Machine Learning Approach
64 V. S. Kulkarni et al.
Fig. 1 Classification of spectrum sensing
techniques [1] in which the harmful interference is prevented with licensed users and the available spectrum is identified to improve the spectrum’s utilization.
In cognitive radio network one of the important techniques is Spectrum sensing and is also the primary task for establishing a cognitive radio system [1].
Spectrum Sensing is of three types. Interference-based sensing, non-cooperative sensing and Cooperative sensing.
In cognitive radio networks, machine learning methods of Cooperative spec- trum sensing is the major leading technique. The capability of decision making can be enhanced by combining machine learning with cognitive radio networks from the previous experiences [2]. Compared to traditional methods unstable cognitive spectrum resources can be used effectively.
Figure1shows the classification of Spectrum sensing. Non-cooperative sensing, interference-based sensing and Cooperative sensing Spectrum sensing are the classification of spectrum sensing.
The performance of detection of spectrum holes is enhanced by cooperative sensing in which secondary user senses the spectrum with the collaboration with primary user. Also known as receiver detection technique.
In non-cooperative spectrum sensing, individual CR scans for primary signal and based on its detection CR decides the presence or absence of primary user.
Cognitive radio measures the interference environment and their transmissions are adjusted such that the interference to PU is not above the regularity limit in interference-based sensing [3].
Cyclo-stationary feature detection, energy detection and match filter detection are the classification of non-Cooperative sensing.
One of the intelligent cooperative sensing methods is machine learning which has the special feature of self-adaptation to environment, high sensing.
Table1shows the comparative advantages and disadvantages of spectrum sensing methods [4]. The author has created this table to familiarize spectrum sensing and for better understanding the state-of-art techniques to the researchers. The performance of any spectrum sensing algorithm is determined by the probability of detection, probability of false alarm, the energy detection for cognitive radio network follows the following formula [5,13].
Machine Learning Approach in Cooperative Spectrum … 65 Table 1 Spectrum sensing methods
Sr. No Techniques of sensing Pros Cons
1 Energy detection 1. There is no requirement of pre knowledge of primary signal characteristics 2. Implementation is easy
1. Noise uncertainty is sensitive
2. False alarm rate is high
2 Cyclostationary feature detection
1. Signal and noise can be distinguished
2. Probability of false alarm decreases at low SNR
1. When the sample size is large, energy
consumption is high 3 Matched filter-based
detection
1. Detection is better at low SNR region
2. Sensing is optimal
1. Knowledge of primary user signal is required in advance, hence impractical 4 Detection based on
Covariance
It is not essential to have the knowledge of primary signal characteristics in advance
Computation is very complex
5 Spectrum sensing based on Machine learning
1. Complex model can be used in easy manner 2. Delay of detection is
minimized
1. Techniques are complex 2. Detection rate is affected
by feature selection
1. Let time of sensing beτs, at the sensing node, when the PU is absent, the signal that is received is given by
s(u)=σn(u) (1)
whereσn(u) is the noise power at duration d.
2. When the PU is present, signal is given as
s(u)=g(u).r(u)+σn(u) (2)
where g(u) is gain of the channel, r(u) is primary signal which is received.
3. Assuming some of the parameters, detection of energy is given by
ED = 1 P0
τs
0
|s(u)|2du (3)
where P0 is the spectral density of powerτs is the time of sensing.
4. Probability of detection and probability of false alarm using energy detection can be given by considering N cooperative sensing nodes. Probability of detection is given by
66 V. S. Kulkarni et al.
pj(det)(αj,βj,τs,CB)=Q
αj−(2τsCB+βjτsCB) 4τsCB+4βjτsCB
where Q(u)=
1
√2π
∞
u
exp −t2
2
(4)
where pj(det) is the probability of detection and is given by P (β≥αj| H1) at j belonging to 1 to N,αj=threshold of detection, H1-presence of PU.
Probability of false alarm is given by
pj(false)(αj,τs,CB)=Q
αj−2τsCB
√4τsCB
(5)
where pj(false)=P (β≥αj| H0) at j belonging to 1 to N, h2-absence of PU.
