This paper presents a vibration–based, observer–driven fault detection scheme for rotary–percussive drilling that is both physics–grounded and robust to operating–regime variability. A compact axial–torsional plant with single–cutter bit–rock interaction captures loading/unloading asymmetry and frictional torque coupling. To compensate for salient nonlinearities and premise uncertainty (e.g., intermittent contact, rate effects), we embed the dynamics in an interval type–2 Takagi–Sugeno (IT2–T–S) fuzzy framework with explicitly defined IF–THEN rules and type reduction, yielding a convex blend of local linear models suitable for analysis and synthesis. An adaptive Luenberger observer is then designed to (i) reconstruct the nominal vibration response, (ii) generate a residual sensitive to faults yet tolerant to modelling errors and measurement noise, and (iii) deliver an online estimate of an unknown axial fault input. A Lyapunov function with vertex LMI conditions guarantees exponential convergence in the fault–free case and uniform ultimate boundedness under bounded faults; the fault–estimation update law is derived to ensure closed–loop stability. Simulations with percussion–style axial forcing demonstrate three key outcomes on a short time horizon: residuals remain within a noise–based threshold pre–fault and cross the band at the fault onset; the estimated fault rapidly converges to the true magnitude with negligible steady bias; and the state–error norm decays quickly pre–fault and exhibits a bounded transient post–fault. The results indicate that the proposed IT2–T–S adaptive observer provides an implementation–ready path to reliable, vibration–based fault detection for drilling systems. The paper concludes with recommendations to migrate to higher–order fuzzy consequents (polynomial/type–2) to further reduce approximation error and tighten residuals in strongly impacting regimes.