Model structure selection based on polygonal curve approximation techniques

Model structure selection is crucial for many applications that are based on identification. This paper presents a selection technique based on polygonal curve approximation to pre-process step-response data, and on a neural network classifier. Only normalized I/O data are employed, so that the network can be trained off-line with simulated data. Model-specific parameterization techniques can be envisaged so that the actual implementation of the complete identification process is not computationally intensive, and its industrial usage (e.g., for regulator autotuning) is affordable. Simulations are reported to show the effectiveness of the proposed method