An end-to-end magnetic anomaly signal denoising method based on deep learning

Abstract

The application discloses an end-to-end magnetic anomaly signal denoising method based on deep learning, which comprises the following steps: firstly, wavelet decomposition and reconstruction are performed on the magnetic anomaly signal to realize band-pass filtering and extract the signal in the frequency band of the magnetic anomaly; then, the signal is decomposed by using an orthogonal basis function (OBF), and the decomposition coefficients on each orthogonal basis are solved; then, the decomposition coefficients of each orthogonal basis are taken as the input of a multi-channel deep learning network, and the corresponding ideal magnetic signal is taken as the output, so that an end-to-end denoising network model is trained; finally, the trained denoising network model is used to process the magnetic signal, and the denoised magnetic anomaly signal is obtained. The method comprehensively utilizes the band-pass filtering based on wavelet analysis, OBF decomposition and end-to-end deep learning network, extracts and denoises the features of the collected magnetic signal, and can effectively improve the signal-to-noise ratio of the magnetic anomaly signal.

Description

Technical Field

[0001] The present invention relates to signal processing, deep learning and other technologies, and belongs to the field of detection of magnetic anomaly signals. Background Art

[0002] Magnetic anomaly detection (MAD) is an important method for detecting ferromagnetic targets and is widely used in geological prospecting, shipwreck detection, and aeromagnetic surveys. The main sources of noise in magnetic signals are geomagnetic noise, inherent sensor noise, test platform interference noise, and geological noise. At long detection distances, the signal-to-noise ratio is extremely low, and magnetic anomaly signals are often buried in strong geomagnetic noise, resulting in a decrease in the performance of traditional magnetic anomaly detection methods.

[0003] With the development of deep learning neural networks, they can now independently learn and extract the deep features hidden in data, replacing the feature extraction process that previously required manual intervention. An end-to-end learning approach eliminates manual subdivision of the problem throughout the learning process. Instead, the deep learning model is fully empowered to autonomously learn the mapping from raw input to desired output, leaving the task of feature extraction to the model. Raw data or lightly preprocessed data is directly fed into the model, allowing it to perform feature extraction itself. This approach allows neural networks to effectively learn feature descriptions, allowing the model to automatically learn feature operators that previously required manual design.

[0004] To address the problem of target signals being interfered with and overwhelmed by noise during magnetic anomaly detection, the present invention discloses an end-to-end magnetic anomaly signal denoising method based on deep learning. This method uses an encoding-decoding structure to design a deep learning model. The decomposition coefficients obtained after wavelet bandpass filtering and OBF decomposition of the noisy magnetic anomaly signal are used as the model input, and the corresponding ideal magnetic anomaly signal is used as the output. Through training with a large amount of data, the features of the magnetic anomaly data are automatically extracted. Finally, the magnetic signal to be denoised is input into the network model, and the output of the model is the denoised magnetic anomaly signal. Compared with traditional methods, this method has better adaptability and robustness, and can greatly improve the signal-to-noise ratio of the magnetic anomaly signal. Summary of the Invention

[0005] This invention discloses an end-to-end magnetic anomaly signal denoising method based on deep learning, which can reduce the noise of magnetic anomaly signals. The implementation steps of this method are as follows:

[0006] Step 1: Based on the magnetic dipole model, construct ideal magnetic anomaly signals under different CPA (Closest Point of Approach) distances and different movement speeds. The calculation formula of the ideal magnetic anomaly signal is as follows:

[0007] B=B x e x+B y e y +B z e z

[0008]

[0009]

[0010]

[0011]

[0012] l=cos I cos D

[0013] m=cos I sin D

[0014] n=sin I

[0015] Where B is the magnetic induction intensity of the target magnetic dipole, B x 、B y 、B z is the component of B on the x, y, and z axes, μ is the vacuum magnetic permeability, p is the magnetic moment of the magnetic dipole, r is the straight-line distance from the magnetic dipole to the sensor, and r x 、r y 、r z is the component of r on the x, y, and z axes, I is the geomagnetic inclination of the detection point, and D is the geomagnetic declination of the detection point;

[0016] Step 2: Construct a noisy magnetic anomaly signal. First, collect a pure geomagnetic background signal at the detection point. By moving the time window, different pure geomagnetic background signals are obtained. The ideal magnetic anomaly signal obtained in step 1 is superimposed with different pure geomagnetic background signals to obtain different noisy magnetic anomaly signals. In this way, a magnetic anomaly signal dataset containing geomagnetic background noise is constructed.

