A Deep Learning-Based Method for Measuring Source Well Depth in Seismic Exploration

Abstract

This invention discloses a deep learning-based method for measuring the depth of seismic wells in seismic exploration. The method divides field construction videos frame by frame into two categories: insertion / removal frames and non-insertion / removal frames. Image recognition and target detection training sets are created based on these two types of video frames. The datasets are input into the image recognition neural network and the target detection neural network, respectively, for training. The neural networks are then tuned to achieve optimal fit. The construction video is input into the trained image neural network for inference. If the video frame shows drill pipe insertion / removal operations, the target detection neural network detects the drill pipe position and records the drill pipe apex change curve. Gaussian filtering is applied to the curve, and a window selection algorithm is executed on the filtered curve to count the drill pipes. Finally, the depth is measured by multiplying the number of drill pipes by the drill pipe length. This invention effectively improves the overall video processing efficiency and significantly reduces the working time for measuring the depth of seismic wells from large amounts of actual construction video.

Description

Technical Field

[0001] This invention relates to the field of construction technology in seismic exploration, and in particular to a method for measuring the depth of seismic source wells based on deep learning. Background Technology

[0002] In oil exploration, it is often necessary to manually drill seismic wells to artificially induce seismic explosions and obtain data. The depth of the seismic well is subject to strict specifications and requirements. To prevent workers from drilling unusable wells due to improper depth, they typically wear recording devices to document the entire process during fieldwork. The number of drill pipes is then manually counted based on the recorded video, and the depth of the seismic well is calculated by multiplying the number of drill pipes by their length to determine if the depth meets the specifications. In simple terms, seismic exploration well depth measurement involves accurately counting the number of drill pipes based on the video footage and then multiplying the number by the pipe length to obtain the well depth. As the number of explorations increases, the amount of recorded video footage also increases dramatically. Manually reviewing and counting drill pipes to measure the well depth in the face of this massive amount of video footage is labor-intensive, time-consuming, inefficient, and lacks accuracy. Currently, there is no effective method for automatically measuring the depth of seismic exploration wells. Summary of the Invention

[0003] The purpose of this invention is to provide a deep learning-based method for measuring the depth of seismic wells in seismic exploration. This method can solve the problems of time-consuming and inefficient manual counting and processing of massive amounts of construction video data in practical engineering, while effectively improving the accuracy of well depth measurement by combining methods from the field of deep learning.

[0004] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0005] This invention includes the following steps:

[0006] S101: Obtain actual source well construction videos in seismic exploration projects;

[0007] S102: Divide the actual seismic source well construction video frame by frame into two categories: insertion / removal behavior frames and non-insertion / removal behavior frames.

[0008] S103: Create an image discrimination dataset for the two types of video frames that have been classified, and create an object detection dataset by labeling the drill rod in the insertion and removal behavior frames;

[0009] S104: Input the image discrimination dataset into the image recognition neural network for training, input the object detection dataset into the object detection network for training, and adjust the parameters of the two neural networks to make them fit each other.

[0010] S105: Input the actual seismic source well construction video from step S101 into the image recognition neural network for inference. If the current video frame is determined to be the operation of inserting and removing drill pipe, then input the video frame into the target detection network to detect the position of drill pipe. Otherwise, input the next frame of the video into the image recognition neural network for inference and judgment.

[0011] S106: Use an object detection network to detect the position and size of the drill rod detection box in the part of the video where the drill rod is being inserted and removed. Then, if the video ends, record the entire drill rod vertex curve. Otherwise, input the next frame of the video into the image recognition neural network for inference and judgment.

[0012] S107: Record the curve of the drill pipe tip change;

[0013] S108: Analyze the drill pipe tip variation curve obtained in step S107. First, perform Gaussian filtering on the drill pipe tip variation curve. Then, execute a window selection algorithm on the filtered curve to obtain the number of drill pipes pulled out and the number of drill pipes inserted. Multiply the maximum value of the two by the pipe length to obtain the source well depth.

[0014] The method for creating the image discrimination dataset in step S103 is as follows: the video frames are divided into two categories according to whether construction is underway, and placed into two folders respectively; the method for creating the target detection dataset is as follows: the annotation tool is used to mark the drill rods in the drill rod insertion and removal video frames with annotation boxes, and then the annotated video frames are saved in YOLO format (Redmon J et al., 2016).

[0015] Step S104 specifically involves: First, building an image recognition neural network, inputting the image recognition dataset into the image recognition neural network for training, and adjusting the parameters of the image recognition neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range; then, building an object detection neural network, inputting the object detection dataset into the object detection neural network for training, and adjusting the parameters of the object detection neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range.

