CN108596327A - A kind of seismic velocity spectrum artificial intelligence pick-up method based on deep learning - Google Patents
A kind of seismic velocity spectrum artificial intelligence pick-up method based on deep learning Download PDFInfo
- Publication number
- CN108596327A CN108596327A CN201810260667.3A CN201810260667A CN108596327A CN 108596327 A CN108596327 A CN 108596327A CN 201810260667 A CN201810260667 A CN 201810260667A CN 108596327 A CN108596327 A CN 108596327A
- Authority
- CN
- China
- Prior art keywords
- neural network
- layer
- velocity spectrum
- velocity
- pooling
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000001228 spectrum Methods 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 20
- 238000013135 deep learning Methods 0.000 title claims abstract description 19
- 238000013528 artificial neural network Methods 0.000 claims abstract description 64
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 63
- 230000000306 recurrent effect Effects 0.000 claims abstract description 46
- 238000000605 extraction Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 75
- 238000011176 pooling Methods 0.000 claims description 68
- 230000006870 function Effects 0.000 claims description 44
- 125000004122 cyclic group Chemical group 0.000 claims description 15
- 230000035945 sensitivity Effects 0.000 claims description 12
- 210000002569 neuron Anatomy 0.000 claims description 10
- 210000004027 cell Anatomy 0.000 claims description 8
- 238000007670 refining Methods 0.000 claims description 7
- 230000002087 whitening effect Effects 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 238000001514 detection method Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 7
- 238000012545 processing Methods 0.000 abstract description 6
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000004364 calculation method Methods 0.000 description 13
- 230000014509 gene expression Effects 0.000 description 12
- 239000013598 vector Substances 0.000 description 11
- 239000011159 matrix material Substances 0.000 description 9
- 230000004913 activation Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 4
- 238000012937 correction Methods 0.000 description 4
- 230000001413 cellular effect Effects 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 101100377706 Escherichia phage T5 A2.2 gene Proteins 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
- 230000017105 transposition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
- Feedback Control In General (AREA)
Abstract
The present invention provides a kind of seismic velocity spectrum artificial intelligence pick-up method based on deep learning, has the characteristics that image based on normal-moveout spectrum, the feature of energy group in being composed by convolutional neural networks extraction rate, and prediction result of the time speed in normal-moveout spectrum corresponding to energy group to sequence is calculated from the feature extracted；Time series feature based on normal-moveout spectrum again, using Recognition with Recurrent Neural Network carry it is described take the time of prediction result and the relationship characteristic of speed, the prediction result is modified according to these relationship characteristics, finally obtains correct time speed to sequence.Beneficial effects of the present invention：High degree of automation is conducive to the time for reducing seism processing, can substantially liberate intelligence and the manual labor of seismic data process personnel completely.
Description
Technical Field
The invention relates to the technical field of seismic data processing of seismic exploration, in particular to a seismic velocity spectrum artificial intelligence picking method based on deep learning.
Background
Velocity analysis is the basis for conventional seismic data processing. In the conventional velocity analysis method, a series of velocities at the same interval are given to scan, and a velocity spectrum (fig. 1) is produced with the superposition energy or the similarity coefficient or the like as a criterion for velocity analysis. Energy boluses of the velocity spectrum correspond to strong reflection information and can provide correct kinetic correction velocity. The accuracy of the speed pick-up can directly influence the results of dynamic correction, offset, AVO and time-depth conversion in oil and gas exploration.
The current velocity spectrum picking mode is completed manually, the working efficiency is low, the time consumption is high, and human errors are easy to occur. Therefore, with the continuous development of seismic exploration and the continuous expansion of exploration fields, the number of seismic data to be processed is increased, and higher requirements on the picking efficiency and precision of velocity spectrum in velocity analysis are also provided.
Through the search of the prior art, the Chinese patent with the publication number of CN105445788A provides an automatic velocity spectrum interpretation method based on a model and global optimization, the method has low picking efficiency, the result is often random, and the method is easy to make mistakes.
Another chinese patent, publication No. CN105572733A, provides an automatic seismic velocity spectrum picking method. The method comprises the steps of firstly converting seismic data into two-dimensional grid data, then setting a calculation time window, determining the position of a maximum value in the area of the calculation time window, then defining a speed search threshold value, recording the position of the maximum value of a speed spectrum, connecting a speed curve, and finally carrying out interpolation on the speed curve. The method needs to manually set reasonable calculation time window and speed search threshold according to prior knowledge, and cannot realize intellectualization and automation.
Disclosure of Invention
In view of this, the embodiment of the present invention provides an artificial intelligence picking method for seismic velocity spectrum based on deep learning.
The embodiment of the invention provides an artificial intelligence picking method of a seismic velocity spectrum based on deep learning, which is characterized in that based on the characteristic that the velocity spectrum has images, the characteristics of energy clusters in the velocity spectrum are extracted through a convolutional neural network, and the prediction result of a time-velocity pair sequence corresponding to the energy clusters in the velocity spectrum is calculated from the extracted characteristics; and then based on the time sequence characteristics of the velocity spectrum, extracting the relation characteristics of the time and the velocity of the prediction result by using a recurrent neural network, and correcting the prediction result according to the relation characteristics to finally obtain an accurate time-velocity pair sequence.
