Evaluation of signal processing tools for improving phased array ultrasonic weld inspection

Abstract

Cast austenitic stainless steel (CASS) that was commonly used in U.S. nuclear power plants is a coarse-grained, elastically anisotropic material. In recent years, low-frequency phased-array ultrasound has emerged as a leading candidate for the inspection of welds in CASS piping, due to the relatively lower interference in the measured signal from ultrasonic backscatter. However, adverse phenomena (such as scattering from the coarse-grained microstructure, and beam redirection and partitioning due to the elastically anisotropic nature of the material) result in measurements with a low signal-to-noise ratio (SNR), and increased difficulty in discriminating between signals from flaws and signals from benign geometric factors. There is therefore a need for advanced signal processing tools to improve the SNR and enable rapid analysis and classification of measurements. This paper discusses recent efforts at PNNL towards the development and evaluation of a number of signal processing algorithms for this purpose. Among the algorithms being evaluated for improving the SNR (and, consequently, the ability to discriminate between flaw signals and non-flaw signals) are wavelets and other time-frequency distributions, empirical mode decompositions, and split-spectrum processing techniques. A range of pattern-recognition algorithms, including neural networks, are also being evaluated for their ability to successfully classify measurements into two or more classes. Experimental data obtained from the inspection of a number of welds in CASS components are being used in this evaluation.