CN110311744A - A Channel Environment Adaptive Spectrum Sensing Method Based on Catboost Algorithm - Google Patents

Abstract

The invention discloses a channel environment adaptive spectrum sensing method based on the Catboost algorithm. The specific steps are: 1. The secondary user collects the energy value in the current channel environment and sends it to a secondary user as a fusion center; 2. The primary user is intermittent 3. The fusion center constructs the information sent by the primary user and the secondary user into a data set, and further constructs a feature vector set; 4. The fusion center uses the Catboost algorithm to train the model; 5. The secondary user continues to send the energy value to the fusion center as a test vector and input it into the trained model; 6. After the fusion center gets the result, it sends whether the channel resource is available to all secondary users, and the secondary user responds according to the judgment of the fusion center ; When the present invention satisfies the false alarm rate of 0.1, the detection rate is increased by 10% compared with SVM, and the misclassification rate and misclassification risk are also significantly reduced.

Description

A Channel Environment Adaptive Spectrum Sensing Method Based on Catboost Algorithm

technical field

The invention belongs to the technical field of wireless communication and artificial intelligence. Specifically, it relates to a channel environment adaptive spectrum sensing method based on a Catboost algorithm.

Background technique

A wireless sensor network is a wireless self-organizing and data-centric network, which is composed of a large number of micro-sensing nodes with computing and communication capabilities. Due to the characteristics of low overhead and low power consumption of wireless sensor network, it has been applied in fields such as industry and agriculture. However, with the development of wireless communication technology, the scarcity of spectrum resources has become the biggest challenge for wireless sensor networks. Now the main wireless sensor network frequency band is 2.4GHz, in the industrial field such as wireless HART WIA-PA and ISA100.11 is based on the IEEE 802.15.4 of the physical layer. Widely used short-range communication technologies like ZigBee, Bluetooth and wifi all work in this frequency band. Therefore, the heavy use of this frequency band has caused channel congestion and inevitable interference. Therefore, cognitive wireless network technology is proposed to solve the problem of scarcity of spectrum resources. Cognitive wireless network technology is an intelligent wireless communication technology that can perceive and learn from its surrounding electromagnetic environment, and then make adjustments to its own working parameter status in response to changes in the electromagnetic environment. Spectrum sensing is the basic way to detect primary user signals. OPAwe et al. proposed a spectrum sensing method using the SVM algorithm for eigenvalue estimation in the sample covariance matrix, where the sensing user equipment is a multi-antenna device. The authors also propose a method for spectrum sensing using Kalman filter channel estimation in slow fading channels. There have been a lot of research and work on this problem. W. Zhang and others have focused on solving the problems of multipath fading, shadow fading and hidden terminals. The main methods include different cooperative spectrum sensing methods and optimized sensing methods. In addition, Umebayashi et al. proposed to use an optimized threshold setting method to improve detection performance, but it requires prior information of the previous state. B.L.Mark et al. mainly use different evaluation algorithms to estimate the position of the primary user transmitter and then determine the spatial position relationship between the primary user and the secondary user and then improve the utilization rate of the spectrum resource space. C.Liu et al proposed to transform the primary user signal detection method, that is, to use the central symmetry feature to improve the space utilization. In recent years, there has gradually been an idea of using machine learning algorithms to solve the problem of spectrum sensing in cognitive wireless networks, and some research results have been made. For example, using the SVM algorithm based on the energy detection method has better performance on spectrum sensing classification problems than other machine learning algorithms, and it is very popular and practical due to its extremely high classification accuracy. However, these works have improved the detection rate of spectrum sensing to a certain extent, but there is still room for improvement. First, the detection rate can continue to increase, and secondly, the misclassification rate and misclassification risk rate also need to continue to increase. The problem addressed by the present invention is carried out under such background.

Contents of the invention

In order to solve the above problems, the present invention provides a channel environment adaptive spectrum sensing method based on Catboost algorithm, which specifically includes the following steps:

Step 1: The front-end energy collection device of the secondary user collects the energy value in the current channel environment, and sends the energy value in the sensing period to a secondary user as a fusion center. Here, the energy detection method is used to calculate the channel environment. The principle is that when the primary user goes online or offline, the energy values collected by the front-end energy collection equipment of the secondary user will have different statistical characteristics, and the energy detection method does not need a priori Information does not require special receiving equipment for each signal like the matched filter method, nor does it require a large amount of calculation like the cyclic spectrum technology.

