CN111860602A - Machine learning-based efficient and rapid industrial spectrum cognition method - Google Patents

Machine learning-based efficient and rapid industrial spectrum cognition method Download PDF

Info

Publication number
CN111860602A
CN111860602A CN202010577278.0A CN202010577278A CN111860602A CN 111860602 A CN111860602 A CN 111860602A CN 202010577278 A CN202010577278 A CN 202010577278A CN 111860602 A CN111860602 A CN 111860602A
Authority
CN
China
Prior art keywords
spectrum
industrial
data
frequency spectrum
cognition
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
CN202010577278.0A
Other languages
Chinese (zh)
Inventor
许驰
李琳
曾鹏
于海斌
金曦
夏长清
刘晓宇
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.)
Shenyang Intelligent Robot Innovation Center Co Ltd
Shenyang Intelligent Robot National Research Institute Co ltd
Shenyang Institute of Automation of CAS
Original Assignee
Shenyang Intelligent Robot Innovation Center Co Ltd
Shenyang Intelligent Robot National Research Institute Co ltd
Shenyang Institute of Automation of CAS
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 Shenyang Intelligent Robot Innovation Center Co Ltd, Shenyang Intelligent Robot National Research Institute Co ltd, Shenyang Institute of Automation of CAS filed Critical Shenyang Intelligent Robot Innovation Center Co Ltd
Priority to CN202010577278.0A priority Critical patent/CN111860602A/en
Publication of CN111860602A publication Critical patent/CN111860602A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover

Abstract

The invention relates to a spectrum sensing technology, in particular to a high-efficiency and rapid industrial spectrum cognition method based on machine learning. Aiming at the problems of large calculated amount, low speed, high energy consumption and the like when the existing spectrum sensing method is used for processing the spectrum big data in the industrial complex electromagnetic environment, the method firstly adopts a Principal Component Analysis (PCA) method to perform dimensionality reduction processing on the industrial spectrum big data, extracts characteristic data, reduces the complexity of data processing and improves the sensing efficiency; then, a single support vector machine (OCSVM) is adopted to carry out classification training learning on the dimensionality reduction data, and a Particle Swarm Optimization (PSO) is utilized to carry out iterative optimization on training parameters, so that the frequency spectrum cognition accuracy is improved, and the situation that the training parameters are in local optimum is avoided; and finally, recognizing the real-time industrial spectrum by adopting an industrial spectrum cognitive model for optimizing learning. The method has the characteristics of low complexity, high sensing precision, high speed and high energy efficiency, and can realize the crossing of the spectrum sensing to the cognition.

