Lowell E. Kolb

Statistical comparison of site-to-site measurement reproducibility

There have been many attempts to assess the reproducibility of EMC measurements between test sites, with varying degrees of success dependent upon implementation details. Hewlett-Packard has a successful program of site-to-site reproducibility evaluations using a custom-designed RF source as a reference, and applying statistical techniques to include the day-to-day variability of each site in the collected data. Comparing the sites to the average (for lack of a reference site), 95% of the measurements agree within 3.8 dB between 30 and 100 MHz and between 550 and 1000 MHz, and within 2.8 dB between 100 and 550 MHz. The program benefits the participating sites by answering the question of how much variability can be expected from day to day and from that site to another, as well as detecting and helping to identify sources of measurement error. The paper also includes evidence that a 1/r correction going from 3 m to 10 m test distance involves errors of about 8 dB.

Opportunities for improved 80%/80% statistical methods with CISPR 32

CISPR 22 and CISPR 32 both require statistical compliance, specifically that products demonstrate with at least 80% confidence that at least 80% of the product population passes the radiated and conducted emissions limits. There have been problems applying the statistical test specified in CISPR 22 to real products because the method assumes a normal distribution, and products often exhibit characteristics which do not fit that model. For example, if the shielding or grounding in electronic equipment is not consistent unit to unit, it can result in bimodal or multi-modal distributions of some emissions. The bimodal distribution may be a common cause behind signal frequencies which appear as significant emissions in some units but are unmeasurable in others. CISPR 32 introduces more options for tools which may be used to demonstrate 80/80 compliance, notably the binomial distribution test, which is robust enough to accommodate a wide variety of population distributions. When CISPR 32 replaces CISPR 22 as the emissions standard, this binomial test method should be available for product qualifications.