CN108169745A - A kind of borehole radar target identification method based on convolutional neural networks - Google Patents
A kind of borehole radar target identification method based on convolutional neural networks Download PDFInfo
- Publication number
- CN108169745A CN108169745A CN201711365529.3A CN201711365529A CN108169745A CN 108169745 A CN108169745 A CN 108169745A CN 201711365529 A CN201711365529 A CN 201711365529A CN 108169745 A CN108169745 A CN 108169745A
- Authority
- CN
- China
- Prior art keywords
- layer
- training sample
- neural networks
- output
- convolutional neural
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/885—Radar or analogous systems specially adapted for specific applications for ground probing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Radar, Positioning & Navigation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Remote Sensing (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Electromagnetism (AREA)
- Probability & Statistics with Applications (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Operations Research (AREA)
- Computer Networks & Wireless Communication (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
The invention discloses a kind of borehole radar target identification method based on convolutional neural networks and machine learning, deep learning application technology, more particularly to application of the deep learning method in borehole radar target identification.Including stream：Sample；Convolutional neural networks model structure designs, network parameter design, over-fitting prevention, activation primitive design, convolutional neural networks model training, target identification.Clutter can be automatically extracted in training process of the present invention, ambient noise, chaff interferent and clarification of objective, greatly reduce expense needed for the selection and extraction of feature, target can be effectively identified in real time, while depth model can extract the two dimensional character of target different levels in the present invention, these features are obtained by the matching of height with training, therefore it can realize the height characterization of target, improve the accuracy of borehole radar echo signal identification.
Description
Technical field
The present invention relates to machine learning, deep learning application technologies, and more particularly to deep learning method is in borehole radar mesh
Application in identifying not.
Background technology
Borehole radar is also known as radar in hole, is a kind of special shape of Ground Penetrating Radar, using pulse signal pulsewidth compared with
Wide therefore frequency spectrum is in relatively low working frequency (10~1000MHz) range, can transfer in wellhole, completes to well week ground
The exploration of matter environment.It can closely descend further object body, have the spy that radial direction detection range is remote and resolution ratio is relatively high
Point.By its technical advantage, borehole radar has been widely used for geologic structure exploration, mining industry, the hydrology and environmental geology
Investigation, civil engineering, nuke rubbish storage site addressing etc..With increasingly increase of the mankind to the subterranean resource demand such as oil, gas,
People descend the research of Detection Techniques to increase increasingly over the ground, provide the underground echo signal processing method of real-time high-efficiency a kind of into
The current active demand studied both at home and abroad.
The processing means of Coherent Noise in GPR Record depend on imaging technique and inversion technique at present.Imaging technique and inverting
Technology, containing more subjective factor, needs operator to have higher technology and experience generally by manually being explained.Imaging technique with
Inversion technique reproduce the shape for burying object, so as to be differentiated according to geometric properties to target.But due to excessively according to
Rely in geometric properties, have ignored original hiding characteristic information in signal, easily misrecognition and the close interference of target shape
Object.In addition imaging technique and inversion technique are computationally intensive, it is difficult to real time discriminating target conditions.Therefore how the identification of real-time high-efficiency
Target echo signal, into it is current in the urgent need to address the problem of.
A kind of borehole radar target identification method based on convolutional neural networks can automatically extract target spy from data
Sign can excavate more and more abstract feature from original input data.Clutter, ambient noise, chaff interferent are extracted by training
And clarification of objective.Network after training can real-time high-efficiency identification target, can be widely applied to borehole radar Data processing.
Invention content
The goal of the invention of the present invention is:In order to overcome in the prior art to the shortcoming of borehole radar target identification,
To reach the more preferable real-time of borehole radar target identification, more high accurancy and precision provides a kind of drilling based on convolutional neural networks
Radar target identification method.
Neural network be it is a kind of be self-adaptation nonlinear dynamic network by a large amount of neurons by being linked to each other to form
System.Neural network has multiple hierarchical structures, and the convolutional layer of higher extracts more advanced feature.With the increase of the number of plies,
Neural network is stronger to the abstracting power of feature, i.e., to the deeper characterization of target.Have benefited from the deep layer table of neural network
Sign ability without being pre-processed to echo data, can substantially reduce the time overhead in terms of pretreatment.Meanwhile neural network
It, can the perfect information environment for being suitable for borehole radar complexity with good learning ability.
