JP7226855B2 - Calculation method, system and storage medium for train passing through tunnel - Google Patents

Description

The present invention relates to the field of artificial intelligence learning, in particular to a method, system and storage medium for calculating the time a train passes through a tunnel.

Improving the ride comfort of passengers on passenger trains throughout their journey will help improve the competitiveness of high-speed rail in terms of passenger transport. Comprehensive information presentation helps passengers to grasp the current train operation status and improve the passenger's ride comfort. At present, the information presented in the train includes temperature, train speed, etc., but it is poor in presenting information related to special working conditions when the train passes through a tunnel or the like.

After the train enters the tunnel, there will be a situation where the communication information is lost, which makes people in the Internet age nervous and panicked, and seriously impairs the comfort of passengers. In addition, the dark closed environment in the tunnel, the discomfort of the ear membrane caused by poor vehicle sealing, and the unknown length of the tunnel make people feel very uneasy. Therefore, designing a method to effectively and accurately measure the time it takes to pass through a tunnel and timely notifying passengers of their waiting time in the tunnel plays a very important role in improving ride comfort. Fulfill. The core of calculating the time it takes a train to pass through a tunnel is the accurate positioning of the train inside the tunnel. There are several types of general train positioning methods as follows.

1. It is a Beidou GPS positioning system, which can use Beidou satellite positioning to provide omnidirectional, all-weather positioning information throughout the day, but is not applicable to train positioning in tunnels.

2. Radio base station type, train information is provided by radio base stations at both ends of the tunnel, effectively reducing the use of trackside equipment along the route, but not applicable to train positioning in the tunnel.

3. It is a response type, and multiple trackside devices are arranged along the railway line, and in-vehicle devices corresponding to the train are installed. Although it has high accuracy, construction and maintenance costs are enormous, and practicality is low. .

The problem to be solved by the present invention is to achieve intelligent sensing of the distance from the train in the tunnel to the tunnel exit, and accurately estimate the running time of the train in the tunnel, for the shortcomings of the prior art. To provide a method, system and storage medium for calculating the time for a train to pass through a tunnel.

In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows. A method for calculating the time for a train to pass through a tunnel, comprising:
Estimate the remaining time for the train to pass through the current tunnel with the formula t=O(i)/(f _Hz (O(i)-O(i-1)));
where the mileage value from the train's current position to the tunnel exit is O(i)= ^OM (i)+ε(i), and O(i−1) is the distance from the train's current position to the tunnel exit. is the distance traveled to the exit, and O ^M (i) is the typical sample integration of the current position obtained by calculating the correlation using the current position feature image and the template image in the typical sample integration template library. The current position feature image is the best matching position in the template library, and the current position feature image is obtained using the collected meteorological parameters of the head and tail sample points of the current position of the train and the meteorological parameter sequence of the current position for a certain period of time before. Obtained by performing phase space reconstruction and RGB color space combination, the template images in the representative sample integrated template library are obtained by performing phase space reconstruction on representative samples of each class tunnel group are color images evolved by meteorological parameters of the head and tail of a train, and a representative sample of each class of tunnel group, depending on the geographic location where the tunnel is located and the weather conditions inside and outside the tunnel. Meteorological parameters are classified, and samples of the same class within the classified area are further divided into samples obtained. The tunnel travel distance prediction error compensation model adopts the data of the tunnel group of the same class, and the temperature sequence, humidity sequence, and atmospheric pressure sequence are input, and the RGB color integration matching model It refers to a model obtained by training a least-squares support vector machine with prediction error as output, where f _Hz is the location update frequency.

The present invention uses the technology of artificial intelligence big data analysis to realize intelligent sensing of the distance from the train to the tunnel exit in the tunnel according to the temperature sequence, humidity sequence and atmospheric pressure sequence data at both ends of the train. It does not require any trackside equipment, is low cost and highly practical.

The specific process of determining the travel distance value O(i) from the current position of the train to the tunnel exit is as follows.
Step 1) of collecting tunnel weather parameters and building a tunnel weather parameter database;
Based on the tunnel weather parameter database, classify the weather parameters in the tunnel weather parameter database according to the geographic location of the tunnel and the weather conditions inside and outside the tunnel, obtain samples of the same class in the area, and group the tunnels with similar attributes. Step 2) for realizing the classification of
step 3) of subdividing the samples of the same class within the zone to obtain a representative sample of tunnel groups of each class;
Phase-space reconstruction was performed on a representative sample of each class of tunnel group to obtain an evolutionary color image of the barometric weather parameters of the head and tail of the train, i.e. the template image, and all representative samples step 4) building a typical sample integrated template library with corresponding template images;
The head and tail sample points of the current position of the train and the sample sequence of the current position over a period of time before are respectively collected, phase space reconstruction and RGB color space combination are performed, and the current position feature image { _ch ,c _f } and perform correlation calculations on the current position feature image and the template image in the typical sample integrated template library to determine the best matching position in the typical sample integrated template library for the current position. 5) and
Adopt the data of tunnel group in the same class, take the temperature sequence, humidity sequence, and pressure sequence as input, and the prediction error of the best matching position as output, train the least squares support vector machine, and compensate for the tunnel travel distance prediction error. step 6) of establishing a model;
Step 7) of collecting real-time weather parameters during train operation and using the tunnel mileage prediction error compensation model to obtain the final mileage value to the tunnel exit, namely O(i). .

The method of the present invention can make full use of changes in weather parameters in tunnels to realize intelligent sensing of train position, does not require rackside equipment, and has the advantages of simplicity and practicality.

The specific realization process of step 2) is
1) dividing the tunnel groups into N classes according to the geographical locations where the tunnel samples are located;
2) Obtain the average temperature distribution in Tmin before the train passes through the tunnel in the current area within one year, adopt the Gaussian distribution function to fit the average temperature distribution in the tunnel in the current area, and perform fitting The subsequent average temperature distribution is divided into K equal parts according to the probability, the temperature samples belonging to one part are defined as temperature samples of the same class in the current area, and the temperature sequence and humidity sequence corresponding to the temperature samples of the same class and defining the pressure sequences as samples of the same class.

The present invention employs geographic location to coarsely classify tunnels, and then employs meteorological parameters to finely classify tunnels in the same area to ensure the reliability of template samples.

