Data-Driven Pressure Sensor Subset Selection for Long-Distance Water Transfer Pipelines: Q-DEIM Benchmarking with...

Area of Science:

• Civil Engineering
• Hydraulic Engineering
• Sensor Networks

Background:

• Conventional long-distance water pipelines utilize dense, often redundant sensor networks for monitoring.
• Existing instrumentation practices lead to over-sampling and increased maintenance costs.
• Effective data reconstruction from sparse sensor networks is crucial for efficiency.

Purpose of the Study:

• To determine the minimum number of pressure sensors required to accurately reconstruct the full pressure field in a long-distance water pipeline.
• To benchmark various sensor selection strategies for optimal network reduction.
• To evaluate the trade-offs between sensor count, reconstruction accuracy, and spatial coverage.

Main Methods:

• Analysis of 7239 hourly pressure snapshots from a 122 km trunk pipeline in northeast China.
• Application of data filtering to retain 60 sensors from an initial 72 deployed sensors.
• Benchmarking of sensor selection strategies: spatial farthest-point, PCA leverage, Q-DEIM, and a hybrid Q-DEIM with spatial diversity penalty.
• Reconstruction using uniform Tikhonov-regularised gappy Proper Orthogonal Decomposition (POD).

Main Results:

• The pressure field exhibits an effective rank of 15 at the 99% cumulative variance level, indicating significant over-sampling.
• Q-DEIM and the hybrid method achieved R2=0.982 (RMSE 0.96 mH2O) with only 15 sensors, representing a 75% reduction.
• Achieving R2≥0.99 required 26 sensors for Q-DEIM and 42 for the hybrid method.
• At 20 sensors, the hybrid method improved worst-case fill distance by 3.3× compared to RMSE increase.

Conclusions:

• Sensor network density in long-distance water pipelines can be substantially reduced without compromising pressure field reconstruction accuracy.
• Q-DEIM and hybrid Q-DEIM strategies offer significant sensor reduction potential, outperforming random and spatial sampling methods.
• Regime coverage within the training data library is identified as a critical constraint for sensor deployment optimization.