CN109033521A - A kind of newly built railway ruling grade Study on Decision-making Method for Optimization - Google Patents

Abstract

The invention discloses a kind of newly built railway ruling grade Study on Decision-making Method for Optimization, the Study on Decision-making Method for Optimization the following steps are included: construct depth convolutional neural networks model first；Then railway case database is established, the every factor that will affect newly built railway ruling grade decision is characterized into grayscale image, and is fused into multichannel image for training network model；Finally propose that a kind of slip sweep, the depth convolutional neural networks model that combined training is completed carry out railway ruling grade decision.Compared with prior art, this method has many advantages, such as that high degree of automation, practical, operational efficiency is high and application prospect is good.

Description

An Optimal Decision-Making Method for Constrained Slopes of Newly Built Railways

technical field

The invention relates to a railway design method, in particular to an optimization decision-making method for a new railway limited slope.

Background technique

Slope restriction is a major railway technical standard with overall significance, which directly affects the transportation capacity, engineering cost, operating cost and driving safety of the line, and may even determine the direction of the line. With the rapid development of my country's economy, the demand for railway transportation continues to increase. At the same time, railway construction gradually shifts from the eastern plains to the western mountainous areas. The complex environment of difficult and dangerous mountainous areas makes the contradiction between railway engineering construction and growing transportation demand more prominent: for better It is an effective means to adapt to complex terrain and geological conditions, shorten the length of the line, save engineering construction costs, and adopt a larger limit slope; however, the line transportation capacity is also affected by the maximum limit slope. That is, when the traction power is the same), using a larger restricted slope will reduce the traction tonnage of the locomotive, thereby reducing the transportation capacity of the line, and will also increase the operating cost and the danger of the downhill section. In addition, the limit slope is a fixed equipment standard, which will be difficult to change once the railway is completed. Therefore, how to scientifically and rationally determine the limit slope that best matches the natural, economic, and social environments is a major problem in the design of railway lines.

The essence of decision-making on the slope limit of new railways is to explore the mapping law between multi-dimensional influencing factors (such as terrain conditions, transportation demand, etc.) and the limit slope, so as to select the best plan. The traditional limited slope optimization decision-making method usually assumes that the law between elements conforms to a certain mathematical model expression, and then obtains the mapping law through statistical regression model parameters. For example, Professor Wang Di of Southwest Jiaotong University obtained the general empirical formula (1) for the mapping law of limiting slope and engineering cost through statistical regression on the design data of thousands of kilometers of mountainous railways in my country. In the formula, A is the project cost, I is the limit slope, and a, b, c are the model parameters related to terrain conditions obtained through statistical regression.

However, the mapping law between multidimensional influencing factors and limiting slope is complex and nonlinear, and it is difficult to express it completely and accurately through a fixed functional relationship. Therefore, there is an urgent need for a method that can comprehensively and accurately identify the mapping rules between multi-dimensional influencing factors and limiting slopes, so as to realize the optimal decision-making of new railway limiting slopes.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method that can comprehensively and accurately identify the mapping law between multi-dimensional influencing factors and limiting slopes, thereby realizing optimal decision-making on limiting slopes of newly built railways.

In order to solve the above-mentioned technical problems, the technical solution adopted in the present invention is: a kind of new-built railway limit slope optimization decision-making method, comprises the following steps:

S ₁ : Construct a deep convolutional neural network model for the optimal decision-making of new railway limit slope;

S ₂ : Establish a training data set D _train and a verification data set D _validate for training a deep convolutional neural network;

S _2-1 : collect N ₁ cases of passenger and freight collinear railways with different restricted slopes, and establish railway case data set D ₁ ;

S _2-2 : Divide the rectangular research area of each railway case based on the start and end positions of each railway line in the railway case data set D ₁ , and extract the grid elevation data information in each rectangular research area to establish the elevation data of the railway case set D2 _;

S _2-3 : Based on the grid elevation data information of each railway case study area in D ₂ , draw the elevation gray map P _elevation of each rectangular study area, and establish the elevation gray used to characterize the terrain elevation change characteristics of each railway case study area Degree atlas D _elevation ;

S _2-4 : Based on the grid elevation data information of each railway case study area in D ₂ , draw the slope grayscale map P _slope of each rectangular study area, and establish the slope grayscale used to characterize the terrain slope characteristics of each railway case study area Atlas D _slope ;