（式中、ａｖｇ（Ｃ_ｉ）がサンプルクラスタＣ_ｉにおけるサンプルの平均距離であり、ａｖｇ（Ｃ_ｊ）がサンプルクラスタＣ_ｊにおけるサンプルの平均距離であり、ｄ_ｃｅｎ（Ｃ_ｉ,Ｃ_ｊ）がサンプルクラスタＣ_ｉとサンプルクラスタＣ_ｊの中心点の間の距離である。）を計算することと、
５）現在の最適化目的関数値ｆｉｔｎｅｓｓに基づいて、灰色オオカミ最適化アルゴリズムを採用して、α、β及びｎ_０の値を更新することと、
６）所定の最適化回数ｍ＝１００に達するまで、ステップ４）とステップ５を繰り返すことであって、この時のα、β及びｎ_０の値極は最適化された最終的なパラメータ値であり、最適化された最終的なパラメータ値でのクラスタリング結果は最終クラスタリング結果であることと、
７）最終クラスタリング結果に従って現在の区域での同一クラスのサンプルをＥ個のクラスタリングサンプルクラスタに分け、各クラスタリングサンプルクラスタのクラスタリング中心及びクラスタリング中心から最も近いＰ個のサンプルに対応する

それぞれ列車頭部で収集された温度シーケンス、列車頭部で収集された温度シーケンス、列車頭部で収集された気圧シーケンス、列車尾部で収集された温度シーケンス、列車尾部で収集された温度シーケンス、列車尾部で収集された気圧シーケンスに対応する。）を取得し、Ｐ＋１個のサンプルに対応する元の時間シーケンスを現在クラストンネルグループの典型的なサンプルと定義することと、を含む。 The specific realization process of step 3) is
1) Modeling an autoregressive integrated moving average model for each time sequence in the same class of samples in the current area, and calculating the features of each time sequence, namely the autoregressive term, the difference term and the extracting the parameters of the moving regression term, wherein all features of the samples of the same class within the zone constitute a feature matrix A, the time sequence comprising a temperature sequence, a humidity sequence and an air pressure sequence;
2) Performing dimension reduction processing on the feature matrix A, selecting M principal components with the largest contribution to characterize the information of the original feature matrix A, and obtaining the transformed matrix A′. ,
3) Kernel function k = αk _rbf + βk _Linear + (1-α-β)k _Laplace (where k _rbf is the radial basis kernel function, k _Linear is the linear kernel function, and k _Laplace is the Laplacian kernel function ), and randomly set the initial values of the kernel function coefficients α, β (α, β∈[0, 1]) and the number of classes n ₀ (n ₀ is a positive integer smaller than 30) to decide;
4) Map the feature values in the matrix A′ to the feature space corresponding to the kernel function k, obtain a new feature matrix A″, use the feature amount in A″ as input, and according to the initial class number n ₀ , k-means A clustering algorithm is used to achieve feature clustering in A", dividing the same class samples in the area into n classes, forming a sample cluster from each class sample, and optimizing the objective function

(where avg(C _i ) is the average distance of samples in sample cluster C _i , avg(C _j ) is the average distance of samples in sample cluster C _j , and d _cen (C _i ,C _j ) is ), which is the distance between the center points of sample clusters C _i and C _j ;
5) based on the current optimization objective function value fitness, adopting the gray wolf optimization algorithm to update the values of α, β and _n0 ;
6) Repeating step 4) and step 5 until a predetermined number of optimization times m=100 is reached, at which time the extremes of α, β and _n0 are the final optimized parameter values. , and the clustering result at the final optimized parameter value is the final clustering result, and
7) Divide the samples of the same class in the current area into E clustering sample clusters according to the final clustering result, corresponding to the clustering center of each clustering sample cluster and the P samples closest to the clustering center;

temperature sequence collected at head of train, temperature sequence collected at head of train, air pressure sequence collected at head of train, temperature sequence collected at tail of train, temperature sequence collected at tail of train, train Corresponds to the barometric sequence collected at the tail. ), and defining the original time sequence corresponding to P+1 samples as typical samples of the current class tunnel group.

The present invention determines typical samples by self-adaptive method and optimizes feature mapping space and feature clustering parameters, which helps to obtain representative typical samples and ensures the effectiveness of templates. .

であることと、を含む。 The specific realization process of step 5) is
1) Determining the current sample point and G previous sample points in the head and tail temperature sequence, humidity sequence and barometric pressure sequence of the train at the current position;
2) Adopting the delayed coordinate method to perform phase space reconstruction, obtain six two-dimensional reconstruction matrices representing the evolutionary characteristics of temperature, humidity and air pressure at the head and tail of the train, and convert the six matrices into combining the head and tail according to the RGB color space to form a current position feature image {c _h ,c _f };
3) performing a convolution operation on the current position feature image and the images in the typical sample integration template library to obtain multiple one-dimensional sequences;
4) sorting the elements in all one-dimensional sequences in descending order and determining the largest L elements as candidate elements, where the pre-sorting position corresponding to the candidate element is the candidate position, and the traveled distance values to the corresponding tunnel exits are s _j (j=1, 2, . . . L);
5) taking the mean of the distance-to-tunnel-exit values corresponding to the candidate positions, where the mean is the best matching position in the representative sample integrated template library for the current position;

including being

The present invention transforms a one-dimensional time sequence into a two-dimensional image by spatial reconstruction, which helps in obtaining detailed information in the time sequence and in accurate template matching of multiple time sequences.

