AFrequency DomainRobustModelValidation Approach toFaultDetection andIsolation withApplications toIdentification ofContaminants in Lubrication andTransformer oils

This paperaddresses theproblemofrobust fault detection andisolation indynamical systems usingfrequency domaindata. Ourmainresult showsthatthis problem canbe reducedto a convexoptimization problemthatcanbe efficiently solved. Theproposed methodisputtotest onanon- trivial, practically relevant problem: detecting thepresence andestimating thecomposition ofcontaminants inlubrication andtransformer oil. I. INTRODUCTION rJTHE problem offault detection andisolation insystems lhasbeengiven considerable attention inrecent years. Themaindrive forthis isthegrowing needforon-board diagnostics andprognostics ofsystems inorder toreduce maintenance costsanddowntime(1-3)focusing on condition basedmaintenance rather thanschedule based maintenance. Manyengineering systems suchasairplane control systems, automotive electronic throttle and transmission systems, powerdistribution systems (such as transformers), chemical process control systems, high performance bearing systems, airborne systems andtanks equipped withgasturbines, allrequire theearly fault detection inorder toprevent catastrophic failure (3-8). Some ofthecommonclasses ofmethods forfault detection and isolation include: model-based analytical methods, physical redundancy andmodelindependent methods. Modelbased analytical approaches arepopular dueto theircost effectiveness. Many havestudied thiscategory using different mathematical models forthefailure prediction (9- 15). However, themajorchallenge withthemodel-based approach isthatitrequires detailed knowledge ofthe physics ofthesystem. Evenincases where adetailed model ofthesystem isavailable, itmaybedifficult toexactly determine thevalues ofallparameters involved, andany mismatch between theassumed andactual dynamics can leadtoincorrectly concluding thatfaults arepresent. To address thisdifficulty, manyresearches havefocused on robust FDI(12, 16-23). Thisreduces thepotential problem withmodelmismatch because itaccounts formodel uncertainty andmeasurement noise. Todate, manyofthe