Machine Learning Methods for Crack Identification in Euler-Bernoulli Beams

Abstract

The objective of this bachelor’s thesis is to evaluate the suitability of various machine learning methods for crack identification in Euler-Bernoulli beams placed on a Pasternak foundation. Methods available in Python libraries are used for this purpose. The machine learning methods of interest are linear regression, kernel ridge, Gaussian process regression, K-nearest neighbours, random forest, gradient boosting, and artificial neural network. The models are trained on three datasets: the one is based on the beam’s natural frequencies, the second one is based on the Haar wavelet coefficients obtained from the mode shape, and the third one is based on the Fourier transformed frequencies. In addition, the models are tested on a dataset containing white noise in the 5% range. The results show that the best results for predicting crack location can be achieved with models trained on the Haar wavelet coefficients, and Gaussian process regression turns out to be the best method for this case. Fourier transformed frequencies give the best results in predicting crack depth, and random forests are the most successful method in this case.

Bakalaureusetöö eesmärk on hinnata erinevate masinõppe meetodite täpsust pragude tuvastamisel Euler-Bernoulli talades, mis on paigaldatud Pasternaki alusele. Selleks kasutatakse Pythoni teekides olevaid meetodeid. Uuritavateks masinõppemeetoditeks on lineaarregressioon, kernel ridge, Gaussi protsessi regressioon, K-lähima naabri meetod (KNN), otsustusmets, gradient boosting, ja tehisnärvivõrk. Mudeleid treeniti kolmel andmehulgal: tala omavõnkumistel põhinev, võnkemoodi kujust teisendatud Haari lainikute kordajatel põhinev, ja Fourier’ teisendatud võnkumistel põhinev andmehulk. Lisaks katsetati mudeleid 5%-list valget müra sisaldaval andmestikul. Tulemused näitavad, et pragude asukoha ennustamisel saavutati parimad tulemused mudelitega, mis olid treenitud Haari lainikute kordajatel, ning parimaks meetodiks osutus Gaussi protsessi regressioon. Sügavuse ennustamisel andis parimaid tulemusi Fourier’ teisendusel põhinevad võnkumised ning edukaimaks meetodiks oli otsustusmets.