Description

Machine learning-based efficient and rapid industrial spectrum cognition method
Technical Field
The invention relates to a spectrum sensing technology, in particular to a high-efficiency and rapid industrial spectrum cognition method based on machine learning.
Background
The rapid development of the 5G technology can not only meet the requirement of large bandwidth of traditional human communication, but also meet the requirements of large connection, low time delay and high reliability of machine communication. Therefore, 5G will support industrial wireless control, and is considered as one of the core technologies of industrial internet, which will promote the rapid development of industry 4.0. However, unlike commercial 5G technology, the licensed spectrum is pure and can be used for centralized scheduling of a single 5G network; industrial applications lack licensed spectrum, can only use unlicensed spectrum for public access, and must share the use with other networks. Therefore, 5G (i.e., industrial 5G) applied in industrial scenes is inevitably interfered by random bursts of other networks. Meanwhile, in an industrial environment, conditions such as temperature, humidity and vibration are variable, the mobile metal equipment is seriously shielded randomly, the industrial radio frequency environment is complicated and variable, the communication quality is seriously influenced, and the real-time performance and the reliability of industrial communication are difficult to ensure.
In order to ensure the communication quality of the industrial 5G, avoid interference to other networks, and achieve harmonious coexistence of heterogeneous networks, the industrial 5G must first perform spectrum sensing to ensure reliable quality of used spectrum. The traditional spectrum sensing method comprises energy detection, cyclostationary feature detection, matched filtering, covariance detection and the like. However, these methods often rely on a fixed channel model and are difficult to handle with massive spectrum state data in broadband communication. Therefore, methods such as compressed sensing and wavelet transformation are also applied to spectrum sensing, and the problem of broadband spectrum sensing is solved. However, when processing industrial spectrum big data, the methods have the disadvantages of large calculation amount, low speed, high energy consumption and the like, and cannot be applied to a dynamically variable industrial electromagnetic environment. In addition, the existing method is mainly based on a classical model, still stays at the level of spectrum sensing, and does not achieve spectrum cognition.
Disclosure of Invention
The invention provides a machine learning-based efficient and rapid industrial spectrum cognition method, aiming at the problems that the existing spectrum perception method is large in calculated amount, low in speed, high in energy consumption and difficult to realize spectrum cognition in a complex electromagnetic environment when processing industrial spectrum big data. The method comprises the steps of firstly, carrying out dimensionality reduction processing on industrial frequency spectrum big data by adopting a Principal Component Analysis (PCA) method, extracting characteristic data, reducing data processing complexity and improving perception efficiency; then, a single support vector machine (OCSVM) is adopted to carry out classification training learning on the dimensionality reduction data, and a Particle Swarm Optimization (PSO) is utilized to carry out iterative optimization on training parameters, so that the frequency spectrum cognition accuracy is improved, and the situation that the training parameters are in local optimum is avoided; and finally, recognizing the real-time industrial spectrum by adopting an industrial spectrum cognitive model for optimizing learning. The method can greatly reduce the spectrum learning calculation amount, improve the spectrum cognition speed and reduce the energy consumption. Compared with the traditional spectrum sensing method, the method has higher sensing sensitivity and lower algorithm complexity.
The technical scheme adopted by the invention for realizing the purpose is as follows: a machine learning-based efficient and rapid industrial spectrum cognition method comprises the following steps:
(1) Industrial spectrum big data acquisition: acquiring electromagnetic spectrum data in an industrial environment, and establishing a binary classification model of a frequency spectrum according to whether the frequency spectrum is occupied;
(2) PCA dimension reduction treatment: carrying out dimensionality reduction on the acquired electromagnetic spectrum data by using a PCA algorithm, and extracting a characteristic vector;
(3) OCSVM data training: for the feature vector after dimensionality reduction, establishing a Lagrange dual decomposition problem based on a Gaussian kernel function by solving a quadratic programming problem about the spectrum occupation condition, and solving an optimal classification hyperplane to obtain a decision function of spectrum classification;
(4) PSO parameter optimization: carrying out iterative optimization on intrinsic parameters and Gaussian kernel function parameters in an OCSVM data training process by adopting PSO (particle swarm optimization), and obtaining an optimized classification decision function as a spectrum cognitive model;
(5) and (3) frequency spectrum cognition: and carrying out PCA (principal component analysis) dimensionality reduction on the frequency spectrum signal to be detected, and then inputting the frequency spectrum signal to be detected into a frequency spectrum cognition classification model to obtain the frequency spectrum occupation condition.
The electromagnetic spectrum data under the industrial environment is acquired under one of the following two conditions:
an idle state: in the industrial frequency spectrum space to be monitored, the industrial wireless system is in a halt state;
busy state: and in the industrial frequency spectrum space to be monitored, the industrial wireless system operates according to industrial requirements.
The spectrum classification binary model is as follows:
Figure BDA0002550183230000021
