CN108344975A - A kind of joint cluster scaling method declined using gradient with included angle cosine - Google Patents

A kind of joint cluster scaling method declined using gradient with included angle cosine Download PDF

Info

Publication number
CN108344975A
CN108344975A CN201810015751.9A CN201810015751A CN108344975A CN 108344975 A CN108344975 A CN 108344975A CN 201810015751 A CN201810015751 A CN 201810015751A CN 108344975 A CN108344975 A CN 108344975A
Authority
CN
China
Prior art keywords
point
vector
supporting vector
cluster
gradient
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201810015751.9A
Other languages
Chinese (zh)
Inventor
侯长波
盛阳
司伟建
曲志昱
邓志安
张春杰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Engineering University
Original Assignee
Harbin Engineering University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN201810015751.9A priority Critical patent/CN108344975A/en
Publication of CN108344975A publication Critical patent/CN108344975A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
    • G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Radar Systems Or Details Thereof (AREA)

Abstract

The present invention provides a kind of joint cluster scaling method declined using gradient with included angle cosine, on the basis of ensureing cluster accuracy rate, by the operating structure of simplifying support vector clustering cluster calibration phase, greatly reduces operation time.Input data is first mapped in higher dimensional space by the method described in the invention, acquire supporting vector point, then the cluster calibration process of supporting vector is improved, it proposes using supporting vector point as initial point, seek Local Extremum by gradient descent method, that is Local Clustering central point, then the similarity of left point and current Local Clustering center is calculated by the method for included angle cosine, left point is belonged in corresponding Local Clustering center, finally local center is merged, to obtain final cluster centre, the sorting of radar signal is completed.Gradient of the present invention declines and included angle cosine cluster calibration algorithm can not only utilize on radar signal sorting, may be also used in the data classification of other adjacent areas.

