A deep autoencoder with structured latent space for process monitoring and anomaly detection in coal-fired power units

Abstract

In the big data era, deep autoencoder (DAE)-based methods for anomaly detection are widely used in monitoring coal-fired power units (CFPUs). However, these methods often overlook essential latent space information crucial for detecting anomalies within the DAE model. This study presents a structured latent space deep autoencoder (SLSDAE) that not only intuitively provides both latent space and reconstruction residual information for anomaly detection but also obviates the need for additional hyperparameters in the model's loss function. Furthermore, by leveraging the support vector data description (SVDD) model, this research extracts anomaly discrimination criteria from the SLSDAE model and introduces an end-to-end, real-time online monitoring framework for CFPUs. Comparative analysis on four public datasets demonstrates that the SLSDAE model enhances the G-mean in anomaly detection by 16.05 % over the DAE model and surpasses the performance of both the βVAE and DAGMM models. When applied to an actual induced draft fan, this framework effectively provides clear status trend tracking and early anomaly detection, up to 20 days in advance.