Resource Optimization of Cognitive Radio Sensor Network Using Hybrid Metaheuristic Optimization and Machine Learning Algorithms

One of the most important uses of wireless communications technologies is their contribution to supporting a wide range of different applications in the fields of electronic health, which contribute to the transfer of medical data and patient information to the hospital and the medical side [1]. However, the use of wireless communications technology in the fields of medical specialties represents a major challenge, as it can affect the electromagnetic interference caused by these vital devices as a result of their wireless devices, which poses a threat to their performance. In addition, different types of e-health applications have different priorities, so these must be given priority to the wireless channel by the relevant corresponding devices influencing its operation [2].

One of the most important basic requirements in the field of wireless communications service is the availability of radio spectrum, and it is considered a major challenge as the spectrum used in the field of wireless radio is a limited and expensive natural resource, and the tremendous growth of the wireless communications market has led to the scarcity of radio spectrum [3]. As for the remaining part of the spectrum, it is unused or little used. Therefore, there is a need to improve the use of the spectrum, and the ideal choice is dynamic access to the spectrum and defining a cognitive radio network [4].

The principle of operation of cognitive radio networks is to adapt to the surrounding radio environment by relying on the knowledge generated from this environment [5]. They can access the licensed spectrum intelligently, opportunistically, and more efficiently without affecting the licensed spectrum, which improves the general use of the radio spectrum [6]. Cognitive radio networks are based on spectrum sensing to discover vacancies and obtain appropriate use of the spectrum through dynamic access It discovers spectrum gaps efficiently and accesses the spectrum in an optimal, opportunistic manner [7].

One of the important uses of radio-based cognitive wireless communications is emergency networks designed for disaster situations and healthcare systems [8]. It is the nature of the radio system's work and perception and its dealing with channels of the spectrum that are not sufficiently used. It transmits data of various types and media and does not take the access of secondary or low-priority users into account [9]. Cognitive radio in healthcare applications is used in sensitive medical devices that are supposed to be protected from interference as a result of wireless transmissions. They are including incubators, infusion pumps, and pacemakers, as medical devices can transmit data in a representative manner by using wireless signal for example, these devices include telemetry monitoring devices, wireless Holter monitors, and wireless electrocardiogram devices [10]. The data transmission of these active medical devices can be interfered with through wireless transmission, relying on other non-medical devices, such as medical devices and information networks [11].

Since every user in radio networks has access to the radio spectrum, the secondary user does not need to be aware of the presence of the primary user [12]. However, restrictions regarding user interference in protected applications still exist, and prioritization between different users and applications will be required to access the channel. To achieve the best service differentiation [13]. This is through the ability to monitor the spectrum and its characteristics and study the behavior of the node that operates in this spectrum and access it dynamically, as the cognitive radio node allows access to the spectrum and taking into account the quality of service during its different health normalization. It also allows relying on a spectrum or supports the spectrum as an ideal associated channel behavior [14].

Radio resources can then be allocated, such as transmission and reception time periods, specifying frequency bands, transmitting antennas, transmission power, etc., according to the current channel condition. Dynamic resource allocation schemes can rely on these variations in the frequency resources and thus lead to much better performance compared to static resource allocation schemes [15]. In addition, dynamic resource allocation works to resolve conflicts between competing nodes to exploit multiple resources and to ensure achieving fairness and avoiding interference and collisions, as dynamic resource allocation schemes deal efficiently with spectrum movement and take care of the quality-of-service requirements and priorities of different services and different nodes [16].

Besides, artificial intelligence learning technologies have gained many applications in the field of cognitive radio devices, as these technologies aim to make machines perform tasks in a way that resembles experts, such as inference, reasoning, problem-solving, knowledge representation, and learning approaches [17].

The main research problems are the dynamic spectrum allocation, how to ensure the quality of service, increase network life, and ensure the lack of interference between nodes and the network elements, reducing sensing information overhead, as well as computational interference operations which affect primary and secondary users.

The objectives of the proposed system are resource allocation, which results in increasing throughput and maximum resource usage of the entire network, and reduced processing time of sent/received packets. Balancing of the sensed spectrum between multiple sensor nodes, and efficient use of spectrum during the sensing process. Optimally meeting QoS requirements reducing or avoiding interference with the primary user network, and eliminating unnecessary transfers of the spectrum to reduce the overall power consumption, reduce the overall delay, and improve the reliability of transmission.

