CN103714382B - A kind of multi-index comprehensive evaluation method for reliability of urban rail train security detection sensor network - Google Patents

Abstract

The invention discloses a kind of multi-index comprehensive evaluation method for reliability of urban rail train security detection sensor network belonging to modern traffic safety technical field.By summing up a series of indexs affecting rail train safety detection sensor network reliability, choose time delay, occupation rate, packet loss and bit error rate four indices as evaluation index.Use fuzzy mathematics theory that every evaluation index measured value is normalized, show that membership function value is as sample, build BP neutral net, training sample, using steepest descent method amendment weights, threshold value, such repetition training sample, until real output value and calculating output valve error are positioned at acceptable scope, training terminates, and the BP neutral net with expertise obtained can carry out comprehensive assessment to rail train safety detection sensor network reliability.The method is that assessment rail train safety detection sensor network performance provides brand-new method, provides theory and practice support for optimizing rail train safety detection sensor network performance.

Description

Multi-index comprehensive evaluation method for reliability of urban rail train safety detection sensor network

Technical Field

The invention belongs to the technical field of modern traffic safety. In particular to a multi-index comprehensive evaluation method for reliability of an urban rail train safety detection sensor network.

Background

The urban rail transit train is used as one of main transportation carriers of urban large-capacity public transportation, and has extreme importance and criticality to the whole urban rail transit for guaranteeing the safe operation of the urban rail transit train. The construction of the urban rail train safety detection sensor network is an important premise for on-the-way monitoring and safety early warning of urban rail trains, and is a necessary basis for formulating traffic safety guarantee measures such as urban rail train safety management strategies, urban rail train system fault detection, urban rail train accident cause analysis and the like, the premise that the urban rail train safety detection sensor network has high reliability is used for constructing the urban rail train safety detection sensor network, and the necessary condition for guaranteeing the safe operation of trains is provided.

At present, the reliability performance evaluation of the urban rail train safety detection sensor network is mostly based on a single index, such as time delay, packet loss rate, bit error rate and the like, or multi-index evaluation introducing subjective factors, such as an expert scoring method. However, the reliability of the safety detection sensor network cannot be comprehensively and objectively reflected by a single index or the score of an expert, so that the reliability of the safety detection sensor network of the urban rail train can be more comprehensively, objectively, reliably and accurately evaluated only by objectively and comprehensively evaluating multiple indexes.

Artificial Neural Networks (ans) are Artificial intelligence technologies that have been developed rapidly in the 80 th century in the 20 th century, are information processing systems that have been proposed based on human knowledge of brain organization structures and activity mechanisms, and are complex network systems that can process information in parallel, perform nonlinear conversion and self-organization adjustment, and have adaptive learning capabilities and high fault tolerance. The neural network takes the input sample and the actual output as training samples, training is carried out on the samples for enough times through calculation, so that the neural network successfully establishes a mapping relation between the input and the output, and the trained neural network model can solve similar problems. The development of the neural network is a system, and the most representative systems are a BP neural network, an adaptive resonance network, a Hopfield network, a self-organizing mapping network and the like. The BP neural network is a feedforward network taking error back propagation learning as an algorithm, a gradient steepest descent method is used for searching, and weight values and threshold values of all layers of the BP neural network are gradually recursively solved according to the principle that the mean square error of a calculated output value and an actual output value is minimum. And if the difference between the calculated output value and the actual output value does not meet the requirement, adjusting the weight value and the threshold value to recalculate until the difference between the calculated output value and the actual output value meets the requirement, and finishing the calculation to obtain the network model.

