Well gravity data intelligent denoising method based on deep residual network

Abstract

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

Description

Technical Field

[0001] This invention relates to the field of geophysical exploration data processing technology, specifically to an intelligent denoising method for wellbore gravity data based on deep residual networks. By introducing deep residual learning technology, this invention improves the signal recovery accuracy and geological feature fidelity of wellbore gravity data in non-stationary, high-noise environments. This method is applicable to deep mineral resource exploration, oil and gas reservoir detection, and other geophysical applications with stringent signal-to-noise ratio requirements for wellbore gravity data. Background Technology

[0002] In deep mineral exploration and oil and gas reservoir detection, wellbore gravity measurement (BHGM) is one of the key technologies for obtaining deep subsurface density imaging. Traditional signal processing algorithms, such as Wiener filtering and wavelet transform, have been widely used in gravity data preprocessing due to their mature theory and ease of implementation. These methods attempt to separate signal from noise through frequency domain or time-frequency domain transformations, which can improve the signal-to-noise ratio of the data to some extent.

[0003] Traditional signal processing algorithms have significant limitations when dealing with complex well observation environments. Well gravity data are inevitably affected by non-stationary environmental noise, including instrument drift, microseismic vibrations, and geological background interference. Traditional linear filters (such as Wiener filters) often blur geological boundaries while suppressing noise; and while wavelet thresholding methods offer multi-resolution advantages, their performance is highly dependent on the manual selection of basis functions and thresholds, making it difficult to adaptively handle non-Gaussian composite noise and prone to artifacts or erroneously filtering out weak, valid signals during denoising.

[0004] Traditional gravity data denoising methods mainly rely on Wiener filtering or wavelet thresholding, which typically lack adaptability to non-stationary noise and cannot cope with complex composite noise interference in wells. Furthermore, they can lead to blurred geological boundaries or artifacts due to improper parameter selection, making it difficult to meet the high-precision requirements for weak signal extraction in deep exploration.

[0005] In recent years, deep learning technology, especially convolutional neural networks (CNNs), has become a research hotspot in the field of geophysical data processing due to its powerful nonlinear feature extraction capabilities, achieving remarkable results in areas such as seismic data denoising. Deep learning methods, by learning the statistical regularities in large amounts of sample data, can overcome the dependence of traditional methods on assumptions about physical models.

[0006] However, the application of deep learning algorithms in wellbore gravity data denoising still has limitations. Existing mainstream denoising models (such as DnCNN) are mostly designed for two-dimensional natural images. When directly applied to one-dimensional wellbore gravity data, they often ignore the inherent physical continuity and long-range dependencies of the depth domain sequence. Furthermore, conventional small convolutional kernels struggle to capture the long-wavelength (low-frequency) features of gravity anomalies, leading to distortion of low-frequency geological trends; and a single mean square error (MSE) loss function can easily cause oversmoothing of signal extrema, making it difficult to meet the stringent data fidelity requirements of high-precision wellbore gravity inversion. Summary of the Invention

[0007] This invention provides an intelligent denoising method for well gravity data based on deep residual networks, aiming to solve the problems of insufficient suppression of non-stationary composite noise, easy blurring of geological boundaries, and difficulty of adapting existing deep learning models to the long-distance dependency characteristics of well data in traditional denoising methods.

[0008] A method for intelligent denoising of well gravity data based on deep residual networks is proposed, which is implemented by the following steps:

[0009] Construct a well-source gravity composite dataset and normalize the noisy gravity data to obtain a normalized dataset.

[0010] Construct a deep residual U-shaped network model, train the network model using the dataset, and obtain the best network model after training;

[0011] During network model training, the objective is to minimize the mixture loss function; in each training round, the network model samples multiple sets of data from the dataset to update the weight parameters of the network model.

[0012] The normalized noisy gravity data is input into the optimal network model, which outputs the estimated background noise distribution. Then, using the residual subtraction formula... Obtain the denoised normalized signal Finally, normalize the signal. Multiply by the corresponding scaling factor This completes the inverse normalization process, ultimately achieving intelligent denoising of gravity data in wells.

