CN109033521B - Newly-built railway slope-limiting optimization decision method - Google Patents

Newly-built railway slope-limiting optimization decision method Download PDF

Info

Publication number
CN109033521B
CN109033521B CN201810658482.8A CN201810658482A CN109033521B CN 109033521 B CN109033521 B CN 109033521B CN 201810658482 A CN201810658482 A CN 201810658482A CN 109033521 B CN109033521 B CN 109033521B
Authority
CN
China
Prior art keywords
layer
railway
convolution
formula
locomotive
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.)
Active
Application number
CN201810658482.8A
Other languages
Chinese (zh)
Other versions
CN109033521A (en
Inventor
蒲浩
张洪
李伟
王雷
宋陶然
李晓明
谢佳
王杰
彭先宝
胡建平
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Central South University
Original Assignee
Central South University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Central South University filed Critical Central South University
Priority to CN201810658482.8A priority Critical patent/CN109033521B/en
Publication of CN109033521A publication Critical patent/CN109033521A/en
Application granted granted Critical
Publication of CN109033521B publication Critical patent/CN109033521B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Geometry (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a newly-built railway slope-limiting optimization decision method, which comprises the following steps: firstly, constructing a deep convolutional neural network model; then establishing a railway case database, representing various factors influencing the decision of limiting the gradient of the newly-built railway into a gray-scale image, and fusing the gray-scale image into a multi-channel image for training a network model; and finally, providing a sliding scanning technology, and performing railway slope limit decision by combining the trained deep convolutional neural network model. Compared with the prior art, the method has the advantages of high automation degree, strong practicability, high operation efficiency, good application prospect and the like.