Technical solution of the present invention is:A kind of borehole radar target identification method based on convolutional neural networks, this method packet
It includes：
Step 1:It obtains and forms training sample set as the radar return data of training sample, wherein training sample set includes
Different classes of identification target, with the classification number of T mark training samples；
The acquisition modes of borehole radar data as training sample are goal-orientation, by adjacent X roads echo data
The 2-D data formed obtains the identical training sample set of size, and sets class for each training sample according to target classification
Other identifier；Adjacent road number X should be less than target size, and one unit of each transverse shifting in center forms one group of new size phase
Same training sample set；
Step 2:Build convolutional neural networks；
The convolutional neural networks depth is M layers, wherein 1~M-1 layers of each layer includes convolutional layer and pond layer, the
M layers are the output layer connected entirely, export the probability matrix for classification results；Training sample is inputted from first layer, last layer nerve
The output of network is as next layer of input, and the size of the convolutional layer convolution kernel is less than the size of input data, the pond
Layer is that convolutional layer is exported to carry out the processing of average value pondization, takes the local mean values when forebay window as the defeated of current window
Go out, the region entirety is replaced with the local mean values in pond region；
Further, the convolution algorithm method of the convolutional layer is：
Si'j′' represent in a manner of the sliding window of default step-length 1 to input data SijIt carries out convolution and obtains the output of corresponding position,
I represents the i-th echo, and j represents j-th point of echo, wnmRepresent convolution filter line n m row parameters；Adjust the size control of w
The size of convolution kernel processed；
The pond method of the pond layer is：The overall output in region is replaced using the local draw value in region：
Wherein eijThe i-th echo, j-th of data value are represented, n is matrix size, and ave [] represents the average value of the matrix,
e0For output valve；Since the network model of deep layer has a fairly large number of convolutional layer, a large amount of parameter can be generated, is not only had big
The redundancy of amount influences computational efficiency, is also easy to the over-fitting for causing neural network.In order to reduce model parameter quantity, improve
Receptive field avoids over-fitting, and the method that we use pond can not only propose a large amount of redundancy parameter, can also reduce
Fitting.
Described M layers connect layer to be complete, i.e., are weighted summation to each element of matrix of M-1 layers of output,Sizes of the wherein P × Q for M-1 layers of output, xiSubscript be used to identify the difference of same training sample
Probability corresponding to classification results, knmFor the parameter that M layers of output layer line n m are arranged, enmFor M-1 layers of output matrix
The element of n rows m row.
Further, activation primitive is set after each convolutional layer, by input of the output as pond layer of activation primitive,
The activation primitive is f (x)=max (0, x), represents each element exported for convolution, takes itself and the maximal term conduct in 0
As a result.
Step 3:Training convolutional neural networks：
Step 301:Setting terminates threshold value, sets learning rate, sets sub- training sample set size；
Step 302:It is sub- training sample set to be concentrated from training sample and randomly choose Ν training sample, and random initializtion is each
Layer convolution kernel initial value, the M layers of output based on depth model obtain the eigenvectors matrix X of each training sample；Successively push away
Calculate the error amount δ of each layer deconvolution parameter：The error amount of M layers of convolution layer parameter is F-X, and desired output F is preset value；Layer afterwards
Error is obtained by the product of the error amount of last layer and the parameter of convolution kernel, wnmConvolution filter line n m row parameters are represented,
Wherein n=1,2 ..., w, m=1,2 ..., w, w represent the size of convolution kernel；Convolution kernel weights are changed with gradient descent method to reduce
Error, the more new formula of convolution layer parameter are as follows：
Wherein a represents learning rate；
Step 303:The cost function of current convolutional neural networks classification results is calculated, i.e. current class result is divided with practical
The mapping of class resultant error, judges whether the knots modification of cost function reaches end threshold value, if so, performing step 304；It is no
Then, then step 302 is performed；
Step 304:Each convolution layer parameter is preserved, obtains the convolutional neural networks of training completion；
Step 4:Input borehole radar data to be identified, goal-orientation, two be made of adjacent X roads echo data
Dimension data obtains the data to be identified of size identical with training sample；Data to be identified are inputted to the convolutional Neural of training completion
Network exports the probability matrix of data to be identified.
Further, activation primitive is set after each convolutional layer in the step 2, by the output of activation primitive as pond
Change the input of layer, the activation primitive is softmax functions；
It is additional to Softmax Parameters in Regression Model θ in step 302j(j=1,2 ... T) is iterated update；
It is primarily based on the category probability matrix h that Softmax regression models calculate each eigenvectors matrix Xθ(x)：
Wherein p (y=t | X, θ) it represents to be predicted as the probability value of a certain classification, vectorial θ=(θ1,θ2,…,θT), it is initial
It is worth for random initializtion, y represents classification recognition result, and e represents the nature truth of a matter,It represents about θjMatrix transposition；
N number of training sample of current iteration is expressed as：(X(1),y(1)),(X(2),y(2)),(X(3),y(3))...(X(N),y(N)), wherein X(i)Represent the eigenvectors matrix of i-th of training sample, this feature vector matrix is by the final of convolutional neural networks
Output obtains, y(i)Represent corresponding X(i)Classification logotype, i.e. y(i)=1,2 ..., T, based on N number of (X(i),y(i)) calculate cross entropy
Function：
Intersect entropy function：
Wherein：x(i)Represent input sample data, m represents total number of samples；
Intersect the cost function of entropy function：
Declined by gradient and calculated
Method realizes the minimum of J (θ)；
It willProduct with learning rate a is as Parameters in Regression Model correction amount：I.e. under
During secondary iteration, using last correction amount as the Parameters in Regression Model of current iteration.
