In this paper, we investigate the use of Bayesian networks to construct large-scale diagnostic systems. In particular, we present a novel analytical and experimental approach to developing large-scale Bayesian networks by composi- tion. This compositional approach reects how (often re- dundant) subsystems are architected to form systems such as electrical power systems. We develop Bayesian networks and clique trees representing 24 different electrical power systems, including the real-world electrical power systems ADAPT. ADAPT is representative of electrical power sys- tems deployed in aerospace, and is located at the NASA Ames Research Center. The largest among these 24 Bayesian networks contains over 1,000 random variables. Related work has used Bayesian networks to diagnose specic elec- trical power systems, however we are not aware of previous research that investigates a wide range of distinct electrical power systems as is done in this paper. While we consider diagnosis of power systems specically, we believe this work has application to numerous health management problems, in particular in dependable systems such as aircraft and space- craft. an EPS by means of Bayesian networks. The real-time rea- soning challenge is associated with the embedding of AI components, including diagnostic reasoners, into hard real- time systems. The scalability challenge, which is the main focus of this paper, is concerned with how different EPSs, with varying architectures and components, can be repre- sented in varying Bayesian networks for diagnosis, and how space requirements and computation times vary accordingly. The modelling challenge has been addressed by a high-level specication language from which Bayesian networks are auto-generated; the real-time reasoning challenge by off- line compilation of Bayesian networks into clique trees or arithmetic circuits which are used on-line. The scalability challenge is addressed by means of composition, in other words by considering the subsystems making up a system. For example, a electrical power systems can be made up by power storage and power distribution subsystems. We pro- vide in this paper several novel analytical and experimen- tal results that shed light on large-scale BNs developed for electrical power system diagnosis. The experimental part in- cludes results for Bayesian networks and clique trees repre- senting 24 different electrical power systems, including the Bayesian network and clique tree models of ADAPT. Indication of application status (e.g., feasibility analy- sis, research prototype, operational prototype, deployed application, etc.): We discuss the development of diagnos- tic applications for 24 different electrical power systems, including the Advanced Diagnostics and Prognostics Test- bed (ADAPT) (see also http://ti.arc.nasa.gov/ adapt/). ADAPT, which has capabilities for power gener- ation, power storage, and power distribution, is a fully oper- ational electrical power system that is representative of such systems in aircraft and spacecraft. We have developed a di- agnostic application that is an operational prototype working on real-world data from ADAPT. This paper takes the next step by considering how what we have learned from ADAPT can be applied to electrical power systems that are similar to ADAPT, but of different sizes and structures. Specically, we discuss the composition of electrical power systems from power storage and power distribution subsystems, and how this composition is reected in the Bayesian network and clique tree models of these EPSs.