First arrival time prediction method and device based on neural network

Abstract

The present disclosure relates to a first arrival time prediction method and device based on a neural network, the method comprising: reading trace head information of a seismic data prediction set, determining a data range on a same geophone line according to geophone coordinates in the prediction set; performing amplitude gain processing, amplitude limiting processing and filtering processing on original data on the same geophone line respectively to generate a plurality of feature maps; inputting the original data and the plurality of feature maps into a pre-trained first arrival time prediction model to obtain model output values; converting the model output values into first arrival times corresponding to the original data according to a preset threshold, and predicting the first arrival times corresponding to data on all geophone lines, wherein the data information is sufficiently mined through feature map extraction, and the weights of different features in first arrival identification are reasonably distributed through the introduction of an attention mechanism in the neural network model, thereby effectively increasing the effectiveness and stability of first arrival automatic picking.

Description

Technical Field

[0001] This disclosure relates to the field of geophysical exploration technology, and in particular to a method and apparatus for predicting the arrival time of first arrival waves based on neural networks. Background Technology

[0002] For onshore seismic exploration, complex near-surface structures severely affect the arrival time and morphology of deep reflection signals, significantly interfering with imaging of deep reservoirs. By utilizing the arrival time information of first arrival waves from seismic data, a near-surface model can be obtained through tomographic inversion. This model can further eliminate the influence of complex near-surface structures on deep reflection signals. Extracting the arrival time information of first arrival waves from observed seismic waveform data is fundamental to subsequent processing and plays a crucial role in seismic data processing.

[0003] Currently, there are two main methods for obtaining the arrival time of first arrival waves. One method involves displaying the seismic waveform on a computer and having seismic data processing experts manually pick out the arrival time of the first arrival wave. The advantage of this method is its high accuracy, but the disadvantages are that it requires significant manpower and resources, and its accuracy depends heavily on the expert's skill level. The other method utilizes automatic picking algorithms. Seismic waveform data is input, and a given algorithm identifies the first arrival waveform and outputs the arrival time of the first arrival wave. Existing automatic picking algorithms can be mainly divided into two categories. One category uses manually designed computer algorithms to obtain the first arrival time from the first arrival waveform, such as the long-short time window method and fractal-based first arrival picking algorithms. Because this type of method relies on manually set algorithms, it is difficult to achieve good results on seismic data with different characteristics. The other type of automatic picking algorithm is based on machine learning. Given waveform data and arrival time labels, machine learning algorithms can automatically obtain the mapping from waveform data to the first arrival time. This mapping relationship can be used for first arrival picking in seismic data.

[0004] With the widespread application of deep learning in general fields such as image processing, many seismic data processing experts have attempted to apply deep learning algorithms to the acquisition of first arrival times. For example, existing technologies include using U-Net convolutional networks to acquire first arrival times, using CNN and RNN networks, and using CWT-CNN classification networks to acquire microseismic arrival times. However, most current methods directly process field data and then input it into the neural network for training. Since some data feature extraction methods are not differentiable, they cannot be automatically learned using deep learning. Therefore, directly applying deep learning to the raw data may not fully extract the information within the data. Consequently, some hidden information in the raw data cannot be extracted through neural network features, or the effective features do not play a sufficient role within the network.

[0005] Therefore, it is essential to research deep learning-based automatic first arrival picking algorithms that can fully utilize hidden features in seismic data. Summary of the Invention

[0006] To solve the above-mentioned technical problems, or at least partially solve them, embodiments of this disclosure provide a method, apparatus, device, and storage medium for predicting the arrival time of first arrival waves based on neural networks.

[0007] In a first aspect, embodiments of this disclosure provide a method for predicting the arrival time of first arrival waves based on neural networks, comprising:

[0008] Read the lead information of the earthquake data prediction set, determine the data range on the same receiver line based on the coordinates of the receiver points in the prediction set, and take the data on the same receiver line as a minimum subset;

[0009] Read the original data of the minimum subset, and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset respectively to generate multiple feature maps;

[0010] The original data of the smallest subset and the multiple feature maps are input into the pre-trained first arrival wave time prediction model to obtain the model output value.

[0011] Based on a preset threshold, the model output value is converted into the arrival time of the first arrival wave corresponding to the original data of the smallest subset. Following the steps of predicting the arrival time of the first arrival wave corresponding to the original data of the smallest subset, the arrival time of the first arrival wave corresponding to the data on all detector lines is predicted.

[0012] In one possible implementation, the first arrival time prediction model is trained through the following steps:

[0013] The surface elevation is constructed based on the spatial location information of the source points and receiver points in the collected seismic data traces, and typical common shot gather data are selected according to the undulation of the surface elevation.