Step 2: The primary user intermittently sends channel resource occupancy to the fusion center.

Step 3: The fusion center constructs the information sent by the primary user and the secondary user into a data set, and further constructs a feature vector set. Using a machine learning algorithm such as Catboost and an energy detection method as the basic perception method, it is natural to think of using the energy value collected by the secondary user as a feature.

Step 4: The Fusion Center trains the model with the Catboost algorithm.

Step 5: The secondary user continues to send the energy value to the fusion center as a test vector and input it into the trained model.

Step 6: After obtaining the result, the fusion center sends whether channel resources are available to all secondary users, and the secondary users respond according to the judgment of the fusion center.

In the above step 1, both the secondary user and the primary user equipment include signal transmitters and receivers, and the secondary user also includes front-end sensing equipment. In the present invention, the focus is on judging whether the primary user transmits a signal in the spectrum sensing stage, that is, the model is simplified to the perception of the signal transmitted by the primary user's transmitter by the sensing device of the secondary user. The present invention uses more realistic unbalanced samples with a ratio of 7:1; while the positive and negative samples in the previous spectrum sensing work based on machine learning algorithms are relatively balanced.

The above step 3 is specifically:

3.1: Data setting and energy normalization:

Using the energy detection method as the basic means of spectrum sensing, the system contains P primary users, denoted as p=1,2,...P and Q secondary users q=1,2,...Q, in order to consider more A general model, so P is at least greater than or equal to 2, and Q can choose a larger number. Assuming that the primary user and the secondary user share frequency band resources without causing interference, there are only two working states of the primary user in the system: online S _p =1 or offline S _p =0; when the primary user goes online, it occupies spectrum resources, The secondary user cannot use it; when the primary user goes offline, it releases the spectrum resource, and the secondary user can use the spectrum resource at this time; as long as there is a primary user occupying the spectrum resource in the system, it is considered as the secondary user and is not allowed to use the spectrum resource; use g _p Represents the geographic location of the PU, and g _q represents the geographic location of the SU.

The energy detector of each SU samples wτ complex baseband signal samples within the sensing time period τ, and the bandwidth is denoted as w; R _q (i) represents the i-th signal sample received by the SU, expressed by the following formula:

Here H ₀ means that there is no PU in the channel, so what the SU perceives the device to receive is only thermal noise, denoted by N _q (i); H ₁ represents the situation when at least one PU is online, W _p (i) represents the emission of PU Signal samples, h _p,q represent the channel gain between PUp and SUq, and S _p is the working state of the PU. Furthermore, SU should make correct decisions within the perception time period. The invention is an energy detection spectrum sensing method based on a machine learning algorithm. In machine learning, feature extraction is the first step. Denote by Y _q the normalized energy level received by SUq:

Here η is the noise power spectral density defined as η=E[|N _q (i)| ² ]; thus, the energy vector contains the energy levels received by all SUs:

Y＝(Y ₁ ,...,Y _Q ) ^T (3)

3.2 After obtaining the energy vector, further analyze its distribution;

Because of the working mode of the PU, each energy value Y _q obeys a non-central chi-square distribution, and the degrees of freedom and non-central parameters are as follows:

r=2wτ (4)

here is the fixed transmission power of the PUp, l _p,q = |h _p,q | ² is the power attenuation, and its calculation formula is as follows:

l _p,q ＝PL(||g _q -g _p ||).ν _p,q ψ _p,q (6)

Here ||.|| represents Euclidean distance, PL(dist)=dist ^-θ represents the path loss about distance dist and loss coefficient θ; ν _p,q and ψ _p,q represent multipath fading and shadow fading respectively; PU and SU meet the 802.22 protocol; in addition, the fading coefficients ν _p,q and ν _p,q remain quasi-static during the perception period, which is 1;

The distribution of the energy level has been described above. When there are enough samples such as wτ, the energy value distribution basically obeys the Gaussian distribution; therefore, the energy vector can be extracted from the multivariate Gaussian distribution, and its mean and variance are as follows:

μ _Yq ＝r+ζ _q (7)

