[en] In this paper, we propose a method for numerically evaluating the probability of failure from the diagnostic results of condition monitoring using the Bayesian theorem. When performing maintenance using real-time monitoring results, a diagnostic result without any inspection error is ideal. However, failure is not caused even if the monitoring method sufficiently overestimates any small damage that can cause failure. Moreover, failure is not caused even if the method slightly underestimates large damage. In other words, for reducing the probability of failure, improving the accuracy of estimating specific damage levels is necessary. In this study, a method for reducing the risk by improving the diagnostic accuracy of specific damage levels by means of controlling the sampling ratio of the training data employed for learning the use of a weight function is proposed. The consequences of overestimation and underestimation of damage differ. The risk caused by underestimation is called failure risk and that caused by overestimation is called economic risk. In this paper, the effect of weighted regression on risk reduction is discussed. The proposed method is validated by employing it to identify delamination in a CFRP beam via the electric potential change method. (author)