This paper is organized into several key sections. Section 1 showed the general introduction about cognitive radio sensor network (CRSN) and resource optimization based on Metaheuristic with Machine Learning Algorithms. Section 2 reviewed common related works of the proposed system. Section 3 showed the proposed methodology based on hybrid Particle Swarm Optimization (PSO) and Tabu Search (TS) algorithms as hybrid approach Particle Swarm-Tabu Search Optimization (HPSTSO) for resource allocation in cognitive radio network. Section 4 presented the proposed implementation method based on OMNET++ for network simulation and Java for trained model of MLP machine learning algorithm. Section 5 and Section 6 showed the proposed system results and system comparison. The proposed system is concluded in Section 7.

The spectrum sensing information metrics are exported into a CSV dataset file as cognitive radio spectrum sensing information attributes. The CRSSI.csv dataset is tested and evaluated with machine learning methods to enhance the resource allocation of cognitive radio spectrum sensing in the health monitoring system. So, the primary objectives of the Hybrid system under consideration are Decrease Cost, High Scalability, High Flexibility, Maximum Resource Utilization, Decrease Processing Time, Increase Availability and High Throughput. Figure 1 shows the flowchart of the proposed system.

1.png

Figure 1. The proposed CAMOA

The machine learning model consists of the following steps:

(1) Data gathering from the OMNET ++ simulation environment: Network simulation records are exported into a dataset (CAMOD.csv) file which contains the same attributes from the first stage of resource allocation parameters in Table 1. The system is executed for 60 minutes to record statistics of network evaluation metrics are 16 attributes and the total number of records is about 207993 records.

(2) Preprocessing of collected CAMOD.csv dataset: The pre-processing methods to improve the performance of the proposed multi-layer neural network include several methods that contribute to improving spectrum sensing, as it is important to build an effective data set before training the system, as the following methods include:

• Data cleaning involves dealing with missing values by removing records that contain missing data or empty values, such as strategies for entering the mean or arithmetic mean to fill in empty records.
• Measuring features is done by unifying the criteria for normalization features to values within the network parameters used in the dataset, and this is done by applying the minimum and maximum normalization measurement by setting values for measuring features with a fixed range, often between zero and one.
• Encoding categorical variables by encoding values within categorical attributes by converting them into numeric formats, which is more compatible with the proposed algorithm.
• The data is divided into a data set, 70% of the training set, 20% of the test set, and 10% of the validation set. The proposed division contributes to the effective evaluation of the model to ensure the highest accuracy in spectrum validation detection.
• Regularize imbalanced data by applying oversampling techniques from minority classes or under-sampling from the least dominant class to ensure that the model is trained effectively across all data classes.

The trained MLP model was built and evaluated in Java using Eclipse Deeplearning4j for detecting primary user appearance with high accuracy, precision, recall, and F1-score to assess the model's performance. The steps of the experiments using this dataset are shown in Figure 2.

2.png

Figure 2. The proposed machine learning algorithm model

The reason for choosing the MLP algorithm over other neural network algorithms is attributed to several factors as strengths of this algorithm, as follows: ease of implementation, as it is relatively simple in construction from connected layers to process the inputs directly, which allows for an easier implementation period and faster training time compared to other more complex algorithms, which makes it suitable for applications with limited resources and require fast decision-making processes. It is also characterized by low demand for processing power without the additional cost associated with building the algorithm, and thus it is quick to respond during the execution of tasks. Furthermore, the MLP algorithm has shown high accuracy when trained in situations where data availability is restricted and linked to the effectiveness of resource management.

The proposed MLP algorithm is a complementary component of the neural network, as it consists of three separate layers, represented by the input layer, the output layer, and the hidden layer. The proposed model is based on relying on different configurations by changing several hidden layers within the neurons of this algorithm, and the best model to accomplish the specific task is determined, as it is proposed within the system with 20 layers. This algorithm is hidden by performing cross-validation and setting the number of folds to ten folds. This model is based on being the most in-depth method for determining general properties. Use the trained MLP model to predict initial resource allocations based on updated input features derived from the current state of the system with current transmission conditions. MLP model incorporates these predictions into the decision-making processes of Hybrid HPSTSO approach of both PSO and Tabu Search algorithms.

The proposed system relied on a variety of data traffic that simulates the application of cognitive radio in healthcare in terms of the nature of data traffic in the network. The proposed system addressed a set of scenarios for different types of data such as text data, image data, and audio, videos and files, with different data sizes passing through radio channels and transmitted between SU nodes in the network. The results summarized the importance of resource management in the network in order to ensure the best scenario for data traffic and ensure the quality of communication for secondary users in cognitive radio networks.

