Journal of Propulsion and Power · 2013 · 50 citations · 30 references
ComputingComputer ScienceSystems EngineeringDynamical SystemsMachine LearningTime Series AnalysisAnomaly Detection In Time SeriesCondition MonitoringFault DetectionState EstimationState-space ModelsSystem IdentificationCombustorsEquivalence RatioGas Turbine OperationLean BlowoutSensor-based Anomaly DetectionState Vector EstimationState Vector RepresentationSymbolic Analysis
This paper develops a novel strategy for prediction of lean blowout in gas turbine combustors based on symbolic analysis of time series data from optical sensors, where the range of instantaneous data is partitioned into a finite number of cells and a symbol is assigned to each cell. Depending on the cell to which a data point belongs, a symbolic value is assigned to the data point. Thus, the set of time series data is converted to a symbol string. The (estimated) state probability vector is computed based on the number of occurrence of each symbol over a given time span. For the purpose of detecting lean blowout in gas turbine combustors, the state probability vector obtained at a condition sufficiently away from lean blowout (reference state) is considered as the reference vector. The deviation of the current state vector from the reference state vector is used as an anomaly measure for early detection of lean blowout. The results showed that the rate of change of the anomaly measure with equivalence ratio changed significantly as the system approached lean blowout. This change in slope of the curve was observed approximately at a similar proximity to lean blowout for different operating conditions and, hence, could be used as an early lean blowout precursor. The actual location of the change of slope depended primarily on the choice of reference state. This technique performed satisfactorily over a wide range of premixing.
30
Symbolic time series analysis via wavelet-based partitioning
Venkatesh Rajagopalan, Asok Ray · Signal Processing · 2006 · 239 citations