Machine learning applied to non-deterministic actions affecting slender structures and their active cancellation

Zusammenfassung

Vibration problems in slender structures pose a significant challenge in modern structural engineering, leading to problems such as structural fatigue, discomfort and safety risks. These undesired vibrations arise from various nondeterministic sources such as dynamic loads, turbulent winds, human activities, and machinery. Understanding and characterizing these actions is crucial for structural design and safety and represents a central research topic in structural engineering. In this context, data-driven methods have emerged as a valuable addition to traditional structural engineering techniques. They use extensive data collection, sensor networks and advanced analytics to provide real-time insights into structural behaviour and accurate forecasts of potential excitations, etc. This doctoral thesis aims to develop and apply data-driven techniques to address vibration challenges in slender structures. Its objectives involve identifying, predicting and characterizing non-deterministic actions affecting these structures by means of data-based non-parametric models as Machine Learning, as well as developing methods to actively mitigate them based on evolutionary computation. The doctoral thesis encompasses three works, where various issues related to different facets of structural vibration analysis have been successfully addressed following the subsequent methodologies. In first place, the focus is on the prediction and characterizing stochastic forces that dynamically influence structures, with particular emphasis on extreme events with potentially significant impacts. Specifically, the work involves the prediction of extreme wind speeds. An intrinsic challenge in predicting such extreme events lies in dealing with highly unbalanced datasets. To address this, in addition to the application of conventional data balancing techniques, a novel three-level Hierarchical Classification/Regression methodology was developed, yielding highly satisfactory results in forecasting extreme wind events while minimizing false alarms. The prediction of stochastic events was conducted across various time prediction horizons, spanning short to long term, ensuring the methodology's robustness and optimal performance across different scenarios. The second work is focused on characterizing non-deterministic forces impacting structures, specifically emphasizing the reproducibility of their temporal series using an electrodynamic shaker. This approach facilitates standardized testing of structural responses to dynamic loads in an objective and repeatable manner. The challenge of dealing with a naturally nonlinear electro-mechanical system, represented by a non-invertible model, was addressed in this work, where the goal was to derive an inverse model for replicating time series signals. To overcome this hurdle, an iterative neural network framework for replicating human-induced ground reaction forces was developed. Within this framework, an inversion-free offline control approach was applied to the electrodynamic shaker, ensuring repeatability and accuracy in dynamic load tests. This proposal was successfully validated, achieving reliable reproduction of ground reaction forces produced by different types, amplitudes, and frequencies of human motion or locomotion activities. The final work involves the successful development, implementation, and experimental validation of an Active Mass Damper control system for a full-scale structure. A genetic evolutionary algorithm was utilized to optimize both the state estimator gain and the feedback gain controlling the actuator in the active control methodology. This demonstrated that the data-based optimization of the control law serves as a viable alternative to classical methods. Various optimization criteria were assessed for this purpose. Additionally, the validation of the control system was carried out by evaluating different parameters in both the time and frequency domains. In terms of the obtained results, the accomplishments achieved throughout the development of this doctoral thesis represent notable contributions to the research field in which it is framed. Developing and successfully applying machine learning and artificial intelligence methods to address challenges arising from structural engineering.