Based on the dynamics models and their identification of mechanical-electronic systems of flexible machining devices, we have built up a real condition monitoring model which includes BAYES fault detection, abrupt fault detection and fault analysis&location. Utilizing the Kullback information distance and the multilevel hierarchical AR model, we have realized the prediction and alarming of faults of flexible machining devices. The experiments on the JCS-020 Machine Center have indicated the models are feasible and valid. They can properly eliminate the influence of the changeful machining mode and process, the factors of time-varying and random disturbances on the condition monitoring, fault diagnosis and prediction, and greatly improve robustness and adaptability of monitoring and diagnosis in the FMS environments.