Column floatation processes are multivariable, complex and difficult to model systems. Usually linear single input/single output (SISO) models are derived from experimental data and process analysis, using prior knowledge of the process. However, it is highly preferable to use multi-input/multi-output (MIMO) models, which are much more accurate. However, this type of models is much more difficult to identify. This paper proposes a methodology to automatically identify a multi-input/multi-output (MIMO) model obtained from experimental data, using a fuzzy modelling strategy. The process has four manipulating variables: feed flow rate, which in normal industrial operation is usually kept constant, washing water, air and rejected flow rates. The outputs of this model, which are normally used to control the grade and the recovery in the flotation column, are the froth layer height, the bias flow rate and the air holdup in the collection zone. By using the regularity criterion, it was possible to determine the structure of the MIMO model without any prior knowledge of the process dynamic. The final model is validated using different experimental data. The experimental data was acquired in a pilot scale laboratory flotation column of 3.2m height by 80mm of diameter. (c) 2004 Elsevier Ltd. All rights reserved.