where x (t) represents a wideband signal received by the transceiver, v (t) represents a noise signal, s (t) represents a signal of another industrial wireless system, H (t) represents a channel gain, and H (t) represents a channel gain0The representative spectrum is free and not occupied by other industrial wireless systems; h1The representative frequency spectrum is busy and is occupied by other industrial wireless systems in the same frequency band.
The PCA dimension reduction processing comprises the following steps:
a. establishing an original data matrix of the industrial frequency spectrum according to the number m of the samples and the characteristic number n of the samples
Figure BDA0002550183230000031
Wherein x isi＝(x1i，x2i，…，xmi) T, i ═ 1, 2, …, n denotes industrial spectrum sample characteristics;
b. standardizing the sample, namely performing zero equalization processing on each column of X to obtain
Figure BDA0002550183230000032
c. Computing covariance matrices for normalized samples
Figure BDA0002550183230000033
And its eigenvalue λ and eigenvector W, satisfying CW ═ λ W;
d. each sample xiConversion into a new sample yi＝WTxiTo obtain a new sample matrix Y ═ (Y)1，y2，...，yn)；
e. Calculating the contribution of variance
Figure BDA0002550183230000034
Wherein λ isiSelecting the first k principal components for the ith element in the characteristic value lambda, ensuring that the accumulated contribution rate of the first k principal components reaches a set value or above, representing the original n-dimensional characteristic by using the k-dimensional characteristic to realize data dimension reduction, and obtaining a matrix after dimension reduction as Y (Y is equal to 1，y2，...，yk)。
The industrial spectrum sample characteristics comprise signal power, time, arrival angle, arrival time, synchronous signals, data packet size, source address, destination address, forwarding address and port number.
The OCSVM data training comprises the following steps:
a. establishing a quadratic programming problem with respect to spectrum occupancy
Figure BDA0002550183230000041
Wherein, the reduced sample y1，…，ymE.g. Y, m is the number of samples; phi: y → H represents the mapping of the original space to the feature space; omega and rho are respectively a normal vector and an offset of a required hyperplane in a feature space; ν epsilon (0, 1)]To balance the parameter, xiiIs a relaxation variable, representing the degree to which the training samples are allowed to be misclassified, R represents a real number;
b. establishing a Lagrangian function
Figure BDA0002550183230000042
Wherein alpha isi，βiIs a Lagrange factor; c. for omega, rho and xiiRespectively calculating partial derivatives to obtain
Figure BDA0002550183230000043
d. Construction of the Gaussian Kernel function K (y)i，yj)＝＜Φ(yi)，Φ(yj)＞＝exp(-g||yi-yj||2) (ii) a Wherein < phi (y)i)，Φ(yj) > represents the calculation of phi (y)i) And Φ (y)j) Inner product of, yi，yjRespectively representing ith and j samples after dimensionality reduction, wherein i and j are 1 and 2.
e. Solving dual decomposition problem
Figure BDA0002550183230000044
Choose any one to satisfy
Figure BDA0002550183230000045
Parameter α of as parameter α*Calculating the offset
Figure BDA0002550183230000046
To obtain alphai *Corresponding vector yiIs a support vector and obtains a decision function of spectrum classification as
Figure BDA0002550183230000047
Where y represents the input sample, f (y) represents the classification decision function, and sgn represents the sign function.
The PSO parameter optimization comprises the following steps:
a. initializing a maximum iteration number M; randomly initializing particle swarm U ═ U1，u2，...，uPP denotes the total number of particles, where the speed of the P-th particle is Vp＝(Vp1，Vp2，...，Vpk)TAt the position Sp＝(Sp1，Sp2，...，Spk)TFor pointing to the positions of k features, k being the number of features reduced in dimension, position Spk＝(Spv，Spg) Is composed of two components, which respectively represent an OCSVM intrinsic parameter v and a Gaussian kernel function parameter g, and the limited ranges are [ Svmin，Svmax]And [ S ]gmin，Sgmax]；Svmin，SvmaxMinimum and maximum values representing the intrinsic parameter v; sgmin，SgmaxRepresenting the minimum and maximum values of the gaussian kernel function parameter g;
b. calculating a fitness value f (Sp) of the current particle based on a classification decision function f (y);
c. updating individual extremum and group extremum:
if the fitness value is
Figure BDA0002550183230000051
Then
Figure BDA0002550183230000052
Otherwise
Figure BDA0002550183230000053
Where h denotes the h-th iteration,
Figure BDA0002550183230000054
representing the current bit of the particle at the h-th iterationThe device is placed in a water tank,
Figure BDA0002550183230000055
representing the position of the individual extreme point at the h iteration;
computing
Figure BDA0002550183230000056
GhExpressing the position of a global extreme point of the whole particle swarm in the h iteration, namely a swarm extreme value;
d. updating the velocity and position of particles
Figure BDA0002550183230000057
Figure BDA0002550183230000058
Wherein the content of the first and second substances,
Figure BDA0002550183230000059
represents the velocity of the particle p at the h-th iteration, c1、c2To be an acceleration factor, r1，r2Is [0, 1 ]]Random numbers, omega, distributed uniformly withinpsoIs the inertial weight;
e. judging whether a termination condition is met:
If the current iteration number exceeds the maximum iteration number M or the change of the fitness function value for N continuous times does not exceed the threshold value sigma, exiting the iteration, and the particle position corresponding to the group extreme value at the moment is the optimal parameter, namely the intrinsic parameter v and the Gaussian kernel function parameter g; and if the exit condition is not met, returning to the step a.
The obtained spectrum occupation condition specifically comprises: inputting the frequency spectrum dimensionality reduction data Z to be detected into a frequency spectrum cognition classification model to obtain f (Z), wherein the specific judgment method comprises the following steps:
if the data collected in the step (1) is the spectrum data in a busy state, judging that the spectrum is busy and occupied when f (Z) > 0; when f (Z) < 0, judging that the frequency spectrum is idle and unoccupied;
