This article reports on results obtained when applying neural networks to the problem of vehicle classification from
SHM measurement data. It builds upon previous work which addressed the issue of reducing vast amounts of data
collected during an SHM process by storing only those events regarded as being "interesting," thus decreasing the stored
data to a manageable size. This capability is extended here by providing a means to group and classify these novel events
using artificial neural network (ANN) techniques. Two types of neural systems are investigated, the first one consists of
two neural layers employing both supervised and unsupervised learning. The second, which is an extension of the first,
has a data pre-processing stage. In this later system, input data presented to the system is first pre-scaled before being
presented to the first network layer. The scaling value is retained and later passed to the second layer as an extra input.
The results obtained for vehicle classification using these two methods showed a success rate of 60% and 90% for the
first and second ANN systems respectively.
One of the greatest challenges in deploying structural health monitoring (SHM) systems is the need to manage the continuous stream of measurements obtained from tens or hundreds of installed sensors. In a practical system the analysis of these measurements must be performed in an automated and robust manner and be completed in real-time. As the first stage in this process, a neural computing based novelty detection system has been developed which is capable of modelling the basic behaviour of a structure and subsequently isolating noteworthy measurements. In this article we examine the trade-off between the system's need to adapt to normal changes in a structures behaviour over the long-term, with the need to maintain a reliable reference model so as to identify important events when they occur. It is demonstrated that extending the existing basic neural processing system, by introducing a 'mixture of experts' approach, can address the contradictory needs of adaptability and model stability. In addition, it is shown that this approach provides a means of incorporating detection of both short-term and long-term phenomena into a single integrated processing system.
This paper explores the use of unsupervised neural computation for event detection (ED) in structural health monitoring (SHM) systems. ED techniques are useful in SHM systems for minimizing the size of SHM data sets, and the costs associated with analyzing, transmitting and storing SHM data. The approach to ED explored here is adaptive, self-configuring and does not require detailed information about the structure being monitored.
A neural network approach known as frequency sensitive competitive learning (FSCL) is used to model the sensor output of an SHM system. SHM system output states which disagree with the model learned are deemed "novel" and detected as SHM events.
The FSCL-ED system is evaluated with SHM data from three structures including the Taylor Bridge, the Portage Creek Bridge and the Golden Boy Statue. Furthermore, this system is able to identify strain gauge events of 0.75, 12.5, 1.25 microstrain or smaller in the SHM measurement data from the Taylor Bridge, the Portage Creek Bridge, and the Golden Boy respectively. The FSCL-ED system is able to identify accelerometer events of .0045g, 0.0020g or smaller in the SHM measurement data from the Portage Creek Bridge, and the Golden Boy respectively.
The FSCL-ED system is compared to a simplified event detection (S-ED) system, which does not use power spectral density estimation or unsupervised neural computation. The S-ED system is shown to be effective but less sensitive than the FSCL-ED system to SHM events. As well, the FSCL-ED system is better able to adapt to noisy environments.
This paper explores the use of unsupervised neural networks and frequency sensitive competitive learning for novel event identification in structural health monitoring (SHM) systems. Our approach assigns a novelty metric based upon the output states of an SHM system. The technique can be applied in data decimation schemes, to enhance the monitoring of such systems, and as an aide to SHM data analysis. Learning units provide a means of characterizing an SHM system, and are subsequently used to assign a novelty metric to new SHM data. The system has been evaluated using data from the Taylor Bridge and Golden Boy statue in Winnipeg, Canada and the Portage Creek bridge in Victoria, Canada. The system is capable of analyzing SHM data from a 14-channel system, recording data at 32 Hz, using 32 learning units at approximately 30 times real-time on an AMD AthlonXP 2500+ based computer. The event identification system is most sensitive to SHM data which exhibits unusual power spectra, including data which shows abrupt changes in sensor outputs. The system may be cascaded in order to perform basic classification of events after identification.