An artificial intelligence automatic detection method for the buried quality of seismic acquisition receiving device

Abstract

The present invention discloses an artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device, comprising the following steps: S1, taking field images of the seismic acquisition receiving device and screening out clear images; S2, using the labelimg annotation tool to feature annotate the clear images, classify the images into qualified and unqualified, and save them as xml tag files; S3, building a YOLOv3 neural network, setting network parameters, performing iterative training, outputting the model after each round of training, and selecting the optimal prediction model according to the evaluation index of the prediction accuracy; S4, using the optimal prediction model to classify and predict the images of the seismic acquisition receiving device to be predicted. The present invention can quickly and effectively carry out construction quality inspection, improve the effectiveness and timeliness of quality control, reduce quality control costs, improve construction efficiency, help quality inspectors to achieve rapid evaluation of instrument installation quality, and ensure the acquisition quality of seismic data.

Description

Technical Field

[0001] The invention belongs to the field of geophysical technology, and in particular relates to an artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device. Background Art

[0002] The acquisition of seismic signals is one of the most important links in the seismic exploration process. Only under the conditions of "the measurement point is guaranteed to be accurate, the excitation factors are consistent, the receiving link is consistent, and the instrument recording factors remain unchanged", can the received seismic reflection information objectively and truly reflect the situation of the underground medium. At present, the seismic acquisition methods such as "four small and one high" with small point spacing, small combined base spacing, small offset distance, small charge and high stacking times, as well as other "high density", "two wide and one high" (broadband, wide azimuth, high density) require the use of a large number of receivers such as geophones and node meters to collect high-quality seismic data, which puts forward higher requirements for the quality assurance of the receiving link. To this end, the on-site construction personnel need to complete the field installation of the receiving device and take photos and records. The quality inspection personnel will subsequently evaluate the installation quality of the receiving device based on the photos to determine whether the photos are clear, whether the instrument is installed correctly, whether the surrounding vegetation and debris are cleared, and whether the burial depth is appropriate. However, due to the large-scale promotion and application of "two wide and one high" seismic acquisition technology, especially high-precision seismic acquisition technology in mountainous areas, the density of gun channels is getting higher and higher, the distance between receiving points is getting smaller and smaller, the number of receiving devices is increasing, and the number of receiving arrangements of more than 10,000 channels has become the norm. Some projects have invested more than 50,000 receiving devices, and the number of photos taken is as high as hundreds of thousands. The huge number faces great challenges for manual quality inspection. Due to its strong subjectivity, low efficiency, long time consumption and high cost, it not only greatly increases labor costs, but also prolongs the quality inspection cycle, affecting the quality and benefits of seismic acquisition projects. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an artificial intelligence automatic detection method for the buried quality of seismic acquisition receiving devices that can quickly and effectively perform construction quality inspections, improve the effectiveness and timeliness of quality control, reduce quality control costs, improve construction efficiency, and help quality inspectors to quickly evaluate the installation quality of instruments.

[0004] The object of the present invention is achieved through the following technical scheme: an artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device, comprising the following steps:

[0005] S1, taking a field image of a seismic acquisition receiving device, and using the Laplace variance algorithm to calculate the variance of the image, determining a threshold value based on the variance, and screening out a clear image;

[0006] S2. Use the labelimg annotation tool to annotate the clear images, classify the images into qualified and unqualified, and save them as XML label files;

[0007] S3. Build the YOLOv3 neural network, set network parameters, perform iterative training, output the model after each round of training, and select the optimal prediction model according to the evaluation index of prediction accuracy;

[0008] S4. Use the optimal prediction model to classify and predict the images of the seismic acquisition receiving device to be predicted.

[0009] Furthermore, the specific implementation method of step S1 is:

[0010] S11, taking field images of the seismic acquisition receiving device and inputting them into a computer;

[0011] S12, perform convolution operation on the gray value P of the image and the Laplace operator H:

[0012]

[0013]

[0014] Where H is the Laplace operator, is a second-order differential operator; f(x,y) is a digital image, x and y are variables; P is the image grayscale matrix, and * is a convolution operation;

[0015] S13, take the matrix after the image gray value is convolved with the Laplacian matrix Calculate the variance s 2 :

[0016]

[0017] In the formula, For the matrix The value of the element in row i and column i in , For the matrix The mean of , n is the number of rows and columns of the matrix;

[0018] S14. Calculate the variance of all images in turn, select the minimum variance from the clear images as the threshold, and then add the images with variance greater than the threshold to the training set.

