CN-122019999-A - Intelligent denoising method for gravity data in well based on depth residual error network

CN122019999ACN 122019999 ACN122019999 ACN 122019999ACN-122019999-A

Abstract

An intelligent denoising method for well gravity data based on a depth residual error network relates to a geophysical exploration data processing method and solves the problems that noise suppression and signal edge reservation are difficult to balance in a traditional signal processing method, and the existing two-dimensional depth learning model is difficult to adapt to the defects of data characteristics in a one-dimensional well. According to the method, by introducing a DnResUnet-based depth residual error network and utilizing a one-dimensional U-shaped network structure and a large convolution kernel design, long-distance dependence characteristics of a depth domain sequence are automatically captured, and effective signal extraction under complex environmental noise is realized. In the method, a global residual error learning strategy is adopted, noise distribution estimation is taken as a core target, background noise is predicted by inputting noise-containing data, and a gravity signal is recovered through residual error subtraction. The method is suitable for the high-precision processing requirement of the gravity data in the well in the fields of deep mineral exploration, geologic structure imaging and the like.

Inventors

LIU YUJUAN
SONG ZIWEN
LIN TINGTING
WEI MENG
Jin Chuhuai

Assignees

吉林大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (9)

1. The intelligent denoising method for the well gravity data based on the depth residual error network is characterized by comprising the following specific implementation processes: Constructing a gravity synthesis data set in the well, and carrying out normalization processing on noisy gravity data to obtain a normalized data set; constructing a depth residual error U-shaped network model, and training the network model by adopting the data set to obtain a trained optimal network model; in each round of training, the network model samples a plurality of groups of data from the data set for updating the weight parameters of the network model; inputting the normalized noise-containing gravity data into an optimal network model, and outputting estimated background noise distribution Then through a residual subtraction formula Obtaining denoised normalized signals And finally normalizing the signal Multiplying by corresponding scaling factor And finally, intelligent denoising of the gravity data in the well is realized.
2. The intelligent denoising method for the well gravity data based on the depth residual error network of claim 1, wherein the specific process for obtaining the gravity synthesis data set is as follows: calculating a real gravity signal by using a gravity forward formula; Injecting composite noise into the true gravity signal to generate noisy gravity data; Preprocessing noise-containing gravity data by adopting a maximum normalization method, wherein the normalization formula is as follows: ; In the formula, And Noise-containing gravity data and true gravity signal of the kth sample respectively, Is a scaling factor whose value is the absolute amplitude maximum of the kth sample.
3. The method for intelligently denoising in-well gravity data based on depth residual error network of claim 2, wherein the composite noise comprises Gaussian white noise, linear drift and colored noise.
4. The intelligent denoising method for the well gravity data based on the depth residual error network of claim 1, wherein the depth residual error U-shaped network model comprises a one-dimensional convolutional encoder and a decoder; The encoder consists of cascade downsampling modules, each downsampling module comprises a maximum pooling layer and a residual block, a large convolution kernel with the size of 7 is utilized to expand a receptive field, and long-distance dependence characteristics of a depth domain sequence are extracted; The residual block comprises a main path and a shortcut branch, wherein the main path consists of two one-dimensional convolution layers, and a batch normalization layer and a ReLU activation function are arranged in the middle; The decoder consists of cascade up-sampling modules, linear interpolation is adopted to restore the signal length, and shallow layer characteristics of the encoder and deep layer characteristics of the decoder are spliced in channel dimension through jump connection so as to fuse high-frequency information.
5. The method for intelligently denoising in-well gravity data based on depth residual error network of claim 1, wherein the mixed loss function is set to minimize a weighted sum of reconstruction errors and total variation regularization terms, and is characterized in that the mixed loss function is set to be a weighted sum of reconstruction errors and total variation regularization terms Expressed by the following formula: ; In the formula, As a mean square error term, For the total variation regularization term, As a true gravity signal, the gravity signal, Is a regularized weight coefficient.
6. The intelligent denoising method for well gravity data based on depth residual error network of claim 5, wherein regularized weight coefficient is set The weight coefficients are determined based on the balanced requirements of signal fidelity and physical smoothness.
7. The method for intelligent denoising of downhole gravity data based on depth residual error network according to claim 5, wherein the mean square error term Noise-removed gravity signal predicted value used as fidelity loss and output by quantification network model With true gravitational signals Amplitude difference between: ; Wherein, N is the total number of sampling points contained in a single sample sequence, i is the serial number index of the sampling points; the true gravity signal of the ith sampling point; a denoising gravity signal predicted value of an ith sampling point output by the denoising network model; the total variation regularization term Defined as the sum of squares of adjacent data point differences: ; Wherein, in the formula, Representing the denoised gravity signal amplitude at the first data point in the b-th batch, c-th channel; And the amplitude of the gravity signal after denoising at the (1+1) th adjacent data point is represented, and B, C, L is the batch size, the channel number and the signal length respectively.
8. The intelligent denoising method for well gravity data based on depth residual error network of claim 1, wherein in the training process of the network model, an Adam optimizer is adopted to update parameters, and initial learning rate is set as follows If the loss function on the verification set does not drop in 10 continuous training periods, the training is stopped and the optimal network model parameters are saved.
9. The intelligent denoising method for the well gravity data based on the depth residual error network of any one of claims 1 to 8 is characterized in that the trained optimal network model is converted into a light format and deployed to a computing platform, and real-time processing of the well gravity data is achieved.

