Deep learning dual-model for predicting and validating sensor Data in undn

Abstract

This thesis presents a deep learning dual-model approach aimed at improving the reliability of sen- sor data in Underwater Named Data Networks (UNDNs)a novel paradigm that integrates Named Data Networking (NDN) into Underwater Wireless Sensor Networks (UWSNs). UWSNs are crit- ical for applications such as environmental monitoring, seismic activity detection, and underwater exploration, yet they suffer from severe challenges including high latency, limited bandwidth, signal attenuation, and sensor faults. NDN enhances underwater communication by enabling content-centric data retrieval with in-network caching and built-in data-centric security. However, NDN alone can- not handle issues like sensor malfunctions, missing readings, or poor data quality. To address these limitations, we propose a dual deep learning framework consisting of two inde- pendent models: a time-series prediction model used to estimate future sensor values and recover missing data, and a classification model designed to validate the correctness of sensor readings. Several architectures were evaluated, including LSTM, GRU, CNN-1D, Transformer, and Temporal Convolutional Net- works (TCN). While most models demonstrated promising performance during testing, only the TCN maintained strong generalization on unseen data. For the classification task, a Multi-Layer Perceptron (MLP) model was employed and showed robust accuracy in distinguishing valid from faulty readings. Both models performed well despite being trained on a limited dataset. Importantly, all models were implemented from scratch without relying on external frameworks. This study contributes toward building intelligent, adaptive, and resilient underwater monitoring systems capable of maintaining data integrity in challenging environ- ments Reliability