Description

A kind of joint cluster scaling method declined using gradient with included angle cosine
Technical field
The present invention relates to a kind of joint cluster scaling methods using gradient decline and included angle cosine, belong to high efficiency radar letter Number deinterleaving algorithm research field high efficiency radar signal sorting algorithm research field.
Background technology
Deinterleaving algorithm based on support vector clustering is a new research direction of radar signal sorting, main task It is the otherness made full use of between different radar parameters, to distinguish different radars.Relative to traditional deinterleaving algorithm, support Vector clusters algorithm eliminates the reliance on this parameter of the arrival time of radar pulse signal, can make full use of arteries and veins intrinsic parameter information, packet Include the information such as pulsewidth carrier frequency angle of arrival.The realization of support vector clustering algorithm includes two parts, and first part is Nonlinear Mapping, Radar parameter is mapped to by high n-dimensional sphere n by gaussian kernel function, and searches out the minimal hyper-sphere for including all data points. Second part is cluster calibration, i.e., is demarcated to radar signal parameter in higher dimensional space, by the parameter from same radar signal It is divided into one kind, is divided into inhomogeneity from different radar signal parameters.
The cluster calibration part of support vector clustering algorithm is the key that entire algorithm, and directly affects the performance of algorithm. According to the structural analysis for demarcating operation to cluster it is found that the time complexity of cluster calibration has exponent relation growth with sample number.Due to Now complicated electromagnetic environment, stream of pulses quantity per second reach million, and the sample number directly handled is very huge, spent when Between be also it is huge.How to improve the structure of cluster calibration algorithm on the basis of ensureing algorithm validity so that be greatly lowered Operation time is the key that support vector clustering is applied to practical radar signal sorting system.For this problem, the present invention carries Go out to decline using gradient and has realized cluster calibration process with cosine similarity algorithm.
Invention content
It is profit the purpose of the invention is to provide a kind of joint cluster scaling method using gradient decline and included angle cosine The Local Clustering central point of supporting vector point is found with gradient descent method and the Local Clustering of left point is judged using included angle cosine Center.
The object of the present invention is achieved like this：Step 1：By the pulsewidth PW, carrier frequency CF, arrival direction DOA of radar signal Form P={ p1,p2,...pk...,pN, N is pulse signal number, pkIncluding multi-Dimensional parameters and the correspondence in radar signal sorting Pulsewidth PW, carrier frequency CF, the arrival direction DOA of radar signal；
Step 2：The value of arrange parameter C and q, wherein C are the soft boundary parameter that minimal hyper-sphere is solved in supporting vector, q For gaussian kernel functionWidth parameter；
Step 3：Data P is mapped in higher-dimension suprasphere using gaussian kernel function, according to the distribution of High dimensional space data, Data P is divided into two classes：
The surface that the first kind is located at higher-dimension suprasphere is known as supporting vector point, with setIt indicates, Middle n1For the number of supporting vector, the second class is located at the inside of suprasphere, referred to as non-supporting vector point, with setIt indicates, wherein n2For the number of non-supporting vector points；
Step 4：Setting update coefficient lr and error amount err；
Step 5：To remaining non-supporting vector pointIt is handled, judges its Local Clustering center；
Step 6：To local cluster centre pointIt calculates, then can determine whether returning for all data Belong to：
Take any two points g in GkAnd gh, judge gkAnd ghWhether same cluster centre is belonged to：Straight line path between two points Up-sample M point, if M values, there are a sampled point s, bring formula into generally between 10~20 | Φ (s)-a | value be more than Suprasphere radius R, then it is assumed that gkAnd ghBelong to different cluster centres.
The invention also includes some such structure features：
1. step 4 specifically includes：
Grad is sought using supporting vector point as initial point, with supporting vector point akFor initial point x0, according to formula：
Obtain Grad x*,
WhereinG(x)-1 With new coefficient lr, to update x0=x0-x*, obtain new x0, until n-th Grad and the N-1 times gradient value difference it is absolute When value is less than err, stop update；The x for claiming n-th to obtain0It is supporting vector point akLocal Clustering central point, in traversal set A All points obtain a series of set of corresponding Local Clustering central pointsBy office in set A Portion's cluster centre point is gkSupporting vector point extract composition setThen gatherClaim the office of E Portion's cluster centre point is gk。
2. step 5 specifically includes：
Appoint and takes a point bkIf it belongs to Local Clustering center gk, then in gkCorresponding set E={ e1,e2,…,en4In must In the presence of point ek, construction vectorAnd vectorUtilize formulaObtain vectorAnd vector's Included angle cosine value, obtained cosine value are that other press maximum in the calculated cosine of vector institute that same method construct goes out , and bkgkMould be less than vector ekgkMould；It will point bkAdd to its local center gkCorresponding setIn, traversal setIn all point, point therein is belonged into its cluster centre In the corresponding set of point.
Compared with prior art, the beneficial effects of the invention are as follows：It is of the present invention poly- using gradient included angle cosine method Class algorithm is located at due to supporting vector point on the spherical surface of higher dimensional space, and after mapping back lower dimensional space, these supporting vector points are proper Benefit surrounds the point for belonging to a cluster centre in the boundary of cluster, and the place closer from cluster centre point, Distance to the suprasphere centre of sphere is closer.Therefore, it can be found and supported using gradient descent method using supporting vector point as initial point The Local Clustering central point of vector point, the processing to remaining non-supporting vector point.The present invention proposes it between cluster centre again M-cosine is formed to judge attaching problem, since these non-supporting vector points are comprised in supporting vector point and Local Clustering Between heart point, so it is necessarily smaller than equidirectional supporting vector point to Local Clustering central point to the distance of cluster centre point Distance, can accurately find the Local Clustering center of left point in this approach, finally these are looked for by gradient and included angle cosine To cluster centre carry out traditional cluster calibration, form final cluster centre.The method demarcates the cluster of support vector clustering Operating structure greatly simplifies, and significantly reduces operation time.
Advantages of the present invention is as follows：
(1) it is put forward for the first time and carries out gradient decline searching Local Clustering center using supporting vector point.
