ENBIS-17 in Naples
9 – 14 September 2017; Naples (Italy)
Abstract submission: 21 November 2016 – 10 May 2017
Maintenance Policies Countering Degradation of Water Supply Networks: Statistical Analysis with a Semi-Markov Model and Panel Data
12 September 2017, 09:20 – 09:40
- Submitted by
- Vincent Couallier
- Vincent Couallier (Institute of Mathematics of Bordeaux), Cyril Leclerc (SUEZ Eau France, le LyRE), Yves Legat (IRSTEA), Karim Claudio (Cetaqua (SUEZ))
- This work addresses the problem of the statistical modeling of maintenance and failure data of water supply networks. Concerning the pipes of the network, empirical studies have shown that a failure is the end-point of a degradation process that may be described by a continuous time multi-state process. Four states are relevant from an engineering point of view:
- State 0 : the pipe is in a “leakage-free” state.
- State 1: Background leakage - initialization of the leak, invisible at the surface with a low and undetectable flow,
- State 2: Detectable leakage - still invisible at the surface, the leak can be detected by acoustic inspection,
- State 3: Visible burst -the leak appears at the surface and must be repaired.
The maintenance lead by water operators includes campaigns of acoustic inspections. These operations yield partial observations of a continuous-time degradation process. In fact only the last state (visible bursts) is exhaustively registered. In such a case, multi-state models using panel data are adequate to model the leakage degradation process. Indeed, the Semi-Markov model with panel data offers an alternative to standard survival analysis of interval censored lifetime data. We show in this work how to fit such a model to data collected since 2010 by the Bordeaux water utility. The network contains more than 30000 pipes, each of them with leak detections and bursts data, as well as characteristics like material, length, diameter, pressure, soil corrosivity, which are used as model covariates for the transition intensities.
Return to programme