Method for vectorizing seismogram paper based on deep meta learning

Abstract

The application discloses a simulation seismogram paper vectorization method based on deep meta learning, which can automatically and quickly determine the size of a scanned paper by using a multi-scale image segmentation technology, and automatically extract various recording parameters in the paper; the best size normalization, type classification, gray scale and image segmentation algorithm are designed for the multi-scale paper, so as to provide high-quality input data for a subsequent vectorization model and solve the single problem of the original vectorization algorithm; a time marker point template matching algorithm and an automatic picking algorithm are used to automatically complete the detection and calibration of the significant time marker points in the paper, so that the human consumption in the process of constructing a time coordinate system is effectively reduced, and the vectorization efficiency of the paper is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, specifically a method for vectorizing model seismic maps based on deep meta-learning. Background Technology

[0002] Analog seismograph waveform records are a type of paper-based seismic monitoring record obtained using early seismic monitoring instruments. They record seismic fluctuations in three directions (east-west, north-south, and up-down) over a period of time (e.g., one day) on a single sheet of paper. Due to the limited length of the paper, the fluctuation information needs to be printed line by line. Manual marking or filling in information such as station information, seismograph model, recording time, clock information, velocities in the three directions, and recorder information is also required. Analog seismograph waveform records are generally preserved in paper form. Currently, wide-format scanners are typically used for scanning, creating a raster format for permanent physical preservation to overcome the effects of temperature, humidity, and mechanical damage under natural conditions.

[0003] In practical applications, vectorized digital model seismic data is required. Currently, three key issues remain to be resolved regarding the transition between scanned drawings and digital data:

[0004] 1) The waveform data of the simulation seismic test has problems such as multiple models, multiple waveforms, multiple scales, and multiple distortions. The traditional single vectorization algorithm cannot demonstrate universal solution capabilities in waveform image segmentation, time coordinate construction and waveform curve sampling. Different algorithms must be designed for different data.

[0005] 2) Effectively remove noisy samples and redundant computational features that reduce the model's learning ability, and construct a vectorized model with low time complexity and low computational cost;

[0006] 3) The overall sample size of the simulated seismic waveform records is large, but the sample size of specific types is small. It is necessary to study how the vectorized model can demonstrate generalization ability on large samples and high accuracy on small samples.

[0007] Meanwhile, considering the large variety of seismographs, and the significant differences in seismic records obtained from different types of seismographs, and even the same type of recorder installed at different stations and at different times, higher requirements are placed on the generalization and robustness of vectorization of model seismograph waveform records. Summary of the Invention

[0008] Based on the above problems, and addressing the issues encountered in the vectorization of simulated seismic waveforms, such as multiple models, multiple waveforms, multiple scales, multiple distortions, and small samples of a single type, this invention provides a method for vectorizing simulated seismic drawings based on deep meta-learning.

[0009] The technical solution to achieve the objective of this invention is:

[0010] A method for vectorizing simulation seismic maps based on deep meta-learning includes:

[0011] (1) Rasterization and image segmentation of seismic waveform records: The paper seismic record paper is rasterized and the obtained raster image is segmented to complete the construction of the image data required for the vectorization model;

[0012] (2) Constructing a time coordinate system in the simulation seismic waveform record: The waveform in the simulation seismic record is calibrated with time markers, and the waveform time coordinate system is accurately modeled. At the same time, the time marker template matching algorithm and automatic picking algorithm are used to shield the algorithm differences caused by various waveforms and machine types.

[0013] (3) Vectorize the entire drawing of the model seismic waveform: Vectorize the seismic waveform in the drawing and accurately splice it with the non-seismic waveform to form a continuous waveform time series. At the same time, carry out earthquake arrival time calculation, earthquake beach ball drawing, geophysical background field analysis, detect and locate seismic waveforms, and remove various non-seismic image noise, smooth waves, interpunct points and other interference information that appear in the seismic waveforms.

[0014] (4) Establish a small sample library of seismic drawings based on noise, distortion, seismic wave type and seismograph instrument type. Using meta-learning technology, establish a waveform correction and vectorization sampling algorithm framework including various deep learning methods. Construct a robust deep meta-learning vectorization model for multi-scale, multi-type and small sample seismic drawings. Correct the distortion and tilt in the waveform record so that the points in the waveform correspond one-to-one with the established time coordinates. Achieve the goal of robust and accurate vectorization of seismic waveforms. At the same time, standardize the above vectorization results according to the international seismic data unified format to achieve consistency with digital seismic waveform records.

[0015] Step (1) involves the rasterization and image segmentation of the simulated seismic waveform record. This is a preprocessing algorithm study conducted on simulated seismic drawings of different types, multiple scales, and non-seismic noise after scanning the paper waveform record. The algorithm includes: automatically and quickly determining the size of the scanned drawing, automatically extracting various recording parameters from the drawing, and designing the best size normalization, type classification, grayscale conversion, and image segmentation (binarization) algorithms for multi-scale drawings to provide high-quality input data for subsequent vectorization models.

