Machine Learning-Based Prediction System for Optimizing Underwater Wireless Sensor Networks

Background/Objectives: UWSNs have significant applications in the marine ecosystem, oceanographic research, and inspection of marine infrastructures underwater. But the rough underwater conditions come with the problem of a long latency, loss of packets, attenuation of signals and a scarcity of energy, which lower network reliability and performance. This study aims at designing a machine learning-based prediction system to enhance the performance, reliability and energy efficiency of UWSNs by predicting the network behaviour and sensor node states given different underwater environmental and communication conditions. Method: The suggested predictive framework uses several machine learning models, such as CatBoost, XGBoost, LightGBM, Random Forest (RF), Decision Tree (DT), Gaussian Naive Bayes (GNB), and K-Nearest Neighbors (KNN). The models have been trained on a dataset of both real and synthetic data of sensors in underwater environments, sensor properties, and environmental conditions (temperature, salinity, pressure), and network variables (signal-to-noise ratio (SNR), traffic load, and battery level). Preprocessing of data using methods like data cleaning, Min-Max scaling, and Principal Component Analysis (PCA) are used to enhance the stability of the model. The dataset is split into an 80/20 training-testing data and 10-fold cross-validation is employed to make the dataset robust. Accuracy, Precision, Recall, F1-Score and ROC-AUC are used to measure model performance. Findings: The results of the experiments prove that CatBoost and XGBoost have the highest predictive performance and reach an accuracy of 92.1, a precision of 0.90, a recall of 0.91, an F1-score of 0.91, and a ROC-AUC of 0.93. LightGBM was able to perform competitively as well with an accuracy of 91.5 percent but at a faster execution and less consumptions of memory. Random Forest was very robust to noisy underwater data. Conversely, Gaussian Naive Bayes had the shortest execution time, but had worse accuracy (79.1) and less reliability in noisy conditions. The findings suggest that the performance of ensemble learning algorithms is superior to that of the traditional models to solve the underwater network complex and uncertain conditions. Novelty: This study makes one of the very first detailed comparative analysis on multiple machine learning algorithms to predict and optimize UWSN performance in realistic underwater conditions. The suggested framework incorporates both environmental sensing parameters and network metrics along with energy-related features in order to achieve predictive optimization of the network with the help of AI. The system allows adaptive and self-optimizing underwater networks, which allow the early warning of the communication breakdowns, better energy control, and increased reliability. The framework can find application in marine ecosystems, underwater pipeline inspection, disaster prediction and maritime security operations due to these capabilities. Keywords: Artificial Intelligence, CatBoost, Machine Learning, XGBoost, Network Optimization, Prediction System, Underwater wireless sensor networks