Effective equipment fault diagnosis can assist to schedule the proper maintenance and reduce breakdown risks for realistic engineering systems. In this paper, a novel two-step precognitive maintenance framework is proposed to diagnose the equipment health conditions based on its real-time Condition Monitoring (CM). The synthetic minority over-sampling technique is implemented firstly to balance a raw CM dataset for training two independent extreme learning machines. Next, our proposed framework will consist of two steps for the fault diagnosis, where Step 1 aims to classify the normal or abnormal categories, and Step 2 further diagnoses the underlying anomaly in details. The effectiveness of our proposed framework is testified on a machine fault simulator with an imbalanced dataset, and it achieves the diagnosis accuracies of more than 97.0%.