Spectral physics-informed neural network for transient pipe flow simulation.

Area of Science:

• Fluid dynamics
• Computational hydraulics
• Machine learning in engineering

Background:

• Accurate wave propagation modeling is crucial for water supply pipeline monitoring and localization.
• Physics-Informed Neural Networks (PINNs) integrate physical laws with data but are typically limited to time-domain transient wave analysis.
• Frequency domain models offer superior sensitivity for system identification and anomaly detection in pipelines.

Purpose of the Study:

• To develop a novel Physics-Informed Neural Network (PINN) model for transient wave propagation in the frequency domain.
• To enhance wave prediction accuracy for pipeline monitoring and assessment applications.
• To investigate the model's capability in handling uncertainties and detecting anomalies like leaks.

Main Methods:

• Development of a Physics-Informed Complex-Valued Neural Network (PICVNN) for frequency domain water hammer modeling.
• Integration of physical principles with complex-valued neural networks to analyze transient wave data.
• Comparative analysis against classical complex-valued neural network (CVNN) benchmarks with varying observation points.

Main Results:

• The PICVNN model accurately reconstructs transient pressures, outperforming classical CVNN models in predictive accuracy.
• The model demonstrates robustness in handling uncertainties in input parameters, mathematical models, and identifying unknown leaks.
• PICVNN achieves higher accuracy but requires a longer training duration compared to traditional CVNNs.

Conclusions:

• The developed PICVNN is an effective tool for frequency domain transient wave analysis in pipelines.
• PICVNN serves as a reliable signal fusion method, enhancing accuracy and reliability by integrating diverse sensor data.
• This approach advances the application of PINNs in pipeline monitoring and system identification.