A Novel Maximum Entropy Inference Engine for Data Fusion in Fault Diagnosis

Period of Performance: 07/01/2004 - 04/30/2005


Phase 1 STTR

Recipient Firm

Intelligent Automation, Inc.
15400 Calhoun Dr, Suite 190
Rockville, MD 20855
Principal Investigator
Firm POC

Research Institution

Pennsylvania State University
110 Technology Center Building
University Park, PA 16802
Institution POC


Inferencing and reasoning algorithms are widely used in various applications where root causes need to be inferenced based on observed evidence. One typical application is fault diagnosis of complex systems based on symptoms reported by various diagnostic monitors. The symptoms may contain both continuous and discrete variables. Intelligent Automation, Inc. and Professor David J. Miller of Penn State University propose to develop a new inference engine based on Maximum Entropy. The key capabilities include: 1) making principled, effective use of both continuous and discrete features; 2) solving general inferencing tasks, wherein any subset of features may need to be inferred, given values for the remaining features; and 3) handling large feature space. The work is built on efficient learning of maximum entropy Probability Mass Functions (PMFs). Our algorithm avoids an artificial mapping of continuous features, leading to a significant edge in solving mixed feature space inference tasks. Our learning algorithm builds the PMF directly from training data and consequently does not require making any explicit conditional independence assessments as compared to Bayesian Networks (BNs). In Phase 1 we will perform comparative studies between our proposed algorithm and other algorithm candidates. In Phase 2, we will implement the algorithm in real-time.