The concept of nonlinear modes has been proved useful to interpret a wide class of nonlinear phenomena in mechanical systems such as energy dependent vibrations and internal resonance. Although this concept was successfully applied to some small scale structures, the computational cost for large-scale nonlinear models remains an important issue that prevents the wider spread of this nonlinear analysis tool in industry. To address this challenge, in this paper, we describe an advanced adaptive reduced order modelling (ROM) technique to compute the damped nonlinear modes for a large scale nonlinear system with frictional interfaces. The principle of this new ROM technique is that it enables the nonlinear modes to be computed in a reduced self-adaptive modal subspace while maintaining similar accuracy to classical reduction techniques. The size of such self-adaptive subspace is only proportional to the number of active slipping nodes in friction interfaces leading to a significant reduction of computing time especially when the friction interface is in a micro-slip motion. The procedure of implementing this adaptive ROM into the computation of steady state damped nonlinear mode is presented. The case of an industrial-scale fan blade system with dovetail joints in aero-engines is studied. Damped nonlinear normal modes based on the concept of extended periodic motion is successfully calculated using the proposed adaptive ROM technique. A comparison between adaptive ROM with the classical Craig-Bampton method highlights the capability of the adaptive ROM to accurately capture the resonant frequency and modal damping ratio while achieving a speedup up to 120. The obtained nonlinear modes from adaptive ROM are also validated by comparing its synthesized forced response against the directly computed ones using Craig-Bampton (CB) method. The study further shows the reconstructed forced frequency response from damped nonlinear modes are able to accurately capture reference forced response over a wide range of excitation levels with the maximum error less than 1% at nearly zero computational cost.