The proposed system results are explained as follows:

5.1 Network evaluation results

The results of the proposed system depend on the performance improvement of the proposed hybrid algorithm, as it showed significant improvements in many performance measures compared to other algorithms, as the packet delivery rate improved as a result of the ease of transferring data more efficiently and maintaining reliable communication between secondary users in the cognitive radio network, as the packet delay was reduced from the beginning of its creation to the second party that receives this packet, which shows the algorithm's ability to accelerate data processing and transmission time and thus make a faster decision in a dynamic environment compared to other methods followed by the two proposed algorithms. The proposed results are based on four case studies: without resource optimization, optimization using the PSO algorithm, optimization using the TS algorithm, and resource optimization using the hybrid HPSTSO algorithm.

1. Results without optimization

In the absence of resource optimization techniques, the results and evaluation of the system showed poorer performance compared to systems using algorithms, as traditional methods struggle to route data through available channels, causing inefficiency for secondary users in the cognitive radio network due to increased packet delay and communication delays, putting it in real-time monitoring capabilities, as well as less effective decision-making processes and poor signal quality due to increased susceptibility to interference, which causes a higher probability of channel collision due to insufficient management of channel access, and thus a decrease in the overall efficiency of spectrum utilization, which causes drawbacks when using this strategy to manage resources in the network.

2. Results of PSO

The results showed resource optimization improvement compared to the system of without resource optimization or the TS algorithm, the noticeable increase in the evaluation metrics as a result of improving the data transmission path, reducing packet loss, and reducing the delay of its arrival time, as well as improving decision-making processes after reducing the waiting time to reach the channel, as well as reducing the processing time in general, as this algorithm has proven its ability to choose the channel with the best quality and the least interference to maintain a strong, stable connection and reduce the possibility of collisions between channels, which clarified the effectiveness of this algorithm in improving the performance of systems based on the health aspect in cognitive radio networks compared to both cases without improvement and the case of the TS algorithm.

3. Results of TS optimization

The results of this algorithm showed an improvement compared to the case without resource allocation and to a lesser degree than the PSO algorithm, as it proved its ability to transfer data, reduce packet delay, enhance data routing efficiency and deliver it faster. It also showed its ability to find solutions to simplify decision-making processes and increase the signal-to-noise ratio, which results in choosing channels with better signal quality and less interference, in addition to maximizing productivity and improving system performance, especially in scenarios that require strong and reliable communications between different network components.

4. Results of HPSTSO

The results area of the hybrid system is an advantage compared to the previous study cases of the PSO and TS algorithm in evaluating the network performance, as it has proven significant improvements in increasing the packet delivery ratio, which reflects the effectiveness and externality of the hybrid in improving data transmission paths and reducing packet loss with a decrease in the latency of arrival. It also showed the ease of making decisions to choose the empty channel and the decrease in the probability of secondary users colliding to obtain the channel, as this algorithm combines the strengths of both the PSO algorithm and the TS algorithm to enhance decision-making processes and simplify computational processes in improving productivity while ensuring effective spectrum use and increasing reliability and efficiency in improving resource utilization for the cognitive radio network.

Figure 7 showed that the hybrid system significantly outperforms the other methods in terms of throughput, achieving 21,379 Mb/s, which is significantly higher than both PSO and TS algorithms, as well as the case without optimization, in which the throughput dropped to only 8624 Mb/s. This performance is attributed to the strength of the hybrid approach, which effectively outperforms all the PSO and TS algorithms, and its more efficient use of resources, which contributed to enhancing the data processing capabilities of the network.

7.png

Figure 7. The average of throughput in Mbps for packets sent and packets received

Figure 8 showed the hybrid HPSTSO is an efficient approach characterized by an average processing time of 30 seconds, which makes it the fastest option compared to the PSO and TS algorithms. This proposed system not only achieves higher productivity, but also significantly reduces the processing time, which enhances the overall performance of the system and the speed of response.

8.png

Figure 8. The average processing time in seconds of the used algorithms

Figure 9 showed the percentage of average packet delivery ratio demonstrates the effectiveness of the hybrid HPSTSO approach compared to other methods, achieving a delivery ratio of 99%, indicating that almost all packets were successfully delivered. These results confirm the HPSTSO system ability to maintain high packet delivery rates, which is critical in ensuring reliable communication and data integrity in a cognitive radio network.

9.png

Figure 9. The average of packet delivery ratio of the proposed hybrid approach

Figure 10 showed the average packet delay for the different algorithms in the proposed system, where the hybrid HPSTSO approach proved an average delay of 30,929 milliseconds as the lowest average delay for packet transmission in the network. The results indicate the efficiency of the hybrid HPSTSO system in reducing packet delay, which is reflected in the efficiency of the network in general compared to other methods.

