Spectra of wavelet scale coefficients from process acoustic measurements as input for PLS modelling of pulp quality

Acoustic and vibration signals are captured by simple standard accelerometers. These can often be mounted directly on operative process equipment, creating a completely non‐invasive measurement system. The signals from the accelerometer are then amplified, digitized by an analogue‐to‐digital converter and stored in some suitable format in a PC. The method most often used for signal processing of acoustic data has been to apply variants of fast Fourier transform (FFT) on sampled data to produce a frequency domain representation. An alternative way tried here is to use the fast wavelet transform (FWT) in combination with FFT. The FWT has the advantage that it produces time‐resolved representations and, on each time scale, different features can be extracted. However, in this case, time resolution has no meaning, since the starting points for data acquisitions were not fixed. The wavelet step can be seen as a series of pre‐filters and it is here followed by FFT on coefficients at each wavelet scale. The results are compared to those obtained after FFT on the complete time series. We have used spectra of wavelet scale coefficients in an attempt to model pulp quality with PLS. In this case the number of points in the resulting wavelet multiresolution spectrum (WT‐MRS) can be limited to a low number, e.g. 255 compared to 1025 with direct FFT on the time series. In the PLS modelling step the advantage is that the first two components describe Y much better than when using the conventional approach, e.g. 72% explained Y variance compared to 40%. A second advantage is that the model requires fewer coefficients. Copyright © 2002 John Wiley & Sons, Ltd.