Further, the method comprises the following steps of:
a1: taking the preprocessed original velocity spectrum and the artificial extraction value of the time-velocity pair sequence corresponding to the original velocity spectrum as training samples, training the convolutional neural network, and obtaining a convolutional neural network training model capable of predicting the prediction result;
a2: taking the prediction result and the manually extracted value of the time-speed pair sequence of the original speed spectrum as a new training sample set, training the recurrent neural network, and obtaining a recurrent neural network training model capable of obtaining an accurate time-speed pair sequence;
a3: and sequentially inputting the original velocity spectrum into the convolutional neural network training model and the cyclic neural network training model to obtain an accurate time-velocity pair sequence.
Furthermore, the convolutional neural network comprises a plurality of convolutional layers and a plurality of pooling layers arranged in a cross way with the convolutional layers, each pooling layer is positioned behind the convolutional layer corresponding to the pooling layer, the cyclic neural network further comprises a plurality of full-connection layers positioned behind the last pooling layer, the full-connection layers are arranged in front and at the back, and the last full-connection layer corresponds to the output of the neural network.
Further, the recurrent neural network includes an input layer, an output layer, and a hidden layer between the input layer and the output layer, the input layer corresponding to an output of the convolutional neural network.
Further, the step of refining of step a1 includes:
a1.1: carrying out whitening pretreatment on an original velocity spectrum to obtain a velocity spectrum, wherein the velocity spectrum and a corresponding time-velocity pair sequence thereof form a training data set of a convolutional neural network;
a1.2: randomly shuffling a training data set of the convolutional neural network, grouping the data in the training set, wherein the number of the data in each group is consistent, and sequentially inputting each group into the convolutional neural network, wherein each group of data sequentially passes through a plurality of convolutional layers;
a1.3: extracting and fusing the features on one convolution layer by using a sliding window to obtain convolution feature maps on the convolution layer, wherein each convolution feature map is related to the weight of the convolution layer corresponding to the convolution feature map;
a1.4: inputting the convolution characteristic map into a subsequent pooling layer adjacent to the convolution characteristic map, wherein the pooling layer adopts a maximum pooling algorithm to obtain a pooling characteristic map, and the pooling characteristic map enters another subsequent convolution layer adjacent to the pooling characteristic map;
a1.5: repeating the steps A1.3 and A1.4 until data enters the last convolutional layer, wherein a Dropout layer is inserted in the last convolutional layer, the Dropout layer is used for randomly leading the value of the pooling feature map passing through the Dropout layer to be zero so as to reduce the phenomenon of network overfitting, then the convolutional layer generates a last convolutional feature map, and the convolutional feature map is finally pooled by the last pooling layer to generate a pooling feature map;
a1.6: performing multiple classification identification on the region position of the energy mass of the velocity spectrum based on the pooling characteristic maps of different pooling layers, screening out the real position of the energy mass, and performing multiple detection and inference on the peak value of the energy mass by using a full-connected layer to obtain the actual position of the peak value of the energy mass, namely the prediction result of the time-velocity pair sequence of the velocity spectrum;
a1.7: calculating a cost function of the convolutional neural network, judging whether the cost function of the last full-connection layer is converged, finishing the training of the convolutional neural network if the cost function of the last full-connection layer is converged, and outputting a prediction result of the time-speed pair sequence, otherwise, performing the next step;
a1.8: calculating gradient values of cost functions of all layers in the convolutional neural network by using a back conduction algorithm, calculating error sensitivity of all layers of the convolutional neural network according to the obtained gradient of each layer by using a momentum-based random gradient descent algorithm, and updating the weight of each layer by using the error sensitivity of each layer;
a1.8: and returning to the step 1.2.
Further, the step of refining of step a2 includes:
a2.1: the prediction result and the artificially extracted value of the time-speed pair sequence of the original speed spectrum form a training sample set of the recurrent neural network, and the training sample set of the recurrent neural network is input into the recurrent neural network;
a2.2: the recurrent neural network extracts the relation characteristic of time and speed in the prediction result through the neurons of the recurrent neural network, the relation characteristic is related to the weight of the corresponding layer, and a more accurate time-speed pair sequence is obtained by utilizing the relation characteristic;
a2.3: calculating a cost function of the convolutional neural network, judging whether the cost function of the output layer is converged, if so, finishing the training of the convolutional neural network, and if not, performing the next step;
a2.4: calculating gradient values of cost functions of all layers in the cyclic neural network by using a reverse conduction algorithm, calculating error sensitivity of all layers in the cyclic neural network according to the gradient values of all layers by using an Adam learning rate self-adaptive algorithm, and updating the weight of each layer by using the error sensitivity of each layer;
a2.5: repeat step A2.2.
Further, the training sample set of the recurrent neural network is formed in an input layer of the recurrent neural network, a hidden layer of the recurrent neural network is used for correcting the prediction result, an output layer of the recurrent neural network is used for outputting a more accurate time-speed pair sequence, and the recurrent neural network is a long-short term memory network structure which comprises a forgetting gate, an input gate, an output gate and a cell state.
Further, the step of refining of step a3 includes:
a3.1: carrying out whitening pretreatment on the original velocity spectrum;
a3.2: inputting the preprocessed velocity spectrum into the convolutional neural network training model, wherein the position corresponding to the energy mass in the processed velocity spectrum is the prediction result of the time-speed pair sequence of the processed velocity spectrum;
a3.3: inputting the prediction result of the processed time-speed pair sequence of the velocity spectrum into the recurrent neural network training model, correcting the prediction result by the recurrent neural network training model, and finally outputting an accurate time-speed pair sequence;
a3.4: and combining a plurality of velocity spectrum intelligent pickup results of the same measuring line at different positions, and forming a superimposed velocity profile through interpolation.