Description

Intelligent denoising method for gravity data in well based on depth residual error network Technical Field The invention relates to the technical field of geophysical exploration data processing, in particular to an intelligent denoising method for well gravity data based on a depth residual error network. The method improves the signal recovery precision and the geological feature fidelity of the gravity data in the well under the non-stable and strong noise environment by introducing a deep residual error learning technology, and is suitable for deep mineral resource exploration, oil and gas reservoir detection and other geophysical application fields with strict requirements on the signal to noise ratio of the gravity data in the well. Background In the fields of deep mineral exploration and hydrocarbon reservoir detection, in-well gravity measurement (BHGM) is one of the key technologies for acquiring subsurface deep density imaging. Conventional signal processing algorithms, such as wiener filtering (WIENER FILTERING) and wavelet transformation (Wavelet transform), have been widely used in preprocessing of gravity data due to their well-established theory and ease of implementation. These methods attempt to separate the signal from noise by a transform in the frequency domain or in the time-frequency domain, which can improve the signal-to-noise ratio of the data to some extent. Conventional signal processing algorithms have significant drawbacks when faced with complex well observation environments. The gravity data in the well is inevitably disturbed by non-stationary environmental noise, including instrument drift, microseismic vibration and geological background disturbances. While the traditional linear filter (such as wiener filtering) tends to blur geological boundaries while suppressing noise, the wavelet threshold-based method has the advantage of multiple resolutions, but the performance of the wavelet threshold-based method is highly dependent on the basis function and the manual selection of the threshold, so that non-Gaussian composite noise is difficult to adaptively process, and artifacts are easily generated in the denoising process or weak effective signals are erroneously filtered. Traditional gravity data denoising methods mainly rely on wiener filtering or wavelet threshold denoising, and generally lack self-adaptability to non-stationary noise, and cannot cope with complex composite noise interference in wells. In addition, the geological boundary is blurred or artifacts are generated due to improper parameter selection, and the high-precision requirement of deep exploration on weak signal extraction is difficult to meet. In recent years, deep learning technology, particularly Convolutional Neural Network (CNN), has become a research hotspot in the field of geophysical data processing due to its strong nonlinear feature extraction capability, and has achieved remarkable results in aspects such as seismic data denoising. The deep learning method can overcome the dependence of the traditional method on the physical model assumption by learning the statistical rule in a large amount of sample data. However, the application of the deep learning algorithm in the field of well gravity data denoising still has limitations. The existing mainstream denoising model (such as DnCNN) is designed for two-dimensional natural images, and when the model is directly applied to gravity data in a one-dimensional well, the inherent physical continuity and long-distance dependence of a depth domain sequence are often ignored. In addition, conventional small convolution kernels are difficult to capture long wavelength (low frequency) features of gravity anomalies, resulting in low frequency geological trend distortion, and single Mean Square Error (MSE) loss functions are easy to cause an overcomplete phenomenon of signal extremum, so that strict requirements on data fidelity of gravity inversion in a high-precision well are difficult to meet. Disclosure of Invention The invention provides an intelligent denoising method for well gravity data based on a depth residual error network, and aims to solve the problems that a traditional denoising method is insufficient in suppression capability for non-stationary composite noise, geological boundaries are easy to blur, and the existing depth learning model is difficult to adapt to data long-distance dependence characteristics in a well. An intelligent denoising method for well gravity data based on a depth residual error network is realized by the following steps: Constructing a gravity synthesis data set in the well, and carrying out normalization processing on noisy gravity data to obtain a normalized data set; constructing a depth residual error U-shaped network model, and training the network model by adopting the data set to obtain a trained optimal network model; in each round of training, the network model samples a plurality of groups of data from th