Applying robust variant of Principal Component Analysis as a damage detector in the presence of outliers

• Outliers (contaminated data) are inevitable part of any system, experiment or data collection.
• PCA is sensitive to outliers, so any strategy based on PCA might not be reliable if these are unconsidered.
• Robust variant of PCA provides more reliable and stable results in SHM field.
• Different robust PCA methods are analyzed and compared with the ordinary counterpart.
• Experiments performed on a real scale structure provide a strong support for the claims.

Using Principal Component Analysis (PCA) for Structural Health Monitoring (SHM) has received considerable attention over the past few years. PCA has been used not only as a direct method to identify, classify and localize damages but also as a significant primary step for other methods. Despite several positive specifications that PCA conveys, it is very sensitive to outliers. Outliers are anomalous observations that can affect the variance and the covariance as vital parts of PCA method. Therefore, the results based on PCA in the presence of outliers are not fully satisfactory. As a main contribution, this work suggests the use of robust variant of PCA not sensitive to outliers, as an effective way to deal with this problem in SHM field. In addition, the robust PCA is compared with the classical PCA in the sense of detecting probable damages. The comparison between the results shows that robust PCA can distinguish the damages much better than using classical one, and even in many cases allows the detection where classic PCA is not able to discern between damaged and non-damaged structures. Moreover, different types of robust PCA are compared with each other as well as with classical counterpart in the term of damage detection. All the results are obtained through experiments with an aircraft turbine blade using piezoelectric transducers as sensors and actuators and adding simulated damages.