（ｋ＝１,２…,Ｎ_１であり、Ｎ_１は検証セットにおけるサンプルペアの数である。）として取得することと、
５）ステップ３）とステップ４）を繰り返し、最適化目的関数

を最小化する入力特徴と最小二乗サポートベクターマシンを決定し、即ち、最適な入力特徴と走行距離予測誤差補償モデルである最適な最小二乗サポートベクターマシンを得ることと、を含む。 The specific realization process of step 6) is
1) Input sample I = (T _h , H _h , P _h , T _f , H _f , P _f ) (where T _h = (t ₁ , t ₂ ..., t _G ) is the head of the train in the tunnel and G previous sample points, where H _h is a head of train humidity sequence of length G and _Ph is a head of train pressure sequence of length G where T _h is a G-length train tail temperature sequence, H _h is a G-length train tail humidity sequence, and _Ph is a G-length train tail pressure sequence. ), defining the output sample as the mileage error value ε corresponding to the current position, and constructing the modeling sample with the combination Y={I, ε} of the input and output samples;
2) dividing the modeling samples into a training set and a validation set;
3) perform binary encoding on each dimensional feature in the input sample I, and if the encoded value corresponding to a dimensional feature is 1, select that feature as an input variable for a least-squares support vector machine; discarding a dimension feature if the corresponding encoded value for that dimension feature is 0;
4) based on the current feature encoding values, update the input samples to obtain an updated training set and a validation set, adopt the updated training set data to train a least squares support vector machine, and update the validation set data; into the trained least-squares support vector machine, and the output sequence of the least-squares support vector machine is

(where k=1,2..., _N1 , where _N1 is the number of sample pairs in the validation set), and
5) Repeat steps 3) and 4) to optimize the objective function

, i.e., obtaining the optimal input features and the optimal least squares support vector machine that is the mileage prediction error compensation model.

The present invention adopts a data-driven method to establish an error correction model, which helps to reduce the error caused by the deviation between the template and the actual value, and further improve the mileage prediction accuracy.

The specific realization process of step 7) is
1) Collect the average temperature distribution data within Tmin before the train passes through the tunnel, obtain the temperature sequence and humidity sequence of the head and tail of the train at the current position, and obtain the template matching output value O ^M (i) and obtaining a compensation error output result ε(i) using the tunnel travel distance prediction error compensation model;
2) Obtaining the mileage value O(i) from the train's current position to the tunnel exit with the formula O(i)=O ^M (i)+ε(i).

The present invention adopts a policy of combining rough prediction by template matching and error correction compensation to realize the calculation of the mileage in the tunnel, which helps to improve the accuracy of the mileage calculation.

The present invention further provides a system for calculating the time a train travels through a tunnel, the system comprising:
An optimum matching position acquisition module for performing correlation calculations on the current position feature image and templates in the typical sample integrated template library to obtain the optimum matching position of the current position in the typical sample integrated template library. Then, the current position feature image is phase space reconstruction and RGB color space using the collected meteorological parameters of the head and tail sample points of the current position of the train and the meteorological parameter sequence of the current position for a certain period of time before. The template images in the typical sample integrated template library are obtained by performing a combination of train head and tail images obtained by performing phase space reconstruction on a representative sample of each class of tunnel group. The typical sample of each class of tunnel group is classified and classified according to the geographical location where the tunnel is located and the weather conditions inside and outside the tunnel. an optimal matching position acquisition module, which is a sample obtained by further dividing samples of the same class within the area;
Adopting the data of the tunnel group of the same class, taking the temperature sequence, humidity sequence and pressure sequence as the input, and the prediction error of the best matching position as the output, obtained by training the least squares support vector machine, the output is the compensation error a tunnel mileage prediction error compensation model that is ε(i);
Formula t=O(i)/(f _Hz (O(i)−O(i−1))) (where O(i)= ^OM (i)+ε(i) and O(i− 1) is the traveled distance value from the position before the train's current position to the tunnel exit, O ^M (i) is the best matching position, and f _Hz is the train position update frequency. a remaining time calculation module for estimating the remaining time t to traverse the tunnel.

である出力ユニットと、を含む。 The optimal matching position acquisition module includes:
a data acquisition unit for determining a current sample point and G previous sample points in the head and tail temperature sequence, humidity sequence and barometric pressure sequence of the current position train;
Adopting the delayed coordinate method to perform phase space reconstruction, obtain six two-dimensional reconstruction matrices representing the evolutionary characteristics of temperature, humidity and air pressure at the head and tail of the train, and divide the six matrices into subdivisions. a feature image calculation unit for forming a current position feature image {c _h , c _f } by combining according to the RGB color space according to the head and tail;
a convolution unit for performing a convolution operation on the current position feature image and the template image in the typical sample integrated template library to obtain a plurality of one-dimensional sequences;
A sorting unit for sorting the elements in all one-dimensional sequences in descending order and determining the largest L elements as candidate elements, wherein the positions before sorting corresponding to the candidate elements are candidate positions, and a sorting unit whose traveled distance values to tunnel exits corresponding to are s _j (j=1, 2, . . . L);
An output unit for taking the mean value of the distance traveled to tunnel exit values corresponding to the candidate positions, wherein the mean value is the best matching position ^OM in the typical sample integrated template library of the current position is

and an output unit that is

（ｋ＝１,２…,Ｎ_１であり、Ｎ_１は検証セットにおけるサンプルペアの数である。）として取得することと、
５）ステップ３）とステップ４）を繰り返し、最適化目的関数

を最小化する入力特徴と最小二乗サポートベクターマシンを決定し、即ち、最適な入力特徴と走行距離予測誤差補償モデルである最適な最小二乗サポートベクターマシンを得ることと、を含む。 The training process of the tunnel mileage prediction error compensation model of the present invention includes:
1) Input sample I = (T _h , H _h , P _h , T _f , H _f , P _f ) (where T _h = (t ₁ , t ₂ ..., t _G ) is the head of the train in the tunnel and G previous sample points, where H _h is a head of train humidity sequence of length G and _Ph is a head of train pressure sequence of length G where _Tf is a G-length train tail temperature sequence, _Hf is a G-length train tail humidity sequence, and _Pf is a G-length train tail pressure sequence. ), defining the output sample as the mileage error value ε corresponding to the current position, and constructing the modeling sample with the combination Y={I, ε} of the input and output samples;
2) dividing the modeling samples into a training set, a validation set, and a test set;
3) perform binary encoding on each dimensional feature in the input sample I, and if the encoded value corresponding to a dimensional feature is 1, select that feature as an input variable for a least-squares support vector machine; discarding a dimension feature if the corresponding encoded value for that dimension feature is 0;
4) based on the current feature encoding values, updated the input samples to obtain an updated training set, validation set and test set, and adopted the updated training set data to train and update a least squares support vector machine; Input the validation set data into a trained least-squares support vector machine, and take the output sequence of the least-squares support vector machine as

(where k=1,2..., _N1 , where _N1 is the number of sample pairs in the validation set), and
5) Repeat steps 3) and 4) to optimize the objective function

, i.e., obtaining the optimal input features and the optimal least squares support vector machine that is the mileage prediction error compensation model.

