Sensor data can be mined to discover a description of an implicit property in an application domain. Main problems in the sensor data domain are the huge volume of noisy data and, therefore, the difficulty to locate the relevant information. KDD is the field where these problems are being addressed. This paper proposes a feature selection approach to discover relevant features, which will characterize the unknown property. The obtained description improves the expert knowledge and makes up a base for the prediction task. The feature selection task becomes intractable when the features set is large. Recurrent neural networks have shown to be very useful to obtain the global minimum of hard problems. This paper proposes two recurrent neural network models for feature selection: the graph model and the cluster model. Experiments show the advantage of the new methods, by comparing them with heuristic methods like RELIEF-F, DTM, and the Wrapper approach. We have selected databases from the UCI Repository and experimental data from steel machining processes. The new methods provide better solutions than the heuristic methods, and solve the tradeoff between accuracy and efficiency, as well as their generalization capability, which gives a solution of the noise problem.