The probability matrix is belonging respectively to the class probability of target to be sorted, and the recognition result of target is corresponding for maximum probability
Classification.
In conclusion the benefit that the present invention can be brought is：Radar initial data can be directly handled, does not need to pre-process, it can
Clutter, ambient noise, chaff interferent and clarification of objective are automatically extracted, greatly reduces and is opened needed for the selection and extraction of feature
Pin, can effectively identify target in real time.
Description of the drawings
Fig. 1 is depth model structure diagram.
Fig. 2 is convolutional layer schematic diagram.
Fig. 3 is average value pond schematic diagram.
Fig. 4 is radar imagery result.
Fig. 5 is target identification result.
Specific embodiment
The present invention is realized using network structure as shown in Figure 1, wherein totally 3 layers of convolutional layer, 3 layers of pond layer, the convolution
Layer built-in activation function.Convolutional layer carries out convolution to input data in a manner of the sliding window of default step-length 1 and obtains the defeated of corresponding position
Go out, as shown in Fig. 2-b；Pond layer exports convolution and carries out dimension-reduction treatment:It takes and works as when the local mean values of forebay window are used as
The output of front window, as shown in Figure 3.7th layer is the output layer connected entirely, using softmax activation primitives.
A kind of borehole radar target identification method based on convolutional neural networks, this method include：
Step 1:It obtains and forms training sample set as the radar return data of training sample, wherein training sample set includes
Different classes of identification target, with the classification number of T mark training samples；
The acquisition modes of borehole radar data as training sample are goal-orientation, by adjacent X roads echo data
The 2-D data formed obtains the identical training sample set of size, and sets class for each training sample according to target classification
Other identifier；Adjacent road number X should be less than target size, and one unit of each transverse shifting in center forms one group of new size phase
Same training sample set；
Step 2:Build convolutional neural networks；
The convolutional neural networks depth is M layers, wherein 1~M-1 layers of each layer includes convolutional layer and pond layer, the
M layers are the output layer connected entirely, export the probability matrix for classification results；Training sample is inputted from first layer, last layer nerve
The output of network is as next layer of input, and the size of the convolutional layer convolution kernel is less than the size of input data, the pond
Layer is that convolutional layer is exported to carry out the processing of average value pondization, takes the local mean values when forebay window as the defeated of current window
Go out, the region entirety is replaced with the local mean values in pond region；
Further, the convolution algorithm method of the convolutional layer is：
Si'j′' represent in a manner of the sliding window of default step-length 1 to input data SijIt carries out convolution and obtains the output of corresponding position,
I represents the i-th echo, and j represents j-th point of echo, wnmRepresent convolution filter line n m row parameters；Adjust the size control of w
The size of convolution kernel processed；
The pond method of the pond layer is：The overall output in region is replaced using the local draw value in region：
Wherein eijThe i-th echo, j-th of data value are represented, n is matrix size, and ave [] represents the average value of the matrix,
e0For output valve；Since the network model of deep layer has a fairly large number of convolutional layer, a large amount of parameter can be generated, is not only had big
The redundancy of amount influences computational efficiency, is also easy to the over-fitting for causing neural network.In order to reduce model parameter quantity, improve
Receptive field avoids over-fitting, and the method that we use pond can not only propose a large amount of redundancy parameter, can also reduce
Fitting.
Described M layers connect layer to be complete, i.e., are weighted summation to each element of matrix of M-1 layers of output,Sizes of the wherein P × Q for M-1 layers of output, xiSubscript be used to identify the difference of same training sample
Probability corresponding to classification results, knmFor the parameter that M layers of output layer line n m are arranged, enmFor M-1 layers of output matrix
The element of n rows m row.
Further, activation primitive is set after each convolutional layer, by input of the output as pond layer of activation primitive,
The activation primitive is f (x)=max (0, x), represents each element exported for convolution, takes itself and the maximal term conduct in 0
As a result.