[0016] The waveforms in the seismometry record described in step (2) consist of smooth waves, seismic waves, and start-stop waves.

[0017] The method for constructing the time coordinate system in the seismic waveform record described in step (2) is as follows: The starting point coordinates d0(x0, y0) and the time marker feature map are manually picked up, where d0 is the starting point, and x0 and y0 are the abscissa and ordinate of the starting point d0, respectively. A Faster R-CNN deep network is used to detect and locate the time markers, and the position d of each time marker in the drawing is calculated. i (x i ,y i (i = 1, 2, 3....) (where d i x is a time marker. i y i (The x and y coordinates of the time markers);

[0018] Next, determine the pixel distance Δd between two adjacent time markers. i The calculation formula is as follows:

[0019] Δd i =d i (x i ,y i )-d i-1 (x i-1 ,y i-1 (1)

[0020] Calculate Δd i Overall waveform length d along the horizontal axis total The ratio θ between them is given by the following formula:

[0021]

[0022] The time length t of the drawing record is obtained by subtracting the clock times from the seismic records. total This allows us to obtain the duration between the two time markers. The calculation formula is as follows:

[0023] Δt i =t total ×θ i (3)

[0024] Then, by combining the detected time marker location information, the time t for all sampling locations in the drawing is calculated. x Then, the construction and deduction of the time coordinate system in the seismic drawings were accurately completed:

[0025]

[0026] The seismic waveforms described in step (3) include S-waves, P-waves, L-waves, and background vibrations.

[0027] The detection of seismic waveforms in step (3) involves using a multi-scale segmentation algorithm to segment the seismic background and seismic window, cutting the training samples according to the scale, and then using the focal-loss function to optimize the model loss calculation process. Finally, a dynamic adjustment strategy for the weights of difficult and easy detection targets is used to improve the problem of missed target detection.

[0028] The localization of seismic waveforms in step (3) involves embedding a multi-scale segmentation algorithm as an operator into the MAML meta-learning framework, dividing waveform samples according to seismic instrument model, drawing size, noise type, etc.; sampling on the sample space of seismic waveform drawings of various types, performing meta-learning on the operator model, improving the robustness of the operator, and obtaining the preliminary generalization parameters of the operator model; for a specific type of seismic drawing sample, starting from obtaining the generalization parameters through meta-learning, continuing to refine the training of the operator to obtain a seismic waveform window detection model with high accuracy.

[0029] Step (3) involves removing smooth waves and time points within the seismic waveform window. This includes selecting the seismic waveform window, traversing the starting point of the waveform curve, scanning and tracking within the window, and extracting and removing smooth waves. The removal method includes:

[0030] 1) Select a region with a vibrating waveform, then detect all grayscale jump points on the straight line to obtain n pairs of coordinates of jump points on the drawing: {[(x0,y0),(x0,y0')],[(x1,y1),(x1,y1')],…,[(x n ,y n ),(x n ,y n If ')]}, then the coordinates of the starting point of the i-th wave are Waveform width h i The calculation is as follows:

[0031]

[0032] 2) Arrange all curves in the window in reverse order according to the y-coordinate of the detected starting point, and then obtain the starting point detection.

[0033] 3) Eliminate smoothing waves. Taking curve i as an example, the overall search direction is horizontally to the right, with a given horizontal search step size step. x Vertical search step size y ;

[0034] Then its next point x in the horizontal direction i+1 The search iteration formula is as follows:

[0035] x i+1 =x i +step x(i = 0, 1, 2, ...) (6)

[0036] For vertical searches, using steps... y The step size is denoted by C, with the center of the waveform curve as the axis point. i Perform an up-and-down search (if the boundary is exceeded, switch to a complex wave search) to obtain the upper bound y. iH and the lower realm y iL , to obtain the jump width:

[0037] h i =y iH -y iL (7)

[0038] When the jump width is greater than the line width h i If no longitudinal jump occurs throughout the entire horizontal path, the search proceeds to a complex wave search. i If so, it will be removed as a smooth wave;

[0039] Meanwhile, within the seismic window, template matching is used to remove the noise from the extracted time stamps.

[0040] Step (4) involves constructing a robust deep meta-learning vectorized model for the seismic mapping drawing, including meta-learning-based vectorization of the seismic waveform, coordinate axis correction of the seismic waveform, curvature correction of the seismic waveform, time axis interpolation, and construction of the waveform time series. After sampling the upper and lower boundary points of the seismic waveform's crankshaft, the coordinates of the center pixel of the waveform curve are calculated to obtain the final sampled seismic waveform curve. The algorithm flow includes:

[0041] 1) First, detect the starting point and line width of the oscillating waveform curve, and dynamically adjust the horizontal traversal step size based on the line width. x Vertical traversal step size y ;

[0042] 2) Next, by sampling the upper and lower boundaries of the oscillating waveform curve from left to right and from bottom to top, the pixel transition point pairs are determined. The formula for calculating the center point C0(x0,y0) of the pixel at the starting point of the curve is as follows:

[0043]

[0044] in The upper and lower bounds of the ordinate of the starting point of the curve;

