Visualizing engineering design data using a modified two-level self-organizing map clustering approach

Engineers tasked with designing large and complex systems are continually in need of decision-making aids able to sift through enormous amounts of data produced through simulation and experimentation. Understanding these systems often requires visualizing multidimensional design data. Visual cues such as size, color, and symbols are often used to denote specific variables (dimensions) as well as characteristics of the data. However, these cues are unable to effectively convey information attributed to a system containing more than three dimensions. Two general techniques can be employed to reduce the complexity of information presented to an engineer: dimension reduction, and individual variable comparison. Each approach can provide a comprehensible visualization of the resulting design space, which is vital for an engineer to decide upon an appropriate optimization algorithm. Visualization techniques, such as self-organizing maps (SOMs), offer powerful methods able to surmount the difficulties of reducing the complexity of n-dimensional data by producing simple to understand visual representations that quickly highlight trends to support decision-making. The SOM can be extended by providing relevant output information in the form of contextual labels. Furthermore, these contextual labels can be leveraged to visualize a set of output maps containing statistical evaluations of each node residing within a trained SOM. These maps give a designer a visual context to the data set’s natural topology by highlighting the nodal performance amongst the maps. A drawback to using SOMs is the clustering of promising points with predominately less desirable data. Similar data groupings can be revealed from the trained output maps using visualization techniques such as the SOM, but these are not inherently cluster analysis methods. Cluster analysis is an approach able to assimilate similar data objects into “natural