Disclosure of Invention

The invention aims to provide a multi-index comprehensive evaluation method for the reliability of an urban rail train safety detection sensor network, which is characterized by selecting indexes influencing the reliability of the urban rail train safety detection sensor network, selecting four indexes of time delay, occupancy rate, packet loss rate and bit error rate as network input samples, and evaluating the reliability of the urban rail train safety detection sensor network by combining a BP neural network and a fuzzy mathematical theory, and comprises the following steps:

(1) processing a sample, measuring four indexes such as time delay, occupancy rate, packet loss rate, error rate and the like, normalizing the measured value by using a membership function in a fuzzy mathematical theory, calculating a corresponding membership function value through a formula to enable the membership function value to have relative comparability, quantizing the membership function value into a comparable interval (-1,1), and taking the membership function value as a sample input value of a BP neural network input layer, wherein a sample output value is a corresponding reliability score of the sensor network;

(2) and (3) constructing a BP neural network, determining a network structure and establishing a BP neural network model. The BP neural network comprises an input layer, a plurality of hidden layers and an output layer; adopting a three-layer BP neural network which is simplest and most widely applied and is provided with an input layer, a hidden layer and an output layer as a constructed BP neural network structure;

(3) training a sample, inputting membership function values of all indexes, calculating an output value of a hidden layer of the BP neural network through a formula, wherein the output value of the hidden layer is an input value of an output layer, and calculating the output value of the output layer through the formula;

(4) adjusting the weight, adjusting the weight according to the difference between the actual output value and the calculated output value, returning to the third step of the scheme, recalculating the hidden layer output value and the output layer output value, adjusting the weight, calculating, adjusting the weight, recalculating until the error between the actual output value and the calculated output value of the BP neural network is less than the set minimum value, namely reaching the network convergence state, finishing the calculation and obtaining the final weight;

(5) and (4) calculating the reliability, wherein the BP neural network obtained at the moment has expert experience knowledge, and the reliability of the urban rail train safety detection sensor network can be calculated according to various index measurement values by using the model.

The BP neural network is constructed in the step (2), because the influence of each evaluation index on the reliability of the urban rail train safety detection sensor network is different, in order to enable each evaluation index to be compared mutually in a system and to have relative comparability, each evaluation index needs to be normalized by using a fuzzy mathematical theory, each processed evaluation index can be used as a training sample of the BP neural network, the time delay, the occupancy rate, the packet loss rate and the error rate are quantitative indexes, and the difference is that the time delay, the packet loss rate and the error rate are reverse indexes, namely the smaller the index is, the better the utilization rate is, the larger the index is, the better the utilization rate is, namely the larger the index is.

The invention has the beneficial effects that: the invention measures a plurality of performance indexes which affect the reliability of the sensor network, tests four index data such as time delay, occupancy rate, packet loss rate and error rate according to the safety detection sensor network of the urban rail train, combines a fuzzy mathematical theory, applies a BP neural network, carries out comprehensive, objective, reliable and accurate evaluation on the reliability of the safety monitoring sensor network of the urban rail train, and scores the reliability of the sensor network, so that the reliability of the sensor network is more intuitive and is convenient for people to understand. The method is characterized in that the measured values of all indexes are normalized by using a fuzzy mathematical theory, membership function values of the measured values are calculated, the measured values are made to be comparable, a BP neural network with expert knowledge is obtained by constructing the BP neural network and carrying out sufficient training on samples, the reliability of the urban rail train safety detection sensor network is evaluated, reference and basis are provided for overall evaluation of the urban rail train safety detection sensor network, and technical support is provided for guaranteeing safe operation of trains.

Drawings

FIG. 1 is a flow chart of the operation of the present invention.

FIG. 2 is a topological structure diagram of a three-layer BP neural network according to the present invention.

Detailed Description

The invention provides a multi-index comprehensive evaluation method for reliability of an urban rail train safety detection sensor network, which is described in detail in the following preferred embodiment with reference to the attached drawings;

example 1:

FIG. 1 is a flow chart of a comprehensive evaluation method for reliability of a sensor network for urban rail train safety detection, which comprises the following steps:

(1) the sample is processed.