[0045] 3) Using step size x Perform a horizontal search. If the sampling point is not black, proceed to step d; if it is black, find the center coordinates C of the next sampling point on the curve. i (x i ,y iThe calculation is as follows: the forward step length of the horizontal axis is step. x With the ordinate unchanged, we obtain point C'. i (x i +step x ,y i The search is performed in both the vertical and horizontal directions around the point to detect black pixels (i.e., color transition points), which are then used as the upper bound. lower bound Then the new coordinates C' are calculated. i (x i ,y i The formula is as follows:

[0046]

[0047] 4) Using step size y Perform a vertical search. If the sampling point is not black, proceed to step c. If it is a black point, find the center coordinates C of the next sampling point on the curve. i (x i ,y i The calculation is as follows: the forward step length on the vertical axis is step. y With the x-coordinate unchanged, we obtain point C'. i (x i ,y i +step y The search is performed in both the left and right directions of the point to detect black pixels (i.e., color transition points), which are then used as the left boundary points. Right boundary point Then the new coordinates C' are calculated. i (x i ,y i The formula is as follows:

[0048]

[0049] 5) Record the above traversal points and determine whether the waveform curve traversal is complete. If not, search in the horizontal or vertical direction of the current point and proceed to step 3) or 4) respectively, until the waveform traversal of the entire window area is completed.

[0050] 6) Finally, visualize the data obtained from the above traversal, and submit it for manual judgment and correction to ensure that the vectorization result of the seismic waveform is accurate and reliable, and finally form the traversal result of the seismic waveform.

[0051] The coordinate axis correction and warp correction of the vibrating waveform include:

[0052] By calibrating the coordinate axes, the coordinate axes of the seismometry record are aligned with the coordinate axes on the drawing, thus obtaining a coordinate (X) perpendicular to the horizontal axis. rig ,Yrig The record and correction formula are as follows:

[0053]

[0054] Where R is the length of the rocker arm, and the angle β between the seismic waveform coordinate axis and the coordinate axis of the model seismic record drawing is obtained by mathematical calculation using the three-point method;

[0055] The above corrections can effectively resolve the issue of non-one-to-one correspondence between the X-axis and Y-axis coordinates of points on the waveform curve caused by the arc-shaped distortion of the curve due to mechanical factors.

[0056] The beneficial effects of this invention are as follows: By using a deep meta-learning algorithm, it is the first time that a single algorithm architecture can effectively locate and detect seismic waveforms under conditions of multiple machine models, multiple scales, and complex noise. It exhibits good robustness and generalization ability across various drawing types when faced with a variety of complex drawing data. By using multi-scale image segmentation technology, the existing architecture can automatically and quickly determine the size of the scanned drawing and automatically extract various recorded parameters from the drawing. Optimal size normalization, type classification, grayscale conversion, and image segmentation (binarization) algorithms are designed for multi-scale drawings, thus providing high-quality input data for subsequent vectorization models and solving the problem of a single target drawing in the original vectorization algorithm. Furthermore, by employing a time marker template matching algorithm and an automatic picking algorithm, significant time markers in the drawing are automatically detected and calibrated, effectively reducing manual labor in the time coordinate system construction process. This approach aims to improve the efficiency of drawing vectorization. The seismic waveform detection algorithm, built upon deep meta-learning, automatically detects complex waves requiring special processing within the drawing and cleans the corresponding windows, providing a favorable working environment for subsequent true waveform vectorization. Finally, the seismic waveform extraction algorithm, based on deep meta-learning, vectorizes and corrects the processed waveforms within the selected true waveform regions. After stitching together the seismic waveforms from the entire drawing and establishing a mapping relationship with the time coordinate system, digitized and usable simulated seismic waveform data is obtained. The visualized image of the extracted digitized point set is compared with the target image used for extraction. Clearly, the existing algorithm framework can effectively vectorize seismic waveforms in almost every waveform extraction process. Attached Figure Description

[0057] Figure 1 This is a technical implementation roadmap for an embodiment of the present invention;

[0058] Figure 2 This is a flowchart of multi-scale modal seismometry drawing segmentation according to an embodiment of the present invention;

[0059] Figure 3 This is a flowchart of the time coordinate system construction method according to an embodiment of the present invention;

[0060] Figure 4 This is a technical roadmap for time marker location and detection in an embodiment of the present invention;

[0061] Figure 5 This is an explanatory table for the simulated seismic waveform recording drawings used in the application of this invention.

[0062] Figure 6 This is a structural diagram of the shock wave extraction framework according to an embodiment of the present invention;

[0063] Figure 7 This is a flowchart illustrating the correction process for the extracted waveform according to an embodiment of the present invention;

[0064] Figure 8 This is a diagram illustrating the effect of waveform curve starting point search in an embodiment of the present invention.

[0065] Figure 9 This is a technical roadmap of the shock wave type vectorization algorithm in the embodiments of the present invention;

[0066] Figure 10 This is a technical roadmap of the learning framework for waveform edge detection in this invention.