S _2-5 : Representing different railway grades as grayscale images with different grayscale values, drawing a railway grade grayscale image P _{classification} corresponding to each railway case according to the actual grades of each railway case in D ₁ , and establishing Railway grade grayscale atlas D _{classification} ;

S _2-6 : Represent different locomotive models as grayscale images with different grayscale values, draw the locomotive model grayscale image P corresponding to each railway case according to the actual locomotive models used in each railway case in D _{1 locomotive} , build locomotive model grayscale atlas D _locomotive ;

S _2-7 : Based on the established elevation grayscale atlas D _elevation , slope grayscale atlas D _slope , railway grade grayscale atlas D _{classification} , and locomotive model grayscale atlas D _locomotive , integrate each railway case in D ₁ The elevation grayscale map P _elevation , slope grayscale map P _slope , railway grade grayscale map P _{classification} and locomotive model grayscale map P _locomotive form a four-channel map P _merge that can represent the information of each railway case, and establish a data set D _merge ;

S _2-8 : Cut the four-channel image representing the information of each railway case in the data set D _merge into a picture with a size of 333×333 pixels, and assign label data. The label data is the limit actually used by each railway case slope value;

S _2-9 : Divide the labeled data graph obtained in S _2-8 at a ratio of 4:1, and establish a training data set D _train and a verification data set D _validate for training a deep convolutional neural network;

S ₃ : Use the training data set D _train established by S ₂ to train the constructed network model, and use the verification data set D _validate established by S ₂ to verify the accuracy of the model, and obtain a trained and verified deep convolutional neural network model;

S ₄ : In addition, collect N ₂ cases of passenger and freight collinear railways that are different from those in the data set D ₁ , and generate a four-channel graph P _merge representing the railway case information according to steps S _2-2 to S _2-7 , and establish Test data set D _test ;

S ₅ : Propose a sliding scanning technology, scan the trained deep convolutional neural network model from left to right and from top to bottom in the data set D _test to represent the elevation information, slope information, railway The four-channel diagram of grade information and locomotive model information determines the recommended limit slope value for each railway case in the D _test according to the output times of each limit slope value.

Further, the deep convolutional neural network model constructed in the step S1 includes ₅ convolutional layers (Conv), 3 pooling layers (Pool), 2 fully connected layers (FC) and 1 Softmax output layer :

(1) The convolution kernel size used in the first convolutional layer (Conv1) is 33×33×3, the stride size is 4, and the number of convolution kernels is 96. After Conv1, the corrected linear unit (ReLU) is connected as the non- The linear activation function makes the model have nonlinear characteristics;

(2) Conv1 is connected to the first pooling layer (Pool1) after nonlinear processing. The pooling kernel size used by Pool1 is 4×4, and the stride size is 2;

(3) Pool1 is connected to the second convolutional layer (Conv2). The convolution kernel size used by Conv2 is 3×3×96, the stride size is 1, the number of convolution kernels is 256, and the modified linear unit is connected after Conv2. (ReLU) for nonlinear processing;

(4) Conv2 is connected to the second pooling layer (Pool2) after nonlinear processing. The pooling kernel size used by Pool2 is 3×3, and the stride size is 2;

(5) The third convolutional layer (Conv3) is connected after Pool2. The convolution kernel size used by Conv3 is 3×3×256, the stride size is 1, the number of convolution kernels is 384, and the corrected linear unit is connected after Conv3. (ReLU) for nonlinear processing;

(6) Conv3 is connected to the fourth convolutional layer (Conv4) after nonlinear processing. The convolution kernel size used by Conv4 is 3×3×384, the stride size is 1, and the number of convolution kernels is 384. After Conv4 Connect the Rectified Linear Unit (ReLU) for nonlinear processing;

(7) Conv4 is connected to the fifth convolutional layer (Conv5) after nonlinear processing. The convolution kernel size used by Conv5 is 3×3×384, the stride size is 1, and the number of convolution kernels is 256. After Conv5 Connect the Rectified Linear Unit (ReLU) for nonlinear processing;

(8) Conv5 is connected to the third pooling layer (Pool3) after nonlinear processing. The pooling kernel size used by Pool3 is 3×3, and the stride size is 2;

(9) Pool3 is connected to the first fully connected layer (FC1). In order to prevent overfitting, the dropout function is used to connect the Pool3 layer to the FC1 layer, and FC1 is connected to the corrected linear unit (ReLU) for nonlinear processing;

(10) FC1 is connected to the second fully connected layer (FC2) after nonlinear processing, and the dropout function is used to prevent overfitting, and FC2 is connected to the corrected linear unit (ReLU) for nonlinear processing;

(11) After nonlinear processing, FC2 is connected to the Softmax output layer, which is used to output the limit slope value recommendation of the new railway.

