In order for a machine to operate safely, it is one of the very important factors to quickly and accurately detect an abnormality. A skilled operator can make a diagnosis from the sound produced by the machine. However, constant monitoring by the operator is costly and inefficient. If abnormalities can be detected and maintained at an early stage, it will lead to avoiding serious failures, and cost reduction can be expected by long-term use of the machine. In this study, we proposed a system that uses machine learning to detect abnormalities from the operating sounds of machines. Instead of the conventional judgment method that requires abnormal sounds in advance, a system that judges them only from learning normal sound data is proposed. In order to confirm the effectiveness of the proposed system, we build a machine operation sound generator and conducted experiments by artificially generating abnormal noise of a bearing. We extracted features called Mel-Frequency Cepstrum Coefficient from the sounds, abnormal sounds were detected by using One Class Support Vector Machine. As a result, it was confirmed that the generated abnormal noise was detected, and it was found that the proposed system was effective for abnormality detection.

機械が安全に動作するうえで，異常を逸早く正確に検知することは非常に重要な要素の1つである。しかし人間による検知では，24時間の監視をすることが困難である。そこで注目されているのが，機械学習である。機械学習では，過去の正常・異常のデータから，機械が傾向を学習し，正常・異常を判別することができる。しかしながら，機械学習を行うためには学習の元となる正常・異常のデータが必要となる。機械動作音の判定を行う際，正常音のデータはともかく，異常音のデータを入手するのは難しい。本研究では，機械学習を用いた機械動作音の正常・異常判定の有効性を確かめることを目的に，機械学習に必要な正常・異常データの収集を行った。具体的には，機械の正常動作音だけでなく，異常動作音も簡単に発生させることが可能な，機械動作音発生装置を用いて音と温度の測定を行い，測定データの解析も行った。