[0067] Figure 11 This is a flowchart for extracting the center of the vibration waveform curve in an embodiment of the present invention;

[0068] Figure 12 This is a simplified geometric diagram illustrating the coordinate axis offset in an embodiment of the present invention;

[0069] Figure 13 This is a corrected simulation seismometry waveform record in an embodiment of the present invention;

[0070] Figure 14 This is a schematic diagram of waveform curve splicing in an embodiment of the present invention. Detailed Implementation

[0071] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0072] Example:

[0073] A deep learning method for vectorization of model seismic maps based on meta-learning algorithms, with its technical implementation roadmap as follows: Figure 1 As shown, the specific steps include the following:

[0074] 1. Image segmentation of simulated seismic waveform recording

[0075] Since seismic waveform records vary in type and size, the image segmentation algorithm should be highly adaptable, meaning it should be able to adapt to input and segmentation of variable-sized scanned raster images. This patent adds a size-adaptive algorithm to the image input module to meet the needs of inputting images of different sizes.

[0076] Simultaneously, a Pyramid Scene Parsing Network is employed for image segmentation. The core of this network is Global Pyramid Pooling, which scales feature maps to several different sizes, resulting in better global and multi-scale information. This is highly suitable for seismic image segmentation with multi-scale characteristics and contributes to a rapid improvement in accuracy. A schematic diagram of the designed Pyramid Scene Parsing Network for image segmentation is shown below. Figure 2 As shown

[0077] Specifically: First, a hierarchical global prior information is designed to reflect the different sizes of seismic maps and their variations in different sub-regions. This prior information can eliminate the constraints of fixed sizes in CNNs. To further reduce the loss of contextual information between different waveform sub-regions, Pyramid pooling is used to fuse features at different levels. First, coarse global-scale pooling features are fused, followed by the fusion of representation layers representing information from different sub-regions. Then, bilinear interpolation is directly performed on all feature layers to restore the length and width of the input. Finally, the features from different layers are concatenated as global features for pyramid pooling, and the image segmentation threshold is calculated.

[0078] Using the aforementioned deep learning algorithm to predict image segmentation thresholds, binarization can be calculated for multi-scale modal seismic waveform images by region. Without overcompensation, the waveform portion (with continuous black lines) can be accurately distinguished from other background and noise, thus achieving the requirement for accurate segmentation of the target image in preprocessing.

[0079] 2. Derivation of the time coordinate system for simulated seismic waveform recording

[0080] To accurately assign time to each waveform sampling point, the starting point must first be manually collected and recorded. Then, deep learning methods are used to automatically detect some time marker points. Finally, a time coordinate system derivation model is constructed to complete the assignment of values ​​to the seismic waveform time series. The process of constructing the time coordinate system and waveform time series is as follows: Figure 3 As shown.

[0081] Specifically, the starting point coordinates d0(x0, y0) and time marker feature maps are manually collected, where d0 is the starting point and x0 and y0 are the x and y coordinates of the starting point d0, respectively. A mature Faster R-CNN deep network is then used to detect and locate the time markers, and the position d of each time marker in the drawing is calculated. i (x i ,y i (i = 1, 2, 3....) (where d i x is a time marker. i y i (These represent the x and y coordinates of the time-marked points). The Faster R-CNN network mainly consists of three deep neural networks: Conv-layers Networks, Region Proposal Networks, and Classification and Regression Networks. The Conv-layers Networks are used for learning and extracting features of the time-marked points, the Region Proposal Networks are used for selecting candidate regions, and the Classification and Regression Networks are used for classifying and locating the coordinates of the candidate regions (precise selection). The time-marked point localization and detection model built based on Faster R-CNN is as follows: Figure 4 As shown.

[0082] Next, determine the pixel distance Δd between two adjacent time markers. i The calculation formula is as follows:

[0083] Δd i =d i (x i ,y i )-d i-1 (x i-1 ,y i-1 (1)

[0084] Calculate Δd i Overall waveform length d along the horizontal axis total The ratio θ between them is given by the following formula:

[0085]

[0086] According to the seismic record description table (such as...) Figure 5 The difference between the clock times shown in the diagram is used to determine the time length t recorded on the drawing. total This allows us to obtain the duration between the two time markers. The calculation formula is as follows:

[0087]

[0088] Then, by combining the detected time marker location information, the time t for all sampling locations in the drawing is calculated. x Then, the construction and deduction of the time coordinate system in the seismic drawings were completed accurately.

[0089]

[0090] 3. Selection and purification of waveform windows (areas)

[0091] Accurately detect and locate the seismic waveform window (region), and clean up the smooth waves, time points, etc. within the seismic waveform window, thereby providing high-quality training samples for subsequent high-precision and robust identification of seismic waveforms. This mainly includes the following three parts:

[0092] (1) Detection and localization of waveform windows

[0093] Specifically, this invention proposes an improved method for seismic window detection: ① It employs a multi-scale segmentation algorithm to segment the seismic background and the seismic window, and cuts the training samples according to the scale bar; ② Then, it uses a focal-loss function to optimize the model loss calculation process; ③ Finally, it adopts a dynamic adjustment strategy for the weights of difficult and easy detection targets to improve the problem of missed target detection. Experimental results show that applying this improved EAST algorithm to seismic window detection improves the accuracy by 5.1% and reduces the time complexity by more than 20% compared to the original algorithm.