[0013] The beneficial effects of this invention are:

[0014] The denoising method described in this invention introduces deep learning technology and is based on a one-dimensional residual U-Net (DnResUnet) model architecture to achieve intelligent gravity data denoising. The DnResUnet model, through the fusion of large convolutional kernels and multi-scale features, can effectively capture the long-distance dependencies and low-frequency geological trends unique to well data, overcoming the low-frequency distortion problem caused by the direct application of existing two-dimensional image denoising models.

[0015] The denoising method described in this invention employs a hybrid loss function tailored to the physical characteristics of gravity anomaly signals. By combining mean square error (MSE) and total variational (TV) regularization, the network can remove strong noise while forcibly constraining the output signal to maintain physical smoothness, avoiding over-smoothing or high-frequency oscillations caused by a single MSE loss. This mechanism ensures that the denoised gravity anomaly curve possesses both an extremely high signal-to-noise ratio and conforms to geological and physical laws.

[0016] The denoising method described in this invention not only demonstrates robustness in high-noise environments (significantly improving PSNR under strong noise) but also possesses excellent geological morphology fidelity, adapting to gravity data processing needs under various geological structural conditions. It also provides high-quality basic data support for deep mineral exploration and geological inversion. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the DnResUnet model in the intelligent denoising method for well gravity data based on deep residual networks described in this invention.

[0018] Figure 2 This is a schematic diagram of the internal structure of the ResBlock in the intelligent denoising method for well gravity data based on deep residual networks described in this invention.

[0019] Figure 3 This is a comparison of logging curves showing the denoising results of the intelligent denoising method for well gravity data based on deep residual networks described in this invention under high-noise environments.

[0020] Figure 4 The corresponding amplitude spectrum comparison diagram;

[0021] Figure 5 This is the time-frequency diagram of the noisy signal;

[0022] Figure 6 This is the time-frequency diagram of the signal after denoising using the DnResUnet denoising network model. Detailed Implementation

[0023] Combination Figures 1 to 6This embodiment describes an intelligent denoising method for wellbore gravity data based on deep residual networks. This intelligent denoising method incorporates deep learning technology, specifically a one-dimensional residual U-Net architecture, namely the DnResUnet denoising network model. This denoising network model combines the encoder-decoder architecture of U-Net with the residual learning module of ResNet, and is designed specifically for the one-dimensional sequence characteristics of wellbore gravity data, achieving adaptive and high-precision denoising. This method transforms the complex signal recovery problem into a noise distribution estimation problem through a global residual learning strategy. Combined with a hybrid loss function based on physical constraints, the DnResUnet denoising network model can automatically learn and optimize the denoising strategy, maximizing the fidelity of geological signals without relying on manual selection of thresholds or basis functions, overcoming the shortcomings of traditional filtering and wavelet denoising methods.

[0024] like Figure 1 As shown, Figure 1 This is the DnResUnet denoising network model, which receives noisy gravity data. As input, the network model employs a U-Net structure, extracting multi-scale features through downsampling, capturing long-range geological trends using large convolutional kernels, and fusing high-frequency details through skip connections. The network model's output is not a direct denoised signal, but rather an estimated background noise distribution. Ultimately, by using noisy gravity data... Compared with the estimated background noise distribution By subtracting the residuals, a high-fidelity, noise-free gravity signal is obtained. .