Step 3:Training convolutional neural networks：
Step 301:Setting terminates threshold value, sets learning rate, sets sub- training sample set size；
Step 302:It is sub- training sample set to be concentrated from training sample and randomly choose Ν training sample, and random initializtion is each
Layer convolution kernel initial value, the M layers of output based on depth model obtain the eigenvectors matrix X of each training sample；Successively push away
Calculate the error amount δ of each layer deconvolution parameter：The error amount of M layers of convolution layer parameter is F-X, and desired output F is preset value；Layer afterwards
Error is obtained by the product of the error amount of last layer and the parameter of convolution kernel, wnmConvolution filter line n m row parameters are represented,
Wherein n=1,2 ..., w, m=1,2 ..., w, w represent the size of convolution kernel；Convolution kernel weights are changed with gradient descent method to reduce
Error, the more new formula of convolution layer parameter are as follows：
Wherein a represents learning rate；
Step 303:The cost function of current convolutional neural networks classification results is calculated, i.e. current class result is divided with practical
The mapping of class resultant error, judges whether the knots modification of cost function reaches end threshold value, if so, performing step 304；It is no
Then, then step 302 is performed；
Step 304:Each convolution layer parameter is preserved, obtains the convolutional neural networks of training completion；
Step 4:Input borehole radar data to be identified, goal-orientation, two be made of adjacent X roads echo data
Dimension data obtains the data to be identified of size identical with training sample；Data to be identified are inputted to the convolutional Neural of training completion
Network exports the probability matrix of data to be identified.
Further, activation primitive is set after each convolutional layer in the step 2, by the output of activation primitive as pond
Change the input of layer, the activation primitive is softmax functions；
It is additional to Softmax Parameters in Regression Model θ in step 302j(j=1,2 ... T) is iterated update；
It is primarily based on the category probability matrix h that Softmax regression models calculate each eigenvectors matrix Xθ(x)：
Wherein p (y=t | X, θ) it represents to be predicted as the probability value of a certain classification, vectorial θ=(θ1,θ2,…,θT), it is initial
It is worth for random initializtion, y represents classification recognition result, and e represents the nature truth of a matter,It represents about θjMatrix transposition；
N number of training sample of current iteration is expressed as：(X(1),y(1)),(X(2),y(2)),(X(3),y(3))...(X(N),y(N)), wherein X(i)Represent the eigenvectors matrix of i-th of training sample, this feature vector matrix is by the final of convolutional neural networks
Output obtains, y(i)Represent corresponding X(i)Classification logotype, i.e. y(i)=1,2 ..., T, based on N number of (X(i),y(i)) calculate cross entropy
Function：
Intersect entropy function：
Wherein：x(i)Represent input sample data, m represents total number of samples；
Intersect the cost function of entropy function：
Declined by gradient and calculated
Method realizes the minimum of J (θ)；
It willProduct with learning rate a is as Parameters in Regression Model correction amount：I.e. under
During secondary iteration, using last correction amount as the Parameters in Regression Model of current iteration.
In the present embodiment, the data source of training sample is in sinopec Logging Company measured data.As shown in figure 4, image
In have very big noise and band interference, by naked eyes can not intuitively judge target.
By goal-orientation, the 2-D data being made of adjacent X roads echo data obtains the identical training of size
Sample, and category identifier is set for each training sample according to target classification, n times target's center is slided, obtains being trained by N groups
The training sample set that sample is formed.
Training set is input to neural network model level 1 volume lamination, successively calculates error amount and with new network as input
Weights.After the completion of training, current network weights are preserved to predict the recognition effect of test sample.
Based on the neural network model that training is completed, input test sample carries out target identification test, defeated in the present embodiment
Go out the class probability matrix that layer calculates the data to be identified of neural network output using Softmax regression models, take maximum probability
Corresponding classification is target identification result.For measured data collection, the detection to targets such as hole, gaps, accuracy rate can reach
To 92.235% discrimination.
Radar return data have very big noise and band interference signal.By conventional Radar Imaging Processing, such as Fig. 4 institutes
Show, people are difficult intuitive, accurately judge the situation in the region well week.After the processing method of this patent, as shown in figure 5,
It clearly can accurately judge well week situation.
The above description is merely a specific embodiment, any feature disclosed in this specification, except non-specifically
Narration, can be replaced by other alternative features that are equivalent or have similar purpose；Disclosed all features or all sides
Method or in the process the step of, other than mutually exclusive feature and/or step, can be combined in any way.