[0017] Step 3: Perform wavelet bandpass filtering on the noisy magnetic anomaly signal obtained in step 2. First, perform wavelet decomposition, then set the detail coefficients belonging to high frequencies and the approximate coefficients belonging to very low frequencies after wavelet decomposition to zero, and then reconstruct the signal to obtain the magnetic anomaly signal after bandpass filtering.

[0018] Step 4: Perform orthogonal basis function (OBF) decomposition on the filtered signal to obtain the decomposition coefficients corresponding to the three orthogonal bases of each signal group;

[0019] Step 5: Build an end-to-end deep learning network, use the decomposition coefficients obtained in step 4 to construct a training sample data set, use the decomposition coefficients as the input of the network model, and the corresponding ideal magnetic anomaly signal as the output of the network model, and train an end-to-end denoising network model;

[0020] Step 6: The magnetic signal to be denoised is processed by wavelet bandpass filtering and OBF decomposition as described in steps 3 and 4, and then the decomposition coefficients are used as the input of the trained denoising network model, and the output of the denoising network model is used as the denoised magnetic anomaly signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of an end-to-end magnetic anomaly signal denoising method based on deep learning of the present invention.

[0022] Figure 2 It is the noisy magnetic anomaly signal mentioned in the present invention.

[0023] Figure 3 It is a set of input signals mentioned in the present invention for training an end-to-end deep learning network.

[0024] Figure 4 It is the output signal for training the end-to-end deep learning network mentioned in the present invention.

[0025] Figure 5 This is a schematic diagram of the network structure of the deep learning training mentioned in the present invention. DETAILED DESCRIPTION

[0026] The following describes in detail a technical solution for an end-to-end magnetic anomaly signal noise reduction method based on deep learning, in conjunction with the accompanying drawings and specific implementation methods. The implementation steps of the method are as follows:

[0027] Step 1: Based on the magnetic dipole model, construct ideal magnetic anomaly signals under different CPA (Closest Point of Approach) distances and different movement speeds. The calculation formula of the ideal magnetic anomaly signal is as follows:

[0028] B=B x e x +B y e y +B z e z

[0029]

[0030]

[0031]

[0032]

[0033] l=cos I cos D m=cos I sin D n=sin I

[0034] Where B is the magnetic induction intensity of the target magnetic dipole, B x 、B y 、B z is the component of B on the x, y, and z axes, μ is the vacuum magnetic permeability, p is the magnetic moment of the magnetic dipole, r is the straight-line distance from the magnetic dipole to the sensor, and r x 、r y 、r z is the component of r on the x, y, and z axes, I is the geomagnetic inclination of the detection point, and D is the geomagnetic declination of the detection point.

[0035] Step 2: Construct a noisy magnetic anomaly signal. First, collect a pure geomagnetic background signal at the detection point. By moving the time window, different pure geomagnetic background signals are obtained. The ideal magnetic anomaly signal obtained in step 1 is superimposed with different pure geomagnetic background signals to obtain different noisy magnetic anomaly signals, such as Figure 2 As shown in Figure 2, a magnetic anomaly signal dataset containing geomagnetic background noise is constructed.

[0036] Step 3: Perform wavelet bandpass filtering on the noisy magnetic anomaly signal obtained in step 2. First, perform wavelet decomposition on the signal into 10 layers. Then, set the detail coefficients D1 to D3 after wavelet decomposition to zero, and set the approximate coefficient A10 to zero, which is equivalent to retaining only the middle detail coefficients D7 to D10. Then, reconstruct the signal to obtain the magnetic anomaly signal after bandpass filtering.

[0037] Step 4: Perform orthogonal basis function (OBF) decomposition on the filtered signal to obtain the decomposition coefficients corresponding to three groups of orthogonal bases for each group of signals. The three orthogonal basis functions are:

[0038]

[0039]

[0040]

[0041] Where w = D / R0, D is the distance between the target and the center point of the trajectory when it moves along the trajectory, and R0 is the shortest distance from the target trajectory to the sensor, that is, the CPA distance.

[0042] The calculation formula of OBF decomposition coefficient is:

[0043]

[0044] Where Sig is the noisy magnetic anomaly signal. Figure 2 The three groups of decomposition coefficients of the signal shown after OBF decomposition are as follows Figure 3 shown.

[0045] Step 5: Build an end-to-end deep learning network, build a training sample data set with the decomposition coefficients obtained in step 4, use the decomposition coefficients as the input of the network model, and the corresponding ideal magnetic anomaly signal as the output of the network model, and train to obtain an end-to-end denoising network model. Figure 2 The decomposition coefficients shown are used as model inputs. Figure 4 The ideal magnetic anomaly signal shown is used as the output of the model for training.

