阅读说明：本技术 基于量子粒子群优化极限学习机的频谱感知方法 (Spectrum sensing method based on quantum particle swarm optimization extreme learning machine ) 是由 张晨洁 郭滨 王志军 李可欣 郭熠 白雪梅 耿小飞 胡汉平 于 2019-11-21 设计创作，主要内容包括：基于量子粒子群优化极限学习机的频谱感知方法，涉及认知无线电领域，解决现有无线信道环境中低信噪比情况下主用户信号检测率较低、传统极限学习机算法仅基于经验风险最小化容易过拟合、网络结构不佳等频谱感知问题，本发明包括提取信号循环谱特征和能量特征；构建训练数据集；根据获得的训练数据集，训练QPSO?ELM频谱感知模型；将所述的提取接收信号的能量特征和循环谱特征作为检测数据输入步骤三中训练好的频谱感知模型，实现对主用户信号的频谱感知，所述频谱感知模型输出为1时，则主用户存在；输出为0时，则主用户不存在。本发明方法经过量子粒子群的优化和引入结构风险使得算法能够更有效的提取输入特征，虚警概率相对较低。(A spectrum sensing method based on a quantum particle swarm optimization extreme learning machine relates to the field of cognitive radio, and solves the spectrum sensing problems that the detection rate of a master user signal is low under the condition of low signal to noise ratio in the existing wireless channel environment, the traditional extreme learning machine algorithm is easy to overfit only based on experience risk minimization, the network structure is not good, and the like; constructing a training data set; training a QPSO-ELM spectrum sensing model according to the obtained training data set; inputting the energy characteristics and the cyclic spectrum characteristics of the extracted receiving signals as detection data into a trained spectrum sensing model in the third step to realize spectrum sensing of signals of the main user, wherein when the output of the spectrum sensing model is 1, the main user exists; and when the output is 0, the master user does not exist. According to the method, the input characteristics can be more effectively extracted by the algorithm through the optimization of the quantum particle swarm and the introduction of the structural risk, and the false alarm probability is relatively low.)

1. The spectrum sensing method based on the quantum particle swarm optimization extreme learning machine is characterized by comprising the following steps: the method is realized by the following steps:

step one, extracting a signal cyclic spectrum characteristic and an energy characteristic;

presence of a Primary user (H)₁) Extracting the cyclic spectrum characteristic and the energy characteristic of the signal under the condition to form a characteristic vector y₁In the absence of primary users (H)₀) Extracting the cyclic spectrum characteristic and the energy characteristic of the signal under the condition to form a characteristic vector y₀；

Step two, constructing a training data set;

the feature vector y obtained in the step one₁And a feature vector y₀Forming a training data set;

step three, training a QPSO-ELM spectrum sensing model by adopting the training data set obtained in the step two; the specific process is as follows:

initializing a particle swarm, a weight matrix A and a bias matrix B;

step two, calculating a weight matrix of the hidden layer and the output layer according to the weight matrix A and the bias matrix B in the step one, and calculating a fitness function of each particle;

the fitness function is represented by the following formula:

where β is the structural risk, γ is a factor that trades off between structural risk and empirical risk, i_bTo imply the number of layers, j_bThe number of samples; i.e. i_b∈[1，2，...，L]，j_b∈[1，2，...，N]N is the number of nodes of the input layer, and L is the number of nodes of the hidden layer;

step three, updating the optimal individual position of each particle;

in the formula:

step three, updating the global optimal position;

in the formula, β_pgOutputting the weight for the global extremum, wherein pg (t-1) is the global optimal position of the particle in t-1 iterations, and pg (t) is the global optimal position in t iterations;

step three, judging whether an iteration termination condition is met, if not, executing the step two to the step three, and if so, obtaining an optimal QPSO-ELM model;

step four, inputting the energy characteristics and the cyclic spectrum characteristics of the extracted receiving signals in the step one as detection data into the trained spectrum sensing model in the step three to realize spectrum sensing of signals of the main user, wherein when the output of the spectrum sensing model is 1, the main user exists; and when the output is 0, the master user does not exist.