Claims (4)
1. a kind of borehole radar target identification method based on convolutional neural networks, this method include：
Step 1:It obtains and forms training sample set as the radar return data of training sample, wherein training sample set includes difference
The identification target of classification, with the classification number of T mark training samples；
The acquisition modes of borehole radar data as training sample are goal-orientation, by adjacent X roads echo data institute structure
Into 2-D data, obtain the identical training sample set of size, and classification mark is set for each training sample according to target classification
Know symbol；Adjacent road number X should be less than target size, and it is identical that one unit of each transverse shifting in center forms one group of new size
Training sample set；
Step 2:Build convolutional neural networks；
The convolutional neural networks depth be M layer, wherein 1~M-1 layers of each layer include convolutional layer and pond layer, M layers
For the output layer connected entirely, the probability matrix for classification results is exported；Training sample is inputted from first layer, last layer neural network
Output as next layer of input, less than the size of input data, the pond layer is for the size of the convolutional layer convolution kernel
Convolutional layer is exported and carries out the processing of average value pondization, takes output of the local mean values when forebay window as current window,
The region entirety is replaced with the local mean values in pond region；
Step 3:Training convolutional neural networks：
Step 301:Setting terminates threshold value, sets learning rate, sets sub- training sample set size；
Step 302:It is sub- training sample set to be concentrated from training sample and randomly choose Ν training sample, and each layer of random initializtion is rolled up
Product core initial value, the M layers of output based on depth model obtain the eigenvectors matrix X of each training sample；It successively calculates each
The error amount δ of layer deconvolution parameter：The error amount of M layers of convolution layer parameter is F-X, and desired output F is preset value；The error of layer afterwards
It is obtained by the product of the parameter of the error amount and convolution kernel of last layer, wnmRepresent convolution filter line n m row parameters, wherein n
=1,2 ..., w, m=1,2 ..., w, w represent the size of convolution kernel；Convolution kernel weights are changed with gradient descent method to reduce error,
The more new formula of convolution layer parameter is as follows：
Wherein a represents learning rate；
Step 303:Calculate the cost function of current convolutional neural networks classification results, i.e. current class result and actual classification knot
The mapping of fruit error, judges whether the knots modification of cost function reaches end threshold value, if so, performing step 304；Otherwise, then
Perform step 302；
Step 304:Each convolution layer parameter is preserved, obtains the convolutional neural networks of training completion；
Step 4:Input borehole radar data to be identified, goal-orientation, the two-dimemsional number being made of adjacent X roads echo data
According to obtaining the data to be identified of size identical with training sample；Data to be identified are inputted to the convolutional neural networks of training completion,
Export the probability matrix of data to be identified.
2. a kind of borehole radar target identification method based on convolutional neural networks as described in claim 1, it is characterised in that
The convolution algorithm method of convolutional layer is in the step 2：
Si'j'' represent in a manner of the sliding window of default step-length 1 to input data SijIt carries out convolution and obtains the output of corresponding position, i generations
The i-th echo of table, j represent j-th point of echo, wnmRepresent convolution filter line n m row parameters；Adjust the size control volume of w
The size of product core；
The pond method of the pond layer is：The overall output in region is replaced using the local draw value in region：
Wherein eijThe i-th echo, j-th of data value are represented, n is matrix size, and ave [] represents the average value of the matrix, e0For
Output valve；
Described M layers connect layer to be complete, i.e., are weighted summation to each element of matrix of M-1 layers of output,
Sizes of the wherein P × Q for M-1 layers of output, xiSubscript for identifying corresponding to the different classifications result of same training sample
Probability, knmFor the parameter that M layers of output layer line n m are arranged, enmThe member of line n m row for M-1 layers of output matrix
Element.
3. a kind of borehole radar target identification method based on convolutional neural networks as described in claim 1, it is characterised in that
Activation primitive is set after each convolutional layer in the step 2, it is described to swash by input of the output of activation primitive as pond layer
Function living is f (x)=max (0, x), represents each element exported for convolution, take its with the maximal term in 0 as a result.