As an inventive concept, the present invention further provides a computer readable storage medium having stored thereon a program arranged to perform the steps of the method of the present invention.

The beneficial effects of the present invention compared to the prior art are as follows.

1. The present invention takes advantage of the technology of artificial intelligence big data analysis to fully unearth the potential law of changes with tunnel depth of barometric weather parameters in the tunnel. Temperature, humidity and barometric pressure sequence data acquired at both ends of the train enables intelligent sensing of the distance from the train in the tunnel to the exit of the tunnel, and based on this, the effective duration of the train's dwell time in the tunnel. realization of accurate estimation.

2. The present invention can provide passengers with accurate information presentation (such as temperature, humidity, stay time information in a tunnel, etc.), which helps improve the passenger's ride comfort.

3. After the modeling is completed, the present invention does not require any trackside equipment, and can realize the collection of input data only with on-vehicle temperature, humidity and air pressure sensors, and is highly worthy of popularization.

4 is a main flow chart of the present invention; Fig. 2 is a schematic diagram of phase space reconstruction of time sequence and matching of RGB integration template of the present invention;

As shown in FIG. 1, the specific steps of Embodiment 1 of the present invention are as follows.

Step 1: Collect tunnel pressure meteorological parameters and build a tunnel pressure meteorological parameter database.

On-board sensors distributed at both ends of the train collect the temperature, humidity and barometric pressure sequences in the tunnel as the train passes in real time. A sequence of mileage and time is collected, with a sampling frequency of 10 Hz. The temperature and humidity averages collected at Tmin before the train enters the tunnel are taken as estimates of the local temperature and humidity averages. The temperature sequence, humidity sequence, and barometric pressure sequence collected by sensors located at both ends of the train when the train passes through a tunnel once, and the average temperature and humidity estimates for the location obtained before entering the tunnel are , constitute a set of tunnel weather parameter samples. Tunnel weather parameter samples collected by all trains operating within a jurisdiction in one year constitute the tunnel weather parameter database. The range of T values is [10,20], where T=10 in the present invention.

Step 2, classify the tunnel group barometric weather parameters.

Based on the tunnel weather parameter database, the weather parameters in the database are classified according to the geographic location of the tunnel and the weather conditions inside and outside the tunnel to achieve classification of tunnel groups with similar attributes. The specific realization process is as follows.

Step A1, according to the division of China's weather features by China's building blocks, classify the tunnel groups into 7 types according to the geographic location of the tunnels in the tunnel weather parameter database.

Step A2, obtain the average temperature distribution within Tmin before the train passes through the tunnel in the current area in one year, and adopt the Gaussian distribution function to fit the average temperature distribution. The average temperature distribution is divided into 10 equal parts according to the probability, the temperature samples belonging to one part are defined as temperature samples of the same class in the current area, and the temperature sequence, humidity sequence and pressure corresponding to the temperature samples of the same class are calculated. Define a sequence as samples of the same class.

Step 3, characterize representative samples of the same class of samples within the area.

Each same-class sample within the area obtained by coarsely classifying the input attribute set based on the class of the tunnel group is further divided. Specifically, it includes the following substeps.

Step B1, perform evolutionary feature extraction on the temperature, humidity and pressure sequences when the train passes through a tunnel in the samples of the same class. The specific steps are to model an autoregressive integrated moving average model (ARIMA) for each time sequence and extract parameters for the autoregressive, differential and moving regression terms of each sequence. Specifically, the time sequence collected each time a train passes through a tunnel can collect six time sequences, and therefore the feature quantity extracted each time a train passes through a tunnel is a=(p ₁ , d ₁ , q ₁ , p ₂ , d ₂ , q ₂ , ..., p ₆ , d ₆ , q ₆ ), where p ₁ is the number of autoregressive terms in time sequence 1, _d1 is the number of differences in time sequence 1, and _q1 is the number of moving average terms in time sequence 1. All the features of the samples of the same class within the area constitute the feature matrix A.

In step B2, a principal component analysis algorithm (PCA) is employed to perform dimensionality reduction processing on the feature matrix A, which is the feature amount of all the samples of the same class within the area. The five principal components with the highest contribution are selected to characterize the information of the original feature matrix A to obtain the transformed matrix A'.

Step B3, define a kernel function.
k = αk _rbf + βk _linear + (1-α-β)k _lap (1)
where _{k_rbf} is the radial basis kernel function, _{k_linear} is the linear kernel function, and _{k_laplace} is the Laplacian kernel function. The feature values of the A' matrix are mapped to the feature space corresponding to the kernel function k.

Step B4, determine the optimization target and adopt the Gray Wolf Optimization Algorithm (GWO) to optimize the kernel function coefficients α, β and the number of classes. where α, βε[0, 1] and the number of classes is a positive integer less than 20.

式中、ａｖｇ（Ｃ_ｉ）がクラスタＣ_ｉにおけるサンプルの平均距離であり、ｄ_ｃｅｎ（Ｃ_ｉ,Ｃ_ｊ）がクラスタＣ_ｉとクラスタＣ_ｊの中心点の間の距離である。 Step B5, determine an optimization objective function.

where avg(C _i ) is the average distance of the samples in cluster C _i and d _cen (C _i ,C _j ) is the distance between the center points of clusters C _i and C _j .

列車頭部で収集された温度シーケンス、列車頭部で収集された温度シーケンス、列車頭部で収集された気圧シーケンス、列車尾部で収集された温度シーケンス、列車尾部で収集された温度シーケンス、列車尾部で収集された気圧シーケンスに対応する。

を現在のクラスの典型的なサンプルと定義する。 Step B6, adopt the k-means clustering algorithm optimized by the gray wolf optimization algorithm according to the above predetermined parameters to realize the clustering of the features after dimensionality reduction. corresponding to the clustering center of each clustering sample cluster and the five nearest samples from the clustering center

Temperature sequence collected at head of train Temperature sequence collected at head of train Pressure sequence collected at head of train Temperature sequence collected at tail of train Temperature sequence collected at tail of train Train tail corresponds to the barometric sequence collected at .

be defined as a typical sample of the current class.

