Self-Supervised Learning for Data Scarcity in a Fatigue Damage Prognostic Problem