LSTM-BIGRU based Spectrum Sensing for Cognitive Radio

. Abstract There is a shortage of wireless spectrum due to developments in the area of wireless communications as well as the number of users that are using resources. Spectrum sensing is a method that solves the issue of shortage. Deep learning surpasses classical methods in spectrum sensing by enabling autonomous feature learning, which enables the adaptive identification of complicated patterns in radio frequency data for cognitive radio in wireless sensor networks. This innovation increases the system's capacity to manage dynamic, real-time circumstances, resulting in increased accuracy over traditional approaches. Spectrum sensing (SS) using LSTM-BIGRU with gaussian noise has been suggested in this article. Long-term dependencies in sequential data are well-preserved by LSTM due to its dedicated memory cells. In addressing and managing long-term dependencies in sequential data, BIGRU's integration enhances the efficacy of the model as a whole. To conduct the investigation, RadioML2016.04C.multisnr open-source dataset was utilized. Whereas, by using RadioML2016.10b open-source dataset, QAM64, QPSK and QAM16 performance evaluation has been investigated. The experimental findings demonstrate that the suggested Spectrum Sensing has better accuracy on the dataset particularly at lower SNRs. The improved spectrum sensing (SS) performance of our suggested model is shown by the evaluation of performance indicators, such as the F1 Score, CKC and Matthew's correlation coefficient, highlighting its potency in the field of spectrum sensing applications.


Introduction
The radio spectrum bands used for wireless communication are frequencies, and spectrum sensing identifies their availability and occupancy state.Cognitive radio networks heavily rely on spectrum sensing, which enables reliant users to access free frequency bands to optimize spectrum efficiency [1].Cognitive radio has been suggested to optimize the utility of wireless networks' spectrum resources.
The use of spectrum sensing strongly influences cognitive radio.Traditional spectrum sensing methods focus on identifying characteristics in a signal received at a given place through techniques like cooperative spectrum sensing, more accurate spectrum sensing is now possible using DL [2].Conventional detectors with their limitations include matched filters, energy detectors, and cyclostationary feature detectors [8] [9].
The requirement for spectrum for 4G and 5G networks has grown due to the widespread use of mobile devices.The present spectrum dilemma is primarily due to the wasteful usage of licensed frequency bands.Cognitive radio (CR) technology has gained popularity as a solution to this issue.The licensed spectrum of primary users (PUs) is available to secondary users (SUs) via CR Wireless communication has typically relied on specialized knowledge and too simplistic models to extract characteristics from the spectrum.These methods mainly depend on human operators and extensively use preconceptions [7].SUs are in charge of ensuring that their access doesn't interfere with PU signals [5].Spectrum sensing (SS), figure out if PUs is on or off, has been a challenging issue, with energy detection being the most preferred method because of its simplicity.There have been many recommendations in recent years to apply machine learning (ML)-based SS algorithms to improve the exposure performance for cooperative radio frequency sensing (CSS) [6].The SUs in the spectrum-aware radio network monitors the frequencies in use and, when appropriate, utilize the free channels [5].Multi-layer perceptron (MLP) networks are among the most widely used deep learning and machine learning frameworks incorporating spectrum sensing.The absence of memory components in MLP networks makes them potentially useless for time-series data.Alternatively, by extracting characteristics from the network itself, deep neural networks (DNN) provide a data-driven approach that may reduce the number of assumptions.Frequency spectrum prediction may reduce time delays by relieving the burden of spectrum sensing.Utilizing a deep recurrent neural network (DRNN) can predict the spectrum for several time slots, as present techniques can only expect the spectrum for a single or fewer time slot [3].Both super-vised and unsupervised machine learning DNN models are now used in RFML.How-ever, supervised machine learning networks are more common since they can be con-figured to do specific tasks rather than merely looking for structure in the data.The recent outcomes show that CNN, and RNN are more effective at spectrum sensing, reducing an algorithm's complexity show how performance might be improved [2].A vital component of the LSTM-based spectrum sensing strategy is training an LSTM network using labeled spectrum data, which imparts the network to discover patterns and correlations in the data.The LSTM network may be used for real-time spectrum sensing after being trained by feeding its received spectrum data; the network then determines whether or not a particular frequency band is open for usage [4].LSTM networks have gained popularity for their ability to learn temporal features from sequential input.[16][17] [18].An LSTMbased SS method use the data's temporal peculiarities [19].Spectrum-aware radio technique based on LSTM that takes use of the extraordinary learning power of LSTM networks [10] [15] to extract latent information from spectrum data, such as temporal correlations between current and timestamps [6].Recurrent neural networks (RNNs) [16] and hybrid CNN-RNN net-works have gained popularity over the last ten years, replacing convolutional neural networks (CNNs) [12], which were previously widely employed in RFML research [13].The Bidirectional Gated Recurrent Units (BiGRUs) neural network architecture is utilized for sequence modeling [20].They improve conventional GRUs by pro-cessing input data forward and backward and incorporating contextual data from previous and future time steps.By taking into account bidirectional context, they provide more detailed representations and increase the model's capacity for precise prediction in sequential data tasks