10.png

Figure 10. The average end-to-end packet delay of the proposed HPSTSO approach

Figure 11 showed the efficiency of the hybrid HPSTSO system compared to the PSO algorithm and the TS algorithm in finding the congestion-free channel rate of 99.17%, which indicates reducing channel congestion and reducing unjustified channel occupation, which reduces network performance and affects the smooth transmission of data.

11.png

Figure 11. The average of channel utilization of the proposed HPSTSO approach

Figure 12 showed average probability of channel collision in the network where the HPSTSO method achieved a low collision probability of 0.0508 which indicates the use of a high efficiency which enhances the overall performance of the network and makes it a superior choice for improving channel utilization in cognitive radio networks.

12.png

Figure 12. The average probability of channel collision of the proposed HPSTSO approach

Figure 13 showed that the hybrid HPSTSO system is effective in maximizing network capacity and resource efficiencies compared to other algorithms, highlighting the efficiency of the proposed hybrid system in improving the overall system performance and network resources.

13.png

Figure 13. The average of network and resource utilization of the proposed HPSTSO approach

Figure 14 showed the differences in performance for the signal-to-noise ratio results. The hybrid HPSTSO approach achieved a ratio of 7.33 dB, which indicates its effectiveness in maintaining a strong signal in a light environment. This is much higher than the proposed PSO and TS algorithms. The hybrid approach shows its superior ability to improve signal strength, making it the most reliable choice for effective communication in noisy environments.

14.png

Figure 14. The average of SNR of the proposed HPSTSO approach

Table 3 showed the proposed Hybrid Particle Swarm-Tabu Search Optimization (HPSTSO) approach compared with other related works associated with the routing approach in cognitive radio network. There is a gap in the relevant work related to the shortage of spectrum resources and low utilization rate without impact on licensed users in cognitive radio networks. The high average results, especially the packet delivery results, indicate that the proposed spectrum management approach is effective in ensuring reliable communication between nodes in a cognitive radio network. The implications for future spectrum management are improved spectrum efficiency due to the effectiveness of the proposed approach in managing the active spectrum while reducing interference between primary and secondary users, as well as supporting dynamic spectrum access, which allows real-time adjustments based on network conditions and data user requirements.

5.2 Results of machine learning algorithm

The proposed system results are based on two datasets (CAMOD and RadioML2016.10B) mainly each dataset is split into (70% training and 30% testing. The preprocessing model used is the same for both datasets. The first created dataset CAMOD contains the total number of columns and rows explained in Table 4.

Table 4. Number of records and attributes of CAMOD dataset

1. Results of hybrid optimization HPSTSO without preprocessing of CAMOD dataset

The results show that applying the hybrid optimization algorithm without preprocessing on the CAMOD dataset, a single performance of 94.14% is considered a high level of predictive accuracy, as well as a false positive rate of 0.9075, which requires room for improvement in reducing false positives based on the total processing time of 132728 ms, which reflects the computational requirements for training and evaluating the complex neural network in general and needs further improvement. Table 5 shows the accuracy and time results of the used dataset.

Table 5. The results without preprocessing of the MLP algorithm on the CAMOD dataset

The results in Table 6 showed the algorithm's ability to handle classification effectively and its ability to achieve accuracy even in the absence of database refinement.

Table 7 showed a set of evaluation parameters for the proposed algorithm on the CAMOD dataset. The results showed that the model predicted the positive class with an accuracy of 0.91791, a detection rate of 0.1503, and a false alarm rate of 90%. These highs indicate a large number of negative cases that were incorrectly classified as positive, as well as a set of other parameters that indicate the presence of discrepancies between the predicted and actual values, a low detection rate, and a high false alarm rate, which require further improvement of the overall performance of the model.

Table 6. Correctly and incorrectly classified the MLP algorithm on the CAMOD dataset

Table 7. Evaluation of MLP algorithm based on the CAMOD dataset

Table 8 showed the performance of the proposed MLP algorithm on the CAMOD dataset. The low Recall value indicates that the model is able to identify 9% of the actual positive cases, which highlights the major challenge in capturing cases. Also, the F-value between precision and recall is relatively low, so the model struggles for overall effectiveness in case classification. As for the Kappa coefficient, which measures the agreement between predicted and actual classifications, it is also low, which requires better model performance.

