Acoustic chemometrics of complex natural systems: flume test rig feasibility studies for monitoring river bed‐load sediment transportation processes

Abstract

AbstractA feasibility study has been carried out on a meso‐scale experimental flume rig with the objective of developing a robust acoustic chemometrics facility of bed‐load mass flux quantification for deployment in rivers and hydroelectric power plant inlet canal systems. The test rig is equipped with an acoustic sensor attached to a steel plate flush with the flume bottom, reproducing natural sedimentary river bed‐load transportation characteristics. Sensor signals were preprocessed by the Fast Fourier Transform (FFT) and subjected to multivariate calibration (PLS1 regression), enabling prediction of sediment bed‐load transport interacting with varying sedimentary characteristics. Based on a comprehensive experimental design on two major factors (F1: mass flux; F2: grain‐size distribution), it is concluded that reliable, test set validated PLS‐prediction models can be established for bed‐load mass fluxes with effective compensation for widely varying sediment size. The main models developed display average relative prediction errors (RMSEP) spanning 4.8–13.2%; the highest errors reflect scenarios modelling for seven non‐overlapping size fraction ranges simultaneously. All real‐world river implementations will experience significantly less varied sediment variability than this worst case scenario, thus realistic models will be characterised by relative RMSEP of some 10% or lower, highly satisfactory for such complex natural systems. All results, interpretations and conclusions reached in the present study are based on the most stringent test set validation protocol. The specific acoustic impact on the sensor plate by sediments characterised by different grain‐size distributions must be expected to lead to significant interaction with the mass flux signals. The experimental design was therefore deployed as a full two‐factor design (7 and 8 levels, respectively), totaling 2 × 56 runs (calibration and test set). Because of a very high sensitivity to the specific time‐varying transportation regimen over the river bottom (successfully duplicated on the flume rig bottom), conventional cross‐validation of a single training set model predicting sediment mass flux runs a high risk of being over‐optimistic, unable to produce a realistic prediction performance assessment. Only true test set validation—an independent second set of all runs in the design covering the same experimental space as the training data—can substantiate a realiable, scientifically acceptable prediction performance validation. This setup allows practical evaluation of the relative merits of cross‐validation versus test set validation in situations in which there is a significant complement of sampling, analysis and measurement errors in both [X,Y] data. Copyright © 2007 John Wiley & Sons, Ltd.