Data-driven fault detection and isolation inspired by subspace identification method

In this paper, we present a data-driven fault detection and isolation method for linear time-invariant system. This approach is inspired by a newly popular subspace identification method called predictor-based subspace identification. Residual estimators can be generated without any prior knowledge about mechanism of the system through the proposed method. These estimators are actually evaluations of different kinds of faults, based on which a data bank of faults can be built. When new operating data of the system is obtained, we input them into the data bank. The fault, if exists, can be detected and isolated. Simulation results based on the benchmark of Tennessee Eastman process demonstrate the validity of the proposed approach.