4. a kind of borehole radar target identification method based on convolutional neural networks as described in claim 1, it is characterised in that
Activation primitive is set after each convolutional layer in the step 2, it is described to swash by input of the output of activation primitive as pond layer
Function living is softmax functions；
It is additional to Softmax Parameters in Regression Model θ in step 302j(j=1,2 ... T) is iterated update；
It is primarily based on the category probability matrix h that Softmax regression models calculate each eigenvectors matrix Xθ(x)：
Wherein p (y=t | X, θ) it represents to be predicted as the probability value of a certain classification, vectorial θ=(θ1,θ2,…,θT), initial value is
Random initializtion, y represent classification recognition result, and e represents the nature truth of a matter,It represents about θjMatrix transposition；
N number of training sample of current iteration is expressed as：(X(1),y(1)),(X(2),y(2)),(X(3),y(3))...(X(N),y(N)),
Middle X(i)Represent the eigenvectors matrix of i-th of training sample, this feature vector matrix is obtained by the final output of convolutional neural networks
It arrives, y(i)Represent corresponding X(i)Classification logotype, i.e. y(i)=1,2 ..., T, based on N number of (X(i),y(i)) calculate and intersect entropy function：
Intersect entropy function：
Wherein：x(i)Represent input sample data, m represents total number of samples；
Intersect the cost function of entropy function：
Pass through gradient descent algorithm reality
The minimum of existing J (θ)；
It willProduct with learning rate a is as Parameters in Regression Model correction amount：Change in next time
Dai Shi, using last correction amount as the Parameters in Regression Model of current iteration.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711365529.3A CN108169745A (en) | 2017-12-18 | 2017-12-18 | A kind of borehole radar target identification method based on convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711365529.3A CN108169745A (en) | 2017-12-18 | 2017-12-18 | A kind of borehole radar target identification method based on convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108169745A true CN108169745A (en) | 2018-06-15 |
Family
ID=62522322
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711365529.3A Pending CN108169745A (en) | 2017-12-18 | 2017-12-18 | A kind of borehole radar target identification method based on convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108169745A (en) |
Cited By (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
| CN109407067A (en) * | 2018-10-13 | 2019-03-01 | 中国人民解放军海军航空大学 | Radar moving targets detection and classification integral method based on time-frequency figure convolutional neural networks |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN110288082A (en) * | 2019-06-05 | 2019-09-27 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
| CN110516566A (en) * | 2019-08-08 | 2019-11-29 | 北京邮电大学 | A kind of filtering method and device based on convolutional layer |
| CN110807428A (en) * | 2019-11-05 | 2020-02-18 | 南方科技大学 | Coal sample identification method and device, server and storage medium |
| CN111025290A (en) * | 2019-12-31 | 2020-04-17 | 电子科技大学 | High-efficiency pre-imaging fusion method and system for drilling radar |
| CN111079559A (en) * | 2019-11-26 | 2020-04-28 | 成都深地领航能源科技有限公司 | Convolutional neural network-based mud pulse signal identification method |
| CN111191583A (en) * | 2019-12-30 | 2020-05-22 | 郑州科技学院 | Space target identification system and method based on convolutional neural network |
| CN111353391A (en) * | 2020-02-17 | 2020-06-30 | 西安电子科技大学 | Radar interference effect evaluation method and device, electronic equipment and storage medium thereof |
| CN111680781A (en) * | 2020-04-20 | 2020-09-18 | 北京迈格威科技有限公司 | Neural network processing method, neural network processing device, electronic equipment and storage medium |
| CN111752754A (en) * | 2020-06-05 | 2020-10-09 | 清华大学 | Method for recovering radar image data in memory |
| CN111948638A (en) * | 2020-07-15 | 2020-11-17 | 南京航空航天大学 | Target length estimation method based on high-resolution range profile and application thereof |
| CN112001270A (en) * | 2020-08-03 | 2020-11-27 | 南京理工大学 | Ground radar automatic target classification and identification method based on one-dimensional convolutional neural network |
| CN112085153A (en) * | 2019-06-12 | 2020-12-15 | 西门子医疗有限公司 | Providing an output signal and providing a trained function by means of a touch-sensitive input unit |
| CN112099014A (en) * | 2020-08-24 | 2020-12-18 | 广东工业大学 | Road millimeter wave noise model detection and estimation method based on deep learning |
| CN112232392A (en) * | 2020-09-29 | 2021-01-15 | 深圳安德空间技术有限公司 | Data interpretation and identification method for three-dimensional ground penetrating radar |
| CN113095354A (en) * | 2021-03-03 | 2021-07-09 | 电子科技大学 | Unknown radar target identification method based on radiation source characteristic subspace knowledge |
| CN113341391A (en) * | 2021-06-01 | 2021-09-03 | 电子科技大学 | Radar target multi-frame joint detection method in unknown environment based on deep learning |
| CN113609970A (en) * | 2021-08-03 | 2021-11-05 | 西北工业大学 | Underwater target identification method based on grouping convolution depth U _ Net |
