A variety of pumps are employed in all aspects of the oilfield life cycle, such as drilling, completion (including hydraulic fracturing), production, and intervention. There is no well-established method to monitor the pump fault states as they are operating based on sensor feedback. As a result, maintenance is performed either prematurely or reactively, both of which result in wasteful downtime and unnecessary expense. A machine learning based neural network model is used for identifying different fault states in a triplex pump from measured pressure sensor data. In a laboratory test, a triplex pump was seeded with eight different fault states (different combinations of damage in discharge valves, stuffing boxes, and suction valves). The time series data for suction and discharge pressures are recorded for 1500 different runs for a combination of eight different fault states and nominal condition. An efficient neural network model is trained on the observed data. This model is tested with multiple cross validation sets and is seen to have over 90% accuracy in predicting the correct fault class solely from the suction and discharge pressure data. In almost all cases, the model was able to correctly differentiate between the nominal response and response with different faults in the system. Different parameter transformation or feature engineering is performed to select optimal input features for the machine learning model. This work demonstrates that machine learning techniques can accurately predict different faults in pumps in operation from monitoring the suction and the discharge pressure. This work also demonstrates the importance using engineering domain knowledge for feature extraction and feature manipulation in creating insightful machine learning models.