[0025] The specific implementation process of the noise reduction method in this embodiment is as follows:

[0026] Step S1. Construct a large-scale well gravity synthesis dataset based on physical forward modeling, and perform normalization operations to finally obtain the normalized dataset;

[0027] This implementation method is based on a geophysical forward modeling framework, generating a density model that includes diverse geological structures, and obtaining real gravity signals through forward modeling calculations. The generated noisy gravity data is then injected with a composite noise consisting of Gaussian white noise, linear drift, and colored noise. Normalization is performed to obtain the normalized dataset; the specific implementation process is as follows:

[0028] Step S11. Construct diverse underground 3D density models; to ensure the generalization ability of the model, generate two types of geological structures;

[0029] 1. Layered geological model: consisting of 2 to 5 horizontal strata of random thickness and density, used to simulate large-scale background fields;

[0030] 2. Discrete anomaly model: Contains blocky, spherical or polyhedral anomalies with random locations, sizes and densities, used to simulate local ore bodies or cavities.

[0031] Step S12. Calculate the true gravity signal using the forward gravity modeling formula. For the discretized underground prism element, the true gravity signal at the i-th sampling point Calculation is all Linear superposition of subdivided units:

[0032] ;

[0033] In the formula, The sensitivity kernel function (geometric factor) represents the gravitational response of a unit density prism to the observation point; This represents the density value of the j-th element; this implementation uses the Nagy analytical formula for calculation. :

[0034] ;

[0035] in, The kernel function's analytical expression. The horizontal distance between the observation point and the boundary of the prism along the x-axis is given. The distance between the observation point and the boundary of the prism along the y-axis is [missing information]. This represents the perpendicular distance between the observation point and the boundary of the prism along the z-axis (depth direction). Unit conversion factor ( ), The directional weight coefficient, with a value of +1 or -1, automatically handles the addition and subtraction relationships of the contributions from the eight vertices based on the Newton-Leibniz formula for triple integrals. This physical modeling ensures the physical correctness of the training data.

[0036] Step S13. Construct a composite noise model that conforms to the well observation environment and generate noisy gravity data. ;

[0037] To simulate a real well measurement environment, this implementation method uses real gravity signals. Three types of non-stationary noise are injected into the mixture to form composite noise. The three types of non-stationary noise are: Gaussian white noise, linear drift, and colored noise.

[0038] In this embodiment, the Gaussian white noise is used to simulate the high-frequency random thermal noise of the sensor; the linear drift is used to simulate the low-frequency baseline drift caused by instrument temperature changes or spring creep, and its amplitude is randomly sampled from the noise level. 5% to 20%; the colored noise is used to simulate space-related environmental disturbances and is generated by applying a moving average filter to white noise; the final noisy gravity data Represented as linear superposition: To cover different exploration scenarios, multiple noise intensity levels were set ( mGal) generated a total of 100,000 noisy-noise-free sample pairs;

[0039] Step S14. Process the noisy gravity data generated in step S13. Normalization is performed to ensure that the input and output of the DnResUnet denoising network model are within a uniform range, forming the final dataset.

[0040] Because the magnitudes of gravity anomalies generated by different geological models vary greatly, direct training can lead to network convergence difficulties. This implementation employs a sample-level maximum value normalization strategy. The specific normalization formula is as follows:

[0041] ;

[0042] In the formula, and Let represent the noisy gravity data and the true gravity signal of the k-th sample, respectively; This represents the maximum absolute amplitude of the current noisy sample sequence, i.e., the scaling factor for each sample. Through this processing, all input data is mapped to the interval [-1, 1], and the scaling factor for each sample... It is preserved so that it can be denormalized after the network denoising model inference to recover the real physical unit (mGal).

[0043] Step S2. Construct a DnResUnet denoising network model. This model is used to establish a nonlinear mapping relationship between noisy gravity data in the well and the background noise distribution, simulating the complex statistical characteristics of gravity noise in the well. The specific construction process of the denoising network model structure is as follows:

[0044] Step S21. Construct the encoder (downsampling path): It consists of cascaded downsampling modules (Down); each Down module includes a max pooling layer (MaxPool1d, stride of 2) and a residual block (ResBlock). The max pooling layer is used to compress the signal length and expand the receptive field, enabling the denoising network model to capture long-range geological dependencies; at the same time, the number of feature channels increases layer by layer (e.g., from 32 to 256) to encode high-level semantic information.