[0019] Furthermore, the specific implementation method of step S2 is: open the images in the training set in turn, mark the target area, and draw a rectangular area manually with a mouse around the center of the detector and the node instrument so that the drawn rectangular area can just enclose the detector and the node instrument; then according to the standard that more than half of the volume of the instrument must be buried underground, the surroundings must be clean and tidy, and not blocked by vegetation and debris, the images that meet the above standards are marked as qualified, otherwise they are marked as unqualified, and the image marks are saved as xml tag files to form a training set.

[0020] Furthermore, the specific implementation method of step S3 is:

[0021] S31. Collect instrument data set: the rectangular area drawn manually in the qualified image is the Right box; the rectangular area drawn manually in the unqualified image is the Wrong box; collect data x, y, w, h, C of the Right box and the Wrong box respectively; where (x, y) is the coordinate of the center point of the box, (w, h) is the height and width of the box, and C is the image category;

[0022] S32. Use K-means clustering method to cluster the width and height of the Right box and the Wrong box. Then calculate 9 anchor points with sizes of (10×13), (16×30), (33×23); (30×61), (62×45), (59×119); (116×90), (156×198), (373×326). Generate a detection box for every three anchor points, and get three detection boxes in total. Calculate the distance from the Right box and the Wrong box to the cluster center, then re-divide the detection box and update the value of the cluster center until the change in the cluster center converges, and the detection box is close to the position of the detected target.

[0023] S33, adding grayscale around the image, fixing the image size to 416×416 pixels to prevent image distortion; then dividing the image into three non-overlapping network units of sizes 13×13, 26×26, and 52×52;

[0024] S34, set YOLOv3 network parameters, process 4 samples in one round of iteration during training, fill the grayscale to 416×416 pixels before training, use BN for regularization at each weight update, set weight decay to 0.0005, momentum decay to 0.9, initial learning rate to 0.001, and number of iterations to 1000;

[0025] S35, input the image obtained in S33 into the backbone network Darknet-53 of YOLOv3 for feature extraction; the YOLO algorithm continuously performs convolution on the input image to extract features, and after multiple downsampling convolution operations, the final output results are feature maps of three different sizes: 52×52×256, 26×26×512, and 13×13×1024;

[0026] S36, evenly distribute the three detection frame sizes obtained by the K-means clustering algorithm to the feature maps of three different sizes, predict a certain amount of prior frames for the feature map of each size, and generate corresponding candidate frames, select the candidate frame with the largest overlap area as the final prediction frame, and then map the prediction frame to the original image according to the relationship between the feature map and the original image, locate the position of the target, and display the confidence probability of the detector and the node instrument on the prediction frame;

[0027] S37, output a prediction model every 20 iterations, and obtain 50 models in total;

[0028] The prediction accuracy evaluation index of the model is defined as:

[0029] Precision=Nr / N (4)

[0030] In the formula, Precision is the prediction accuracy of the model, Nr is the number of samples predicted correctly, and N is the total number of samples;

[0031] S38. Input the test set into each model in turn, and the one with the highest accuracy is the optimal prediction model.

[0032] The beneficial effects of the present invention are as follows: in view of the problems of cumbersome manual quality inspection of buried receiving devices in seismic exploration, heavy workload, low efficiency, and easy errors, the YOLOv3 target detection algorithm based on deep learning and computer vision technology is used, and clear images are screened with the help of the Laplace operator in edge detection. Field placement images of receiving devices such as detectors and node meters are used as training sets and prediction sets to establish an automatic detection method that can quickly evaluate the field installation quality of receiving devices, and the subjectivity, low efficiency, low precision and other unfavorable factors that affect efficient exploration in manual operations are gradually eliminated by utilizing the advantages of intelligence; the present invention can quickly and effectively carry out construction quality inspections, improve the effectiveness and timeliness of quality control, reduce quality control costs, improve construction efficiency, help quality inspectors to quickly evaluate the installation quality of instruments, and ensure the acquisition quality of seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the artificial intelligence automatic detection method of the buried quality of the seismic acquisition receiving device of the present invention;

[0034] Figure 2 It is a schematic diagram of an unclear image and a clear image;

[0035] Figure 3 It is a feature annotation diagram of a qualified device and an unqualified device, wherein (a) is a qualified device and (b) is an unqualified device;

[0036] Figure 4 Schematic diagram of grid division;

[0037] Figure 5 This is the YOLOv3 network structure diagram

[0038] Figure 6 It is a graph showing the change of accuracy with the number of iterations;

