Advancements in Gas Turbine Fault Detection: A Machine Learning Approach Based on the Temporal Convolutional Network–Autoencoder Model

Abstract

To tackle the complex challenges inherent in gas turbine fault diagnosis, this study uses powerful machine learning (ML) tools. For this purpose, an advanced Temporal Convolutional Network (TCN)–Autoencoder model was presented to detect anomalies in vibration data. By synergizing TCN capabilities and Multi-Head Attention (MHA) mechanisms, this model introduces a new approach that performs anomaly detection with high accuracy. To train and test the proposed model, a bespoke dataset of CA 202 accelerometers installed in the Kirkuk power plant was used. The proposed model not only outperforms traditional GRU–Autoencoder, LSTM–Autoencoder, and VAE models in terms of anomaly detection accuracy, but also shows the Mean Squared Error (MSE = 1.447), Root Mean Squared Error (RMSE = 1.193), and Mean Absolute Error (MAE = 0.712). These results confirm the effectiveness of the TCN–Autoencoder model in increasing predictive maintenance and operational efficiency in power plants.