[0094] Next, the above algorithm is embedded as an operator into the MAML meta-learning framework. Specifically: ① Waveform samples are divided according to seismic instrument model, drawing size, noise type, etc.; ② Sampling is performed on the multi-type seismic waveform drawing sample space to conduct meta-learning on the operator model to improve the robustness of the operator and obtain the generalization (preliminary) parameters of the operator model; ③ For a specific type of seismic drawing sample, starting from the generalization parameters obtained by meta-learning, the operator is further refined and trained to finally obtain a seismic waveform window detection model with high accuracy. The entire seismic waveform window detection and localization algorithm meta-learning framework is as follows: Figure 7 As shown.

[0095] (2) Remove smooth waves and time stamps within the waveform window.

[0096] There are usually a certain amount of interference information such as smooth waves and time stamps in the waveform window, which can be used as noise to remove noise.

[0097] The smooth wave removal process includes: selecting a waveform window; traversing the starting point of the waveform curve; scanning and tracking the smooth waves within the window; and extracting and removing the smooth waves, as illustrated below. Figure 7 As shown.

[0098] Specifically, first, select a region with a vibrating waveform, then detect all grayscale jump points on the straight lines to obtain n pairs of coordinates of the jump points on the drawing: {[(x0,y0),(x0,y0')],[(x1,y1),(x1,y1')],…,[(x n ,y n ),(x n ,y n If ')]}, then the coordinates of the starting point of the i-th wave are Waveform width h i The calculation is as follows:

[0099]

[0100] Then, all curves within the window are sorted in reverse order according to the y-coordinate of the detected starting point, using Shell sort, to obtain the starting point detection result, as shown below. Figure 8 As shown.

[0101] Next, smoothing is performed. Taking curve i as an example, the search direction is generally horizontal to the right, with a given horizontal search step size step. x Vertical search step size y .

[0102] Then its next point x in the horizontal direction i+1 The search iteration formula is as follows:

[0103] x i+1 =x i +step x (i = 0, 1, 2, ...) (6)

[0104] For vertical searches, using steps... y The step size is denoted by C, with the center of the waveform curve as the axis point. i Perform an up-and-down search (if the boundary is exceeded, switch to a complex wave search) to obtain the upper bound y. iH and the lower realm y iL , to obtain the jump width:

[0105] h i =y iH -y iL (7)

[0106] When the jump width is greater than the line width h i If no longitudinal jump occurs throughout the entire horizontal path, the search proceeds to a complex wave search. iIf so, it is treated as a smooth wave and removed.

[0107] Meanwhile, within the seismic window, template matching is used to remove the noise from the extracted time stamps.

[0108] 4. A robust vectorization algorithm for seismic waveforms based on meta-learning

[0109] Here, addressing the key scientific problem of "accurate vector sampling of seismic waveforms from small samples of drawings across multiple aircraft models, scales, types, and single-type drawings," the designed technical process includes: seismic waveform vectorization based on meta-learning; coordinate axis correction of the seismic waveform; warping correction of the seismic waveform; time axis interpolation processing; and construction of the waveform time series. A schematic diagram of the technical roadmap is shown below. Figure 9 .

[0110] I. Vibration waveform curve completion and tracking sampling.

[0111] After removing smooth waves and eliminating noise from the seismic waveform window, the following algorithms are used for seismic waveform curve completion and seismic waveform curve boundary tracking sampling:

[0112] (1) Complete the waveform curve with vibration:

[0113] In seismic simulation drawings, smooth waves and seismic waves may intersect. When smooth waves or time markers are removed, seismic waves may be partially erased, resulting in waveform breaks. This invention designs an interactive and user-friendly completion program to manually complete the breaks in the purified seismic waveform, thereby forming a continuous seismic waveform curve, which facilitates the accurate extraction of seismic waves in the next step.

[0114] (2) Sampling of boundary tracking of seismic waveform curves:

[0115] To accurately extract seismic waves, it is necessary to obtain the complete boundary of the seismic wave curve, i.e., the set of upper and lower boundary points of gray-level transitions on the curve. The meta-learning framework will adjust the hyperparameters in the operator model used for upper and lower edge detection of the curve to adapt it to edge detection of seismic waveforms of various types, sizes, and small samples, thereby obtaining more accurate and continuous sampling of the upper and lower boundaries of the waveform curve and achieving robust vectorization.

[0116] The general process is as follows: ① First, the drawing type must be labeled according to the seismic wave type (S-wave, P-wave, L-wave, and background vibration), seismograph model, size, distortion form, etc., to establish a relatively large sample library of different types of waveforms (current data 1.5T); ② Then, using the mature MAML meta-learning framework as an optimizer, a deep learning operator hyperparameter training platform for curve edge detection is built; ③ Next, using the trained hyperparameter deep learning operator model, for a small sample library of specific types of waveforms, the internal parameters of the model are further trained and finely adjusted to predict the upper and lower edges of the generated waveform curves. The entire meta-learning algorithm framework for edge detection is as follows: Figure 11 .

