Breaking the Homogeneity Assumption: Specialized Multi-Generator Adversarial Learning for Rare Failure Detection in Predictive Maintenance