A recursive Bayesian approach to small fatigue crack propagation and detection modeling

Abstract

Engineers have witnessed much advancement in the study of fatigue crack detection and propagation modeling. More recently, the use of certain damage precursors such as acoustic emission signals to assess the integrity of structures has been proposed for application to prognosis and health management of structures. However, due to uncertainties associated with small crack detection of damage precursors and crack size measurement errors of the detection technology used, applications of prognosis and health management assessments have been limited. In this article, a methodology is developed for the purpose of assessment of crack detection and propagation parameters and the minimization of uncertainties including detection and sizing errors associated with a series of known crack detection and propagation models that use acoustic emission as the precursor to fatigue cracking. The methodology is facilitated by the Bayesian inference of a joint-likelihood model which includes sizing and detection models. Examples where several dog-bone Al 7075T6 specimens are tested to produce fatigue crack initiation and propagation data and estimates based on remaining useful life support the effectiveness and usefulness of the proposed methodology.

Estimating damage size and remaining useful life in degraded structures using deep learning-based multi-source data fusion

“…24,45,46 Moreover, the AE cumulative count and energy have shown promising results in the literature on crack detection and growth estimation using AE. 20,29,40 Consequently, the AE cumulative count and energy are used as the input features in our time-series analysis ( N ft = 2, see section “Customized LSTM-based regression model for fatigue crack size estimation”) where crack detection and length estimation tasks are carried out in a twofold process: (1) Crack detection: considering a previous work, 45 abrupt change in the AE count per signal is considered for crack initiation detection. It turned out that this value is in the range of [65–150] for all of Yun’s experiments, where crack initiation is defined to be the time at which crack reaches 250 µm of length.…”

“…For example, CNN-based models for binary detection of various damages are introduced in several reported works. [25][26][27][28] However, automated, but not manual, 29 image processing has not been used for online damage sizing and subsequent RUL estimation. This gap in the literature might be due to the structure of standard CNNs, which can result in significant loss of resolution in the output due to successive pooling layers.…”