The project will focus the development of a robust automated approach for identifying damage in structural systems. The research thrusts are: 1) fusion of complementary algorithms and 2) optimal sensor distributions for the fused set. Selection of complementary algorithms involves identification of methods whose sensitivity to damage and to the sources that cloud damage detection differs with damage scenarios and operating conditions. In a first phase the project inspects the fusion of detection filters that operate on residual correlations with filters that work with amplitude dependent residual metrics. The research expects to demonstrate that the optimized fused detector will have a damage detection threshold that, for a fixed probability of false alarm, is significantly better than that of the individual algorithms. Intimately connected with the algorithmic fusion is research on the selection of sensor layouts that are optimal, given the fused interrogation scheme. Following the analytical work the research progresses into an experimental phase where the performance of the fused algorithms is tested on a one quarter scale steel structure that is exposed to the weather and thus subjected to realistic environmental changes.
Algorithm fusion has proven fruitful in Automatic Target Recognition and various other areas but a systematic examination in the context of Structural Health Monitoring is first carried out in this project. If successful, this research will not only offer a robust damage detection scheme for applications to civil structures but it will also point to the merit of algorithmic fusion for other objectives such as the localization and the quantification of damage. Educational activities connected with the project include: 1) interactions with Olin College, an undergraduate engineering school of excellence, through introduction of multi-week research activities based on topics from the project 2) involvement with the program Girls Get Connected (GGC), a science and technology outreach for middle school girls in the Boston area and 3) an afternoon of hands-on activities on the Harvard?s Medical School explorations program, which is attended each fall by over 200 middle school students from Cambridge and Boston. The graduate student working on the project will also receive advanced training on the topic of damage detection in civil structures which is of high engineering importance.
Albeit most evident in the automotive and airline industry, maintenance for adequate performance and safety is also needed in structural systems. Structural Health Monitoring (SHM) attempts to provide information on the state of a structural system from the analysis of signals. The outcome of an interrogation to a SHM system is either: 1) that the data shows no reason for concern or 2) that an anomaly is detected. In the second case a full characterization of the fault or damage involves, in addition to detection, localization, severity quantification and prognosis, the last entry referring to what can be expected if actions A or B or C …. are taken, including no action at all. Detection is the essential pillar on which any SHM system is built and it is the focus of this project. Qualitatively all detection problems face the same difficulty. Namely, telling apart observations that point to what one is interested in, from observations that do not; whales from submarines, armed missiles from decoys, etc. In monitoring of civil structures the issue is telling apart changes arising from fluctuations in the environment from changes associated with damage. The environment being the temperature field, the humidity and the wind conditions. For a detector to be effective the probability that damage is identified when it occurs must be high and the rate of false alarms must be low. Fig.1 illustrates the matter schematically and shows that adequate performance requires that the probability distributions in the healthy and the damage state that one is interested in detecting are "sufficiently" separated. With this background we now outline the contributions of the present project. Change Detection and Damage Detection The project shows that a critical issue for arriving at successful detection is to fuse the damage detection (DD) framework with "change detection" (CD). Specifically, in DD the question is whether one data set can be used to determine (with an acceptable Type I and Type II error probability) that a structure is healthy or damaged. The answer for very large damage may be yes, but for the small damage that one is often interested in conditioned based maintenance, the answer is typically NO. In CD the question is switched from whether the present data set shows damage to whether the "accumulated evidence" from the past N data sets indicates that the structure has changed. Since the damage produces permanent changes (as opposed to those from the environment which fluctuate), after sufficient time even small damage can be separated from the clutter with appropriate processing. The reason to stating that DD and CD are best fused, instead of switching to CD is that large sudden damage can be identified by the DD module. Environmental Fluctuations in Change Detection In its basic form the CD framework detects trends. If a model that represents the healthy condition is formulated from data collected at temperature T0 the system reacts as the mean temperature deviates from T0. One approach commonly considered in the literature uses Principal Components (PCA) but the approach was found to reduce sensitivity to small damage drastically. A simple but effective method to eliminate the environmental trend based on scrambling the data in the time axis was developed and shown to work effectively. Fig.2 illustrates how the implementation of PCA makes the detection of a particular damage impossible while the scrambling approach in fig.3 detects the damage without difficulty. The price paid by the simplicity of scrambling being the addition of some more waiting time to the announcement of damage. Single Model Detectors It is contended that robustness and practicality is promoted by operating in a data-driven fashion and within a format that does not require the formulation of multiple models. Specifically, the proposed scheme works with a single model and residuals that depend on how well the currently recorded data "fits" the healthy model. The project puts forth two data-driven residual based detectors: 1) A subspace residual approach and 2) A Kalman filter innovations based detector. The skeleton of these techniques existed prior to this project but neither proved adequate for application in damage detection (or change detection) under changing statistics of the ambient excitation. The project developed modified versions robust against these changes. Fig.4 illustrates results for the subspace scheme and fig.5 for the Kalman filter. Details of the developments are reported in papers that appear in the Journal of Mechanical Systems and Signal Processing. Algorithm Fusion and Optimal Sensor Placements Fusion of the subspace and the Kalman filter detectors did not show significant improvements in the Type I Type II error performance. The optimal sensor placement work showed there is a relatively large subset of conditions that are "close to optimal" and a set of "poor positions" that can be identified using observability based criteria. An illustration appears in fig.6.
