Learning in the Model Space for Fault Diagnosis

AbstractThe emergence of large scaled sensor networks facilitates the collec-tion of large amounts of real-time data to monitor and control complexengineering systems. However, in many cases the collected data may be in-complete or inconsistent, while the underlying environment may be time-varying or un-formulated. In this paper, we have developed an innova-tive cognitive fault diagnosis framework that tackles the above challenges.This framework investigates fault diagnosis in the model space insteadof in the signal space. Learning in the model space is implemented byfitting a series of models using a series of signal segments selected witha rolling window. By investigating the learning techniques in the fittedmodel space, faulty models can be discriminated from healthy models us-ing one-class learning algorithm. The framework enables us to constructfault library when unknown faults occur, which can be regarded as cog-nitive fault isolation. This paper also theoretically investigates how tomeasure the pairwise distance between two models in the model spaceand incorporates the model distance into the learning algorithm in themodel space. The results on three benchmark applications and one simu-lated model for the Barcelona water distribution network have confirmedthe effectiveness of the proposed framework.