The technical scheme provided by the embodiment of the invention has the following beneficial effects: the seismic velocity spectrum artificial intelligence picking method based on the deep learning can (1) carry out artificial intelligence picking on a seismic velocity spectrum by adopting the seismic velocity spectrum artificial intelligence picking method based on the deep learning, and corresponding original seismic velocity spectrum data are input into a convolutional neural network training model and a cyclic neural network training model obtained after training, so that an accurate time-velocity pair sequence can be obtained without any artificial intervention in the period, and the degree of automation is high; (2) the invention needs to spend a lot of time when training the model; after training is finished, the time-velocity curve can be automatically and quickly obtained by inputting the original velocity spectrum data into the training model, the time for processing the seismic data is favorably shortened, and the intelligence and physical labor of seismic data processing personnel can be basically and completely liberated.
Drawings
FIG. 1 is a prior art seismic velocity spectrum;
FIG. 2 is a block diagram of a convolutional neural network and a cyclic neural network of the present invention;
FIG. 3 is a comparison schematic of the velocity spectrum pickup of the present invention;
FIG. 4 is a frame diagram of the seismic velocity spectrum artificial intelligence picking method based on deep learning of the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be further described with reference to the accompanying drawings.
Referring to fig. 2 and 4, an embodiment of the present invention provides an artificial intelligence picking method for seismic velocity spectrum based on deep learning, which mainly includes two stages of learning and intelligent picking, where the learning stage mainly trains a convolutional neural network to obtain a convolutional neural network training model and trains a cyclic neural network to obtain a cyclic neural network training model; in the intelligent picking stage, the accurate time-velocity pair sequence in the original velocity spectrum is obtained by using the convolutional neural network training model and the cyclic neural network training model.
The seismic velocity spectrum artificial intelligence picking method based on deep learning is characterized in that a velocity spectrum has the characteristics of an image, the characteristics of energy clusters in the velocity spectrum are extracted through the convolutional neural network training model, and the prediction result of a time-velocity pair sequence corresponding to the energy clusters in the velocity spectrum is calculated from the extracted characteristics; and based on the time sequence characteristics of the velocity spectrum, extracting the relation characteristics of the time and the velocity of the prediction result by using the recurrent neural network training model, and correcting the prediction result according to the relation characteristics to finally obtain an accurate time-velocity pair sequence.
Specifically, the learning phase includes:
a1: and taking the preprocessed original velocity spectrum and the artificial extraction value of the time-velocity pair sequence corresponding to the original velocity spectrum as training samples, training the convolutional neural network, and obtaining a convolutional neural network training model capable of predicting the prediction result.
Referring to fig. 2, the convolutional neural network includes a plurality of convolutional layers and a plurality of pooling layers arranged to intersect with the convolutional layers, each pooling layer is located behind the convolutional layer corresponding to the convolutional layer, the convolutional neural network further includes a plurality of fully-connected layers located behind the last pooling layer, the plurality of fully-connected layers are arranged in front of and behind each other, and the last fully-connected layer corresponds to an output of the neural network.
In this embodiment, the convolutional neural network at least includes 4 convolutional layers, 4 pooling layers, and 3 fully-connected layers.
The convolutional neural network has the specific structure as follows:
the first layer is a convolution layer, 16 convolution kernels are used in the convolution layer, the window size of the convolution kernels is 3 x 3 grids, the moving step length is set to be 2, the activation function is set to be a ReLU function, and 16 convolution characteristic graphs are output;
the second layer is a pooling layer, the kernel window size of the second layer is 2 multiplied by 2 grids, the moving step length is set to be 2, the size of the obtained pooling characteristic graph is changed to be one fourth of the original size after the convolution characteristic graph passes through the pooling layer, the number of the characteristic graphs is not changed, and 16 pooling characteristic graphs are output;
the third layer is a convolution layer, the layer uses 32 convolution kernels, the window size of the convolution kernels is 3 x 3 grids, the moving step length is set to be 2, the activation function is set to be a ReLU function, and 32 convolution characteristic graphs are output;
the fourth layer is a pooling layer, the size of a kernel window of the pooling layer is 2 multiplied by 2 grids, the moving step length is set to be 2, and 32 pooling feature maps are output;
the fifth layer is a convolution layer, the convolution layer uses 64 convolution kernels, the window size of the convolution kernels is 3 x 3 grids, the moving step length is set to be 2, the activation function is set to be a ReLU function, and 64 convolution characteristic graphs are output;
the sixth layer is a pooling layer, the size of a kernel window of the pooling layer is 2 multiplied by 2 grids, the moving step length is set to be 2, and 64 pooling feature maps are output;
the seventh layer is a convolution layer, the layer uses 256 convolution kernels, the window size of the convolution kernels is 3 x 3 grids, the moving step size is set to be 2, the activation function is set to be a ReLU function, meanwhile, a Dropout layer is added, and 256 convolution characteristic graphs are output;
the eighth layer is a pooling layer, the size of a kernel window of the layer is 2 multiplied by 2 grids, the moving step length is set to be 2, and 256 pooling characteristic graphs are output;
the ninth layer is a full connection layer, and 256 neurons are arranged in the layer;
the tenth layer is a full connection layer, and 1024 neurons are arranged in the layer;
the eleventh layer is a fully connected layer, and the number of neurons in the layer corresponds to the number of time-velocity pairs and energy boluses in the velocity spectrum.
The process of training the convolutional neural network is as follows:
a1.1: and performing whitening pretreatment on the original velocity spectrum to obtain a velocity spectrum, wherein the velocity spectrum and a corresponding time-velocity pair sequence form a training data set of the convolutional neural network.