[0039] Figure 7 This is a graph of (partial) prediction results. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0041] like Figure 1 As shown, an artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device of the present invention comprises the following steps:

[0042] S1. Take a field image of the seismic acquisition receiving device, and use the Laplace variance algorithm to calculate the variance of the image, determine the threshold value according to the variance, and filter out a clear image; the specific implementation method is:

[0043] S11, taking field images of the seismic acquisition receiving device and inputting them into a computer;

[0044] S12, perform convolution operation on the gray value P of the image and the Laplace operator H:

[0045]

[0046]

[0047] Where H is the Laplace operator, is a second-order differential operator; f(x,y) is a digital image, x and y are variables; P is the image grayscale matrix, and * is a convolution operation;

[0048] S13, take the matrix after the image gray value is convolved with the Laplacian matrix Calculate the variance s 2 :

[0049]

[0050] In the formula, For the matrix The value of the element in row i and column i in , For the matrix The mean of , n is the number of rows and columns of the matrix;

[0051] S14, variance s 2 It can represent the blurriness of the image. If the variance is large, it means that the image is clear, the focus is accurate, and the image frequency response range is wide; if the variance is small, it means that the image is blurred, the number of edges is small, and the image frequency response range is narrow. Figure 2 This reflects the relationship between the variance size and the image clarity. Images with variances of 60 and 179 are not clear, while images with variances of 1385 and 1230 are very clear. The variances of all images are calculated in turn, the minimum variance is selected as the threshold from the clear images, and then the images with variances greater than the threshold are added to the training set. In this embodiment, a total of 500 images were collected. Through the above processing, 372 clear images were screened out, 272 of which were selected as the training set, accounting for 73%, and the remaining 100 images were selected as the test set, accounting for 23%.

[0052] S2. Use labelimg annotation tool to annotate the clear images, classify the images into qualified and unqualified, and save them as xml tag files. The specific implementation method is: open the images in the training set one by one, annotate the target area, and draw a rectangular area around the center of the detector and the node instrument manually with the mouse so that the drawn rectangular area can just enclose the detector and the node instrument, such as Figure 3 As shown; then, according to the standard that more than half of the volume of the instrument must be buried underground, the surroundings must be clean and tidy, and it must not be blocked by vegetation or debris, the images that meet the above standards are marked as qualified, otherwise they are marked as unqualified, and the image tags are saved as xml tag files to form a training set.

[0053] The image classification in the training set and the test set of this embodiment is shown in Tables 1 to 3.

[0054] Table 1

[0055] Dataset Proportion / (%) Quantity / sheet Training set 73 272 Test Set 27 100

[0056] Table 2

[0057] Training set Pass / piece Pass / piece Detector 132 60 Node Instrument 63 17

[0058] Table 3

[0059] Test Set Pass / piece Unqualified / Detector 50 23 Node Instrument 22 6

[0060] S3. Build the YOLOv3 neural network, set network parameters, perform iterative training, output the model after each round of training, and select the optimal prediction model according to the evaluation index of prediction accuracy; the specific implementation method is:

[0061] S31. Collect instrument data set: the rectangular area drawn manually in the qualified image is the Right box; the rectangular area drawn manually in the unqualified image is the Wrong box; collect data x, y, w, h, C of the Right box and the Wrong box respectively; where (x, y) is the coordinate of the center point of the box, (w, h) is the height and width of the box, and C is the image category;

[0062] S32. Use K-means clustering method to cluster the width and height of Right box and Wrong box. Then find the anchor box of appropriate size and calculate 9 anchor points with sizes of (10×13), (16×30), (33×23); (30×61), (62×45), (59×119); (116×90), (156×198), (373×326). Generate a detection box for every three anchor points, and get three detection boxes in total. Calculate the distance from Right box and Wrong box to the cluster center, then re-divide the detection box and update the value of the cluster center until the change of the cluster center converges, and the detection box is close to the position of the detected target.

[0063] S33, add grayscale around the image, fix the image size to 416×416 pixels to prevent image distortion; then divide the image into three non-overlapping network units of size 13×13, 26×26, and 52×52, respectively. Figure 4 As shown, each network unit is responsible for the detection of an area. As long as the center point of the object falls within the area, the object is determined by the grid point.