Further, the railway cases collected in the step _S2-1 cover different grades of railways and different locomotive models.

Further, the division method of the railway rectangular research area in the step _S2-2 is as follows: set the start and end points of a certain railway case line as S _i : (x _Si , y _Si ) and E _i : (x _Ei , y _Ei ), the research area of the railway case is a rectangular area with S _i and E _i as the diagonal points, |x _Ei -x _Si | as the length, and |y _Ei -y _Si | as the width.

Further, in the steps S _2-3 and S _2-4 , the elevation grayscale map and gradient grayscale map of each railway case rectangular research area are drawn by Global Mapper software.

Further, in the step _S2-5 , the size of the railway grade grayscale map P _{classification} is the same as the size of the rectangular research area of the railway case.

Further, in the step _S2-6 , the size of the locomotive model grayscale map P _locomotive is the same as the size of the rectangular research area using the locomotive model railway case.

Further, the four-channel map P _merge of each railway case in the step _S2-7 is the elevation grayscale map P _elevation and slope grayscale map P _slope of each railway case using the merge function in the computer vision library OpenCV , the railway grade grayscale map P _{classification} and the locomotive type grayscale map P _locomotive are obtained after fusion.

Further, the network model constructed by training in the step S3 is the label data set _{D train} established based _on S2, and the connection weights between the layers in the network model are continuously updated through the gradient descent algorithm, as follows:

(1) Softmax layer connection weight update

The Softmax layer is used to output the limit slope value recommended by the model. This layer calculates the output probability of each limit slope value according to the output value of each neuron in the previous layer, so that the slope value with the highest output probability is selected as the limit slope value recommended by the model. Its function expression is shown in formula (2):

In the formula: P(y ⁽ⁱ⁾ =j|x ⁽ⁱ⁾ ; W) is the i-th picture as the input data, and the j-th value is selected as the probability of limiting the slope in the Softmax layer, and x ⁽ⁱ⁾ is the Softmax layer The input data (that is, the output data of the previous layer), W is the connection weight between the Softmax layer and the previous layer.

The model loss function E is established based on the Softmax function, and its function expression is shown in formula (3):

In the formula: 1{y ⁽ⁱ⁾ =j} is a logical expression, if the i input picture is labeled as the jth limit slope, then 1{y ⁽ⁱ⁾ =j}=1, otherwise 1{y ^{(i) )} =j}=0, λ is the weight attenuation coefficient.

Based on the loss function E, the residual error of each neuron in the Softmax layer can be calculated according to formula (4):

The connection weights of each neuron in the Softmax layer are updated according to formula (5) and formula (6):

(2) Fully connected layer connection weight update

Each neuron in the fully connected layer is connected to all neurons in the previous layer, and its connection weight update formula is as follows:

In the formula: W ^l is the connection weight matrix of each neuron in the current layer (fully connected layer), b ^l is the connection bias vector of each neuron in the current layer, and α is the learning rate.

The partial derivative of the loss function to the connection weights of each neuron in the fully connected layer and the partial derivative of the connection bias of each neuron in the fully connected layer It can be calculated according to formula (9) and formula (10) respectively.

In the formula: x ^l-1 is the output vector of the previous connection layer of the current layer (full connection layer), δ ^l is the residual error of each neuron in the current layer (full connection layer), which can be calculated according to the neurons of the subsequent connection layer The residual δ ^l+1 is calculated.

In the formula: W ^l+1 is the connection weight matrix of each neuron in the connection layer after the current layer (full connection layer), and f(·) is the ReLU activation function.

(3) Convolutional layer connection weight update

Each neuron in the convolution layer is connected to the previous layer through the convolution kernel, and the weight update formula of each convolution kernel connection is as follows:

In the formula: is the connection weight matrix of the dth convolution kernel of the current layer (convolutional layer), is the connection bias vector of the dth convolution kernel of the current layer (convolutional layer), and α is the learning rate.

