Blind Identification of Structural Damage via Independent Component Analysis

The novel unsupervised blind source separation (BSS) technique is able to recover the underlying sources and their associated characteristic factors using only the measured mixture signals. This paper proposes to use the BSS learning algorithm— independent component analysis (ICA)—for blind identification of structural damage. Under the weak assumption of independence, ICA is derived to bias towards sparse components, which typically signify damage information. The measured structural responses are first transformed into the wavelet domain and then fed as mixtures into the BSS model. ICA is used to solve the BSS model, simultaneously yielding the “interesting” source with sparse representation that indicates damage instant, as well as its associated spatial signature in the recovered mixing matrix revealing the damage location. Numerical simulations and laboratory experiments show that the proposed ICA algorithm performs accurate blind identification of structural damage, whether in single or multiple events. The proposed ICA damage identification algorithm is straightforward and efficient to implement. It is output-only and unsupervised, requiring no excitation information or any knowledge with respect to the structure. Besides, it is robust to measurement noise. The FastICA algorithm enjoys cubic convergence rate, which enhances the proposed method with computational efficiency for online monitoring of structures.