[0014] Pick up the first arrival time information corresponding to the common shot set data, save it to the track head, and form a label dataset of common shot set data - first arrival time;

[0015] Select a training set from the label dataset, determine the data range on the same detector line based on the detector point coordinates in the training set, take the data on the same detector line as a minimum subset, and record the start and end channel numbers corresponding to the minimum subset;

[0016] Read the original data of the minimum subset and its corresponding arrival and arrival time information, and convert the arrival and arrival time information into an arrival and arrival time matrix corresponding to the number of lanes in the minimum subset, wherein the matrix and the arrival and arrival time information have an index relationship;

[0017] The original data of the smallest subset is subjected to automatic gain control, amplitude limiting, and filtering respectively to generate multiple feature maps;

[0018] Using the minimum subset of original data and the multiple feature maps as input, and the first arrival time matrix as output, a neural network model is trained to obtain the first arrival time prediction model.

[0019] In one possible implementation, the neural network model includes an attention mechanism network and a first-arrival detection network. The process of training the neural network model using a minimum subset of the original data and the multiple feature maps as input, and the first-arrival time matrix as output, to obtain a first-arrival wave arrival time prediction model, includes:

[0020] Obtain the starting and ending indices of the original data of the smallest subset, read the seismic trace information corresponding to the starting and ending indices, and store the seismic trace information in a matrix Dat1 with H rows and W columns, where H is the number of sampling points for each trace of the seismic trace information, and W is the number of traces of the seismic trace.

[0021] Input the matrix Dat1 and its corresponding feature map into the attention mechanism network, and output the weighted matrix Dat;

[0022] Input the weighted matrix Dat into the initial arrival detection network, and the output matrix FB is H rows and W columns. out , where the matrix elements represent the probability that the current sample value is the first arrival;

[0023] Calculate matrix FB out with initial arrival time matrix FB label The first residual between the two is used to update the parameters of the attention mechanism network and the first arrival detection network until the first residual is less than a preset threshold.

[0024] Select a validation set from the label dataset and calculate the second residual corresponding to the data in the validation set. When the second residual is less than a preset threshold, use the first arrival time prediction model composed of the current attention mechanism network and the first arrival detection network as the trained model.

[0025] In one possible implementation, the starting and ending indices of the original data of the minimum subset include:

[0026] Acquire the relative coordinates RX, RY, or receiver point signals in the seismic data header;

[0027] Based on the preset range of RY or the receiver signal, the seismic data is divided into multiple receiver gathers;

[0028] Record the current number of common detection line gathers as NRecLine, and the start index Idx_Start and end index Idx_End of each common detection line gather.

[0029] In one possible implementation, converting the first-arrival / arrival time information into a first-arrival / arrival time matrix corresponding to the minimum subset number of lanes includes:

[0030] The initial arrival time information is set as a vector FB with W rows;

[0031] The vector FB is expanded into a matrix FB with H rows and W columns by following these steps. label :

[0032] FB label All elements are initialized to 0. The first arrival time (FAT) of each seismic trace is compared with the ratio of FAT to the sampling rate for each of the H sampling points in that trace. This determines the matrix FB of FAT's FAT in rows H and columns W for each seismic trace. 1abel Set the element at the specified position in the table to 1.

[0033] In one possible implementation, the step of converting the model output value into the arrival time of the first wave corresponding to the smallest subset of the original data according to a preset threshold includes:

[0034] The first arrival time of each seismic data point is determined in the matrix FB output by the model using the following expression. out Position in the middle:

[0035]

[0036] Where, samplerate is the sampling rate, p min For a preset threshold, the maxindex() function returns the index of the element containing the maximum value in a vector;

[0037] Obtain the vector FB of W rows as the first arrival time of the W seismic traces in the smallest subset.

[0038] In one possible implementation, the process of performing amplitude gain processing, amplitude limiting processing, and filtering processing on the original data of the minimum subset to generate multiple feature maps includes:

[0039] The amplitude gain of the original data of the minimum subset is processed using the following expression:

[0040]

[0041]

[0042] Where nw is the size of the time window. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace.

[0043] The following expression is used to perform amplitude limiting on the original data of the smallest subset:

[0044] Calculate the maximum value for each question, and then apply amplitude limiting to the data. The method is as follows:

[0045]

[0046] Among them, maxamp j Let p be the maximum amplitude of the j-th channel, and p be the amplitude limiting parameter. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace.

[0047] Bandpass processing is used to perform bandpass filtering on each trace, and the characteristic map after amplitude gain control is obtained based on the sample value of each sample point of each filtered seismic trace.

[0048] Secondly, embodiments of this disclosure provide a first-arrival wave arrival time prediction device based on a neural network, comprising:

[0049] The determination module is used to read the trace information of the earthquake data prediction set, determine the data range on the same detector line based on the coordinates of the detector points in the prediction set, and take the data on the same detector line as a minimum subset.