The process of whitening preprocessing the raw velocity spectrum is: an amplitude threshold value of the original velocity spectrum grid points is set, and when the amplitude value of one of the grid points is smaller than the threshold value, the amplitude value of the grid point is set to zero, thereby obtaining a preprocessed velocity spectrum. And manually extracting a time-speed pair sequence of the processed velocity spectrum, wherein the time-speed pair sequence is an estimated value and has relatively low accuracy.
A1.2: randomly shuffling the training data set of the convolutional neural network, grouping the data in the training set, wherein the number of the data in each group is consistent, and sequentially inputting each group into the convolutional neural network, wherein each group of data sequentially passes through a plurality of convolutional layers.
A1.3: and extracting and fusing the features on one convolution layer by using a sliding window to obtain convolution feature maps on the convolution layer, wherein each convolution feature map is related to the weight of the convolution layer corresponding to the convolution feature map.
The forward implementation process of the convolutional layer is as follows:
ai，k＝f(xi*Wk+bk) (1)
wherein, ai，kIs the ith tensor velocity spectral data x of the training set input via the kth convolution kernel in the convolutional layeriConvolution characteristic graph obtained by convolution processing, representing convolution operation, WkWeight representing the kth convolution, bkFor the threshold corresponding to the kth convolution kernel, f (-) is the ReLU-type activation function:
f(x)＝max(0，x) (2)
as can be seen from formula (2), when x is less than or equal to zero, the function value is zero; when x is greater than zero, the function value is still x.
A1.4: and inputting the convolution characteristic map into the subsequent pooling layer adjacent to the convolution characteristic map, wherein the pooling layer adopts a maximum pooling algorithm to obtain a pooling characteristic map, and the pooling characteristic map enters another subsequent convolution layer adjacent to the pooling characteristic map.
And inputting the convolution characteristic diagram generated by the convolution layer into the pooling layer, wherein the pooling layer adopts maximum pooling, the pooling dimension is set to be 2, the moving step length is 2, the size of the pooling characteristic diagram obtained after the convolution characteristic diagram passes through the pooling layer becomes one fourth of the original size, and the number of the characteristic diagrams is unchanged. The maximum pooling uses the following formula:
cj＝maxpool(ai，k) (3)
wherein, cjA jth pooling profile generated for the pooling layer.
A1.5: repeating the steps A1.3 and A1.4 until the data enters the last convolution layer, wherein a Dropout layer is inserted in the last convolution layer and used for randomly enabling the value of the pooling feature map passing through the Dropout layer to be zero so as to reduce the phenomenon of network overfitting, and then the convolution layer generates a last convolution feature map which is finally pooled by the last pooling layer to generate a pooling feature map.
The calculation process for randomly letting the pooling profile across the Dropout layer be set to zero is:
Dropout(x)＝RandomZero(p)×x (4)
the RandomZero (p) indicates that the value in the data matrix x of the layer is set to be 0 according to the set probability p, and the Dropout (x) indicates the data matrix obtained after the Dropout layer is passed through in the training stage.
A1.6: and performing multiple classification and identification on the region position of the energy mass of the velocity spectrum based on the pooling characteristic maps of different pooling layers, screening out the real position of the energy mass, and performing multiple detection and inference on the peak value of the energy mass by using the full-connected layers to obtain the actual position of the peak value of the energy mass, namely the prediction result of the time-velocity pair sequence of the velocity spectrum.
Wherein the forward calculation process of the full connection layer is as follows:
ai，l＝f(Wlai，l-1+bl) (5)
in the formula, ai，lAnd the characteristic vector represents the ith tensile velocity spectrum data after passing through the full connecting layer, and the characteristic vector contains the actual position of the energy mass peak.
A1.7: and calculating a cost function of the convolutional neural network, judging whether the cost function of the last full-connection layer is converged, finishing the training of the convolutional neural network if the cost function of the last full-connection layer is converged, and outputting a prediction result of the time-speed pair sequence, otherwise, performing the next step.
The calculation formula of the cost function of the convolutional neural network is as follows:
wherein, yiDenotes xiThe corresponding objective, f (-) represents the function of the forward computation of all convolutional layers, L (-) represents the loss function, m is the number of input vectors, and θ represents the parameter of the convolutional neural network.
A1.8: calculating gradient values of cost functions of all layers in the convolutional neural network by using a back conduction algorithm, calculating error sensitivity of all layers of the convolutional neural network according to the obtained gradient of each layer by using a momentum-based random gradient descent algorithm, and updating the weight of each layer by using the error sensitivity of each layer.
If the current l layer is a fully connected layer, the residual calculation formula of the current l layer is as follows:
δl＝(Wl)Tδl+1f′(al) (7)
parameter WlThe gradient of (a) is calculated as:
the gradient calculation formula of the parameter b is as follows:
wherein, deltal+1Is the residual error of the (l + 1) th layer in the network, J (W, b; x, y) is the cost function, W, b, x, y respectively represent the weight, the threshold, the training input and the corresponding target, and T is the matrix transposition.
If the current layer is a convolutional layer, the residual calculation formula of the current layer is as follows:
δl＝δl+1*rot180(Wl)f′(al) (10)
wherein, rot180 (W)l) Represents WlA rotation of 180 degrees is required.
If the current l layer is a pooling layer, the residual calculation formula of the current l layer is as follows:
δl＝upsample((Wl)Tδl+1)f′(al) (11)
where the upsamplle (-) function represents a function of the amplification and error redistribution of the pooled error matrix.