[0117] Specifically, the Holistically-Nested Edge Detection (HED) algorithm is adopted as the deep learning model for double edge detection of waveform curves. Its advantages are: ① Inspired by fully convolutional neural networks, it can take the entire seismic waveform as input and directly generate a double edge map as output; ② It employs nested multi-scale feature learning, which aligns with the multi-scale features of the established seismic waveform sample library, helping to overcome the influence of different scales in the simulated seismic maps to some extent; ③ Simultaneously, inspired by deep supervision networks, it performs deep supervision to "guide" early classification results, further reducing gradient loss in the fine iteration of waveform edge detection. In the preliminary research phase, it was found that utilizing these features of HED to extract the double edges of seismic waveforms exhibits high accuracy and computational efficiency.

[0118] (3) Determining the center of the seismic waveform curve:

[0119] After sampling the upper and lower boundary points of the vibrating waveform curve, the coordinates of the center pixel of the waveform curve should be calculated to obtain the final vibrating waveform curve sample, i.e., the precise direction and trend of the curve. The algorithm flow is as follows: Figure 11 As shown.

[0120] First, detect the starting point and line width of the oscillating waveform curve, and then dynamically adjust the horizontal traversal step size based on the line width. x Vertical traversal step size y .

[0121] Next, by sampling the upper and lower boundaries of the oscillating waveform curve from left to right and from bottom to top, the pixel transition point pairs are determined. The formula for calculating the center point C0(x0,y0) of the pixel at the starting point of the curve is as follows:

[0122]

[0123] in The upper and lower bounds of the ordinates of the curve's starting point are given.

[0124] c uses step size x Perform a horizontal search. If the sampling point is not black, proceed to step d; if it is black, find the center coordinates C of the next sampling point on the curve. i (x i ,y i The calculation is as follows: the forward step length of the horizontal axis is step. x With the ordinate unchanged, we obtain point C'. i (x i +step x ,y i The search is performed in both the vertical and horizontal directions around the point to detect black pixels (i.e., color transition points), which are then used as the upper bound. lower bound Then the new coordinates C' are calculated. i (x i ,y i The formula is as follows:

[0125]

[0126] d with step size step y Perform a vertical search. If the sampling point is not black, proceed to step c. If it is a black point, find the center coordinates C of the next sampling point on the curve. i (x i ,y i The calculation is as follows: the forward step length on the vertical axis is step. y With the x-coordinate unchanged, we obtain point C'. i (x i ,y i +step y The search is performed in both the left and right directions of the point to detect black pixels (i.e., color transition points), which are then used as the left boundary points. Right boundary point Then the new coordinates C' are calculated. i (x i ,y i The formula is as follows:

[0127]

[0128] e records the above traversal points and determines whether the waveform curve traversal is complete. If not, it searches in the horizontal or vertical direction of the current point and proceeds to step c or d respectively, until the waveform traversal of the entire window area is completed.

[0129] Finally, the data obtained from the above traversal is visualized and submitted for manual judgment and correction to ensure that the vectorized results of the seismic waveform are accurate and reliable, thus forming the traversal results of the seismic waveform.

[0130] 2. Perform coordinate axis and curvature correction on the waveform after vector sampling of the model seismic record.

[0131] Correction of seismic waveforms is another key aspect of this embodiment, directly affecting the effectiveness of the final vectorized data results in subsequent practical applications such as earthquake arrival time calculation and earthquake beach ball correction.

[0132] The rocker arm seismograph generates the rocker arm swing amplitude in proportion to the magnitude of the earthquake, which drives the recording pen on the rocker arm to record the earthquake waveform on the drawing. In the process, it may produce an arc rather than a deflection perpendicular to the time axis.

[0133] A simplified geometric representation of the coordinate axis offset in this embodiment is shown below. Figure 12 As shown. The rocker arm recording pen and the time axis of the drawing have an angle of β. In order to record the coordinates of the distorted point (X... dig ,Y dig Converted to a series of undistorted point coordinates (X) cor ,Y cor This patent employs the following process:

[0134] First, coordinate axis calibration is performed to ensure that the coordinate axes of the seismometry record fall on the coordinate axes of the drawing (this position is usually not exactly on the X-axis), thus obtaining a coordinate perpendicular to the horizontal axis (X). rig ,Y rig Record. The correction formula is as follows:

[0135]

[0136] Where R is the length of the rocker arm. The angle β between the seismic waveform coordinate axis and the coordinate axis of the model seismic record drawing is obtained mathematically using the three-point method.

[0137] The above corrections can effectively resolve the issue of non-one-to-one correspondence between the X-axis and Y-axis coordinates of points on the waveform curve caused by the arc-shaped distortion of the curve due to mechanical factors.