[0050] The generation module is used to read the original data of the minimum subset and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset to generate multiple feature maps.

[0051] The input module is used to input the minimum subset of raw data and the multiple feature maps into the pre-trained first arrival wave arrival time prediction model to obtain the model output value;

[0052] The conversion module is used to convert the model output value into the first arrival time corresponding to the original data of the minimum subset according to a preset threshold, and predict the first arrival time corresponding to the data on all detector lines according to the steps of predicting the first arrival time corresponding to the original data of the minimum subset.

[0053] Thirdly, embodiments of this disclosure provide an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0054] Memory, used to store computer programs;

[0055] When the processor executes the program stored in the memory, it implements the above-mentioned method for predicting the arrival time of the first arrival wave based on a neural network.

[0056] Fourthly, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the above-described neural network-based first-arrival wave arrival time prediction method.

[0057] Compared with the prior art, the technical solutions provided in this disclosure have at least some or all of the following advantages:

[0058] The first arrival time prediction method based on neural networks described in this embodiment reads the trace information of the earthquake data prediction set, determines the data range on the same receiver line based on the receiver coordinates in the prediction set, and treats the data on the same receiver line as a minimum subset; reads the original data of the minimum subset, and performs amplitude gain processing, amplitude limiting processing, and filtering processing on the original data of the minimum subset to generate multiple feature maps; inputs the original data of the minimum subset and the multiple feature maps into a pre-trained first arrival time prediction model to obtain the model output value; converts the model output value into the first arrival time corresponding to the original data of the minimum subset according to a preset threshold, and predicts the first arrival time corresponding to the data on all receiver lines according to the steps of predicting the first arrival time corresponding to the original data of the minimum subset. By fully mining the information contained in the data through the feature map extraction in the early stage, and reasonably allocating the weights of different features in first arrival recognition by introducing an attention mechanism in the neural network model, the effectiveness and stability of automatic first arrival picking can be effectively increased. Attached Figure Description

[0059] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0060] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0061] Figure 1 A schematic flowchart of a neural network-based first-arrival wave arrival time prediction method according to an embodiment of the present disclosure is shown.

[0062] Figure 2 The schematic diagram illustrates a flowchart of a neural network-based first arrival time prediction method according to another embodiment of the present disclosure;

[0063] Figure 3 The schematic diagram illustrates the structure of a first-arrival wave time prediction model according to an embodiment of the present disclosure;

[0064] Figure 4 The schematic diagram illustrates a flowchart of a neural network-based method for predicting the arrival time of first arrival waves according to yet another embodiment of the present disclosure.

[0065] Figure 5 This is the first test result of the first arrival wave arrival time prediction method based on neural networks according to the embodiments of this disclosure;

[0066] Figure 6 This is the second test result of the first arrival wave arrival time prediction method based on neural networks according to the embodiments of this disclosure;

[0067] Figure 7 A schematic block diagram of a device for predicting the arrival time of first arrival waves based on a neural network according to an embodiment of the present disclosure is shown; and

[0068] Figure 8 A schematic block diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0070] See Figure 1 The embodiments of this disclosure provide a method for predicting the arrival time of first arrival waves based on neural networks, including the following steps:

[0071] S1, Read the trace information of the earthquake data prediction set, determine the data range on the same detector line based on the coordinates of the detector points in the prediction set, and take the data on the same detector line as a minimum subset;

[0072] S2, read the original data of the minimum subset, and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset respectively to generate multiple feature maps;

[0073] S3, input the minimum subset of original data and the multiple feature maps into the pre-trained first arrival wave arrival time prediction model, and obtain the model output value;

[0074] S4. Based on a preset threshold, convert the model output value into the arrival time of the first arrival wave corresponding to the original data of the smallest subset, and predict the arrival time of the first arrival wave corresponding to the data on all detector lines according to the steps of predicting the arrival time of the first arrival wave corresponding to the original data of the smallest subset.

[0075] like Figure 2 As shown, an exemplary embodiment is provided to explain the neural network-based first-arrival wave arrival time prediction method of this disclosure, including:

[0076] S21. Extract the spatial location information of the source point and receiver point from the seismic data obtained in the field, construct the surface elevation and use computer software to display it in three dimensions. Combine the surface elevation undulation (the surface condition of the current data work area) and select representative shot gather data, which is usually one-twentieth to one-fiftieth of the original data shot number. The more complex the surface conditions, the more shot gathers are needed.

[0077] In practical applications, surface elevation undulations are categorized into various types, such as plains, mountains, loess plateaus, deserts, and piedmont zones. Based on the surface elevation morphology extracted from the current data, the traversed surface conditions are manually distinguished. Then, representative shot collections are selected from each surface condition and saved separately as Segy format files. The number of shot collections in each surface condition should be relatively consistent, and the distribution of shot collections within each surface condition should be relatively uniform.

