You are building a model to predict whether a telecom network site will experience a service-impacting fault in the next 24 hours. You have access to operational counters, alarm history, site metadata, and recent maintenance activity, and the main challenge is turning these raw signals into useful model features.
How would you approach feature engineering for this machine learning problem, and how would you decide which engineered features are likely to improve model performance without introducing leakage or unnecessary complexity?