A GAN-based fault detection for dynamic process with deconvolutional networks

Abstract

Aiming to overcome the difficulty to obtain the fault data of practical system, a fault detection approach using health data only is proposed based on the whole space of the system being divided into the fault status and the fault-free status. Firstly the time series of observation window is generated by a deconvolutional network with an input of initial data obtained by Monte Carlo method. The probability distribution of generated data approximates to the actual sample data by discriminator of generative adversarial network. Through continuous iteration, the health probability distribution is finally obtained in the whole space. Concurrently the discriminator is evolved into a fault detector which realizes the detection of new data. The effectiveness of the algorithm is demonstrated by a numerical simulation example based on a wind turbine benchmark model.