In large-scale earthquake disasters, urban monitoring system would be effective in decision-making for emergency responses. For this purpose, we have proposed a data-driven technique for portfolio-monitoring of structures that adopts a machine-learning method called Correlation Anomaly Detection (CAD). This technique detects damaged structures in a portfolio by detecting a change in correlation property between dynamic characteristics of different structures. The key features of our proposed technique are summarized as follows.
(i) The methodology can be categorized into data-driven techniques which require no priori information about the physical properties of systems, and therefore has versatility in terms of applied objectives.
(ii) In the context of structural health monitoring (SHM), the methodology can be categorized into output-only techniques, which do not use input ground accelerations. This advantage allows its easy and cost-efficient implementation.
(iii) Most of output-only SHM methodologies adopt assumptions on the characteristics of excitations or mode shapes of vibrations. Our proposed technique, however, does not make such assumptions and thus has applicability even to strong motion observations.
The paper presents an improvement of CAD as Extended Correlation Anomaly Detection (ECAD), which utilizes the frequency content of time-series data, and examines its applicability to portfolio-monitoring of structures even in the case of different ground motions input to structures.
We first show the summary of CAD and then formulates the ECAD algorithm, which could utilize the frequency content of time series. In ECAD, the covariance structure corresponding to each discretized frequency is modelled as a co-spectrum matrices among time-series data, in order to capture correlation anomalies in the frequency domain. The series of anomaly scores computed by CAD for different discretized freqeuncies is defined as correlation anomaly spectrum, whose integration gives anomaly score for each monitored objective.
Second, we examine the applicability of CAD and ECAD to portfolio-monitoring of structures by numerical experiments. The portfolio is modelled by SDOF oscillators to which ground accelerations on different observation stations are input. The experiment shows (i) the good applicability of ECAD and its superiority to CAD even in the case of different excitations to structures, and (ii) the ability of ECAD to capture changes of natural frequencies.
Finally, we demonstrate ECAD by using actual strong motion observations. The aim is to detect structural damages resulted from the main shock of The 2011 off the Pacific coast of Tohoku Earthquake. The demonstration reveals the following: (i) ECAD has the applicability even to actual structures; (ii) ECAD tends to underestimate structural damage of structures whose response is a wide-band or multi-modal process; (iii) ECAD tends to overestimate structural damage when a structure shows different vibration modes between in reference and test data.