The update formula of the momentum random gradient descent algorithm is as follows:
wherein,representing the gradient of a parameter theta calculated from x, y in the training set, α being the learning rate, gamma being the momentum parameter, vtIs the current velocity vector, vt-1Is the velocity vector in the previous iteration. a is initially set to 0.1, vtThe initial setting is 0, gamma is between the intervals (0,1), gamma is set to 0.5 in the initial training stage, and becomes larger as the number of iterations increases.
A1.8: and returning to the step 1.2.
A2: and taking the prediction result and the manually extracted value of the time-speed pair sequence of the original speed spectrum as a new training sample set, training the recurrent neural network, and obtaining a recurrent neural network training model capable of obtaining an accurate time-speed pair sequence.
The recurrent neural network comprises an input layer, an output layer and a hidden layer positioned between the input layer and the output layer, wherein the input layer corresponds to the output of the convolutional neural network. The input layer vector dimensions correspond to the output of a convolutional neural network; the hidden layer uses at least 20 neurons; the output layer is a time-speed pair and has the same number as the neurons of the eleventh layer of the convolutional neural network.
The process of training the recurrent neural network is as follows:
a2.1: and (3) forming a training sample set of the recurrent neural network by the prediction result and the artificially extracted value of the time-velocity pair sequence of the original velocity spectrum, and inputting the training sample set of the recurrent neural network into the recurrent neural network.
A2.2: and the recurrent neural network extracts the relation characteristic of time and speed in the prediction result through the neurons of the recurrent neural network, the relation characteristic is related to the weight of the corresponding layer, and a more accurate time-speed pair sequence is obtained by utilizing the relation characteristic.
The recurrent neural network is a long-short term memory (LSTM) network structure, which comprises a forgetting gate, an input gate, an output gate and a cell state, and the forward calculation formula is as follows:
f(t)＝σ(Wfh(t-1)+Ufx(t)+bf) (13)
i(t)＝σ(Wih(t-1)+Uix(t)+bi) (14)
a(t)＝tanh(Wah(t-1)+Uax(t)+ba) (15)
C(t)＝C(t-1)⊙f(t)+i(t)⊙a(t)(16)
o(t)＝σ(Woh(t-1)+Uox(t)+bo) (17)
h(t)＝o(t)⊙tanh(C(t)) (18)
wherein, the formula (13) is a mathematical expression of the LSTM forgetting door, f(t)Indicates the probability of forgetting the state of the previous layer of hidden cells, h(t-1)Representing the hidden state of the previous sequence, x(t)Representing the input of the current sequence, Wf，Uf，bfRespectively, a weight matrix and an offset vector, and sigma is a sigmoid activation function.
The LSTM input gate is composed of two parts, and the expressions (14) and (15) are mathematical expressions of the two parts respectively. i.e. i(t)，a(t)Respectively representing the outputs of the two parts. WiAnd UiWeight matrix representing input gates, biRepresenting the offset vector of the input gate, WaAnd UaWeight matrix representing input gates, baRepresenting the offset vector of the input gate.
Equation (16) is a mathematical expression for updating the state of cells, where C(t)Indicating the current cellular state, C(t-1)indicates the cell state of the previous sequence, ⊙ indicating the Hadamard product.
Equations (17), (18) represent two-part mathematical expressions for the output gate, the first part being o(t)I.e. a hidden neuron which is hidden by the previous sequence of hidden states h(t-1)And the input vector of the current sequence and the activation function sigmoid, and the second part h(t)Indicating the current hidden state, which is represented by the current cell state C(t)Tan h activation function and o(t)And (4) forming. Hidden neurons and hidden states are speed, timeAnd (4) information.
Equation (19) is a mathematical expression for updating the prediction output of the current sequence.For the prediction output of the current sequence, V and c represent the weight matrix and the offset vector of the prediction output, respectively.
A2.3: and calculating a cost function of the convolutional neural network, judging whether the cost function of the output layer is converged, finishing the training of the convolutional neural network if the cost function of the output layer is converged, and otherwise, performing the next step.
A2.4: and calculating gradient values of cost functions of all layers in the cyclic neural network by using a reverse conduction algorithm, calculating error sensitivity of all layers in the cyclic neural network according to the gradient values of all layers by using an Adam learning rate self-adaptive algorithm, and updating the weight of each layer by using the error sensitivity of each layer.
The gradient value update formula is as follows:
equation (20) represents the hidden state h(t)The expression of the gradient calculation of (a),in a hidden state h(t)L represents a cost function, and expression (21) represents a cell state C(t)The expression of the gradient calculation of (a),is in a cellular state C(t)The gradient value of (a). Equation (22) represents the weight matrix W in the forgetting gatefThe gradient update expression of (1):
where τ denotes an index value of the current sequence, and the gradient update expression of other parameters is similar to expression (22).
The updating formula of the Adam learning rate self-adaptive algorithm is as follows:
where η represents the step size, set to 0.001,. epsilon.a constant for numerical stability, set to 10-8,denotes vtThe correction value for the first moment estimate,represents mtCorrection values of second-order moment estimates, vtAnd mtThe following were used:
vt＝ρ1vt-1+(1-ρ1)L
mt＝ρ2mt-1+(1-ρ2)L⊙L
where ρ is1Exponential decay Rate, ρ, representing an estimate of the first moment2Exponential decay Rate, ρ, representing an estimate of the second moment1And ρ2in the interval [0,1), L represents a cost function, and L represents a Hadamard product.
A2.5: repeat step A2.2.
The intelligent pick-up phase comprises:
a3: and sequentially inputting the original velocity spectrum into the convolutional neural network training model and the cyclic neural network training model to obtain an accurate time-velocity pair sequence.
