This paper presents a failure diagnosis of a V-belt thorough vibration monitoring. The belt vibration is monitored at points on a driven bearing. Seven types of basic causes of belt failures are considered. Power spectra of the vibration data are calculated through a noise reduction called a cross-spectrum method. Six parameters characterizing the vibration data are identified, and seventeen typical combinations of basic causes are diagnosed by a Bayes' discriminant function approach. Two types of incorrect diagnoses are considered : type I leaves a failed belt not repaired, and type II causes overmaintenance. Two types of losses are considered accordingly, and the ratio of them is determined so as to minimize the type I error. It turns out that the Bayes' approach using the six characteristic parameters successfully diagnoses the V-belt failure.