Fault Detection Using Autoencoder with Linear Hidden Layer Dynamics

This paper presents a novel structure for data-driven fault detection using multi-layer autoencoder whose hidden layer trajectory is embedded into a linear space. The hidden layer is constructed using the view of system behaviour and its trajectories are described by a kernel representation. Fault detection is carried out by monitoring the residue of the hidden layer. The proposed structure has the ability to capture complex nonlinear dynamical features while providing a clear representation for the hidden layer. Simulation studies on continuously stirred reactor (CSTR) is also conducted, which demonstrates that the proposed approach can detect faults effectively while distinguishing between faults and unknown disturbances.