Abstract : Runtime verification of formal specifications provides the means to generate error detectors with detection capabilities depending mostly on the kind of formalism considered. The stronger the formalism is the easier the speration between correct and erroneous execution is. Nevertheless, two recurring issues have to be considered before using such error detection mechanisms. First, the cost, at run-time, of such error detector has to be assessed. Then, we have to ensure that the execution of such detectors has a well defined semantics. This thesis aims at better understanding the conditions of use of such detectors within critical real-time software application. Given formal behavioural specification, we defined the notion of "behavioural error". Then, we identify the class of early detectors that optimize the detection latency between the occurence of such errors and their signalling. The whole generation process has been implemented for specifications provided as timed automata. The prototype achieves early error detection thanks to a preprocessing of the automaton to generate its temporal abstraction. Our contributions are threefold :formalisation of early detection, algorithms for timed automata run-time verification, andprototyping of such detectors on a real-time kernel, Xenomai.