Learning to discriminate phases in gas-liquid flow

Abstract

Discrimination of phases, i.e. substances of different aggregate states, in gas-liquid flow is a sensor data interpretation problem. Probes emitting two-state signals are usually used to detect local phase changes through a property, like conductivity, which is different for the two phases [1]. The probe signal then needs to be processed to obtain a particular flow characteristic. The processing methods typically reconstruct the two-state signal from a probe signal by static application of predefined threshold values. Although adequate to acquire certain flow features, the existing methods may in signal interpretation significantly disagree with expert understanding of the underlying physical phenomena.

To alleviate this problem, a discrimination technique has been introduced that incorporates human expertise into probe signal interpretation [2]. Development of the technique includes three stages: designing a prototype phase discrimination procedure, providing training probe and two-state signals, and tuning procedure thresholds. Unlike the existing procedures, the prototype discrimination procedure relates threshold values to local extrema detected in the input signal, what makes it less sensitive to noise and signal drift. In preparing training signals, a human expert is involved demonstrating his/her phase discrimination skill on a selected probe signal sequence. The resulting expert binary signal serves as a reference for threshold tuning. Phase discrimination procedure thresholds are finally tuned using a genetic algorithm. The optimization is performed so that the resulting phase discrimination procedure generates the binary output as close to the one defined by expert as possible.

The experimental verification on air-water flow under laboratory conditions has shown the new technique is suitable for gas-liquid flow measurements and applicable to various flow regimes appearing in industrial processes.