The loss function is the partial derivative of each connection weight of the dth convolution kernel of the current layer (convolution layer) The calculation formula is as follows:

In the formula: is the output value of the d′th feature map of the previous connection layer of the current layer (convolutional layer), D ^l-1 is the number of feature maps of the previous connection layer of the current layer (convolutional layer), is the residual matrix of the dth feature map of the current layer (convolutional layer).

The partial derivative of the loss function to each connection bias of the dth convolution kernel of the current layer (convolutional layer) The calculation formula is as follows:

In the formula: is the connection bias vector of the dth feature map in the current layer (convolutional layer), and Respectively, the number of rows and columns of the dth feature map in the current layer (convolutional layer), is the residual value of row i and column j in the dth feature map of the current layer (convolutional layer).

The residual of the current layer (convolutional layer) is calculated based on the residual of the next connected layer through backpropagation. If the pooling layer is connected after the current layer (convolutional layer), then the residual matrix of the dth feature map of the current layer (convolutional layer) is calculated according to formula (17).

In the formula: X ^l-1 is the output matrix of the previous connection layer of the current layer (convolutional layer), The residual matrix of the dth feature map in the next connection layer after the current layer (convolutional layer).

If the current layer (convolutional layer) is connected to a convolutional layer, then the weight matrix of the current layer (convolutional layer) is calculated according to formula (18).

In the formula: is the residual matrix of the d′th feature map in the next connection layer after the current layer (convolutional layer), is the weight matrix of the dth layer of the d″th convolution kernel of the next connection layer after the current layer (convolutional layer), The output matrix of the dth feature map of the current layer (convolutional layer).

Further, the sliding scanning technique in step S5 is specifically as follows: when scanning a certain _four -channel image in the test data set D _test , each scan can output the limited slope of the 333×333 pixel area in the four-channel image Recommended value, after scanning the entire four-channel map, select the slope value with the most output times as the recommended limit slope value for the railway case represented by the four-channel map.

The beneficial effect of the present invention is that deep learning simulates the hierarchical structure of the brain, and can automatically obtain hierarchical multi-layer feature expressions from massive data, and explore input data and output without giving mathematical expressions. underlying patterns in the data. The scheme of the present invention adopts the deep learning algorithm to make the decision-making of limiting the slope of the newly-built railway, which is practical and feasible. Based on the convolutional neural network in the deep learning algorithm, the present invention proposes an optimal decision-making method for limiting the slope of newly-built railways. The method learns manual decision-making experience, identifies the mapping law between multi-dimensional influencing factors and limiting slopes, and realizes the determination of limiting slopes for newly-built railways. Human decision making. The scheme of the present invention adopts the sliding scanning technology, and realizes the decision-making of restricting the slope of different railway cases. The method of the invention has high degree of automation, strong practicability and high operating efficiency, and has good prospects for popularization and application.

Description of drawings

Fig. 1 is the schematic flow sheet of new-built railway limit slope optimization decision-making method of the present invention;

Fig. 2 is the deep convolutional neural network model of the embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a sliding scanning technology according to an embodiment of the present invention.

Detailed ways

In order to describe the technical content, achieved goals and effects of the present invention in detail, the following descriptions will be made in conjunction with the embodiments and accompanying drawings.

Embodiment 1 of the present invention is a kind of newly-built railway limit slope optimization decision-making method, as shown in Figure 1, this optimization decision-making method comprises the following steps:

S ₁ : Construct a deep convolutional neural network model for the optimization decision-making of new railway limit slopes. The constructed network model includes 5 convolutional layers (Conv), 3 pooling layers (Pool), and 2 fully connected layers ( FC) and 1 Softmax output layer:

(1) The convolution kernel size used in the first convolutional layer (Conv1) is 33×33×3, the stride size is 4, and the number of convolution kernels is 96. After Conv1, the corrected linear unit (ReLU) is connected as the non- The linear activation function makes the model have nonlinear characteristics;

(2) Conv1 is connected to the first pooling layer (Pool1) after nonlinear processing. The pooling kernel size used by Pool1 is 4×4, and the stride size is 2;

(3) Pool1 is connected to the second convolutional layer (Conv2). The convolution kernel size used by Conv2 is 3×3×96, the stride size is 1, the number of convolution kernels is 256, and the modified linear unit is connected after Conv2. (ReLU) for nonlinear processing;

(4) Conv2 is connected to the second pooling layer (Pool2) after nonlinear processing. The pooling kernel size used by Pool2 is 3×3, and the stride size is 2;