if the data collected in the step (1) is spectrum data in an idle state, judging that the spectrum is idle and unoccupied when f (Z) > 0; and when f (Z) < 0, the frequency spectrum is judged to be busy and occupied.
The invention has the following advantages and beneficial effects:
1. in the industrial spectrum big data processing process, the PCA method is adopted to perform dimensionality reduction on the mass spectrum data, the key characteristic vector is extracted, the irrelevant vector can be greatly compressed, the data calculation amount is reduced, the spectrum cognition speed and accuracy are improved, and the energy consumption is reduced.
2. The OCSVM is adopted for rapid spectrum cognition, only one sample data in the condition of idle spectrum or busy spectrum is needed to be acquired for training, and data in various conditions are not needed to be acquired, so that the calculation amount can be reduced, the training time is shortened, the training efficiency is improved, and the OCSVM has high recognition accuracy and strong robustness.
3. The method adopts PSO to optimize the parameters of the OCSVM spectrum sensing classification model, avoids local optimization of the parameters through iterative updating of the particles, and can greatly improve the accuracy of classification model training.
Drawings
FIG. 1 is an overall flow chart of a machine learning-based efficient and rapid industrial spectrum cognition method;
FIG. 2 is a flow chart of the PCA data dimension reduction process;
FIG. 3 is a flow chart of OCSVM data training;
fig. 4 is a PSO parameter optimization flow chart.
Detailed Description
The present invention will be described in detail with reference to the accompanying drawings.
The rapid development of the artificial intelligence technology provides a new approach for frequency spectrum cognition. Through continuous learning of the spectrum space, the defect that a traditional sensing method is dependent on a model in a transition mode can be overcome, and accurate spectrum cognition without the model is achieved. Therefore, the machine learning method is adopted to train industrial spectrum big data and recognize industrial complex electromagnetic spectrum space so as to realize high-efficiency and rapid industrial spectrum recognition. The invention can provide basic technical support for industrial 5G, provides basis for industrial 5G communication through accurate cognition on industrial frequency spectrum space, and ensures the characteristics of low time delay and high reliability of the communication.
The invention provides a machine learning-based high-efficiency and rapid industrial frequency spectrum cognition method, which comprises the following steps: (1) acquiring industrial frequency spectrum big data; (2) PCA dimension reduction treatment; (3) training OCSVM data; (4) PSO parameter optimization; (5) and (5) recognizing the frequency spectrum. Fig. 1 is an overall flowchart of an efficient and fast industrial spectrum cognition method based on machine learning. The following detailed description of the present invention is provided in connection with the accompanying drawings.
(1) Industrial spectrum big data acquisition
Under the industrial environment to be monitored, a large-scale antenna array is adopted, radio transceivers are arranged at multiple points, and complex electromagnetic spectrum big data under the industrial environment are collected. The industrial frequency spectrum big data is only acquired under one of the following two conditions: firstly, in an industrial environment to be monitored, all wireless systems (generally any wireless system in the industrial environment, including WiFi, Bluetooth and the like) are in a shutdown state, so that the spectrum space is ensured to be clean; secondly, in the frequency spectrum space to be monitored, the wireless system normally operates according to industrial requirements, namely the frequency spectrum space is dynamically changeable.
On the basis, according to the busy or idle state of the frequency spectrum, a binary hypothesis test model of frequency spectrum classification is established as follows:
Figure BDA0002550183230000071
Where x (t) represents a wideband signal received by the transceiver, v (t) represents a noise signal, s (t) represents a signal of another wireless system, and h (t) represents a channel gain. H0The representative spectrum is free and not occupied by other wireless systems; h1Indicating that the frequency spectrum is busy and occupied by other same-frequency band wireless systems.
(2) PCA dimensionality reduction processing
And carrying out dimensionality reduction on the industrial spectrum big data obtained by sampling by utilizing a PCA algorithm, and extracting a characteristic vector.
As shown in FIG. 2, the specific process is as follows
a. Establishing an original data matrix of the industrial frequency spectrum according to the number m of the samples and the characteristic number n of the samples
Figure BDA0002550183230000072
Wherein xi＝(x1i，x2i，…，xmi)TAnd i is 1, 2, …, and n represents industrial spectrum sample characteristics, including signal power, time, arrival angle, arrival time, synchronization signal, packet size, source address, destination address, forwarding address, port number, and the like.
b. Standardizing the sample, namely performing zero equalization processing on each column of X to obtain
Figure BDA0002550183230000081
c. Computing covariance matrices for normalized samples
Figure BDA0002550183230000082
And its eigenvalue lambda and eigenvector W, satisfy
CW＝λW；
d. Each sample xiConversion into a new sample yi＝WTxiTo obtain a new sample matrix Y ═ (Y)1，y2，...，yn)；
e. Calculating the contribution of variance
Figure BDA0002550183230000083
Wherein λ isiSelecting the first k principal components for the ith element in the characteristic value lambda, ensuring that the accumulated contribution rate of the first k principal components reaches a set value or above, representing the original n-dimensional characteristic by using the k-dimensional characteristic to realize data dimension reduction, and obtaining a matrix after dimension reduction as Y (Y is equal to 1，y2，...，yk)。
