 realizes channel grouping; the 1x1 convolution kernel in step S302 suppresses high-frequency random noise; the 3x3 convolution kernel in step S303 processes medium-frequency structural noise; the 5x5 convolution kernel in step S304 captures low-frequency global noise; the progressive residual model uses a decomposition and fusion residual structure to decompose noise estimation error into different scale subspaces, realizes spatial dimension expansion instead of channel grouping expansion of an ordinary residual module through 1x1, 3x3 and 5x5 multi-scale convolution kernels, and realizes the goal of hierarchical receiving domain expansion.

[0087] As Figure 2As shown, in the progressive residual model, the first group of 1x1 convolution kernels only receives the input feature map and groups the feature map, outputting three sub-feature maps X1, X2 and X3, the feature map X1 is input into the second group of 1x1 convolution kernels, the second group of 1x1 convolution kernels is used to process and suppress high-frequency random noise, and after the second group of 1x1 convolution kernels captures and eliminates the high-frequency random noise in the feature map X1, a high-frequency denoised signal Y1 is output; the feature map X2 and the high-frequency denoised signal Y1 are jointly input into a 3x3 convolution kernel, the 3x3 convolution kernel is used to process and suppress mid-frequency structural noise, and after the 3x3 convolution kernel captures and eliminates the mid-frequency noise in the feature map X2 and Y1, a mid-frequency denoised signal Y2 is output, the feature map X3, the high-frequency denoised signal Y1 and the mid-frequency denoised signal Y2 are jointly input into a 5x5 convolution kernel, the 5x5 convolution kernel is used to process and suppress low-frequency global noise, and after the 5x5 convolution kernel captures and eliminates the low-frequency global noise in the feature map X3, Y1 and Y2, Y3 is output; the first group of 1x1 convolution pre-compression reduces the number of channels of the subsequent 3x3 and 5x5 convolution, so that the calculation amount of the 5x5 convolution is reduced to 37% of the ordinary residual, the high-frequency denoised signal Y1 output by the 1x1 convolution kernel, the mid-frequency denoised signal Y2 output by the 3x3 convolution kernel and the low-frequency global denoised signal Y3 output by the 5x5 convolution are jointly input into a third group of 1x1 convolution kernels, and cross-scale feature interaction is realized by point-by-point addition of the third group of 1x1 convolution kernels, which improves the calculation efficiency and realizes feature fusion.

[0088] Verification test

[0089] SNR and RMSE are a set of complementary indicators, SNR focuses on the degree of noise suppression; RMSE (root mean square error) focuses on the accuracy of signal restoration, the lower the RMSE value, the smaller the difference between the denoised signal and the true signal, that is, the higher the restoration degree, the better the denoising effect; for example Figure 3 、 Figure 4 As shown, the SNR and RMSE of U-net, CCRN, EDCC-EMD and the present application are shown in Table 1.

[0090] Table 1 Comparison of SNR and RMSE results of four algorithms

[0091] Dry record U-net CCRN EDCC-EMD The method of the present application SNR (dB) -7.0031 -1.9768 3.5764 6.5891 15.3415 RMSE (rad) 0.1034 0.0923 0.0658 0.0422 0.0235

[0092] The dry content record in Table 1 represents the original dry content signal, and according to Table 1, the SNR in the method of the application is higher than that of the other three methods, that is, the de-noising effect of the application is better than that of the other three methods, the SNR value is improved from-7.0031dB of the dry content record to 15.3415dB of the method of the application, which shows that the signal after de-noising by the de-noising method of the application is purer; the value of RMSE of the method of the application is significantly lower than that of the other methods, which shows that the restored signal of the application is close to the true value, the model has high accuracy, and the RMSE is reduced from 0.1034 of the dry content record to 0.0235 of the application, which shows that the prediction accuracy after de-noising is significantly improved.

[0093] The above is only a preferred embodiment of the application, and does not limit the application, and any simple modification, change and equivalent structure change of the above embodiment according to the technical essence of the application still belong to the protection scope of the technical solution of the application.

Claims

1. A denoising optimization method based on wavelet basis and dynamic adaptive threshold, characterized in that, Includes the following steps: Step S1: Based on the DAS signal and the Morlet wavelet basis, obtain the DAS-Morlet wavelet basis function; Step S2: Construct a dynamic threshold decision model based on channel statistics and perform dynamic threshold processing on the DAS-Morlet wavelet signal; Step S3: Use an asymptotic residual model to perform multi-scale fusion on the threshold-processed DAS-Morlet wavelet signal and output the reconstructed DAS signal.

2. The denoising optimization method based on wavelet basis and dynamic adaptive threshold according to claim 1, characterized in that, The DAS-Morlet wavelet basis function is expressed by the following equation: In the formula: s is the original wavelet scaling parameter; ζ is the scaling smoothing factor; t is the time variable, representing the sampling time point of the input signal; C is the envelope gain, used to control the steepness of the Sigmoid function; f DAS represents the dominant frequency of the target sound wave; u represents the translation factor, which controls the position of the wavelet on the time axis.

3. The denoising optimization method based on wavelet basis and dynamic adaptive threshold according to claim 1, characterized in that, The channel statistics include μ c and σ c The μ c It can be expressed by the following formula: The σ c It can be expressed by the following formula: In the formula: T represents time; μ c This indicates that the output feature map F(c,t) for channel c is averaged over the time dimension T; σ c This indicates the degree of fluctuation in the characteristics of channel c.

4. A denoising optimization method based on wavelet basis and dynamic adaptive threshold as described in claim 3, characterized in that, The dynamic threshold decision model includes: Step S201: Based on the statistic μ c and σ c Dynamically sense the characteristic distribution of each channel; Step S202: Introduce a dual-channel attention mechanism to generate the threshold parameter α c η c ; Step S203: Adjust the threshold parameter α c and η c Perform gradient update adjustments; Step S203: Based on the statistic μ from step S201 c and σ c and α in step S203 c and η c Thus, a dynamic threshold formula based on channel statistics is obtained.

5. A denoising optimization method based on wavelet basis and dynamic adaptive threshold as described in claim 1, characterized in that, The dynamic threshold decision model is represented by the following equation: In the formula: F(c,t) represents the input signal of channel c at time position t, η c Indicates the threshold offset; α c Indicates the degree of softness or hardness.

6. A denoising optimization method based on wavelet basis and dynamic adaptive threshold as described in claim 1, characterized in that, The asymptotic residual model includes: Step S301: Input the feature map and use a 1×1 convolution kernel to split the feature map into three sub-feature maps [X1, X2, X3]; Step S302: Input the sub-feature map X1 into a 1×1 convolution kernel for processing, and output Y1; Step S303: Input the sub-feature map X2 and the output Y1 from step S302 into a 3×3 convolution kernel to output Y2; Step S304: Input the sub-feature map X3, the output Y1 from step S302, and the output Y2 from step S303 into a 5×5 convolution kernel to output Y3; Step S305: Input Y1 output from step S302, Y2 output from step S303, and Y3 output from step S304 into a 1×1 convolution kernel for feature fusion, and output the reconstructed DAS signal.

7. A denoising optimization method based on wavelet basis and dynamic adaptive threshold as described in claim 6, characterized in that, The number of 1×1 convolutional kernels is multiple, and the multiple 1×1 convolutional kernels are used to perform feature map splitting, convolution processing and feature fusion respectively.

Images