Although efforts in this field are significant around the world, real prognostics systems are still scarce in industry. Indeed, it is hard to provide efficient approaches that are able to handle with the inherent uncertainty of prognostics and nobody is able to a priori ensure that an accurate prognostic model can be built. As for an example of remaining problems, consider datadriven prognostics approaches: how to ensure that a model will be able to face with inputs variation with respect to those ones that have been learned, how to ensure that a learned-model will face with unknown data, how to ensure convergence of algorithms, etc. In other words, robustness, reliability and applicability of a prognostic approach are still open areas. Following that, the aim of this paper is to address these challenges by proposing a new neural network (structure and algorithm) that enhances reliability of RUL estimates while improving applicability of the approach. Robustness, reliability and applicability aspects are first discussed and defined according to literature. On this basis, a new connexionist system is proposed for prognostics: the Improved-Extreme Learning machine (Imp-ELM). This neural network, based on complex activation functions, enables to reduce the influence of human choices and initial parameterization, while improving accuracy of estimates and speeding the learning phase. The whole proposition is illustrated by performing tests on a real industrial case of cutting tools from a Computer Numerical Control (CNC) machine. This is achieved by predicting tool condition (wear) in terms of remaining cuts successfully made. Thorough comparisons with adaptive neuro fuzzy inference system (ANFIS) and existing ELM algorithm are also given. Results show improved robustness, reliability and applicability performances.