![[External Link]](https://webarchive.library.unt.edu/eot2008/20090114092941im_/http://institute.lanl.gov/_images/icons/external_site.gif){kind=icon}
Listed below are topics relating to state of the art SHM research in the areas of Feature Extraction, Statistical Modeling and Data Normalization.
Building upon previous work at the Los Alamos National Laboratory (LANL), a distributed damage detection system with wireless sensor units is being implemented. The wireless sensor unit consists of accelerometer sensors, an onboard processor, and a radio-frequency transmitter. Microelectromechanical systems (MEMS) accelerometers are being tested on various structures to monitor the dynamic behavior of the joints. Basic statistical damage detection algorithms are programmed onto the processor onboard the wireless sensor units. Changes in correlation of the statistical properties are used to detect damage in the joint. The data is processed locally onboard the wireless sensor units and only the final diagnosis results are transmitted to the base station.
The impedance method monitors the variations in mechanical impedance resulting from damage, which is coupled with the electrical impedance of the PZT sensor/actuator. Using this coupling property, this powerful and robust methodology records the variations in structural mechanical impedance in response to high-frequency excitations and structural damage is detected by analyzing variations from baseline impedance signatures. In order to ensure high sensitivity to incipient damage, the electrical impedance is measured at high frequencies (typically higher than 30 kHz). At such high frequencies, the wavelength of the excitation is small and is sensitive enough to detect minor changes in the structural integrity.
For the Lamb wave propagations, one PZT patch acting as an actuator launches elastic waves through the structure, and responses are measured by an array of the other PZT patches acting as sensors. The structure can be systematically surveyed by sequentially using each of the PZT patches as an actuator and the remaining PZT patches as sensors. The technique looks for the possibility of damage by tracking changes in transmission velocity and wave attenuation/ reflections. LANL researchers developed a unique wavelet-based data processing method to identify the delamination in composite plates.
LANL also investigates the applicability of time reversal concept in modern acoustics to structural damage identification. In the time reversal acoustics method, an input signal at an excitation point can be reconstructed if a response signal measured at another point is reemitted to the original excitation point after being reversed in a time domain. This time reversibility of Lamb waves is violated when wave distortion/scattering is caused by a defect along a direct wave path. Examining the deviation of the reconstructed signal from the known initial input signal allows instantaneous identification of damage without requiring the baseline signal for comparison.
|
The Sequential Probability Ratio Test (SPRT) develops a statistical method that quickly arrives at a decision between these two hypotheses and is applicable to continuous monitoring. In the original formulation of the SPRT algorithm, the feature is assumed to be Gaussian and thresholds are set based on the center mass properties of the distribution. It is likely, however, that the feature used for damage identification is sensitive to the tails of the distribution and that the tails may not necessarily be governed by Gaussian characteristics. By modelling the tails using the technique of Extreme Value Statistics (EVS), the hypothesis decision thresholds for the SPRT algorithm may be set avoiding the normality assumption. |
The first and most important objective of any damage identification algorithm is to ascertain with confidence if damage is present. Considering most real world applications of damage detection, this detection must be accomplished in an unsupervised learning mode. Here, the term unsupervised learning implies that data from a damaged state are not used to aid in the damage detection process. Many methods have been proposed for unsupervised damage detection based on ideas of novelty detection, which is founded in the fields of pattern recognition and multivariate statistics. However, novelty detection is limited to distributions of data that are Gaussian or near Gaussian in some sense. Specifically, the assumption of normality imposes potentially misleading behaviour on the extreme values of the data, namely, those points in the tails of the distribution. As the problem of novelty detection specifically focuses attention on these tails, the assumption of normality is likely to lead the analysis astray. An alternative approach can be based on extreme value statistics. This branch of statistics was developed to specifically model behaviour in the tails of the distribution of interest.
In this application of the statistical pattern recognition paradigm, a prediction model of a chosen feature is developed from a baseline structure. The feature is extracted from the time domain response of the structure. After the model is developed, subsequent feature sets are tested against the model to determine if a change in the feature has occurred. In the proposed statistical inference for damage identification there are two basic hypotheses; (1) the model can predict the feature, in which case the structure is undamaged or (2) the model can not accurately predict the feature, suggesting that the structure is damaged.