[0016] In step S107, the change curve of the drill rod vertex is recorded as follows: the position and size of the drill rod detection box are obtained according to the output of the target detection neural network, the coordinates of the drill rod vertex are calculated according to the coordinate origin and the position and size of the detection box, and then recorded.

[0017] The Gaussian filtering formula in step S108 is:

[0018]

[0019] Where x′ is the filtered curve, x is the original curve, σ is the standard deviation of the Gaussian kernel, and μ is the curve mean.

[0020] The window selection algorithm in step S108 includes the following steps:

[0021] S1081: Set the window threshold σ t and the merging threshold σ i ;

[0022] S1082: Traverse the curve; if the curve value is greater than the window threshold σ t Then, starting from the beginning of the window, record the coordinate data of the window's starting point on the curve;

[0023] S1083: Starting from the position recorded in step S1082, traverse the curve. When the curve value is less than the window threshold σ... t Then the window ends, and the coordinates of the window's end point on the curve are recorded;

[0024] S1084: If the end of the curve is not reached, S1082 and S1083 will be executed again starting from the end point recorded in S1083. If the curve ends, jump to S1085.

[0025] S1085: Traverse all selected windows. If the distance between two windows is less than the merging threshold σ... i The two windows will be merged, with the starting point of the left window used as the starting point of the merged window and the ending point of the right window used as the ending point of the merged window.

[0026] S1086: Find the maximum value in all windows, and the coordinate data of the maximum value on the curve;

[0027] S1087: Traverse all windows. If the maximum value is closer to the beginning of the window, count it as inserting the drill rod. If the maximum value is closer to the end of the window, count it as pulling out the drill rod.

[0028] The beneficial effects of this invention are:

[0029] This invention is a deep learning-based method for measuring the depth of seismic source wells in seismic exploration. Compared with existing technologies, this invention has the following technical advantages:

[0030] (1) This invention combines the ideas of image recognition and target detection in deep learning to achieve accurate counting of drill rods. It utilizes the powerful data analysis and representation capabilities of deep learning networks to detect the size and position of drill rods, thereby improving the overall accuracy and efficiency of construction video processing through image recognition.

[0031] (2) This invention designs a multi-stage, multi-scale neural network that can accurately determine whether the current video frame is in the process of inserting or removing drill rods, and accurately locate the position and size of the drill rods during the insertion or removal process, based on the characteristics of deep learning target detection and image recognition tasks. This effectively solves the interference of non-insertion / removal drill rod frames on the final curve and improves the accuracy of counting during the final curve analysis.

[0032] (3) When performing curve analysis and counting, the window counting method used in this invention can accurately capture the drill rod change signal, accurately capture the characteristics of inserting and removing the drill rod, and perform accurate counting.

[0033] (4) The neural network model and training environment used in this invention are not limited to a specific framework and can be selected according to the actual application scenario. Therefore, this invention can be deployed on platforms with different operating systems and different compilation environments. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the specific implementation of the present invention;

[0035] Figure 2 This is a schematic diagram of the target detection dataset annotation in an embodiment of the present invention;

[0036] Figure 3 This is a diagram of the ConvNeXt network structure in an embodiment of the present invention;

[0037] Figure 4 This is a diagram of the YOLOv5 network structure in an embodiment of the present invention;

[0038] Figure 5 This is a target detection result diagram in an embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram of curve analysis in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0041] This example provides a deep learning-based method for measuring the depth of seismic source wells in seismic exploration. In this embodiment, the construction video is an actual construction recording from a work area in Southwest China. Please refer to... Figure 1 The method of the present invention specifically includes the following steps:

[0042] S101: Obtain actual source well construction videos in seismic exploration projects;

[0043] S102: Divide the video frame by frame into two categories: plug-in behavior frames and non-plug-in behavior frames;

[0044] In this embodiment, in step S102, the 15 construction videos are first processed into video frames, and the video frames are divided into two categories according to whether the drill rod is being inserted or removed.

[0045] S103: Create an image discrimination dataset for the two types of video frames that have been classified, and create an object detection dataset by labeling the drill rod in the insertion and removal behavior frames;

[0046] The specific steps for creating an image recognition dataset are as follows:

[0047] The video frames were divided into two categories based on whether construction was underway, and placed into two separate folders.

[0048] Figure 2 This is a schematic diagram of a video frame and an labeled video frame in the training set of this embodiment. The target detection dataset that can be used for network training in step S103 can be created using the two. The specific steps are: use the annotation tool to mark the drill rod in the drill rod insertion and removal video frame with annotation boxes; save the annotated video frame in YOLO format.