(2) be put forward for the first time using to left point using the model split of included angle cosine to corresponding Local Clustering center.
(3) structure for greatlying simplify the operation of 2 cluster of support vector clustering algorithm steps calibration, reduces operation time, Algorithm engineering is set to be implemented as possibility.
Description of the drawings
Fig. 1 is the flow chart of the present invention；
Fig. 2 is the two-dimentional input data distribution map after normalization；
Fig. 3 is the supporting vector point schematic diagram obtained by Nonlinear Mapping；
Fig. 4 is the supporting vector Local Clustering central point declined by gradient；
Fig. 5 is as umber of pulse increases, and cluster demarcates operation time schematic diagram.
Specific implementation mode
Present invention is further described in detail with specific implementation mode below in conjunction with the accompanying drawings.
In conjunction with Fig. 1 to Fig. 5, step of the invention is as follows：
Step (1)：By the parameters such as the pulsewidth of radar signal, carrier frequency, arrival direction composition P={ p1,p2,...pk...,pN} Shown, N is pulse signal number.pkIncluding multi-Dimensional parameters, correspond to pulsewidth PW, the carrier frequency of radar signal in radar signal sorting The information such as CF, arrival direction DOA, p-shaped at data input；
Step (2)：The value of arrange parameter C and q, wherein C are the soft boundary parameter that minimal hyper-sphere is solved in supporting vector, Q is gaussian kernel functionWidth parameter.
Step (3)：Data P is mapped in higher-dimension suprasphere using gaussian kernel function, according to point of High dimensional space data Data P is divided into two classes by cloth.The surface that the first kind is located at higher-dimension suprasphere is known as supporting vector point, with setIt indicates, wherein n1For the number of supporting vector, the second class is located at the inside of suprasphere, referred to as non-supporting Vector point, with setIt indicates, wherein n2For the number of non-supporting vector points.
Step (4)：Setting update coefficient lr and error amount err.Grad is sought by initial point of supporting vector point, i.e., with branch Hold vector point akFor initial point x0, utilize formulaObtain Grad x*, thereinFor x points to the suprasphere centre of sphere Distance, G (x)-therein1With new coefficient lr, to update x0=x0-x*, obtain new x0.Until the Grad and N-1 of n-th When secondary Grad absolute value of the difference is less than err, stop update.The x for claiming n-th to obtain0It is supporting vector point akLocal Clustering Central point traverses point all in set A, obtains a series of set of corresponding Local Clustering central pointsIt is g by Local Clustering central point in set AkSupporting vector point extract composition setThen gatherThe Local Clustering central point of E is referred to as gk。
Step (5)：To remaining non-supporting vector pointIt is handled, is judged in its Local Clustering The heart, the specific method is as follows：Appoint and takes a point bkIf it belongs to Local Clustering center gk, then in gkCorresponding setIn must have point ek, construction vectorAnd vectorUtilize formulaIt calculates VectorAnd vectorIncluded angle cosine value, this cosine value is that other are pressed the vector that same method construct goes out and are calculated It is maximum in the cosine gone out, and bkgkMould be less than vector ekgkMould.It at this time will point bkAdd to its local center gkIt is corresponding SetIn, traversal setIn all point, point therein is belonged to it and is gathered In the corresponding set of class central point.
Step (6)：The Local Clustering center that step 4 and step 5 are formed can represent the cluster property of total data.So To local cluster centre pointIt calculates, then can determine whether the ownership of all data.Take any two points in G gkAnd gh, judge gkAnd ghWhether the method for same cluster centre is belonged to：Straight line path up-samples M point between two points, and M takes Value generally between 10~20, if there are a sampled point s, bring formula into | Φ (s)-a | value be more than suprasphere radius R, then Think gkAnd ghBelong to different cluster centres.
The descent coefficient lr and error function err, it is characterised in that the value of lr determines the rate that gradient declines, mistake It is big then decrease speed is fast, it is easy to skip Local Clustering center, it is too small, increase time complexity, general value is 0.1~0.2. Error function err is with to determine whether dropping to Local Clustering central point, and value size determines the position of true cluster centre It sets, between general value 0.05~0.2.
The sampling number M, which is characterized in that it is the sampling carried out between obtained Local Clustering central point, The size of its value determines the accuracy rate of cluster, excessive, increases the time, and too small accuracy rate reduces, and general M values of choosing are Between 10~20.
Referring to Fig.1, it is the cluster scaling method flow chart declined based on gradient using described with included angle cosine.
With reference to Fig. 2, it is the figure of supporting vector point, initial data is mapped to higher dimensional space, finds and is located in higher dimensional space These points are mapped in luv space, just constitute supporting vector point by the point of spherical surface, form the boundary of each cluster block, will Data are surrounded.
It is that the method passes through the office to decline as initial point gradient with reference to the supporting vector point in Fig. 2 with reference to Fig. 3 Portion's cluster centre.
With reference to Fig. 4, the method is with the increase of umber of pulse, the time of CPU consumption.
It is above-mentioned for the present invention it is special for embodiment, be not limited to the present invention.The present invention is applicable not only to two-dimensional parameter, The case where being also applied for multi-Dimensional parameters simultaneously；Be applicable not only to radar signal parameter sorting, while be also applied for its it is similar with it is logical It is the case where initial data classifies to target to cross extraction characteristic parameter.It is not departing from the spirit and scope of the invention, it can It does a little adjustment and optimization and protection scope of the present invention is subject to claim.
To sum up, the invention discloses a kind of methods declined using gradient and included angle cosine carries out cluster calibration.This method exists On the basis of ensureing cluster accuracy rate, by the operating structure of simplifying support vector clustering cluster calibration phase, fortune is greatly reduced Evaluation time.Input data is first mapped in higher dimensional space by the method described in the invention, acquires supporting vector point, then to supporting The cluster calibration process of vector is improved, and proposes, using supporting vector point as initial point, to seek local extremum by gradient descent method Then it is similar to current Local Clustering center to calculate left point by the method for included angle cosine for point, i.e. Local Clustering central point Degree, left point is belonged in corresponding Local Clustering center, is finally merged to local center, final to obtain Cluster centre, complete the sorting of radar signal.Gradient of the present invention declines and included angle cosine cluster calibration algorithm not only may be used With using in the data classification that on radar signal sorting, may be also used in other adjacent areas.