(5) The third convolutional layer (Conv3) is connected after Pool2. The convolution kernel size used by Conv3 is 3×3×256, the stride size is 1, the number of convolution kernels is 384, and the corrected linear unit is connected after Conv3. (ReLU) for nonlinear processing;

(6) Conv3 is connected to the fourth convolutional layer (Conv4) after nonlinear processing. The convolution kernel size used by Conv4 is 3×3×384, the stride size is 1, and the number of convolution kernels is 384. After Conv4 Connect the Rectified Linear Unit (ReLU) for nonlinear processing;

(7) Conv4 is connected to the fifth convolutional layer (Conv5) after nonlinear processing. The convolution kernel size used by Conv5 is 3×3×384, the stride size is 1, and the number of convolution kernels is 256. After Conv5 Connect the Rectified Linear Unit (ReLU) for nonlinear processing;

(8) Conv5 is connected to the third pooling layer (Pool3) after nonlinear processing. The pooling kernel size used by Pool3 is 3×3, and the stride size is 2;

(9) Pool3 is connected to the first fully connected layer (FC1). In order to prevent overfitting, the dropout function is used from the Pool3 layer to the FC1 layer, and FC1 is connected to the corrected linear unit (ReLU) for nonlinear processing;

(10) FC1 is connected to the second fully connected layer (FC2) after nonlinear processing, and the dropout function is used to prevent overfitting, and FC2 is connected to the corrected linear unit (ReLU) for nonlinear processing;

(11) After nonlinear processing, FC2 is connected to the Softmax output layer, which is used to output the recommended value of the new railway limit slope.

S ₂ : Establish a training data set D _train and a verification data set D _validate for training a deep convolutional neural network;

S _2-1 : Collect 246 cases of passenger and freight collinear railways with 6‰, 12‰, and 24‰ as limit slopes. The collected railway cases cover four railway grades: I, II, III, and IV. Three types of locomotives, Shaoshan 1, Shaoshan 3, and Shaoshan 4, establish a railway case data set D ₁ ;

S _2-2 : Divide the rectangular research area of each railway case based on the start and end positions of each railway line in the railway case data set D ₁ , and extract the grid elevation data information in each rectangular research area to establish the elevation data of the railway case set D2 _;

S _2-3 : Based on the grid elevation data information of each railway case study area in D ₂ , draw the elevation gray map P _elevation of each rectangular study area, and establish the elevation gray used to characterize the terrain elevation change characteristics of each railway case study area Degree atlas D _elevation ;

S _2-4 : Based on the grid elevation data information of each railway case study area in D ₂ , draw the slope grayscale map P _slope of each rectangular study area, and establish the slope grayscale used to characterize the terrain slope characteristics of each railway case study area Atlas D _slope ;

S _2-5 : Represent four kinds of railway grades with grayscale images with grayscale values of 0, 40, 80, and 120, and draw the corresponding railway cases according to the actual grades of each railway case in D ₁ P _{classification} of the railway grade grayscale map, and the establishment of the railway grade grayscale atlas D _{classification} ;

S _2-6 : The three types of electric locomotives of Shaoshan ₁ , Shaoshan 3 and Shaoshan 4 are represented by grayscale images with grayscale values of 160, 200, and 240 respectively, and according to the For the actual locomotive model, draw the locomotive model grayscale map P _locomotive corresponding to each railway case, and establish the locomotive model grayscale atlas D _locomotive ;

S _2-7 : Based on the established elevation grayscale atlas D _elevation , slope grayscale atlas D _slope , railway grade grayscale atlas D _{classification} , and locomotive model grayscale atlas D _locomotive , integrate each railway case in D ₁ The elevation grayscale map P _elevation , slope grayscale map P _slope , railway grade grayscale map P _{classification} and locomotive model grayscale map P _locomotive form a four-channel map P _merge that can represent the information of each railway case, and establish a data set D _merge ;

S _2-8 : Cut the four-channel image representing the information of each railway case in the data set D _merge into a picture with a size of 333×333 pixels, and assign label data. The label data is the limit actually used by each railway case slope value;

S _2-9 : Divide the labeled pictures obtained in S _2-8 in a ratio of 4:1, and establish a training data set D _train and a verification data set D _validate for training a deep convolutional neural network;

S ₃ : Use the training data set D _train established by S ₂ to train the constructed network model, and use the verification data set D _validate established by S ₈ to verify the accuracy of the model, and obtain a trained and verified deep convolutional neural network model. The training and verification took 9 hours and 35 minutes (i7 processor, 16G memory and GTX 1080 graphics card), and the deep convolutional neural network model with an accuracy of 83.35% was obtained.