σ ² _Yq ＝2(r+2ζ _q ) (8)

So the mean vector and covariance matrix of the energy vector are as follows:

μ _Yq ＝(μ _Y1 ,...,μ _YQ ) ^T (9)

The above-mentioned Catboost algorithm is specifically as follows: in the sensing period, the secondary user sends the perceived energy value in the channel to the fusion center as a feature energy vector, and the primary user intermittently sends information about whether spectrum resources are occupied or not to the fusion center as a label, thus completing construction of the training data set. The Catboost algorithm is used to train the model in the fusion center. Now introduce the Catboost algorithm: The Catboost algorithm is proposed by Yandex. This algorithm optimizes the processing of category features and is processed in the training phase instead of the data preprocessing phase. Another advantage of this algorithm is that A new method is used to calculate leaf node values when selecting a tree model, which helps reduce overfitting. The Catboost algorithm has two running methods: CPU and GPU. The GPU method is faster than the most popular Xgboost GPU method at present, and the CPU method is also the same. The Catboost algorithm, like the gradient boosting method, also builds a new tree to fit the residual of the current model. However, traditional gradient boosting methods are affected by biased point gradient estimates, which can easily lead to overfitting. The gradient is evaluated each time using the same data points so that the model is built. This will cause the feature space distribution of the gradient to be fitted to be shifted compared to the true gradient distribution space, which will lead to overfitting. Many GBDT methods (such as Xgboost) to build the next tree mainly include two steps: selecting the tree structure and setting the value of the leaf. In order to choose the optimal number structure, the algorithm will perform different split enumerations, build trees, set leaf values, score and choose splitting methods. In both stages leaf values are computed to fit the gradient. Catboost is the same as the traditional method in the second stage, but the improved method is used in the first stage. Use F ^k to represent the first k-tree constructed, and g ^k (X _h , Z _h ) to represent the model constructed by the gradient value of the k-tree constructed in the hth training sample. In order to make the gradient g ^k (X _h , Z _h ) unbiased, we need to train the model F ^k without X _h , which makes the training process impossible. Consider the following trick to solve this problem: For each realization of X _h , we Train a model M _h and update it without gradient estimation. Use M _h to fit the gradient on the basis of X _h , and use this method to get the score. Catboost generates s random permutations in the training set. It uses a variety of permutations to obtain the gradient of the residual to enhance the robustness of the algorithm, uses different permutations to train the model, and then uses multiple permutations to avoid overfitting. For each permutation σ, n models M _k are trained. This means that when building a new tree, you need to store and recalculate O(n ² ) to fit the permutation σ. For each model M _k , you need to update M _k (X ₁ ),...,M _k (X _k ) , so the computational complexity is O(sn ² ), an important trick is used in the implementation process to reduce the time complexity of each tree construction to O(sn): for each permutation, instead of storing and updating time-complex The value M _k (X _j ) of degree O(n ² ), but keeps the value M _k '(X _j ), k=1,...,[log ₂ (n)], j<2 ^k+1 , Here M _k (X _j ) is a fitted approximation of sample j based on the first ^2k samples. Then, the predicted M _k (X _j ) will be lower than The gradient on _Xh is used to select the tree structure.

Compared with prior art, the beneficial effect of the present invention:

When the present invention uses a false alarm rate of 0.1 that meets the requirements of IEEE 802.11, the detection rate is increased by 10% compared with SVM, the overall classification performance is better than SVM, and the misclassification rate and misclassification risk are also significantly reduced. Knowing that the transmission power of the primary user is different, that is, the performance of the present invention is basically stronger than that of the SVM algorithm under different signal-to-noise ratios for the secondary user, that is, the machine learning model under different channel environments can be retrained to adapt to the current channel, which has stronger practicability, even In the case of low signal-to-noise ratio, the performance is also better, and the robustness is stronger. It is of great significance for spectrum sensing applications.

Description of drawings

Fig. 1 is a step diagram of the spectrum sensing method based on machine learning in the present invention.

Figure 2 is a structural model of the 7*7 cooperative spectrum sensing system.

Figure 3 is the ROC curve of the Catboost algorithm and the SVM algorithm (linear kernel function) in the 7*7 model.

Detailed ways

The implementation of the present invention will be further described below in conjunction with the accompanying drawings.

