Monitoring algorithms for detecting changes in the ozone concentrations

Abstract

The quality of data collected by air pollution monitoring networks is often affected by inaccuracies and missing data problems, mainly due to breakdowns and/or biases of the measurement instruments. In this paper we propose a statistical method to detect, as soon as possible, biases in the measurement devices, in order to improve the quality of collected data on line. The technique is based on the joint use of stochastic modelling and statistical process control algorithms. This methodology is applied to the mean hourly ozone concentrations recorded from one monitoring site of the Bologna urban area network. We set up the monitoring algorithm through Monte Carlo simulations in such a way to detect anomalies in the data within a reasonable delay. The results show several out of control signals that may be caused by problems in the measurement device. Copyright © 2000 John Wiley & Sons, Ltd.

ESTIMATION OFCpFOR AUTOCORRELATED DATA AND MEASUREMENT ERRORS

“…4. State space models are often used to model dynamic systems which describe manufacturing processes and suitable formulations are useful to monitor measurement devices (Basseville and Nikoforv, [12] Bordignon and Scagliarini [13] ).…”

Inference on C pk for autocorrelated data in the presence of random measurement errors

“…State space models are often used to model dynamic systems that describe manufacturing processes, and suitable formulations are useful to monitor measurement devices [1,2,14].…”

Design of change detection algorithms based on the generalized likelihood ratio test

“…The data set consists of the mean hourly average of the ozone, O 3 , and hourly nitrogen dioxide concentrations fNO 2 n g and hourly average temperature values fT n g, (see Bordignon and Scagliarini, 2000). The data were measured at one monitoring site of the Bologna urban area network, from June 1993 to December 1996.…”

“…As Lai (1995) pointed out, performance of scheme (2) is sensitive to the choice of M and of the other design parameter h. Both values can be chosen on the basis of asymptotical considerations (Lorden, 1971;Lai, 1995). An alternative procedure is to derive a convenient value of the threshold h, conditionally to a ®xed value of M (Bordignon and Scagliarini, 2000). Unfortunally, a satisfactory choice of both the window size and the threshold is very dif®cult in a ®nite context.…”