A deep learning-based pumping unit noise positioning method

Abstract

The application discloses a pumping unit noise positioning method based on deep learning, and steps are as follows: S1, local seismic trace sets of several fixed sizes are obtained by using sliding window single-step sliding to block seismic data by columns, if the number of seismic traces containing pumping unit noise reaches threshold Th1, it is marked as 1, and if the number of seismic traces containing pumping unit noise is lower than threshold Th2, it is marked as 0; S2, the local seismic trace sets are pretreated, the energy spectrum is calculated, mean filtering is carried out, then down-sampling is carried out, and part of data is randomly selected as a training set, and the remaining part is used as a test set; S3, a deep CNN network is built, and the training set obtained in step S2 is used to train the CNN network, and in the training process, the misclassified data is used to supplement the training set for repeated training; S4, the test sample is used to test the CNN network trained in step S3, and the positioning function is quantitatively evaluated; S5, the width of the positioned pumping unit noise is estimated according to the positioning result in step S4.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical and seismic data detection analysis, and particularly relates to a pumping unit noise positioning method based on deep learning. BACKGROUND

[0002] Pumping unit noise is an important influencing factor for secondary high-precision exploration and deep seismic exploration in old oil fields, and its form is as shown in the drawing. Figure 1 The existence of pumping unit noise seriously reduces the resolution of seismic data, making subsequent seismic data processing more difficult. Closing the pumping unit will seriously affect the production efficiency and cost is high, so the subsequent processing of seismic data is very important. The current pumping unit noise suppression method is mainly blind source separation and abnormal amplitude attenuation. These methods use fixed parameters for noise filtering, which can easily cause noise residue and loss of effective signals. Therefore, the present application proposes a method for automatically positioning pumping unit noise based on deep convolutional neural network (CNN) and estimating the width of pumping unit noise using image post-processing, which is used to adaptively estimate the parameters of noise filtering. SUMMARY

[0003] The present application aims to solve the above problems, and provides a pumping unit noise positioning method based on deep learning, which can automatically position the pumping unit noise in secondary high-precision exploration and deep seismic exploration data, so as to adaptively estimate the parameters of noise filtering and avoid the loss of effective signals to a certain extent. The present application positions the pumping unit noise in seismic data containing pumping unit noise based on deep convolutional neural network (CNN), uses a fixed-size sliding window to slide and block the seismic data, and labels the degree of containing pumping unit noise based on the blocked data.

[0004] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0005] A pumping unit noise positioning method based on deep learning, the steps of which are as follows:

[0006] S1, using a fixed-size sliding window to slide and block the seismic data by column, obtaining a plurality of local seismic trace sets of fixed size, each local seismic trace set containing a fixed number of seismic traces, labeling according to whether the number of seismic traces containing a section of sweeping pumping unit noise reaches a set threshold, if the number of seismic traces containing pumping unit noise reaches the threshold Th1, labeling it as 1, if the number of seismic traces containing pumping unit noise is lower than the threshold Th2, labeling it as 0, and the rest is not used for training;

[0007] S2, pre-processing the local seismic trace set obtained in step S1, calculating its energy spectrum, performing mean filtering, then down-sampling, randomly selecting part of the data after pre-processing as a training set, and the remaining part as a test set;

[0008] S3, building a deep CNN network, and training the CNN network using the training set obtained in step S2, using misclassified data to supplement the training set during training to repeatedly train;

[0009] S4, using the test sample to test the CNN network trained in step S3, and quantitatively evaluating its positioning function;

[0010] S5, according to the positioning result in step S4, estimating the width of the pumping unit noise positioned.

[0011] As an improvement to the above technical solution, in step S1, the column direction of the sliding window is fixed at w; w seismic data is selected each time.

