https://www.mdu.se/

Measurements of the dissolved oxygen (DO) concentration are central in aeration control strategies at Water Resource Recovery Facilities (WRRFs). Despite this, more research has been focused on DO-control strategies, see e.g. (Amand et al. 2013), than on fault detection (FD) methods to for DO-measurements. One FD-method was proposed in (Carlsson & Zambrano 2016), where the ratios between airflows at different aerated zones were monitored to detect bias in DO-sensors. However, this was argued to be inadequate to distinguish large process disturbances from sensor bias. It is a general problem that process disturbances are hard or impossible to separate from sensor faults.

One approach that potentially could be used to distinguish between sensor and process fault is active fault detection, in contrast to traditional or passive fault detection. In active fault detection an auxiliary signal is designed exclusively for fault detection and injected into the system (Esna Ashari et al. 2012). In this paper, we used the impulse from an automatic air cleaning system of the DO-sensor as design signal, and monitored the impulse response, see Figure 1 for an example. A similar approach was suggested already in 1992 (Spanjers & Olsson 1992), where a changed time constant of the DO-sensor was shown to be a good indication of an artificially fouled DO-sensor. More recently, Andersson and Hallgren showed that the impulse response from an air-cleaning procedure could be used to detect a biologically fouled DO-sensor (Andersson S. & Hallgren F. 2015).

However, none of the previous studies made repeated experiments of fouled versus cleaned sensors in order to characterize the variation between the impulse responses. This is needed to compare different FD-methods and their performance to distinguish fouled (faulty) from clean (normal) impulse responses.

In this paper we made detailed experiments with artificial fouling and used the results to compare two fault detection methods, Rise time estimation (RTE), and Gaussian process regression (GPR) (Rasmussen & Williams 2005).