[0046] The end-to-end deep learning network is an encoder-decoder network with the following structure: Figure 5 As shown in Figure 1, the model consists of four convolutional layers and four deconvolutional layers connected in series. The four convolutional layers use 128, 256, 512, and 512 4×1 convolution kernels, respectively, while the four deconvolutional layers use 512, 512, 256, and 128 4×1 convolution kernels, respectively. The input signal passes through the convolutional and deconvolutional layers in sequence before reaching the output layer. The loss function in this model is the cross entropy function, which is calculated as follows:

[0047]

[0048] Where x is the true label value, is the predicted probability, i.e. Represents the difference between the predicted probability and the sample label.

[0049] Step 6: The magnetic signal to be denoised is processed by wavelet bandpass filtering and OBF decomposition as described in steps 3 and 4, and then the decomposition coefficients are used as the input of the trained denoising network model, and the output of the denoising network model is used as the denoised magnetic anomaly signal.

[0050] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. An end-to-end magnetic anomaly signal denoising method based on deep learning, characterized in that: The implementation steps are as follows: Step 1: Based on the magnetic dipole model, construct ideal magnetic anomaly signals at different CPA (Closest Proximity Approach) distances and different movement speeds. The calculation formula of the ideal magnetic anomaly signal is as follows: B=B x e x +B y e y +B z e z l=cos I cos D m=cos I sin D n=sin I Where B is the magnetic induction intensity of the target magnetic dipole, B x 、B y 、B z is the component of B on the x, y, and z axes, μ is the vacuum magnetic permeability, p is the magnetic moment of the magnetic dipole, r is the straight-line distance from the magnetic dipole to the sensor, and r x 、r y 、r z is the component of r on the x, y, and z axes, I is the geomagnetic inclination of the detection point, and D is the geomagnetic declination of the detection point; Step 2: Construct a noisy magnetic anomaly signal. First, collect a pure geomagnetic background signal at the detection point. By moving the time window, different pure geomagnetic background signals are obtained. The ideal magnetic anomaly signal obtained in step 1 is superimposed with different pure geomagnetic background signals to obtain different noisy magnetic anomaly signals. In this way, a magnetic anomaly signal dataset containing geomagnetic background noise is constructed. Step 3: Perform wavelet bandpass filtering on the noisy magnetic anomaly signal obtained in step 2. First, perform wavelet decomposition, then set the detail coefficients belonging to high frequencies and the approximate coefficients belonging to very low frequencies after wavelet decomposition to zero, and then reconstruct the signal to obtain the magnetic anomaly signal after bandpass filtering. Step 4: Perform orthogonal basis function (OBF) decomposition on the filtered signal to obtain the decomposition coefficients corresponding to the three orthogonal bases of each signal group; Step 5: Build an end-to-end deep learning network, use the decomposition coefficients obtained in step 4 to construct a training sample data set, use the decomposition coefficients as the input of the network model, and the corresponding ideal magnetic anomaly signal as the output of the network model, and train an end-to-end denoising network model; Step 6: The magnetic signal to be denoised is processed by wavelet bandpass filtering and OBF decomposition as described in steps 3 and 4, and then the decomposition coefficients are used as the input of the trained denoising network model, and the output of the denoising network model is used as the denoised magnetic anomaly signal.

2. The end-to-end magnetic anomaly signal denoising method based on deep learning according to claim 1, characterized in that: There are three orthogonal basis functions in step 4, and the formulas are: Where w = D / R0, D is the distance between the target and the center point of the trajectory when it moves along the trajectory, and R0 is the closest distance from the target trajectory to the sensor, that is, the CPA distance; The calculation formula of OBF decomposition coefficient is: Where Sig is the noisy magnetic anomaly signal.

3. The end-to-end magnetic anomaly signal denoising method based on deep learning according to claim 1, characterized in that: The end-to-end deep learning network constructed in step 5 is an encoding-decoding network, which consists of four convolutional layers and four deconvolutional layers connected in series. The number of convolution kernels in the four convolutional layers is n, 2n, 4n, and 4n, respectively, and the number of convolution kernels in the deconvolutional layer is 4n, 4n, 2n, and n, respectively. The size of the convolution kernel is 4×1. The input signal passes through the convolutional layer and the deconvolution layer in sequence and finally reaches the output layer. The loss function in the model is the cross entropy function, which is calculated as follows: Where x is the true label value, is the predicted probability, i.e. Represents the difference between the predicted probability and the sample label.

Images