In recent years, Bayesian networks have been drawing attention of the industrial and research community especially in the field of diagnostics for the reasoning capabilities they offer under conditions of uncertainty. Given the system of interest, a Bayesian network represents a graphical model of the system itself, in which the different players are linked to each other through probabilistic and causal relations. If the model is queried with appropriate statistical techniques, the whole approach can present several advantages over other data analysis methods. Among the others: 1) the approach can provide outputs even if some entries to the model are missing, due to the above mentioned dependencies between the players of the system; 2) the approach represents an ideal environment to include prior knowledge during the building up of the model, given the causal and probabilistic semantics; 3) a Bayesian network provides the possibility to learn causal relationships and gives therefore the possibility to improve the domain knowledge. Airbus Defence and Space has been working on improving the aircraft diagnostics capabilities at component, subsystem and system level in terms of fault detection and isolation. The focus has been also to develop means for reasoning about the remaining operational and functional capabilities of the aircraft. The initial outcomes have been tested on a simulation platform featuring a Data Acquisition Processing Unit, various computing nodes, on which the different aircraft systems (like the fuel system, the hydraulic system, the actuation systems, etc...) run. The data communication architecture of the platform is based on OSA-CBM (Open System Architecture for Condition-Based Maintenance). Initial objectives of the project are: 1) to demonstrate the feasibility of integration of the concept within the above described simulation framework; 2) to develop means to allow an easy and structured translation of the system engineer knowledge in terms of a Bayesian network with associated conditional probabilities; 3) to provide a modular architecture for the concept facilitating effective coordination between the development-departments and efficient development and maintenance of the software and 4) to prove the scalability of the concept (i.e. applicability to systems of different sizes and reasoning on different levels from component to system level). The candidate systems selected for the proof of concept are the fuel and the hydraulic systems of a generic aircraft. The results obtained so far look promising with respect to the above mentioned objectives of the project.
[1]
Randall Davis,et al.
The Origin of Rule-Based Systems in AI
,
2005
.
[2]
Balaram Das,et al.
Generating Conditional Probabilities for Bayesian Networks: Easing the Knowledge Acquisition Problem
,
2004,
ArXiv.
[3]
Adnan Darwiche.
Argument Calculus and Networks
,
1993,
UAI.
[4]
Jesús Cerquides,et al.
Tractable Bayesian Learning of Tree Augmented Naive Bayes Classifiers
,
2003
.
[5]
Randall Davis,et al.
Model-based reasoning: troubleshooting
,
1988
.
[6]
Adnan Darwiche,et al.
Compiling Bayesian Networks Using Variable Elimination
,
2007,
IJCAI.
[7]
Sou-Sen Leu,et al.
BAYESIAN-NETWORK-BASED HYDRO-POWER FAULT DIAGNOSIS SYSTEM DEVELOPMENT BY FAULT TREE TRANSFORMATION
,
2013
.
[8]
Adnan Darwiche,et al.
Compiling Bayesian Networks with Local Structure
,
2005,
IJCAI.
[10]
Davis E. King,et al.
Dlib-ml: A Machine Learning Toolkit
,
2009,
J. Mach. Learn. Res..
[11]
Mark Schwabacher,et al.
A Survey of Data -Driven Prognostics
,
2005
.