[0049] S104: Input the image discrimination dataset into the image recognition neural network for training, input the object detection dataset into the object detection network for training, and adjust the parameters of the two neural networks to make them fit each other.

[0050] In this embodiment, the image recognition neural network uses the ConvNeXt neural network (Liu Z et al., 2022), such as... Figure 3 As shown, the object detection neural network uses the YOLOv5 neural network (Jocher G et al., 2022), as follows: Figure 4 As shown, the specific steps for training two neural networks include:

[0051] S1041: ConvNeXt neural network is built as an image recognition neural network. The input of the network is a 224×224 video frame image, and the final output of the network is the classification of the image.

[0052] S1042: Input the image recognition dataset into the ConvNeXt neural network. The self-supervised learning label is the folder name used for classification data. The dataset is randomly divided into training and test sets (80% and 20% respectively). The training set is used to train the network, and the test set is used to evaluate the network's fit. The number of training epochs is set to 100, the batch size to 8, and the learning rate to 0.0005.

[0053] S1043: Adjust the parameters of the ConvNeXt neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range.

[0054] S1044: Build a YOLOv5 neural network as an object detection neural network. The input of the network is a 640×640 video frame image. The final output of the network is the location of the detected object, the size of the object box, and the type of the detected object.

[0055] S1045: Input the object detection dataset into the YOLOv5 neural network. The labels for supervised learning are the labels generated during the annotation process. The dataset is randomly divided into training and test sets, with 80% and 20% of the data allocated to each. The training set is used to train the network, and the test set is used to evaluate the network's fit. The number of training epochs is set to 100, the batch size is set to 16, and the learning rate is set to 0.0005.

[0056] S1046: Adjust the parameters of the YOLOv5 neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range.

[0057] S105: Input the video from step S101 into the image recognition neural network for inference. If the current video frame is determined to be a drill rod insertion / removal operation, then input the video frame into the target detection network to detect the drill rod position. Otherwise, input the next video frame into the image recognition neural network for inference and judgment.

[0058] S106: Use an object detection network to detect the position of the drill rod in the video frame of the drill rod insertion / removal operation, such as... Figure 5 As shown. Then, if the video ends, the entire curve of the drill rod vertex in the video is recorded; otherwise, the next frame of the video is input into the image recognition neural network for inference and judgment.

[0059] S107: Record the curve of the drill pipe tip change;

[0060] In this embodiment, the specific steps for recording the change curve of the drill rod vertex are as follows: the position and size of the drill rod detection frame are obtained according to the output of step S106; the coordinates of the drill rod vertex are calculated based on the coordinate origin of the upper left corner and the position and size of the detection frame, and then recorded.

[0061] S108: Analyze the obtained curves. First, apply Gaussian filtering to the curves. Then, execute a window selection algorithm on the filtered curves to obtain the number of drill pipes pulled out and the number of drill pipes inserted. Multiply the maximum of these two numbers by the pipe length to obtain the source well depth. The Gaussian filtering formula is:

[0062]

[0063] Where x′ is the filtered curve, x is the original curve, σ is the standard deviation of the Gaussian kernel, and μ is the curve mean. In this embodiment, the Gaussian filter parameter σ is set to 2.

[0064] The specific steps of the window selection algorithm in this embodiment are as follows:

[0065] S1081: Set the window threshold σ t and the merging threshold σ i In this embodiment, the window threshold is set to 50, and the merging threshold is set to 10.

[0066] S1082: Traverse the curve; if the curve value is greater than the window threshold σ t Then the window begins, recording the coordinates of the window's starting point on the curve.

[0067] S1083: Start traversing the curve from the position recorded in step two. When the curve value is less than the window threshold σ... t Then the window ends, and the coordinates of the window's end point on the curve are recorded.

[0068] S1084: If the end of the curve is not reached, S1082 and S1083 will be executed again starting from the end point recorded in S1083. If the curve ends, jump to S1085.

[0069] S1085: Traverse all selected windows. If the distance between two windows is less than the merging threshold σ... i The two windows will be merged, with the starting point of the left window becoming the starting point of the merged window, and the ending point of the right window becoming the ending point of the merged window.

[0070] S1086: Find the maximum value in all windows, and the coordinate data of the maximum value on the curve;

[0071] S1087: Traverse all windows. If the maximum value is close to the beginning of the window, count it as inserting the drill rod. If the maximum value is close to the end of the window, count it as pulling out the drill rod.