| CN113687326A (en) * | 2021-07-13 | 2021-11-23 | 广州杰赛科技股份有限公司 | Vehicle-mounted radar echo noise reduction method, device, equipment and medium |
| CN114049551A (en) * | 2021-10-22 | 2022-02-15 | 南京航空航天大学 | ResNet 18-based SAR raw data target identification method |
| CN116821651A (en) * | 2023-05-04 | 2023-09-29 | 广州海洋地质调查局 | Method and device for setting natural gas hydrate distant view zone and finely identifying ore body |
| CN116520277B (en) * | 2023-04-26 | 2023-12-08 | 中国人民解放军93209部队 | Radar target recognition method and system based on separated convolution network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718957A (en) * | 2016-01-26 | 2016-06-29 | 西安电子科技大学 | Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network |
| CN106407931A (en) * | 2016-09-19 | 2017-02-15 | 杭州电子科技大学 | Novel deep convolution neural network moving vehicle detection method |
| CN106407986A (en) * | 2016-08-29 | 2017-02-15 | 电子科技大学 | Synthetic aperture radar image target identification method based on depth model |
| CN106710293A (en) * | 2016-10-31 | 2017-05-24 | 中原智慧城市设计研究院有限公司 | Dynamic and intelligent vehicle guidance method for underground parking lot |
| US20170337471A1 (en) * | 2016-05-18 | 2017-11-23 | Nec Laboratories America, Inc. | Passive pruning of filters in a convolutional neural network |
-
2017
- 2017-12-18 CN CN201711365529.3A patent/CN108169745A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105718957A (en) * | 2016-01-26 | 2016-06-29 | 西安电子科技大学 | Polarized SAR image classification method based on nonsubsampled contourlet convolutional neural network |
| US20170337471A1 (en) * | 2016-05-18 | 2017-11-23 | Nec Laboratories America, Inc. | Passive pruning of filters in a convolutional neural network |
| CN106407986A (en) * | 2016-08-29 | 2017-02-15 | 电子科技大学 | Synthetic aperture radar image target identification method based on depth model |
| CN106407931A (en) * | 2016-09-19 | 2017-02-15 | 杭州电子科技大学 | Novel deep convolution neural network moving vehicle detection method |
| CN106710293A (en) * | 2016-10-31 | 2017-05-24 | 中原智慧城市设计研究院有限公司 | Dynamic and intelligent vehicle guidance method for underground parking lot |
Non-Patent Citations (2)
| Title |
|---|
| 张昭昭等: "《模块化神经网络结构分析与设计》", 30 April 2014 * |
| 艾玲梅等: "基于循环卷积神经网络的目标检测与分类", 《计算机技术与发展》 * |
Cited By (35)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109407067A (en) * | 2018-10-13 | 2019-03-01 | 中国人民解放军海军航空大学 | Radar moving targets detection and classification integral method based on time-frequency figure convolutional neural networks |
| CN109444863A (en) * | 2018-10-23 | 2019-03-08 | 广西民族大学 | A kind of estimation method of the narrowband ultrasonic echo number based on convolutional neural networks |
| CN109375186A (en) * | 2018-11-22 | 2019-02-22 | 中国人民解放军海军航空大学 | Radar target identification method based on the multiple dimensioned one-dimensional convolutional neural networks of depth residual error |
| CN110288082A (en) * | 2019-06-05 | 2019-09-27 | 北京字节跳动网络技术有限公司 | Convolutional neural networks model training method, device and computer readable storage medium |
| CN112085153A (en) * | 2019-06-12 | 2020-12-15 | 西门子医疗有限公司 | Providing an output signal and providing a trained function by means of a touch-sensitive input unit |
| CN110516566B (en) * | 2019-08-08 | 2021-07-27 | 北京邮电大学 | Filtering method and device based on convolutional layer |
| CN110516566A (en) * | 2019-08-08 | 2019-11-29 | 北京邮电大学 | A kind of filtering method and device based on convolutional layer |
| CN110807428A (en) * | 2019-11-05 | 2020-02-18 | 南方科技大学 | Coal sample identification method and device, server and storage medium |
| CN110807428B (en) * | 2019-11-05 | 2023-05-30 | 南方科技大学 | Coal sample identification method, device, server and storage medium |
| CN111079559A (en) * | 2019-11-26 | 2020-04-28 | 成都深地领航能源科技有限公司 | Convolutional neural network-based mud pulse signal identification method |
| CN111191583A (en) * | 2019-12-30 | 2020-05-22 | 郑州科技学院 | Space target identification system and method based on convolutional neural network |
| CN111191583B (en) * | 2019-12-30 | 2023-08-25 | 郑州科技学院 | Space target recognition system and method based on convolutional neural network |
| CN111025290A (en) * | 2019-12-31 | 2020-04-17 | 电子科技大学 | High-efficiency pre-imaging fusion method and system for drilling radar |
| CN111353391A (en) * | 2020-02-17 | 2020-06-30 | 西安电子科技大学 | Radar interference effect evaluation method and device, electronic equipment and storage medium thereof |
| CN111680781A (en) * | 2020-04-20 | 2020-09-18 | 北京迈格威科技有限公司 | Neural network processing method, neural network processing device, electronic equipment and storage medium |
| CN111680781B (en) * | 2020-04-20 | 2023-07-25 | 北京迈格威科技有限公司 | Neural network processing method and device, electronic equipment and storage medium |
| CN111752754A (en) * | 2020-06-05 | 2020-10-09 | 清华大学 | Method for recovering radar image data in memory |
| CN111948638B (en) * | 2020-07-15 | 2024-01-09 | 南京航空航天大学 | Target length estimation method based on high-resolution range profile and application thereof |
| CN111948638A (en) * | 2020-07-15 | 2020-11-17 | 南京航空航天大学 | Target length estimation method based on high-resolution range profile and application thereof |