Step 4, build a typical sample integrated template library.

Perform phase space reconstruction on a typical sample, set delay time to 1, window length to 5, and adopt the delayed coordinate method to perform spatial reconstruction on the temperature, humidity and pressure sequences. , three two-dimensional reconstruction matrices representing the evolution characteristics of temperature, humidity and pressure at the head of the train and the evolution characteristics of temperature, humidity and pressure at the tail of the train, and converting the three matrices into the RGB color space to form an evolutionary color image of the barometric weather parameters of the train head and tail, which is the template image {h _i , f _i } (i=0,1,2...5) .

Step 5, train a template RGB color integration matching model (shown in Figure 2).

Collect the meteorological data of the head and tail sample points of the current position of the train and the meteorological sample sequence of a certain period of time respectively, perform phase space reconstruction and RGB color space combination, and obtain the current position feature image { _ch ,c _f }. A correlation calculation is performed on the images in the current location feature module and the template library to determine the best matching location in the template library for the current location. Specifically, it includes the following steps.

Step C1, obtain the relevant meteorological data of the current sample point and 19 previous sample points in the head and tail temperature, humidity and pressure sequence of the current position of the train.

Step C2, set the delay time to 1 and the window length to 5, adopt the delay coordinate method to perform spatial reconstruction, and create 6 two-dimensional representations of the evolution characteristics of temperature, humidity and air pressure at the head of the train. The reconstruction matrices are obtained, and the 6 matrices are combined according to the RGB color space for each head and tail to form the current position feature image { _ch , _cf }.

を行い、式中、各ｇ_ｉは１個の一次元シーケンスである。 Step C3, a convolution operation on the current position feature image and the image in the template library

where each g _i is a one-dimensional sequence.

Step C4, sort the elements in all g _i sequences in descending order, determine the maximum 5 elements as candidate elements, the position before sorting corresponding to the candidate element is the candidate position, and the tunnel corresponding to the candidate position The traveled distance values to the exit are s _j (j=1, 2, . . . 5).

である。 Step C5, take the average value of the distance traveled to the tunnel exit corresponding to the candidate location, and the average value is the current template matching output value, that is, the formula for the template matching output value is:

is.

Step 6, establish an integrated template matching error compensation model.

Adopting the data of the tunnel group of the same class, taking the temperature sequence, humidity sequence, and pressure sequence as input, and the prediction error of the template matching model as output, trains the least squares support vector machine (LSSVM) to predict the tunnel travel distance. Establish an error compensation model. Specifically, it includes the following steps.

Step D1, define training samples, define input samples I=(T _h ,H _h ,P _h ,T _f ,H _f ,P _f ), where T _h =(t ₁ ,t ₂ . . . , t ₁₉ , t ₂₀ ) is the temperature sequence of the current and 19 previous sample points of the train head in the tunnel. Similarly, H _h is a train head humidity sequence of length 20, Ph is a train head pressure sequence of length 20, _{T h is} a train tail temperature sequence of length 20, H _h is Length 20 train tail humidity sequence, _Ph is length 20 train tail pressure sequence. The output sample is the odometer error value ε corresponding to the current position. We construct a modeling sample with an input-output combination Y={I, ε}. For each co-class tunnel group, 3000 samples are selected to establish a mileage prediction error compensation model.

Step D2, divide the modeling samples into a training set and a validation set. 70% of the number of 3000 sample pairs among the modeling samples are selected as the training set and 30% as the validation set by adopting the non-reconstructive random sampling scheme.

Step D3, determine the optimization target and initialize the optimization value. Adopting the binary alligator algorithm to optimize the input features of the model, i.e. performing binary encoding on each dimension feature in the input sample I, if the encoding value corresponding to a dimension feature is 1, The feature is selected as the input variable of the LSSVM model, and if the encoded value corresponding to a dimension feature is 0, the dimension feature is discarded. The 60 dimensional features are randomly initialized and coded to 0 or 1.

（ｋ＝１,２…９００）として取得する。最適化目的関数を定義する。 Step D4, determine an optimization objective function. Based on the current feature encoding values, input features are determined and the training set data is employed to train the LSSVM model. Input the validation set data into the trained LSSVM model and write the model output sequence as

(k=1, 2...900). Define an optimization objective function.

Step D5, output the optimized prediction model. A binary alligator algorithm is adopted to perform iterative optimization operations to determine the optimal input features and the LSSVM model, which is the LSSVM mileage prediction error compensation model.

Step 7, get the test data and call the mileage prediction model

During train operation, real-time barometric weather parameters are collected, current state feature images are constructed, three-dimensional image templates are searched and matched, and error correction input variables and error compensation models are determined. The template matching output and the error compensation output are fused to obtain the final mileage value to the tunnel exit. Specifically, it includes the following steps.

Step E1, collect the average temperature distribution within Tmin before the train passes through the tunnel. Determine the sample class to which the current state (ie the weather parameters of the train's current position) belongs. Temperature and humidity sensors mounted on the head and tail of the train are used to obtain the current temperature and humidity sequences.

Step E2, determine the sample class to which the current state belongs according to the current average temperature, and extract the corresponding template library.

Step E3, refer to the process in step 3 to determine the best matching position in the template library, and output the template matching output value O ^M (i) of the current sample point.

Step E4, refer to the process of process C1 in step 6, get the current state model input vector I and substitute it into the trained LSSVM model to get the compensation error output result ε(i).

Step E5, fuse the 3D template matching output value and the LSSVM model output value to obtain the final output value of the distance from the current position to the tunnel exit according to the following formula.
O(i)= ^OM (i)+ε(i) (5)

Step 8, Calculate the estimated remaining time to exit the tunnel.

The remaining time until the train passes through the current tunnel can be estimated by the following formula.
t=O(i)/(f _Hz (O(i)-O(i-1)))) (6)
where f _Hz is the location update frequency and in the present invention f _Hz =50 Hz.

