A high energy collider ring needs a full network of sensors to operate. In particular, beam position monitors (BPMs) provide the transverse position of the beam in the chamber at multiple locations around the ring, and one category of BPMs, the Turn-by-turn BPMs (TbT-BPMs), is used to detect any magnetic defects in the lattice. Several methods exist to reconstruct this lattice and the associated optical functions, but each of them needs to have several functional TbT-BPMs dispersed all along the ring. In the context of the FCC, this involves several thousand BPMs scattered along a 91-km ring, operating in a very adverse environment, especially due to the effects of radiation on the electronics. Moreover, in order to maximize the duty cycle in this very large-scale accelerator, operation may have to occur even when some of its sensors are down. Consequently, the fast detection of faulty TbT-BPMs and the adaptation of the optics functions reconstruction are paramount, as is having sufficient confidence in the actual measurement, which needs to be sensitive and precise enough (i.e., having a high signal-to-noise ratio) to support decisions on the correction of the behavior of the multi-GeV beam. Indeed, the ability to function with very high precision in a very noisy environment is also a challenge that SuperKEKB, the largest e+/e- collider currently in operation, but also future colliders and light sources, need in order to measure rapidly and with high precision the optics functions. We present several machine-learning methods, adapted to the accelerator context, for detecting faulty TbT BPMs and for denoising the tracks in SuperKEKB, with the aim of identifying an approach that is efficient and scalable to the scale of the FCC accelerator.