A Machine Learning Approach for Automatic Operational Modal Analysis

Structural Health Monitoring (SHM) has been one of the main research topics in the area of civil, mechanical and aerospace engineering for the past few years. Modal parameters and their evolution in time can be used as features and indicators of the damage which a structure is subjected to and might also allow for a prediction of the residual useful life of the same. SHM needs of long time-series of data to be efficient. For this reason, output-only techniques of system identification, which can record data continuously and without the constant supervision of an operator, are particularly suitable for this aim and are typically referred to as Operational Modal Analysis (OMA). In the output-only techniques, the knowledge of the input is replaced by the assumption that the input is a realisation of a stochastic process (white noise). The determination of a model that fits the measured data is named stochastic system identification. Despite the existence of a large number of OMA algorithms developed during the last decades, this work is exclusively focused on the Covariance-driven Stochastic Subspace Identification (SSI-COV). SSI-COV needs the definition of a model order n which is directly linked to the number of modes identified in the analysis (n/2). Because of the uncertainty of the SSI method, it’s necessary to conduct this analysis considering a range of model orders which has, as result, the identification of several modes, generally called poles. The poles identified for a certain model order may have similarities or dissimilarities with the poles identified for a different model order. These are evaluated in function of the modal parameters which characterise each pole. The poles which show a low variation of the modal values with the changing of the model order are defined as stable (physical modes). On the other hand, the poles with high variation of the modal parameters with the changing of the model order are defined as unstable (spurious modes). This evidence has led to the ideation of stabilization diagrams which show the variation of modal parameters in function of the model order. In a stabilization diagram the stability can be estimated by a visual analysis with the aim of determining a model order where the poles manifest a low variation of parameters. The adoption of a stabilization diagram results in a manual identification of modal parameters which can lead to results affected by user experience. In this study, an automated identification method is proposed with the aim of providing a process which is completely independent from the user experience, objective and based on the latest statistical methods of analysis. Consequently, a multistage cluster process is developed on the basis of the definition of the physical parameters characterising the modal properties. The proposed method is tested on a numerical case considering the influence of the model order and the dimensions of the Henkel matrix on the results. Once all the parameters which control the process are validated, two experimental cases are analysed with the aim of verifying and quantifying the performance of the proposed method. In the first case a helicopter blade is used as a simplified experimental case; then, a scale reproduction of a masonry arch bridge is analysed as a complex application on a Civil Engineering structure.