In this study, we propose a knowledge-based approach for detection and isolation of predefined and nonpredefined sensor faults in fault tolerant control (FTC) of a three tank system. Farthest first traversal algorithm (FFTA) of data mining is used for the first time for the classification of faults in a FTC system. Predefining here means that features of a fault and its effects are known before the fault is seen on the system. Therefore, if a predefined fault is detected on the system, it is isolated into a known fault cluster and predefined action for that cluster can be taken to tolerate the fault. However, in a working system, there may be some other faults, which are not predefined. Those may be inaccurately isolated into available known clusters, since the clusters are determined according to predefined faults instead of non predefined ones. In our work, we also propose a method for isolating the non predefined faults that rearranges the clusters of predefined faults, online. In order to show the efficiency of proposed method, seven predefined and thirteen non predefined fault scenarios are applied to a closed loop FTC system. While three of the non-predefined faults are not accurately isolated without the proposed method, all of the faults are isolated correctly with the proposed method.