The organization of the paper is as follows. Literature survey of various method- ologies of cooperative spectrum sensing are described in Sect.2. Conclusion in Sect.3 is preceded by a table of comparative methodologies.
2 Literature Survey
Zan Li et al. [6] proposed CSS model of machine learning in which user grouping concept is shown. CSS framework based on optimization model which had four modules. Group scheduling module, SVM classification module, user grouping module and SVM training module are the four modules of grouping concept. Without the reduction in accuracy of sensing and the functions like grouping of optimized users, redundant and abnormal are achieved by using three algorithms of grouping.
In the algorithm of first grouping there are groups of two which are normal and abnormal. Then in second grouping algorithm redundant and non-redundant users are distinguished and finally optimized model is established. As per the requirement of cooperative sensing optimization of specific number of groups are done where the users are divided, selecting each time. Finally, the user group are non-redundant and secured that involve in the recognition of pattern. Optimization problem can be solved by Binary particle swarm optimization. By using this grouping algorithms, the CR networks can be avoided the harm of abnormal and the redundant users.
Assigning silent state for long term or temporarily for many users reduces the unnec- essary energy consumption. The efficiency of sensing can be improved by increasing the grouping. This is done by increasing the speed of classification in model of SVM in machine learning. Centralized CSS model network consisting of fusion center is used in the simulation with 10 or 20 CRs. Here the simulation result shows that there is 17.98% rate of improvement for the speed of average classification. Along with the entire network security they also showed the operational efficiency of SVM model increased using group algorithm 1. Few non-highly redundant users and highly redun- dant users are used in grouping algorithm 2 with network of higher signal quality.
Machine Learning Approach in Cooperative Spectrum … 67 SVM achievement 100% accuracy. In algorithm 3, the users of CR are divided into cooperation groups of two or more which are optimized. Cooperative sensing can be performed independently by each group. The classification speed increases with the number of groups being more.
Yingqi Lu et al. [7] proposed machine learning based classification which uses the probability vector of dimension being low. This method has the time for classifica- tion shorter and duration for training smaller. So, energy vector with N-dimensions is not used here. Keeping the same spectrum performance and by converting a high dimensional feature vector, a feature vector of constant two-dimensional is achieved.
With M secondary users and primary user selecting single SVM and K-means algo- rithms of machine learning are used. Small scale CRN uses primary user as single and M=2SUs and large-scale CRN using single primary user (PU) and M=9SUs are considered.
In Small scale CRN, the techniques of machine learning are trained and then new feature vectors are tested, which gives the result of (PFA) probability of false alarm and (PD) the probability of detection. They showed that The accuracy of detection compared to algorithms of AND rule and OR rule which are hard fusion, classification based on machine learning is better. In terms of energy vector, there is better performance by SVM based classification using probability vector. With the highest transmitting power of PU, performance is better.
In large scale CRN, through plot of Receiver operating characteristics curves accuracy is shown better in both SVM-ploy and SVM-linear when the probability vector is combined. Here energy vector is replaced by probability vector and proved the accuracy to be better.
Through simulation they proved that
1. With the equal training samples, the duration of training for K-means clustering is longest.
2. SVM-ploy takes more training time than SVM-linear as more time is spent on feature vector mapping by functions of polynomial kernel.
3. SVM-linear is more efficient than SVM-ploy.
4. With the probability vector, the performance is best of SVM linear compared to various decision-making scheme such as K-means, SVM-ploy. These deci- sions making schemes have duration of training and classification delay low, probability of detection high.
Xiaolin Ma et al. [8] proposed Extreme Learning machine (ELM) for numerous primary users to implement the algorithm of sensing cooperative spectrum with the network of cognitive radio. ELM which is an algorithm of machine learning that has three step free tuning process is used to make very accurate the channel sensing and identification which involves. In this paper initially the identification of multiple primary users is established by spectrum sensing model. A classification scheme of channel pattern can be obtained by ELM with the help of energy detection and channel model. Channel model as well as sensing model, energy model, fusion center is also implemented.
