Reliability of adaptive multivariate software sensors for sewer water quality monitoring

This study investigates the use of a multivariate approach, based on Principal Component Analysis PCA), as software sensor for fault detection and reconstruction of missing measurements in on-line monitoring of sewer water quality. The analysis was carried out on a 16-months dataset of five commonly available on-line measurements (flow, turbidity, ammonia, conductivity and temperature).

The results confirmed the great performance of PCA (up to 10 weeks after parameter estimation) when estimating a measurement from the combination of the remaining four variables, a useful feature in data validation. However, the study also showed a dramatic drop in predictive capability of the software sensor when used for reconstructing missing values, with performance quickly deteriorating after 1 week since parameter estimation.

The software sensor provided better results when used to estimate pollutants mainly originated from wastewater sources (such as ammonia) than when used for pollutants affected by several processes (such as TSS). Overall, this study provides a first insight in the application of multivariate methods for software sensors, highlighting drawback and potential development areas.

A combination of (i) advanced methods for on-line data validation, (ii) frequent parameter estimation, and (iii) automatic method for classification of dry/wet periods may provide the needed background for a successful application of these software sensors.