CN102521651A - Bow net contact force prediction method based on NARX neural networks - Google Patents

Abstract

The invention discloses a pantograph-catenary contact force prediction method based on a NARX neural network in the technical field of railway safety operation control. Including: obtaining test data through simulation tests; normalizing the test data; establishing a NARX neural network prediction model; extracting the first set number of data pairs as training samples from the normalized test data, using The Bayesian regularization algorithm trains the NARX neural network prediction model; and then extracts a second set number of data pairs as test samples from the normalized test data, and inputs the test samples as input data into the trained NARX In the neural network prediction model, the output results are denormalized and used as the prediction value of pantograph-catenary contact force. The invention adopts the NARX neural network model to predict the pantograph-catenary contact force, and improves the prediction accuracy of the panto-catenary contact force.

Description

Prediction method of pantograph-catenary contact force based on NARX neural network

technical field

The invention belongs to the technical field of railway safety operation control, in particular to a pantograph-catenary contact force prediction method based on a NARX neural network.

Background technique

The development of high-speed railways is an inevitable trend of my country's railway modernization, and electric locomotives have become the main force in high-speed locomotives due to their advantages such as large capacity, high speed, low energy consumption, low pollution, cheap transportation, and safety and reliability. In the high-speed electrified railway system, the pantograph-catenary current receiving system is directly related to the train speed, that is, when the train is running at high speed, it must maintain a stable current receiving state, that is to say, there must be a certain contact between the pantograph and the contact wire pressure. When the contact pressure is too small, it is easy to cause off-line, that is, the pantograph breaks away from the contact wire and generates an arc; when the contact pressure is too high, the contact wire lifts too much, causing the contact wire to locally bend, causing fatigue damage to the contact wire, and making the contact Wire wear increases, causing pantograph-catenary accidents in severe cases. Therefore, the measurement of pantograph-catenary contact force is of great significance to ensure the safety of trains and the development of high-speed railways in my country.

Yuan Hua, Bi Jihong and others used the method of establishing a "contact pair" between pantograph and catenary to realize pantograph-catenary coupling, established a rigid suspension catenary coupling model, and performed nonlinear transient dynamic analysis on the model to obtain the contact force and the vertical direction of the skateboard. The displacement curve. Xu Haidong, Xu Min and others established a pantograph-catenary dynamics model based on the dynamics theory of large railway systems, and applied the vibration response at the pantograph base on the top of the locomotive as the excitation of the pantograph-catenary system to the dynamic model. Effect of rail-coupled vibration on pantograph-catenary contact pressure.

Scholars at home and abroad have researched pantograph-catenary methods on-site, semi-physical and semi-virtual (the pantograph is the real object, and the catenary is obtained by computer simulation) test and computer simulation. Field tests have extremely high requirements for measurement methods and data processing, and only a few countries such as Germany can directly measure the contact force between the pantograph and the catenary. With the emergence of high-speed electronic computers, it is possible to use numerical methods to fully simulate the huge structure of catenary, and computer simulation methods have become the most general research method.

There are basically two methods to realize the coupling between catenary and pantograph: one is proposed by Wu Tianxing et al. to realize the coupling by equal displacement of bow and net at the contact point, and the other is to achieve coupling by spring proposed by Zhang Weihua et al. The former does not need to select the contact stiffness, but offline cannot be considered; for the latter, because the actual relationship between the dynamic contact force between the pantograph and the catenary and the displacement of the slider is: when they are in contact, the relationship between the two maintains the contact stiffness; once offline, the contact force will always be zero, there is no longer any connection between the two. Therefore, using the spring to simulate the pantograph-network coupling cannot study the vibration of the bow-network after offline.

It is also possible to use the life-death element technology in the engineering design software ANSYS to simulate the movement of the pantograph along the catenary. Activate or kill the corresponding pantograph. The biggest defect of this method is that it cannot consider the nodal velocity of the pantograph, the continuity and transferability of displacement, so the calculation results are very different from the facts and are very inaccurate.

From the above introduction, it can be seen that the existing pantograph-catenary contact force prediction methods have different degrees of defects, so it is necessary to propose a new pantograph-catenary contact force prediction method to improve the accuracy of pantograph-catenary contact force prediction.

Contents of the invention

The purpose of the present invention is to provide a pantograph-catenary contact force prediction method based on NARX neural network to solve the problem that the precision of the panto-catenary contact force calculated by the commonly used pantograph-catenary contact force prediction method is not high.