[0072] Figure 6 The curve analysis process in this embodiment is shown. (a) is the recorded curve of the drill pipe tip change, (b) is the curve after Gaussian filtering, and (c) is all the windows selected after the window selection algorithm is executed, as well as the window start point and maximum value point.

[0073] This embodiment illustrates the deep learning-based seismic exploration source well depth measurement method of the present invention, which effectively improves the overall efficiency of construction video processing, greatly reduces the working time of source well depth measurement tasks with a large number of construction videos, and provides an efficient and universal framework template for drilling pipe counting processing in similar geological engineering projects in the future.

[0074] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for measuring the depth of seismic source wells based on deep learning, characterized in that, Includes the following steps: S101: Obtain actual source well construction videos in seismic exploration projects; S102: Divide the actual seismic source well construction video frame by frame into two categories: insertion / removal behavior frames and non-insertion / removal behavior frames. S103: Create an image discrimination dataset for the two types of video frames that have been classified, and create an object detection dataset by labeling the drill rod in the insertion and removal behavior frames; S104: Input the image discrimination dataset into the image recognition neural network for training, input the object detection dataset into the object detection network for training, and adjust the parameters of the two neural networks to make them fit each other. S105: Input the actual seismic source well construction video from step S101 into the image recognition neural network for inference. If the current video frame is determined to be the operation of inserting and removing drill pipe, then input the video frame into the target detection network to detect the position of drill pipe. Otherwise, input the next frame of the video into the image recognition neural network for inference and judgment. S106: Use an object detection network to detect the position and size of the drill rod detection box in the part of the video where the drill rod is being inserted and removed. Then, if the video ends, record the entire drill rod vertex curve. Otherwise, input the next frame of the video into the image recognition neural network for inference and judgment. S107: Record the change curve of the drill pipe vertex: Obtain the position and size of the drill pipe detection box based on the output of the target detection neural network, calculate the coordinates of the drill pipe vertex based on the coordinate origin and the position and size of the detection box, and record them; S108: Analyze the drill pipe tip variation curve obtained in step S107. First, perform Gaussian filtering on the drill pipe tip variation curve. Then, execute a window selection algorithm on the filtered curve to obtain the number of drill pipes pulled out and the number of drill pipes inserted. Multiply the maximum value of the two by the pipe length to obtain the source well depth.

2. The method for measuring the depth of seismic source wells based on deep learning according to claim 1, characterized in that: The method for creating the image discrimination dataset in step S103 is as follows: the video frames are divided into two categories according to whether construction is underway, and placed into two folders respectively; the method for creating the target detection dataset is as follows: the annotation tool is used to mark the drill rods in the drill rod insertion and removal video frames with annotation boxes, and then the annotated video frames are saved in YOLO format.

3. The method for measuring the depth of seismic source wells based on deep learning according to claim 2, characterized in that: Step S104 specifically involves: First, building an image recognition neural network, inputting the image recognition dataset into the image recognition neural network for training, and adjusting the parameters of the image recognition neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range; then, building an object detection neural network, inputting the object detection dataset into the object detection neural network for training, and adjusting the parameters of the object detection neural network to achieve a fit, i.e., the test set loss fluctuates within a certain range.

4. The method for measuring the depth of seismic source wells based on deep learning according to claim 3, characterized in that: The Gaussian filtering formula in step S108 is: in The filtered curve is shown below. The curve before filtering. The standard deviation of the Gaussian kernel. It is the mean of the curve.

5. The deep learning-based method for measuring the depth of seismic source wells in seismic exploration according to claim 4, characterized in that: The window selection algorithm in step S108 includes the following steps: S1081: Set window threshold and merging threshold ; S1082: Traverse the curve; if the curve value is greater than the window threshold... Then, starting from the beginning of the window, record the coordinate data of the window's starting point on the curve; S1083: Starting from the position recorded in step S1082, traverse the curve. When the curve value is less than the window threshold... Then the window ends, and the coordinates of the window's end point on the curve are recorded; S1084: If the end of the curve is not reached, S1082 and S1083 will be executed again starting from the end point recorded in S1083. If the curve ends, jump to S1085. S1085: Traverse all selected windows; if the distance between two windows is less than the merge threshold... The two windows will be merged, with the starting point of the left window used as the starting point of the merged window and the ending point of the right window used as the ending point of the merged window. S1086: Find the maximum value in all windows, and the coordinate data of the maximum value on the curve; S1087: Traverse all windows. If the maximum value is closer to the beginning of the window, count it as inserting the drill rod. If the maximum value is closer to the end of the window, count it as pulling out the drill rod.

Images