The method of digital filtering by auto-associative neural network (AANN) is studied. The AANN adopts a special structure with abottle-nect layer and has an unit overall gain. The internal relation between the variables can be built after the training of a large amount of noise samples. The quintessence of the information can be extracted through the compression before the bottle-nect layer and the decompression after that. And then the redundant information can be used to restrain the noise during the measurement. The negtive effect of the noise upon the paramenter measurement of the aero-engine can be reduced maximally.