In order to achieve the above object, the technical solution adopted in the present invention is, a kind of pantograph-catenary contact force prediction method based on NARX neural network, it is characterized in that described method comprises:

Step 1: Obtain test data through simulation tests; where the test data includes contact line irregularity data and corresponding pantograph-catenary contact force data;

Step 2: Normalize the test data;

Step 3: Establish NARX neural network prediction model;

Step 4: From the normalized test data, extract the first set number of data pairs as training samples; wherein, the data pair refers to the normalized contact line irregularity data and the corresponding normalized Pantograph-catenary contact force data after processing;

Step 5: Take the normalized contact line irregularity data in the training sample and the corresponding normalized pantograph-catenary contact force data as the input data and output data respectively, and use the Bayesian regularization algorithm to train NARX neural network prediction model;

Step 6: From the normalized test data, extract a second set number of data pairs as test samples; wherein, the data pairs refer to the normalized contact line irregularity data and the corresponding normalized Pantograph-catenary contact force data after normalization processing;

Step 7: Input the normalized contact line irregularity data in the test sample into the NARX neural network prediction model trained in step 5, denormalize the output results, and denormalize The output after processing is used as the predicted value of pantograph-catenary contact force.

Specifically, the acquisition of test data through simulation tests is to first establish a pantograph-catenary coupling dynamic model, and then use MATLAB/Simulink software to perform dynamic simulation to obtain contact line irregularity data and corresponding pantograph-catenary contact force data.

The described experiment data is carried out normalization process specifically is to utilize formula

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; scale</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}$

n为试验数据的个数。Normalize the experimental data x _i ; where,

n is the number of test data.

The middle layer node of described NARX neural network forecasting model adopts tan-sigmoid function, the output layer node adopts linear function, the number of input layer nodes is 1, the number of middle layer nodes is 15, the number of output layer nodes is 1, and the input and output delays are both is 45; wherein, the tan-sigmoid function is x is the input data of the hidden layer, T is the scaling coefficient, and θ is the displacement coefficient.

The first set number of data pairs is 1300 data pairs.

The second set number of data pairs is 700 data pairs.

After said step 7, also include the step of adopting the root mean square error method to evaluate the performance of the NARX neural network prediction model, specifically using the formula

$\mathrm{<span\; class="notranslate">\; RMSE</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Evaluate the performance of the NARX neural network prediction model after training; among them, y(i) is the target value in the test sample, y _m (i) is the predicted value after denormalization processing, and N is the number of data in the test sample number.

The invention adopts the NARX neural network model to predict the pantograph-catenary contact force, and improves the prediction accuracy of the panto-catenary contact force.

Description of drawings

Fig. 1 is the flow chart of pantograph-catenary contact force prediction method based on NARX neural network;

Figure 2 is a structural diagram of the NARX neural network;

Figure 3 is a comparison chart between pantograph-catenary contact force test data output and NARX neural network output;

Figure 4 is the correlation analysis between pantograph-catenary contact force test data output and NARX neural network output.

Detailed ways

The preferred embodiments will be described in detail below in conjunction with the accompanying drawings. It should be emphasized that the following description is only exemplary and not intended to limit the scope of the invention and its application.

Figure 1 is a flowchart of the pantograph-catenary contact force prediction method based on NARX neural network. In Figure 1, the pantograph-catenary contact force prediction method based on NARX neural network includes:

Step 1: Obtain test data through simulation tests; where the test data includes contact line irregularity data and corresponding pantograph-catenary contact force data.

To obtain test data through simulation experiments, the pantograph-catenary coupling dynamic model is established first, and then dynamic simulation is performed using MATLAB/Simulink software to obtain contact line irregularity data and corresponding pantograph-catenary contact force data.

Step 2: Normalize the test data.

The normalization of the test data includes the normalization of the contact line irregularity data and the corresponding pantograph-catenary contact force data. The normalization process specifically uses the formula

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; scale</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}$

n为试验数据的个数，x_i为接触线不平顺数据/弓网接触力数据。Normalize the contact line irregularity data and pantograph-catenary contact force data; among them,

n is the number of test data, and _xi is the contact line irregularity data/pantograph-catenary contact force data.

Step 3: Establish NARX neural network prediction model.

x为隐层的输入数据，T为缩放系数，θ为位移系数。The structure of NARX neural network (Nonlinear Auto-Regressive with eXogenous input NeuralNetworks) is shown in Figure 2. The middle layer node of the NARX neural network prediction model uses the tan-sigmoid function, the output layer node uses the linear function, the number of input layer nodes is 1, the number of middle layer nodes is 15, the number of output layer nodes is 1, and the input and output delays are both 45 . Among them, the tan-sigmoid function is

x is the input data of the hidden layer, T is the scaling coefficient, and θ is the displacement coefficient.

Step 4: From the normalized test data, extract 1300 data pairs as training samples; where, the data pair refers to the normalized contact line irregularity data and the corresponding normalized Pantograph-catenary contact force data.

Step 5: Take the normalized contact line irregularity data in the training sample and the corresponding normalized pantograph-catenary contact force data as the input data and output data respectively, and use the Bayesian regularization algorithm to train NARX Neural Network Forecasting Model.