[0064] S34, set YOLOv3 network parameters, one iteration (Iteration) processes 4 samples during training, the image is filled with grayscale to 416×416 pixels before training, BN (Batch normalization) is used for regularization at each weight update, weight decay (Decay) is set to 0.0005, momentum decay (Momentum) is set to 0.9, the initial learning rate (Learning rate) is set to 0.001, and the number of iterations is 1000;

[0065] S35, input the image obtained in S33 into the YOLOv3 backbone network Darknet-53 for feature extraction; the YOLOv3 network structure is as follows Figure 5As shown in Figure 1. The YOLOv3 network first normalizes the training image to 416 pixels × 416 pixels, and then divides it into 13 pixels × 13 pixels grids. Each grid is responsible for detecting B bounding boxes and their confidences, as well as the probabilities of C categories. Darknet-53 (Residual Network-53) with 53 convolutional layers is used as the feature extraction network to extract the features of the receiving device, as shown in Figure 1. Figure 5 As shown in the figure. The YOLO algorithm continuously performs convolution on the input image to extract features. The height and width of the image will also be continuously compressed as the number of convolutions increases. This is manifested in the downsampling process where the image size is compressed from 416×416 to 13×13 and the number of channels is expanded from 3 to 1024. Three feature maps of different sizes, 52×52×256, 26×26×512, and 13×13×1024, are obtained, which can reflect the features of the input image. The obtained 13×13×1024 is divided into two directions. One direction is used for classification prediction and regression prediction. After two convolutions, it is decomposed into 13×13×3×25 images. Each image is 13×13 in size and contains 3 prior frames. The image tensor is 25 [20 (object category) + 1 (bounding box confidence) + 4 (prediction frame parameters))]. The other direction is used for upsampling, and the feature layer of 26×26×512 is stacked to construct a feature pyramid for multi-scale feature fusion to extract more effective features. The stacked result is convolved multiple times for classification prediction to obtain a 26×26×3×25 image. The convolution result is also upsampled and stacked with the 52×52×256 feature layer, and the same convolution operation is performed to obtain a 52×52×3×25 image.

[0066] S36, evenly distribute the three detection frame sizes obtained by the K-means clustering algorithm to the feature maps of three different sizes, predict a certain amount of prior frames for the feature map of each size, and generate corresponding candidate frames, select the candidate frame with the largest overlap area as the final prediction frame, and then map the prediction frame to the original image according to the relationship between the feature map and the original image, locate the position of the target, and display the confidence probability of the detector and the node instrument on the prediction frame;

[0067] S37, output a prediction model every 20 iterations, and obtain 50 models in total;

[0068] The prediction accuracy evaluation index of the model is defined as:

[0069] Precision=Nr / N (4)

[0070] In the formula, Precision is the prediction accuracy of the model, Nr is the number of samples predicted correctly, and N is the total number of samples;

[0071] S38. Input the test set into each model in turn. The model with the highest accuracy is the optimal prediction model, which is used as the basis for judging whether the installation of the detection instrument is qualified, that is, more than half of the volume of the instrument is buried underground, the surroundings are clean and tidy, and are not blocked by vegetation and debris.

[0072] Figure 6 It reflects the relationship between the accuracy Precision and the number of iterations Interations. As can be seen from the figure, when the number of iterations reaches 600, the accuracy has stabilized, and the accuracy between adjacent models is not much different. The maximum accuracy is 93%, corresponding to the 49th model, the number of iterations is 980, and it is regarded as the optimal model.

[0073] After training, the test set is input into the optimal model. Some prediction results are as follows: Figure 7 As shown, the top two rows are correctly placed pictures, the instrument is buried at a suitable depth and surrounded by soil; the bottom two rows are incorrectly placed pictures, and the vegetation and debris around the instrument have not been cleared. In this embodiment, all 100 images of the test set are input into the optimal model for classification and discrimination, 93 images are correctly identified, 4 images are incorrectly identified, and 3 images are missed, with an accuracy rate of 93%, which meets the error requirements.

[0074] This example selects 372 clear pictures from 500 manually taken pictures of seismic signal detection instruments placed in the field for the experiment. 272 ​​pictures are randomly selected as the training set, divided into "qualified" and "unqualified" according to the standard, and a label atlas of the seismic signal receiving device is created through manual annotation; the remaining 100 pictures are used as the test set. After 1000 iterations of training, the optimal prediction model is obtained with an accuracy rate of 93%. The model test results show that the YOLOv3 algorithm successfully realizes the judgment of whether the installation of the seismic signal receiving device is qualified or not, thereby ensuring the consistency of the quality control of the device installation and reducing the cost of manual quality inspection during oil exploration.

