Change-point detection methods have been the subject of
intensive research investigations. Due to their sequential
nature and their adequacy to address online problems,
sequential probability ratio test (SPRT; Wald 1945),
cumulative sum (CUSUM; Page 1954), and generalized likelihood
ratio (GLR; Willsky & Jones 1976) are among the
most popular frequentist change-point detection methods.
The commonality of these widely used algorithms is that
the logarithm of the likelihood ratio between two consecutive
intervals, for the same time series, is monitored and a
change-point is declared whenever these intervals possess
different statistical properties. In spite of their widespread
success for practical applications, their major limitation
comes from their dependence on the knowledge of
probability distribution functions (pdf) of the data. Furthermore,
other approaches, such as spectral based methods
(Adak 1998), maximum likelihood estimation (Guralnik &
Srivastava 1999), and subspace identification (Katayama
2005) have also been explored. The limitation of all these
change-point detection methods is that they rely on prespecified
thresholds, which are difficult to establish a priori.
However, most of the Bayesian change-point detection
(BCPD) approaches have been used in an offline setting
for retrospective studies, which means that the entire data
set is required before computation of the probability of
change-point. To overcome these limitations, and those of
the frequentist methods, online BCPD (OBCPD)
approaches have been introduced by Adams & MacKay
(2007).
The purpose of this paper is to provide an efficient methodology
for water quality monitoring and contamination
event detection. We review a recent development in time
series forecasting and use it to forecast different water quality
time series. The performance of four change-point
detection algorithms is compared using residuals from the
forecast. Note that the change-point detection algorithms
examined here can be applied to residuals derived from
any predictive modeling algorithm. We place emphasis on
the smallest detectable contaminant, and the rate of false
and true positives, which are characterized by the receiver
operating characteristic (ROC) curve. Particularly, we
show that our methodology is able to distinguish normal
operation changes from contamination events. For each
change-point detection method, and a particular contamination
event, we provide the ROC curve. The remainder of
the paper is organized as follows. The following section,
‘Water quality time series prediction’, presents the time
series prediction method used to learn the parameters of
8 A. Ba & S. A. McKenna | Water quality monitoring with online change-point detection methods Journal of Hydroinformatics | 17.1 | 2015
the autoregressive (AR) model describing the water quality
data. Next is ‘Water quality time series monitoring’, which
introduces the frequentist approaches used in this paper,
namely SPRT and CUSUM as well as the binomial event discriminator
(BED) and the OBCPD. This is followed by
‘Application of change-point detection methods to water
quality data’ in which the experiments with water quality
time series are examined. The final section provides conclusions
and future work.