Urban electromagnetic noise suppression method based on multi-source noise

Abstract

The invention discloses an urban electromagnetic noise suppression method based on multi-source noise. The method comprises the following steps: 1, suppressing power frequency noise in noise by using a Wavelet-NGSE algorithm; 2, random noise in the noise is suppressed through minimum noise separation; (3) residual random noise in the noise in the step (2.3) is suppressed by using a CNN-GRU model; according to the invention, the noise in the collected transient electromagnetic signals is classified in a refined manner, and each type of noise features are removed. In order to solve the problem of power frequency harmonic noise interference, an accurate fundamental frequency value is extracted through a Wavelet-NGSE algorithm, and the suppression of power frequency noise with high correlation is realized by combining the periodicity of power frequency and adjusting the transmitting frequency and waveform; further suppressing random noise by using superposition average; and finally, the minimum noise separation + CNN-GRU network calculates data components and reconstructs signals after minimum noise separation, so that high-precision residual noise separation is realized, electromagnetic data with a high signal-to-noise ratio is obtained, and a foundation is laid for realizing accurate interpretation of deep signals.

Description

Technical Field

[0001] The present invention relates to the field of transient electromagnetic signal processing, and specifically to a method for suppressing urban electromagnetic noise based on multi-source noise. Background Art

[0002] In geophysical field fixed-point exploration work, the research on noise elimination of electromagnetic method instruments has made relatively great progress. Internationally, a large amount of research has been done on improving the signal-to-noise ratio of transient electromagnetic methods. However, due to the very serious interference of urban environmental human noise, the obtained electromagnetic signals are contaminated with power frequency harmonic noise, random noise and residual noise. Among them, the power frequency harmonic noise accounts for a relatively large proportion and has a high amplitude. The random noise has a complex composition and a wide frequency band, polluting the late-stage transient electromagnetic signals. The methods in the prior art are no longer applicable to the urban environment. In order to ensure high efficiency, high-frequency emission TEM instruments bring a larger amount of data. However, the urban random noise has a complex composition and a wide frequency band, polluting the late-stage transient electromagnetic signals. Traditional single processing means are difficult to effectively suppress. Therefore, when applying high-frequency emission TEM instruments in the urban environment, there is an urgent need for a method for suppressing urban electromagnetic noise based on multi-source noise. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for suppressing urban electromagnetic noise based on multi-source noise, which can remove each type of noise characteristic respectively and improve the quality of transient electromagnetic signals, so as to solve the technical defects and unachievable technical requirements in the prior art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: A method for suppressing urban electromagnetic noise based on multi-source noise, comprising the following steps: I. Suppress the power frequency noise in the noise by using the Wavelet-NGSE algorithm; 1.1), Collect a group of environmental noises and preprocess the noises by using the wavelet threshold algorithm; 1.2), Estimate the fundamental frequency of the power frequency component in the noise by using the Wavelet-NGSE frequency estimator f ; 1.3), Adjust the period of the transmitted current to an even multiple of the power frequency period as the transient electromagnetic emission mode; 1.4), Collect signals, perform a difference operation on adjacent signals, and remove the power frequency harmonics in the original data; In the present application, on the basis of the traditional transient electromagnetic power frequency suppression method, the volatility factor of the power frequency fundamental frequency is taken into account, so that the method is more in line with the actual situation and improves the power frequency suppression effect during the power frequency processing. At the same time, combining the positive and negative bipolar signal difference averaging ensures the high efficiency of the entire data processing process.

