One of the major problem for the identification of industrial multi-input multi-output (MIMO) systems is the too small variance of some signals collected during several experiments. Usually, sufficient information exist but are often dispersed among many data-files. In this case, it is difficult to conciliate this information to build a single well adapted model. The ALS procedure (Alternate Least Squares) has been performed in order to solve this kind of problem. In the first part of this paper, the authors present the principle of the ALS procedure; the details of the complete algorithm are given in part two. Then, based on some examples, a comparison of the Prediction Error Method (PEM) with our new method is presented.