A novel method of process dead-time identification: support vector machine approach

Performance and robustness of model-based control system are sensitive to the modeling error, especially to the dead-time identification error. Support vector machine (SVM) employs structure risk minimization principle to control model complexity and the upper bound of generalization risk. If the seeking dead-time contained in training data equals dead time of actual plant, the trained SVM has the lowest complexity. The identification procedure is described as follows. Firstly, specify a dead-time seeking range based on the prior process knowledge. Secondly, construct training data sets from input-output data according to different dead times in seeking range and train SVMs respectively. Finally, the estimated dead-time can be obtained through comparing the numbers of support vectors of all trained SVMs. A lot of discrete simulations for the first order plus dead-time system have been done to illuminate the effectiveness of proposed method.