[0005] The NGSE algorithm is an effective time - series frequency estimation algorithm. This method can obtain the fundamental frequency of the main periodic components in the time series. However, when the irrelevant components in the sequence interfere severely, the estimation accuracy of the frequency will be greatly affected. For transient electromagnetic noise data, random noise, as a component unrelated to the power frequency, will seriously affect the estimation accuracy of the power - frequency fundamental frequency. Therefore, the data is processed as described above in this application. Specifically: 1.1.1), Collect a group of environmental noise, and the acquisition duration of this environmental noise is greater than or equal to 60 s; 1.1.2), Then use the wavelet threshold algorithm to regard the power - frequency component in the noise as an effective signal, pre - process the pure noise, and remove the noise components other than the power frequency. The related formula is as follows: Among them, the wavelet basis is selected as 'db8', and the threshold function is selected as the soft threshold. n (t) is the pure environmental noise. n’(t) is the noise after wavelet pre - processing. 1.2.1), After the noise pre - processing, establish the NGSE frequency estimation equation: Among them, k = 1, 2,..., k max , k max is equal to the total number of power - frequency harmonics to be estimated. n is the sampling point in the time domain, while N corresponds to the total number of sampling points. N is proportional to the sampling frequency and the total noise duration. Δt is the time interval between sampling points, which is inversely proportional to the sampling frequency. n’(n) is the pre - processed noise data obtained in step 1.2). ω is the angular frequency. f The initial frequency of A k and B k is set to 50 Hz, which is consistent with the ideal power - frequency fundamental frequency. In the formula, T represents the total duration of the collected noise in the time domain. The variables Xk and Yk are intermediate transition variables. They are sensitive to the offset between the currently estimated angular frequency ω and the true power - frequency fundamental frequency, and are the key intermediate variables in the frequency estimation process. Among them, represents the angular frequency offset, that is, the correction amount between the current estimated frequency value and the next estimated frequency value, which is calculated from the above four estimated intermediate variables. , are all transient variables; 1.2.2) Substitute the preprocessed noise data in step 1.1.2) into the NGSE frequency estimation equation and iterate continuously to make the angular frequency ω offset continuously approach 0, so that the angular frequency ω continuously approaches the true power frequency fundamental frequency of the noise. 1.2.3) Calculate the power frequency fundamental frequency ω according to the estimated angular frequency f。 The NGSE frequency estimation algorithm is an iterative optimization process. The angular frequency ω gets closer and closer to the true power frequency fundamental frequency of the noise with each iteration optimization, while the angular frequency offset gets closer and closer to 0 with the progress of the iterative optimization. When the NGSE frequency estimation algorithm finishes the estimation, both the angular frequency ω and the angular frequency offset tend to be stable. Finally, the power frequency fundamental frequency f is obtained by converting the estimated angular frequency ω .

[0006] Preferably, according to the power frequency fundamental frequency f obtained in step 1.2.3), adjust the period of the transmitted current waveform to 2 / f , that is, the transmission frequency is f / 2 , as the set transient electromagnetic emission mode, where the transmitted current waveform is a bipolar square wave with a duty cycle of 25%.

[0007] After obtaining the actual estimated value of the power frequency fundamental frequency, the next step is to set the transient electromagnetic emission mode. The transmitted current waveform is a bipolar square wave with a duty cycle of 25%, and use the power frequency fundamental frequency f obtained by the Wavelet-NGSE algorithm to adjust the period of the bipolar transmitted square wave to 2 / f (the transmission frequency is f / 2 ).

[0008] It should be explained here that the meaning of adjusting the period of the transmitted current to an even multiple of the power frequency period as the transient electromagnetic emission mode in step 1.3) of this application is not limited to 2 / f (the transmission frequency is f / 2 ), but to adjust the transmission frequency to f / 2, f / 4… etc. 2 / f (the transmission frequency is f / 2 ) is just one of the preferred ways.

[0009] Preferably, 1.4.1), start the survey device according to the set transient electromagnetic emission mode in step 1.3); 1.4.2), collect the transient electromagnetic secondary field signals generated at the moment of current turn-off through the receiving coil. In each emission cycle, a set of positive signals and a set of negative signals can be obtained; 1.4.3), subtract and average the adjacent positive and negative signals collected to complete power frequency suppression.

[0010] Preferably, in step 1.4.2), a set of positive signals and a set of negative signals obtained in each emission cycle include: Signal s + (n) is a positive signal, and signal s - (n) is a negative signal. The two sets of signals have the same amplitude and opposite positive and negative directions. n T+ (n) And n T- (n) both represent the power frequency harmonics coupled in the signal; the specific operation process in step 1.4.3) is: In this application, after the transmitting coil works, the receiving coil can collect the transient electromagnetic secondary field signals generated at the moment of current turn-off. In each emission cycle, we can obtain a set of positive signals x+(n) and a set of negative signals x-(n) . Since the duty cycle of the bipolar square wave is adjusted to 25% and the emission cycle is an even multiple of the power frequency cycle, the power frequency harmonics n T+ (n) and n T- (n) have the same phase, so the amplitudes are also the same. Finally, by performing a subtraction operation on adjacent signals, the power frequency harmonics in the original data can be removed. By this method, we only need to perform one subtraction operation to effectively remove the power frequency harmonics and will not cause TEM signal distortion.

