Session: 01-10-01: Machine Learning, Artificial Intelligence, and Deep Learning in Dynamics, Vibrations, and Control

Paper Number: 167273

Classification of Spatially Distributed Data Harvesting Positions of the Vibrations of a Carbon-Fiber Cylindrical Shell. Model Compression and Weight Space Analysis via the Framework of the Advanced Proper Orthogonal Decomposition Algorithm 

Machine learning algorithms and neural models have evolved rapidly through the past decade. Despite the success of deep learning models and other state-of-the-art algorithms in a plethora of complex tasks, optimal selection of model order is yet to be addressed by a robust method of generalized performance. In this paper, vibration datasets are collected simultaneously during the free vibrations of a carbon fiber cylindrical shell by three piezoelectric accelerometers placed strategically upon its surface. Structural anomalies stored in the core structure of the component, resulting from damages induced during manufacturing and operation or its complex nature, create a landscape of dynamics. This landscape is scanned by the oncoming wave’s perturbation after the hammer-hit excitation of the cylinder at 51 equally distributed positions over its overall length. In the transit phenomena contained in the initial moments following excitation, stochasticity is evident as a result of the existence of uncertainties in the system - as in all real-world structures. Additionally, the position of data collection is certainly influencing the apparent stochasticity of scattering as certain locations may amplify the stochasticity effects due to the microstructural randomness. The classification task of this study targets in the prediction of data collection sites based on the dynamics extracted from each piezoelectric sensor. Early-time window datasets are split in 51 time series vectors corresponding to the overall number of excitation positions. In continuous datasets are sorted and labeled in three different classes based on the sensor position harvesting the cylinders vibrations. A multi-perceptron network with one hidden layer is constructed for the classification. The time series vectors are processed through the Fast Fourier Transform for extracting their frequency components and the training database is completed. Despite, a relevant fast success, the network struggled to achieve high performance and plateaued in an 88% accuracy of predictions. The cylinders vibrations are highly dominated by the existence of a dominant frequency of much higher magnitude in comparison with the rest. The latter, in addition with the uniformity of the structures geometry imposes difficulties in capturing any data patterns of lower energy caused by the existent stochasticity. Here, the advanced Proper Orthogonal Decomposition algorithm was used as a computational tool of great novelty in handling high dimensional data and feature extraction. After, analyzing the weights of the first layer in the network, the POD energy spectrum of the POD transformed weight matrixes indicate that the cylinders dynamics evolve in a much lower dimensional space than the weights matrixes dimensions suggest. In addition the POD mode shapes exhibit overfitting along the POD spatial dimensions. All the above indicate a model order reduction based on the findings provided by the POD. As a result the size of the datasets early time window is decreased furtherly to achieve higher time resolution of the FFT vectors for the transient dynamic phenomena evolving exactly after the excitation and the training is repeated from scratch. The network improved in classifying the three different datasets properly, achieving an accuracy percentage of 94%. The main motivation and contribution of this paper is the construction of an optimal POD model order selection framework able to adjust a networks hyperparameters in the physical constraints of system in the case of structural dynamics and the development of algorithms able to perform classification tasks of undoubted practicality in control and structural health monitoring processes.

Presenting Author: Ioannis Georgiou National Technical University of Athens, NTUA

Presenting Author Biography: Dr. Ioannis (Yannis) T. Georgiou is a Faculty Member of the National Technical University of Athens, GR and Adjunct Faculty Member of Purdue University, West Lafayette, Indiana . His scientific endeavors include the development of advanced diagnostics of composite material structures by using advanced proper orthogonal decomposition tools in combination with Artificial Intelligence and Machine learning algorithms.

Authors:

Konstantinos Liontos National Technical University of Athens
Ioannis Georgiou National Technical University of Athens, NTUA