Corresponding to the method of embodiment 1, embodiment 2 of the present invention provides a system for calculating the time for a train to pass through a tunnel, which system comprises:
An optimum matching position acquisition module for performing correlation calculations on the current position feature image and templates in the typical sample integrated template library to obtain the optimum matching position of the current position in the typical sample integrated template library. Then, the current position feature image is phase space reconstruction and RGB color space using the collected meteorological parameters of the head and tail sample points of the current position of the train and the meteorological parameter sequence of the current position for a certain period of time before. The template images in the typical sample integrated template library are obtained by performing a combination of train head and tail images obtained by performing phase space reconstruction on a representative sample of each class of tunnel group. The typical sample of each class of tunnel group is classified and classified according to the geographical location where the tunnel is located and the weather conditions inside and outside the tunnel. an optimal matching position acquisition module, which is a sample obtained by further dividing samples of the same class within the area;
Adopting the data of the tunnel group of the same class, taking the temperature sequence, humidity sequence and pressure sequence as the input, and the prediction error of the best matching position as the output, obtained by training the least squares support vector machine, the output is the compensation error a tunnel mileage prediction error compensation model that is ε(i);
Formula t=O(i)/(f _Hz (O(i)−O(i−1))) (where O(i)= ^OM (i)+ε(i) and O(i− 1) is the traveled distance value from the position before the train's current position to the tunnel exit, O ^M (i) is the best matching position, and f _Hz is the train position update frequency. a remaining time calculation module for estimating the remaining time t to traverse the tunnel.

である出力ユニットと、を含む。 The best matching position acquisition module is
a data acquisition unit for determining a current sample point and G previous sample points in the head and tail temperature sequence, humidity sequence and barometric pressure sequence of the current position train;
Adopting the delayed coordinate method to perform phase space reconstruction, obtain six two-dimensional reconstruction matrices representing the evolutionary characteristics of temperature, humidity and air pressure at the head and tail of the train, and divide the six matrices into subdivisions. a feature image calculation unit for forming a current position feature image {c _h , c _f } by combining according to the RGB color space according to the head and tail;
a convolution unit for performing a convolution operation on the current position feature image and the template image in the typical sample integrated template library to obtain a plurality of one-dimensional sequences;
A sorting unit for sorting the elements in all one-dimensional sequences in descending order and determining the largest L elements as candidate elements, wherein the positions before sorting corresponding to the candidate elements are candidate positions, and a sorting unit whose traveled distance values to tunnel exits corresponding to are s _j (j=1, 2, . . . L);
An output unit for taking the mean value of the distance traveled to tunnel exit values corresponding to the candidate positions, wherein the mean value is the best matching position ^OM in the typical sample integrated template library of the current position is

and an output unit that is

The data collection units in this example are various sensors attached to the head and tail of the train, such as temperature sensors, humidity sensors, barometric pressure sensors.

（ｋ＝１,２…,Ｎ_１であり、Ｎ_１は検証セットにおけるサンプルペアの数である。）として取得することと、
５）ステップ３）とステップ４）を繰り返し、最適化目的関数

を最小化する入力特徴と最小二乗サポートベクターマシンを決定し、即ち、最適な入力特徴と走行距離予測誤差補償モデルである最適な最小二乗サポートベクターマシンを得ることと、を含む。 The training process of the tunnel mileage prediction error compensation model includes:
1) Input sample I = (T _h , H _h , P _h , T _f , H _f , P _f ) (where T _h = (t ₁ , t ₂ ..., t _G ) is the head of the train in the tunnel and G previous sample points, where H _h is a head of train humidity sequence of length G and _Ph is a head of train pressure sequence of length G where _Tf is a G-length train tail temperature sequence, _Hf is a G-length train tail humidity sequence, and _Pf is a G-length train tail pressure sequence. ), defining the output sample as the mileage error value ε corresponding to the current position, and constructing the modeling sample with the combination Y={I, ε} of the input and output samples;
2) dividing the modeling samples into a training set and a validation set;
3) perform binary encoding on each dimensional feature in the input sample I, and if the encoded value corresponding to a dimensional feature is 1, select that feature as an input variable for a least-squares support vector machine; discarding a dimension feature if the corresponding encoded value for that dimension feature is 0;
4) based on the current feature encoding values, update the input samples to obtain an updated training set and a validation set, adopt the updated training set data to train a least squares support vector machine, and update the validation set data; into the trained least-squares support vector machine, and the output sequence of the least-squares support vector machine is

(where k=1,2..., _N1 , where _N1 is the number of sample pairs in the validation set), and
5) Repeat steps 3) and 4) to optimize the objective function

, i.e., obtaining the optimal input features and the optimal least squares support vector machine that is the mileage prediction error compensation model.

Embodiment 3 of the present invention provides a computer readable storage medium storing a program configured to perform the steps of the method of Embodiment 1 of the present invention.

Claims (10)

