Method for Suppressing Airborne Transient Electromagnetic In-Band Vibration Noise

Abstract

Disclosed in the present invention is a method for suppressing airborne transient electromagnetic in-band vibration noise, comprising: dividing the data after current turn-off into two segments according to whether the useful signal is attenuated to the system noise level: the segment A is the useful signal segment, and the segment B is the pure noise segment; limiting the bandwidth of the data of the segment B according to the frequency range of the in-band noise, and labeling the result as BL; training a neural network using the BL, utilizing the well trained neural network to predict the in-band vibration noise contained in the data of the segment A, and labeling the prediction result as PNA; and subtracting the PNA from the data of the segment A to suppress the in-band vibration noise contained in the data of the segment A.

Claims

1. A method for suppressing airborne transient electromagnetic high-frequency vibration in-band noise, characterized in that the method comprises the following steps: S1, the airborne transient electromagnetic signal after current turn-off containing in-band vibration noise, dividing the signal into two segments according to whether the useful signal is attenuated to the system noise level: the segment A is the useful signal segment, and the segment B is the pure noise segment; S2, processing data of the segment B, limiting a bandwidth of the data of the segment B to be just greater than a bandwidth of the in-band vibration noise, and labeling the result as B.sub.L; S3, training a wavelet neural network using the data B.sub.L, utilizing the well trained wavelet neural network to predict the in-band vibration noise contained in the data of the segment A, and labeling a prediction result as PN.sub.A; and S4, subtracting the PNA from the data of the segment A to suppress the in-band noise contained in the data of the segment A.

2. The method of claim 1, characterized in that processing the data of the segment B in S2 comprises low pass filtering.

3. The method of claim 1, characterized in that processing the data of the segment B in S2 comprises empirical mode decomposition.

4. The method of claim 1, characterized in that training the wavelet neural network using the data BL and then utilizing the well trained wavelet neural network to predict the in-band vibration noise contained in the data of the segment A in S3 comprises: S3.1, arranging the data B.sub.L in reverse that is labeled as B.sub.LR; S3.2, utilizing the B.sub.LR to train the wavelet neural network; S3.3, utilizing the well trained wavelet neural network to predict the in-band vibration noise contained in the data of the segment A, and labeling a prediction result as PN.sub.AR; and S3.4, reversing the data sequence PN.sub.AR to obtain the final prediction result PN.sub.A.

5. The method of claim 4, characterized in that utilizing the B.sub.LR to train the wavelet neural network in S3.2 comprises: constructing two sets of data Input and Output respectively as an input and an output of the wavelet neural network, $\mathrm{Input}=\left[\begin{array}{cccc}{B}_{L.Math.R}\left(1\right)& B_{L.Math.R}\left(2\right)& \mathrm{.Math.}& B_{L.Math.R}\left(n\right)\\ B_{L.Math.R}\left(2\right)& B_{L.Math.R}\left(3\right)& \mathrm{.Math.}& B_{L.Math.R}\left(n+1\right)\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ B_{L.Math.R}(N-n)& B_{L.Math.R}\left(N-n+1\right)& \mathrm{.Math.}& B_{L.Math.R}\left(N-1\right)\end{array}\right]$ $\mathrm{Output}=\left[\begin{array}{c}{B}_{L.Math.R}\left(n+1\right)\\ B_{L.Math.R}(n+2)\\ \mathrm{.Math.}\\ B_{L.Math.R}\left(N\right)\end{array}\right]$ and constructing the wavelet neural network, and making prediction accuracy reach convergence requirements through iterative training, wherein n is the number of nodes in an input layer of the wavelet neural network.

6. The method of claim 5, characterized in that the value of n is determined based on sequence characteristics of the high-frequency vibration in-band noise.

7. The method of claim 4, characterized in that in S3.2, when the number of nodes in the input layer of the wavelet neural network is n, a hidden layer contains 2 n-3 n nodes and the number of nodes in an output layer is 1.

8. The method of claim 7, characterized in that utilizing the trained wavelet neural network to predict the high-frequency vibration in-band noise contained in the data of the segment A in S3.3 comprises: utilizing [B.sub.LR(N-n+1) [B.sub.LR(N-n+2) . . . B.sub.LR(N)] as input data to predict a value of the high-frequency vibration in-band noise contained in a last date point in the date sequence of the segment A, and advancing the prediction point by point until completing the prediction of values of the high-frequency vibration in-band noise of all data points in the data sequence of the segment A, in order to obtain the sequence PN.sub.AR.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The drawings are used to provide a further understanding of the technical solution of the present invention and constitute a part of the specification. Also, the drawings serve to explain the technical solution of the present invention together with the embodiments of the present invention, and do not constitute a limitation on the technical solution of the present invention.

