Fault detection in complex plants has to cope with substantial problems due to the very large data amount. In many cases, adequate plant descriptions are not available, so that models has to be built up on line. To achieve this in a sensible time, data have to be sorted and this almost always leads to an information compression. While this proves very helpful to detect faults, it represents a serious obstacle for the identification of the faulty channel, as the existing partial models do not usually span a full measurement space or do it with a very poor condition. This paper proposes to use a double technique to achieve this end, first improving the fault isolation process through a gradient based method, but then recurring to model-on-demand methods which can be used to complete the required measurement space to yield the precise fault channel information.