You are building a binary classifier to identify rare fault events from spacecraft telemetry and subsystem health data. The operations team wants a model that can flag likely faults early enough to support triage, but true fault cases are much less common than normal operating behavior.
How would you approach training and evaluating this model so that the class imbalance does not produce a system that looks accurate but misses the events that matter? Explain how you would choose the model, handle the skewed labels, and decide whether the model is good enough for operational use.