A kind of Railway Roadbed intelligent identification Method based on GPR
Technical field
The invention belongs to inspection of railway subgrade technical field, is related to a kind of Railway Roadbed intelligence based on GPR
Recognition methods.
Background technology
The evaluation of railway bed state-detection is the key link of railway maintenance maintenance.At present in detection means, GPR
(GPR)It is a kind of ideal detection method.But GPR data process and interpretations rely primarily on artificial interpretation both at home and abroad at present,
Efficiency is low, poor in timeliness.Therefore, a kind of fast and accurately railway bed GPR data processing methods are established, are China railways roads
Urgent problem to be solved in base detection.
From 2008, start the research for a small amount of Railway Roadbed intelligent identification Method occur.Research is more with railway at present
Professional is the technical background such as main body, machine vision, pattern-recognition weakness, and the Damage Types of consideration are less, and do not examine
Consider the main units such as track switch, bridge;Using one-dimensional single track radar data as recognition unit, it is segmented along depth direction, extraction one
The radar signal characteristic value of dimension;The accuracy rate for finally carrying out pattern-recognition is not high, it is difficult to puts into actual engineer applied.
The content of the invention
It is an object of the invention to provide a kind of Railway Roadbed intelligent identification Method based on GPR, realizes fast
Fast, accurate, lossless Railway Roadbed identification.
The present invention solves the technical scheme that its technical problem uses:A kind of Railway Roadbed intelligence based on GPR
Energy recognition methods, is specifically included:
(1)Establish Railway Roadbed Intelligent Recognition software systems
a)Normal railway bed, the railway bed comprising different type subgrade defect, railroad bridge, road are detected using GPR
Trouble, preserve detection data;
b)Pretreatment:Detection data are subjected to zero line correction, are converted to gray level image;
c)Two-dimensional discrete:The ground penetrating radar image of acquisition is divided into several recognition units along mileage direction, each identification is single
Member include 50 ~ 150 track datas, then each recognition unit is divided into several along depth direction and identifies subelement, and adjacent two
Individual identification subelement has 50% overlapping region;
d)Feature extraction:Various features value, all identification subelements of each recognition unit are extracted in units of identifying subelement
Characteristic value form the characteristic vector of the recognition unit, original dimensions M=identification subelement number × characteristic value of characteristic vector
Number;
e)Feature Dimension Reduction:Determine dimensionality reduction dimension N, using principal component analysis to characteristic vector carry out dimensionality reduction, structure low-dimensional feature to
Amount;
f)Build identification model:Support vector machine classifier is established, low-dimensional characteristic vector is input in grader, trains this point
Class device, build the Railway Roadbed intelligent recognition model based on GPR;
(2)Carry out Railway Roadbed identification
Railway bed to be identified is detected using GPR, preserves detection data;Carried out by the Intelligent Recognition software established pre-
Processing, two-dimensional discrete and feature extraction, Feature Dimension Reduction is carried out with dimensionality reduction dimension N;The section of railway track roadbed is entered using identification model
Row identification, obtain the roadbed type of each recognition unit of section of railway track roadbed.
Described subgrade defect includes rising soil, ballast contamination, sinking, aqueous, empty.
Described characteristic value includes the signal characteristic of radar image, such as energy, variance;Histogram statistical features, as average,
Standard variance, smoothness, third moment, uniformity, entropy.
Described determination dimensionality reduction dimension N detailed process is as follows:
(1)A series of dimension values are set in the range of 8 ~ M, principal component analysis is utilized respectively and dimensionality reduction is carried out to characteristic vector, often
The data set after one group of dimensionality reduction is obtained under individual dimension values;
(2)It is utilized respectively the Data set reconstruction raw data set after each group dimensionality reduction, calculates reconstructed data set and raw data set
Root-mean-square error;
(3)In dimension values of the root-mean-square error less than 0.5%, the dimension values of minimum are selected as dimensionality reduction dimension N.
