Incipient fault detection and process monitoring of thermal power plant pulverizing system based on deep representation learning

Abstract

Process data in contemporary thermal power plants shows the characteristics of large capacity, strong coupling, and high-order nonlinearity, which brings great challenges to process monitoring and fault detection. A deep representation learning fault detection scheme based on stacked sparse denoising autoencoder (SSDAE) is proposed in this paper. Specifically, to enhance the capabilities of noise reduction and feature representation for complex industrial data, the sparse denoising autoencoder (SDAE) is proposed by considering both noise and sparsity constraints. Then, a deep learning architecture is constructed by stacking multiple SDAEs layer by layer to achieve a highly nonlinear representation capability. Based on the low-dimensional representation and residual distance of SSDAE, three monitoring indicators, RE2, MD2, and ZD2, are designed by different distance metrics and the k-nearest neighbor (KNN) discriminant rule. The effectiveness of the proposed method is validated by studying a nonlinear numerical case and a practical power plant pulverizing system. The experimental results demonstrate that the proposed method can effectively detect incipient and slight faults that are difficult to detect with traditional methods.