Sensor fault diagnosis based on on-line random forests

In order to reduce the memory requirement, and obtain real-time status of equipment, the paper proposed to use the the on-line random forests (ORFs) algorithm to identify sensor fault. The sample set is derived from Tennessee Eastman (TE) process. The models are updated by a group of sensor data, which are collected in each interval. As models are real-time and dynamic, the equipment could be tested at any time. Moreover, the samples obtained at previous intervals are not need to store. The results of experiments show that the accuracies of ORFs and Random Forests (RFs) are similar in sensor fault diagnosis process. And in some fast changing process, ORFs distinguishes fault types with higher accuracy, better adaptable and faster than RFs.