Because the influence of each evaluation index on the reliability of the urban rail train safety detection sensor network is different, in order to enable each evaluation index to be compared mutually in a system and have relative comparability, each evaluation index needs to be normalized by applying a fuzzy mathematical theory, and each processed evaluation index can be used as a training sample of the BP neural network. The delay, the utilization rate, the packet loss rate and the bit error rate are all quantitative indexes, the difference is that the delay, the packet loss rate and the bit error rate are reverse indexes, namely the smaller the index is, the better the utilization rate is, and the larger the index is, the better the utilization rate is. Before the reliability of the train safety detection sensor network is comprehensively evaluated by using a BP neural network, the actual measurement values of all evaluation indexes are normalized by using a fuzzy mathematical theory, and the measurement values are normalized to be within a dimensionless interval of (-1,1) according to a membership function.

Normalization formula of forward indicator:

$A_k^t=(a_k^t-{\stackrel{\‾}{p}}_k)/\left|{\stackrel{\‾}{p}}_k\right|---\left(1\right)$

normalization formula of reverse index:

$A_k^t=({\stackrel{\‾}{p}}_k-a_k^t)/\left|{\stackrel{\‾}{p}}_k\right|---\left(2\right)$

converting the index measurement value into a membership function value on an interval [ -1,1 ]:

$x_k^t=(1-e^{-A_k^t})/(1+e^{-A_k^t})---\left(3\right)$

wherein,for the ith measurement value of the kth evaluation index,the average of all measured values for the kth evaluation index,in order to convert the variables in the middle of the run,and the corresponding membership function value of the ith measurement value of the kth evaluation index in the interval (-1,1), namely the input value of the BP neural network input layer.

It can be seen from the above specific conversion process that the larger the measurement value of the forward indicator is, the larger the membership function value corresponding thereto is, and when the measurement value reaches a certain height, the membership function value approaches "saturation", and the application of equation (3) can prevent the inaccurate comprehensive evaluation caused by the excessively large measurement value of the evaluation indicator.

(2) And (3) constructing a BP neural network, determining a network structure and establishing a network model.

The BP neural network structure includes an input layer, an output layer, and one or more hidden layers. In general, a hidden layer can enable a neural network to have certain precision and expression capability; the invention adopts a three-layer BP neural network with an input layer, a hidden layer and an output layer as a constructed BP neural network structure.

The number of the input layer units of the BP neural network is determined by the number of reliability evaluation indexes of the train safety detection sensor network. The invention adopts four indexes of time delay, utilization rate, packet loss rate and error rate to evaluate. Let input layer vector X ═ X¹，x²，...，xⁿ)^T，k is (1, 2., p), and p is the number of input layer units, and in the present invention, the number of input layer units p is 4. Wherein xⁱFor the (i) th sample,the input layer input value is the membership function value corresponding to the ith measurement value of the kth evaluation index in the interval (-1,1), and n is the sample number.

One output layer unit is used for reliability score of the train safety detection sensor network, and a percentile grading method is adopted in the invention. The output vector is given by (Y)¹，y²，...，yⁿ)^T，s is (1,2,. multidot., m), and m is the number of output layer units; in the invention, the number m of the output layer units is 1. WhereinAnd outputting the output value of the s unit of the output layer in the ith training.

The hidden layer exists for calculation, and has no specific meaning, and the number q of the hidden layer units can be calculated through a formula.

$q=\sqrt{p+m}+a---\left(4\right)$

Wherein a is [1,10 ]]Is constant. Let the hidden layer output vector, i.e. the output layer input vector, be U ═ U¹，u²，...，uⁿ)^T，t ═ 1,2,. q. WhereinAnd the output value of the t unit of the hidden layer in the ith training is shown.

Let the weight vector from input layer to hidden layer be v and the weight vector from hidden layer to output layer be w. The BP neural network topology is shown in fig. 2.

(3) And training the sample.

The BP neural network belongs to a feedforward network, and the training process comprises two stages of forward propagation and error backward propagation. First, the input layer input samples are processed by the hidden layer and passed to the output layer, which is the forward propagation stage. If the actual output value is not consistent with the calculated output value, the error is reversely propagated in a certain form, transmitted to the hidden layer and then transmitted to the input layer, the error is uniformly distributed to each layer of unit, error signals of each layer of unit are obtained, and weight values of each layer are modified, which is an error reverse propagation stage. The process of continuously adjusting the weight is the training process of the network.

