Being in the category of data driven approaches, both adaptive neuro-fuzzy inference system (ANFIS) and principal component analysis (PCA) have been widely used in literature for fault detection and isolation when the whole things that we know about and have from the systems are some measurements corrupted by noise. In spite of promising applications of both methods, it is an unanswered question that which method must be considered as the first option when there is a possibility of designing and implementing fault detection systems using both methods. In this research work, we implement these methods over an unknown nonlinear system and assess performance of each method for detecting small plant component faults. In order to find the best arrangements of inputs and outputs for creating ANFIS and PCA models, different possibilities are examined. Simulation results for different cases have been presented in the paper and those clearly suggest that PCA method is generally more reliable for fault detection and more robust to measuring noise than ANFIS.