TRIDENT: Three-Lever Privileged Self-Distillation for Audio-Visual Source Separation in Flow-Chemistry Pilot Plants

Its weakness is that one fume-hood bay runs several unit operations at once, so the pump, the switching valve, the gas bubbler and the agitator arrive as a single mixed channel; the loudest of them owns most of the acoustic energy, and the components that carry the earliest fault signatures are exactly the quiet ones. We cast this as audio-visual source separation in which the viewing window supplies the second modality and the supervision is privileged: single-unit calibration recordings and 15-min at-line HPLC samples exist during training and never at deployment. Dense privileged supervision then reproduces a failure mode already documented for language models: the separator collapses onto the dominant source, and the weak stream is silently abandoned. We treat that collapse as a measurable symptom and govern it with three levers: where the signal is applied (per time-frequency band credit rather than per sample), what the teacher is shown (a ladder of privileged views admitted only after probe-based verification), and when the signal changes (guidance scheduled by a reaction-phase posterior with alarm-triggered refund). On mixtures built from pump, valve, fan and slide-rail recordings, the resulting separator raises weak-source SI-SNRi from 3.9 to 6.4 dB relative to a contribution-balanced baseline, cuts the collapse index from 0.48 to 0.21, and lifts downstream symptom detection from 0.81 to 0.85 AUC, with a distillation budget of 29% of all time-frequency bins.