Beneficial effect
Compared with background technology, the invention has the advantages that:
(1)Using machine vision, mode identification technology, quick, accurate, the lossless intelligence of a variety of Railway Roadbeds can be achieved
Identification;
(2)Polytype two dimensional character such as two-dimensional discrete, extraction signal characteristic, histogram statistical features is carried out to GPR images
Value, optimize the character representation of subgrade defect so that discrimination significantly improves;
(3)Dimensionality reduction dimension is determined using the root-mean-square error of reconstructed data set and raw data set, both can guarantee that low after dimensionality reduction
Dimensional feature vector reduces data redudancy, lifts recognition efficiency again to the sign performance of raw data set.
Brief description of the drawings
The software systems that Fig. 1 is the present invention build flow chart.
Fig. 2 is the two-dimensional discrete process schematic of the present invention.
Embodiment
The invention will be further described with embodiment below in conjunction with the accompanying drawings.
A kind of Railway Roadbed intelligent identification Method based on GPR, including establish Railway Roadbed and intelligently know
Other software systems, progress Railway Roadbed identify two parts.
First, establish Railway Roadbed Intelligent Recognition software systems.
(1)Normal railway bed, the railway bed comprising different type subgrade defect, railway bridge are detected using GPR
Beam, track switch, preserve detection data;Subgrade defect includes rising soil, ballast contamination, sinking, aqueous, empty;
(2)Pretreatment:Detection data are subjected to zero line correction, are converted to gray level image;
(3)Two-dimensional discrete:As shown in Fig. 2 the ground penetrating radar image A of acquisition is divided into several recognition units along mileage direction
{A1, A2..., Ak, each recognition unit includes 50 ~ 150 track datas, if then being divided into each recognition unit along depth direction
Dry identification subelement, adjacent two identification subelements have 50% overlapping region, as shown in Fig. 2 recognition unit A2It is divided into { A21,
A22, A23, A24...;
(4)Feature extraction:Various features value is extracted in units of identifying subelement, includes the signal characteristic of radar image, if
Amount, variance;Histogram statistical features, such as average, standard variance, smoothness, third moment, uniformity, entropy;Each recognition unit
The characteristic values of all identification subelements form the characteristic vector of the recognition unit, the original dimensions M of characteristic vector=for identification it is single
First number × characteristic value number;
(5)Feature Dimension Reduction:Different dimension values { D is set in the range of 8 ~ M1, D2..., Dk, it is utilized respectively principal component analysis
Dimensionality reduction is carried out to characteristic vector, the data set after one group of dimensionality reduction is obtained under each dimension values;The number being utilized respectively after each group dimensionality reduction
Raw data set is reconstructed according to collection, calculates the root-mean-square error { E of reconstructed data set and raw data set1, E2..., Ek};Equal
In dimension values of the square error less than 0.5%, the dimension values of minimum are selected as dimensionality reduction dimension N;
(6)Build identification model:Support vector machine classifier is established, low-dimensional characteristic vector is input in grader, training should
Grader, the Railway Roadbed intelligent recognition model based on GPR is built, it is soft to complete Railway Roadbed Intelligent Recognition
The structure of part system.
Second, carry out Railway Roadbed identification.
(1)Railway bed to be identified is detected using GPR, preserves detection data;
(2)Pre-processed by the Intelligent Recognition software established, two-dimensional discrete, feature extraction;
(3)Feature Dimension Reduction is carried out with dimensionality reduction dimension N, builds low-dimensional characteristic vector;
(4)Low-dimensional characteristic vector is input in identification model, the section of railway track roadbed is identified using identification model, is obtained
The roadbed type of each recognition unit of section of railway track roadbed.
Embodiment described above is only used for the present invention, rather than limitation of the present invention, person skilled in the relevant technique,
In the case of not departing from the spirit and scope of the present invention, various conversion or modification, therefore all equivalent technologies can also be made
Scheme should also belong to scope of the invention, should be limited by each claim.