A data-driven approach of fault detection for LTI systems

This paper proposes a modified subspace aided data-driven fault detection method for linear time-invariant systems. The main merit of this method lies in the avoidance of identifying the mechanism-based model of a system. Inspired by subspace identification method, we construct parameterized matrices of residual signal directly from input and output data without any prior knowledge about mechanisms of a plant. Modified measures are adopted to reduce computational complexity of the algorithm. Fault detection then can be implemented successfully. Simulation studies on the benchmark of Tennessee Eastman process demonstrate the validity of the proposed approach.