Claims (3)

1. a kind of joint cluster scaling method declined using gradient with included angle cosine, it is characterised in that：Steps are as follows：
Step 1：By the pulsewidth PW of radar signal, carrier frequency CF, arrival direction DOA composition P={ p1,p2,...pk...,pN, N is arteries and veins Rush signal number, pkIncluding multi-Dimensional parameters and corresponding to the pulsewidth PW, carrier frequency CF, arrival side of radar signal in radar signal sorting To DOA；
Step 2：The value of arrange parameter C and q, wherein C are the soft boundary parameter that minimal hyper-sphere is solved in supporting vector, and q is height This kernel functionWidth parameter；
Step 3：Data P is mapped in higher-dimension suprasphere using gaussian kernel function, according to the distribution of High dimensional space data, will be counted It is divided into two classes according to P：
The surface that the first kind is located at higher-dimension suprasphere is known as supporting vector point, with setIt indicates, wherein n1 For the number of supporting vector, the second class is located at the inside of suprasphere, referred to as non-supporting vector point, with set It indicates, wherein n2For the number of non-supporting vector points；
Step 4：Setting update coefficient lr and error amount err；
Step 5：To remaining non-supporting vector pointIt is handled, judges its Local Clustering center；
Step 6：To local cluster centre pointIt calculates, then can determine whether the ownership of all data：
Take any two points g in GkAnd gh, judge gkAnd ghWhether same cluster centre is belonged to：Straight line path up-samples between two points If M point, M values, there are a sampled point s, bring formula into generally between 10~20 | Φ (s)-a | value be more than suprasphere Radius R, then it is assumed that gkAnd ghBelong to different cluster centres.
2. a kind of joint cluster scaling method declined using gradient with included angle cosine according to claim 1, feature are existed In：Step 4 specifically includes：
Grad is sought using supporting vector point as initial point, with supporting vector point akFor initial point x0, according to formula：
Obtain Grad x*,
WhereinG(x)-1For with New coefficient lr, updates x0=x0-x*, obtain new x0, until Grad and the N-1 times Grad absolute value of the difference of n-th are small When err, stop update；The x for claiming n-th to obtain0It is supporting vector point akLocal Clustering central point, it is all in traversal set A Point, obtain a series of set of corresponding Local Clustering central pointsTo locally it gather in set A Class central point is gkSupporting vector point extract composition setThen gatherClaim the part of E poly- Class central point is gk。
3. a kind of joint cluster scaling method declined using gradient with included angle cosine according to claim 2, feature are existed In：Step 5 specifically includes：
Appoint and takes a point bkIf it belongs to Local Clustering center gk, then in gkCorresponding setIn must exist a little ek, construction vectorAnd vectorUtilize formulaObtain vectorAnd vectorAngle more than String value, it is maximum in the calculated cosine of vector institute that same method construct goes out that obtained cosine value is that other are pressed, and bkgk Mould be less than vector ekgkMould；It will point bkAdd to its local center gkCorresponding setIn, traversal SetIn all point, point therein is belonged in its corresponding set of cluster centre point.
CN201810015751.9A 2018-01-08 2018-01-08 A kind of joint cluster scaling method declined using gradient with included angle cosine Pending CN108344975A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810015751.9A CN108344975A (en) 2018-01-08 2018-01-08 A kind of joint cluster scaling method declined using gradient with included angle cosine