[0078] S22, Manually pick representative shot collection data using manual picking software, and save the first arrival time to the track head to form a label dataset.

[0079] In practical applications, the domestically developed seismic data processing software GeoEast is used to acquire first arrival and arrival times. Alternatively, other software can be used.

[0080] S23, the generated label dataset is divided into a training set and a validation set;

[0081] In practical applications, 70% of the gun set data in the labeled dataset can be randomly selected as the training set and saved to a separate file; the remaining 30% of the data can be used as the validation set and saved to a separate file.

[0082] S24, obtain the common detection line index of the training set and the validation set respectively, read the trace head information of the training set and the validation set respectively, determine the data range of the same detection line according to the detection point coordinates, take the data on the same detection line as a minimum subset, and save the start and end trace numbers corresponding to the minimum subset to a separate file;

[0083] In practical applications, the relative coordinates RX and RY of the seismic data tracehead are first obtained. For the same receiver line, RY is similar. For observations of fixed data, the spacing between different receiver lines is relatively stable, denoted as RecLine_Interval, which is typically between 50 and 500. A relative tolerance Err_Relative is defined. The difference in RY between the current seismic trace and the previous seismic trace is detected, denoted as RY_Error. If RY_Error > Err_Relative * RecLine_Interval, it indicates that the current seismic trace has entered the next receiver line. Simultaneously, the number of co-receiver line gathers in the current dataset can be obtained, denoted as NRecLine. Finally, NRecLine, along with the start index Idx_Start and end index Idx_End of each co-receiver line gather, are saved to a file for later reading.

[0084] S25, construct the SE-UNET neural network, as follows Figure 3 As shown, the overall structure consists of two parts. The first part is the attention mechanism network, which takes the original data and multi-channel data composed of multiple features as input. Through a global pooling layer, a fully connected layer (FC), a ReLU layer, a fully connected layer, and a sigmoid layer, the weight of each channel can be calculated. The output is the data after weighting the multi-channel data using these weights. This disclosure currently extracts 13 features, which, together with the original data, form a total of 14 channels of data. The output is also the weighted data of 14 channels.

[0085] If the weight coefficients are represented by adjacent values...

[0086] w = [w1, w2, ..., w 14 ]

[0087] The output data is represented as

[0088] Dat = [D1, F1, ..., F] 13 ]

[0089] Where D1 represents the original data, F i The calculated features are all matrices of size H rows and W columns. After passing through the attention mechanism network, the data is...

[0090] Dat = [w1D1, w2F1, ..., w 13 F 13 ]

[0091] The second part of the neural network layer is the first arrival detection network. This disclosure uses a conventional U-Net network, which is input to the output of the attention mechanism network. The output is an H-row, W-column matrix, which specifically represents the probability that the current sample value is the first arrival.

[0092] S26, Train the neural network using training and test set data: Use the training and validation sets generated in step S23 and the common detection line index generated in step S24 to train on the network constructed in step S25, and obtain neural network parameters that can match the output data and the first and first arrival times through iterative inversion.

[0093] Training a network model in practical applications includes the following steps:

[0094] 6.1 Extract the original data and corresponding first and first arrival times from the training set data prepared in step S23 according to the minimum subset index information recorded in step S24.

[0095] 6.2 The original data is processed with different AGC gains, amplitude limiting, filtering and other methods to generate multiple feature maps;

[0096] 6.3 Input the raw data and multiple feature maps obtained in step 6.2 into the SE-UNET network simultaneously to obtain its output;

[0097] 6.4 The first arrival time information is expanded into matrix information of the same dimension as the neural network output data, and the residuals between the two are calculated;

[0098] 6.5 The residuals are backpropagated using the back projection algorithm to update the model parameters in the neural network;

[0099] 6.6 The residuals of the validation set can be calculated using steps 6.1 to 6.4. By comparing the parameters of the validation set with those of the training set, it can be evaluated whether the current model is overfitting or underfitting.

[0100] 6.7 Repeat steps 6.1 to 6.6 until the residual is small enough, and save the various parameters in the neural network model to a file.

[0101] S27, Repeat step S24 for the data to be applied, loading the neural network to predict the initial arrival time. Load the neural network model generated in step S26, input the processed raw data into the neural network, and after a series of processing steps, obtain the final initial arrival information and write the final information to a file.

[0102] In practical applications, using a trained neural network model to predict unpicked data includes the following steps:

[0103] 7.1 Read the lead information of the prediction set data, determine the data range on the same detector line based on the detector point coordinates, and take the data on the same detector line as a minimum subset. Save the start and end number of the minimum subset to a separate file.

