PSI - Issue 28

F. Conrad et al. / Procedia Structural Integrity 28 (2020) 2195–2205 F. Conrad, A. Blug et al. / Structural Integrity Procedia 00 (2019) 000–000

2198

4

3. GPU-based DIC-sensor The left side of Fig. 2 shows the sensor head of the real-time DIC system in front of the biaxial testing machine with a cruciform specimen from Fig. 1. It consists of a sensor head combining a fast CoaXPress camera (Optronis CP70-16-M-148, 4.672 x 3.416 pixel, 3.9 µm pixel pitch, 148 fps for full image) in combination with a blue 23 W LED ring light. Either a telecentric lens (Sill S5LPJ1290/216, magnification 1:0.46, working distance 190 mm) or a standard lens (QIOPTIQ APO-RODAGON HR 75/5.6, magnification 1:1.15, working distance 150 mm) were used. The resulting image resolution is 8.5 µm/pixel with the telecentric lens and approximately 4.5 µm/pixel with the standard lens. A CoaxPress frame grabber (microEnable 5 ironman AQ8-CXP6D) transfers image data from camera to an Intel Core i9-7900 computer equipped with a Nvidia GeForce RTX 2080 TI GPU with 11 GB RAM and 4352 CUDA cores. In contrast to the system described in Blug et. al. (2019), this hardware combination increases data bandwidth from 0.7 GB/s by a factor of three to roughly 2.1 GB/s. In addition, size and position of the imaging area on the 16 MPixel image sensor can be selected in a wide range without loss of data bandwidth. This means that the frame rate of the system is adaptable by image size, i.e. from 850 Hz with an image size of 2176 x 1080 Pixel to 1200 Hz with 2176 x 864 Pixel. The first case corresponds to a field-of-view (FoV) of 18.5 x 9.2 mm², the latter to a FoV of 18.5 x 7.3 mm².

Fig. 2. Left: DIC sensor head in front of the biaxial testing machine equipped with a cruciform specimen. Right: camera image of specimen surface superimposed by the four regions of interest (ROI) for strain-control and by the full-field evaluation area. The (x,y)-coordinates (yellow) mark the camera coordinate system, the (a,b) coordinates correspond to the A and B axes of the biaxial testing system. The system is operated either in ‘strain-control’ mode where global strain along the A and B axes is calculated from four regions of interest (ROI) or in ‘strain-field’ mode, where the displacement is calculated in real time. Both cases are shown on the right side of Fig. 2 where a camera image is superimposed by the four ROIs for strain-control and the full-field DIC area surrounding the crack-starter notch, which consists of up to 40,000 ROIs with a typical size of 21 x 21 pixel, each. The outer ROIs consist of two concentric squares: the inner ones mark the area of the reference templates (typically 63 x 63 pixels), the outer ones determine the so-called search areas with typically 255 x 255 pixel within that displacement is measured path-independently, i.e. without tracking. This path independence makes strain-control robust against single mismeasurements, as it does not depend on previous images or evaluation results. Furthermore, the sample surface is not speckle-painted as it is necessary for commercial systems. Instead, it works marker-free by correlation the microstructure of the specimen surface as described by Blug et. al. (2019). In strain-control mode with four search ROIs of 255 x 255 pixel are computed with a correlation rate of approximately 12 kHz, thus the strain measurement rate is limited by camera frame rate. At large cracks with a length of up to 7 mm the outer ROIs must have a sufficient distance to the crack to measure global strain along the A and B axes independent from crack length. As FEM simulations indicate, the minimum base length � for strain calculation is 10 mm, similar to that of the mechanical extensometer. To optimize the frame rate for a given image data transfer rate of 2 GB/s, the camera was tilted by 45 degrees. Thus, A and B axes run diagonally so that a FoV of 18.5 x 7.3 mm²