Steam turbine anomaly detection: an unsupervised learning approach using enhanced long short-term memory variational autoencoder

Steam turbines, pivotal to thermal power generation, incur substantial costs and operational disruptions from downtime, maintenance, and damage. Precise anomaly detection is essential for their safe and stable operation. However, challenges such as inherent anomalies, limited temporal modeling, and high-dimensional sensor data often hinder existing approaches. This study proposes an Enhanced Long Short-Term Memory Variational Autoencoder with Deep Advanced Features and Gaussian Mixture Model (ELSTMVAE-DAF-GMM), an unsupervised framework tailored for anomaly detection in unlabeled steam turbine time-series data. Specifically, the model integrates Long Short-Term Memory Variational Autoencoder to map high-dimensional time-series data into a compact latent space. A Deep Autoencoder Local Outlier Factor algorithm filters inherent anomalies during training, sharpening the model's discriminative power. Additionally, we introduce Deep Advanced Features, which combine latent representations and reconstruction errors to provide a non-overlapping and structured data representation. Anomaly detection of the representation distribution is then estimated using a Gaussian Mixture Model. Comparative and ablation studies on real industrial steam turbine data collected from a thermal power plant in China demonstrate superior performance, with high accuracy (94.6%) and low false alarm rate (5.43%), outperforming baseline methods. © 2025 Elsevier Ltd.