The method specifically comprises the following steps:
a3.1: the raw velocity spectrum is subjected to a whitening pre-processing.
A3.2: and inputting the preprocessed velocity spectrum into the convolutional neural network training model, wherein the position corresponding to the energy mass in the processed velocity spectrum is the prediction result of the time-speed pair sequence of the processed velocity spectrum.
A3.3: and inputting the prediction result of the processed time-speed pair sequence of the velocity spectrum into the recurrent neural network training model, correcting the prediction result by the recurrent neural network training model, and finally outputting an accurate time-speed pair sequence.
A3.4: and combining a plurality of velocity spectrum intelligent pickup results of the same measuring line at different positions, and forming a superimposed velocity profile through interpolation.
FIG. 1 is a prior art seismic velocity spectrum and FIG. 3 is a comparison schematic of the velocity spectrum pickup of the present invention.
In this document, the terms front, back, upper and lower are used to define the components in the drawings and the positions of the components relative to each other, and are used for clarity and convenience of the technical solution. It is to be understood that the use of the directional terms should not be taken to limit the scope of the claims.
The features of the embodiments and embodiments described herein above may be combined with each other without conflict.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (8)
1. An artificial intelligence picking method of seismic velocity spectrum based on deep learning is characterized in that: extracting the characteristics of energy clusters in the velocity spectrum through a convolutional neural network based on the characteristics of the velocity spectrum with images, and calculating the prediction result of a time-velocity pair sequence corresponding to the energy clusters in the velocity spectrum from the extracted characteristics; and then based on the time sequence characteristics of the velocity spectrum, extracting the relation characteristics of the time and the velocity of the prediction result by using a recurrent neural network, and correcting the prediction result according to the relation characteristics to finally obtain an accurate time-velocity pair sequence.
2. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 1, characterized in that: the method comprises the following refining steps:
a1: taking the preprocessed original velocity spectrum and the artificial extraction value of the time-velocity pair sequence corresponding to the original velocity spectrum as training samples, training the convolutional neural network, and obtaining a convolutional neural network training model capable of predicting the prediction result;
a2: taking the prediction result and the manually extracted value of the time-speed pair sequence of the original speed spectrum as a new training sample set, training the recurrent neural network, and obtaining a recurrent neural network training model capable of obtaining an accurate time-speed pair sequence;
a3: and sequentially inputting the original velocity spectrum into the convolutional neural network training model and the cyclic neural network training model to obtain an accurate time-velocity pair sequence.
3. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 2, characterized in that: the convolutional neural network comprises a plurality of convolutional layers and a plurality of pooling layers which are arranged in a crossed mode with the convolutional layers, each pooling layer is located behind the convolutional layer corresponding to the pooling layer, the convolutional neural network further comprises a plurality of full-connection layers located behind the last pooling layer, the full-connection layers are arranged in a front-back mode, and the last full-connection layer corresponds to the output of the neural network.
4. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 3, characterized in that: the recurrent neural network comprises an input layer, an output layer and a hidden layer positioned between the input layer and the output layer, wherein the input layer corresponds to the output of the convolutional neural network.
5. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 4, characterized in that: the step of refining step a1 includes:
a1.1: carrying out whitening pretreatment on an original velocity spectrum to obtain a velocity spectrum, wherein the velocity spectrum and a corresponding time-velocity pair sequence thereof form a training data set of a convolutional neural network;
a1.2: randomly shuffling a training data set of the convolutional neural network, grouping the data in the training set, wherein the number of the data in each group is consistent, and sequentially inputting each group into the convolutional neural network, wherein each group of data sequentially passes through a plurality of convolutional layers;
a1.3: extracting and fusing the features on one convolution layer by using a sliding window to obtain convolution feature maps on the convolution layer, wherein each convolution feature map is related to the weight of the convolution layer corresponding to the convolution feature map;
a1.4: inputting the convolution characteristic map into a subsequent pooling layer adjacent to the convolution characteristic map, wherein the pooling layer adopts a maximum pooling algorithm to obtain a pooling characteristic map, and the pooling characteristic map enters another subsequent convolution layer adjacent to the pooling characteristic map;
a1.5: repeating the steps A1.3 and A1.4 until data enters the last convolutional layer, wherein a Dropout layer is inserted in the last convolutional layer, the Dropout layer is used for randomly leading the value of the pooling feature map passing through the Dropout layer to be zero so as to reduce the phenomenon of network overfitting, then the convolutional layer generates a last convolutional feature map, and the convolutional feature map is finally pooled by the last pooling layer to generate a pooling feature map;
a1.6: performing multiple classification identification on the region position of the energy mass of the velocity spectrum based on the pooling characteristic maps of different pooling layers, screening out the real position of the energy mass, and performing multiple detection and inference on the peak value of the energy mass by using a full-connected layer to obtain the actual position of the peak value of the energy mass, namely the prediction result of the time-velocity pair sequence of the velocity spectrum;
a1.7: calculating a cost function of the convolutional neural network, judging whether the cost function of the last full-connection layer is converged, finishing the training of the convolutional neural network if the cost function of the last full-connection layer is converged, and outputting a prediction result of the time-speed pair sequence, otherwise, performing the next step;
a1.8: calculating gradient values of cost functions of all layers in the convolutional neural network by using a back conduction algorithm, calculating error sensitivity of all layers of the convolutional neural network according to the obtained gradient of each layer by using a momentum-based random gradient descent algorithm, and updating the weight of each layer by using the error sensitivity of each layer;
a1.8: and returning to the step 1.2.
6. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 4, characterized in that: the step of refining step a2 includes:
a2.1: the prediction result and the artificially extracted value of the time-speed pair sequence of the original speed spectrum form a training sample set of the recurrent neural network, and the training sample set of the recurrent neural network is input into the recurrent neural network;
a2.2: the recurrent neural network extracts the relation characteristic of time and speed in the prediction result through the neurons of the recurrent neural network, the relation characteristic is related to the weight of the corresponding layer, and a more accurate time-speed pair sequence is obtained by utilizing the relation characteristic;
a2.3: calculating a cost function of the convolutional neural network, judging whether the cost function of the output layer is converged, if so, finishing the training of the convolutional neural network, and if not, performing the next step;
a2.4: calculating gradient values of cost functions of all layers in the cyclic neural network by using a reverse conduction algorithm, calculating error sensitivity of all layers in the cyclic neural network according to the gradient values of all layers by using an Adam learning rate self-adaptive algorithm, and updating the weight of each layer by using the error sensitivity of each layer;
a2.5: repeat step A2.2.
7. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 6, characterized in that: the training sample set of the recurrent neural network is formed on an input layer of the recurrent neural network, a hidden layer of the recurrent neural network is used for correcting the prediction result, an output layer of the recurrent neural network is used for outputting a more accurate time-speed pair sequence, and the recurrent neural network is a long-short term memory network structure which comprises a forgetting gate, an input gate, an output gate and a cell state.
8. The seismic velocity spectrum artificial intelligence picking method based on deep learning of claim 4, characterized in that: the step of refining step a3 includes:
a3.1: carrying out whitening pretreatment on the original velocity spectrum;
a3.2: inputting the preprocessed velocity spectrum into the convolutional neural network training model, wherein the position corresponding to the energy mass in the processed velocity spectrum is the prediction result of the time-speed pair sequence of the processed velocity spectrum;
a3.3: inputting the prediction result of the processed time-speed pair sequence of the velocity spectrum into the recurrent neural network training model, correcting the prediction result by the recurrent neural network training model, and finally outputting an accurate time-speed pair sequence;
a3.4: and combining a plurality of velocity spectrum intelligent pickup results of the same measuring line at different positions, and forming a superimposed velocity profile through interpolation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810260667.3A CN108596327B (en) | 2018-03-27 | 2018-03-27 | Seismic velocity spectrum artificial intelligence picking method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810260667.3A CN108596327B (en) | 2018-03-27 | 2018-03-27 | Seismic velocity spectrum artificial intelligence picking method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108596327A true CN108596327A (en) | 2018-09-28 |
| CN108596327B CN108596327B (en) | 2021-08-03 |
Family
ID=63624703
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810260667.3A Active CN108596327B (en) | 2018-03-27 | 2018-03-27 | Seismic velocity spectrum artificial intelligence picking method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108596327B (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110007366A (en) * | 2019-03-04 | 2019-07-12 | 中国科学院深圳先进技术研究院 | A kind of life searching method and system based on Multi-sensor Fusion |
| CN110471111A (en) * | 2019-09-06 | 2019-11-19 | 中国海洋大学 | A kind of normal-moveout spectrum automatic pick method based on convolutional neural networks |
| CN110688722A (en) * | 2019-10-17 | 2020-01-14 | 深制科技(苏州)有限公司 | Automatic generation method of part attribute matrix based on deep learning |
| CN110826691A (en) * | 2019-10-14 | 2020-02-21 | 中国地质大学（武汉） | Intelligent seismic velocity spectrum pickup method based on YOLO and LSTM |
| CN110889494A (en) * | 2019-11-29 | 2020-03-17 | 北京龙田华远科技有限公司 | Coal mining machine positioning method and device |
| CN111257934A (en) * | 2020-01-17 | 2020-06-09 | 哈尔滨工业大学 | Seismic oscillation peak acceleration prediction method based on second-order neuron deep neural network |
| CN111461172A (en) * | 2020-03-04 | 2020-07-28 | 哈尔滨工业大学 | Lightweight-class feature fusion method of hyperspectral remote sensing data based on two-dimensional point group convolution |
| CN112285776A (en) * | 2020-10-23 | 2021-01-29 | 中国矿业大学(北京) | Seismic velocity automatic picking method based on deep learning |
| CN112540404A (en) * | 2019-09-20 | 2021-03-23 | 中国石油化工股份有限公司 | Automatic speed analysis method and system based on deep learning |
| CN113705084A (en) * | 2021-08-02 | 2021-11-26 | 北京工业大学 | Seismic simulation vibration table closed-loop control method based on deep learning |
| CN116047604A (en) * | 2023-02-23 | 2023-05-02 | 成都理工大学 | Deep seismic phase rapid pickup method based on amplitude statistics and time-frequency analysis |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080133140A1 (en) * | 2006-12-05 | 2008-06-05 | Ralf Ferber | Processing seismic data using interferometry techniques |
| CN102073064A (en) * | 2009-11-25 | 2011-05-25 | 中国石油天然气集团公司 | Method for improving velocity spectrum resolution by using phase information |
| CN102606151A (en) * | 2012-04-01 | 2012-07-25 | 中国石油大学(北京) | Method and device for predicting rock drillability of wildcat well before drilling |