[0028] FIG. 1 is a schematic diagram of a traditional central loop transient electromagnetic detection device;

[0029] FIG. 2 is a schematic diagram of an airborne transient electromagnetic system based on the fixed-wing aircraft, in which 1 represents the transmitter loop, 2 represents the bird, and the arrow represents the flight direction of the aircraft;

[0030] FIG. 3 is a schematic diagram of an airborne transient electromagnetic system based on the helicopter, in which 3 represents the sensor, 4 represents the bearing cable, and the arrow represents the flight direction of the aircraft;

[0031] FIG. 4 is an effect diagram of the traditional motion noise processing, in which 5 represents the ATEM raw observation data, and 6 represents the data processed by the traditional motion noise suppression method;

[0032] FIG. 5 is a spectrum analysis diagram of the data, in which 7 represents a signal containing no in-band vibration noise, and 8 represents a signal containing the in-band vibration noise;

[0033] FIG. 6 is a diagram of the data segment, in which 9 represents the ATEM useful signal, 10 represents the segment point, and 11 represents the IBV noise;

[0034] FIG. 7 is a topology diagram of the wavelet neural network;

[0035] FIG. 8 is prediction results (PN.sub.AR) of the IBV of the data in the segment A, in which 12 represents the original signal, and 13 represents the predicted IBV signal; and

[0036] FIG. 9 is a removal result of IBV noise, in which 14 represents a signal containing no in-band noise in adjacent periods, and 15 represents a signal after the in-band noise processed based on the wavelet neural network.

DETAILED DESCRIPTION OF THE INVENTION

[0037] The following description of the embodiments of the present invention will clearly and completely describe the purpose, the technical schemes and the advantages of the present invention with reference to the drawings. It should be noted that the embodiments of the present invention and the features in the embodiments can be in any combination with each other without conflict.

[0038] The technical schemes of the present invention will be described in detail below by particular embodiments.

[0039] The method for suppressing the airborne transient electromagnetic in-band vibration noise in the embodiments can include the following steps.

[0040] S1, the ATEM data after current turn-off containing the IBV noise can be divided into two segments according to whether the useful signal is attenuated to the system noise level: the segment A is the useful signal segment, and the segment B is the pure noise segment.

[0041] The general form of the ATEM useful signal is in an exponential decay form. As shown in FIG. 6, the ATEM data containing the IBV noise can be divided into two segments, and the segment point is based on the system noise level. That is, the data can be divided into two segments: segment A and segment B. The useful signal contained in the segment A has not yet been fully attenuated (the signal amplitude is higher than the system noise level). While in the segment B, the ATEM useful signal has been fully attenuated (i.e., the ATEM useful signal cannot be contained in the segment B), and the data of the segment B can mainly consist of the IBV noise and other wide-band stationary random noise.

[0042] S2, the data of the segment B can be processed to limit the bandwidth of the data of the segment B to be just greater than the bandwidth of the IBV noise. Specifically, the low-pass filtering, the empirical mode decomposition, and other methods can be applied to limit the bandwidth of the data of the segment B to be slightly greater than the bandwidth of the IBV noise, thereby avoiding the influence of the wide-band random noise on the subsequent processing. After the processing, the bandwidth-limited data of the segment B can be obtained and labeled as BL.

[0043] S3, the wavelet neural network can be trained using the data BL and utilized to predict the IBV noise contained in the data of the segment A, and the prediction result can be labeled as PNA.

[0044] The purpose of this step is to train a wavelet neural network (WNN) utilizing the data B.sub.L, in order to achieve the prediction to the IBV noise contained in the data of the segment A. The so-called prediction usually uses the early data to predict the late data in the time sequence. But in practice, the data of the segment A is located in the early stage of the data of the segment B. Therefore, in order to realize the prediction of IBV in the data of the segment A using the data of the segment B, the data of the segment B (i.e., data B.sub.L) may need to be reversely arranged and to be used to train the WNN to realize the prediction of the IBV noise contained in the data of the segment A. In fact, the result of this prediction actually corresponds to the reverse sequence of the IBV noise contained in the data of the segment A (labeled as PN.sub.AR). By reversing the prediction result sequence PN.sub.AR, the prediction sequence in correct time direction (PN.sub.A) of the IBV noise contained in the data of the segment A can be obtained.

