Magnetotelluric signal-noise separation method based on artificial neural network

Abstract

The invention discloses a magnetotelluric signal-noise separation method based on an artificial neural network. The magnetotelluric signal-noise separation method comprises the following steps: constructing massive sample libraries conforming to magnetotelluric weak signals and strong interference characteristics; inputting the sample library into a one-dimensional convolutional neural network for training to obtain a corresponding training set and a test set, and defining the network and related training parameters to obtain a training model; performing signal-noise identification on the simulated and actually measured magnetotelluric data by using the obtained mathematical model; noise suppression is only carried out on signals identified as strong interference by using a wavelet soft threshold method; and combining the electromagnetic signal segment after strong interference suppression with the weak electromagnetic signal segment to obtain a reconstructed magnetotelluric signal. According to the method, data segments of strong interference and weak electromagnetic signals can be identified in a more adaptive and high-precision mode, more useful magnetotelluric signals are reserved, and geoelectricity information of a measuring point is better reflected.

Description

Technical Field

[0001] The present invention belongs to the technical field of magnetotelluric signal processing, and particularly relates to a magnetotelluric signal-to-noise separation method based on an artificial neural network. Background Art

[0002] Numerous geophysicists have proven through a large number of field investigations and subsequent research that China has great potential for deep resource development. The magnetotelluric sounding method occupies an important position in geophysical exploration work due to its large exploration depth, low exploration cost, convenient construction, and mature data processing and interpretation technologies.

[0003] The acquisition, processing, and interpretation of measured data are three inseparable and important stages of geophysical exploration work. The area is densely populated, and various human electromagnetic interference sources such as transmission towers, highways, and high-speed railways crisscross, resulting in the collected data containing various complex noise interferences. Therefore, how to correctly and reasonably separate the signal from noise of the measured data collected in the field, and extract reliable and useful magnetotelluric information from strong interferences has become a prerequisite for the magnetotelluric sounding method to carry out work in strong electromagnetic interference areas. Summary of the Invention

[0004] The purpose of the present invention is to design a magnetotelluric signal-to-noise separation method based on an artificial neural network to adaptively and accurately classify weak magnetotelluric signals and strong interference sequences, retain more required magnetotelluric signals, perform targeted noise suppression processing on the classified abnormal sequences, greatly improve the noise suppression ability of the magnetotelluric method in a strong electromagnetic interference environment, and maximize the retention of the slow-changing information in the low-frequency band of magnetotellurics.

[0005] The present invention provides a magnetotelluric signal-to-noise separation method based on an artificial neural network, including the following steps:

[0006] Step 1: Use the amplitude of the magnetotelluric signal time series as a characteristic parameter, and construct a sample library containing a large number of samples with typical weak magnetotelluric signals and strong interference characteristics;

[0007] Step 2: Input the data of the sample library obtained in Step 1 into a one-dimensional convolutional neural network for training to obtain a signal-to-noise classification mathematical model;

[0008] Step 3: Use the signal-to-noise classification mathematical model trained in Step 2 to classify the magnetotelluric data. Classified as "1" indicates a weak magnetotelluric signal sequence, and classified as "2" indicates a magnetotelluric signal sequence affected by strong interference;

[0009] Step 4: Use the wavelet soft threshold method to suppress the noise of the magnetotelluric signal sequence classified as "2" in Step 3, and obtain a denoised signal sequence;

[0010] Step 5: Merge and reconstruct the denoised signal sequence after noise suppression with the magnetotelluric signal sequence classified as "1" in Step 3.

[0011] The present invention uses the amplitude of magnetotelluric signals as characteristic parameters, constructs a large number of sample library data that conform to weak magnetotelluric signals and strong interference based on this, inputs these sample library data into a one-dimensional convolutional neural network for training to obtain a trained signal-to-noise classification mathematical model, then adaptively and highly accurately classifies simulated and measured magnetotelluric signal sequences, only performs noise suppression processing on the magnetotelluric signal sequences classified as "2" using the wavelet soft threshold method, and finally merges and reconstructs the signal sequences classified as "1" and the signal sequences classified as "2" and processed by denoising. A series of processes can effectively process the magnetotelluric data affected by interference and can retain more required signals.

[0012] Further preferably, the purpose of constructing the sample library in Step 1 is to better represent the time-domain waveform characteristics of magnetotelluric measured data and subsequent model training. Therefore, signals of square waves, triangular waves, and pulse interferences, as well as signals without the above waveforms, are constructed as the sample library. The length of each segment of the simulated signal is 50, there are 1600 each of square waves, triangular waves, and pulses, and the amplitude is between 10 -5 and 10 5 ; there are 3200 signals without the above waveforms, and the amplitude is between -2000 and 2000.

[0013] Further preferably, in Step 2, the one-dimensional convolutional neural network uses a convolution kernel size of 1×3, a quantity of 100, a feature map size of 1×48, and a pooling layer size of 1×2. The maximum pooling method is selected, and the size of the pooled feature map is 1×24. The number of neurons in the fully connected layer is 100×24, and the output neurons are 2.

