| Περίληψη: | Modern technology is heavily dependent upon structural and mechanical systems. Early damage detection and classification of the structure’s stability can prolong structures’ health and useful lifetime. Damage diagnosis is an important process related to increased safety, reduced costs and predictive maintenance in vibrating structures. Gathering vibration signals is usually an easy task since there is no need for specialized sensors and expensive equipment or disrupting the structure's continuous operation. Therefore, vibration-based Structural Health Monitoring (SHM) has gained considerable popularity over the past years as it combines effectiveness, ease of use, and cost-effective implementation. A critical barrier of SHM is the dynamic changes due to factors other than incipient damage, including varying Environmental and Operating Conditions (EOCs), such as temperature, humidity, payload, and other uncertainty sources. In addition, SHM needs an extensive signal pre-processing for a proper feature extraction, which is time-consuming. Despite SHM’s efficiency in damage detection, the feature extraction becomes significantly more complicated in damage classification tasks. Moreover, in many cases, it is difficult to acquire faulty conditions from a structure, as it must be damaged on purpose or previous real damaged cases should be used. Therefore, it is common to have much less damaged than healthy experiments (imbalanced datasets) for SHM. In this thesis, we are dealing at first with the issue of uncertainty and damage detection and classification by approaching SHM using Supervised Learning techniques during the oscillatory operation of a UAV tail boom. Secondly, we are dealing with the problem of imbalanced datasets using Reinforcement Learning. Data were gathered from an aircraft’s simulated tail boom by a lab-scale structure representing a composite beam - which was being re-positioned between two consecutive experiments for incorporating the response signals with uncertainty - and response signals were recorded for different structural states (one healthy and six damage scenarios). On the first hand, a Supervised Learning model identified the healthy or damaged state and then classified the possible damaged condition into a specific category. The received signals were used as time-series input to a 1-Dimensional Convolutional Neural Network (CNN1D). Using an equal number of healthy/damaged experiments, signal preprocessing and hyperparameters’ tuning the model achieved 100% True Positive Rate in binary classification problem (healthy/damaged) with a false alarm level equal to 0.5% or more, and more than 97% as the worse scenario in categorical classification (six different damaged scenarios) , which is a very difficult task for SHM, with a false alarm equal to 3% or more. On the other hand, a Reinforcement Learning model used for the identification of a binary condition with an imbalanced dataset. Although Supervised Learning struggled to identify correctly the structure’s condition with imbalanced experiments, using a combination of Deep Q-Networks and CNN1D, we proved that our Reinforcement Learning approach could classify correctly more than 99% of the experiments.
|
|---|