| CN103675900A (en) * | 2012-09-04 | 2014-03-26 | 中国石油天然气集团公司 | Method for determining optimum velocity profile of converted-wave pre-stack time migration during seismic data processing process |
-
2018
- 2018-03-27 CN CN201810260667.3A patent/CN108596327B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080133140A1 (en) * | 2006-12-05 | 2008-06-05 | Ralf Ferber | Processing seismic data using interferometry techniques |
| CN102073064A (en) * | 2009-11-25 | 2011-05-25 | 中国石油天然气集团公司 | Method for improving velocity spectrum resolution by using phase information |
| CN102606151A (en) * | 2012-04-01 | 2012-07-25 | 中国石油大学(北京) | Method and device for predicting rock drillability of wildcat well before drilling |
| CN103675900A (en) * | 2012-09-04 | 2014-03-26 | 中国石油天然气集团公司 | Method for determining optimum velocity profile of converted-wave pre-stack time migration during seismic data processing process |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110007366A (en) * | 2019-03-04 | 2019-07-12 | 中国科学院深圳先进技术研究院 | A kind of life searching method and system based on Multi-sensor Fusion |
| CN110471111A (en) * | 2019-09-06 | 2019-11-19 | 中国海洋大学 | A kind of normal-moveout spectrum automatic pick method based on convolutional neural networks |
| CN112540404B (en) * | 2019-09-20 | 2024-04-12 | 中国石油化工股份有限公司 | Automatic speed analysis method and system based on deep learning |
| CN112540404A (en) * | 2019-09-20 | 2021-03-23 | 中国石油化工股份有限公司 | Automatic speed analysis method and system based on deep learning |
| CN110826691A (en) * | 2019-10-14 | 2020-02-21 | 中国地质大学（武汉） | Intelligent seismic velocity spectrum pickup method based on YOLO and LSTM |
| CN110688722A (en) * | 2019-10-17 | 2020-01-14 | 深制科技(苏州)有限公司 | Automatic generation method of part attribute matrix based on deep learning |
| CN110688722B (en) * | 2019-10-17 | 2023-08-08 | 深制科技(苏州)有限公司 | Automatic generation method of part attribute matrix based on deep learning |
| CN110889494A (en) * | 2019-11-29 | 2020-03-17 | 北京龙田华远科技有限公司 | Coal mining machine positioning method and device |
| CN111257934B (en) * | 2020-01-17 | 2022-03-11 | 哈尔滨工业大学 | Seismic oscillation peak acceleration prediction method based on second-order neuron deep neural network |
| CN111257934A (en) * | 2020-01-17 | 2020-06-09 | 哈尔滨工业大学 | Seismic oscillation peak acceleration prediction method based on second-order neuron deep neural network |
| CN111461172A (en) * | 2020-03-04 | 2020-07-28 | 哈尔滨工业大学 | Lightweight-class feature fusion method of hyperspectral remote sensing data based on two-dimensional point group convolution |
| CN112285776A (en) * | 2020-10-23 | 2021-01-29 | 中国矿业大学(北京) | Seismic velocity automatic picking method based on deep learning |
| CN113705084A (en) * | 2021-08-02 | 2021-11-26 | 北京工业大学 | Seismic simulation vibration table closed-loop control method based on deep learning |
| CN116047604A (en) * | 2023-02-23 | 2023-05-02 | 成都理工大学 | Deep seismic phase rapid pickup method based on amplitude statistics and time-frequency analysis |
| CN116047604B (en) * | 2023-02-23 | 2023-10-17 | 成都理工大学 | Deep seismic phase rapid pickup method based on amplitude statistics and time-frequency analysis |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108596327B (en) | 2021-08-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108596327B (en) | Seismic velocity spectrum artificial intelligence picking method based on deep learning | |
| CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
| CN109816095B (en) | Network flow prediction method based on improved gated cyclic neural network | |
| CN110321603B (en) | Depth calculation model for gas path fault diagnosis of aircraft engine | |
| CN106407649B (en) | Microseismic signals based on time recurrent neural network then automatic pick method | |
| CN109508655A (en) | The SAR target identification method of incomplete training set based on twin network | |
| CN109800875A (en) | Chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine | |
| CN110728698B (en) | Multi-target tracking system based on composite cyclic neural network system | |
| CN106022954B (en) | Multiple BP neural network load prediction method based on grey correlation degree | |
| CN112949828A (en) | Graph convolution neural network traffic prediction method and system based on graph learning | |
| CN110766060B (en) | Time series similarity calculation method, system and medium based on deep learning | |
| CN114548591A (en) | Time sequence data prediction method and system based on hybrid deep learning model and Stacking | |
| CN110276441A (en) | A kind of trapezoidal overlap kernel impulse response estimation method based on deep learning | |
| CN115689008A (en) | CNN-BilSTM short-term photovoltaic power prediction method and system based on ensemble empirical mode decomposition | |
| CN111723523A (en) | Estuary residual water level prediction method based on cascade neural network | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| CN114166509A (en) | Motor bearing fault prediction method | |
| CN113935489A (en) | Variational quantum model TFQ-VQA based on quantum neural network and two-stage optimization method thereof | |
| CN110223342B (en) | Space target size estimation method based on deep neural network | |
| CN114357372A (en) | Aircraft fault diagnosis model generation method based on multi-sensor data driving | |
| CN114596726A (en) | Parking position prediction method based on interpretable space-time attention mechanism | |
| CN115761654B (en) | Vehicle re-identification method | |
| CN105740815A (en) | Human body behavior identification method based on deep recursive and hierarchical condition random fields | |
| CN115083511A (en) | Peripheral gene regulation and control feature extraction method based on graph representation learning and attention | |
| CN113361476B (en) | Zhang Heng one-number pre-earthquake abnormal signal identification method based on artificial intelligence technology |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |