Failures on control valves often stop a part of the process or shutdown the whole plant if it occurs on critical systems. The detection of these failures is often dependent on operator supervision and experience. Thereby, the topic studies methods to detect fault on control valves, in order to warn the technical unit early to have timely corrective and preventive maintenance methods. The study topic is to build a structure to utilized the fault that occur in the Electro-Pneumatic Control Valve object, which is valve Fisher using positioner DVC6200 on experimental KIT of Control System at the Fertilizer Ca Mau Plant, with two typical fault on the control valve are instrument air leakage (fail air) and mechanical jam due to friction (fail friction). Valve operating parameters are collected by DCS system and connected to MATLAB/SIMULINK via OPC communication. Along with that, the study project also proposes three methods to detect fault on control valves proposed are: algorithm using valve linear characteristic, algorithm using model reference valve and algorithm using machine learning. In particular, the algorithms not only stop at simulation verification but also applied in practice, by running real-time experiments on the built model to evaluate the detection ability. Early failure in upper control valve object. The experimental results show the average fault detection accuracy of algorithms linear characteristic is the highest (75.45%), but it is not possible to classify specific fault, and algorithm machine learning not only has high accuracy (74.20 %), but also classify specific fault.