Predictive maintenance strategies which estimate remaining useful life of system components to prevent breakdowns and down-times by timely and well-scheduled maintenance ensures the reliable availability of assets and lowers total costs of ownership. The focus on the components’ life times falls short, however, to infer the system-level capability to achieve upcoming tasks, especially if these tasks vary either in the strain they cause for the system or in the environmental conditions in which the system needs to perform. Such an assessment of the health and mission readiness of a system is crucial for mobile assets like seafaring vessels undertaking long-term operations without the option to easily come in for repairs or for industrial assets that need to complete long production runs in one go under varying circumstances. We propose a multi-step methodology to achieve such assessments using both Bayesian reasoning for diagnosis and prognosis and physics-based simulation models. First, we construct an appropriate Bayesian network in an object-oriented way by fitting a pre-compiled library of network fragments to the system’s schematics using generative techniques. We then parameterize the obtained network using a combination of expert knowledge and machine learning to fine tune system-level interactions between components and their link to the system’s performance. The learning step uses past operational data that we augment or complement with synthetic data, created by a physics-based simulation model, where needed. Finally, we use the trained Bayesian network to assess the mission readiness of the system given the probabilistics of its diagnosed state, expected impact of possible maintenance interventions, and the estimated profile of the future use. We illustrate and verify our methodology on a cooling system with an active feedback control loop, but our approach for mission readiness assessment is domain-independent, universally applicable, and typically feasible where operational data and engineering knowledge can be brought together to solve its challenge.
The usefulness of Bayesian network technology for expert-systems for diagnosis, prediction, and analysis of complex technical systems has been shown by several examples in the past. Yet, diagnosis systems using Bayesian networks are still not being deployed on an industrial scale. One reason for this is that it is seldom feasible to generate networks for thousands of systems either by manual construction or by learning from data. In this paper, we present a systematic approach for the generation of Bayesian networks for technical systems which addresses this issue. We use existing system specifications as input for a domain-dependent translation process that results in networks which fulfil our requirements for model-based diagnosis and system analysis. Theoretical considerations and experiments show that the quality of the networks in terms of correctness and consistency depends solely on the specifications and translation rules and not on learning parameters or human factors. We can significantly reduce time and effort required for the generation of Bayesian networks by employing a rules-based expert system for generation, assembly and reuse of components. The resulting semi-automatic process meets the major requirements for industrial employment and helps to open up additional application scenarios for expert systems based on Bayesian networks.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.