The channel environment adaptive spectrum sensing method based on the Catboost algorithm in the present invention has a system structure frame diagram as shown in Fig. 1 .

The cooperative spectrum sensing model based on geographic location in the present invention will be described below in conjunction with accompanying drawing 2, and the ROC curve, detection rate, misclassification risk and Performance on misclassification rate. The simulation is carried out in a 64-bit PC with python3.6.2, memory RAM 16G, six-core i7 (3.2GHz) environment.

The performance index that the present invention compares is as follows:

a) ROC curve (Receiver operating characteristic curve), the result is the average curve after running the experiment 200 times independently, this indicator reflects the overall classification performance of the algorithm.

b) Detection Probability, the probability that an unauthorized user is successfully detected when an authorized user goes online, and the result is the average detection rate of 200 experiments.

c) Misclassification Risk, when the authorized user goes online, the classifier gives the probability that the authorized user goes offline, and the result is the average misclassification risk rate of 200 experimental runs.

d) Misclassification Error Rate (Misclassification Error Rate), the probability of misjudgment by the classifier, that is, the probability of judging the authorized user offline as online and the authorized user online as offline. The result is the average misclassification rate of 200 experiments. .

Example

In order to verify the usability and feasibility of the present invention based on the Catboost algorithm to solve the spectrum sensing problem under the cognitive wireless network, a simulation experiment is carried out and the algorithm performance is compared with the SVM algorithm. The simulation parameters are set as follows: the perception period τ is 100μs, the bandwidth is 5MHz, the noise power spectral density is -174dBm, the transmit power of each PU is 200mW, the path loss coefficient is 4, and the multipath fading and shadow fading coefficients are both 1. The probability of a PU going online is 0.5. The kernel function of SVM is selected as the linear kernel function, because the excellent performance of the linear kernel function in this problem has been proved in the previous work. There are 160 training vectors and 640 testing vectors. The ratio of positive and negative samples is 7:1. In the structural model of the 7*7 cooperative spectrum sensing system in Figure 2, there are 49 SUs, evenly distributed in the 7*7 grid, and 3 PUs are respectively in (-1100m ,-1000m), (750m, 890m), (1500m, -1000m) positions. As can be seen from the ROC curve in Figure 3, the SVM algorithm is marked with a solid line, and the Catboost algorithm is marked with a dotted line. When the false alarm rate is 0.1, the detection rate of Catboost is 10% higher than that of SVM, and the overall classification performance is better than SVMs.

Compared with the SVM algorithm under different transmission powers of the PU, the present invention has a higher detection rate, misclassification risk and misclassification rate as shown in Table 1:

Table 1 Performance indicators of Catboost algorithm and SVM algorithm in the case of different primary user transmit power in 7*7 model

As the signal-to-noise ratio increases, both algorithm indicators will increase, and Catboost will converge when the signal-to-noise ratio is high. All these prove the feasibility and usability of the present invention, and the present invention can be used to solve the spectrum sensing problem under the cognitive wireless network.

Claims (4)

1. a kind of channel circumstance adaptive spectrum cognitive method based on Catboost algorithm, which is characterized in that including following step It is rapid:

Step 1: the energy value in secondary user front end energy acquisition equipment acquisition present channel environment, and the energy in the period will be perceived Magnitude is dealt into one user as fusion center；

Step 2: primary user sends fusion center for occupied channel resource situation by phased manner；

Step 3: the information structuring that fusion center sends primary user and time user is at data set, and further construction feature vector Collection；

Step 4: fusion center Catboost algorithm training pattern；

Step 5: secondary user continues to send fusion center for energy value, as test vector and inputs into training pattern；

Step 6: fusion center obtain after result will whether available channel resources are sent to all secondary users, secondary user is according to fusion The judgement at center is made a response.

2. a kind of channel circumstance adaptive spectrum cognitive method based on Catboost algorithm according to claim 1, It is characterized in that, non-equilibrium sample, ratio 7:1 is used in the step 1；And frequency spectrum is carried out based on machine learning algorithm before Positive negative sample is balanced in the work of perception.

