Soft sensors are used to estimate process variables that are difficult to measure online. However, the predictive accuracy gradually decreases with changes in the state of chemical plants. Regression models can be updated, but if the model is updated with abnormal data, the predictive ability deteriorates. In practice, when the prediction error of an objective variable exceeds a threshold, an abnormal situation is detected. However, no effective method exists to decide this threshold. We have proposed a method to estimate the relationships between applicability domains and the accuracy of prediction of soft sensor models quantitatively. The larger the distances to models (DMs), the lower the estimated accuracy of prediction. Hence, the model between DMs and accuracy can separate variations in process variables and y-analyzer fault. This method was applied to real industrial data. The fault detection ability of the proposed method was better than that of the traditional one. (C) 2010 American Institute of Chemical Engineers AIChE J, 57: 1506-1513, 2011