The evolving digital oilfield offers new possibilities for data based approaches to pump monitoring. Electrical submersible pumps equipped with downhole sensors provide data which can be used for automatic malfunction detection. The proposed method of intelligent interpretation facilitates pump monitoring and integrates data more effectively. In this thesis, an approach based on data driven model builders was chosen to create neural networks which are capable of classifying ESP conditions and modelling operating parameters. In order to obtain a clean data set for machine learning, data cleansing including different types of filtering operations was applied. Techniques included: outlier removal, time stamp handling, logical and moving average filtering. The downhole sensor data in combination with surface data was used to model gross production rates during measurement gaps with an accuracy of 3 - 6 m3/d. Different ways of data manipulation and data arrangements are presented to overcome difficulties related to the observer effect when using surface separator measurements for production rate determination. Sensor data was also applied to generate realistic artificial data sets for failure classification. Real life examples of pump failures were then used to evaluate the capability of feed forward neural networks for failure classification. Individual and combinatorial training data sets were investigated to analyze sensitivities. Classification of pump failure with artificial neural networks can be carried out with an accuracy of greater than 80%. To conclude, a brief outlook for future research regarding failure classification and prediction is given.