Description

Newly-built railway slope-limiting optimization decision method
Technical Field
The invention relates to a railway design method, in particular to a newly-built railway slope-limiting optimization decision method.
Background
The grade limit is a main technical standard of railways with global significance, and directly influences the transportation capacity, engineering cost, operation cost and traffic safety of a line, and even possibly determines the trend of the line. With the rapid development of economy in China, the railway transportation demand is continuously increased, meanwhile, railway construction is gradually changed from the eastern plain to the western mountain area, and the contradiction between railway engineering construction and the increasing transportation demand is more prominent due to the complex environment of the hard mountain area: in order to better adapt to complex terrain and geological conditions, shorten the line length and save the engineering construction cost, the adoption of a larger limited gradient is an effective means; however, the line transportation capacity is also affected by the maximum limit gradient, and in the case of the same model number (i.e. the same traction power), the use of a larger limit gradient will reduce the traction tonnage of the locomotive, thereby reducing the line transportation capacity, and also increasing the operation cost and the risk of the downhill section. In addition, the limiting gradient is a fixed equipment standard and is difficult to modify once the railway is built. Therefore, how to scientifically and reasonably decide the limit gradient which is optimally matched with the natural, economic and social environments is a great problem in the design of the railway line at present.
The decision of limiting the gradient of the newly-built railway is essentially to explore the mapping rule of multi-dimensional influence factors (such as terrain conditions, transportation requirements and the like) and the limiting gradient, so as to select the optimal scheme. In a traditional limited gradient optimization decision method, a rule between elements is assumed to conform to a certain mathematical model expression, and then a mapping rule is obtained by counting regression model parameters. For example, the Wang mansion of southwest university of transportation obtains a general empirical formula (1) of a limited slope and engineering cost mapping rule by performing statistical regression on design data of railways in thousands of kilometers of mountain areas in China. Wherein A is engineering cost, I is limiting gradient, and a, b and c are model parameters related to terrain conditions obtained through statistical regression.
Figure BDA0001706037240000011
However, the mapping rule between the multidimensional influencing factor and the limiting gradient is complex and nonlinear, and is difficult to completely and accurately express through a fixed functional relation. Therefore, a method for comprehensively and accurately identifying the mapping rule between the multidimensional influence factors and the limited gradient is urgently needed, and the optimization decision of the limited gradient of the newly-built railway is realized.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: the method can comprehensively and accurately identify the mapping rule between the multidimensional influence factors and the limited gradient, and further realize the optimization decision of the limited gradient of the newly-built railway.
In order to solve the technical problems, the invention adopts the technical scheme that: a newly-built railway slope-limiting optimization decision method comprises the following steps:
S1: constructing a deep convolution neural network model for newly building a railway slope limiting optimization decision;
S2: establishing a training data set D for training a deep convolutional neural networktrainAnd validating the data set Dvalidate；
S2-1: collecting N1Establishing a railway case data set D by adopting built passenger-cargo collinear railway cases with different limiting slopes1；
S2-2: based on the railway case data set D1Dividing rectangular research areas of each railway case at the starting and ending positions of each railway line, extracting grid elevation data information in each rectangular research area, and establishing a railway case elevation data set D2；
S2-3: based on D2Drawing the elevation gray-scale map P of each rectangular research area according to the grid elevation data information of each railway case research areaelevationEstablishing an elevation gray level atlas D for representing the terrain elevation change characteristics of each railway case research areaelevation；
S2-4: based on D2Drawing gradient gray level graph P of each rectangular research area according to grid elevation data information of each railway case research areaslopeAnd establishing a gradient gray level atlas D for representing the terrain gradient characteristics of the railway case research areaslope；
S2-5: representing different railway grades as grey-scale maps with different grey-scale values according to D1Each of which isActual grade of railway case, drawing railway grade gray-scale map P corresponding to each railway caseclassificationEstablishing a railway grade gray level map set Dclassification；
S2-6: characterizing different locomotive models as gray-scale maps with different gray-scale values according to D1The actual locomotive model used by each railway case is drawn, and a locomotive model gray scale map P corresponding to each railway case is drawnlocomotiveEstablishing a model gray level map set D of the locomotivelocomotive；
S2-7: elevation gray scale atlas D based on establishmentelevationGradient gray scale atlas DslopeRailway grade gray scale atlas DclassificationLocomotive model gray scale atlas DlocomotiveFused to D1Elevation gray-scale map P of each railway caseelevationGradient gray scale map PslopeRailway grade gray scale map PclassificationGray scale map P of motor vehicle modellocomotiveForming a four-channel map P capable of representing the information of each railway casemergeAnd creating a data set Dmerge；
S2-8: data set DmergeCutting a four-channel image representing information of each railway case into images with the size of 333 multiplied by 333 pixels, and giving label data, wherein the label data is a limited gradient value actually used by each railway case;
S2-9: will S2-8Dividing the obtained labeled data graph according to the proportion of 4:1, and establishing a training data set D for training the deep convolutional neural networktrainAnd validating the data set Dvalidate；
S3: by using S2Established training data set DtrainTraining the constructed network model and adopting S2Created validation data set DvalidateVerifying the model precision to obtain a trained and verified deep convolutional neural network model;
S4: in addition, collecting N2Bars and data sets D1In different built passenger-cargo collinear railway cases and according to stepsStep S2-2To S2-7Generating a four-channel map P characterizing railway case informationmergeEstablishing a test data set Dtest；