[0014] Further preferably, the signal-to-noise classification mathematical model in Step 2 is a one-dimensional convolutional neural network.

[0015] Essentially, a convolutional neural network is a mapping from input to output. It can learn a large number of mapping relationships between input and output without any precise mathematical expressions between input and output. As long as the convolutional network is trained with known patterns, the network has the mapping ability between input and output.

[0016] Further preferably, the magnetotelluric interference data similar to the sample library constructed in Step 1 are square wave signals, triangular wave signals, and pulse signals whose amplitudes increase positively or negatively with the number of samples.

[0017] Advantages of the present invention:

[0018] 1. The present invention proposes to use the amplitude of the magnetotelluric signal itself as a characteristic parameter, construct a massive sample library data that conforms to magnetotelluric weak signals and strong interferences, and introduce a custom one-dimensional convolutional neural network into the field of magnetotelluric data processing, providing more refined guidance for subsequent noise suppression.

[0019] 2. The present invention provides a method for separating magnetotelluric signal from noise based on an artificial neural network. Compared with existing traditional methods, it can not only avoid losing too much required magnetotelluric information, but also be more adaptive and accurate compared with small-sample learning algorithms.

[0020] 3. The present invention provides to input the sample library data into a one-dimensional convolutional neural network for training to obtain a trained signal-to-noise classification model. After saving the model, the signal-to-noise classification of magnetotelluric simulated and measured signal sequences can be carried out, and only the signal sequences classified as "2" are subjected to noise suppression using the wavelet soft threshold method. The method proposed by the present invention can not only retain more useful information of magnetotelluric data, but also denoise more thoroughly, and the final result is closer to the ideal effect.

[0021] 4. The present invention directly uses the amplitude of the magnetotelluric signal itself as a characteristic parameter, avoiding introducing too many other characteristic parameters that affect the accuracy of subsequent signal-to-noise classification. At the same time, using a one-dimensional convolutional neural network for training can avoid small-sample learning algorithms such as "support vector machines" from getting into a chaotic state when facing a large amount of data during training.

[0022] 5. A convolutional neural network is a type of feedforward neural network that contains convolutional calculations and has a deep structure, and is one of the representative algorithms of deep learning. A convolutional neural network is essentially a mapping from input to output. It can learn a large number of mapping relationships between input and output without any precise mathematical expression between input and output. As long as the convolutional network is trained with known patterns, the network will have the mapping ability between input and output. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flowchart of the present invention;

[0024] Figure 2 is a structural diagram of a one-dimensional convolutional neural network;

[0025] Figure 3 is a set of time-domain waveform diagrams randomly selected from the constructed sample library data, where Figure (a), Figure (b), Figure (c), and Figure (d) are weak magnetotelluric signals, square wave signals, triangular wave signals, and pulse signals respectively;

[0026] Figure 4 is a model training diagram;

[0027] Figure 5It is the effect diagram of separating the signal from noise of the analog signal, where Fig. (a), Fig. (b) and Fig. (c) are respectively the effect diagrams of separating the signal from noise of the analog large-scale square wave interference, triangular wave interference and pulse interference;

[0028] Figure 6 It is the effect diagram of separating the signal from noise of the measured magnetotelluric signal, where Fig. (a), Fig. (b) and Fig. (c) are respectively the effect diagrams of separating the signal from noise of the measured large-scale square wave interference, triangular wave interference and pulse interference;

[0029] Figure 7 It is the apparent resistivity-phase curve diagram of the original data of the measured point EL22193A, the remote reference method and the method of the present invention. Among them, curves "1", "2" and "3" respectively represent the apparent resistivity-phase curves of the original data, the remote reference method processing and the method of the present invention; Figure 7 Among the 4 figures are R xy the apparent resistivity curve in the R xy direction, P xy the phase curve in the P yx direction, Detailed implementation manner

[0030] The present invention will be further described below in conjunction with embodiments.

[0031] As Figure 1 shown, the present invention discloses a magnetotelluric signal-to-noise separation method based on an artificial neural network, including the following steps:

[0032] Step1: Use the amplitude of the magnetotelluric signal time series as a characteristic parameter, and construct a sample library containing a large number of samples with typical weak magnetotelluric signals and strong interference characteristics.

[0033] Specifically, signals of square wave, triangular wave and pulse interference, as well as signals without the above waveforms are constructed as the sample library. The length of each segment of the analog signal is 50, with 1600 each of square wave, triangular wave and pulse, and the amplitude is between 10 -5 and 10 5 ; 3200 signals without the above waveforms, and the amplitude is between -2000 and 2000.

[0034] Step2: Input the sample library data into a one-dimensional convolutional neural network for training.

[0035] Specifically, the one-dimensional convolutional neural network uses a convolutional kernel size of 1×3, with a quantity of 100, a feature map size of 1×48, and a pooling layer size of 1×2. The maximum pooling method is selected, and the size of the feature map after pooling is 1×24. The number of neurons in the fully connected layer is 100×24, and the number of output neurons is 2.

[0036] Step3: After the training in Step2, a signal-to-noise classification mathematical model is obtained, and the model is saved and applied to subsequent magnetotelluric simulations and measured signal sequences.