[0045] The specific method of training the WNN using the data segment B.sub.LR is as follows. Assuming that the data length of B.sub.LR is N, firstly two groups of data are constructed as the input and output of the WNN:

[00002]$\mathrm{Input}=\left[\begin{array}{cccc}{B}_{L.Math.R}\left(1\right)& B_{L.Math.R}\left(2\right)& \mathrm{.Math.}& B_{L.Math.R}\left(n\right)\\ B_{L.Math.R}\left(2\right)& B_{L.Math.R}\left(3\right)& \mathrm{.Math.}& B_{L.Math.R}\left(n+1\right)\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ B_{L.Math.R}(N-n)& B_{L.Math.R}\left(N-n+1\right)& \mathrm{.Math.}& B_{L.Math.R}\left(N-1\right)\end{array}\right]$$\mathrm{Output}=\left[\begin{array}{c}{B}_{L.Math.R}\left(n+1\right)\\ B_{L.Math.R}(n+2)\\ \mathrm{.Math.}\\ B_{L.Math.R}\left(N\right)\end{array}\right]$

[0046] That is to say, the (n+1)th data can be predicted by using any continuous n data in the B.sub.LR sequence, in other words, the (n+1)th data is related to the previous n data. The value of n is related to the sequence characteristics of the specific IBV to be suppressed. After determining the value of n, a WNN is constructed with n nodes in the input layer, 2 n-3 n nodes in the hidden layer and 1 node in the output layer. Through iterative training, the prediction accuracy can meet the convergence requirement.

[0047] After the WNN training is completed, the IBV value contained in the last data point of the data sequence of the segment A (i.e., the first data in the reverse sequence A.sub.R of the data of the segment A) is predicted using [B.sub.LR(N-n+1) . . . B.sub.LR(N)] as the input data. Then the prediction is advanced point by point until completing the prediction of the IBV values of all data points of the segment A to obtain the PN.sub.AR sequence. The PN.sub.AR sequence can be reversed to obtain the prediction result of the high-frequency motion noise of the IBV in the data of the segment A.

[0048] Fourth, the PN.sub.A can be subtracted from the data of the segment A to suppress the in-band noise contained in the data of the segment A.

[0049] The measured data of the 2.5th period in FIG. 4(b) can be taken as an example. First, the data can be divided into two segments. In the previous segment, the transient electromagnetic response is attenuated to approximately the same amplitude as the IBV noise signal. In the latter segment, the transient electromagnetic response is completely attenuated, and the IBV noise signal dominates the data.

[0050] According to the characteristics of the motion noise, the wavelet neural network is selected for suppressing the IBV noise. The wavelet neural network (WNN) is developed from the back propagation (BP) neural network. The wavelet basis function is used as the transfer function of the nodes in the hidden layer of the BP neural network. The basic topology of the WNN is shown in FIG. 7.

[0051] The topology of the WNN is mainly divided into the input layer, the hidden layer and the output layer. Signals X.sub.1, X.sub.2 to X.sub.n contained in the input layer are input signals of the WNN, and signals Y.sub.1 to Y.sub.m in the output layer are the predicted input signals of the WNN. The hidden layer is between the input layer and the output layer. Compared with the traditional artificial neural network, the WNN uses the wavelet basis function to replace the traditional Sigmoid function in the hidden layer. In this example, the wavelet basis function is as follows:

(x)=cos (1.75x)e.sup.x.sup.2.sup./2

[0052] The arrows between each element of the input layer and each element of the hidden layer represent input connection weights, and arrows between each element of the hidden layer and each element of the output layer represent output connection weights. It can be seen from the above arrows that: (1) there is no interconnection between elements of the same layer; and (2) all layers are interconnected, that is, any element of any layer and all elements of other layers are interconnected. It can be seen from the topology that the WNN is actually equivalent to using wavelet basis function as the core, and the mapping relationship between the input function and the output function is constructed through the parameter training.

[0053] A WNN with a 5-9-1 structure is designed. That is, the input layer contains 5 nodes, the hidden layer contains 9 nodes, and the output layer contains 1 node. The above structure also means that each data value is related to the previous 5 data values (the previous selection of n).

[0054] After completing the training of WNN, it is used to predict the IBV noise contained in the data of the segment A. The result is shown in FIG. 8.

[0055] The horizontal axis of FIG. 8 is the sampling points in reverse chronological order. The prediction data in FIG. 8 is reversed, and the reversed data of the prediction data is subtracted from the raw data to obtain the result after the IBV noise is processed. The result is compared with the data without the IBV noise in the adjacent period, which is shown in FIG. 9. It can be seen from FIG. 9 that the above method can effectively suppress the IBV noise and lay a good data foundation for subsequent data processing and inversion interpretation.

[0056] While the embodiments of the present invention have been described above, it may be understood that they are only for the understanding of the present invention and they are not intended to limit the invention to these embodiments. Any modifications and variations in the form and details of the embodiments can be made by those skilled in the art within the spirit and scope of the invention. However, the scope of patent protection of the invention may still be defined by the appended claims.