S ₄ : Collect another 36 cases of passenger and freight collinear railways that are different from those in data set D ₁ , and follow steps S _2-2 to S _2-7 to establish a four-channel graph P _merge representing the railway case information, and establish a test Data set _Dtest ;

S ₅ : Propose a sliding scanning technology, scan the trained deep convolutional neural network model from left to right and from top to bottom in the data set D _test to represent the elevation information, slope information, railway The four-channel diagram of grade information and locomotive model information, and according to the output times of each limit slope value, determine the limit slope recommendation value of each railway case in the D _test . Among the 36 railway cases tested in this test, the limit slopes of 34 railway cases have been accurately decided (that is, the limit slope values recommended by the model are the same as the limit slope values determined manually), and the accuracy rate can reach 94.44%.

The sliding scanning technology referred to in the present invention refers to scanning the entire picture, and determining the limit slope value according to the output times of different limit slopes.

In summary, the present invention provides a new railway limit slope optimization decision-making method, first constructs a deep convolutional neural network model, then establishes a railway case database, and represents various factors that affect the new railway limit slope decision into a grayscale map, And fused into a multi-channel image for training the network model; finally, a sliding scanning technology is proposed, combined with the trained deep convolutional neural network model to make a new railway limit slope decision.

The above description is only an embodiment of the present invention, and does not limit the patent scope of the present invention. All equivalent transformations made by using the description of the present invention and the contents of the accompanying drawings, or directly or indirectly used in related technical fields, are all included in the same principle. Within the scope of patent protection of the present invention.

Claims (9)

1. a kind of newly built railway ruling grade Study on Decision-making Method for Optimization, it is characterised in that: the following steps are included:

S₁: building is used for the depth convolutional neural networks model of newly built railway ruling grade Optimal Decision-making；

S₂: establish the training dataset D for training depth convolutional neural networks_trainWith validation data set D_validate；

S_2-1: collect N₁Item uses the built mixed passenger and freight railway case of different ruling grades, establishes railway case data collection D₁；

S_2-2: it is based on the railway case data collection D₁In the start, end position of each rail track divide the square of each railway case Shape survey region, and the grid altitude data information in each rectangle survey region is extracted, establish railway case altitude data collection D₂；

S_2-3: it is based on D₂In each railway case study region grid altitude data information, draw the elevation of each rectangle survey region Grayscale image P_elevation, establish the elevation gray scale atlas for characterizing each railway case study region landform altitude variation characteristic D_elevation；

S_2-4: it is based on D₂In each railway case study region grid altitude data information, draw the gradient of each rectangle survey region Grayscale image P_slope, establish the gradient gray scale atlas D for characterizing each railway case study region terrain slope feature_slope；

S_2-5: different classifications of rail are characterized as the different grayscale image of gray value, according to D₁In each railway case actual grade, Draw classification of rail grayscale image P corresponding with each railway case_{classification}, establish classification of rail gray scale atlas D_{classification}；

S_2-6: different type of locomotive are characterized as the different grayscale image of gray value, according to D₁In it is real used in each railway case Border type of locomotive draws type of locomotive grayscale image P corresponding with each railway case_locomotive, establish type of locomotive gray scale Atlas D_locomotive；

S_2-7: the elevation gray scale atlas D based on foundation_elevation, gradient gray scale atlas D_slope, classification of rail gray scale atlas D_{classification}, type of locomotive gray scale atlas D_locomotive, merge D₁In each railway case elevation grayscale image P_elevation、 Gradient grayscale image P_slope, classification of rail grayscale image P_{classification}With type of locomotive grayscale image P_locomotive, formation can characterize respectively The four-way figure P of railway case information_merge, and establish data set D_merge；

S_2-8: by data set D_mergeIt is 333 × 333 pixels that the four-way figure of each railway case information of middle characterization, which is cut into size, Picture, and assign label data, the ruling grade value that label data is actually used by each railway case；

S_2-9: by S_2-8Middle gained tape label datagram presses the ratio cut partition of 4:1, establishes for training depth convolutional neural networks Training dataset D_trainWith validation data set D_validate；

