Spectrum analysis of mixing power curves for neural network prediction of dough rheological properties

Two spectrum analysis techniques, namely Fast Fourier Transform (FFT) and power spectral density (PSD), were used for data preprocessing as part of development of neural networks for prediction of the rheological properties of a cookie dough. The raw data were extracted from the mixing power consumption curves acquired during mixing of the dough. The rheological properties studied were farinograph peak, extensibility and maximum resistance. Two types of neural networks, namely back propagation (BP) and general regression neural network (GRNN) with different architectures were developed for this study. The prediction accuracy of these networks when trained with the raw data, or FFT or PSD treated data was evaluated. The results indicate that the FFT and PSD treated data retained most of the characteristics of the raw data but with reduced noise and size. The performance of the networks trained with FFT or PSD treated data was found to be comparable with that of the networks trained with the raw data. For the BP network, the FFT treatment slightly improved the network performance while the PSD treatment resulted in a slightly higher but acceptable APE. It was noted that the farinograph peak and extensibility were better predicted than the maximum resistance by the neural network techniques. The prediction of maximum resistance was greatly improved by using a GRNN trained with the FFT or PSD treated data. It is therefore concluded that the spectrum analysis techniques used in the study can greatly improve the efficiency of the neural networks for the prediction of dough rheology while preserve prediction accuracy.