Table 8. Recall, the F-Measure and Kappa coefficient of the without preprocessing of MLP algorithm based on the CAMOD dataset

2. Results of hybrid optimization HPSTSO with preprocessing of CAMOD dataset

This case depends on refining the CAMOD database that was created in the first part of the system using resource optimization techniques and managing the data transfer process in the network, as the proposed algorithm is based on the MLP machine learning model, which achieved detection accuracy in sectarian sensing, whether sensing the radio spectrum, whether from secondary users or primary users, as Table 9 showed a percentage of 99.45%, a high level of performance according to a low false positive rate of 0.0085, which indicates the effectiveness of the proposed refinement system in correctly identifying categories and high prediction capabilities according to a very short processing time that contributes to investing time to move between the channels proposed by FC.

Table 9. The accuracy of hybrid optimization HPSTSO with preprocessing of CAMOD dataset

Table 10 showed the exceptional performance in classifying data from the proposed database and the low error rate as a result of the reliability of the model and its strength in making predictions related to radio spectrum sensing. It is considered suitable for applications that require very accurate classification results and high-level performance applicable in realistic scenarios.

Table 10. Correctly and incorrectly classified of HPSTSO with preprocessing of CAMOD dataset

The results in Table 11 and Table 12 showed that all the positive predictions made by the model were correct, which indicates complete accuracy that contributed to a detection rate that correctly identifies all actual positive cases, as well as a low false alert rate, as the model generates very few false positives, and thus false alerts are few and do not cost the system any acquisition operations. The model's ability to distinguish between categories was high, as it succeeded in identifying all positive cases as well as identifying negative cases.

The results indicate a significant improvement in resource allocation efficiency when employing the hybrid PSTSO optimization compared to traditional methods. Specifically, the hybrid approach demonstrated a marked increase in throughput and resource utilization while effectively reducing interference among SUs. The evaluation metrics revealed that the hybrid optimization outperformed both the PSO and TS methods individually, showcasing the benefits of combining these two powerful algorithms. Furthermore, the preprocessing of data prior to applying the MLP model yielded superior results in terms of accuracy and F-measure, indicating that data quality plays a crucial role in enhancing machine learning outcomes. Overall, the results highlight the effectiveness of the HPSTSO model in optimizing spectrum allocation within CRNs, providing a robust framework for future research and practical applications in cognitive radio systems.

Table 11. Evaluation of HPSTSO with preprocessing of CAMOD dataset

Table 12. Recall, F-Measure and Kappa coefficient of HPSTSO with preprocessing of CAMOD dataset

3. Results of the preprocessing approach of RadioML2016.10B dataset

The proposed algorithm is tested with the RadioML2016.10B dataset with a total sample sum of 240,000. The entire sample sets, which include both signal and noise, have been labelled for training, validation, and classification. The RadioML2016.10b dataset includes several specific attributes that characterize the radio signals it contains. The main attributes are as follows:11 Modulation Type, SNR, Sample Frame Size, Number of Samples, and Channel Effects. Table 13 showed number of records and attributes of RadioML2016.10B dataset.

Table 13. Number of records and attributes of RadioML2016.10B dataset

In cognitive radio networks, spectrum sensing is crucial for identifying unused frequency bands that can be utilized by secondary users without interfering with primary users. Table 14 showed the resulting accuracy of 98.7% in this context means that the MLP can reliably detect these opportunities, leading to better resource utilization and reduced interference.

Table 14. Accuracy, confusion matrix and time of the RadioML2016.10B dataset after the preprocessing approach

Table 15. Correctly and incorrectly classified the RadioML2016.10B dataset after the preprocessing approach

Table 16. Evaluation of the RadioML2016.10B dataset after the preprocessing approach

Table 17. Recall, the F-Measure and Kappa coefficient of the RadioML2016.10B dataset after the preprocessing approach

Tables 15-17 show the extent of improvement shown by the proposed system based on correctly classified cases at a rate of 98.7% of the total data set. This is a result of the effectiveness of the preprocessing method in increasing the efficient network performance on the database, as it led to increasing the accuracy of the model in classifying and identifying signal modifications within the RadioML2016.10B database and reducing the low error rate. In addition, the vast majority of positive predictions provided by the model are correct, which indicates an exceptional level of accuracy and the detection rate of actual positive cases is normal, as well as the low percentage of false alert rates and the ability of the proposed system to distinguish between categories based on the proposed algorithm contributed to the increase in the true positive rate with a decrease in the average absolute error, which contributed to improving the algorithm's ability to find efficient practical detection in classifying wireless signals for the RadioML2016.10B database.