Vibrations are concomitant to operational industrial mechanical systems and in long-term are a potential threat for faults such as misalignment, increased wear or even severe damage of the machine parts. One of the physical phenomena behind is vibrational resonance, in which several vibration modes are merging into a single preference one with the risk of unbalanced forces self-amplification. The prognosis of such states in real-life is challenging, especially for complex and/or big scale machines, while their stochastic nature may reduce the remaining lifetime, disrupt functionality and increase the running costs. Recent advances in computational analysis propose early detection or even short-term forecasting of such operational regimes, however limited number of works address active damping of those, which is of a high industrial expectation. This work enriches this field, proposing an architecture that analyses complex vibrational data and damps resonances in real-time. Our particular focus is chatter marks a specific type of surface damage caused by the exposure of the contact surface by resonance induced strong periodical forces. Abrupt formations of chatter marks are reproduced on a twin disk tribometer, where two rotational axes are driven by independent servo motors with a tight line contact between both. The resonance cancelation is provided by a closed loop control that changes the axis relative rotation speeds with an objective to minimize the resonance frequencies. The feedback is provided by a supervisory trained classifier that predicts the intensity of resonance frequencies as a function of axes rotation speed, utilizing a mode decomposition of the measured vibrations as input. Our study investigates the applicability of optimal control and Bayesian optimization for real-time failure frequencies damping. Experimental tests show that damping control increases the remaining lifetime of the parts, preventing their severe damage even during chatter marks occurrence. The presented framework is generic, applicable to a broader number of industrial machines and operation conditions and scalable in temporal domain.