[0011] II. Suppress the random noise in the noise through minimum noise separation 2.1), adjust the data processed in step 1.6) through noise covariance to generate minimum noise separation components; 2.2), form a rotation matrix based on the use of noise covariance estimation and data covariance; 2.3), use the rotation matrix to associate the transformed data with the signal-to-noise ratio; On the basis of filtering power frequency harmonic noise, in order to further suppress random noise, this application explores a data processing method based on minimum noise separation. Minimum Noise Fraction (MNF) is a statistical analysis method based on noise covariance estimation. Its decomposition idea is to decompose the original observed data into multiple components with certain rules and relationships. When they are applied to noise reduction processing, signal-noise separation is achieved through reconstruction with a certain threshold, so as to achieve the purpose of improving the signal-to-noise ratio. The transformation process is to use a rotation matrix to associate the transformed data with the signal-to-noise ratio, and the noise covariance is used to adjust the original data to be processed into the Minimum Noise Fraction (NF) arranged from high to low signal-to-noise ratio. The rotation matrix is formed based on noise covariance estimation and the use of data covariance. Since the low-order minimum noise separation components have significantly higher signal-to-noise ratios, strong filtering procedures can be applied to them without seriously reducing the signal quality of the data. When the transient electromagnetic instrument makes multiple measurements with the same number within the unit equivalent measurement point, multiple approximately identical information components will be reflected in the magnitude of the error value of the signal covariance matrix, where the covariance in linear algebra characterizes the correlation degree of the data vector. Therefore, signal-noise separation can be performed through mathematical transformation.

[0012] Preferably, the specific content in step 2.1) is: adjusting the content processed in step 1.4.3) through noise covariance to be processed into the minimum noise separation components arranged from high to low signal-to-noise ratio.

[0013] Preferably, the specific content of step 2.2) is: 2.2.1), generating multiple secondary field data matrices within the equivalent measurement point: Among them, the noisy data X each channel has been zero-meaned, and the matrix contains n secondary field signals, and each signal contains m time channels, X S represents the theoretical electromagnetic response of the secondary field, X N represents noise, and the signal and noise are uncorrelated with each other; 2.2.2), generating X The covariance matrix of is: Among them C S represents X S The covariance matrix of, C N is X N The covariance matrix of; 2.2.3), the covariance matrix CN It is decomposed into eigenvalues and eigenvectors as follows: where D N represents the diagonal matrix of eigenvalues arranged in descending order, U N represents the eigenvectors corresponding to each eigenvalue.

[0014] Preferably, in step 2.3), the specific content is: The minimum noise separation component is defined as the ratio of the variance of the noise and the total variance of each minimum noise separation component, which is related to the signal-to-noise ratio: where represents the noise variance of the i th component, represents the variance of the noisy signal of the i th component; The noise covariance matrix C N is composed of elements γ kq as follows: where, where represents the average value of the k th time channel, represents the average value of the q th time channel; Construct the matrix: such that P T C N P is the identity matrix; then perform the following calculations: where D X and P represent the eigenvalue matrix and the eigenvector matrix respectively; construct the second rotation matrix R = PV, and perform a linear transformation on the electromagnetic data matrix X to obtain the minimum noise separation component NF, that is, the matrix ψ: Finally, by retaining the first L ( L < m ) minimum noise separation components for data reconstruction, the reconstructed signal is expressed as: is the same as the transient electromagnetic data before processing X , and the matrix size remains unchanged.

[0015] III. Use the CNN-GRU model to suppress the residual random noise in the noise in step 2.3) 3.1), preprocess the transient electromagnetic signal by using the power frequency resonance self-canceling algorithm and the MNF algorithm; 3.2), use a deep learning model to complete residual noise suppression. The traditional MNF algorithm can only suppress random noise to a certain extent, and when the random noise intensity is high, the effect of the MNF algorithm is poor, resulting in some random noise components remaining in the data. Therefore, based on the traditional MNF algorithm for processing data, this application introduces a deep learning algorithm, uses a neural network model to identify the effective signal features, and on the basis of suppressing power frequency interference and part of the random noise by the traditional algorithm, uses a deep learning model to suppress the residual random noise and complete the strong noise suppression in the urban environment.

[0016] Preferably, the specific content in step 3.2) is: the main structure of the network in the deep learning model used is composed of a convolutional neural network (CNN) and a gated recurrent unit (GRU). The convolutional neural network and the gated recurrent unit are used to extract the spatial features and time features of the transient electromagnetic data respectively, and at the same time, a symmetric high-dimensional feature extraction of an autoencoder is adopted.

