Fault detection and classifications using supervised learning algorithms are widely studied; however, lesser attention is given to fault detection using unsupervised learning. This work focused on the integration of unsupervised learning with cognitive modelling to detect and diagnose unknown fault conditions. It is achieved by integrating two techniques: (i) incremental one class algorithm to identify anomaly condition and introduce a new state of fault to the current fault states if an unknown fault occurs, and (ii) dynamic shallow neural network to learn and classify the fault state. The proposed framework is applied to the well-known Tennessee Eastman process and achieved significantly better results compared to results reported by earlier studies. Laboratory experiments are also performed using a pilot-scale system to test the validity of the approach. The results confirm the proposed framework as an effective way to detect and classify known and unknown faults in process operations. © 2019