This thesis documents a PhD research program undertaken at Swinburne University of Technology between the years 2000 and 2004. The research was funded by the Cooperative Research Centre for Cast Metals Manufacturing and was undertaken in collaboration with Nissan Casting Plant Australia Pty Ltd and the Ford Motor Company Australia Limited. This thesis reports on the investigation of the possibility of using an ultrasonic sensing-based, non-destructive testing system to detect gas porosity defects in aluminium die casting parts with rough surfaces. The initial intention was to develop a procedure to obtain ultrasonic signals with the maximum possible amplitude from defects within the rough surface areas of the castings. A further intention was to identify defects with the application of a suitable signal processing technique to the raw ultrasonic signal. The literature review has indicated that ultrasonic techniques have the potential to be used to detect subsurface defects in castings. The possibility of classifying very weak ultrasonic signals obtained from rough surface sections of castings through a neural network approach was also mentioned in the literature. An extensive search of the literature has indicated that ultrasonic sensing techniques have not been successfully used to detect sub-surface defects in aluminium die castings with rough surfaces. Ultrasonic inspection of castings is difficult due to the influence of microstructural variations, surface roughness and the complex shape of castings. The design of the experimental set-up used is also critical in developing a proper inspection procedure. The experimental set-up of an A-scan ultrasonic inspection rig used in the research is described in this thesis. Calibration of the apparatus used in the inspection rig was carried out to ensure the reliability and repeatability of the results. This thesis describes the procedure used to determine a suitable frequency range for the inspection of CA313 aluminium alloy castings and detecting porosity defects while accommodating material variations within the part. The results obtained from ultrasonic immersion testing indicated that focused probes operating at frequencies between 5 MHz and 10 MHz are best suited for the inspection of castings with surface roughness Ra values varying between 50 [micro milli] and 100 [micro milli]. For the purpose of validating the proposed inspection methodology, gas porosity defects were simulated through side-drilled holes in the in-gate section of selected sample castings. Castings with actual porosity defects were also used in this research. One of the conclusions of this research was that it was extremely difficult to detect defects in castings with surface roughness above 125 [micro milli]. Once the ultrasonic signal data was obtained from the sample aluminium die castings with different surface roughness values ranging from 5 [micro milli] to 150 [micro milli] signal analysis was carried out. Signal feature extraction was achieved using Fast Fourier Transforms (FFT), Principal Component Analysis (PCA) and Wavelet Transforms (WT) prior to passing the ultrasonic signals into a neural network for defect classification. MATLAB tools were used for neural network and signal pre-processing analysis. The results indicated that poor classification (less than 75%) was achieved with the WT, PCA and combination of FFT/PCA and WT/PCA pre-processing techniques for rough surface signals. However, the classification of the signals pre-processed with the combination of WT/FFT, FFT/WT and FFT/WT/PCA classifiers provided much better classification of more than 90% for smooth surface signals and 78% to 84% for rough surface signals. The results obtained from ultrasonic testing of castings with both real and simulated defects were validated with X-ray analysis of the sample castings. The results obtained from this research encourage deeper investigation of the detection and characterisation of sub-surface defects in castings at the as-cast stage. Implications for the industrial application of these findings are discussed and directions for further research presented in this thesis.