[0017] 1. This application refines and classifies the noise in the collected transient electromagnetic signal, and removes each type of noise feature separately. A noise reduction strategy based on bipolar square wave emission, wavelet-NGSE, stacked averaging, minimum noise separation, and deep learning is proposed to form a filtering process for suppressing power frequency harmonic noise, random noise, and residual noise. For the problem of power frequency harmonic noise interference, the precise fundamental frequency value is extracted by the wavelet-NGSE algorithm, and then combined with the periodicity of the power frequency, the emission frequency and waveform are adjusted to achieve the suppression of highly correlated power frequency noise; further use stacked averaging to suppress random noise; finally, the minimum noise separation + CNN-GRU network calculates the data components, reconstructs the signal after minimum noise separation, so as to separate the high-precision residual noise and obtain electromagnetic data with a relatively high signal-to-noise ratio, laying a foundation for the accurate interpretation of deep signals. Description of the Drawings

[0018] Figure 1 It is the waveform diagram of the simulation signal in the embodiment of the present invention; Figure 2 It is the waveform diagram of the signal-noise coupled data in the embodiment of the present invention; Figure 3 It is the waveform diagram after power frequency suppression in the assumed bipolar power frequency suppression method in the embodiment of the present invention; Figure 4 It is the waveform diagram after power frequency suppression in the actual simulation fundamental frequency bipolar power frequency suppression method in the embodiment of the present invention; Figure 5 It is the schematic diagram of the minimum noise separation processing result in the embodiment of the present invention; Figure 6 Schematic diagram of the CNN-GRU model structure in the embodiments of the present invention; Figure 7 Schematic diagram of the suppression results of three algorithms on urban transient electromagnetic random noise in the embodiments of the present invention; Detailed implementation manners

[0019] Next, in combination with the accompanying drawings in the embodiments of the present invention Figures 1 - 7 , the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] Please refer to Figures 1 - 7 , embodiments of the present invention: Embodiment

[0021] As Figures 1 - 7 shown: The NGSE algorithm is an effective time series frequency estimation algorithm, and this method can obtain the fundamental frequency of the main periodic components in the time series. However, when the irrelevant components in the sequence interfere severely, the estimation accuracy of the frequency will be greatly affected. For transient electromagnetic noise data, random noise, as a component unrelated to the power frequency, will seriously affect the estimation accuracy of the power frequency fundamental frequency. Therefore, we propose a Wavelet-NGSE algorithm to perform frequency estimation on transient electromagnetic noise data. First, collect a group of environmental noises to ensure that the collection duration is greater than or equal to 60 ms. Then, use the wavelet threshold algorithm to regard the power frequency component in the noise as an effective signal, preprocess the pure noise, and remove the noise components other than the power frequency. The wavelet basis is selected as 'db8', and the threshold function is selected as the soft threshold. Among them, the wavelet basis is selected as 'db8', the threshold function is selected as the soft threshold, n(t) is the pure environmental noise, and n'(t) is the noise after wavelet preprocessing. After the noise preprocessing, establish the NGSE frequency estimation equation set. Among them, k = 1, 2,..., k max , k max is equal to the total number of power frequency harmonics to be estimated, n is the sampling point in the time domain while N corresponds to the total number of sampling points, N is proportional to the sampling frequency and the total noise duration, Δt is the time interval between sampling points and is inversely proportional to the sampling frequency, n’(n) is the noise data after preprocessing obtained in step 1.2),ω is the angular frequency, f the initial frequency of f is set to 50 Hz, which is consistent with the ideal power frequency fundamental frequency. A k and B k are similar to the Fourier coefficients and represent the estimated values of the phase and amplitude of the power frequency harmonics respectively; In the formula, T represents the total duration of the acquisition noise in the time domain, and the variable Xk and Yk are intermediate transition variables. They are sensitive to the offset between the currently estimated angular frequency ω and the true power frequency fundamental frequency, and are key intermediate variables in the frequency estimation process; Among them, represents the angular frequency offset, that is, the correction amount between the current estimated frequency value and the next estimated frequency value, which is calculated from the above four estimated intermediate variables. , are all transition variables; The NGSE frequency estimation algorithm is an iterative optimization process. The angular frequency ω gets closer and closer to the true power frequency fundamental frequency of the noise with each iteration optimization, while the angular frequency offset gets closer and closer to 0 with the progress of the iterative optimization. When the NGSE frequency estimation algorithm ends, the angular frequency ω and the angular frequency offset both tend to be stable. Finally, the power frequency fundamental frequency f is obtained by using the estimated angular frequency ω .

