Diagnostic system for boilers and furnaces using CFD and neural networks

Computational fluid dynamics (CFD) tools were used to build a ''virtual'' furnace, validated with experimental data. This model was used to simulate both normal and ''faulty'' behaviours, regarding parameters such as energy conversion efficiency, fouling and steam leakage. A database was developed comprising normal situations and simulated fault sets, characterized by virtual sensor outputs used in the evaluation of diagnostic parameters patterns to be processed and recognized by the diagnostic system. Artificial neural networks (ANN) reduce modelization requirements compared with quantitative model-based approaches, while being relatively immune to noise and uncertainties and capable of extrapolating beyond their training scope. Training adjusts network parameters in order to best represent relationships underlying simulated sensor readings (network inputs) and diagnostic parameters used as outputs. Neural networks were used to process the database, with satisfactory results even in their most simple form (backpropagation networks) trained using standard algorithms. Pattern recognition was thus performed, identifying root causes of simulated anomalies. The interaction with related research areas and future proposed developments and improvements are also discussed.