| Summary: | In Non-Destructive Evaluation (NDE) inspection, a range of materials exist that exhibit a heterogeneous or acoustically scattering microstructure, for example austenitic steels and Inconel alloys, and are used extensively across many industrial sectors. When inspected using conventional ultrasound techniques, defect signals are significantly corrupted by noise from randomly distributed scatterer associated with the material microstructure. In some cases, even defects that are much larger than the grain size distribution in the microstructure can be difficult to detect. This Thesis presents an investigation into the development of new algorithms to suppress backscattering noise in the received ultrasonic echoes associated with both individual transducers and phased arrays. Using the fact that structural noise in these difficult materials is frequency coherent, frequency diversity based techniques like the well-known Split Spectrum Processing have been developed. However, conventional algorithms are either ineffective or sensitive to the variations of material characteristics, especially when the signal to noise ratio (SNR) is low. A frequency diversity based technique, Moving Bandwidth Split Spectrum Processing (MB-SSP), has been developed, which is less influenced by material characteristics. MB-SSP first selects an ascending series of frequency bands and a trace is reconstructed for each selected band in which a defect is present: this occurs when all frequency components are in uniform sign. Combining all reconstructed signals through averaging gives a probability profile of potential defect positions. A range of supervised machine learning techniques has also been employed to further improve detect capability, if the pre-acquired training data is available. Instead of investigating the structure and pattern of the spectrum of an individual echo, the proposed method focuses on the distinction between the ensembles of defect signals and clutter noise. A training process is used to establish the statistical analysis, based on which a hypothesis test is then applied to the received echoes to indicate defects. The approach is designed to be adaptive to the material microstructure and characteristics due to the statistical training aspect of the technique. The concept of applying clustering algorithms to further reduce the influence of artefact noise remaining in A-scan data after processing by MB-SSP or a conventional defect detection algorithm is also discussed. The segmental signals that potentially contain defects in the processed A-scans are clustered into groups. The distinction and similarity between each group and the ensemble of randomly selected noise segments can be observed by applying a classification algorithm. Each class will then be labelled as either a 'legitimate reflector' or 'artefact' based on this observation and the expected probability of detection (PoD) and the probability of false alarm (PFA) determined. Finally, the developed A-scan based noise reduction algorithms have been extended into phased array imaging. Here the techniques are applied to the raw Full Matrix Capture (FMC) datasets prior to processing by an appropriate imaging algorithm. Total Focusing Method (TFM) and the focused B-scan imaging is applied to both standard and pre-processed FMC datasets on both simulated data and experimental data from coarse-grained materials. Importantly, the background noise is significantly suppressed in every case using the pre-processed FMC data.
|
|---|
