Machines are trained by humans to seek patterns in large volumes of data in order to discriminate between different data types, classes, or categories. A pattern recognition classifier (PRC), an artificial intelligence tool, is trained using data that can be traced back to a given class. Multiple data sets are used to train the PRC to differentiate between classes. This type of generic description applies to any type of information, or data, but how does one apply artificial intelligence and pattern recognition to non-destructive testing or more specifically, phased array ultrasonic testing (PAUT). The practice has been around since the 1980’s but we revisit the importance of input robust data nd principles here for phased array ultrasonic testing of welds. The article summarizes some of the challenges that are being assessed to automated PAUT weld defect classification.
Weld Defects and Phased Array Patterns
Weld defects can be generally categorized as volumetric and planar flaws. Volumetric weld flaws include porosity, slag, and inclusions. Planar weld flaws include lack of fusion (LOF), heat affected zone cracks, center line cracks, base metal cracks, and others. Some examples of weld defects that are routinely found with phased array ultrasonic testing (PAUT) are show in Figure 1.
Figure 1: Weld defects detected by phased array ultrasonic testing (PAUT); slag (top left), root incomplete fusion (top right), center line crack (bottom left), and bevel lack of fusion (bottom right).
Example data, generated using an Olympus MX-2 Omni-scan, from a slag inclusion is shown Figure 2. The data is generated using a 45 – 70 degree shear wave S-scan on a single vee 0.625” thick plate. The phased array A-scan and S-scan are shown on the left and right, respectively. The slag inclusion is detected in the second leg of the scan towards the top of the weld. Multiple individual peaks, or a cluster of blended peaks, may be observed in the A-scan depending on the inclusion shape. Similarly, multiple zones of high intensity reflections are observed in the phased array S-scan.
Figure 2: Phased array ultrasonic testing (PAUT) data from slag inclusion.
Phased array data from a root lack of fusion in a 0.375” thick steel plate is shown in Figure 3. The weld in case is a single vee with a very narrow root gap. An extremely strong corner trap indication at very high intensity is observed at the plate bottom – roughly 0.10 inches from the weld centerline. The focal law cursor in the S-scan is placed at the point of maximum intensity – roughly 90% full screen height (%FSH).Compared to the slag inclusion, this type of defect is detected at excellent signal-to-noise ratio (SNR) and is comparably narrow in width in both the S-scan and A-scan.The excellent SNR is due to the geometry of the corner trap. Ultrasound is reflected efficiently from the sharp corner caused by the root lack of fusion.Interestingly, there is a smaller secondary indication in the S-scan that is traced back to ultrasonic energy traveling up the surface of the LOF and diffracting from the LOF tip.This secondary indication, at lower SNR, is comparable to the crack tip diffraction phenomenon.Also notice, this tip diffraction is located further in time, or distance, as additional time is required to travel up the LOF surface to the tip. In fact, this delay can be measured to accurately provide the height of the LOF in this case.