From Not Enough to Too Much: using SEA and Principle Components Analysis to predict and handle data variability. by John Holt

Several of the technical papers presented at the 2005 SAE Noise and Vibration conference focused on the statistical evaluation of data. I sat in on the presentation of two papers that spanned the extremes of data handling, and how to interpret and use data that is available to the NVH engineer.

On the Not Enough end of the spectrum, Lear Corporation presented a paper exploring SEA modeling techniques used to predict vehicle interior NVH performance based on a very small sample size, utilizing Gage R&R studies to solidify many of the testing variables. At times where vehicle availability limits the capability to produce statistically robust data sets, it is necessary to employ as many strategies as possible to identify and predict vehicle and system variation by reducing errors in testing techniques and system variables. The authors used a controlled Noise Reduction test utilizing a reciprocal path method to establish a predictive model and specification width. These predictions were then used to evaluate an operational vehicle test for rear-tire noise reduction capabilities. They found that while these efforts help to build a reasonable predictive model, SEA is still incapable at this level to provide reasonable vehicle-level variation estimates when looking at several vehicles with respect to NR performance.

Paper 2005-01-2553 : A Study of NVH Vehicle Testing Variability by Terence Connelly, Jud Knittel and Mark Jay, Lear Corporation

Conversely, problems can arise when you have Too Much data to digest in an efficient manner. Christian Fernholz from Visteon Corporation described how in benchmarking studies where many data points are available, it becomes cumbersome to sift through data sets from numerous samples to draw reasonable conclusions if the spread of data is wide. Mr. Fernholz presented a benchmarking case where 99 steering pumps were tested using 12 different NVH measurements. The challenge lies in evaluating critical measurements and establishing rating techniques to quantify major contributors to NVH performance. Principle Components Analysis (PCA), a multivariate statistical tool, employs an eigenvalue approach to rank principle components, or measurement variables (i.e vector sum point acceleration, SPL, line pressure, etc.), according to their influence on test sample variation. This information can then be used to rank the test specimens in terms of these two or three identified principle components, and gross conclusions can be drawn about a large field of samples. At a high level, this method can identify noticeable differences and trends in large sample sizes that would warrant further, more detailed investigation.

Paper 2005-01-2518: Multivariate Statistical Methods for the Analysis of NVH Data by Christian Fernholtz, Visteon Corporation