An adaptive learning damage estimation method for structural health monitoring

Structural health monitoring is an important problem of interest in many civil infrastructure and aerospace applications. In the last few decades, many techniques have been investigated to address the detection, estimation, and classification of damage in structural components. One of the key challenges in the development of real-world damage identification systems, however, is variability due to changing environmental and operational conditions. Conventional statistical methods based on static modeling frameworks can prove to be inadequate in a dynamic and fast changing environment, especially when a sufficient amount of data is not available. In this paper, a novel adaptive learning structural damage estimation method is proposed in which the stochastic models are allowed to perpetually change with the time-varying conditions. The adaptive learning framework is based on the use of Dirichlet process (DP) mixture models, which provide the capability of automatically adjusting to structure within the data. Specifically, time-frequency features are extracted from periodically collected structural data (measured sensor signals), that are responses to ultrasonic excitation of the material. These are then modeled using a DP mixture model that allows for a growing, possibly infinite, number of mixture components or latent clusters. Combined with input from physically based damage growth models, the adaptively identified clusters are used in a state-space setting to effectively estimate damage states within the structure under varying external conditions. Additionally, a data selection methodology is implemented to enable judicious selection of informative measurements for maximum performance. The utility of the proposed algorithm is demonstrated by application to the estimation of fatigue-induced damage in an aluminum compact tension sample subjected to variable-amplitude cyclic loading.