[0138] However, the following problems remain: the curve traces have a higher density at bends, while research on the seismograph principle of mechanical seismographs indicates that, due to the uniform movement of the paper, the model seismograph records should also be evenly distributed across the entire time axis within the window. This contradicts the theory, so further adjustments are needed to the above correction results to ensure a uniform distribution of each recording point across the time axis. This patent uses time axis interpolation to process the data. Specifically, it applies first-order linear interpolation to the entire selected window based on the number of collected coordinate points, ensuring a uniform distribution of the collected points across the entire selected area's time length. Then, the adjusted coordinate points are connected. After applying this method, the corrected result is as follows: Figure 13 As shown.

[0139] Finally, the final seismic waveform sequence, after multiple corrections, is attached to the previously obtained time coordinate system, that is, the waveform amplitude is integrated with the time coordinate system, thereby completing the construction of the time series of the entire drawing. Combined with the previously obtained title information, the vectorization sampling of the model seismic record drawing is completed.

[0140] 5. Standardization, inversion, and application of simulated seismic waveforms

[0141] After the seismic waveform is vectorized, the sampled vector points of the seismic waveform curve are stored in the database in the form of multiple two-dimensional tables. Each discrete point set (two-dimensional table) represents a curve in the drawing, and these curves need to be stitched together. Figure 14 A schematic diagram of the waveform curve splicing algorithm is given, that is, the tail of each curve is connected to the head of the curve immediately above it, and so on.

[0142] Next, in accordance with the internationally standardized seismic data format (SAC), further sampling, interpolation, and format conversion were performed, and information on the seismic instrument and waveform acquisition function was supplemented to standardize the data, which was then published in SAC format.

[0143] Then, a waveform inversion (display) SAC interface program was written to quantitatively invert seismic (P-wave, S-wave, L-wave, background vibration) waveforms by time (accurate to the minute) and direction, realizing "intelligent reading" of historical seismic data.

[0144] Finally, based on the vectorized seismic waveform data in SAC format, the arrival times of past major earthquakes were recalculated and earthquake beach ball mapping was performed at seismic stations in Chengde, Tangshan, Urumqi, and Chengdu, and the earthquake catalog of past major earthquakes was corrected. Existing digital seismic data were further integrated to provide computer-readable geophysical field background data over a longer time range for earthquake prediction and forecasting research. The relevant vectorization methods and models were extended to the vectorization of seismic monitoring data such as groundwater level and geostress.

Claims

1. A method for vectorizing model seismic maps based on deep meta-learning, characterized by including: (1) Rasterization and image segmentation of seismic waveform records: The paper seismic record paper is rasterized and the obtained raster image is segmented to complete the construction of the image data required for the vectorization model; (2) Constructing the time coordinate system in the simulation seismic waveform record: The waveform in the simulation seismic record is calibrated with time markers, and the waveform time coordinate system is accurately modeled. At the same time, the time marker template matching algorithm and the automatic picking algorithm are used to shield the algorithm differences caused by various waveforms and machine types. (3) Vectorize the entire seismic waveform on the drawing: Vectorize the seismic waveform in the drawing and accurately splice it with the quiescent waveform to form a continuous waveform time series. At the same time, calculate the earthquake arrival time, draw the earthquake beach ball, and analyze the geophysical background field to detect and locate the seismic waveform, and remove various non-seismic image noise, smooth waves, intercalation points and other interference information that appear in the seismic waveform. The location of the seismic waveform is to embed the multi-scale segmentation algorithm as an operator into the MAML meta-learning framework, and divide the waveform samples according to the seismic instrument model, drawing size, noise type, etc. Sampling is performed on the sample space of seismic waveform drawings of various types, and meta-learning is performed on the operator model to improve the robustness of the operator and obtain the preliminary generalization parameters of the operator model. For a specific type of seismic drawing sample, starting from the generalization parameters obtained by meta-learning, the operator is further refined and trained to obtain a seismic waveform window detection model with high accuracy. (4) Establish a small sample library of seismic drawings based on noise, distortion, seismic wave type, and seismograph instrument type. Use meta-learning technology to establish a waveform correction and vectorization sampling algorithm framework that includes multiple deep learning methods. Construct a robust deep meta-learning vectorization model for multi-scale, multi-type, and small-sample simulated seismic drawings. Correct distortion and tilt in waveform records to ensure a one-to-one correspondence between the points in the waveform and the established time coordinates. Achieve robust and accurate vectorization of seismic waveforms. Standardize the vectorization results according to the international seismic data unified format to achieve consistency with digital seismic waveform records. The construction of the robust deep meta-learning vectorization model for simulated seismic drawings includes meta-learning-based seismic waveform vectorization, coordinate axis correction of seismic waveforms, warp correction of seismic waveforms, time axis interpolation processing, and construction of waveform time series. After sampling the upper and lower boundary points of the seismic waveform crankshaft, calculate the center pixel coordinates of the waveform curve to obtain the final seismic waveform curve sampling. The algorithm flow includes: 1) First, detect the starting point and line width of the oscillating waveform curve, and dynamically adjust the horizontal search step size according to the line width. Vertical search step size ; 2) Next, by sampling the upper and lower boundaries of the oscillating waveform curve from left to right and from bottom to top, the pixel transition point pairs and the center point of the pixel at the curve's starting point are determined. The calculation formula is as follows: （8） in The upper and lower bounds of the ordinate of the starting point of the curve; 3) Using step size Perform a horizontal search; if the point is black, find the center coordinates of the next sampling point on the curve. The calculation is as follows: the forward step length of the horizontal axis is... With the ordinate unchanged, we obtain the point. The search is performed in both the vertical and horizontal directions above and below the point to detect black pixels, i.e., color transition points, which are used as the upper bound. lower bound Then, the new coordinates are calculated. The formula is as follows: （9） 4) Using step size Perform a vertical search; if the point is black, find the center coordinates of the next sampling point on the curve. The calculation is as follows: the forward step length of the vertical axis is... With the x-coordinate unchanged, we obtain the point. The search is performed in both the left and right directions of the point to detect black pixels, i.e., color transition points, which are used as the left boundary points. Right boundary point Then, the new coordinates are calculated. The formula is as follows: （10） 5) Record the above traversal points and determine whether the waveform curve traversal is complete. If not, search in the horizontal or vertical direction of the current point and proceed to step 3) or 4) respectively, until the waveform traversal of the entire window area is completed. 6) Finally, visualize the data obtained from the above traversal, and submit it for manual judgment and correction to ensure that the vectorization result of the seismic waveform is accurate and reliable, and finally form the traversal result of the seismic waveform.

