Fault diagnosis for dynamic system based on the independent latent space reconstruction of generative adversarial network

Abstract

Data-driven fault diagnosis for dynamic process faces three challenges. Firstly, models are hard to establish to describe the multivariate coupled correlations. Secondly, in an actual industrial process, the cost of fault data collection is huge, and fault training data is often scarce. Fortunately, a lot of normal operating data can be sorted, however, traditional fault diagnosis methods have insufficient processing capabilities for big data which is the third challenge. In view of the above problems, an improved long short-term memory generative adversarial network (LSTM-GAN) is introduced to utilize the time sequence normal big data to train a fault detection model and the reconstruction error of generator and inverter is proposed as the fault indicators. Then, independent latent space is obtained by constructing an inverter and the fault related subspace is extracted for fault type diagnosis. Experiment results based on the benchmark dataset of the Tennessee Eastman Process (TEP) show the effectiveness of the proposed method from the aspects of false alarm rate, miss alarm rate and diagnostic accuracy.