[0104] 7.2 Load the neural network model output from step 6.7;

[0105] 7.3 Read the original data of the smallest subset and generate multiple feature maps in the manner described in step 6.2;

[0106] 7.4 Input the output of step 7.3 into the neural network model loaded in step 7.2, and obtain its output;

[0107] 7.5 Based on a manually given threshold, convert the output of the neural network into first and second arrival times and save it to a file;

[0108] 7.6 Repeat steps 7.3 to 7.5 until all data has been processed, completing the task of predicting the first arrival time on the new data using the neural network model.

[0109] In this embodiment, in step S3, the first arrival wave arrival time prediction model is trained through the following steps:

[0110] The surface elevation is constructed based on the spatial location information of the source points and receiver points in the collected seismic data traces, and typical common shot gather data are selected according to the undulation of the surface elevation.

[0111] Pick up the first arrival time information corresponding to the common shot set data, save it to the track head, and form a label dataset of common shot set data - first arrival time;

[0112] Select a training set from the label dataset, determine the data range on the same detector line based on the detector point coordinates in the training set, take the data on the same detector line as a minimum subset, and record the start and end channel numbers corresponding to the minimum subset;

[0113] Read the original data of the minimum subset and its corresponding arrival and arrival time information, and convert the arrival and arrival time information into an arrival and arrival time matrix corresponding to the number of lanes in the minimum subset, wherein the matrix and the arrival and arrival time information have an index relationship;

[0114] The original data of the smallest subset is subjected to automatic gain control, amplitude limiting, and filtering respectively to generate multiple feature maps;

[0115] Using the minimum subset of original data and the multiple feature maps as input, and the first arrival time matrix as output, a neural network model is trained to obtain the first arrival time prediction model.

[0116] In practical applications, the neural network model includes an attention mechanism network and a first-arrival detection network. The neural network model is trained using a minimum subset of the original data and the multiple feature maps as input, and the first-arrival time matrix as output, to obtain a first-arrival wave arrival time prediction model, including:

[0117] Obtain the starting and ending indices of the original data of the smallest subset, read the seismic trace information corresponding to the starting and ending indices, and store the seismic trace information in a matrix Dat1 with H rows and W columns, where H is the number of sampling points for each trace of the seismic trace information, and W is the number of traces of the seismic trace.

[0118] Input the matrix Dat1 and its corresponding feature map into the attention mechanism network, and output the weighted matrix Dat;

[0119] Input the weighted matrix Dat into the initial arrival detection network, and the output matrix FB is H rows and W columns. out , where the matrix elements represent the probability that the current sample value is the first arrival;

[0120] Calculate matrix FB out with initial arrival time matrix FB label The first residual between the two is used to update the parameters of the attention mechanism network and the first arrival detection network until the first residual is less than a preset threshold.

[0121] Select a validation set from the label dataset and calculate the second residual corresponding to the data in the validation set. When the second residual is less than a preset threshold, use the first arrival time prediction model composed of the current attention mechanism network and the first arrival detection network as the trained model.

[0122] In practical applications, the starting and ending indices of the original data of the minimum subset include:

[0123] Acquire the relative coordinates RX, RY, or receiver point signals in the seismic data header;

[0124] Based on the preset range of RY or the receiver signal, the seismic data is divided into multiple receiver gathers;

[0125] Record the current number of common detection line gathers as NRecLine, and the start index Idx_Start and end index Idx_End of each common detection line gather.

[0126] In this embodiment, step S4, converting the first-arrival / arrival time information into a first-arrival / arrival time matrix corresponding to the minimum subset number of lanes, includes:

[0127] The initial arrival time information is set as a vector FB with W rows;

[0128] The vector FB is expanded into a matrix FB with H rows and W columns by following these steps. label :

[0129] FB labelAll elements are initialized to 0. The first arrival time (FAT) of each seismic trace is compared with the ratio of FAT to the sampling rate for each of the H sampling points in that trace. This determines the matrix FB of FAT's FAT in rows H and columns W for each seismic trace. label Set the element at the specified position in the table to 1.

[0130] In this embodiment, step S2 involves performing amplitude gain processing, amplitude limiting processing, and filtering processing on the original data of the smallest subset to generate multiple feature maps, including:

[0131] The amplitude gain of the original data of the minimum subset is processed using the following expression:

[0132]

[0133]

[0134] Where nw is the size of the time window. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace.

[0135] The following expression is used to perform amplitude limiting on the original data of the smallest subset:

[0136] Calculate the maximum value for each question, and then perform amplitude limiting on the data. The method is as follows:

[0137]

[0138] Among them, maxamp j Let p be the maximum amplitude of the j-th channel, and p be the amplitude limiting parameter. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace.

[0139] Bandpass processing is used to perform bandpass filtering on each trace, and the characteristic map after amplitude gain control is obtained based on the sample value of each sample point of each filtered seismic trace.

