From industrial process control to network security: deployable data-driven approaches for anomaly and fault detection

The proliferation of interconnected systems within Industry 4.0, the Internet of Things (IoT), and advanced network infrastructures has generated an unprecedented demand for intelligent and autonomous monitoring to ensure reliability, efficiency, and security. However, the deployment of complex data-driven models for these tasks is often limited by the high-dimensional nature of industrial data and the computational constraints of embedded and edge devices. The thesis tackles these challenges through the development and evaluation of data-driven frameworks for predictive maintenance, anomaly detection, and network security across diverse application domains. The contributions range from the optimization of industrial monitoring processes—where managing data complexity is as crucial as improving model efficiency—to the deployment of lightweight, high-performance solutions for edge computing. The motivation behind this research stemmed from my curiosity about how data-driven methods could transform industrial systems and redefine the way we approach data analysis and anomaly detection. The first part of this thesis embodies this perspective, focusing on anomaly detection in critical pharmaceutical processes such as lyophilization and batch production to make industrial systems more efficient, reliable, and autonomous. The goal is to develop a model capable not only of identifying faults but also of anticipating them, thereby transforming raw data into actionable indicators of a system’s health. Starting from traditional data-driven approaches, such as Principal Component Analysis (PCA) [1], and extending the study to mathematical models based on explicit knowledge of the governing physical laws—often incomplete or built upon simplifying assumptions—the advantages of data-driven methods are highlighted. Building on these foundations, the study proposes a predictive framework using AutoEncoder (AE) model [3] as a core component for anomaly detection in dynamic production processes. Unlike linear data-driven approaches, DL models, such as AE, can learn complex, non-linear relationships within industrial processes, providing a more comprehensive understanding of system behavior. This led to the development of a Long Short-Term Memory (LSTM)-based framework specifically designed to capture temporal dependencies in multivariate signals, enabling accurate time-series forecasting. By combining these two tasks, a framework is defined to effectively distinct anomalous operating conditions in a predictive way. The second part focuses on the deployment of DL architectures on constrained hardware. A comparative analysis of Machine Learning (ML) and DL models is conducted for air quality forecasting [2] on IoT devices, demonstrating the advantages of Post-Training Quantization (PTQ) as an effective strategy to compress models while preserving accuracy. This investigation highlights the trade-off between model performance and computational efficiency, emphasizing the importance of optimization for real-world edge deployment. Finally, the research extends its scope to cybersecurity applications, assessing the performance of DL-based intrusion detection systems for IoT and Software-Defined Network (SDN) environments. Through the evaluation of quantization and compression strategies, the study demonstrates how Artificial Intelligence (AI) models can achieve high detection performance even under strict resource limitations, analyzing several ML and DL architectures under compression techniques like quantization and pruning. Overall, this thesis represents both a scientific contribution and a personal journey toward understanding how AI can evolve into a sustainable, deployable, and effective enabler for the next generation of intelligent, autonomous, and interconnected systems.