(3) OCSVM data training
Establishing a quadratic programming problem about the spectrum occupation condition, establishing a Lagrange dual decomposition problem based on a Gaussian kernel function, and solving an optimal classification hyperplane to obtain a decision function of spectrum classification. As shown in fig. 3, the specific process is as follows:
a. establishing a quadratic programming problem with respect to spectrum occupancy
Figure BDA0002550183230000084
Wherein, the reduced sample y1，…，ymE.g. Y, m is the number of samples; phi: y → H represents the mapping of the original space to the feature space; omega and rho are respectively a normal vector and an offset of a required hyperplane in a feature space; v is an element (0, 1)]To balance the parameter, xiiIs a relaxation variable, representing the degree to which the training samples are allowed to be misclassified, R represents a real number;
b. establishing a Lagrangian function
Figure BDA0002550183230000091
Wherein alpha isi，βiIs a Lagrange factor;
c. for omega, rho and xiiRespectively calculating partial derivatives to obtain
Figure BDA0002550183230000092
d. Construction of the Gaussian Kernel function K (y)i，yj)＝＜Φ(yi)，Φ(yj)＞＝exp(-g||yi-yj||2) (ii) a Wherein < phi (y)i)，Φ(yj) > represents the calculation of phi (y)i) And Φ (y)j) Inner product of, yi，yjRespectively representing ith and j samples after dimensionality reduction, wherein i and j are 1 and 2.
e. Solving dual decomposition problem
Figure BDA0002550183230000093
Choose any one to satisfy
Figure BDA0002550183230000094
Alpha of (A)*Calculating the offset
Figure BDA0002550183230000095
To obtain alphai *Corresponding vector yiIs a support vector and obtains a decision function of spectrum classification as
Figure BDA0002550183230000096
Where y represents the input sample and f (y) represents the classification decision function.
(4) PSO parameter optimization
In the OCSVM training process, the inherent parameters and Gaussian kernel function parameters of the OCSVM are iteratively optimized by adopting the PSO. As shown in fig. 4, the specific process is as follows:
a. initializing a maximum iteration number M; randomly initializing particle swarm U ═ U1，u2，...，uPWherein the velocity of the p-th particle is Vp＝(Vp1，Vp2，...，Vpk)TAt the position Sp＝(Sp1，Sp2，...，Spk)TFor pointing to the positions of k features, k being the number of features reduced in dimension, position Spk＝(Spv，Spg) Is composed of two components, which respectively represent an OCSVM intrinsic parameter v and a Gaussian kernel function parameter g, and the limited ranges are [ Svmin，Svmax]And [ S ]gmin，Sgmax]；Svmin，SvmaxMinimum and maximum values representing the intrinsic parameter v; sgmin，SgmaxRepresenting the minimum and maximum values of the gaussian kernel function parameter g;
b. based on the classification decision function f (y), dividing SpSubstituting into a classification decision function, calculating the fitness value f (S) of the current particlep)；
c. Updating individual extremum and group extremum:
if the fitness value is
Figure BDA0002550183230000101
Then
Figure BDA0002550183230000102
Otherwise
Figure BDA0002550183230000103
Where h denotes the h-th iteration, ShRepresents the current position of the particle at the h-th iteration, LhRepresenting the position of the individual extreme point at the h iteration;
calculation of Gh＝max(Lh)，GhRepresenting the position of a global extreme point of the whole particle swarm during the h iteration;
d. updating the velocity and position of particles
Vh+1＝ωpsoVh+c1r1(Ph-Sh)+c2r2(Gh-Sh)
Sh+1＝Sh+Vh+1
Wherein, VhRepresenting the velocity of the particle at the h-th iteration, c1、c2To be an acceleration factor, r 1，r2Is [0, 1 ]]Random numbers, omega, distributed uniformly withinpsoIs the inertial weight;
e. judging whether a termination condition is met:
if the current iteration times exceed the maximum iteration times M or the change of the fitness function value for N continuous times does not exceed the threshold value sigma, exiting the iteration, and the position of the particle corresponding to the group extreme value at the moment is the optimal parameter; and if the exit condition is not met, returning to the step a.
(5) Spectrum cognition
Based on an industrial spectrum classification model obtained by training, firstly, PCA (principal component analysis) dimensionality reduction processing is carried out on a spectrum signal to be detected to obtain dimensionality reduction data Z, then the dimensionality reduction data Z is input into a PSO-OCSVM (particle swarm optimization-online statistics virtual machine) spectrum cognition classification model for optimizing learning, and spectrum occupation conditions are judged according to a decision function.
If the acquired and trained data is large spectrum data in a busy state, if f (Z) > 0, judging that the spectrum is busy and occupied; and when f (Z) < 0, judging that the frequency spectrum is idle and unoccupied.
If the acquired and trained data is large spectrum data in an idle state, judging that the spectrum is idle and unoccupied when f (Z) > 0; and when f (Z) < 0, the frequency spectrum is judged to be busy and occupied.
Through the process, the method can support high-efficiency and quick processing of mass spectrum data, and realize accurate cognition of real-time spectrum. Through the dimensionality reduction processing of PCA, the key characteristic vector is extracted, the irrelevant vector can be greatly compressed, the data calculation amount is reduced, the frequency spectrum cognition speed and accuracy are improved, and the energy consumption is reduced. On the basis, the OCSVM can be trained by only acquiring one sample data in the condition that the frequency spectrum is free or busy without acquiring data of various conditions, so that the calculated amount can be reduced, the training time can be shortened, the training efficiency can be improved, and the OCSVM has high recognition accuracy and strong robustness. Meanwhile, parameters of the OCSVM spectrum sensing classification model are optimized by adopting the PSO, and the particles are updated iteratively, so that the local optimization of the parameters can be avoided, and the accuracy of classification model training can be greatly improved.