S5: providing a sliding scanning technology, scanning a data set D by a trained deep convolution neural network model from left to right and from top to bottomtestRepresenting four-channel map of elevation information, gradient information, railway grade information and locomotive model information of each railway case, and determining D according to output times of each limited gradient valuetestAnd (4) the recommended limit gradient value of each railway case.
Further, the step S1The deep convolutional neural network model constructed in (1) comprises 5 convolutional layers (Conv), 3 pooling layers (Pool), 2 full-link layers (FC) and 1 Softmax output layer:
(1) the convolution kernel size adopted by the first convolution layer (Conv1) is 33 multiplied by 3, the step size is 4, the number of convolution kernels is 96, and a modified linear unit (ReLU) is connected behind Conv1 to be used as a nonlinear activation function, so that the model has nonlinear characteristics;
(2) conv1 is connected with a first pooling layer (Pool1) after nonlinear treatment, the size of a pooling core adopted by Pool1 is 4 x 4, and the step size is 2;
(3) a second convolutional layer (Conv2) is connected behind the Pool1, the size of a convolution kernel adopted by the Conv2 is 3 multiplied by 96, the step size is 1, the number of the convolution kernels is 256, and a modified linear unit (ReLU) is connected behind the Conv2 for nonlinear processing;
(4) conv2 was treated non-linearly and then connected to a second pooling layer (Pool2), Pool2 with pooling kernel size of 3 × 3 and step size of 2;
(5) a third convolutional layer (Conv3) is connected behind the Pool2, the size of a convolution kernel adopted by the Conv3 is 3 multiplied by 256, the step size is 1, the number of the convolution kernels is 384, and a modified linear unit (ReLU) is connected behind the Conv3 for nonlinear processing;
(6) conv3 is connected with a fourth convolutional layer (Conv4) after nonlinear processing, the size of a convolution kernel adopted by Conv4 is 3 multiplied by 384, the step size is 1, the number of the convolution kernels is 384, and a modified linear unit (ReLU) is connected after Conv4 for nonlinear processing;
(7) conv4 is connected with a fifth convolutional layer (Conv5) after nonlinear processing, the size of a convolution kernel adopted by Conv5 is 3 multiplied by 384, the step size is 1, the number of convolution kernels is 256, and a modified linear unit (ReLU) is connected after Conv5 for nonlinear processing;
(8) conv5 was treated non-linearly and then connected to a third pooling layer (Pool3), Pool3 with pooling kernel size of 3 × 3 and step size of 2;
(9) a first full connection layer (FC1) is connected behind the Pool3, in order to prevent an overfitting phenomenon, a dropout function is adopted for connecting the Pool3 layer to the FC1 layer, and a modified linear unit (ReLU) is connected behind the FC1 layer for nonlinear processing;
(10) FC1 is connected with a second full connection layer (FC2) after nonlinear processing, a dropout function is adopted to prevent the over-fitting phenomenon, and a correction linear unit (ReLU) is connected with the FC2 for nonlinear processing;
(11) and the FC2 is connected with a Softmax output layer after nonlinear processing and is used for outputting the newly-built railway limit gradient value recommendation.
Further, the step S2-1The railway cases collected in (1) cover different grades of railway and different locomotive models.
Further, the step S2-2The middle railway rectangular research area division method comprises the following steps: setting the starting point and the ending point of a certain railway case line as Si：(xSi,ySi) And Ei：(xEi,yEi) Then the area of investigation of the railway case is SiAnd EiAs diagonal point, with | xEi-xSiL is long, | yEi-ySiAnd | is a wide rectangular area.
Further, the step S2-3、S2-4And drawing the elevation gray-scale map and the gradient gray-scale map of the rectangular research area of each railway case by adopting Global Mapper software.
Further, the step S2-5Middle and railway grade gray scale map PclassificationIs the same size as the rectangular study area of the railway case.
Further, the stepsS2-6Grayscale map P of middle and middle locomotive modellocomotiveIs the same size as the rectangular study area of the railway case using the locomotive model.
Further, the step S2-7Four-channel map P of each railway casemergeThe elevation gray-scale map P of each railway case is obtained by adopting merge function in computer vision library OpenCVelevationGradient gray scale map PslopeRailway grade gray scale map PclassificationAnd a locomotive type gray scale map PlocomotiveAnd obtaining the fusion protein after fusion.
Further, the step S3The network model constructed by the middle training is based on S2Created tag data set DtrainContinuously updating the connection weight between each layer in the network model by a gradient descent algorithm, which comprises the following specific steps:
(1) softmax layer connection weight update
The Softmax layer is used for outputting the limited gradient value recommended by the model, calculating the output probability of each limited gradient value according to the output value of each neuron in the previous layer, and selecting the gradient value with the maximum output probability as the limited gradient value recommended by the model, wherein the function expression of the gradient value is shown as the formula (2):
Figure BDA0001706037240000051
in the formula: p (y)(i)＝j|x(i)(ii) a W) is the probability that the ith picture is taken as input data, the jth value is selected as the limiting gradient in the Softmax layer, and x(i)Is input data of the Softmax layer (i.e. output data of the previous layer), and W is a connection weight of the Softmax layer and the previous layer.
Establishing a model loss function E based on a Softmax function, wherein the function expression of the model loss function E is shown as a formula (3):
Figure BDA0001706037240000061
in the formula: 1{ y(i)J is logicExpression, if the i input pictures are marked as the jth limiting gradient, 1{ y }(i)1, otherwise 1{ y }(i)J is 0, and λ is a weight attenuation coefficient.
Based on the loss function E, the residuals of the neurons in the Softmax layer can be calculated as equation (4):
Figure BDA0001706037240000062
the connection weights of the neurons in the Softmax layer are updated according to equations (5) and (6):
Figure BDA0001706037240000063
Figure BDA0001706037240000064
(2) full connection layer connection weight update
Each neuron of the full connection layer is connected with all neurons of the previous layer, and the connection weight updating formula is as follows:
Figure BDA0001706037240000065
Figure BDA0001706037240000066
in the formula: wlA connection weight matrix for each neuron of the current layer (full connection layer), blThe connecting bias vector of each neuron in the current layer is alpha, which is the learning rate.
Partial derivative of loss function to neuron connection weight of full connection layer
Figure BDA0001706037240000071
And partial derivatives of bias for neuron connections of full connection layer
Figure BDA0001706037240000072
Can be calculated according to equation (9) and equation (10), respectively.
Figure BDA0001706037240000073
Figure BDA0001706037240000074
