To prevent the accidents caused from the trouble of machines in the field of the welfare and/or medical treatment, health monitoring equipment that detects the trouble automatically on the earlier stage is desirable. The judgments by conventional acoustical health monitoring equipments are made by neural network with the analyzed frequency data. This article proposes a new pre-processing method in which the amplitude of frequency analyzed data is compressed based on the loudness function. Preparing for defect bottles as difficult examination examples to distinguish, it has been confirmed that the new method is able to distinguish the difference. The result of the inspection shows the percentages of the correct answer rise 66.3% to 95.9% by implementing non-linear processing. Therefore, it is concluded that this non-linear processing is effective to detect the defect characterized by relatively faint signals.