Claims (8)

1. A machine learning-based efficient and rapid industrial spectrum cognition method is characterized by comprising the following steps:
(1) industrial spectrum big data acquisition: acquiring electromagnetic spectrum data in an industrial environment, and establishing a binary classification model of a frequency spectrum according to whether the frequency spectrum is occupied;
(2) PCA dimension reduction treatment: carrying out dimensionality reduction on the acquired electromagnetic spectrum data by using a PCA algorithm, and extracting a characteristic vector;
(3) OCSVM data training: for the feature vector after dimensionality reduction, establishing a Lagrange dual decomposition problem based on a Gaussian kernel function by solving a quadratic programming problem about the spectrum occupation condition, and solving an optimal classification hyperplane to obtain a decision function of spectrum classification;
(4) PSO parameter optimization: carrying out iterative optimization on intrinsic parameters and Gaussian kernel function parameters in an OCSVM data training process by adopting PSO (particle swarm optimization), and obtaining an optimized classification decision function as a spectrum cognitive model;
(5) and (3) frequency spectrum cognition: and carrying out PCA (principal component analysis) dimensionality reduction on the frequency spectrum signal to be detected, and then inputting the frequency spectrum signal to be detected into a frequency spectrum cognition classification model to obtain the frequency spectrum occupation condition.
2. The efficient and fast industrial spectrum cognition method based on machine learning according to the claim 1 is characterized in that the electromagnetic spectrum data in the industrial environment is collected under one of the following two conditions:
An idle state: in the industrial frequency spectrum space to be monitored, the industrial wireless system is in a halt state;
busy state: and in the industrial frequency spectrum space to be monitored, the industrial wireless system operates according to industrial requirements.
3. The efficient and fast industrial spectrum cognition method based on machine learning according to claim 1 is characterized in that the spectrum classification bivariate model is as follows:
Figure FDA0002550183220000011
where x (t) represents a wideband signal received by the transceiver, v (t) represents a noise signal, s (t) represents a signal of another industrial wireless system, H (t) represents a channel gain, and H (t) represents a channel gain0The representative spectrum is free and not occupied by other industrial wireless systems; h1The representative frequency spectrum is busy and is occupied by other industrial wireless systems in the same frequency band.
4. The efficient and fast industrial spectrum cognition method based on machine learning according to the claim 1, characterized in that the PCA dimensionality reduction process comprises the following steps:
a. establishing an original data matrix of the industrial frequency spectrum according to the number m of the samples and the characteristic number n of the samples
Figure FDA0002550183220000021
Wherein x isi＝(x1i,x2i,…,xmi)TI is 1,2, …, n represents industrial spectrum sample characteristics;
b. standardizing the sample, namely performing zero equalization processing on each column of X to obtain
Figure FDA0002550183220000022
c. Computing covariance matrices for normalized samples
Figure FDA0002550183220000023
And its eigenvalue λ and eigenvector W, satisfying CW ═ λ W;
d. each sample xiConversion into a new sample yi＝WTxiTo obtain a new sample matrix Y ═ (Y)1,y2,...,yn)；
e. Calculating the contribution of variance
Figure FDA0002550183220000024
Wherein λ isiSelecting the first k principal components for the ith element in the characteristic value lambda, ensuring that the accumulated contribution rate of the first k principal components reaches a set value or above, representing the original n-dimensional characteristic by using the k-dimensional characteristic to realize data dimension reduction, and obtaining a matrix after dimension reduction as Y (Y is equal to1,y2,...,yk)。
5. The method of claim 1, wherein the industrial spectrum sample characteristics include signal power, time, angle of arrival, time of arrival, synchronization signal, packet size, source address, destination address, forwarding address, and port number.
6. The method for efficient and fast industrial spectrum cognition based on machine learning according to claim 1, wherein the OCSVM data training comprises the following steps:
a. establishing a quadratic programming problem with respect to spectrum occupancy
Figure FDA0002550183220000031
Wherein, the reduced sample y1,…,ymE.g. Y, m is the number of samples; y → H represents the mapping from the original space to the feature space; omega and rho are respectively a normal vector and an offset of a required hyperplane in a feature space; ν epsilon (0, 1) ]To balance the parameter, xiiIs a relaxation variable, representing the degree to which the training samples are allowed to be misclassified, R represents a real number;
