Keywords: fault detection, shear buildings, neural networks, earthquake, hysteresis.

Health monitoring of civil structures during severe seismic events is an active field of research. Civil structures, such as buildings and bridges are subjected to base accelerations, and in the case of severe motions, may show cracking or yielding in the resistant elements, even collapse. It is worth noting that on-line or near-real-time algorithms may be useful for structural control applications, simulations, decision making strategies, and fault detection. Consider the following example: suppose the case of telephonic centres with on-line structural health monitoring. Suppose that during a severe earthquake event, real-time algorithms could detect a given fault in that particular building, and a re-routing strategy of the telecommunication lines to another centre could be implemented, without the loss of communication signals for users and emergency systems. That is why it is important to find out if a building was subjected to damage in real-time. In this way, an alarm could be triggered, warning the people inside the building, the owner, the authorities, and the emergency services.

Fault detection is the first stage of damage assessment. In this research, the process in this stage is to rule out the presence of a fault during seismic events. That is, the structure is assumed to remain in the linear range throughout the entire earthquake motion. If the structure undergoes some sort of damage (cracking, yielding, hysteretic, etc.), the methodology should be able to determine the existence of a fault. The present research proposes the use of a neural network to forecast, on-line, the behaviour of a structure subjected to base excitation, trained only with the first part of a seismic acceleration record. This window could give the parameters of a linear structure, and in the case of fault or damage, the output of the network will be different to that of the sensor output, giving rise to fault detection.

The present research shows the use of a neural network and the back-propagation learning algorithm, for fault detection of shear buildings. The building considered in this research is a single degree of freedom system (SDOF), excited with the Loma Prieta acceleration record, registered at Treasure Island at Santa Cruz station (1989). The base acceleration record contains 2000 samples at 50 samples per second.

In order to observe the fault detection analysis, a reference simulation was performed, in other words, no fault was present during the training and forecasting of the displacement output, and the structure was set to be linear elastic throughout the entire motion.

Later on, two different types of faults were induced to the SDOF in the simulations. Firstly, the stiffness of the SDOF was reduced to half the initial stiffness, assuming severe sudden damage, during the intense part of the earthquake. The neural network was trained with the first 12 seconds of the displacement response, and the weights were used to forecast the response of the linear structure. If the structure remains linear during the earthquake, the forecast should be very close to that of the theoretical behaviour. Nonetheless, if the stiffness of the structure is reduced, the responses would be different and a discrepancy could be observed between the forecast and the real displacements (fault detection).

Secondly, a non-linear SDOF was considered. The Bouc-Wen differential model was used to simulate the hysteretic component. Two features of this model are that it has the ability to represent several hysteretic shapes and it correlates well with laboratory tests. Again, the neural network was trained with the first 12 seconds of the displacement response. In this case, it could be noted that the training time showed an almost linear behaviour (mild non-linearity). Afterwards, the hysteretic loops became wider and a stronger non-linear relationship between the displacement and the stiffness is observed. Then, the weights were used to forecast the response of that identified structure. Since the behaviour was non-linear, the forecast gave a response very different to that of the theoretical output, giving rise again to fault detection.

The aim of the paper is to develop a technique that could be used in fault detection analysis and structural health monitoring. The data acquired during the first 12 seconds of an earthquake motion could be used to train the network, later, the artificial neural network with the base acceleration data as input, could predict the responses and compare them to the real on-line data. Therefore, some criteria as to whether the structure behaved differently to that of a linear elastic structure could be established. The implementation of the proposed algorithm could be done using digital processors or analogue electronics. The present research is also aimed to the implementation of fault detection algorithms on a partially instrumented building in Mexico City.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £80 +P&P)