A guided wave ultrasound (GWU) based monitoring system has been developed and is currently being used to detect rail breaks in an operational rail line. The development of more sophisticated systems to detect defects at long ranges requires accurate wave propagation models of the rail. This is complicated because many different dispersive modes are excited and propagate in the rail. Changes in environmental conditions, such as temperature, also influence the propagation characteristics of the rail. The semi-analytical finite element (SAFE) method can be used to predict the dispersion characteristics. This paper will present a method for estimating the defect growth signatures in field-measured spectrograms using linear and nonlinear ML models trained using simulated spectrograms generated from a SAFE model. The defect growth is measured using piezo-electric transducers mounted under the railhead to transmit and receive ultrasonic guided waves in the pulse-echo mode. These were installed on an operational heavy haul rail track [1]. An artificial defect was also introduced for two weeks by glueing a small mass under the railhead. The deterioration of the glue joint caused variation in the measured defect signature of the mass.To allow for a model that can generalize across known and unknown conditions, we consider an unsupervised training approach using generative modelling. Generative modelling approaches identify a lower-dimensional description of the variations in the data that can be used to reproduce the data. These lower-dimensional descriptions are referred to as latent spaces. Generative statistical and machine learning approaches can be designed to efficiently reconstruct signals or identify interpretable variations in the data.Reconstruction focused approaches compactly capture the maximum variance in the data to reconstruct signals. These include Principal Component Analysis (PCA), Autoencoders (AE), Variational Autoencoders (VAE) and Generative Adversarial Networks (GANs) [2]. In turn, latent focused approaches aim to reconstruct the data by identifying underlying variations in the informative and statistically independent data. Latent focused approaches include Independent Component Analysis (ICA) and beta-Variational Autoencoder (beta-VAE). ICA has been used to identify growing defects in GWU time-series signals [1]. This paper focuses on whether linear techniques are sufficient to identify source components of GWU measurements or whether nonlinear machine learning approaches provide additional benefits.Time-domain vibration signals from experimental and simulated measurements are converted into spectrograms using the Short-Time Fourier Transform. Spectrograms or image data allow for comparison with established nonlinear ML image-driven approaches. ICA and beta-VAE models are trained on simulated spectrograms and compared in their ability to extract informative components of the artificial defect from experimental spectrograms. We deliberately choose spectrograms over raw time-series data, as spectrograms provide an improved setting for linear statistical learning and nonlinear machine learning approaches. Firstly, spectrograms include highly nonlinear transformations that are higher-dimensional than windowed raw time-series data. This gives an improved setting for linear methods. Secondly, research on image data using nonlinear machine methods is vastly more mature than raw time-series data.The growing defect measurements are used as input to the ML models trained using numerous simulated spectrograms. The models learn to map variations in spectrograms, including multiple weld reflectors, into an interpretable latent space. Notably, the approach allows for the isolation and identification of salient underlying variations in the latent space, such as the growing artificial defect. Using our training approach, we show that GWU source characteristics such as defects can be untangled in the latent space, while original field measured data can also be reconstructed accurately.Acknowledgement:This work was supported by the Council of Scientific and Industrial Research (CSIR) and the Center for Asset and Integrity Management, the University of Pretoria in South Africa.References[1] P. W. Loveday, C. S. Long, D. A. Ramatlo, Ultrasonic guided wave monitoring of an operational rail track, Structural Health Monitoring 19 (6) (2020) 1666–1684.[2] W. Booyse, D.N. Wilke and P.S. Heyns. Deep digital twins for detection, diagnostics and prognostics. Mechanical Systems and Signal Processing 140, 106612.

Identification of Defect Growth in Guided Wave Measurements

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.