A method for calculating the time for a train to pass through a tunnel, comprising:
estimating the remaining time for the train to pass through the current tunnel with the formula t=O(i)/(f _Hz (O(i)-O(i-1)));
where the mileage value from the train's current position to the tunnel exit is O(i)=O ^M (i)+ε(i), and O(i−1) is the previous position of the train's current position. to the tunnel exit, and O ^M (i) is the typical current position obtained by calculating the correlation using the current position feature image and the template image in the typical sample integrated template library. The current position feature image is the collected meteorological parameters of the head and tail sample points of the current position of the train and the meteorological parameters of the current position of a previous period of time. Obtained by performing phase space reconstruction and RGB color space combination using sequences, the template images in the typical sample integration template library are phase space 4 is a color image of the meteorological parameters of the head and tail of the train obtained by performing a reconstruction, wherein a representative sample of the tunnel group of each class includes the geographic locations where the tunnels are located and It is a sample obtained by classifying the weather parameters according to the weather conditions inside and outside the tunnel, and further dividing the weather parameters of the same class in the classified area, and ε(i) was collected during the operation of the train. It is a compensation error obtained by inputting the weather parameters into a tunnel travel distance prediction error compensation model . sequence, pressure sequence as input, RGB color integration matching model prediction error as output, refers to a model obtained by training a least squares support vector machine, f _Hz is the location update frequency, the weather parameters are: said temperature sequence corresponding to the head of said train in operation; said humidity sequence corresponding to said head of said train; said pressure sequence corresponding to said head of said train; said temperature sequence corresponding to said tail of said train. , said humidity sequence corresponding to said train tail, and said barometric pressure sequence corresponding to said train tail. A specific process of determining the travel distance value O(i) from the current position of the train to the tunnel exit is as follows:
step 1) of collecting the weather parameters and building a weather parameter database;
Classify the weather parameters in the weather parameter database according to the geographic location of the tunnel and the weather conditions inside and outside the tunnel according to the weather parameter database to obtain samples of the same class within the area and have similar attributes. a step 2) of realizing the tunnel group classification;
step 3) of subdividing the samples of the same class within the zone to obtain a representative sample of said tunnel groups of each class;
Phase space reconstruction is performed on representative samples of the tunnel group for each class to obtain color images of the meteorological parameters of the head and tail of the train, i.e. template images, and all representative samples step 4) building said exemplary sample integration template library with template images corresponding to
A current sample acquisition time of the head and tail of the current position of the train, and a sample sequence at a plurality of previous sample acquisition times, respectively, are collected, phase space reconstruction and RGB color space combination are performed, and the current position of the train is constructing a position feature image {c _h , c _f }, performing correlation calculations on the current position feature image and template images in the typical sample integration template library, and performing the typical sample integration of the current position; step 5) of determining the best matching position in the template library;
Adopting the data of the tunnel group of the same class, taking the temperature sequence, the humidity sequence and the air pressure sequence as input, and the prediction error of the best matching position as output , training the least squares support vector machine, and tunneling Step 6) of establishing a mileage prediction error compensation model;
Step 7) of collecting the weather parameters in real time during train operation and using the tunnel mileage prediction error compensation model to obtain the final mileage value to the tunnel exit, namely O(i); 2. The method of calculating the time for a train to pass through a tunnel according to claim 1, characterized in that it comprises: The specific realization process of step 2) is
1) dividing the tunnel groups into N classes according to the geographical locations where the tunnel samples are located;
2) Obtain the average temperature distribution within T (T is a predetermined time) min before the train passes through the tunnel in one year in the current area, and adopt the Gaussian distribution function to obtain the temperature distribution of the current area Fit the average temperature distribution in the tunnel, divide the average temperature distribution after fitting into K equal parts according to the probability, define the temperature samples belonging to one part as temperature samples of the same class in the current area, and defining the temperature sequence, the humidity sequence and the pressure sequence corresponding to the air temperature samples of the same class as samples of the same class. The specific realization process of step 3) is
1) Modeling an autoregressive integrated moving average model for each time sequence in the same class of samples in the current area, and calculating the features of each time sequence, namely the autoregressive term, the difference term and the extracting the parameters of the moving regression term, wherein all features of the samples of the same class in the area form a feature matrix A, the time sequence includes the temperature sequence, the humidity sequence and the pressure sequence; and
2) Performing dimension reduction processing on the feature matrix A, selecting M principal components with the largest contribution to characterize the information of the original feature matrix A, and obtaining the transformed matrix A′. ,
3) Kernel function k = αk _rbf + βk _Linear + (1-α-β)k _Laplace (where k _rbf is the radial basis kernel function, k _Linear is the linear kernel function, and k _Laplace is the Laplacian kernel function ), and randomly set the initial values of the kernel function coefficients α, β (α, β∈[0, 1]) and the number of classes n ₀ (n ₀ is a positive integer smaller than 30) to decide;
4) Map the feature values in the matrix A′ to the feature space corresponding to the kernel function k, obtain a new feature matrix A″, use the feature amount in A″ as input, and according to the initial class number n ₀ , k-means A clustering algorithm is used to achieve feature clustering in A", dividing the same class samples in the area into n classes, forming a sample cluster from each class sample, and optimizing the objective function

(where avg(C _i ) is the average distance of samples in sample cluster C _i , avg(C _j ) is the average distance of samples in sample cluster C _j , and d _cen (C _i ,C _j ) is ), which is the distance between the center points of sample clusters C _i and C _j ;
5) based on the current optimization objective function value fitness, adopting the gray wolf optimization algorithm to update the values of α, β and _n0 ;
6) Repeating step 4) and step 5 until a predetermined number of optimization times m=100 is reached, at which time the extremes of α, β and _n0 are the final optimized parameter values. , and the clustering result at the final optimized parameter value is the final clustering result, and
7) Divide the samples of the same class in the current area into E clustering sample clusters according to the final clustering result, corresponding to the clustering center of each clustering sample cluster and the P samples closest to the clustering center;

The temperature sequence collected at the head of the train, the humidity sequence collected at the head of the train , the barometric pressure sequence collected at the head of the train, and the temperature collected at the tail of the train, respectively . sequence, the humidity sequence collected at the tail of the train , the barometric pressure sequence collected at the tail of the train. ) and defining the original time sequence corresponding to P+1 samples as a typical sample of the current class tunnel group how to calculate the time it takes to travel through a tunnel. The specific realization process of step 5) is
1) determining a current sample time and a plurality of previous sample time points in the temperature sequence, the humidity sequence and the pressure sequence of the head and tail of the train at the current position;
2) Adopting the delayed coordinate method to perform phase space reconstruction, obtaining six two-dimensional reconstruction matrices representing the evolution characteristics of temperature, humidity and air pressure at the head and tail of the train, and obtaining six matrices are combined according to the RGB color space for each head and tail to form a current position feature image { _ch ,c _f };
3) performing a convolution operation on the current location feature image and the images in the exemplary sample integration template library to obtain a plurality of one-dimensional sequences;
4) sorting the elements in all one-dimensional sequences in descending order and determining the largest L elements as candidate elements, where the pre-sorting position corresponding to the candidate element is the candidate position, and the traveled distance values to the corresponding tunnel exits are s _j (j=1, 2, . . . L);
5) taking the average of the traveled distance values to the tunnel exit corresponding to the candidate locations, where the average value is the best matching location ^OM in the exemplary sample integrated template library of the current location is