From input layer to hidden layer:

and setting an initial weight v and an initial threshold theta to be random minimum values.

And inputting the membership function value of the evaluation index as an input sample of the BP neural network, and calculating the output values of the hidden layer and the output layer.

Let the t-th neuron activation value of the hidden layer in the ith training be

$s_t^i={\mathrm{\Σ}}_{k=1}^p{v}_{kt}{x}_k^t-{\mathrm{\θ}}_t,(t=1,2,\mathrm{...},q)---\left(5\right)$

In the formula (5), v_kτThe connection weight value of the kth neuron of the BP neural network input layer and the t-th neuron of the hidden layer is input,the corresponding membership function value of the ith measurement value of the kth evaluation index in the interval (-1,1), i.e. the input layer input value theta_τThreshold for the t-th neuron of the hidden layer.

The hidden layer neuron activation function adopts a sigmoid function:

f(x)＝1/[1+exp(-x)](6)

t-th neuron output value of hidden layer

From the hidden layer to the output layer:

output value of hidden layer neuron at this timeIs an input value to an output layer neuron, the output layer neuron activation value being

$l_s^t={\mathrm{\Σ}}_{t=1}^q{w}_{ts}{u}_t^i-r_s,(s=1,2,\mathrm{...}m)---\left(8\right)$

W in formula (7)_τsThe connection weight value r of the t-th neuron of the BP neural network hidden layer and the s-th neuron of the output layer is_sThreshold for the s-th neuron of the output layer.

The neuron activation function of the output layer is the same as that of the hidden layer, and a sigmoid function is adopted:

f(x)＝1/[1+exp(-x)](9)

the s-th neuron of the output layer outputs a value of

$y_s^t=f\left(l_3^t\right),(t=1,2,\mathrm{...}q)---\left(10\right)$

(4) And adjusting the weight value.

Let the known actual output value beThe difference between the actual output value and the calculated output value is

Adjusting the weight omega of the output layer according to a formula by using the steepest descent method_τs：

${\mathrm{\Δ\ω}}_{t3}=\mathrm{\η}(d_3^t-y_3^t)y_3^t(1-y_3^t)u_t^t$

ω_τs(i+1)＝ω_τs(i)+Δω_τs

Wherein η is a scale factor, i.e., learning rate, and is set to [0, 1]]A small number in between. Omega_τs(i +1) is the hidden layer to output layer weight in the (i +1) th training.

Using the steepest descent method to adjust the weight v of the output layer according to the formula_kτ：

$\begin{array}{c}{\mathrm{\Δv}}_{kt}=\mathrm{\η}\underset{s=1}{\overset{m}{\Σ}}(d_s^i-y_s^i)y_s^i(1-y_s^i){\mathrm{\ω}}_{ts}{u}_t^i(1-u_t^i)x_k^i\\ v_{kt}(i+1)=v_{kt}\left(i\right)+{\mathrm{\Δv}}_{kt}\end{array}$

Wherein v is_kτ(i +1) is the layer-to-hidden layer weight input in the (i +1) th training.

And (4) returning to the step (3), and performing next training until the error between the actual output value and the calculated output value of the BP neural network is smaller than the set minimum value, namely, the network convergence state is achieved.

(5) Computing reliability

The BP neural network model obtained at the moment has expert experience knowledge, the reliability of any safety detection sensor network can be evaluated according to the provided index measurement value, and the reliability degree of the sensor network can be known according to the output reliability score.

In order to more comprehensively, objectively, reliably and accurately reflect the reliable performance of the safety detection sensor network, a brand new method is provided for evaluating the performance of the safety detection sensor network of the urban rail train, and theoretical and practical support is provided for optimizing the performance of the safety detection sensor network of the urban rail train.