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810015751.9A CN108344975A (en) 2018-01-08 2018-01-08 A kind of joint cluster scaling method declined using gradient with included angle cosine

Publications (1)

Publication Number Publication Date
CN108344975A true CN108344975A (en) 2018-07-31

Family

ID=62961334

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810015751.9A Pending CN108344975A (en) 2018-01-08 2018-01-08 A kind of joint cluster scaling method declined using gradient with included angle cosine

Country Status (1)

Country Link
CN (1) CN108344975A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109613486A (en) * 2018-12-03 2019-04-12 中国人民解放军空军工程大学 A kind of Radar Signal Sorting Method based on core cluster support vector clustering
CN113033615A (en) * 2021-03-01 2021-06-25 电子科技大学 Radar signal target real-time association method based on online micro-cluster clustering
CN113775929A (en) * 2021-09-28 2021-12-10 上海天麦能源科技有限公司 Urban gas pipe network layout area division method
CN113850995A (en) * 2021-09-14 2021-12-28 华设设计集团股份有限公司 Event detection method, device and system based on tunnel radar vision data fusion

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109613486A (en) * 2018-12-03 2019-04-12 中国人民解放军空军工程大学 A kind of Radar Signal Sorting Method based on core cluster support vector clustering
CN113033615A (en) * 2021-03-01 2021-06-25 电子科技大学 Radar signal target real-time association method based on online micro-cluster clustering
CN113850995A (en) * 2021-09-14 2021-12-28 华设设计集团股份有限公司 Event detection method, device and system based on tunnel radar vision data fusion
CN113775929A (en) * 2021-09-28 2021-12-10 上海天麦能源科技有限公司 Urban gas pipe network layout area division method

Similar Documents

Publication Publication Date Title
CN108344975A (en) A kind of joint cluster scaling method declined using gradient with included angle cosine
Li et al. Toward convolutional neural networks on pulse repetition interval modulation recognition
WO2021142916A1 (en) Proxy-assisted evolutionary algorithm-based airfoil optimization method and apparatus
Dai et al. Combination of DNN and improved KNN for indoor location fingerprinting
CN104346481B (en) A kind of community detection method based on dynamic synchronization model
CN111062425B (en) Unbalanced data set processing method based on C-K-SMOTE algorithm
CN106326923B (en) A kind of position data clustering method of registering taking position into account and repeating with density peaks point
CN106780458B (en) Point cloud framework extraction method and device
CN110443810A (en) Point cloud plane dividing method based on quick adjacent voxel inquiry
CN110489419A (en) A kind of k nearest neighbor approximation querying method based on multilayer local sensitivity Hash
CN111027140B (en) Airplane standard part model rapid reconstruction method based on multi-view point cloud data
CN101908213B (en) SAR image change detection method based on quantum-inspired immune clone
CN107703480A (en) Mixed kernel function indoor orientation method based on machine learning
El‐Sayed et al. Plane detection in 3D point cloud using octree‐balanced density down‐sampling and iterative adaptive plane extraction
CN105654483A (en) Three-dimensional point cloud full-automatic registration method
CN113344019A (en) K-means algorithm for improving decision value selection initial clustering center
CN102243641A (en) Method for efficiently clustering massive data
WO2022088390A1 (en) Image incremental clustering method and apparatus, electronic device, storage medium and program product
CN104850712B (en) Surface sampled data topology Region Queries method in kind
CN109271427A (en) A kind of clustering method based on neighbour's density and manifold distance
CN108566620A (en) A kind of indoor orientation method based on WIFI
Song et al. Binary open clusters in the Gaia data
CN109284409A (en) Picture group geographic positioning based on extensive streetscape data
CN104102718A (en) Method for unbalanced data classification by unbalanced data set oriented cluster boundary sampling and dynamic pruning
CN104123382B (en) A kind of image set abstraction generating method under Social Media

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20180731

RJ01 Rejection of invention patent application after publication