Fault prediction is a significant issue for ensuring industrial process safety and reliability. In practical processes, due to complexity and non-linearity, this leads to many difficulties for process fault prediction. Aiming to improve the fault prediction accuracy in procedure-oriented systems, a new feedback differential evolution-optimized extreme learning machine (FDE-ELM) with a time delay-based extended finite state machine (TD-EFSM) approach is proposed. The proposed method is exemplified in the complicated Tennessee Eastman (TE) benchmark process. The results show that the new joint time-delay EFSM-based FDE-ELM shows superiority not only in modelling stability but also in detection sensitivity.