2. The spectrum sensing method based on the quantum-behaved particle swarm optimization extreme learning machine according to claim 1, wherein the spectrum sensing method comprises the following steps: in the third step, initializing a particle swarm, a weight matrix A and a bias matrix B specifically comprise the following steps:

B＝[b₁，b₂，…，b_L]_1×L

randomly generating initial particle swarms, wherein each particle consists of a group of input weights and hidden layer bias values, and the initialization range is reduced to be between-1 and 1;

3. The spectrum sensing method based on the quantum-behaved particle swarm optimization extreme learning machine according to claim 1, wherein the spectrum sensing method comprises the following steps: in the third step, the weight matrix of the hidden layer is given by the following formula:

A＝[A₁,A₂,...A_L]as input weights, x₁,x₂,...x_NInput for each neuron node respectively; the weight matrix for the connection of the hidden layer and the output layer is given by:

in the formula, H⁺The Moore-Penrose generalized inverse of the hidden layer output matrix H, T is the desired output.

4. The spectrum sensing method based on the quantum-behaved particle swarm optimization extreme learning machine according to claim 1, wherein the spectrum sensing method comprises the following steps: in the third and fourth steps, the method also comprises the step of_aFor each dimension of the particle, the local attractor is calculated according to the following formula

in the formula (I), the compound is shown in the specification,

the mean value of the optimal positions of the individual particles is expressed by the following formula:

in the formula (I), the compound is shown in the specification,

the evolutionary equation for the particle is:

in the formula (I), the compound is shown in the specification,

5. The spectrum sensing method based on the quantum-behaved particle swarm optimization extreme learning machine according to claim 1, wherein the spectrum sensing method comprises the following steps: the iteration termination condition is as follows: and when the maximum iteration times are selected and the iteration model reaches the global optimal position or the individual optimal position, the iteration is terminated.

Technical Field

The invention relates to the field of cognitive radio, in particular to a spectrum sensing algorithm based on a quantum particle swarm optimization extreme learning machine.

Background

With the development of the communication industry and the increasing requirements of people on the speed and quality of networks, radio spectrum resources are increasingly scarce, and fixed frequency bands are allocated to fixed services by various countries according to the technical characteristics of the radio services, service capabilities, broadband requirements and other factors. The utilization rate of the frequency spectrum is low, and even a busy frequency band has a lot of available idle frequency spectrum. The spectrum waste is reduced, the spectrum utilization rate is improved, and the problem to be solved urgently is solved.

The existing spectrum sensing algorithm can achieve good identification effect under high signal-to-noise ratio, but the identification performance under low signal-to-noise ratio is not ideal. The spectrum sensing can be regarded as a binary classification problem from the classification perspective, and can be regarded as a linear classification problem under the condition of high signal-to-noise ratio, and the traditional spectrum sensing algorithm can well solve the problem by setting a linear threshold. In a wireless channel with a low signal-to-noise ratio, the research direction of spectrum sensing is to solve the problem of nonlinear threshold signal classification, which is the problem of machine learning algorithm research.

Although a cooperative spectrum sensing method based on machine learning (including supervised and unsupervised machine learning) obtains better detection performance, when the noise power is higher, the robustness detection is seriously influenced by energy as characteristic input; the spectrum sensing method based on the Artificial Neural Network (ANN) provided by a researcher takes signal energy and cyclostationary features as input features, but for large-scale training data, the ANN is easy to have an overfitting problem, so that the spectrum sensing performance is reduced; researchers have also proposed a single hidden layer feed forward neural network (SLFN) algorithm: and (4) an extreme learning machine algorithm. The algorithm randomly generates the weight and the hidden layer deviation, and the learning speed is much faster than that of the traditional gradient descent algorithm, so that the global optimal solution can be obtained and the generalization capability is good. Although the extreme learning machine has better generalization capability, the algorithm accelerates the training process by randomly selecting input weights and hidden layer bias, which may result in selecting more hidden nodes and poor weights instead of the optimal network structure, increasing the complexity of the network. Conventional ELM algorithms are based only on empirical risk minimization making the algorithms relatively easy to overfit.

Disclosure of Invention

The invention provides a spectrum sensing method based on a quantum particle swarm optimization extreme learning machine, which aims to solve the problems of low detection rate of a master user signal, easy overfitting of a traditional extreme learning machine algorithm based on experience risk minimization, poor network structure and the like in the existing wireless channel environment under the condition of low signal to noise ratio.

