An anomaly detection method for gas turbines in power plants using conditional variational autoencoder optimized with self-attention

An innovative anomaly detection (AD) framework designed for gas turbines (GTs) in power plants working under variable operating conditions (OCs) was proposed based on conditional variational autoencoder (CVAE) optimized with self-attention (SA) in this paper. Three key factors contribute to the mode...

Full description

Saved in:

Bibliographic Details
Published in:	Reliability engineering & system safety Vol. 267; p. 111894
Main Authors:	Chen, Dongchao, Li, Xiuxia, Xu, Jingquan, Wang, Zhong
Format:	Journal Article
Language:	English
Published:	Elsevier Ltd 01.03.2026
Subjects:	Anomaly detection Conditional variational autoencoder Gas turbine Self-attention Variable operating conditions COAT HI RCAT GTP TET RDCT CIAT Self-attention LS DL LV LRP FNR AMFormer KDE OCSVM EGDP HI-FM PCC DS Variable operating conditions MD LSD ML Pr FWT AD EH AE DCD IS EP LOP Anomaly detection LPT_flow_mod PCA ET RA LOT MAE AS LPT_eff_mod FGMMP FGTFCVP FGHOT ANI SIS HPT_eff_mod IGVCSO CBVP BPT FM CVAE AUROC FPR SA SVR Conditional variational autoencoder R Re OC PHM MFPLCSO SA-CVAE GD SP MSE Acc CC CSP OS COP GL RUL CPFM CP GT LFCPF TE EGT CPD CPF IGV IGVP Gas turbine GCV VAE TEP TIT HRSG
ISSN:	0951-8320
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	An innovative anomaly detection (AD) framework designed for gas turbines (GTs) in power plants working under variable operating conditions (OCs) was proposed based on conditional variational autoencoder (CVAE) optimized with self-attention (SA) in this paper. Three key factors contribute to the model's superior performance in AD. First, the marginal distributions (MDs) of parameters were transformed into conditional probability distributions (CPDs) by incorporating OC variables as conditional inputs in the variational autoencoder (VAE) model, and decoupled modeling of equipment states and external OCs is achieved. Second, by embedding SA modules, the method dynamically adjusts weight allocation of the parameters to capture complex dependencies among them, which can enhance the capability of the model to represent intricate coupling relationships in multi-dimensional parameters. Third, the logarithmic reconstruction probability (LRP), derived from the probabilistic output of SA-CVAE, serves as a unified health indicator (HI) to evaluate the health condition of the GT. The indicator not only accounts for the discrepancy between the reconstructed and original inputs but also incorporates the influence of latent variable (LV) variance on the reconstruction results. Experiments were carried out on one synthetic, one public, and two real-world datasets, with SA-CVAE compared against six leading AD methods. The results demonstrate that under varying OCs, SA-CVAE exhibits superior performance in AD tasks, achieving leading or competitive results across key evaluation metrics.
ISSN:	0951-8320
DOI:	10.1016/j.ress.2025.111894