Claims (5)

1. A multi-index comprehensive evaluation method for reliability of an urban rail train safety detection sensor network is characterized by comprising the following steps:

(1) processing a sample, measuring four indexes of time delay, occupancy rate, packet loss rate and error rate, normalizing the measured value by using a membership function in a fuzzy mathematical theory, calculating a corresponding membership function value through a formula to enable the membership function value to have relative comparability, quantizing the membership function value to a comparable interval (-1,1), and taking the membership function value as a sample input value of a BP neural network input layer, wherein a sample output value is a corresponding reliability value of a sensor network;

(2) constructing a BP neural network, determining a network structure, and establishing a BP neural network model, wherein the BP neural network comprises an input layer, a plurality of hidden layers and an output layer; adopting a three-layer BP neural network which is simplest and most widely applied and is provided with an input layer, a hidden layer and an output layer as a constructed BP neural network structure;

(3) training a sample, inputting membership function values of all indexes, calculating an output value of a hidden layer of the BP neural network through a formula, wherein the output value of the hidden layer is an input value of an output layer, and calculating the output value of the output layer through the formula;

(4) adjusting the weight, adjusting the weight according to the difference between the actual output value and the calculated output value, returning to the step (3), recalculating the hidden layer output value and the output layer output value, adjusting the weight, calculating, adjusting the weight, recalculating until the error between the actual output value and the calculated output value of the BP neural network is less than a set minimum value, namely reaching a network convergence state, finishing the calculation and obtaining the final weight;

(5) and (4) calculating the reliability, wherein the BP neural network obtained at the moment has expert experience knowledge, and the reliability of the urban rail train safety detection sensor network can be calculated according to various index measurement values by using the model.

2. The method for comprehensively evaluating the reliability of the urban rail train safety detection sensor network with multiple indexes according to claim 1, wherein the BP neural network is constructed in the step (2), because the evaluation indexes have different influences on the reliability of the urban rail train safety detection sensor network, in order to enable the evaluation indexes to be compared with each other in a system, the evaluation indexes have relative comparability, the evaluation indexes are normalized by using a fuzzy mathematical theory, the processed evaluation indexes can be used as training samples of the BP neural network, the time delay, the occupation rate, the packet loss rate and the bit error rate are quantitative indexes, the difference is that the time delay, the packet loss rate and the bit error rate are reverse indexes, the smaller the probability is, and the larger the occupation rate is, the better the probability is.

3. The method for comprehensively evaluating the reliability of the sensor network for the safety detection of the urban rail train according to claim 2, characterized in that each evaluation index is normalized by applying a fuzzy mathematical theory to establish an evaluation index normalization model,

normalization formula of forward indicator:

$A_k^i=(a_k^i-\stackrel{\‾}{p_k})/\left|\stackrel{\‾}{p_k}\right|---\left(1\right)$

normalization formula of reverse index:

$A_k^i=(\stackrel{\‾}{p_k}-a_k^i)/\left|\stackrel{\‾}{p_k}\right|---\left(2\right)$

and (3) converting the index measurement value into a membership function value on an interval (-1, 1):

$x_k^i=(1-e^{-A_k^i})/(1+e^{-A_k^i})---\left(3\right)$

wherein,for the ith measurement value of the kth evaluation index,the average of all measured values for the kth evaluation index,in order to convert the variables in the middle of the run,and the corresponding membership function value of the ith measurement value of the kth evaluation index in the interval (-1,1), namely the input value of the BP neural network input layer.

4. The urban rail train safety detection sensor network reliability multi-index comprehensive evaluation method according to claim 1, characterized in that the step (2) is carried out to construct a BP neural network model

Let input layer vector X ═ X¹,x²,...,xⁿ)^T，k ═ 1,2,. cndot, p), where x isⁱFor the (i) th sample,a membership function value corresponding to the ith measurement value of the kth evaluation index in an interval (-1,1) is an input layer input value, and n is the number of samples; p is the number of input layer units;

the output vector is given by (Y)¹,y²,...,yⁿ)^T，s ═ 1,2,. multidot.m), whereOutputting the output value of the s unit of the output layer in the ith training; the output layer unit is a reliability score of a train safety detection sensor network; m is the number of output layer units;

the hidden layer exists for calculation, has no specific meaning, the number q of hidden layer units can be calculated by a formula,

