Industrial companies usually own assets which are prone to breaking down. Each year vast quantities of money are spent on maintenance to fix these assets after they have failed. This study presents a set of predictive maintenance techniques which could help combat this problem. A large set of sensor data taken from a subsystem of a commercial gas terminal in the UK was analysed. We first used a recently developed machine learning technique called conformal clustering to show that there are patterns in this data, indicating predictability, and also to identify anomalies in the data. We then proposed a way to identify which metrics may explain why the subsystem has failed. Finally we built a model to predict failure of the subsystem in probabilistic form, and ran experiments to optimise it. We believe that the methodology developed in this project could be generalised to many such cases when large data sets from industrial sensors are available.