[0022] After obtaining the actual estimated value of the power frequency fundamental frequency, the second step is to set the transient electromagnetic emission mode. The emission current waveform adopts a bipolar square wave with a duty cycle of 25%, and uses the power frequency fundamental frequency f obtained by the Wavelet-NGSE algorithm to adjust the period of the bipolar emission square wave to 2 / f (the emission frequency is f / 2 ). When the transmitting coil works, the receiving coil can collect the transient electromagnetic secondary field signal generated at the moment of current turn-off. In each emission cycle, we can obtain a set of positive signals x +(n) and a set of negative signals x-(n) . The signal s + (n) is a positive signal, and the signal s- (n) is a negative signal. The two sets of signals have the same amplitude and opposite positive and negative directions. n T+ (n) andn T- (n) all represent the power frequency harmonics coupled in the signal; since the duty cycle of the bipolar square wave is adjusted to 25% and the emission period is an even multiple of the power frequency period, the power frequency harmonics n T+ (n) and n T- (n) have the same phase, so the amplitudes are also the same.

[0023] Finally, by performing a difference operation on adjacent signals, the power frequency harmonics in the original data can be removed. The specific formula is as follows: By this method, we only need to perform one subtraction operation to effectively remove the power frequency harmonics without causing distortion of the TEM signal.

[0024] To verify the performance of the power frequency algorithm based on Wavelet-NGSE frequency estimation, we conduct experimental tests.

[0025] We set the parameters according to the above power frequency harmonic formula as follows: the power frequency fundamental frequency is 51 Hz, the total number of harmonics is 3, and the initial power frequency phase is generated by a random number. The experimental signal is a one-dimensional forward modeling signal generated by a homogeneous earth model. The waveform diagrams of the simulation signal and the signal-noise coupled data are as Figures 1 - 2 shown. The coupling method of the signal and the noise is the analog bipolar emission signal method, and the simulated emission frequency is 50.5 Hz (half of the power frequency fundamental frequency, even multiple minus). Among them, the signal-to-noise ratio of the first positive signal is -38.31 dB. Further, it should be explained that Figures 1 - 4 the abscissa time in Figure 5 represents time, with the unit of millisecond (ms), while the ordinate Amplitude represents amplitude, with the unit of volt (V). It can also be referred to the content of the abscissa and ordinate.

[0026] Experiments are conducted using the traditional bipolar power frequency suppression method based on the 50 Hz assumption and the bipolar power frequency suppression method based on the actual simulation fundamental frequency of 51 Hz respectively. First, assuming that the fundamental frequency of the power frequency harmonics is 50 Hz, using the same simulation signal, simulating bipolar positive and negative signals to ensure that the cycle period of the positive and negative signals is 40 ms, which is twice the ideal power frequency period. Then, the difference between adjacent positive and negative signals is taken and divided by two to complete the power frequency suppression. The waveform diagram after power frequency suppression is as Figure 3 shown. It can be seen from the locally enlarged waveform of the signal from 2 ms to 10 ms in the figure that there are still many power frequency harmonics remaining after this power frequency suppression. Through calculation, this power frequency suppression increases the signal-to-noise ratio of the signal from -38.31 dB to -5.65 dB.

[0027] Then, we used the true simulated fundamental frequency to conduct a bipolar power frequency suppression experiment. Taking 51 Hz as the power frequency fundamental frequency, we used the same simulated signal as in the above experiment to simulate bipolar positive and negative signals, ensuring that the cycle of the positive and negative signals was 39.2 ms, which was twice the actual simulated power frequency cycle. Then, we subtracted adjacent positive and negative signals and divided the result by two to complete the power frequency suppression. The waveform diagram after power frequency suppression is as shown in Figure 4 . As can be seen from the locally enlarged waveform of the signal from 2 ms to 10 ms in the figure, almost all power frequency harmonic components were suppressed after this power frequency suppression, and only random noise remained. Through calculation, the signal-to-noise ratio of the signal was increased from -38.31 dB to 1.55 dB after this power frequency suppression. This simulation experiment proved that when the power frequency fundamental frequency fluctuated from the ideal 50 Hz to 51 Hz, accurately obtaining the actual fundamental frequency of the power frequency and performing power frequency suppression based on this had a significant improvement compared to the traditional 50 Hz fundamental frequency assumption.