The spectrum sensing method based on the quantum particle swarm optimization extreme learning machine is realized by the following steps:

step one, extracting a signal cyclic spectrum characteristic and an energy characteristic;

presence of a Primary user (H)₁) Extracting the cyclic spectrum characteristic and the energy characteristic of the signal under the condition to form a characteristic vector y₁In the absence of primary users (H)₀) Extracting the cyclic spectrum characteristic and the energy characteristic of the signal under the condition to form a characteristic vector y₀；

Step two, constructing a training data set;

the feature vector y obtained in the step one₁And a feature vector y₀Forming a training data set;

step three, training a QPSO-ELM spectrum sensing model according to the training data set obtained in the step two; the specific process is as follows:

initializing a particle swarm, a weight matrix A and a bias matrix B;

step two, calculating a weight matrix of the hidden layer and the output layer according to the weight matrix A and the bias matrix B in the step one, and calculating a fitness function of each particle;

the fitness function is represented by the following formula:

where β is the structural risk, γ is a factor that trades off between structural risk and empirical risk, i_bTo imply the number of layers, j_bThe number of samples; i.e. i_b∈[1，2，...，L]，j_b∈[1，2，...，N]N is the number of nodes of the input layer, and L is the number of nodes of the hidden layer;

is jth_bThe expected output value of the one sample,

to connect with the ith_bThe weight vectors of the hidden layer neurons and the output layer neurons,to connect with the ith_bThe input weights of the individual input nodes and the hidden layer node,

is the ith_bThe bias of individual neurons is the threshold of the hidden layer neurons,

to represent

And

g (-) is the activation function of the hidden layer;

step three, updating the optimal individual position of each particle;

in the formula:

are respectively the ith_aOutputting the weight of the current position of each particle and outputting the weight of the individual extreme value;

is the ith_aThe individual best positions of the particles in t iterations,

is the ith_aIndividual positions of the particles in t iterations;is the ith_aThe individual most position of each particle in t-1 iterations,

is the ith iteration in t-1_aIndividual positions of particles;

step three, updating the global optimal position;

in the formula, β_pgOutputting the weight for the global extremum, wherein pg (t-1) is the global optimal position of the particle in t-1 iterations, and pg (t) is the global optimal position in t iterations;

step three, judging whether an iteration termination condition is met, if not, executing the step two to the step three, and if so, obtaining an optimal QPSO-ELM model;

step four, inputting the energy characteristics and the cyclic spectrum characteristics of the extracted receiving signals in the step one as detection data into a trained spectrum sensing model in the step three to realize spectrum sensing of signals of a main user, wherein when the output of the spectrum sensing model is 1, the main user exists; and when the output is 0, the master user does not exist.

The invention has the beneficial effects that:

1) the method of the invention is superior to ANN and SVM methods in the correct detection probability under the conditions of different signal-to-noise ratios of-25 dB to-5 dB. The detection probability is still higher than 70% in-10 dB, and the detection performance is better under the condition of low signal-to-noise ratio;

2) the method introduces structural risk, reduces empirical risk, improves generalization performance of the algorithm, overcomes the defects that the traditional ANN algorithm is easy to fall into local optimal solution and the SVM is easy to over-fit under the condition of low signal to noise ratio to cause larger classification precision error, and improves the detection accuracy of the traditional ELM in spectrum sensing;

3) according to the method, the input characteristics can be more effectively extracted by the algorithm through the optimization of the quantum particle swarm and the introduction of the structural risk, and the false alarm probability is relatively low;

4) when the number of the neurons in the hidden layer reaches a certain value, the detection probability only floats in a small range. The traditional ELM algorithm can cause great change of detection probability when the number of neurons is different;

5) the identification precision of the method is higher than that of an ANN (artificial neural network), an SVM (support vector machine) and a traditional ELM (element-by-element) algorithm, the speed is higher than that of the traditional neural network algorithm adopting a gradient descent method, repeated forward calculation and reverse calculation errors are not needed to be corrected, and the learning efficiency is greatly improved.

Drawings

FIG. 1 is a flow chart of an extreme learning machine spectrum sensing algorithm for quantum particle swarm optimization;

FIG. 2 is a flow chart of algorithm model training;

FIG. 3 is a graph showing the detection probability of each algorithm under different signal-to-noise ratios;

FIG. 4 is a graph illustrating false alarm probability performance versus false alarm probability performance for different SNR;

FIG. 5 is a diagram of the spectrum sensing detection probability effect of the ELM algorithm under-15 dB under different hidden layer neuron numbers;

FIG. 6 is a diagram of the probability effect of spectrum sensing detection of the QPSO-ELM algorithm under-15 dB under different hidden layer neuron numbers.

Detailed Description