3. a kind of channel circumstance adaptive spectrum cognitive method based on Catboost algorithm according to claim 1, It is characterized in that, the step 3 specifically:

3.1: data setting and energy normalized:

Use energy measuring method as frequency spectrum perception basic means, includes in systems P primary user, be denoted as p=1,2 ... P With Q user q=1,2 ... Q；Primary user and time user sharing band resource will not interfere simultaneously herein, be In system there are two types of the working conditions of primary user: online S_p=1 or offline S_p=0；Which occupies frequency spectrum moneys when primary user is online Source, secondary user cannot use；It discharges frequency spectrum resource when primary user is offline, and frequency spectrum resource can be used in secondary user at this time；System There is a primary user to occupy frequency spectrum resource in as long as, then regarding as time user does not allow to reuse frequency spectrum resource；Use g_pIt represents The geographical location of PU, g_qRepresent the geographical location of SU；

The energy detector of each SU samples w τ complex baseband signal sample in detecting period period tau, and bandwidth is expressed as w；R_q (i) i-th of sample of signal that SU is received is represented, is indicated with following formula:

H herein₀Representing does not have PU in channel, so the only thermal noise that SU awareness apparatus receives, uses N_q(i) it indicates；H₁It represents The case where when at least one PU is online, W_p(i) the transmitting sample of signal of PUp, h are represented_p,qRepresent the channel between PUp and SUq Gain, S_pThe as working condition of PU；Use Y_qIt is horizontal to represent the normalized energy that SUq is received:

Herein η be noise power spectral density be defined as η=E [| N_q(i)|²]；Therefore, energy vectors include what all SU were received Energy level:

Y=(Y₁,...,Y_Q)^T (3)

3.2 after obtaining energy vectors, further analyze its distribution；

Because of the operating mode of PU, each energy value Y_qNon-central chi square distribution is obeyed, freedom degree and non-centrality parameter are as follows:

R=2w τ (4)

HereIt is the fixed transmission power of PUp, l_p,q=| h_p,q|²It is power attenuation, calculation formula It is as follows:

l_p,q=PL (| | g_q-g_p||).ν_p,qψ_p,q (6)

Here | | | | represent Euclidean distance, PL (dist)=dist^-θIt represents and is damaged about the path of distance dist and loss coefficient θ It loses；ν_p,qAnd ψ_p,qRespectively represent multipath fading and shadow fading；PU and SU meets 802.22 agreements；In addition, in detecting period section Interior fading coefficients ν_p,qAnd ν_p,qIt is constant to be quasi-static, as 1；

Reach w τ, energy Distribution value Gaussian distributed in sample size；Therefore energy vectors can be from multivariate Gaussian distribution It extracts, mean value and variance are as follows:

μ_Yq=r+ ζ_q (7)

σ² _Yq=2 (r+2 ζ_q) (8)

Therefore the mean vector of energy vectors and covariance matrix are as follows:

4. a kind of channel circumstance adaptive spectrum cognitive method based on Catboost algorithm according to claim 1, It is characterized in that, the Catboost algorithm training pattern specifically:

The second stage of Catboost is identical with conventional method, and the first stage has used improved method: using F^kIt represents the of building One k tree, g^k(X_h,Z_h) represent the model constructed in h-th of training sample building k tree gradient value；In order to make g^k(X_h,Z_h) nothing Partially, for realizing X each time_h, we train a model M_h, go to update without the mode that gradient is estimated；Use M_hIn X_hBasis Upper fitting gradient makes to be scored in this way；Catboost produces s random alignment in training set, uses Various arrangement sampling strengthens the robustness of algorithm to obtain the gradient of residual error, carrys out training pattern using different arrangements, then Over-fitting is avoided using various arrangement；

For each arrangement σ, n model M of training_k, store when constructing new tree and recalculate O (n²) come be fitted arrangement σ, For each model M_k, need to update M_k(X₁),...,M_k(X_k), so computation complexity is O (sn²), make during realization The time complexity of each tree building is dropped to O (sn) with following method: not being storage and update for each arrangement Time complexity O (n²) value M_k(X_j), but retention value M_k'(X_j), k=1 ..., [log₂(n)], j < 2^k+1, M here_k(X_j) It is based on first 2^kThe fitting of the sample j of sample is approximate；Then, the M of prediction_k(X_j) can be lower than In X_hOn gradient be used to select tree construction.