[0140] like Figure 4 As shown, an exemplary embodiment is provided to explain the neural network training method of this disclosure, including:

[0141] S41, Obtain the data of a certain common receiver gather and the corresponding first arrival and arrival times: Read the file generated in step S24, obtain the start index Idx_Start and end index Idx_End of the corresponding receiver gather; Read the information of the seismic traces from Idx_Start to Idx_End in the corresponding training set, and save the seismic data in a matrix of H rows and W columns, where H is equal to the number of sampling points of each trace of the seismic data, and W is Idx_End-Idx_Start+1, denoted as Dat1; At the same time, obtain the first arrival and arrival time information of each trace from the seismic trace head, and save it in a vector of W rows, denoted as FB.

[0142] S42, calculate multiple feature maps of the original data: obtain 13 feature maps using various processing methods such as amplitude gain control (4 parameters), amplitude limiting (5 parameters), and filtering (4 parameters). However, it is not limited to the currently used methods.

[0143] The main method for amplitude gain control is to control the amplitude gain at each sample point of each seismic trace. Calculate weights

[0144]

[0145] Where nw is the size of the time window, applied to each point.

[0146]

[0147] The characteristic map after amplitude gain control is obtained. Here, four characteristic maps are generated by selecting 400, 200, 100, and 50 time windows respectively.

[0148] The amplitude limiting is calculated by determining the maximum value for each channel, and then the data is subjected to amplitude limiting processing. The method is as follows:

[0149]

[0150] maxamp j denoted as the maximum amplitude of the j-th channel, and p as the amplitude limiting parameter. This disclosure uses 0.9, 0, 8, 0.6, 0.4, and 0.2 to generate five feature maps.

[0151] The filtering process uses common bandpass filtering to perform bandpass filtering on each channel. This disclosure uses 3-60Hz, 4-50Hz, 5-40Hz, and 6-30Hz to generate four feature maps.

[0152] The parameters used here are not limited to those listed above, and the feature extraction methods used are not limited to those listed.

[0153] S43, Input the original data and corresponding feature maps into the neural network to obtain the output: Input Dat into the neural network, and calculate through the forward propagation of the neural network to obtain a matrix FB with H rows and W columns. out .

[0154] S44 expands the manually picked first and arrival times into a matrix. This disclosure expands the W-row vector FB into an H-row, W-column matrix FB. label First, set up FB. label All elements are initialized to 0. For each seismic data point, the matrix FB, located at row H and column W, can be calculated based on the first arrival time. label In the formula, the element at the corresponding position is set to 1. This can be expressed as follows:

[0155]

[0156] Where samplerate is the sampling rate.

[0157] S45, calculate the residual between the neural network output and the first-arrival time. This disclosure uses a binary classification cross-entropy loss function to define the neural network output FB. out Compared to manual pickup of FB label The residuals between them are specifically represented as follows:

[0158]

[0159] S46, Update the neural network model parameters. This disclosure uses the automatic gradient optimization feature of the neural network optimization system PyTorch to obtain the gradient of the objective function with respect to the neural network model parameters, and uses the adaptive moment estimation method (Adam) to iteratively optimize and update the model parameters. The updated model is output to a file, and the iteration loop continues until the residual is sufficiently small.

[0160] S47, Calculate the residuals of the validation set data. Repeat steps 201 to 205 for the validation set data to calculate the residuals and output them. The performance of the current neural network model can be evaluated by comparing the residuals of the training set and the validation set data.

[0161] In this embodiment, step S4, which involves converting the model output value into the arrival time of the first wave corresponding to the smallest subset of the original data based on a preset threshold, includes:

[0162] The first arrival time of each seismic data point is determined in the matrix FB output by the model using the following expression. out Position in the middle:

[0163]

[0164] Where, samplerate is the sampling rate, pmin For a preset threshold, the maxindex() function returns the index of the element containing the maximum value in a vector;

[0165] Obtain the vector FB of W rows as the first arrival time of W seismic traces in the smallest subset, output the first arrival time to an external file, and update the seismic data trace header with the finally obtained first arrival file to complete the prediction of the first arrival time.

[0166] The first arrival time prediction based on neural networks provided in this disclosure provides basic data for subsequent seismic data processing methods such as tomographic modeling and static correction.

[0167] Figure 5 To test the results of the method described in this disclosure on data with a high signal-to-noise ratio, Figure 6 To test the results of the method described herein on data with low signal-to-noise ratios (SNR), the curves in the figure represent the first arrivals predicted by this method, and the background waveform image represents the plotted seismic waveform data. The comparison shows that the picking results are relatively stable and the picking accuracy is high for both high and low SNR data. For the first arrival waveforms that can be picked in low SNR data, the accuracy is high, and extremely low SNR data is well identified, with no picking performed in this area. Compared to picking unreliable first arrivals, this result is more beneficial for subsequent practical applications.