In the formula: x is the number ofl-1Is the output vector, delta, of a connected layer above the current layer (full connected layer)lThe residual error of each neuron in the current layer (full connection layer) can be determined according to the residual error delta of each neuron in the next connection layerl+1And (4) calculating.
Figure BDA0001706037240000075
In the formula: wl+1F (-) is the ReLU activation function, which is the connection weight matrix of each neuron of the posterior connection layer of the current layer (full connection layer).
Figure BDA0001706037240000076
(3) Convolutional layer connection weight update
Each neuron of the convolution layer is connected with the previous layer through a convolution kernel, and the connection weight updating formula of each convolution kernel is as follows:
Figure BDA0001706037240000077
Figure BDA0001706037240000078
in the formula:
Figure BDA0001706037240000079
the connection weight matrix of the (d) th convolution kernel of the current layer (convolutional layer),
Figure BDA00017060372400000710
the connected offset vector of the d-th convolution kernel of the current layer (convolution layer) is denoted by α as the learning rate.
The d-th convolution kernel of the current layer (convolution layer) is connected with the weight partial derivative by the loss function
Figure BDA00017060372400000711
The calculation formula of (a) is as follows:
Figure BDA0001706037240000081
in the formula:
Figure BDA0001706037240000082
is the output value of the D' th characteristic diagram of the previous connection layer of the current layer (convolution layer), Dl-1The number of feature maps of the previous connection layer of the current layer (convolution layer),
Figure BDA0001706037240000083
the residual matrix is the d-th characteristic diagram of the current layer (convolutional layer).
The d convolution kernel of the current layer (convolution layer) is connected with the bias partial derivative by the loss function
Figure BDA0001706037240000084
The calculation formula of (a) is as follows:
Figure BDA0001706037240000085
in the formula:
Figure BDA0001706037240000086
the connected offset vector of the d-th feature map in the current layer (convolutional layer),
Figure BDA0001706037240000087
and
Figure BDA0001706037240000088
respectively the number of rows and columns of the d-th feature map in the current layer (convolutional layer),
Figure BDA0001706037240000089
the residual values of i row and j column in the d-th feature map in the current layer (convolutional layer) are shown.
The residual of the current layer (convolutional layer) is calculated based on the layer residual of the next connection by back propagation. If the current layer (convolutional layer) is connected to the subsequent pooling layer, the residual matrix of the d-th feature map of the current layer (convolutional layer) is calculated by equation (17).
Figure BDA00017060372400000810
In the formula: xl-1Is the output matrix of the previous connection layer of the current layer (convolutional layer),
Figure BDA00017060372400000811
residual matrix of the d-th characteristic diagram in a connected layer behind the current layer (convolutional layer).
If a convolutional layer is connected after the current layer (convolutional layer), the weight matrix of the current layer (convolutional layer) is calculated by equation (18).
Figure BDA00017060372400000812
In the formula:
Figure BDA00017060372400000813
is the residual matrix of the d' th characteristic diagram in the next connection layer after the current layer (convolution layer),
Figure BDA00017060372400000814
a d-th layer weight matrix of a d' th convolution kernel of a connection layer subsequent to the current layer (convolutional layer),
Figure BDA00017060372400000815
the output matrix of the d-th characteristic diagram of the current layer (convolutional layer).
Further, the step S5The sliding scanning technique in (1) is specifically as follows: when scanning the test data set DtestAnd when a certain four-channel image is scanned, the recommended limit gradient value of a 333 x 333 pixel area in the four-channel image can be output each time, and after the whole four-channel image is scanned, the gradient value with the maximum output times is selected as the recommended limit gradient value of the railway case represented by the four-channel image.
The invention has the beneficial effects that: the deep learning simulates the hierarchical structure of the brain, can automatically acquire hierarchical multi-layer feature expression from massive data, and explores the potential rules existing between input data and output data without giving mathematical expressions. The scheme of the invention adopts a deep learning algorithm to make the slope limiting decision of the newly-built railway feasible. The invention provides a newly-built railway limited gradient optimization decision method based on a convolutional neural network in a deep learning algorithm. The scheme of the invention adopts a sliding scanning technology, and realizes the decision of limiting the gradient of different railway cases. The method has the advantages of high automation degree, strong practicability, high operation efficiency and good popularization and application prospects.
Drawings
FIG. 1 is a schematic flow chart of a newly built railway slope limiting optimization decision method of the invention;
FIG. 2 is a deep convolutional neural network model according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a sliding scanning technique according to an embodiment of the present invention.
Detailed Description
In order to explain technical contents, achieved objects, and effects of the present invention in detail, the following description is made with reference to the accompanying drawings in combination with the embodiments.
An embodiment of the present invention is a newly-built railway slope-limiting optimization decision method, as shown in fig. 1, the optimization decision method includes the following steps:
S1: the method comprises the following steps of constructing a deep convolutional neural network model for newly-built railway slope-limiting optimization decision, wherein the constructed network model comprises 5 convolutional layers (Conv), 3 pooling layers (Pool), 2 full-connection layers (FC) and 1 Softmax output layer:
(1) the convolution kernel size adopted by the first convolution layer (Conv1) is 33 multiplied by 3, the step size is 4, the number of convolution kernels is 96, and a modified linear unit (ReLU) is connected behind Conv1 to be used as a nonlinear activation function, so that the model has nonlinear characteristics;
(2) conv1 is connected with a first pooling layer (Pool1) after nonlinear treatment, the size of a pooling core adopted by Pool1 is 4 x 4, and the step size is 2;
(3) a second convolutional layer (Conv2) is connected behind the Pool1, the size of a convolution kernel adopted by the Conv2 is 3 multiplied by 96, the step size is 1, the number of the convolution kernels is 256, and a modified linear unit (ReLU) is connected behind the Conv2 for nonlinear processing;
(4) conv2 was treated non-linearly and then connected to a second pooling layer (Pool2), Pool2 with pooling kernel size of 3 × 3 and step size of 2;
(5) a third convolutional layer (Conv3) is connected behind the Pool2, the size of a convolution kernel adopted by the Conv3 is 3 multiplied by 256, the step size is 1, the number of the convolution kernels is 384, and a modified linear unit (ReLU) is connected behind the Conv3 for nonlinear processing;