S₃: use S₂The training dataset D of foundation_trainThe constructed network model of training, and use S₂The validation data set of foundation D_validateModel accuracy is verified, is obtained by training and the depth convolutional neural networks model verified；

S₄: in addition collect N₂Item and data set D₁Middle different built mixed passenger and freight railway case, and according to step S_2-2To S_2-7It is raw At the four-way figure P of characterization railway case information_merge, establish test data set D_test；

S₅: a kind of slip sweep is proposed, by trained depth convolutional neural networks model by from left to right, from top to bottom Sequential scan data set D_testMiddle characterization each railway case elevation information, grade information, classification of rail information, type of locomotive The four-way figure of information determines D according to the output times of each ruling grade value_testIn the ruling grade of each railway case push away Recommend value.

2. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S₁In, The depth convolutional neural networks model includes 5 convolutional layers, 3 pond layers, 2 full articulamentums and 1 Softmax output Layer:

1) the convolution kernel size that first convolutional layer uses is 33 × 33 × 3, step size 4, and convolution kernel number is 96, first Connection amendment linear unit makes model have nonlinear characteristic as nonlinear activation function after a convolutional layer；

2) first convolutional layer connects first pond layer, the pond core size that first pond layer uses after Nonlinear Processing It is 4 × 4, step size 2；

3) second convolutional layer is connected after first pond layer, the convolution kernel size that second convolutional layer uses for 3 × 3 × 96, Step size is 1, and convolution kernel number is 256, and connection amendment linear unit carries out Nonlinear Processing after second convolutional layer；

4) second convolutional layer connects second pond layer, the pond core size that second pond layer uses after Nonlinear Processing It is 3 × 3, step size 2；

5) third convolutional layer is connected after second pond layer, the convolution kernel size that third convolutional layer uses for 3 × 3 × 256, Step size is 1, and convolution kernel number is 384, and connection amendment linear unit carries out Nonlinear Processing after third convolutional layer；

6) third convolutional layer connects the 4th convolutional layer, the convolution kernel size that the 4th convolutional layer uses after Nonlinear Processing It is 3 × 3 × 384, step size 1, convolution kernel number is 384, and connection amendment linear unit carries out non-after the 4th convolutional layer Linear process；

7) the 4th convolutional layer connects the 5th convolutional layer, the convolution kernel size that the 5th convolutional layer uses after Nonlinear Processing It is 3 × 3 × 384, step size 1, convolution kernel number is 256, and connection amendment linear unit carries out non-after the 5th convolutional layer Linear process；

8) the 5th convolutional layer connects third pond layer, the pond core size that third pond layer uses after Nonlinear Processing It is 3 × 3, step size 2；

9) first full articulamentum is connected after the layer of third pond, to prevent over-fitting, layer is arrived in third pond First full articulamentum connection uses dropout function, and connection amendment linear unit carries out non-linear after first full articulamentum Processing；

10) first full articulamentum connects second full articulamentum after Nonlinear Processing, and prevents from producing using dropout function Over-fitting is given birth to, connection amendment linear unit carries out Nonlinear Processing after second full articulamentum；

11) second full articulamentum connects Softmax output layer after Nonlinear Processing, for exporting newly built railway ruling grade Recommendation.

3. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S_2-1 In, the railway case of collection covers different brackets railway and different type of locomotive.

4. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S_2-2 In, the method for dividing rectangle survey region based on rail track start, end position is as follows:

If certain railway case route start, end are respectively S_i: (x_Si,y_Si) and E_i: (x_Ei,y_Ei), then the research of the railway case Region is with S_iAnd E_iFor angle steel joint, with | x_Ei-x_Si| to grow, | y_Ei-y_Si| it is wide rectangular area.

5. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S_2-5 In, classification of rail grayscale image P_{classification}Size it is identical as the rectangle survey region size of the railway case.

6. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S_2-6 In, type of locomotive grayscale image P_locomotiveSize with using the type of locomotive railway case rectangle survey region size phase Together.

7. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S_2-7 In each railway case four-way figure P_mergeTo use the merge function in the OpenCV of computer vision library to each railway case The elevation grayscale image P of example_elevation, gradient grayscale image P_slope, classification of rail grayscale image P_{classification}With locomotive type gray scale Scheme P_locomotiveIt is obtained after being merged.

8. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S₃In The constructed network model of training is based on S₂The label data collection D of foundation_train, network is constantly updated by gradient descent algorithm The connection weight of each interlayer in model, specific as follows:

1) Softmax layers of connection weight update:

The Softmax layers of ruling grade value recommended for output model, this layer are calculated according to the output valve of each neuron of preceding layer The output probability of each ruling grade value, thus the ruling grade value for selecting the maximum value of slope of output probability to recommend as model, Shown in its function representation such as formula (2):

In formula: P (y⁽ⁱ⁾=j | x⁽ⁱ⁾；It W) is to select j-th of value as limit in Softmax layer choosing using the i-th picture as input data The probability of the gradient processed, x⁽ⁱ⁾For Softmax layers of input data, W is the connection weight of Softmax layers with preceding layer；

Model loss function E is established based on Softmax function, shown in function expression such as formula (3):

In formula: 1 { y⁽ⁱ⁾=j } it is logical expression, if i-th input picture mark is j-th of ruling grade value, 1 { y⁽ⁱ⁾=j }=1, otherwise 1 { y⁽ⁱ⁾=j }=0, λ be weight attenuation coefficient；

Based on loss function E, the residual error of Softmax layers of each neuron is calculated by formula (4):

The connection weight of Softmax layers of each neuron is updated by formula (5), formula (6):

2) full articulamentum connection weight updates:

Each neuron of full articulamentum is connected with upper one layer of all neurons, and connection weight more new formula is as follows:

In formula: W^lFor the connection weight matrix of each neuron of current layer, b^lFor the connection bias vector of each neuron of current layer, α is Learning rate；

Partial derivative of the loss function to each neuron connection weight of full articulamentumIt is connected partially with to each neuron of full articulamentum The partial derivative setIt is calculated respectively by formula (9) and formula (10)；

In formula: x^l-1For the output vector of an articulamentum on current layer, δ^lIt, can be according to connecting thereafter for the residual error of each neuron of current layer Meet the residual error δ of each neuron of layer^l+1It calculates；

δ^l=(W^l+1)δ^l+1⊙f′(W^lx^l-1+b^l) (11)

In formula: W^l+1For the connection weight matrix of each neuron of articulamentum after current layer, f () is ReLU activation primitive；

3) convolutional layer connection weight updates:

Each neuron of convolutional layer is connected by convolution kernel with preceding layer, and each convolution kernel connection weight more new formula is as follows:

In formula:For the connection weight matrix of d-th of convolution kernel of current layer,It is biased for the connection of d-th of convolution kernel of current layer Vector, α are learning rate；

Loss function is to each connection weight partial derivative of d-th of convolution kernel of current layerCalculation formula it is as follows:

In formula:For the output valve of the previous a characteristic pattern of articulamentum d ' of current layer, D^l-1For the spy of the previous articulamentum of current layer Map number is levied,For the residual matrix of d-th of characteristic pattern of current layer；

Loss function respectively connects biasing partial derivative to d-th of convolution kernel of current layerCalculation formula it is as follows:

In formula:For the connection bias vector of d-th of characteristic pattern in current layer,WithD-th of feature respectively in current layer The line number and columns of figure,For i row in d-th of characteristic pattern in current layer, the residual values of j column；

The residual error of current layer is the layer residual computations by backpropagation, based on latter connection；If latter linked current layer is pond Change layer, then the residual matrix of d-th of characteristic pattern of current layer is calculated by formula (17)；

In formula: X^l-1For the output matrix of the previous articulamentum of current layer,D-th characteristic pattern is residual in the latter articulamentum of current layer Poor matrix；

If latter linked current layer is convolutional layer, the weight matrix of current layer is calculated by formula (18):

In formula:For the residual matrix of a characteristic pattern of d ' in the latter articulamentum of current layer,For the latter company of current layer D layers of weight matrix of a convolution kernel of d " of layer are connect,The output matrix of d-th of characteristic pattern of current layer.

9. newly built railway ruling grade Study on Decision-making Method for Optimization according to claim 1, it is characterised in that: the step S₅In, As scan test data collection D_testIn certain four-way figure when, it is big to scan 333 × 333 pixels in the exportable four-way figure every time The ruling grade recommendation of zonule selects the value of slope conduct for then taking output times most after completing whole four-way figure scanning The ruling grade recommendation of the characterized railway case of the four-way figure.