b. establishing a Lagrangian function
Figure FDA0002550183220000032
Wherein alpha isi，βiIs a Lagrange factor; c. for omega, rho and xiiRespectively calculating partial derivatives to obtain
Figure FDA0002550183220000033
d. Construction of the Gaussian Kernel function K (y)i,yj)＝＜Φ(yi),Φ(yj)＞＝exp(-g||yi-yj||2) (ii) a Wherein < phi (y)i),Φ(yj) > represents the calculation of phi (y)i) And Φ (y)j) Inner product of, yi,yjRespectively representing ith and j samples after dimensionality reduction, wherein i and j are 1 and 2 … m;
e. solving dual decomposition problem
Figure FDA0002550183220000034
Choose any one to satisfy
Figure FDA0002550183220000035
Parameter α of as parameter α*Calculating the offset
Figure FDA0002550183220000038
To obtain alphai *Corresponding vector yiIs a support vector and obtains a decision function of spectrum classification as
Figure FDA0002550183220000037
Where y represents the input sample, f (y) represents the classification decision function, and sgn represents the sign function.
7. The efficient and fast industrial spectrum cognition method based on machine learning according to claim 1, wherein the PSO parameter optimization comprises the following steps:
a. initializing a maximum iteration number M; randomly initializing particle swarm U ═ U1,u2,...,uPP denotes the total number of particles, where the speed of the P-th particle is Vp＝(Vp1,Vp2,...,Vpk)TAt the position Sp＝(Sp1,Sp2,...,Spk)TFor pointing to the positions of k features, k being the number of features reduced in dimension, position Spk＝(Spv,Spg) Is composed of two components respectively representing the inherent parameter v and the Gaussian kernel function parameter g of the OCSVM, and the limited ranges are [ S vmin,Svmax]And [ S ]gmin,Sgmax]；Svmin,SvmaxMinimum and maximum values representing the intrinsic parameter v; sgmin,SgmaxRepresenting the minimum and maximum values of the gaussian kernel function parameter g;
b. calculating the fitness value f (S) of the current particle based on the classification decision function f (y)p)；
c. Updating individual extremum and group extremum:
if the fitness value is
Figure FDA0002550183220000041
Then
Figure FDA0002550183220000042
Otherwise
Figure FDA0002550183220000043
Where h denotes the h-th iteration,
Figure FDA0002550183220000044
representing the current position of the particle at the h-th iteration,
Figure FDA0002550183220000045
representing the position of the individual extreme point at the h iteration;
computing
Figure FDA0002550183220000046
GhExpressing the position of a global extreme point of the whole particle swarm in the h iteration, namely a swarm extreme value;
d. updating the velocity and position of particles
Figure FDA0002550183220000047
Figure FDA0002550183220000048
Wherein the content of the first and second substances,
Figure FDA0002550183220000049
represents the velocity of the particle p at the h-th iteration, c1、c2To be an acceleration factor, r1，r2Is [0,1 ]]Random numbers, omega, distributed uniformly withinpsoIs the inertial weight;
e. judging whether a termination condition is met:
if the current iteration times exceed the maximum iteration times M or the change of the fitness function value for N continuous times does not exceed the threshold value sigma, exiting the iteration, and the particle position corresponding to the group extreme value at the moment is the optimal parameter, namely the inherent parameter v and the Gaussian kernel function parameter g; and if the exit condition is not met, returning to the step a.
8. The efficient and rapid industrial spectrum cognition method based on machine learning according to claim 1 is characterized in that the obtained spectrum occupation condition specifically comprises: inputting the frequency spectrum dimensionality reduction data Z to be detected into a frequency spectrum cognition classification model to obtain f (Z), wherein the specific judgment method comprises the following steps:
If the data collected in the step (1) is the spectrum data in a busy state, judging that the spectrum is busy and occupied when f (Z) > 0; when f (Z) < 0, judging that the frequency spectrum is idle and unoccupied;
if the data collected in the step (1) is spectrum data in an idle state, judging that the spectrum is idle and unoccupied when f (Z) > 0; and when f (Z) < 0, the frequency spectrum is judged to be busy and occupied.
CN202010577278.0A 2020-06-22 2020-06-22 Machine learning-based efficient and rapid industrial spectrum cognition method Pending CN111860602A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010577278.0A CN111860602A (en) 2020-06-22 2020-06-22 Machine learning-based efficient and rapid industrial spectrum cognition method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010577278.0A CN111860602A (en) 2020-06-22 2020-06-22 Machine learning-based efficient and rapid industrial spectrum cognition method