68 V. S. Kulkarni et al.
In channel model, the primary user which uses the channel detects the signal consists of an authorized signal and noise which is called Rayleigh channel. Using multiple primary users, the signals can be detected by secondary user in the network if the number of primary users is more than one. In energy model, usage of channel by primary users can be sensed by energy. The fusion center aims at determination of usage of channel. Based on the transmission of energy vectors from secondary user channel usage is determined.
Simulation is carried out with multiple primary users and the results are compared with SVM, it was shown that training time of ELM is shorter. The detection probability being higher the probability of overall detection will also be better.
SVM and ELM algorithms were proposed to train the samples with 1000 and 2000 as the sample size. It is proved that ELM needs only 92 ms to reach probability of detection compared to SVM which requires 8 times more time than ELM. Accuracy of SVM is also less than ELM.
Olusegun Peter Awe et al. [9] worked in cognitive radio network in order to solve the problem of sensing the spectrum on the condition of multiple primary users and proposed algorithm based on based multi-class support vector machine (SVM) which is ECOC (error correcting output codes). Spectrum holes is detected by Joint spatial–temporal detection which is implemented using this algorithm. To study the attributes of each state, through one-versus-all (OVA) and one-versus- one (OVO) Multi-class support vector algorithms are used [10–12]. Error correcting output codes (ECOC) scheme is used to solve the optimization problem. The multi class problem can be dealt with ECOC which is a framework that creates multiple binary classification task by breaking it [9] and [10]. Initially the condition is consid- ered as detection of multiple class signal in which more than one sub-class are formed from each class. Then multi-class SVM algorithms checks for performance of the energy-based features and ECOC (error correcting output codes). Detection of performance is mainly based on accuracy of classification and curves of receiver operating characteristics. They proved that unused spatial spectrum can be detected by using ECOC.
Karaputugala Madushan Thilina et al. [1] proposed technique of machine learning for cognitive radio in which both supervised and un-supervised methods are used.
weighted K-nearest-neighbor and support vector machine learning is used in super- vised and gaussian mixture model, K-means clustering is used in unsupervised where cooperative spectrum sensing is implemented. Initially by using the energy levels of vector the availability is checked for channel which is considered as feature vector.
The decision is taken by feeding this to the classifier. A classifier categorizes into two categories the availability of channel with available class and the unavailable class for each feature vector. The training phase has to be covered by classifier before the classification being online. Partitioning of training feature vectors into K clus- ters is done by implementing the algorithm of K-means in which test energy vector class and primary user state represented by each cluster is determined by classifier.
Mapping of each cluster determines the available and unavailable class of channel.
Training feature vectors are obtained from Gaussian mixture distributions in GMM which corresponds to a cluster. The SVM and the KNN is also proposed due to the
Machine Learning Approach in Cooperative Spectrum … 69 capability of higher prediction. In the SVM, the margin between separating hyper- planes and feature vectors are maximized to obtain support vectors. Performance is determined by evaluating all the classification techniques such as the ROC curve, the classification delay and training time.
The result shows that
1. Capability of classification type K-nearest neighbor for 1000 energy vectors is shown by the time taken for uploading the training energy vector to the classifier as 50 μsas compared to GMM having high training duration of 12,796 s for 1000 samples and SVM having highest training duration of 1. 65,817 s for 1000 samples
2. KNN classifier has high classification delay even though training time is lowest compared to Fisher linear discriminant which shows lowest classification delay.
Also, no change is observed in classification delay in GMM, K-means clustering and Fisher linear discriminant classifiers with the training energy vector with different batch.
3. K-means clustering is better approach compared to another classifier because of capability PU detection being higher and classification delay and lower training.
4. Compared to other algorithms performance of detection is found highest in SVM classifier by using Kernel functions like polynomial kernel and linear kernel which is used to map feature space with higher dimensional space.
They also concluded that the classifiers can be trained by obtaining the energy vectors one to one by improving the CSS approaches. Comparative Table Method- ologies.
Table2shows the comparative methodologies used by various authors and their findings which led to a better result.