In collaboration with industrial partners, this research leverages the concept of a “Digital Twin” to create a virtual replica of industrial machinery that monitors health, identifies anomalies, and predicts potential failures. Analysing time-series data from multiple sensors poses significant challenges, particularly as machines dynamically adjust their operating conditions to meet production demands, making traditional forecasting algorithms ineffective. Additionally, the absence of labelled data distinguishing between healthy and faulty states necessitates unsupervised approaches that focus on the evolving correlations among sensor-measured physical properties as the machine ages. A series of experiments were conducted to evaluate multiple techniques for fault detection and predictive maintenance. Based on these evaluations, two effective approaches were selected: a statistical method combining Principal Component Analysis (PCA) with Mahalanobis Distance (MD) and a deep learning approach using Autoencoder Neural Networks (AE). Both methodologies consistently identified deviations from normal machine behaviour, capturing early signs of degradation and transitions from optimal performance to failure. The findings demonstrate the potential of these techniques for predictive maintenance, enabling timely alarms to prompt preemptive actions and minimise unplanned downtime. This research underscores the value of integrating data-driven anomaly detection methods within a Digital Twin framework to enhance the reliability, efficiency, and operational resilience of industrial systems.