Publications (1)

Publication Number Publication Date
CN111860602A true CN111860602A (en) 2020-10-30

Family

ID=72988042

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010577278.0A Pending CN111860602A (en) 2020-06-22 2020-06-22 Machine learning-based efficient and rapid industrial spectrum cognition method

Country Status (1)

Country Link
CN (1) CN111860602A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112422213A (en) * 2020-11-11 2021-02-26 杭州电子科技大学 Efficient spectrum sensing method based on support vector machine

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105704103A (en) * 2014-11-26 2016-06-22 中国科学院沈阳自动化研究所 Modbus TCP communication behavior abnormity detection method based on OCSVM double-contour model
CN105703963A (en) * 2014-11-26 2016-06-22 中国科学院沈阳自动化研究所 PSO-OCSVM based industrial control system communication behavior anomaly detection method
CN110830124A (en) * 2019-11-21 2020-02-21 长春理工大学 Spectrum sensing method based on quantum particle swarm optimization extreme learning machine

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105704103A (en) * 2014-11-26 2016-06-22 中国科学院沈阳自动化研究所 Modbus TCP communication behavior abnormity detection method based on OCSVM double-contour model
CN105703963A (en) * 2014-11-26 2016-06-22 中国科学院沈阳自动化研究所 PSO-OCSVM based industrial control system communication behavior anomaly detection method
US20170329314A1 (en) * 2014-11-26 2017-11-16 Shenyang Institute Of Automation, Chinese Academy Of Sciences Modbus tcp communication behaviour anomaly detection method based on ocsvm dual-outline model
CN110830124A (en) * 2019-11-21 2020-02-21 长春理工大学 Spectrum sensing method based on quantum particle swarm optimization extreme learning machine

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
李琳: "基于OCSVM 的工业控制系统入侵检测算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》, pages 21 - 54 *
翟旭平;杨兵兵;孟田;: "基于PCA和混合核函数QPSO_SVM频谱感知算法", 电子测量技术, no. 09, pages 87 - 90 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112422213A (en) * 2020-11-11 2021-02-26 杭州电子科技大学 Efficient spectrum sensing method based on support vector machine
CN112422213B (en) * 2020-11-11 2022-06-14 杭州电子科技大学 Efficient spectrum sensing method based on support vector machine

Similar Documents

Publication Publication Date Title
CN109547133B (en) SVM high-efficiency spectrum sensing method based on Cholesky decomposition sampling covariance matrix
CN112036433B (en) CNN-based Wi-Move behavior sensing method
CN109525994B (en) High-energy-efficiency spectrum sensing method based on support vector machine
CN110120926B (en) Communication signal modulation mode identification method based on evolution BP neural network
CN109379153B (en) Spectrum sensing method
Li et al. Automatic modulation classification based on bispectrum and CNN
Zhang et al. Signal detection and classification in shared spectrum: A deep learning approach
CN110830124A (en) Spectrum sensing method based on quantum particle swarm optimization extreme learning machine
CN112036239B (en) Radar signal working mode identification method and system based on deep learning network
CN110932807B (en) Spectrum sensing method and system for MIMO (multiple input multiple output) system under non-Gaussian noise
Shen et al. Node identification in wireless network based on convolutional neural network
CN112949820A (en) Cognitive anti-interference target detection method based on generation of countermeasure network
Gong et al. Multi-task based deep learning approach for open-set wireless signal identification in ISM band
CN111860602A (en) Machine learning-based efficient and rapid industrial spectrum cognition method
Cao et al. Improving prediction accuracy in LSTM network model for aircraft testing flight data
CN108631817B (en) Method for predicting frequency hopping signal frequency band based on time-frequency analysis and radial neural network
Geng et al. Spectrum sensing for cognitive radio based on feature extraction and deep learning
CN114548170A (en) Radar signal sorting method based on GPU (graphics processing Unit) end fuzzy C-means clustering algorithm
CN113343801A (en) Automatic wireless signal modulation and identification method based on lightweight convolutional neural network
CN116865884A (en) Broadband spectrum sensing method based on online learning
West et al. A wideband signal recognition dataset
Li et al. TFF_aDCNN: a pre-trained base model for intelligent wideband spectrum sensing
CN114337881B (en) Wireless spectrum intelligent sensing method based on multi-unmanned aerial vehicle distribution and LMS
CN115345207A (en) Self-adaptive multi-meteorological-element prediction method
CN113378718A (en) Action identification method based on generation of countermeasure network in WiFi environment

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