Due to frequent changes in thermal load and drift of online oxygen analyzer, the heater's thermal efficiency optimization system with limited maintenance resources seldom works in long term. To solve this problem, a novel and practical optimization method combing RIMER (i.e. belief rule-base inference methodology using the evidential reasoning) approach and SA (stochastic approximation) on-line self optimization is proposed. The optimization scheme consists of (i) an off-line expert system that determines the optimal steady state operation for a given thermal load, and (ii) an on-line optimization system that further improves the thermal efficiency to alleviate the influence caused by sensors' drift. In more details, at the off-line stage, a belief-rule-base (BRB) expert system is constructed to determine good initial references of operating conditions for a specified thermal load, which quickly drives the system to a near optimal operation point when confronted with the thermal load change; this is based on RIMER. During on-line operation, these off-line determined initial values are further refined by using the SA approach to optimize the thermal efficiency. The newly obtained optimal operating condition then is updated online to compensate the sensor's drift. The optimized profile is implemented through a practical control strategy for the flue gas air system of fired heaters, which is applied to the flue gas oxygen concentration and chamber negative pressure control on the basis of flue gas-air control system. Simulation results on thelJniSim (TM) Design platform demonstrate the feasibility of the proposed optimization scheme. Furthermore, the field implementation results at a real process illustrate the effectiveness of this optimization system. Both simulation and field application show that the thermal efficiency can be nearly improved by c. 1%. (C) 2017 Published by Elsevier Ltd.