[0045] Step S22. Decoder Construction (Upsampling Path): Consists of cascaded upsampling modules (Up). Each Up module uses linear interpolation for upsampling, gradually restoring the original resolution of the signal. At each upsampling stage, a skip connection is introduced to concatenate the shallow features of the corresponding layer of the encoder with the deep features of the decoder along the channel dimension. This mechanism directly transmits high-frequency detail information, preventing the loss of edge features related to gravity anomalies in deep networks.

[0046] Step S23. Constructing Residual Blocks (ResBlocks): The residual blocks serve as the core units of the denoising network model. For example... Figure 2 As shown, the residual block includes a main path and a shortcut branch; the main path consists of two one-dimensional convolutional layers (Conv1d) sandwiched between batch normalization (BN) and ReLU activation functions; the shortcut branch includes a convolutional layer for aligning the number of channels. To capture the long-wavelength features (low-frequency trend) of gravity anomalies, the first layer of the denoising network model and the convolutional layers within the residual block are designed with large convolutional kernels (Kernel Size=7) to significantly expand the effective receptive field.

[0047] Step S24. Output Layer Design: The last layer of the denoising network model is an output convolutional layer (OutConv), which does not use any non-linear activation function. This is because the task of the denoising network model is to estimate the background noise distribution. Its value range is unrestricted (it can be positive or negative), and the linear output layer ensures that the denoising network model can accurately adapt to the amplitude changes of the input signal, thereby supporting accurate residual subtraction operations.

[0048] Step S3. Use the normalized dataset obtained in Step S1 to train and evaluate the denoising network model, and deploy the network model for actual denoising; the specific implementation process is as follows:

[0049] Step S31. Design a hybrid loss function to optimize and update the DnResUnet denoising network model. To preserve the physical form of gravity anomalies while removing noise, this implementation abandons the single mean square error loss and designs a hybrid loss function that includes physical constraints. The specific implementation process is as follows:

[0050] First, define the mean square error term ( As a fidelity loss, it is used to quantize the denoised gravity signal prediction value output by the denoising network model. (Noise-free gravity signal) and real gravity signal The difference in amplitude between them can be expressed by the following formula:

[0051] ;

[0052] In the formula, N is the total number of sampling points contained in a single sample sequence; i is the index of the sampling point. This represents the actual gravity signal at the i-th sampling point; Let be the predicted value of the denoised gravity signal at the i-th sampling point output by the denoising network model. This mean square error term helps the output of the denoising network model numerically approximate the true value.

[0053] Then define the total variation regularization term ( As a smoothness constraint, TV regularization is introduced because the gravitational field has physical continuity on a large scale, effectively suppressing high-frequency artifacts and spurious oscillations. It is defined as the sum of squares of the differences between adjacent data points:

[0054] ;

[0055] In the formula, This represents the denoised gravity signal amplitude at the l-th data point in the b-th batch and c-th channel. The amplitude of the denoised gravity signal at the (l+1)th adjacent data point is represented; B, C, and L are the batch size, number of channels, and signal length, respectively.

[0056] Finally, a hybrid loss function is constructed; expressed as follows:

[0057] ;

[0058] In this embodiment, regularization weights are set. The weight setting ensures that the network model focuses on both the accuracy of signal reconstruction and the physical smoothness of the output curve during training, which conforms to geological laws.

[0059] Step S32. Training the denoising network model: Divide the dataset into a training set (80%), a validation set (10%), and a test set (10%). Use the hybrid loss function designed in Step S31, based on the PyTorch framework, and iterate the parameters using the Adam optimizer. Set the initial learning rate to... The batch size is 64;

[0060] Step S33. Early Stopping and Saving of Denoising Network Model: An early stopping mechanism is introduced during training. If the loss function on the validation set does not decrease within 10 consecutive training epochs, training is automatically terminated, and the parameters of the denoising network model with the lowest validation loss are saved as the optimal denoising network model. This strategy effectively prevents overfitting and ensures the generalization ability of the denoising network model to unknown data.

