Efficient integration of reverse engineering (RE) tools into the product development cycle remains constrained by the cost of specialized hardware and the need for expert operators, especially as geometric complexity increases. 3D CNN semantic segmentation pipelines that use voxelated signed distance fields (SDFs) as network inputs present a promising path toward reducing manual involvement and high-performance computing dependency through automated, spatially decomposed analysis. However, segmentation performance is strongly influenced by dataset- and pipeline-specific parameters inherent to voxelated SDF processing. This work identifies and analyses these parameters and ranks their relative influence on segmentation robustness and accuracy in a exemplary CNN-based segmentation pipeline. Insights are derived using a Taguchi L9 experimental design combined with ANOVA, estimating parameter contributions under varied dataset conditions and training initializations. We observe that voxel size has the strongest influence on segmentation robustness, while narrowband width delivers consistent but smaller gains, and sliding window overlap (SWO) provides comparatively modest benefits despite maintaining robust system behaviour across tested conditions. From these findings, we formulate practical guidelines for constructing robust and reliable SDF-based training datasets, reducing tuning overhead and accelerating future deployment of automated 3D segmentation for surface reconstruction workflows in RE.