EPIC-IoT 2026 Abstracts

Area 1 - Edge-Intelligent and Privacy-Aware IoT Systems for Industry

Short Papers

Paper Nr:	5
Title:	Hybrid Gas Sensing with Adaptive Sampling
Authors:	Francisco J. G. da Silva, Tomás O. Rodrigues, Diogo Santos, Miguel dos Santos, João Longras, Emanuel Maia, Rui Pinto and Gil Gonçalves
Abstract:	Continuous gas monitoring in Internet of Things (IoT) environments requires balancing fast detection of hazardous events with long-term energy efficiency and sensor durability. Traditional fixed-rate sensing guarantees responsiveness but produces redundant measurements during stable periods, leading to unnecessary energy consumption and accelerated sensor wear. This paper proposes a hybrid, confidence-driven adaptive gas sensing architecture that dynamically adjusts sampling frequency according to environmental variability and model uncertainty. The system combines physical gas sensor nodes with virtual data generated through a Conditional Variational Autoencoder (CVAE), enabling safe augmentation of rare and hazardous gas scenarios that are difficult to capture in practice. A Random Forest (RF) classifier performs multi-class gas identification and outputs confidence scores, which are used as state inputs to a Proximal Policy Optimization (PPO) Reinforcement Learning (RL) agent responsible for controlling sampling intervals in real time. All models are trained offline and deployed for inference on a resource-constrained Raspberry Pi gateway, demonstrating feasibility at the edge. Experimental results using mixed real and synthetic datasets show that the adaptive PPO policy maintains classification accuracy up to 97% while reducing sensor readings by approximately 40% compared to fixed-rate and random sampling strategies under an equivalent sensing budget. A sudden gas event case study further confirms rapid sampling escalation under uncertainty and automatic return to energy-efficient operation once stability is restored, validating the practicality of adaptive sampling for reliable, energy-aware IoT gas monitoring.

Paper Nr:	6
Title:	Evaluating Feature Selection Techniques for Efficient Anomaly Detection in IoT Security Systems
Authors:	Edward Nepolo, Mkhuseli Ngxande and Attlee Gamundani
Abstract:	The Internet of Things (IoT) is said to have been the main driver of the Fourth Industrial Revolution, and is now considered a crucial enabling technology for the Fifth Industrial Revolution due to its ability to contribute to intelligent systems. This revolution has introduced challenges in securing IoT infrastructure against anomalies and zero-day attacks. Feature selection (FS) plays an important role in improving anomaly detection by reducing data dimensionality, computational resource needs, and enhancing model performance. This paper investigates the influence of Variance Threshold, ANOVA, Autoencoder (AE), Chi-square, and ensemble feature selection techniques on the performance of four anomaly detection models using the ToN_IoT and NSL_KDD datasets. The performances of the four anomaly detection models Isolation Forest, One Class SVM (Support Vector Machine), Variational Autoencoder, and standard Autoencoder are evaluated using precision, recall, F1 score, and accuracy. The paper demonstrates that statistical models ANOVA, and Chi-square achieved competitive performance, while the ensemble method achieved marginal accuracy, precision, and F1 score on the ToN_IoT dataset.

Paper Nr:	7
Title:	Real-Time Water Leakage Detection: A Multi-Tiered IIoT Architecture Using Machine Learning
Authors:	David Teixeira, Hugo Castro, Michele Pozzozengaro, Pharrell Tratsaert, Tomás Vinhas, José Oliveira, Rui Pinto and Gil Gonçalves
Abstract:	As industrial and residential water management systems evolve toward Internet of Things (IoT) paradigms, real-time leakage detection becomes essential for resource efficiency and infrastructure protection. This paper presents a multi-tiered IoT architecture for intelligent water leak detection, integrating edge simulation, Message Queuing Telemetry Transport (MQTT)-based middleware, and Machine Learning (ML)-driven analytics. A multi-threaded simulation environment was developed to emulate distributed sonar and temperature sensors under stable, leaking, and refilling conditions, enabling controlled validation of the proposed approach. At the edge, a Random Forest (RF) classifier deployed on a Raspberry Pi performs real-time inference on streaming telemetry. A key contribution is a differential feature engineering strategy that computes the rate of change between consecutive distance measurements, allowing the model to distinguish static low water levels from active leakage events—an ambiguity common in threshold-based systems. Experimental results on synthetic datasets demonstrate a 99% F1-score with sub-millisecond inference latency, confirming the feasibility of edge-based intelligence. The paper also discusses the transition toward a physical Proof of Concept (PoC), addressing sensor noise and model adaptation challenges. The complete pipeline, supported by InfluxDB and Grafana, provides a scalable and resilient framework for modern water infrastructure monitoring.