[0168] See Figure 7 Embodiments of this disclosure provide a neural network-based first-arrival wave arrival time prediction device, comprising:

[0169] The determination module 11 is used to read the trace information of the earthquake data prediction set, determine the data range on the same detector line according to the coordinates of the detector points in the prediction set, and take the data on the same detector line as a minimum subset.

[0170] The generation module 12 is used to read the original data of the minimum subset and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset to generate multiple feature maps.

[0171] Input module 13 is used to input the minimum subset of original data and the multiple feature maps into the pre-trained first arrival wave arrival time prediction model to obtain the model output value;

[0172] The conversion module 14 is used to convert the model output value into the arrival time of the first arrival wave corresponding to the original data of the minimum subset according to a preset threshold, and to predict the arrival time of the first arrival wave corresponding to the data on all detector lines according to the step of predicting the arrival time of the first arrival wave corresponding to the original data of the minimum subset.

[0173] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0174] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] In the above embodiments, any and more of the determining module 11, generating module 12, input module 13, and conversion module 14 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in one module. At least one of the determining module 11, generating module 12, input module 13, and conversion module 14 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the determining module 11, generating module 12, input module 13, and conversion module 14 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0176] Reference Figure 8 As shown, the electronic device provided in the fourth exemplary embodiment of this disclosure includes a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140.

[0177] Memory 1130 is used to store computer programs;

[0178] When processor 1110 executes the program stored in memory 1130, it implements the following neural network-based first-arrival wave arrival time prediction method:

[0179] Read the lead information of the earthquake data prediction set, determine the data range on the same receiver line based on the coordinates of the receiver points in the prediction set, and take the data on the same receiver line as a minimum subset;

[0180] Read the original data of the minimum subset, and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset respectively to generate multiple feature maps;

[0181] The original data of the smallest subset and the multiple feature maps are input into the pre-trained first arrival wave time prediction model to obtain the model output value.

[0182] Based on a preset threshold, the model output value is converted into the arrival time of the first arrival wave corresponding to the original data of the smallest subset. Following the steps of predicting the arrival time of the first arrival wave corresponding to the original data of the smallest subset, the arrival time of the first arrival wave corresponding to the data on all detector lines is predicted.

[0183] The aforementioned communication bus 1140 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, it is represented by only one thick line in the figure, but this does not indicate that there is only one bus or one type of bus.

[0184] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0185] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1130 may also be at least one storage device located remotely from the aforementioned processor 1110.

[0186] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0187] Embodiments of this disclosure also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements the neural network-based first-arrival wave arrival time prediction method described above.

[0188] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus. The computer-readable storage medium carries one or more programs that, when executed, implement the neural network-based first-arrival wave arrival time prediction method according to embodiments of this disclosure.

[0189] According to embodiments of this disclosure, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0190] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0191] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for predicting the arrival time of first arrival waves based on neural networks, characterized in that, The method includes: Read the lead information of the earthquake data prediction set, determine the data range on the same receiver line based on the coordinates of the receiver points in the prediction set, and take the data on the same receiver line as a minimum subset; Read the original data of the minimum subset, and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset respectively to generate multiple feature maps; The minimum subset of original data and the multiple feature maps are input into a pre-trained first arrival time prediction model to obtain the model output value. The first arrival time prediction model is a neural network model, which includes an attention mechanism network and a first arrival detection network. Based on a preset threshold, the model output value is converted into the arrival time of the first arrival wave corresponding to the original data of the smallest subset. Following the steps of predicting the arrival time of the first arrival wave corresponding to the original data of the smallest subset, the arrival time of the first arrival wave corresponding to the data on all detector lines is predicted. The original data of the minimum subset is subjected to amplitude gain processing, amplitude limiting processing, and filtering processing respectively to generate multiple feature maps, including: The amplitude gain of the original data of the minimum subset is processed using the following expression: in, The size of the time window. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace. The following expression is used to perform amplitude limiting on the original data of the smallest subset: Calculate the maximum value for each question, and then perform amplitude limiting on the data. The method is as follows: in, Let p be the maximum amplitude of the j-th channel, and p be the amplitude limiting parameter. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace. Bandpass processing is used to perform bandpass filtering on each trace, and the characteristic map after amplitude gain control is obtained based on the sample value of each sample point of each filtered seismic trace.

2. The method according to claim 1, characterized in that, The first arrival wave arrival time prediction model is trained through the following steps: The surface elevation is constructed based on the spatial location information of the source points and receiver points in the collected seismic data traces, and typical common shot gather data are selected according to the undulation of the surface elevation. Pick up the first arrival time information corresponding to the common shot set data, save it to the track head, and form a label dataset of common shot set data - first arrival time; Select a training set from the label dataset, determine the data range on the same detector line based on the detector point coordinates in the training set, take the data on the same detector line as a minimum subset, and record the start and end channel numbers corresponding to the minimum subset; Read the original data of the minimum subset and its corresponding arrival and arrival time information, and convert the arrival and arrival time information into an arrival and arrival time matrix corresponding to the number of lanes in the minimum subset, wherein the matrix and the arrival and arrival time information have an index relationship; The original data of the smallest subset is subjected to automatic gain control, amplitude limiting, and filtering respectively to generate multiple feature maps; Using the minimum subset of original data and the multiple feature maps as input, and the first arrival time matrix as output, a neural network model is trained to obtain the first arrival time prediction model.

