Managing large fleets of machines in a cost-effective way is becoming more important as corporations own increasingly large amounts of assets. The steady improvement in cost and reliability of sensors, processors and communication devices has helped the spread of a new paradigm: the Internet of Things. This paradigm allows for real-time monitoring of countless physical objects, obtaining data that can be fed to machine learning algorithms to predict their future state and take managerial decisions.

Despite rapid technological change, industries have been slow to react, and it has been only recently that many have transitioned towards a new business model: servitisation. Servitisation is based on selling the services that assets provide, instead of the assets themselves. Although more companies are adopting this business model, there is a lack of solutions aimed to maximise its economic value. This thesis presents one such solution capable of predicting failures in real time, thus reducing a crucial cost contribution to asset ownership: unexpected failures. This new approach, Distributed Collaborative Prognostics, consists of providing each machine with its own particular agent, that enables it to communicate with other similar machines in order to improve its failure predictions.

This thesis implements Distributed Collaborative Prognostics in three different scenarios: (i) using a multi-agent simulation framework, (ii) using synthetic data from a well-established prognostics data set, and (iii) using real data from a fleet of industrial gas turbines. Each of these scenarios is used to study different elements of the prognostics problem. Multi-agent simulations allow for the calculation of the cost of predictive maintenance coupled with Distributed Collaborative Prognostics, and for the estimation of the cost of agent failures in different architectures. Synthetic data is used as a test bench and to study assets operating in dynamic situations. Real industrial data from the Siemens industrial gas turbine fleet serves to test the applicability of the tool in a real scenario.

This thesis concludes that Distributed Collaborative Prognostics is the adequate solution for large and heterogeneous fleets of assets operating dynamically. Its cost effectiveness depends on the value of the assets; in general, highly-valued assets are more conducive to Distributed Collaborative Prognostics, as the savings from improved failure predictions compensate the cost of enabling them with Internet of Things technologies.