Related Work
Convolutional operations in the convolutional neural network (CNN) extract complicated characteristics from the preprocessed input that are crucial for regression tasks, according to studies by Mahak Kalra [11] et al.The CNN's input layer must first receive the pre-processed data, which must then perform convolution operations to extract complex features relevant to regression.The model consists of layers such as the sequence layer, LSTM layer, fully connected layer, and regression layer.The LSTM layer incorporates LSTM memory cells to enhance the model's memory capacity.The CNN-RNN model suggested by this research yields the following results the rate is 0.9895, the error proportion is 0.0105.CNN and SAM-CNN detectors can manage spectrum sensing in situations with intermittent prominent signal presence, according to studies by Zhan Cong et al. [12].Numerical findings demonstrate that the recommended detectors perform much better than the stateof-the-art detectors, with higher detection probabilities and improved ROC performance.This method also indicates that the SAM-CNN detector, which utilizes switch feature extraction from channel state data, is superior to the CNN detector in spotting abnormalities.The best performing of the other detectors achieves detection probabilities below 0.7.In contrast, the CNN and the SAM-CNN detectors achieve around 0.8 and 0.87 detection percentages with a false alarm probability (Pf) of 0.1, respectively.. The best performance of the other detectors drops to 0.55 when Pf is adjusted to 0.05, but the SAM-CNN detector retains a detection probability of around 0.84.When Pf is set to 0.05, the SAM-CNN detector improves the detection probability when compared to the CNN detector by 19%.The Generalized Likelihood Ratio Test (GLRT) was examined in the work of Q. Cheng et al. [13].The GLRT scheme's underlying Maximum Likelihood (ML) estimations were devised and expressed in closed forms, allowing for real-time implementation appropriate for automotive applications.The main goals of this method were to evaluate algorithm performance in dynamic random processes characteristic of urban traffic applications and to advance GLRT spectrum sensing for a generic rank-K situation, where 1 ≤ K < M. The research notably addressed the complicated vehicular-application spectra that are present in areas such as New York City.In the work by M. Karimzadeh et al. [14] it was shown that the suggested system performs better with heavy-tailed Generalized Gaussian Noise (GGN) than its predecessor PCA, and the well-known ED method.The slope parameter was optimized in order to obtain this improved performance.The research used a GGD with shape parameter η=0.9 to account for noise in numerical results.This distribution is some-what heaviertailed than a Laplacian distribution, which is generally regarded as the standard heavy-tailed distribution.The best value in the fading situation is considered.to be the same as that in the no-fading scenario, despite of the difficulty in calculating this value when the primary signal suffers fading.This presumption illustrates how difficult it may be to determine the ideal settings when there is signal fading.(5) and in .(6) [20]. denotes the momentary value, ❋ denotes element wise multiplication, and  denotes sigmoid activation.Grget, Gnpt, and Gtpt denote the weight matrices [21] which belong to forget, input, output gates.

Proposed Method
activation in LSTM is described in .(7) .BIGRU operates in both directions and extracts context features in the spectrum sensing model [21][22].The output is then passed through a dense layer where the features will be extracted, and finally, sigmoid activation is used for binary classification to classify signal and noise classes.F1 Score and MCC equations have been represented in eq. ( 1) and ( 2).Where φ indicates True, and θ indicates False.Suffixes p and n indicates positive and negative respectively.

Results and Discussions
The suggested method's SNR vs accuracy is shown in Fig. 2. The TP, TN, FP, and FN values have been determined using the confusion matrix.The categorization results are greater than 84% for both true positives and true negatives.This demonstrates how the model more accurately separates signals from noise.When the value is greater than -3 dB, the accuracy reaches 95%, and at lower SNR, it reaches 60%.The representation of confusion matrices in SNR wise is shown in Fig. 3.It shows that the categorization improves as SNR rises; even at lower SNRs, the classification is comparably better.Fig. 4. shows the SNR wise comparison of Performance metrics.It has been observed that at lower SNRs the MCC value is higher.And CKC value for the proposed system is above 0.9 which indicates that the performance of the system is reasonably good at all SNR.Similarly, MCC and F1 Scores have also been indicated SNR wise.The system performance assessment utilizing the RadioML 2016.04C.multisnrdataset is shown in Fig. 2, Fig. 3, and Fig. 4. Table 1 compares the suggested method's performance metrics.When it comes to spectrum sensing, a range of performance measures provide a thorough

Conclusion
In contrast to conventional techniques, spectrum sensing using deep learning in Cognitive Radio (CR) performs better because of advancements in AI and DL technology.This paper proposed the LSTM-BIGRU strategy as an enhanced DL-based spectrum sensing method that provides a sophisticated knowledge of spectrum sensing.Our model carefully examined performance metrics, such as MCC, F1, and CKC scores, to provide a thorough assessment of its effectiveness using, RadioML 2016.04C.multisnrdataset, and performance evaluation for QAM64, QPSK, and QAM16 have been performed using RadioML 2016.10bdataset.The suggested method

2 EAI
Endorsed Transactions on Internet of Things | Volume 10 | 2024 |

Figure 1 .
Figure 1.Proposed system Block diagram Equation (1) gives the model a problem conceptualization.Spectrum sensing is a representation

Figure 3 .
Figure 3. SNR wise Confusion matrices from -20dB to 16dB assessment of the detection procedure.The precision value of 0.8469 indicates the dependability of properly detected cases when it comes to assessing the accuracy of positive predictions.The percentage of true positives that are mistakenly categorized as negatives, or the false omission incidence, is 0.1384, which indicates a very low incidence of missed detections.The effective sensitivity in detecting positive cases is shown by the miss rate, which is 0.1357.With an observed value of 0.1530, the false discovery rate which quantifies the proportion of false positives to all anticipated positives indicates a modest degree of overprediction.The model's ability to

Figure 2 .
Figure 2. Confusion matrix and accuracy percentage comparison plot