Residual Life Prediction of Pneumatic Control Valves Based on Trans-TCN-GRU Modeling

Abstract

Control valves, as key actuating components in process industrial systems, are widely used in major industrial fields such as coal chemical, petrochemical, and nuclear power. In recent years, with the gradual increase in industry scale and industrial technology, the operating conditions of control valves have become increasingly harsh. To ensure their safe and stable operation, accurately predicting their remaining useful life (RUL) is an important task. In current research on RUL prediction for control valves, challenges such as difficulty in obtaining fault data, limited real-world data, and information loss due to the use of only time-series data can lead to reduced prediction accuracy. To address these issues, this study collected a variety of simulated fault data based on a DAMADICS model built in Simulink, as well as real fault data generated from a pneumatic control valve test bench. The approach involved using an improved Transformer model for feature extraction from the data, applying a Temporal Convolutional Network (TCN) to mine temporal relationships among the data, and finally using a GRU module to extract features from non-time-series operating condition data. This resulted in a hybrid neural network model that uses full-scale data for predicting the remaining useful life of control valves. To validate the effectiveness of this model, the collected data was applied to it and compared with single models such as TCN and Transformer. The results showed that the Root Mean Square Error (RMSE) was reduced by 49.8% and 63.7%, the Mean Absolute Error (MAE) was reduced by 61.1% and 75%, and the Mean Absolute Percentage Error (MAPE) was reduced by 50.6% and 69.6%, respectively, demonstrating that the proposed method offers higher accuracy.