2. The method for vectorizing model seismic survey drawings according to claim 1, characterized in that: Step (1) involves the rasterization and image segmentation of the simulated seismic waveform record. This is a preprocessing algorithm study conducted on simulated seismic drawings of different types, multiple scales, and non-seismic noise after scanning the paper waveform record. The algorithm includes: automatically and quickly determining the size of the scanned drawing, automatically extracting various recording parameters from the drawing, and designing the best size normalization, type classification, grayscale conversion, and image segmentation algorithms for multi-scale drawings to provide high-quality input data for subsequent vectorization models.

3. The method for vectorizing model seismic survey drawings according to claim 1, characterized in that: The waveforms in the seismometry record described in step (2) consist of smooth waves, seismic waves, and start-stop waves.

4. The method for vectorizing model seismic survey drawings according to claim 1, characterized in that: The method for constructing the time coordinate system in the simulated seismic waveform record in step (2) is as follows: the starting point coordinates are manually picked up. The time-marked point feature map is generated, where d0 is the starting point for picking, and x0 and y0 are the x and y coordinates of the starting point d0, respectively. A Faster R-CNN deep network is used to detect and locate the time-marked points, calculating the position of each time-marked point in the drawing. Let i = 1, 2, 3..., where d i x is a time marker. i y i The x and y coordinates of the time marker points; Next, determine the pixel distance between two adjacent time markers. The calculation formula is as follows: (1) calculate Overall waveform length in the horizontal direction The ratio between The formula is as follows: (2) The length of the time recorded on the drawing is obtained by subtracting the clock times from the seismic records. This allows us to obtain the duration between the two time markers. The calculation formula is as follows: (3) Then, by combining the detected time marker location information, the time of all sampling locations in the drawing was calculated. Then, the construction and deduction of the time coordinate system in the seismic drawings were accurately completed: (xi≤x<xi+1) (4)。 5. The method for vectorizing model seismic survey drawings according to claim 1, characterized in that: The detection of seismic waveforms in step (3) involves using a multi-scale segmentation algorithm to segment the seismic background and seismic window, cutting the training samples according to the scale, and then using the focal-loss function to optimize the model loss calculation process. Finally, a dynamic adjustment strategy for the weights of difficult and easy detection targets is used to improve the problem of missed detection of targets.

6. The method for vectorizing model seismic survey drawings according to claim 1, characterized in that: Step (3) involves removing smooth waves and time points within the seismic waveform, including selecting the seismic waveform window, traversing the starting point of the waveform curve, scanning and tracking within the window, and extracting and removing smooth waves. This removal method includes: 1) Select a region with a vibrating waveform, then detect all grayscale jump points on the straight line to obtain n pairs of coordinates for the jump points on the drawing: Then the coordinates of the starting point of the i-th wave are waveform width The calculation is as follows: (5) 2) Arrange all curves within the window in reverse order of the y-coordinate of the detected starting point using Shell sort, and then obtain the starting point detection. 3) Eliminate smoothing waves. Taking curve i as an example, the overall search direction is horizontally to the right, with a given horizontal search step size. Vertical search step size ; Then the next point in its horizontal direction The search iteration formula is as follows: （6） For vertical search, The step size is the axis point with respect to the center of the waveform curve. Perform an up-and-down search; if the boundary is exceeded, switch to a complex wave search to obtain the upper bound. and the lower realm , to obtain the jump width: (7) When the jump width is greater than the line width h i If no longitudinal jump occurs throughout the entire horizontal path, the search proceeds to a complex wave search. i If so, it will be removed as a smooth wave; Meanwhile, within the seismic window, template matching is used to remove the noise from the extracted time stamps.

Images