Abstract: To effectively identify the faults occurring in the operation of auxiliary equipment, the multivariate state estimation technique was applied for the fault warning of related auxiliaries. The historical state matrix of these auxiliaries was divided into three categories according to the power level of the unit, while typical states were selected to establish submodels based on isometric sampling. For input observation vectors, corresponding estimation vectors were given, and the deviation between them was reflected by a similarity function. Based on interval estimation, the adaptive threshold of similarity degree was designed. Numerical simulations were finally conducted with the data obtained before coal clogging in the medium speed mill of a 350 MW thermoelectric unit. Results show that the model is able to send an early warning 264 s before coal mill tripping, indicating that the method has a high efficiency in fault detection, which helps to reduce the false alarm rate, compared with traditional fixed threshold methods.