$q=\sqrt{p+m}+a---\left(4\right)$

wherein a is [1,10 ]]Constant in between, let the hidden layer output vector, i.e. the output layer input vector, be U ═ U¹,u²,...,uⁿ)^T，t ═ 1,2,. q; whereinThe output value of the t unit of the hidden layer in the ith training is obtained; q is the number of hidden layer units;

let the weight vector from input layer to hidden layer be V and the weight vector from hidden layer to output layer be W.

5. The method for comprehensively evaluating the reliability of the sensor network for the safety detection of the urban rail train according to the claim 1, characterized in that the step (3) trains samples, establishes a sample training model,

from input layer to hidden layer:

setting an initial weight v and an initial threshold theta, which are set as random minimum values;

inputting membership function values of the evaluation indexes as input samples of the BP neural network, and calculating output values of the hidden layer and the output layer;

let the t-th neuron activation value of the hidden layer in the ith training be

$s_t^i=\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{v}_{kt}{x}_k^i-{\mathrm{\θ}}_t,(t=1,2,...,q)---\left(5\right)$

In the formula (5), v_ktThe connection weight value of the kth neuron of the BP neural network input layer and the t-th neuron of the hidden layer is input,the corresponding membership function value of the ith measurement value of the kth evaluation index in the interval (-1,1), i.e. the input layer input value theta_tThe threshold value of the t-th neuron of the hidden layer is shown, and p is the number of input layer units; q is the number of hidden layer units;

the hidden layer neuron activation function adopts a sigmoid function:

f(x)＝1/[1+exp(-x)](6)

t-th neuron output value of hidden layer

$u_t^i=f\left(s_t^i\right),(t=1,2,...,q)---\left(7\right)$

From the hidden layer to the output layer:

output value of hidden layer neuron at this timeIs an input value to an output layer neuron, the output layer neuron activation value being

$l_s^i=\underset{t=1}{\overset{q}{\mathrm{\Σ}}}{w}_{ts}{u}_t^i-r_s,(s=1,2,...,m)---\left(8\right)$

W in formula (7)_tsThe connection weight value r of the t-th neuron of the BP neural network hidden layer and the s-th neuron of the output layer is_sA threshold value for the s-th neuron of the output layer; m is the number of output layer units;

the neuron activation function of the output layer is the same as that of the hidden layer, and a sigmoid function is adopted:

f(x)＝1/[1+exp(-x)](9)

the s-th neuron of the output layer outputs a value of

$y_s^i=f\left(l_s^i\right),(s=1,2,...,m)---\left(10\right);$

The adjusting of the weight in the step (4) comprises: let the known actual output value beThe difference between the actual output value and the calculated output value is

Using the steepest descent method to adjust the weight w of the output layer according to the formula_ts：

${\mathrm{\Δw}}_{ts}=\mathrm{\η}(d_s^i-y_s^i)y_s^i(1-y_s^i)u_t^i$

w_ts(i+1)＝w_ts(i)+Δw_ts

Wherein η is a scale factor, i.e., learning rate, and is set to [0, 1]]A smaller number in between; w is a_ts(i +1) is the weight from the hidden layer to the output layer in the (i +1) th training;

adjusting the weight v of the input layer according to a formula by using the steepest descent method_kt：

${\mathrm{\Δv}}_{kt}=\mathrm{\η}\underset{s=1}{\overset{m}{\Σ}}(d_s^i-y_s^i)y_s^i(1-y_s^i)w_{ts}{u}_t^i(1-u_t^i)x_k^i$

v_kt(i+1)＝v_kt(i)+Δv_kt

Wherein v is_kt(i +1) inputting the weights from the layer to the hidden layer in the (i +1) th training;

and (4) returning to the step (3), and performing next training until the error between the actual output value and the calculated output value of the BP neural network is smaller than the set minimum value, namely, the network convergence state is achieved.