Machine learning techniques for impact characterization and vibration based damage detection for composite structures

Abstract

This thesis investigates the detection and localisation of two significant challenges facing Structural Health Monitoring of composite structures: Impact Detection and Damage Identification. To address Impact Detection, Localization, and Classification, a novel Deep Learning methodology is proposed for categorising impact data based on impact type (Hard or Soft). This is achieved by capturing voltage signals from Piezo-Electric sensors installed on a composite panel. The data undergoes further processing for classification according to energy, location, and material. A streamlined and automated feature extraction and selection process is accomplished through a sophisticated deep learning algorithm. Convolutional Neural Networks (CNN) are leveraged to discern critical features from the voltage data. Once these features are identified, impacts are categorized as either Hard Impacts (simulated using steel impactors in a controlled lab environment) or Soft Impacts (simulated using silicon impactors in a controlled lab environment), along with their corresponding location and energy levels. Moreover, to expedite the process, data for training is extracted as anomalies from the signals via Isolation Forests (IF). This novel approach yields highly accurate identification of Hard and Soft Impacts, their respective locations, and associated energy levels. Next the Damage detection in composite material is investigated by introducing a novel statistical vibration-based method. This approach accounts for uncertainties in the measured resonance frequencies. The methodology relies on utilizing resonance frequencies, recognized as the most precise and easily measurable vibration feature. To validate the effectiveness of this method, case studies were conducted using two identical composite plates: one with delamination and the other in pristine condition. In this context, the Frequency Response Functions (FRF) were gauged and employed as the primary input for the Resonance Detection Algorithm in the proposed method. Applying these FRFs to a Resonance Detector Function facilitates the determination of Resonant Frequencies and their statistical distribution. By assessing the statistical distributions of the corresponding resonant frequencies, the reliability in detecting damage is derived using the Beta Distribution. Upon scrutinizing the damage detection reliability of the two sets of corresponding resonant frequencies, it has been established that alterations in natural frequencies stem from structural changes, rather than random measurement errors. The concluding part of this thesis is dedicated to Damage localization. Here, the natural frequencies identified from the impaired plate are employed to train a Principal Component Analysis (PCA) algorithm. By constructing a vector space using PCA for the frequencies associated with damaged locations, distinct from the vector space of frequencies where damage is absent, it becomes possible to ascertain the precise location of the damage. This methodology establishes that damage localization can be achieved by observing alterations in the vector space of natural frequencies affected by damage.