PSI - Issue 64

Nikhil Holsamudrkar et al. / Procedia Structural Integrity 64 (2024) 580–587 Holsamudrkar Nikhil et al./ Structural Integrity Procedia 00 (2019) 000 – 000

582

3

The dataset is filtered to exclude any secondary failure modes, such as shear or friction. The waveforms recorded up to 40% of the peak load of the specimens were ignored from the final dataset to avoid any load-seating phenomenon. A total of 11328 filtered signals belonging to different classes were considered in the model development. Concrete and mortar tensile cracking were found to have similar signal characteristics and hence are considered a single class, which can be a limitation of this study. Future work includes considering tensile and shear cracking for concrete and mortar as separate classes. The four classes considered in the study include fiber rupture, matrix/concrete tensile cracking, fabric-matrix debonding, and steel yielding. A bar plot in Fig. 3 represents the data distribution in these classes.

Fig.1. AE Data collection for the (a) Fabric rupture, (b) Fabric-matrix debonding, (c) Steel yielding, (d) Concrete Cracking, (e) Mortar matrix cracking

Table 1. Properties of the materials used in the study

Test

Material

Size of Specimen (L x B x H) mm

No of specimens

Loading rate (mm/min)

Material Grade (MPa)

Codal Guidelines

Direct Tension Direct Tension Single Lap-shear Four-point bending Four-point bending

FRCM mesh impregnated

300 x 30 x 0.184 *

20 20 10

0.5

2000

ASTM D3039 ASTM A615

Steel rebar

16 ¥

1

550

FRCM-concrete interface

350 x 100 x 100 § 800 x 150 x 150 500 x 100 x 100

0.2 0.2 0.2

-

RILEM 250-CSM

Concrete

2 2

35 †

ASTM C78 ASTM C78

FRCM Mortar matrix

70

* Equivalent thickness ¥ Diameter of rebar § Concrete substrate size † Compressive Strength

2.1. Discrete Wavelet Transform (DWT)

The discrete wavelet transform (DWT) is chosen to convert waveform data into image data. It facilitates efficient data compression and denoising by representing signals in terms of wavelet coefficients, enabling the removal of redundant or noisy information while preserving important features. The DWTs preserve both time and frequency domain characteristics, thereby assisting in focused analysis and interpretation. Figure 2 shows waveforms and their corresponding DWTs for different failure modes captured through different tests. DA14 wavelet is used in the current study, a member of the Daubechies family. It is generally utilized in signal processing tasks like denoising, compression, and feature extraction due to its orthogonal properties and effectiveness in capturing signal characteristics. A threshold of 90% was selected while carrying out the DWT transform. The current representation of DWT in Fig. 2 is done to improve the interpretability of the time domain signal and its corresponding DWT. However, CNN models work efficiently if the image's aspect ratio is 1. Therefore, data to the CNN model is fed by resizing the DWT image to 224 x 224 in RGB format. The axis labels, plot titles, and other features, such as the color bar, are removed, and only the DWT image is retained and fed into the model.