[0028] On the basis of filtering out power frequency harmonic noise, in order to further suppress random noise, this application explored a data processing method based on minimum noise separation. Minimum Noise Fraction (MNF) is a statistical analysis method based on noise covariance estimation. Its decomposition idea is to decompose the original observed data into multiple components with certain rules and relationships. When applying them to noise reduction processing, through reconstruction with a certain threshold, signal-to-noise separation is achieved to improve the signal-to-noise ratio. The transformation process is to use a rotation matrix to associate the transformed data with the signal-to-noise ratio, and the noise covariance is used to adjust the original data to process it into the Minimum Noise Fraction (NF) components arranged in descending order of signal-to-noise ratio. The rotation matrix is formed based on noise covariance estimation and the use of data covariance. Since the low-order minimum noise separation components have significantly higher signal-to-noise ratios, strong filtering procedures can be applied to them without seriously reducing the signal quality of the data. When the transient electromagnetic instrument makes the same number of multiple measurements within the unit equivalent measurement point, the presence of multiple approximately identical information components will be reflected in the magnitude of the error value of the signal covariance matrix, where the covariance in linear algebra characterizes the correlation degree of the data vectors. Therefore, signal-to-noise separation can be performed through mathematical transformation. The multiple secondary field data matrices within the equivalent measurement point can be expressed as: where the noisy data X has been zero-meaned for each channel, the matrix contains n secondary field signals, each signal contains m time channels, and X S represents the theoretical electromagnetic response of the secondary field, and X N represents the noise, and the signal and noise are uncorrelated with each other.

[0029] The covariance matrix of X is expressed as: where C S representsX S Covariance matrix of C N is X N Covariance matrix of C N is decomposed into eigenvalues and eigenvectors as shown in the following formula: where D N represents the diagonal matrix of eigenvalues arranged in descending order, U N represents the eigenvectors corresponding to each eigenvalue. The minimum noise separation component is defined as the ratio of the variance of the noise and the total variance of each minimum noise separation component, which is related to the signal-to-noise ratio: where, represents the i noise variance of the th i component, C N Covariance matrix C N is composed of elements γ kq as follows: where, where represents the k average value of the th q time channel, such that P T C N P is the identity matrix. Then perform the following calculations: where D X and P represent the eigenvalue matrix and the eigenvector matrix respectively. Construct the second rotation matrix R = PV, and perform a linear transformation on the electromagnetic data matrix X to obtain the minimum noise separation component NF, i.e., the matrix ψ . Finally, by retaining the first L ( L < m ) minimum noise separation components for data reconstruction, the reconstructed signal is expressed as: and the transient electromagnetic data before processingX Same The matrix size remains unchanged.

[0030] The minimum noise separation simulation test is used to verify the effects of the two in the noise reduction process. As Figure 5 shown, the minimum noise separation and stacked average test results are presented. Among them, the best effect of the minimum noise separation is obtained by taking the first 1 minimum noise separation component. The ESNR of the noisy signal before processing is 3.6 dB. After using the minimum noise separation in combination with the stacked average technology, the ESNR are 17.1 dB and 22.6 dB respectively. It can be seen that the signal-to-noise ratio is improved with each step of processing. Compared with the stacked average, the minimum noise separation improves the signal-to-noise ratio more significantly. Before and after processing, it can be seen that the random noise suppression effect is obvious, and within a limited number of stacking times, the transient electromagnetic data quality is significantly improved.

[0031] The traditional MNF algorithm can only suppress random noise to a certain extent. When the random noise intensity is high, the effect of the MNF algorithm is poor, resulting in some random noise components remaining in the data. Therefore, in this chapter, on the basis of processing data with the traditional MNF algorithm, a deep learning algorithm is introduced. The neural network model is used to identify the effective signal features. On the basis of suppressing power frequency interference and part of the random noise by the traditional algorithm, the deep learning model is used to suppress the residual random noise to complete the strong noise suppression in the urban environment.

[0032] Combined with the characteristics of the transient electromagnetic signal and the special network structure proposed above, in order to suppress the residual noise, we propose a new model, the CNN-GRU model, to suppress the residual random noise of the transient electromagnetic. The model structure is as Figure 6 shown, Figure 6Among them, Noisy signal is the noise signal, MNF is Minimum Noise Fraction - Minimum Noise Separation, GRU is Gated Recurrent Unit - Gated Recurrent Unit, Input layer is the input layer, Conv layer is the convolutional layer, reconstruction layer is the reconstruction layer, Processed signal is the processed signal, Theoretical signal is the theoretical signal, LossA and LossB are both loss functions. The model has two loss functions. Among them, LossA is the loss between the reconstructed signal (from the reconstruction layer) and the processed theoretical signal, and LossB is the loss between the noise signal and the processed theoretical signal. Loss is the total loss. The final loss is the weighted sum of LossA and LossB, and the weights are α and (1 - α) respectively. The CNN - GRU model suppresses noise based on traditional denoising algorithms. First, the power - frequency resonance self - cancellation algorithm and the MNF algorithm are used to pre - process the transient electromagnetic signal to initially suppress random noise, and then a deep - learning model is used to complete the suppression of residual noise, greatly improving the quality of the transient electromagnetic signal. The main structure of the network is composed of a combination of a convolutional neural network (CNN) and a recurrent neural unit (GRU). The convolutional neural network and the recurrent neural unit are used to extract the spatial features and temporal features of the transient electromagnetic data respectively, and at the same time, the advantage of symmetric high - dimensional feature extraction of the auto - encoder is adopted.