The method for calculating the time for a train to pass through a tunnel according to any one of claims 1 to 4, characterized by comprising: The specific realization process of step 6) is
1) Input samples I = (T _h , H _h , P _h , T _f , H _f , P _f ) (where T _h = (t ₁ , t ₂ ..., t _G ) is the H is the temperature sequence of the head current sample point and G previous sample points, where _Hh is the humidity sequence of the train head of length G, _Ph is the length of G; is the pressure sequence at the head of a certain train, T _f is the temperature sequence at the tail of the train of length G, and H _f is the humidity sequence of the tail of the train of length G. and P _f is the pressure sequence at the tail of the train of length G), define an output sample to be the mileage error value ε corresponding to the current position, and define an input sample and an output sample constructing a modeling sample with the combination Y={I, ε} of
2) dividing the modeling samples into a training set and a validation set;
3) perform binary encoding on each dimensional feature in the input sample I, and if the encoded value corresponding to a dimensional feature is 1, select that feature as an input variable for the least squares support vector machine; , discarding a feature of a dimension if the corresponding encoded value is 0;
4) based on the current feature encoding values, update the input samples to obtain an updated training set and a validation set, adopt the updated training set data to train the least squares support vector machine, and update the validation set; Input the data into the trained least -squares support vector machine, and the output sequence of the least-squares support vector machine as

(where k=1,2..., _N1 , where _N1 is the number of sample pairs in the validation set), and
5) Repeat steps 3) and 4) to optimize the objective function

determining the input features and the least -squares support vector machine that minimize A method for calculating the time for a train to pass through a tunnel according to any one of claims 1 to 5. The specific realization process of step 7) is
1) Collect the average temperature distribution data within Tmin before the train passes through the tunnel , obtain the temperature sequence and the humidity sequence of the head and tail of the train at the current position, and obtain the template matching output value O obtaining ^M (i) and utilizing a tunnel mileage prediction error compensation model to obtain a compensation error output result ε(i);
2) obtaining the traveled distance value O(i) from the train's current position to the tunnel exit with the formula O(i)= ^OM (i)+ε(i). A method for calculating the time for a train to pass through a tunnel according to any one of 1 to 6. A best matching position acquisition module for performing correlation calculations on the current position feature image and templates in the typical sample integrated template library to obtain the best matching position of the current position in the typical sample integrated template library. wherein the current position feature image is phase space reconstruction and RGB color using the collected meteorological parameters of the head and tail sample points of the current position of the train and the meteorological parameter sequence of the current position for a certain period of time before; Obtained by performing spatial combination, the template images in the typical sample integrated template library are the head of the train obtained by performing phase-space reconstruction on the representative samples of each class of tunnel group. and a color image of the weather parameters of the tail and a representative sample of the tunnel group for each class classifying the weather parameters according to the geographic location where the tunnel is located and the weather conditions inside and outside the tunnel . , an optimal matching position acquisition module, which is a sample obtained by further dividing samples of the same class within the classified area;
Adopting the data of the tunnel group in the same class, taking the temperature sequence, humidity sequence, and pressure sequence as input, and the prediction error of the best matching position as output, obtained by training a least squares support vector machine, and the output is compensated. a tunnel mileage prediction error compensation model with error ε(i);
Formula t=O(i)/(f _Hz (O(i)−O(i−1))) (where O(i)= ^OM (i)+ε(i) and O(i− 1) is the traveled distance value from the position before the train's current position to the tunnel exit, O ^M (i) is the best matching position, and f _Hz is the position update frequency of the train. and a remaining time calculation module for estimating the remaining time t for the train to pass through the current tunnel. The optimal matching position acquisition module includes:
a data acquisition unit for determining a current sample point and G previous sample points in the temperature sequence, the humidity sequence and the barometric pressure sequence of the head and tail of the train at the current position;
Phase space reconstruction is performed by adopting the delayed coordinate method to obtain six two-dimensional reconstruction matrices representing the evolution characteristics of temperature, humidity and air pressure in the head and tail of the train, and dividing the six matrices. a feature image calculation unit for forming a current position feature image {ch, cf} by combining the head part and the tail part according to the RGB color space;
a convolution unit for performing a convolution operation on the current position feature image and template images in the exemplary sample integration template library to obtain a plurality of one-dimensional sequences;
A sorting unit for sorting the elements in all one-dimensional sequences in descending order and determining the largest L elements as candidate elements, wherein the positions before sorting corresponding to the candidate elements are candidate positions, and a sorting unit whose traveled distance values to tunnel exits corresponding to are s _j (j=1, 2, . . . L);
An output unit for taking the mean value of distance traveled to tunnel exit values corresponding to candidate positions, wherein the mean value is the best matching position ^OM in the exemplary sample integrated template library of the current position is

and an output unit that is
The training process of the tunnel mileage prediction error compensation model includes :
1) Input samples I = (T _h , H _h , P _h , T _f , H _f , P _f ) (where T _h = (t ₁ , t ₂ ..., t _G ) is the H is the temperature sequence of the head current sample point and G previous sample points, where _Hh is the humidity sequence of the train head of length G, _Ph is the length of G; is the pressure sequence at the head of a certain train, T _f is the temperature sequence at the tail of the train of length G, and H _f is the humidity sequence of the tail of the train of length G. and P _f is the pressure sequence at the tail of the train of length G), define an output sample to be the mileage error value ε corresponding to the current position, and define an input sample and an output sample constructing a modeling sample with the combination Y={I, ε} of
2) dividing the modeling samples into a training set and a validation set;
3) perform binary encoding on each dimensional feature in the input sample I, and if the encoded value corresponding to a dimensional feature is 1, select that feature as an input variable for the least squares support vector machine; , discarding a feature of a dimension if the corresponding encoded value is 0;
4) based on the current feature encoding values, update the input samples to obtain an updated training set and a validation set, adopt the updated training set data to train the least squares support vector machine, and update the validation set; Input the data into the trained least -squares support vector machine, and the output sequence of the least-squares support vector machine as

(where k=1,2..., _N1 , where _N1 is the number of sample pairs in the validation set), and
5) Repeat steps 3) and 4) to optimize the objective function

determining the input features and the least -squares support vector machine that minimize 9. The system for calculating the time for a train to pass through a tunnel according to claim 8. A computer readable storage medium, characterized in that it stores a program arranged to perform the steps of the method according to any one of claims 1-7.

Images