| CN112001270A (en) * | 2020-08-03 | 2020-11-27 | 南京理工大学 | Ground radar automatic target classification and identification method based on one-dimensional convolutional neural network |
| CN112001270B (en) * | 2020-08-03 | 2022-10-14 | 南京理工大学 | Ground radar automatic target classification and identification method based on one-dimensional convolutional neural network |
| CN112099014A (en) * | 2020-08-24 | 2020-12-18 | 广东工业大学 | Road millimeter wave noise model detection and estimation method based on deep learning |
| CN112099014B (en) * | 2020-08-24 | 2023-08-22 | 广东工业大学 | Road millimeter wave noise model detection estimation method based on deep learning |
| CN112232392A (en) * | 2020-09-29 | 2021-01-15 | 深圳安德空间技术有限公司 | Data interpretation and identification method for three-dimensional ground penetrating radar |
| CN112232392B (en) * | 2020-09-29 | 2022-03-22 | 深圳安德空间技术有限公司 | Data interpretation and identification method for three-dimensional ground penetrating radar |
| CN113095354A (en) * | 2021-03-03 | 2021-07-09 | 电子科技大学 | Unknown radar target identification method based on radiation source characteristic subspace knowledge |
| CN113341391B (en) * | 2021-06-01 | 2022-05-10 | 电子科技大学 | Radar target multi-frame joint detection method in unknown environment based on deep learning |
| CN113341391A (en) * | 2021-06-01 | 2021-09-03 | 电子科技大学 | Radar target multi-frame joint detection method in unknown environment based on deep learning |
| CN113687326A (en) * | 2021-07-13 | 2021-11-23 | 广州杰赛科技股份有限公司 | Vehicle-mounted radar echo noise reduction method, device, equipment and medium |
| CN113687326B (en) * | 2021-07-13 | 2024-01-05 | 广州杰赛科技股份有限公司 | Vehicle-mounted radar echo noise reduction method, device, equipment and medium |
| CN113609970A (en) * | 2021-08-03 | 2021-11-05 | 西北工业大学 | Underwater target identification method based on grouping convolution depth U _ Net |
| CN114049551B (en) * | 2021-10-22 | 2022-08-05 | 南京航空航天大学 | ResNet 18-based SAR raw data target identification method |
| CN114049551A (en) * | 2021-10-22 | 2022-02-15 | 南京航空航天大学 | ResNet 18-based SAR raw data target identification method |
| CN116520277B (en) * | 2023-04-26 | 2023-12-08 | 中国人民解放军93209部队 | Radar target recognition method and system based on separated convolution network |
| CN116821651A (en) * | 2023-05-04 | 2023-09-29 | 广州海洋地质调查局 | Method and device for setting natural gas hydrate distant view zone and finely identifying ore body |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108169745A (en) | A kind of borehole radar target identification method based on convolutional neural networks | |
| CN108182450A (en) | A kind of airborne Ground Penetrating Radar target identification method based on depth convolutional network | |
| Zhu et al. | Intelligent logging lithological interpretation with convolution neural networks | |
| CN103955702B (en) | SAR image terrain classification method based on depth RBF network | |
| CN104616664B (en) | A kind of audio identification methods detected based on sonograph conspicuousness | |
| CN104570083B (en) | Geologic body automatic identifying method based on multi-dimensional earthquake attribute | |
| CN103487832B (en) | Supervision waveform classification is had in a kind of 3-D seismics signal | |
| CN104408481B (en) | Classification of Polarimetric SAR Image method based on depth wavelet neural network | |
| CN105975931A (en) | Convolutional neural network face recognition method based on multi-scale pooling | |
| CN106874956A (en) | The construction method of image classification convolutional neural networks structure | |
| CN104008551B (en) | A kind of Citrus Huanglongbing pathogen detection method based on visible images | |
| CN104680180B (en) | Classification of Polarimetric SAR Image method based on K mean values and sparse own coding | |
| CN106096561A (en) | Infrared pedestrian detection method based on image block degree of depth learning characteristic | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN104182985B (en) | Remote sensing image change detection method | |
| CN107944483B (en) | Multispectral image classification method based on dual-channel DCGAN and feature fusion | |
| CN103247059A (en) | Remote sensing image region of interest detection method based on integer wavelets and visual features | |
| CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
| CN104453875B (en) | Shale gas reservoir identification method based on self-organizing competitive neural network | |
| CN109766936A (en) | Image change detection method based on information transmitting and attention mechanism | |
| CN107844751A (en) | The sorting technique of guiding filtering length Memory Neural Networks high-spectrum remote sensing | |
| CA2839728A1 (en) | Apparatus and methods of analysis of pipe and annulus in a wellbore | |
| CN107563411A (en) | Online SAR target detection method based on deep learning | |
| CN110609320A (en) | Pre-stack seismic reflection pattern recognition method based on multi-scale feature fusion | |
| CN108122003A (en) | A kind of Weak target recognition methods based on deep neural network |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180615 |
|
| RJ01 | Rejection of invention patent application after publication |