(6) conv3 is connected with a fourth convolutional layer (Conv4) after nonlinear processing, the size of a convolution kernel adopted by Conv4 is 3 multiplied by 384, the step size is 1, the number of the convolution kernels is 384, and a modified linear unit (ReLU) is connected after Conv4 for nonlinear processing;
(7) conv4 is connected with a fifth convolutional layer (Conv5) after nonlinear processing, the size of a convolution kernel adopted by Conv5 is 3 multiplied by 384, the step size is 1, the number of convolution kernels is 256, and a modified linear unit (ReLU) is connected after Conv5 for nonlinear processing;
(8) conv5 was treated non-linearly and then connected to a third pooling layer (Pool3), Pool3 with pooling kernel size of 3 × 3 and step size of 2;
(9) a first full connection layer (FC1) is connected behind the Pool3, in order to prevent an overfitting phenomenon, a dropout function is adopted from a Pool3 layer to an FC1 layer, and a modified linear unit (ReLU) is connected behind the FC1 layer for nonlinear processing;
(10) FC1 is connected with a second full connection layer (FC2) after nonlinear processing, a dropout function is adopted to prevent the over-fitting phenomenon, and a correction linear unit (ReLU) is connected with the FC2 for nonlinear processing;
(11) and the FC2 is connected with a Softmax output layer after nonlinear processing and is used for outputting the slope limit recommended value of the newly-built railway.
S2: establishing a training data set D for training a deep convolutional neural networktrainAnd validating the data set Dvalidate；
S2-1: 246 passenger-cargo collinear railway cases with the gradient limited by 6 per thousand, 12 per thousand and 24 per thousand are collected, the collected railway cases cover four railway grades of grade I, grade II, grade III and grade IV, three locomotive models of Shaoshan model 1, Shaoshan model 3 and Shaoshan model 4, and a railway case data set D is established1；
S2-2: based on the railway case data set D1Dividing rectangular research areas of each railway case at the starting and ending positions of each railway line, extracting grid elevation data information in each rectangular research area, and establishing a railway case elevation data set D2；
S2-3: based on D2Drawing the elevation gray-scale map P of each rectangular research area according to the grid elevation data information of each railway case research areaelevationEstablishing an elevation gray level atlas D for representing the terrain elevation change characteristics of each railway case research areaelevation；
S2-4: based on D2Drawing gradient gray level graph P of each rectangular research area according to grid elevation data information of each railway case research areaslopeAnd establishing a gradient gray level atlas D for representing the terrain gradient characteristics of the railway case research areaslope；
S2-5: with gray-scale values of 0, 40, 80, 120 respectivelyThe grey scale maps characterize four railway classes and are based on D1The actual grade of each railway case is drawn, and a railway grade gray scale map P corresponding to each railway case is drawnclassificationEstablishing a railway grade gray level map set Dclassification；
S2-6: representing three electric locomotive models of Shaoshan 1 type, Shaoshan 3 type and Shaoshan 4 type by gray scale graphs with gray scale values of 160, 200 and 240 respectively, and according to D1The actual locomotive model used by each railway case is drawn, and a locomotive model gray scale map P corresponding to each railway case is drawnlocomotiveEstablishing a model gray level map set D of the locomotivelocomotive；
S2-7: elevation gray scale atlas D based on establishmentelevationGradient gray scale atlas DslopeRailway grade gray scale atlas DclassificationLocomotive model gray scale atlas DlocomotiveFused to D1Elevation gray-scale map P of each railway caseelevationGradient gray scale map PslopeRailway grade gray scale map PclassificationGray scale map P of motor vehicle modellocomotiveForming a four-channel map P capable of representing the information of each railway casemergeAnd creating a data set Dmerge；
S2-8: data set DmergeCutting a four-channel image representing information of each railway case into images with the size of 333 multiplied by 333 pixels, and giving label data, wherein the label data is a limited gradient value actually used by each railway case;
S2-9: will S2-8Dividing the obtained picture with the label according to the proportion of 4:1, and establishing a training data set D for training the deep convolutional neural networktrainAnd validating the data set Dvalidate；
S3: by using S2Established training data set DtrainTraining the constructed network model and adopting S8Created validation data set DvalidateAnd verifying the model precision to obtain a trained and verified deep convolution neural network model. The training and verification takes 9 hours and 35 hoursIn minutes (i7 processor, 16G memory and GTX 1080 video card), a deep convolutional neural network model with the accuracy of 83.35% is obtained.
S4: in addition, 36 pieces of data were collected together with the data set D1In different built passenger-cargo collinear railway cases and according to the step S2-2To S2-7Establishing a four-channel map P representing railway case informationmergeEstablishing a test data set Dtest；
S5: providing a sliding scanning technology, scanning a data set D by a trained deep convolution neural network model from left to right and from top to bottomtestRepresenting four-channel map of elevation information, gradient information, railway grade information and locomotive model information of each railway case, and determining D according to output times of each limited gradient valuetestAnd (4) the recommended limit gradient value of each railway case. In 36 railway cases tested at this time, the limited gradient of 34 railway cases is accurately decided (that is, the limited gradient value recommended by the model is the same as the limited gradient value of the manual decision), and the accuracy can reach 94.44%.
The sliding scanning technology refers to that the limited gradient value is determined according to the output times of different limited gradients by scanning the whole picture.
In summary, the invention provides a newly-built railway limited gradient optimization decision method, which comprises the steps of firstly constructing a deep convolution neural network model, then establishing a railway case database, representing various factors influencing the newly-built railway limited gradient decision into a gray-scale image, and fusing the gray-scale image into a multi-channel image for training the network model; and finally, providing a sliding scanning technology, and combining the trained deep convolutional neural network model to perform newly-built railway slope limit decision.
The above description is only an embodiment of the present invention, and not intended to limit the scope of the present invention, and all equivalent changes made by using the contents of the present specification and the drawings, or applied directly or indirectly to the related technical fields, are included in the scope of the present invention.

