Robust Principal Component Thermography for Defect Detection in Composites

Abstract

Pulsed Thermography (PT) data are usually affected by noise and as such most of the research effort in the last few years has been directed towards the development of advanced signal processing methods to improve defect detection. Among the numerous techniques that have been proposed, principal component thermography (PCT)-based on principal component analysis (PCA)-is one of the most effective in terms of defect contrast enhancement and data compression. However, it is well-known that PCA can be significantly affected in the presence of corrupted data (e.g., noise and outliers). Robust PCA (RPCA) has been recently proposed as an alternative statistical method that handles noisy data more properly by decomposing the input data into a low-rank matrix and a sparse matrix. We propose to process PT data by RPCA instead of PCA in order to improve defect detectability. The performance of the resulting approach, Robust Principal Component Thermography (RPCT)-based on RPCA, was evaluated with respect to PCT-based on PCA, using a CFRP sample containing artificially produced defects. We compared results quantitatively based on two metrics, Contrast-to-Noise Ratio (CNR), for defect detection capabilities, and the Jaccard similarity coefficient, for defect segmentation potential. CNR results were on average 40% higher for RPCT than for PCT, and the Jaccard index was slightly higher for RPCT (0.7395) than for PCT (0.7010). In terms of computational time, however, PCT was 11.5 times faster than RPCT. Further investigations are needed to assess RPCT performance on a wider range of materials and to optimize computational time.

Keywords: CFRP; OIALM; Orthogonal IALM; PCP; RPCA; Robust PCA; noise reduction; pulsed thermography.

Figure 1

(a) Research block diagram; and (b) proposed method block diagram.

Figure 2

(a) CTA CFRP plate, where Z is the defect depth and labels are used to identify the location of each defect; and (b) pulsed thermography setup. a, PC; b, IR camera; c1 and c2, left and right flashes; d, CFRP specimen.

Figure 3

(a) Jaccard Index similarity definition; and (b) similarity between the ground-truth and the detected area.

Figure 4

Segmentation and Jaccard index computation flow graph.

Figure 5

Examples of reference and defect regions. The boundaries of the reference region are between the green and red lines, while the defective region is inside the blue line area.

Figure 6

(a) Raw data at Frame 5; (b) fifth component of PCT data; and (c) fifth component of RPCT data.

Figure 7

Thermal profiles (temperature vs. time plots on a semi-logarithmic scale) of a pixel inside defect FBH-4M (red plot); a sound pixel close to this defect (blue plot); and the absolute thermal contrast between these two (green plot).

Figure 8

PCT components: (a) first; (b) second; (c) third; (d) 21st; (e) 500th; and (f) 3500th.

Figure 9

Comparative CNR curves for defects at the same depth (Z = 1.43 mm) but different types (FBH, pull-outs and inserts): (Row 1) raw data; (Row 2) PCT; and (Row 3) RPCT. Columns 1–5 represent FBH-4M, FBH-4G, PO10-F2, PO15-H, and Tef-H, respectively.

Figure 10

CNRmax by defect type as a function of the defect depth for all data sequences (PCT and RPCT).

Figure 11

CNRmax by defect depth as a function of the defect type for all data sequences (PCT and RPCT).