Multi-Leak Diagnosis in Pipelines – A Comparison of Approaches C. Verde Instituto de Ingenier´ ıa UNAM Coyoac´ an DF, M´ exico verde@servidor.unam.mx R. Morales-Menendez, L.E. Garza, Research and Graduate Program Tecnol´ ogico de Monterrey Monterrey NL, M´ exico {rmm, legarza}@itesm.mx A. Vargas, P. Velasquez-Roug, C. Rea, C.T. Aparicio, J.O. De la Fuente MSc Students Tecnol´ ogico de Monterrey Monterrey NL M´ exico {A00777924,A00362720,A00797489,A00952040,A00608453}@itesm.mx Abstract Leaks on pipelines can cause strong economic losses and environmental problems if these are not detected on time. The problem of detecting leaks is even more complicated when the pipelines are too large, difficult to reach by main- tenance personnel, and equipped with minimum instrumen- tation. A comparison of four fault diagnosis approaches based on Output Observers, Artificial Neural Networks, Particle Filtering and Principal Components Analysis are presented. Simulated results of multi-leaks in pipelines showed that Particle Filtering techniques outperform the other approaches. However, a combined solution is pro- posed. 1 Introduction The leak diagnosis is a challenge for the engineering sci- ences, since the economic losses and environmental dam- age produced by a deteriorated network are very significant [23]. Thus, different point of views, tools and backgrounds have been explored to implement monitoring systems for aqueducts, oil lines, networks of pipelines etc. As Wang pointed out [31]: no single method can always meet all the requirements and each technique has its advantages and disadvantages in different circumstances. An overview of diverse diagnosis tools for pipelines considering hardware and software tools is presented in [31], [18] and [21]. In the case of distribution networks, genetic algorithms [27] and wavelet analysis [1] have been used to detect leaks position. The community of control has developed a general framework for the fault diagnosis of dynamic systems which allows the generation of fault symptoms, called resid- uals, by software [25] and [17]. Diverse mathematical model-based procedures have been developed in particular for leak diagnosis. Billman [2], Shields [26], Korbicz [18] and [13] designed residual generators using a finite dimen- sion model and assuming fix space discretization in the set of partial differential equations which describes the fluid be- havior. However, these methods can be only applied to de- tect leaks in limited cases. Assuming sequential leaks and scanning the pipeline with an adaptive law to estimate the leak position is a novel approach given in [29]. This proce- dure is successful if a new leak appears after the previous one has been located. However if a new leak occurs before the previous one is located the adaptive algorithm fails. It is important to note that the location methods based on steady state conditions without test signals (for instance [9]) have a common property: if they are applied for the multiple leaks case, all of them delivered wrong location of the faults. In [19] a proposal is made to capture the leaks patterns with the frequency response diagram considering a valve perturbation in the line. Ferrante [12] presented a method based on transient response using the Fourier Transform of pressure signals. However, all these methods have practical limitations if dominant nonlinear effects are presented in the fluid and a test signal is required. Brunone [3] reported a method based on unsteady state tests with the capacity to detect two leaks, assuming one leak is presented after the other. 2008 Seventh Mexican International Conference on Artificial Intelligence 978-0-7695-3441-1/08 $25.00 © 2008 IEEE DOI 10.1109/MICAI.2008.33 352