Real-time Structural Health Monitoring Using Machine Learning Algorithm

Abstract

The 2010-2012 Canterbury earthquakes, the 2013 Cook Strait earthquakes and the 2016 Kaikoura earthquake highlighted that damaging earthquake scenarios are very real and they can occur at anytime and anywhere in New Zealand. It is desirable to ascertain and track the structural performance and integrity for buildings for public safety and emergency management. Seismic instrumentation has often been promoted as the solution to this requirement. Real-time Structural Health Monitoring (SHM) algorithms in a seismic instrumentation system continuously translate structural monitoring data into building state prediction. The overall goal of this study was to evaluate and improve the current damage detection algorithms implemented in civil structures, which made real-time SHM possible. The aim was to improve occupant safety and accelerate the rate of recovery for individual buildings, and by extension improved community resilience to extreme events. A key issue influencing the performance of the current algorithms is varying operational and environmental conditions. This project studied a year's worth of instrumented building data from the GNS Science building in Lower Hutt, with the aim of evaluating the influence of different environmental conditions on predicted building motion and corresponding dynamic characteristics, including building acceleration amplitudes and modal frequencies. The environmental conditions considered were temperature, wind speed, relative humidity and human activity. The results of the analysis demonstrated operational and environmental conditions have a noticeable effect on building dynamic properties estimation. In 2016, the CentrePort BNZ building suffered severe damage to its structural members and its non-structural members as a result of the 2016 Mw 7.8 Kaikoura earthquake. Coincidently, extensive non-structural damage also occurred in this building in the 2013 Cook Strait earthquakes. The recorded earthquake response of the CentrePort BNZ building during the two earthquakes was analysed to gain a comprehensive understanding of the building dynamic responses during earthquakes. A damage detection algorithm using autoregressive (AR) models with Mahalanobis squared distance (MSD) was applied to the instrumented building data for two years of data. It successfully detected the change in building damage state correlating to actual observations due to the two earthquakes. A parametric study was conducted to consider the effect of AR orders and exceedance probability on an optimum threshold for signalling damage in MSD-based damage detection algorithms. The results indicated that both factors are important parameters affecting the - ii - detection accuracy of the algorithm. Besides, this algorithm detected damage accurately under varying operational and environmental effects. A new damage detection method called the MSDAANN algorithm was proposed based on combining MSD and auto-associative neural network (AANN) approaches. The performance of the proposed algorithm was evaluated using data from the ASCE benchmark structure and through an analysis of receiver operating curves (ROCs). The results showed that the proposed MSD-AANN algorithm performed better than MSD-based algorithm or AANN-based algorithm. In addition, proposed MSD-AANN algorithm is selfcalibrated, it can be automatically applied to datasets and obtained damage detection results in a very short time with high accuracy even when the structure was operating under varying operational and environmental conditions. These represent improvements beyond current solutions and present great potential for real-time SHM applications. Microsoft Azure Machine Learning Studio (MLS) is a powerful machine learning tool, with which data scientists and developers can quickly build, test, and develop predictive models using state-of-the-art machine learning algorithms. But the effective application of machine learning algorithms in SHM applications remains a challenge for researchers. A parametric study of a cloud-based machine learning damage detection algorithm using two-class boosted decision tree was therefore conducted to investigate the effects of input length and the number of sensors on damage detection accuracy of a cloud-based machine learning algorithm. To facilitate a comparison, an MSD-based damage detection algorithm was also applied to the same data sets. The parametric study showed that both input data length and sensor numbers greatly affected the damage detection accuracy. The detection accuracy of both cloud-based machine learning and MSD-based algorithms increased when more data was used. More data in this instance means greater length of input data or longer time-duration preceding a prediction. Cloud-based machine learning algorithm was more accurate than traditional MSD algorithm for the same input data length. Moreover, cloud-based machine learning algorithm reached to 80% of detection accuracy using only 160-second of input data which there is a significant proof of concept and achievement towards real-time damage detection in a real-world SHM scenario. The parametric analysis also found that only three sensors, located at the top, middle, and bottom of the building, were sufficient to achieve over 85% damage detection accuracy when cloud-based machine learning algorithm was used. For 90% accurate damage detection, the cloud-based machine learning algorithm required 10 minutes of input data. Accounting for 2-minute computation time, it meant that 90% accurate damage prediction for a very complex building could be achieved within 12 minutes. The cloud-based machine learning algorithm, therefore, have great potential for achieving very near real-time damage detection. - iii - Significant contributions of this work include (i) understanding how varying operational and environmental conditions affect building response measurements; (ii) demonstrating successful time-critical damage detection in a real-world building damaged in two major earthquakes using MSD-based damage detection algorithm; (iii) providing a framework to guide the selection of input parameters when using MSD-based damage detection algorithm; (iv) proposing a novel real-time MSD-AANN damage detection algorithm that is more accurate and faster than traditional algorithms. (v) introducing a cloud-based machine learning algorithm which utilises Microsoft Azure Machine Learning Studio (MLS) to execute damage detection processes.