[0061] Step S34. Reasoning and inverse normalization;

[0062] After training, the saved optimal denoising network model is used to process the noisy gravity data. Processing is then performed. First, based on the normalization strategy described in step S14, normalized noisy gravity data is obtained. Then the noisy gravity data Input a DnResUnet network model, output the estimated background noise distribution Then, using the residual subtraction formula... The denoised normalized signal is obtained. Finally, normalize the signal. Multiply by the scaling factor corresponding to the sample This completes the inverse normalization process, restoring the final denoising result with physical units.

[0063] This embodiment also includes performance verification of the denoising network model: evaluating the performance of the denoising network model on a test set. For example... Figures 3 to 6 As shown, this demonstrates the effect in a high-noise environment ( The noise reduction effect under ( ). Among them, Figure 3 A comparison chart of well logging curves showing the noise reduction results. Figure 4 Here is a comparison chart of the corresponding amplitude spectra. Figure 5 This is the time-frequency plot of the noisy signal. Figure 6 The image shows the time-frequency plot of the DnResUnet denoised signal. It can be seen that the method of this invention can effectively remove strong clutter covering the signal. The denoised curve (blue solid line) highly overlaps with the real geological signal (black dashed line), with no significant amplitude attenuation or phase lag. Quantitative indicators show that the peak signal-to-noise ratio (PSNR) of this invention at this noise level is improved by approximately 22 dB compared to the noisy data, and the goodness-of-fit coefficient (GFC) is increased to 0.8257, verifying its robustness and effectiveness in complex environments.

[0064] This embodiment also includes verifying the denoising performance of the DnResUnet denoising network model after training. Verification includes the model's generalization ability under different signal-to-noise ratios and its accuracy in recovering weak geological signals. Verification metrics are set as mean squared error (MSE), peak signal-to-noise ratio (PSNR), and goodness-of-fit coefficient (GFC) to comprehensively evaluate the amplitude accuracy and waveform similarity of the denoising results. The weights of the trained DnResUnet denoising network model are converted and saved in a lightweight inference format (such as ONNX or TorchScript) to improve the processing efficiency of large-scale well logging data. Then, the denoising network model is deployed to a geophysical data processing workstation or high-performance computing cluster to achieve rapid denoising processing of gravity data from field well logging.

[0065] In deploying applications, through scaling factors Inverse scaling is performed to ensure that the output strictly corresponds to the actual physical units of gravity anomalies (mGal), meeting the strict requirements of geological interpretation and inversion for the physical meaning of data.

[0066] The denoising method described in this embodiment introduces deep residual learning technology based on the DnResUnet denoising network model. Utilizing one-dimensional convolution and a large receptive field design, it automatically captures long-range dependency features of depth-domain sequences, thereby achieving high-precision and adaptive denoising of well gravity data. The denoising method employs a global residual learning strategy, transforming the complex signal recovery task into an estimation problem of background noise distribution. Noise is predicted from input noisy data, and the effective signal is recovered through residual subtraction. To ensure the geological rationality of the denoising results, the network combines encoder-decoder multi-scale feature fusion to avoid distortion of low-frequency geological trends. The loss function design adopts a hybrid form of mean squared error and total variational regularization, minimizing reconstruction error while forcibly constraining the physical smoothness of the output signal, aiming to balance denoising fidelity and the integrity of geological morphology. Through training and validation on large-scale physical simulation datasets, the DnResUnet denoising network model maintains robustness in complex environments with strong noise interference, ultimately achieving high-quality gravity signal extraction.

