Magda Ruiz Ordóñez

Ensemble learning as approach for pipeline condition assessment

The algorithms commonly used for damage condition monitoring present several drawbacks related to unbalanced data, optimal training requirements, low capability to manage feature diversity and low tolerance to errors. In this work, an approach based on ensemble learning is discussed as alternative to obtain more efficient diagnosis. The main advantage of ensemble learning is the use of several algorithms at the same time for a better proficiency. Thereby, combining simplest tree decision algorithms in bagging scheme, the accuracy of damage detection is improved. It takes advantage by combining prediction of preliminary algorithms based on regression models. The methodology is experimentally validated on a carbon steel pipe section, where mass adding conditions are studied as possible failures. Data from an active system based on piezoelectric sensors are stored and characterized through the T2 and Q statistical indexes. Then, they are the inputs to the ensemble learning. The proposed methodology allows determining the condition assessment and damage localizations in the structure. The results of the studied cases show the feasibility of ensemble learning for detecting occurrence of structural damages with successful results.

Leak detection and localization on hydrocarbon transportation lines by combining real-time transient model and multivariate statistical analysis

Safety and reliability of hydrocarbon transportation lines (pipelines) around the world represents a critical aspect for industry, operators and population. Lines failures caused by external agents, corrosion, inadequate designs, among others, generate impacts on population, environment, infrastructure and economy, besides it may be catastrophically. Therefore, it is essential to constantly monitor operating conditions and hydraulic lines to faults and thus to take measures to mitigate the failure. Localization of leakage is more than comparison between simulated and measured flows, from the dynamic of these flows it can be inferred the localization of the leakage, and even its magnitude. One option is to develop an inverse Transient Model (TM) able to calculate parameters of the pipeline by using the measured flow. However, if the calculation of flows is computational expensive, the inverse calculation is even more. These phenomenological models reproduce as closely the response (flow and pressure) of the pipeline. The simulation contains information to optimize the pumping rate, the momentum and energy including a high number of inputs and constraints to consider that growing exponentially with the level of detail to get in the pipeline. Therefore, this method has a high computational cost. The other option is to simulate several scenarios by using TM and train some kind of classifier or predictor with the simulated measurements. The first phase of our complete proposed methodology under development is presented in this work. We have focused on carrying out simulations of pressure along a pipeline using TM and applying Principal Component Analysis (PCA) as a tool to recognize hided patterns which allow classify leakages in different locations and different magnitudes. doi: 10.12783/SHM2015/292

Application of the PCA to guided ultrasonic waves to evaluate tensile stress in a solid rod

In this work, the use of Principal Component Analysis (PCA), through T2 and Q-statistics, is proposed to evaluate the tensile stress in a bar rod by means of ultrasonic guided waves. This method can be potentially used in-situ and real time for continuous condition monitoring. The specimen tested is a 1020 steel bar of 1” cold finished with yield strength of 441 MPa. Piezoelectric transducers are used in a through-transmission configuration to excite and detect axis-symmetric waves at the free ends of the bar. The load on the bar is performed using a MTS® testing machine in load control to maximum load of 80% of the yield strength. A dedicated Matlab® software is utilized to perform ultrasonic signal generation, signal acquisition and processing. A 5 cycles Gaussian-modulated sinusoidal pulse is used to generate narrow-band waves. Data captured by the piezoelectric sensor for each load condition are projected into the PCA model. The first two projections, Qstatistic and T2-statistic indices are analyzed. Results of each load scenario are presented and discussed demonstrating the feasibility and potential of using this formulation in the evaluation of the tensile stress in structural elements. doi: 10.12783/SHM2015/226