[0075] S4. Use the optimal prediction model to classify and predict the images of the seismic acquisition receiving device to be predicted.

[0076] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. An artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device, characterized in that: The following steps are involved: S1, taking a field image of a seismic acquisition receiving device, and using the Laplace variance algorithm to calculate the variance of the image, determining a threshold value based on the variance, and screening out a clear image; S2. Use the labelimg annotation tool to perform feature annotation on clear images, classify images into qualified and unqualified, and save them as xml label files; the specific implementation method is: open the images in the training set in turn, mark the target area, and draw a rectangular area around the center of the detector and the node instrument manually with the mouse so that the drawn rectangular area can just envelop the detector and the node instrument; then, according to the standard that more than half of the volume of the instrument must be buried underground, the surroundings must be clean and tidy, and not blocked by vegetation and debris, mark the images that meet the above standards as qualified, otherwise mark them as unqualified, save the image marks as xml label files, and form a training set; S3. Build the YOLOv3 neural network, set network parameters, perform iterative training, output the model after each round of training, and select the optimal prediction model according to the evaluation index of prediction accuracy; the specific implementation method is: S31. Collect instrument data set: the rectangular area drawn manually in the qualified image is the Right box; the rectangular area drawn manually in the unqualified image is the Wrong box; collect data x, y, w, h, C of the Right box and the Wrong box respectively; where (x, y) is the coordinate of the center point of the box, (w, h) is the height and width of the box, and C is the image category; S32. Use K-means clustering method to perform cluster analysis on the width and height of the Right box and the Wrong box; then calculate 9 anchor points with sizes of (10×13), (16×30), and (33×23); (30×61), (62×45), (59×119); (116×90), (156×198), (373×326), generate a detection frame for every three anchor points, and get three detection frames in total; calculate the distance from the Right frame and the Wrong frame to the cluster center, then re-divide the detection frame and update the value of the cluster center until the change in the cluster center converges, and the detection frame is close to the position of the detected target; S33, adding grayscale around the image, fixing the image size to 416×416 pixels to prevent image distortion; then dividing the image into three non-overlapping network units of sizes 13×13, 26×26, and 52×52; S34, set YOLOv3 network parameters, process 4 samples in one round of iteration during training, fill the grayscale to 416×416 pixels before training, use BN for regularization at each weight update, set weight decay to 0.0005, momentum decay to 0.9, initial learning rate to 0.001, and number of iterations to 1000; S35, input the image obtained in S33 into the backbone network of YOLOv3 for feature extraction; the YOLO algorithm continuously performs convolution on the input image to extract features, and after multiple downsampling convolution operations, the final output results are feature maps of three different sizes: 52×52×256, 26×26×512, and 13×13×1024; S36, evenly distribute the three detection frame sizes obtained by the K-means clustering algorithm to the feature maps of three different sizes, predict a certain amount of prior frames for the feature map of each size, and generate corresponding candidate frames, select the candidate frame with the largest overlap area as the final prediction frame, and then map the prediction frame to the original image according to the relationship between the feature map and the original image, locate the position of the target, and display the confidence probability of the detector and the node instrument on the prediction frame; S37, output a prediction model every 20 iterations, and obtain 50 models in total; The prediction accuracy evaluation index of the model is defined as: Precision=Nr / N (4) In the formula, Precision is the prediction accuracy of the model, Nr is the number of samples predicted correctly, and N is the total number of samples; S38, input the test set into each model in turn, and the one with the highest accuracy is the optimal prediction model; S4. Use the optimal prediction model to classify and predict the images of the seismic acquisition receiving device to be predicted.

2. The artificial intelligence automatic detection method for the buried quality of a seismic acquisition receiving device according to claim 1 is characterized in that: The specific implementation method of step S1 is: S11, taking field images of the seismic acquisition receiving device and inputting them into a computer; S12, perform convolution operation on the gray value P of the image and the Laplace operator H: Where H is the Laplace operator, is a second-order differential operator; f(x,y) is a digital image, x and y are variables; P is the image grayscale matrix, and * is a convolution operation; S13, take the matrix after the image gray value is convolved with the Laplacian matrix Calculate the variance s 2 : In the formula, For the matrix The value of the element in row i and column i in , For the matrix The mean of , n is the number of rows and columns of the matrix; S14. Calculate the variance of all images in turn, select the minimum variance from the clear images as the threshold, and then add the images with variance greater than the threshold to the training set.

Images