[0012] As an improvement to the above technical solution, in step S2, the local seismic trace set is denoted as X, first, the column direction mean filter is used to obtain the filtered data where A is the mean filter, then the bicubic interpolation is used for down-sampling to obtain

[0013] As an improvement to the above technical solution, in step S3, the CNN network includes four convolutional layers and three fully connected layers; the training set is used to train the CNN network, the test set is used to test the trained CNN network, and the misclassified samples are used to fine-tune the CNN network; the CNN network uses the cross-entropy loss function:

[0014]

[0015] where p represents the true probability distribution, and q represents the probability distribution output by the CNN network.

[0016] As an improvement to the above technical solution, in step S4, the seismic trace where the pumping unit noise is positioned is extended left and right to obtain N, which ensures that it contains complete noise, and is obtained in a similar manner to pre-processing:

[0017]

[0018] where Th mean represents binaryzation using energy mean selection threshold, I open represents image opening operation, and then in The maximum connected domain is found in the middle to obtain the horizontal width W, and 1.2W is used as the final width to compensate for the noise edges lost in the image opening operation.

[0019] Compared with the prior art, the application has the advantages and positive effects that:

[0020] Compared with the traditional manual pumping unit noise positioning method, the pumping unit noise positioning method based on deep learning can automatically position the pumping unit noise, improve the intelligentization of seismic data processing, and effectively save time cost and manpower and material resources. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0022] Figure 1 is a schematic diagram of the noise form of the pumping unit;

[0023] Figure 2 is a flowchart of the present application;

[0024] Figure 3 is a schematic diagram of the CNN network structure of the present application;

[0025] Figure 4 is a schematic diagram of the pumping unit positioning test result of single-shot seismic data;

[0026] Figure 5 is a schematic diagram of the image post-processing result for width estimation; DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, any modification, equivalent replacement, improvement, etc. obtained by those skilled in the art without creative labor should be included in the protection scope of the present application.

[0028] As shown in Figures 1 to 5 ,

[0029] The pumping unit noise positioning method based on deep learning of the present application has the following steps:

[0030] S1, the seismic data is divided into blocks by column using a fixed-size sliding window, obtaining a plurality of local seismic trace sets of fixed size, each local seismic trace set contains a fixed number of seismic traces, and the current local seismic trace set is labeled according to whether the number of seismic traces containing a section of sweep-like pumping unit noise reaches a set threshold, if the number of seismic traces containing pumping unit noise reaches the threshold Th1, it is labeled as 1, if the number of seismic traces containing pumping unit noise is lower than the threshold Th2, it is labeled as 0, and the rest is not used for training;

[0031] S2, the local seismic trace set obtained in step S1 is preprocessed, the energy spectrum is calculated, mean filtering is performed, and then down-sampling is performed, part of the data after preprocessing is randomly selected as a training set, and the remaining part is used as a test set;

[0032] S3, a deep CNN network is built, and the training set obtained in step S2 is used to train the CNN network, and the misclassified data is used to supplement the training set during training to repeatedly train;

[0033] S4, the trained CNN network in step S3 is tested by using the test sample, and the positioning function is quantitatively evaluated;

[0034] S5, according to the positioning result in step S4, the width of the positioned pumping unit noise is estimated.

[0035] In step S1, the fixed size of the sliding window in the column direction is w; w seismic data is selected each time. In step S2, the local seismic trace set is denoted as X, first, the mean filter is used to filter the data in the column direction to obtain the filtered data where A is the mean filter, then the data is down-sampled by using bicubic interpolation to obtain

[0036] In step S3, the CNN network includes four convolutional layers and three fully connected layers; the training set is used to train the CNN network, the test set is used to test the trained CNN network, and the misclassified samples are used to fine-tune the CNN network; the CNN network uses the cross-entropy loss function:

[0037]

[0038] where p represents the true probability distribution, and q represents the probability distribution output by the CNN network.