3. The method according to claim 2, characterized in that, The process of training a neural network model using the minimum subset of original data and the multiple feature maps as input, and the first arrival time matrix as output, to obtain a first arrival time prediction model includes: Obtain the starting and ending indices of the original data of the minimum subset, read the seismic trace information corresponding to the starting and ending indices, and store the seismic trace information into a matrix of H rows and W columns. In the above, H represents the number of sampling points for each seismic trace, and W represents the number of seismic traces. matrix And in the corresponding feature map input attention mechanism network, the output weighted matrix ; The weighted matrix Input an initial arrival detection network and output a matrix of H rows and W columns. , where the matrix elements represent the probability that the current sample value is the first arrival; Calculate matrix Arrival and arrival time matrix The first residual between the two is used to update the parameters of the attention mechanism network and the first arrival detection network until the first residual is less than a preset threshold. Select a validation set from the label dataset and calculate the second residual corresponding to the data in the validation set. When the second residual is less than a preset threshold, use the first arrival time prediction model composed of the current attention mechanism network and the first arrival detection network as the trained model.

4. The method according to claim 3, characterized in that, The starting and ending indices of the original data in the minimum subset include: Acquire the relative coordinates RX, RY, or receiver point signals in the seismic data header; Based on the preset range of RY or the receiver signal, the seismic data is divided into multiple receiver gathers; Record the current number of common detection line gathers as NRecLine, and the start index Idx_Start and end index Idx_End of each common detection line gather.

5. The method according to claim 2, characterized in that, The step of converting the first arrival time information into a first arrival time matrix corresponding to the minimum subset number of lanes includes: The initial arrival time information is set as a vector of W rows. ; The vector will be processed through the following steps. Expanded into a matrix of H rows and W columns : Will All elements are initialized to 0. The first arrival time (FAT) and the ratio of FAT to the sampling rate for each of the H sampling points in that FAT are compared to determine the matrix of FAT FAT in rows H and columns W for each FAT. Set the element at the specified position in the table to 1.

6. The method according to claim 5, characterized in that, The step of converting the model output value into the arrival time of the first wave corresponding to the smallest subset of the original data according to a preset threshold includes: The first arrival time of each seismic data point is determined in the matrix output by the model using the following expression. Position in the middle: in, Sampling rate, For a preset threshold, the maxindex() function returns the index of the element containing the maximum value in a vector; Get the vector of W rows , which represents the first arrival time of the W seismic traces in the smallest subset.

7. A device for predicting the arrival time of first arrival waves based on a neural network, characterized in that, include: The determination module is used to read the trace information of the earthquake data prediction set, determine the data range on the same detector line based on the coordinates of the detector points in the prediction set, and take the data on the same detector line as a minimum subset. The generation module is used to read the original data of the minimum subset and perform amplitude gain processing, amplitude limiting processing and filtering processing on the original data of the minimum subset to generate multiple feature maps. The input module is used to input the minimum subset of original data and the multiple feature maps into the pre-trained first arrival time prediction model and obtain the model output value. The first arrival time prediction model is a neural network model, which includes an attention mechanism network and a first arrival detection network. The conversion module is used to convert the model output value into the first arrival time corresponding to the original data of the minimum subset according to a preset threshold, and predict the first arrival time corresponding to the data on all detector lines according to the steps of predicting the first arrival time corresponding to the original data of the minimum subset. The original data of the minimum subset is subjected to amplitude gain processing, amplitude limiting processing, and filtering processing respectively to generate multiple feature maps, including: The amplitude gain of the original data of the minimum subset is processed using the following expression: in, For the size of the time window, For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace. The following expression is used to perform amplitude limiting on the original data of the smallest subset: Calculate the maximum value for each question, and then perform amplitude limiting on the data. The method is as follows: in, Let p be the maximum amplitude of the j-th channel, and p be the amplitude limiting parameter. For each sample point value of each seismic trace, a characteristic map after amplitude gain control is obtained based on each sample point value of each seismic trace. Bandpass processing is used to perform bandpass filtering on each trace, and the characteristic map after amplitude gain control is obtained based on the sample value of each sample point of each filtered seismic trace.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the neural network-based first arrival time prediction method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the neural network-based first arrival time prediction method according to any one of claims 1-6.

Images