What is the inference accuracy loss rate of an industrial AI edge gateway for small target detection models?

Industrial AI edge gateways deploying small target detection models via INT8 quantization typically experience accuracy loss of <3%. However, under electrical fast transient (EFT) interference, DDR access errors can cause a sudden drop in accuracy. Countermeasures include connecting a PBZ1608E600Z0T ferrite bead in series with the DDR_VTT terminating resistor and adding a 2.2μF capacitor to stabilize the reference voltage. Simultaneously, a CMZ2012A-900T common-mode filter is added to the DDR clock line to reduce crosstalk. Actual testing on a security inspection device showed a 15% accuracy drop with EFT ±2kV without protection; after protection, the drop was <4%, and the recovery time was <1ms.