[0039] In step S4, the seismic trace where the pumping unit noise is located is extended to the left and right to obtain N, which ensures that it contains complete noise, and is obtained in a similar way to preprocessing:

[0040]

[0041] where Th is the threshold value mean represents the binaryzation using the energy mean selection threshold value, I open represents the image opening operation, and then in the maximum connected domain is found to obtain its horizontal width W, and 1.2W is used as the final width to compensate for the loss of noise edges in the image opening operation.

[0042] The experimental seismic data contains 17640 seismic traces, each containing 3501 sampling points, the window width is fixed at 41, so the patch size obtained by blocking is 3501*1, and the size is 350*41 after downsampling. The specific structure of the CNN network is shown in Figure 3 The initial learning rate of the CNN network is lr=0.005, the step size of each convolutional layer is 2*2, the batch_size used for training is 10 each time, and a total of 30 rounds of training are performed. Figure 4 The pumping unit noise positioning results of the present application are shown, the local seismic trace set classified as 1 is marked with a red frame, and the local seismic trace set classified as 1 is directly merged using a larger frame. At the same time, the accuracy of positioning is quantitatively calculated, the error rate is calculated by the number of seismic traces covered by the local seismic trace set labeled with errors containing pumping units, and the proportion of the entire seismic data used for testing, the accuracy = 1-error rate, and the accuracy of the two shot gathers used for testing is 95.09%, 86.87% respectively. And the labeling errors are mostly multi-labels, and the phenomenon of missing labels is less. The width of the positioned pumping unit noise is estimated, the threshold is used for binaryzation, Figure 5 is an example of image post-processing results for width estimation.

Claims

1. A method for locating noise in an oil pumping unit based on deep learning, characterized in that: The steps of this method are: S1. Using a fixed-size sliding window, the seismic data is divided into blocks by column in a single step to obtain several local seismic gathers of fixed size. Each local seismic gather contains a fixed number of seismic traces. The local seismic gather is marked according to whether the number of traces containing a broom-shaped pumping unit noise reaches a set threshold. If the number of seismic traces containing pumping unit noise reaches the threshold Th1, it is marked as 1. If the number of seismic traces containing pumping unit noise is less than the threshold Th2, it is marked as 0. S2. Preprocess the local seismic gathers obtained in step S1, calculate their energy spectrum, perform mean filtering, then downsample, randomly select a portion of the preprocessed data as the training set, and use the remaining portion as the test set. S3. Build a deep CNN network and train the CNN network using the training set obtained in step S2. During the training process, use misclassified data to supplement the training set and train repeatedly. S4. Test the CNN network trained in step S3 using test samples to quantitatively evaluate its localization function. S5. Based on the positioning results in step S4, estimate the width of the located pumping unit noise.

2. The deep learning-based method for locating pumping unit noise as described in claim 1, characterized in that: In step S1, the column direction of the sliding window is fixed at a size of w; w seismic data are selected each time.

3. The deep learning-based method for locating pumping unit noise as described in claim 1, characterized in that: In step S2, the local seismic gather is denoted as X. First, it is subjected to mean filtering along the column direction to obtain the filtered data. Where A is the mean filter, and then bicubic interpolation is used to downsample it to obtain...

4. The deep learning-based method for locating pumping unit noise as described in claim 1, characterized in that: In step S3, the CNN network includes four convolutional layers and three fully connected layers; the CNN network is trained using the training set, tested using the test set, and fine-tuned using misclassified samples; the CNN network uses the cross-entropy loss function. Where p represents the true probability distribution and q represents the probability distribution output by the CNN network.

5. The deep learning-based method for locating pumping unit noise as described in claim 3, characterized in that: In step S4, the seismic trace where the pumping unit noise is located is extended to the left and right to obtain N, ensuring that it contains the complete noise. This is achieved using the following method: Where A is the mean filter, and R... bicubic To perform downsampling using bicubic interpolation, where Th... mean The representative uses the energy mean to select a threshold for binarization, I open Indicates the image opening operation, then... The horizontal width W is obtained by finding the largest connected component in the image. To compensate for the noise edges lost in the image opening operation, 1.2W is used as the final width.

Images