[0067] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0068] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An intelligent denoising method for well gravity data based on deep residual networks, characterized by: The specific implementation process of this method is as follows: Construct a well-source gravity composite dataset and normalize the noisy gravity data to obtain a normalized dataset. Construct a deep residual U-shaped network model, train the network model using the dataset, and obtain the best network model after training; During network model training, the objective is to minimize the mixture loss function; in each training round, the network model samples multiple sets of data from the dataset to update the weight parameters of the network model. The normalized noisy gravity data is input into the optimal network model, which outputs the estimated background noise distribution. ; Then, using the residual subtraction formula... Obtain the denoised normalized signal Finally, normalize the signal. Multiply by the corresponding scaling factor This completes the inverse normalization process, ultimately achieving intelligent denoising of gravity data in wells.

2. The intelligent denoising method for well gravity data based on deep residual networks according to claim 1, characterized in that: The specific process for obtaining the gravity synthesis dataset is as follows: Calculate the real gravity signal using the forward gravity modeling formula; Injecting composite noise into the real gravity signal to generate noisy gravity data; The noisy gravity data are preprocessed using the maximum value normalization method. The normalization formula is as follows: ； In the formula, and These are the noisy gravity data and the true gravity signal for the k-th sample, respectively. is the scaling factor, and its value is the maximum absolute magnitude of the k-th sample.

3. The intelligent denoising method for well gravity data based on deep residual networks according to claim 2, characterized in that: The composite noise includes Gaussian white noise, linear drift, and colored noise.

4. The intelligent denoising method for well gravity data based on deep residual networks according to claim 1, characterized in that: The deep residual U-shaped network model includes a one-dimensional convolutional encoder and a decoder; The encoder consists of cascaded downsampling modules, each of which includes a max pooling layer and a residual block. It utilizes a large convolutional kernel with a kernel size of 7 to expand the receptive field and extract long-range dependency features of the deep domain sequence. The residual block includes a main path and a shortcut branch. The main path consists of two one-dimensional convolutional layers with a batch normalization layer and a ReLU activation function in between. The shortcut branch includes a convolutional layer for aligning the number of channels. The decoder consists of cascaded upsampling modules, which use linear interpolation to recover the signal length and splice the shallow features of the encoder with the deep features of the decoder in the channel dimension through skip connections to fuse high-frequency information.

5. The intelligent denoising method for well gravity data based on deep residual networks according to claim 1, characterized in that: The hybrid loss function is set to minimize the weighted sum of the reconstruction error and the total variation regularization term. This can be expressed as follows: ； In the formula, This is the mean square error term. For total variational regularization, This is a true gravity signal. This is the regularization weight coefficient.

6. The intelligent denoising method for well gravity data based on deep residual networks according to claim 5, characterized in that: Set regularization weight coefficients The weighting coefficients are determined based on the balance requirements between signal fidelity and physical smoothness.

7. The intelligent denoising method for well gravity data based on deep residual networks according to claim 5, characterized in that: The mean square error term As a fidelity loss, it is used to quantify the denoised gravity signal prediction value output by the network model. Compared with real gravity signals Differences in amplitude between them: ； In the formula, N is the total number of sampling points contained in a single sample sequence; i is the index of the sampling point. This represents the actual gravity signal at the i-th sampling point; This is the predicted value of the denoised gravity signal at the i-th sampling point output by the denoising network model. The total variation regularization term Defined as the sum of squares of the differences between adjacent data points: ； In the formula, in the formula, This represents the denoised gravity signal amplitude at the l-th data point in the b-th batch and c-th channel. The amplitude of the denoised gravity signal at the (l+1)th adjacent data point is represented; B, C, and L are the batch size, number of channels, and signal length, respectively.

8. The intelligent denoising method for well gravity data based on deep residual networks according to claim 1, characterized in that: During the training of the network model, the Adam optimizer is used for parameter updates, and the initial learning rate is set to... An early stopping mechanism is introduced: if the loss function on the validation set does not decrease within 10 consecutive training epochs, training is terminated and the optimal network model parameters are saved.

9. A method for intelligent denoising of well gravity data based on deep residual networks according to any one of claims 1-8, characterized in that: The best trained network model is then converted into a lightweight format and deployed to a computing platform to enable real-time processing of gravity data in wells.

Images