Deep feature extraction and fault diagnosis of solenoid valves in propulsion systems based on temporal-attention GraphLSTM model

Fault diagnosis plays a crucial role in monitoring and maintaining industrial processes and equipment such as discovering solenoid valve faults. Given the intricate nature of complex industrial processes characterized by nonlinearity and dynamics, this paper introduces a novel Temporal-Attention Graph Long Short-Term Memory (TA-GraphLSTM) model for fault diagnosis, which smoothly integrates a hybrid GraphLSTM network module with a temporal-attention block. The novel architecture leverages the strengths of both graph and LSTM neural networks to effectively handle the complexities inherent in industrial data. To construct the input graph structure data, we develop a variable correlation analysis strategy based on the Maximum Information Coefficient (MIC), which facilitates accurate representation of the relationships between variables. Furthermore, the incorporation of the temporal-attention module allows the model to dynamically assign weights to hidden variables across time steps to capture temporal dependencies effectively. The proposed TA-GraphLSTM fault diagnosis method is validated through the application of a spacecraft propulsion system. The experiment results prove the effectiveness and robustness of the model.