[0033] To verify the random noise suppression performance of the CNN - GRU model, actual measured urban noise is used for data - processing experiments. After power - frequency pre - processing, it is then coupled with the same transient electromagnetic data, and the signal - to - noise ratios of the data before and after processing are calculated to quantify the performance of the algorithm. Noise suppression comparison experiments are carried out using the SFSDSA algorithm, the LSTM - Autoencoder algorithm, and the CNN - GRU model respectively. The processing results of the three algorithms are as Figure 7 shown Figure 7 in which, amplitude (V): represents the amplitude of the signal, with the unit of volt (V), t (ms): represents time, with the unit of millisecond (ms), signal with noise: represents the signal with noise, signal: represents the signal without noise.

[0034] Through calculation, the signal-to-noise ratio of the transient electromagnetic signal before processing is -14.3 dB. After random noise suppression by the SFSDSA algorithm, the LSTM-Autoencoder algorithm, and the CNN-GRU model, the signal-to-noise ratios are successively increased to 11.1 dB, 15.8 dB, and 23.5 dB. All three algorithms effectively improve the quality of the transient electromagnetic data, and the CNN-GRU model has a larger increase in signal-to-noise ratio. In addition, it can be clearly seen from the waveform diagrams before and after signal processing that the transient electromagnetic signal processed by the CNN-GRU model is the closest to the ideal transient electromagnetic signal, which is consistent with the results of the simulation noise suppression experiment. Therefore, the CNN-GRU model has the best effect in processing urban random noise after power frequency suppression, and the quality of the transient electromagnetic data is improved significantly.

[0035] Based on the urban transient electromagnetic noise suppression strategy proposed in the application, a noise suppression experiment was carried out on a set of measured transient electromagnetic data. According to the above content, three treatments of strong power frequency suppression, random noise suppression, and residual random noise suppression were carried out respectively. By analyzing the transient electromagnetic trace profile data before and after processing, the effectiveness of the noise suppression strategy in this paper was verified.

[0036] The above shows and describes the basic principles, main features, and advantages of the present invention. For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic features of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0037] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for suppressing urban electromagnetic noise based on multi-source noise, characterized in that: The following steps are involved:

1. Use Wavelet-NGSE algorithm to suppress power frequency noise in the noise; 1.1) Collect a set of environmental noises and pre-process the noises using the wavelet threshold algorithm; 1.2) Use Wavelet-NGSE frequency estimator to estimate the fundamental frequency of the power frequency component in the noise f ; 1.3) Adjust the period of the emission current to an even multiple of the power frequency period as the transient electromagnetic emission mode; 1.4) Collect signals, perform difference calculation on adjacent signals, and remove power frequency harmonics in the original data; 2. Suppress random noise in noise through minimum noise separation 2.1) Adjust the data processed in step 1.6) by noise covariance to generate the minimum noise separation component; 2.2) Forming a rotation matrix based on the use of noise covariance estimation and data covariance; 2.3) Use the rotation matrix to associate the transformed data with the signal-to-noise ratio; 3. Use the CNN-GRU model to suppress the residual random noise in the noise in step 2.3) 3.1) Use the power frequency resonance self-cancellation algorithm and MNF algorithm to pre-process the transient electromagnetic signal; 3.2) Use deep learning model to suppress residual noise.

2. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 1, characterized in that: The specific contents of step 1.1) are: 1.1.1) Collect a set of environmental noises, the collection time of which is greater than or equal to 60 seconds; 1.1.2) Then use the wavelet threshold algorithm to treat the power frequency component in the noise as a valid signal, and pre-process the pure noise using the following formula: Among them, the wavelet base is selected as 'db8', and the threshold function is selected as the soft threshold. n(t) is pure ambient noise, n'(t) is the noise after wavelet preprocessing.

3. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 2 is characterized in that the specific contents of step 1.2) are: 1.2.1) Establish the NGSE frequency estimation equation: in, k =1,2,..., k max , k max is equal to the total number of power frequency harmonics to be estimated, n For the sampling points in the time domain N The corresponding number of sampling points is N It is proportional to the sampling frequency and the total duration of the noise. Δt is the time interval between sampling points, which is inversely proportional to the sampling frequency. n'(n) is the preprocessed noise data obtained in step 1.2), ω is the angular frequency, f The initial frequency is set to 50Hz. A k and B k Similar to the Fourier coefficients, they represent the phase and amplitude estimates of the power frequency harmonics; In the formula, T Represents the total duration of the collected noise in the time domain, and the variable Xk and Y are intermediate transition variables, and their effect on the current estimated angular frequency ω The offset from the actual power frequency fundamental frequency is relatively sensitive and is a key intermediate variable in the frequency estimation process; in, Represents the angular frequency offset, that is, the correction between the current estimated frequency value and the next estimated frequency value, which is calculated by the above four estimated intermediate variables. , All are excessive variables; 1.2.2) by substituting the pre-processed noise data in step 1.1.2) into the NGSE frequency estimation equation and iterating continuously so that the angular frequency ω The offset is constantly approaching 0, making the angular frequency ω Continuously approaching the true power frequency fundamental frequency of the noise; 1.2.3) According to the estimated angular frequency ω Obtain the industrial frequency fundamental frequency f。 4. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 3 is characterized in that: The specific content of step 1.3) is: according to the power frequency base frequency obtained in step 1.2.3) f , adjust the period of the emission current waveform to 2 / f , that is, the transmitting frequency is f / 2 , as the setting transient electromagnetic emission mode, in which the emission current waveform adopts a bipolar square wave with a duty cycle of 25%.

5. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 4 is characterized in that: The specific content of step 1.4) is: 1.4.1) Start the survey device according to the transient electromagnetic emission mode set in step 1.3); 1.4.2) The transient electromagnetic field signals generated at the moment of current shutoff are collected by the receiving coil. In each transmission cycle, a group of positive signals and a group of negative signals can be obtained; 1.4.3) Subtract the collected adjacent positive and negative signals and take the average to achieve power frequency suppression.

6. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 5, characterized in that: In step 1.4.2), a set of positive signals and a set of negative signals obtained in each transmission cycle include: Signal s + (n) is a positive signal, signal s - (n) is a negative signal, the two sets of signals have the same amplitude and opposite positive and negative directions. n T+ (n) and n T- (n) Both represent the power frequency harmonics coupled into the signal; the calculation process in step 1.4.3) is specifically as follows: .

7. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 6, characterized in that: The specific content of the step 2.1) is: adjusting the content processed in step 1.4.3) by noise covariance to process it into the minimum noise separation component with the signal-to-noise ratio arranged from high to low.

8. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 7 is characterized in that: The specific contents of step 2.2) are: 2.2.1) Generate multiple secondary field data matrices within the equivalent measurement point: Among them, noisy data X Each channel has been zero-meaned, and the matrix contains n Secondary field signals, each containing m A time road, X S represents the theoretical electromagnetic response of the secondary field, X N Represents noise, signal and noise are unrelated; 2.2.2), generate X The cosquare matrix is: in C S express X S The covariance matrix of C N yes X N The covariance matrix of 2.2.3) The covariance matrix C N The decomposition into eigenvalues ​​and eigenvectors is as follows: in D N is a diagonal matrix representing the eigenvalues ​​in descending order, U N Denote the eigenvector associated with each eigenvalue.

9. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 8, characterized in that: In the step 2.3), the specific content is: The minimum noise separation component is related to the signal-to-noise ratio: in, Indicates i The noise variance of the components, Indicates i The noise covariance matrix C N By element γ kq constitute: Among them, Indicates k The average value of the time channel, Indicates q Average value of the time trace; construct the matrix: Make P T C N P is the identity matrix; then ; Then perform the following calculation: in D X and P Represent the eigenvalue matrix and eigenvector matrix respectively; construct the second rotation matrix R = PV, And the electromagnetic data matrix X Perform linear transformation to obtain the minimum noise separation component NF, that is, the matrix ψ: Finally, by retaining L ( L < m ) minimum noise separation components are used to reconstruct the data, and the reconstructed signal is expressed as: Transient electromagnetic data before processing X same, The matrix size remains unchanged.

10. The method for suppressing urban electromagnetic noise based on multi-source noise according to claim 9, characterized in that: The specific content of the step 3.2) is: the main structure of the network in the deep learning model is composed of a convolutional neural network (CNN) and a GRU (GRU), and the convolutional neural network and the GRU respectively extract the spatial features and temporal features of the transient electromagnetic data, while the symmetric high-dimensional feature extraction of the autoencoder is adopted.

Images