Claims (9)

1. A newly-built railway slope-limiting optimization decision method is characterized by comprising the following steps: the method comprises the following steps:
S1: constructing a deep convolution neural network model for newly building a railway slope limiting optimization decision;
S2: establishing a training data set D for training a deep convolutional neural networktrainAnd validating the data set Dvalidate；
S2-1: collecting N1Establishing a railway case data set D by adopting built passenger-cargo collinear railway cases with different limiting slopes1；
S2-2: based on the railway case data set D1Dividing rectangular research areas of each railway case at the starting and ending positions of each railway line, extracting grid elevation data information in each rectangular research area, and establishing a railway case elevation data set D2；
S2-3: based on D2Drawing the elevation gray-scale map P of each rectangular research area according to the grid elevation data information of each railway case research areaelevationEstablishing an elevation gray level atlas D for representing the terrain elevation change characteristics of each railway case research areaelevation；
S2-4: based on D2Drawing gradient gray level graph P of each rectangular research area according to grid elevation data information of each railway case research areaslopeAnd establishing a gradient gray level atlas D for representing the terrain gradient characteristics of the railway case research areaslope；
S2-5: representing different railway grades as grey-scale maps with different grey-scale values according to D1The actual grade of each railway case is drawn, and a railway grade gray scale map P corresponding to each railway case is drawnclassificationEstablishing a railway grade gray level map set Dclassification；
S2-6: characterizing different locomotive models as gray-scale maps with different gray-scale values according to D1The actual locomotive model used by each railway case is drawn, and a locomotive model gray scale map P corresponding to each railway case is drawnlocomotiveEstablishing a model gray level map set D of the locomotivelocomotive；
S2-7: elevation gray scale atlas D based on establishmentelevationGradient gray scale atlas DslopeRailway grade gray scale atlas DclassificationLocomotive model gray scale atlas DlocomotiveFused to D1Elevation gray-scale map P of each railway caseelevationGradient gray scale map PslopeRailway grade gray scale map PclassificationGray scale map P of motor vehicle modellocomotiveForming a four-channel map P capable of representing the information of each railway casemergeAnd creating a data set Dmerge；
S2-8: data set DmergeCutting a four-channel image representing information of each railway case into images with the size of 333 multiplied by 333 pixels, and giving label data, wherein the label data is a limited gradient value actually used by each railway case;
S2-9: will S2-8Dividing the obtained labeled data graph according to the proportion of 4:1, and establishing a training data set D for training the deep convolutional neural networktrainAnd validating the data set Dvalidate；
S3: by using S2Established training data set DtrainTraining the constructed network model and adopting S2Created validation data set DvalidateVerifying the model precision to obtain a trained and verified deep convolutional neural network model;
S4: in addition, collecting N2Bars and data sets D1In different built passenger-cargo collinear railway cases and according to the step S2-2To S2-7Generating a four-channel map P characterizing railway case informationmergeEstablishing a test data set Dtest；
S5: scanning the data set D by the trained deep convolution neural network model from left to right and from top to bottomtestRepresenting four-channel map of elevation information, gradient information, railway grade information and locomotive model information of each railway case according to each limitThe number of outputs of the gradient value, determining DtestAnd (4) the recommended limit gradient value of each railway case.
2. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S1The deep convolutional neural network model comprises 5 convolutional layers, 3 pooling layers, 2 full-link layers and 1 Softmax output layer:
1) the size of a convolution kernel adopted by the first convolution layer is 33 multiplied by 3, the step size is 4, the number of the convolution kernels is 96, and a correction linear unit is connected behind the first convolution layer to be used as a nonlinear activation function, so that the model has nonlinear characteristics;
2) the first convolutional layer is connected with a first pooling layer after nonlinear processing, the size of a pooling kernel adopted by the first pooling layer is 4 multiplied by 4, and the step size is 2;
3) the second convolution layer is connected behind the first pooling layer, the size of a convolution kernel adopted by the second convolution layer is 3 multiplied by 96, the step size is 1, the number of the convolution kernels is 256, and the second convolution layer is connected with a correction linear unit for nonlinear processing;
4) the second convolutional layer is connected with a second pooling layer after nonlinear processing, the size of a pooling kernel adopted by the second pooling layer is 3 multiplied by 3, and the step size is 2;
5) the second pooling layer is connected with a third convolution layer, the size of convolution kernels adopted by the third convolution layer is 3 x 256, the step size is 1, the number of the convolution kernels is 384, and the third convolution layer is connected with a correction linear unit for nonlinear processing;
6) the third convolutional layer is connected with a fourth convolutional layer after nonlinear processing, the size of a convolutional kernel adopted by the fourth convolutional layer is 3 multiplied by 384, the step size is 1, the number of the convolutional kernels is 384, and the fourth convolutional layer is connected with a correction linear unit for nonlinear processing;
7) the fourth convolutional layer is connected with a fifth convolutional layer after nonlinear processing, the size of a convolutional kernel adopted by the fifth convolutional layer is 3 multiplied by 384, the step size is 1, the number of the convolutional kernels is 256, and the fifth convolutional layer is connected with a correction linear unit for nonlinear processing;
8) the fifth convolutional layer is connected with a third pooling layer after nonlinear processing, the size of a pooling core adopted by the third pooling layer is 3 multiplied by 3, and the step size is 2;
9) a first full-connection layer is connected behind the third pooling layer, in order to prevent the over-fitting phenomenon, a dropout function is adopted for connecting the third pooling layer to the first full-connection layer, and a correction linear unit is connected behind the first full-connection layer for nonlinear processing;
10) the first full-connection layer is connected with the second full-connection layer after nonlinear processing, a dropout function is adopted to prevent the over-fitting phenomenon, and the second full-connection layer is connected with a correction linear unit for nonlinear processing;
11) and the second full-connection layer is connected with the Softmax output layer after nonlinear processing and is used for outputting the slope-limiting recommended value of the newly-built railway.
3. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S2-1The collected railway cases cover different grades of railway and different locomotive models.
4. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S2-2In the middle, the method for dividing the rectangular research area based on the starting and ending positions of the railway line is as follows:
setting the starting point and the ending point of a certain railway case line as Si：(xSi,ySi) And Ei：(xEi,yEi) Then the area of investigation of the railway case is SiAnd EiAs diagonal point, with | xEi-xSiL is long, | yEi-ySiAnd | is a wide rectangular area.
5. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S2-5Middle and railway grade gray scale map PclassificationSize of and the railwayThe rectangular study area of the cases was the same size.
6. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S2-6Grayscale map P of middle and middle locomotive modellocomotiveIs the same size as the rectangular study area of the railway case using the locomotive model.
7. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S2-7Four-channel map P of each railway casemergeThe elevation gray-scale map P of each railway case is obtained by adopting merge function in computer vision library OpenCVelevationGradient gray scale map PslopeRailway grade gray scale map PclassificationAnd a locomotive type gray scale map PlocomotiveAnd obtaining the fusion protein after fusion.
8. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S3The network model constructed by the middle training is based on S2Created tag data set DtrainContinuously updating the connection weight between each layer in the network model by a gradient descent algorithm, which comprises the following specific steps:
1) softmax layer connection weight update:
the Softmax layer is used for outputting the limited gradient value recommended by the model, calculating the output probability of each limited gradient value according to the output value of each neuron in the previous layer, and selecting the gradient value with the maximum output probability as the limited gradient value recommended by the model, wherein the function expression of the gradient value is shown as the formula (2):
Figure FDA0002952295360000051
in the formula: p (y)(i)＝j|x(i)(ii) a W) is the probability that the ith picture is taken as input data, the jth value is selected as the limiting gradient in the Softmax layer, and x(i)The input data of the Softmax layer, W is the connection weight of the Softmax layer and the previous layer;
establishing a model loss function E based on a Softmax function, wherein the function expression of the model loss function E is shown as a formula (3):
Figure FDA0002952295360000052
in the formula: 1{ y(i)J is a logic expression, if the ith input picture is marked as the jth limited gradient value, 1{ y }(i)1, otherwise 1{ y }(i)J is 0, and lambda is a weight attenuation coefficient;
based on the loss function E, the residuals of the neurons in the Softmax layer are calculated according to equation (4):
Figure FDA0002952295360000053
the connection weights of the neurons in the Softmax layer are updated according to equations (5) and (6):
Figure FDA0002952295360000061
Figure FDA0002952295360000062
2) updating the connection weight of the full connection layer:
each neuron of the full connection layer is connected with all neurons of the previous layer, and the connection weight updating formula is as follows:
Figure FDA0002952295360000063
Figure FDA0002952295360000064
in the formula: wlA connection weight matrix for each neuron of the current layer, blConnecting bias vectors of each neuron of the current layer, wherein alpha is a learning rate;
partial derivative of loss function to neuron connection weight of full connection layer
Figure FDA0002952295360000065
And partial derivatives of bias for neuron connections of full connection layer
Figure FDA0002952295360000066
Respectively calculating according to the formula (9) and the formula (10);
Figure FDA0002952295360000067
Figure FDA0002952295360000068
in the formula: x is the number ofl-1Is the output vector, delta, of a connection layer above the current layerlThe residual error of each neuron in the current layer can be determined according to the residual error delta of each neuron in the next connection layerl+1Calculating;
δl＝(Wl+1)δl+1⊙f′(Wlxl-1+bl) (11)
in the formula: wl+1A connection weight matrix of each neuron of a posterior connection layer of the current layer, wherein f (·) is a ReLU activation function;
ReLU：
Figure FDA0002952295360000071
3) convolutional layer connection weight update:
each neuron of the convolution layer is connected with the previous layer through a convolution kernel, and the connection weight updating formula of each convolution kernel is as follows:
Figure FDA0002952295360000072
Figure FDA0002952295360000073
in the formula:
Figure FDA0002952295360000074
the connection weight matrix for the d-th convolution kernel of the current layer,
Figure FDA0002952295360000075
connecting offset vectors of the d-th convolution kernel of the current layer, wherein alpha is a learning rate;
each connection weight partial derivative of the d-th convolution kernel of the current layer by the loss function
Figure FDA0002952295360000076
The calculation formula of (a) is as follows:
Figure FDA0002952295360000077
in the formula:
Figure FDA0002952295360000078
is the output value of the D' th characteristic diagram of the previous connection layer of the current layer, Dl-1The number of feature maps of the connection layer before the current layer,
Figure FDA0002952295360000079
a residual error matrix of the d-th characteristic diagram of the current layer;
the d-th convolution kernel of the current layer is connected with the bias partial derivative by the loss function
Figure FDA00029522953600000710
The calculation formula of (a) is as follows:
Figure FDA00029522953600000711
in the formula:
Figure FDA00029522953600000712
for the concatenated offset vector of the d-th feature map in the current layer,
Figure FDA00029522953600000713
and
Figure FDA00029522953600000714
respectively the number of rows and columns of the d-th feature map in the current layer,
Figure FDA00029522953600000715
residual values of i rows and j columns in the d-th feature map in the current layer are obtained;
the residual error of the current layer is calculated based on the residual error of the next connected layer through back propagation; if the current layer is connected with the pooling layer, calculating a residual error matrix of the d-th characteristic diagram of the current layer according to the formula (17);
Figure FDA0002952295360000081
in the formula: xl-1The output matrix of the connection layer previous to the current layer,
Figure FDA0002952295360000082
a residual error matrix of the d-th characteristic diagram in a connecting layer behind the current layer;
if the convolution layer is connected behind the current layer, the weight matrix of the current layer is calculated according to the formula (18):
Figure FDA0002952295360000083
in the formula:
Figure FDA0002952295360000084
is the residual matrix of the d' th feature map in the next connection layer after the current layer,
Figure FDA0002952295360000085
a layer d weight matrix for the layer d "of the convolution kernel of the next layer after the current layer,
Figure FDA0002952295360000086
and (4) an output matrix of the d-th feature map of the current layer.
9. The newly-built railway slope limiting optimization decision method according to claim 1, characterized in that: said step S5When scanning the test data set DtestAnd when a certain four-channel image is scanned, the recommended limit gradient value of a 333 x 333 pixel area in the four-channel image can be output each time, and after the scanning of the whole four-channel image is completed, the gradient value with the maximum output times is selected as the recommended limit gradient value of the railway case represented by the four-channel image.
CN201810658482.8A 2018-06-25 2018-06-25 Newly-built railway slope-limiting optimization decision method Active CN109033521B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810658482.8A CN109033521B (en) 2018-06-25 2018-06-25 Newly-built railway slope-limiting optimization decision method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810658482.8A CN109033521B (en) 2018-06-25 2018-06-25 Newly-built railway slope-limiting optimization decision method