Bayesian regularization (BR, Bayesian Regularization) algorithm refers to the establishment of a special performance parameter composed of output errors, weights and thresholds of each layer in the training process in order to improve the network promotion ability, and optimize it according to Levenberg-Martquartdt Theory adjusts the weights and thresholds of the network to minimize this parameter.

Step 6: From the normalized test data, extract 700 data pairs as test samples; where, the data pairs refer to the normalized contact line irregularity data and the corresponding normalized The pantograph-catenary contact force data.

Step 7: Input the normalized contact line irregularity data in the test sample into the NARX neural network prediction model trained in step 5, denormalize the output results, and denormalize The output after processing is used as the predicted value of pantograph-catenary contact force.

The root mean square error method can also be used to evaluate the performance of the above-mentioned NARX neural network prediction model, specifically using the formula

$\mathrm{<span\; class="notranslate">\; RMSE</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

Evaluate the performance of the NARX neural network prediction model after training; among them, y(i) is the target value in the test sample, y _m (i) is the predicted value after denormalization processing, and N is the number of data in the test sample number.

The smaller the RMSE value, the higher the prediction accuracy of the model, and the closer the predicted value is to the target value. Secondly, the curve fitting between the model output and the target output can intuitively reflect the degree of approximation between the target output value and the model output value, as shown in Figure 3. Finally, linear regression analysis is performed on the model output and the target output, and the correlation coefficient between the target output value and the model output value can be accurately calculated, as shown in Figure 4, where A neural network output data, T represents the test output data .

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present invention can easily think of changes or Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (7)

1. A bow net contact force prediction method based on a NARX neural network is characterized by comprising the following steps:

step 1: obtaining test data through a simulation test; the test data comprises contact line irregularity data and bow net contact force data corresponding to the contact line irregularity data;

step 2: carrying out normalization processing on the test data;

and step 3: establishing a NARX neural network prediction model;

and 4, step 4: extracting a first set number of data pairs from the test data after the normalization processing to be used as training samples; the data pairs refer to contact line irregularity data after normalization processing and pantograph-catenary contact force data after normalization processing corresponding to the contact line irregularity data;

and 5: respectively taking the contact line irregularity data after normalization processing in the training sample and the bow net contact force data after normalization processing corresponding to the contact line irregularity data as input data and output data, and training an NARX neural network prediction model by adopting a Bayesian regularization algorithm;

step 6: extracting a second set number of data pairs from the normalized test data to serve as test samples; the data pairs refer to contact line irregularity data after normalization processing and pantograph-catenary contact force data after normalization processing corresponding to the contact line irregularity data;

and 7: inputting the contact line irregularity data after normalization processing in the test sample as input data into the NARX neural network prediction model trained in the step 5, performing inverse normalization processing on an output result, and taking the output result after the inverse normalization processing as a bow net contact force prediction value.

2. The method as claimed in claim 1, wherein the obtaining of the test data through the simulation test is specifically that a bow-net coupling dynamics model is established, and then MATLAB/Simulink software is used for performing dynamic simulation to obtain contact line irregularity data and bow-net contact force data corresponding to the contact line irregularity data.

3. The method of claim 1, wherein the normalizing the test data is performed using a formula

$x_i^{\mathrm{scal}}=\frac{x_i-x_{\min }}{x_{\max }-x_{\min }}$

For test data x_iCarrying out normalization processing; wherein,

n is the number of test data.

4. The method of claim 1, wherein the NARX neural network prediction model employs tan-sigmoid function for intermediate layer nodes, linear function for output layer nodes, number of input layer nodes is 1, number of intermediate layer nodes is 15, number of output layer nodes is 1, input and output delays are all 45; wherein the tan-sigmoid function isx is input data of the hidden layer, T is a scaling coefficient, and theta is a displacement coefficient.

5. The method of claim 1, wherein the first set number of data pairs is 1300 data pairs.

6. The method of claim 1, wherein the second set number of data pairs is 700 data pairs.

7. The method as claimed in claim 1, wherein said step 7 is followed by the step of evaluating the performance of the NARX neural network prediction model using root mean square error, in particular using a formula

<math> <mrow> <mi>RMSE</mi> <mrow> <mo>(</mo> <mi>y</mi> <mo>,</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mo>)</mo> </mrow> <mo>=</mo> <msqrt> <mfrac> <mn>1</mn> <mi>N</mi> </mfrac> <msubsup> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </msubsup> <msup> <mrow> <mo>(</mo> <mi>y</mi> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>-</mo> <msub> <mi>y</mi> <mi>m</mi> </msub> <mrow> <mo>(</mo> <mi>i</mi> <mo>)</mo> </mrow> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mrow> </math>

Evaluating the performance of the trained NARX neural network prediction model; wherein y (i) is a target value in the test sample, y_m(i) And N is the number of data in the test sample.

Images