[0037] Step4: Use the trained model to perform signal-to-noise classification on the magnetotelluric signal sequence. Classifying as "1" indicates a weak magnetotelluric signal sequence, and classifying as "2" indicates a strong interference sequence.

[0038] Step5: Only use the wavelet soft threshold method to suppress noise for the strong interference sequence classified as "2" and obtain the denoised signal.

[0039] Step6: Merge and reconstruct the signal sequence classified as "1" and the signal sequence classified as "2" and subjected to noise suppression.

[0040] The technical problem solved by the present invention is to propose a magnetotelluric signal-to-noise separation method based on an artificial neural network for problems such as excessive loss of required information during the processing of existing traditional methods and the easy confusion of small-sample intelligent classification algorithms in the face of a large amount of data. The present invention uses the amplitude of the magnetotelluric signal itself as a characteristic parameter, without the need to introduce too many other characteristic parameters, and artificially constructs a massive magnetotelluric sample library data; at the same time, a one-dimensional convolutional neural network customized and suitable for one-dimensional magnetotelluric signals is introduced, and the magnetotelluric sample library data is input into the one-dimensional convolutional neural network for training to obtain a signal-to-noise classification mathematical model; finally, only the wavelet soft threshold method is used to suppress noise for the magnetotelluric signal sequence classified as "2". The present invention introduces the one-dimensional convolutional neural network belonging to large-sample learning into the processing of magnetotelluric data, which will achieve adaptive and high-precision signal-to-noise classification, providing a guarantee for improving the quality of magnetotelluric data and further analysis and research.

[0041] The effect of the present invention is evaluated by calculating the apparent resistivity-phase curve of the reconstructed magnetotelluric data. Attached Figure 7 Shown is a comparison diagram of the apparent resistivity-phase curve of the original data of the magnetotelluric measurement point, the apparent resistivity-phase curve after processing by the comparative original reference method, and the apparent resistivity-phase curve processed by the method of the present invention. It can be seen from the comparison that after being processed by the method of the present invention, the apparent resistivity curve is smoother and more continuous, the near-source effect of the 45° asymptote rise of the apparent resistivity curve is significantly improved, and more low-frequency slow-varying information in the time domain sequence is retained, and the result more truly reflects the inherent geoelectric information of the measurement point itself.

[0042] Analysis of the above specific implementation shows that the adaptability and high precision in the present invention are reflected in the introduction of a one-dimensional convolutional neural network for large-sample learning, which is suitable for processing one-dimensional magnetotelluric signals; the wavelet soft threshold method is used for noise suppression to avoid the defects of the hard threshold method and highlight the thoroughness of the method of the present invention in the noise suppression process.

[0043] The above specific implementation manners have further elaborated in detail on the technical field, background, purpose, solution and beneficial effects of the present invention. It should be understood that this implementation manner is only the preferred manner of the present invention and is not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for separating magnetotelluric signals from noise based on artificial neural network, characterized in that: The steps include: Step 1: Take the amplitude of the magnetotelluric signal time series as the characteristic parameter and construct a sample library containing a large number of samples with typical weak magnetotelluric signals and strong interference characteristics; Step 2: Input the sample library data obtained in step 1 into a one-dimensional convolutional neural network for training to obtain a signal-noise classification mathematical model; Step 3: Use the signal-to-noise classification mathematical model trained in step 2 to classify the magnetotelluric data. Classification "1" indicates a weak magnetotelluric signal sequence, and classification "2" indicates a magnetotelluric signal sequence with strong interference. Step 4: Use the wavelet soft threshold method to suppress the noise of the magnetotelluric signal sequence classified as "2" in step 3, and obtain the denoised signal sequence; Step 5: The denoised signal sequence after noise suppression is combined and reconstructed with the magnetotelluric signal sequence classified as "1" in step 3.

2. The method for separating magnetotelluric signals from noise based on artificial neural network according to claim 1, characterized in that: In step 1, in order to better characterize the time domain waveform characteristics of the magnetotelluric measured data and the subsequent model training, square wave, triangle wave and pulse interference signals, as well as signals without the above waveforms, were constructed as sample libraries; the length of each simulated signal was 50, with 1600 square waves, triangle waves and pulses, and the amplitude was between 10 -5 to 10 5 There are 3200 signals not containing the above waveforms, with amplitudes between -2000 and 2000.

3. The method for separating magnetotelluric signals from noise based on artificial neural network according to claim 1, characterized in that: In step 2, the one-dimensional convolutional neural network uses a convolution kernel size of 1×3, the number is 100, the feature map size is 1×48, and the pooling layer size is 1×2; the maximum pooling method is selected, and the size of the feature map after pooling is 1×24, the number of neurons in the fully connected layer is 100×24, and the number of output neurons is 2.

4. The method for separating magnetotelluric signals from noise based on artificial neural network according to claim 1, characterized in that: The geomagnetic interference-like data in the sample library constructed in step 1 are square wave signals, triangular wave signals and pulse signals whose amplitudes increase positively or negatively with the number of samples.

Images