This paper presents a new condition monitoring method based on a latent process model. The method consists of three steps. First, a sensor signal is modeled by a latent process model that is a combination of a time-varying auto-regression model and a dynamic linear model, which decomposes the signal into several components, and each component represents a different part of the monitored system with different time-frequency behavior. Based on the latent process model, important features are extracted. Finally, using the generative topographic mapping, the selected features are mapped to a lower (two)- dimension space for classification. The proposed method is tested in condition monitoring of sheet metal stamping processes. A large number of experiments were conducted. In particular, two cases are presented in detail. From the testing results, it is found that the proposed method is able to detect various defects with a success rate around 98%. This result is significantly better than the conventional artificial neural network method. In addition, the new method is a self-organizing method and hence, little training is necessary. These advantages make the method very attractive for practical applications.