Flow regime identification for air valves failure evaluation in water pipelines using pressure data

Area of Science:

• Hydraulics and Fluid Mechanics
• Signal Processing
• Machine Learning Applications

Background:

• Air accumulation in water transmission pipelines due to air valve failure can lead to reduced capacity, vibrations, and catastrophic failures.
• Downward sloping pipes are prone to air accumulation, causing flow regime transitions that are indicative of air valve issues.
• Current research on flow regime identification for air valve fault diagnosis is limited.

Purpose of the Study:

• To develop a reliable method for identifying air valve operational states in freshwater pipelines.
• To utilize pressure signals and machine learning for fault diagnosis of air valves.
• To investigate the effectiveness of Support Vector Machines (SVMs) in classifying different flow regimes.

Main Methods:

• Laboratory experiments were conducted to collect pressure data for four common flow regimes: bubbly, plug, blow-back, and stratified flow.
• Two SVM models were developed and trained to identify specific flow regimes (bubbly and plug flow).
• Analysis of signal features, including Power Spectral Density and Short-Zero Crossing Rate, was performed to determine optimal indicators for SVM classification.

Main Results:

• Pressure signals are effective indicators for identifying flow regimes and thus the operational status of air valves.
• Power Spectral Density and Short-Zero Crossing Rate were identified as the most effective features for SVM-based flow regime classification.
• Optimal selection of SVM features and pressure signal parameters resulted in identification accuracies exceeding 93%.

Conclusions:

• SVM-based flow regime identification using pressure signals is a promising approach for the fault diagnosis of air valve failures in water pipelines.
• The method provides a reliable means to distinguish between functioning and malfunctioning air valves.
• Further optimization of signal processing and machine learning parameters can enhance diagnostic accuracy.