Publications (2)

Publication Number Publication Date
CN109033521A CN109033521A (en) 2018-12-18
CN109033521B true CN109033521B (en) 2021-04-20

Family

ID=64610700

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810658482.8A Active CN109033521B (en) 2018-06-25 2018-06-25 Newly-built railway slope-limiting optimization decision method

Country Status (1)

Country Link
CN (1) CN109033521B (en)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105243233A (en) * 2015-11-04 2016-01-13 中南大学 Line-station collaborative optimization method for railway in complicated mountain area
CN106777752A (en) * 2016-12-30 2017-05-31 华东交通大学 A kind of bullet train follows the trail of operation curve Optimal Setting method
CN106844949A (en) * 2017-01-18 2017-06-13 清华大学 A kind of training method for realizing the controllable two-way LSTM models of locomotive section
CN106910185A (en) * 2017-01-13 2017-06-30 陕西师范大学 A kind of DBCC disaggregated models and construction method based on CNN deep learnings
CN107133960A (en) * 2017-04-21 2017-09-05 武汉大学 Image crack dividing method based on depth convolutional neural networks
CN107689035A (en) * 2017-08-30 2018-02-13 广州华多网络科技有限公司 A kind of homography matrix based on convolutional neural networks determines method and device

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106033555A (en) * 2015-03-13 2016-10-19 中国科学院声学研究所 Big data processing method based on depth learning model satisfying K-dimensional sparsity constraint

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105243233A (en) * 2015-11-04 2016-01-13 中南大学 Line-station collaborative optimization method for railway in complicated mountain area
CN106777752A (en) * 2016-12-30 2017-05-31 华东交通大学 A kind of bullet train follows the trail of operation curve Optimal Setting method
CN106910185A (en) * 2017-01-13 2017-06-30 陕西师范大学 A kind of DBCC disaggregated models and construction method based on CNN deep learnings
CN106844949A (en) * 2017-01-18 2017-06-13 清华大学 A kind of training method for realizing the controllable two-way LSTM models of locomotive section
CN107133960A (en) * 2017-04-21 2017-09-05 武汉大学 Image crack dividing method based on depth convolutional neural networks
CN107689035A (en) * 2017-08-30 2018-02-13 广州华多网络科技有限公司 A kind of homography matrix based on convolutional neural networks determines method and device

Also Published As

Publication number Publication date
CN109033521A (en) 2018-12-18

Similar Documents

Publication Publication Date Title
CN113240580B (en) Lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation
CN110766063B (en) Image classification method based on compressed excitation and tightly connected convolutional neural network
CN109584161A (en) The Remote sensed image super-resolution reconstruction method of convolutional neural networks based on channel attention
CN111583165A (en) Image processing method, device, equipment and storage medium
CN114092832B (en) High-resolution remote sensing image classification method based on parallel hybrid convolutional network
CN108197326A (en) A kind of vehicle retrieval method and device, electronic equipment, storage medium
CN106228185A (en) A kind of general image classifying and identifying system based on neutral net and method
CN111275171B (en) Small target detection method based on parameter sharing multi-scale super-division reconstruction
CN110889450B (en) Super-parameter tuning and model construction method and device
CN108682009A (en) A kind of Alzheimer's disease prediction technique, device, equipment and medium
CN110751644B (en) Road surface crack detection method
CN112115967B (en) Image increment learning method based on data protection
CN112330050A (en) Power system load prediction method considering multiple features based on double-layer XGboost
CN113469119B (en) Cervical cell image classification method based on visual converter and image convolution network
CN113449787B (en) Chinese character stroke structure-based font library completion method and system
CN109376763A (en) Sample classification method, system and medium based on multisample reasoning neural network
CN113159067A (en) Fine-grained image identification method and device based on multi-grained local feature soft association aggregation
CN115966010A (en) Expression recognition method based on attention and multi-scale feature fusion
CN113807362B (en) Image classification method based on interlayer semantic information fusion depth convolution network
CN113781385B (en) Combined attention-seeking convolution method for automatic classification of brain medical images
CN114647760A (en) Intelligent video image retrieval method based on neural network self-temperature cause and knowledge conduction mechanism
CN109033521B (en) Newly-built railway slope-limiting optimization decision method
CN112101438A (en) Left and right eye classification method, device, server and storage medium
CN115346055A (en) Multi-kernel width map based neural network feature extraction and classification method
Zhao et al. Recognition and Classification of Concrete Cracks under Strong Interference Based on Convolutional 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
GR01 Patent grant
GR01 Patent grant