Bubble recognition algorithms for the processing of wire-mesh sensor data

Wire-mesh sensors deliver sequences of instantaneous, two-dimensional gas fraction distributions from a measuring cross-section. The spatial and temporal resolutions are high enough to reflect individual bubbles in much more than one measuring frame during their passage through the sensor plane. The bubbles can therefore be detected by a pattern recognition method based on the identifications of connected areas of gas-filled elements in the array of local instantaneous gas fractions, available as a function of the lateral co-ordinates x,y and as function of time. Algorithms were presented in earlier publications, which are based on a recursive fill procedure that assigns individual bubble numbers to the elements of the gas fraction array. This procedure has to rely on a threshold for the cancellation of the recursive fill, because the gas fraction data are affected by noise. The choice of the correct threshold is difficult, since a too high gas fraction threshold leads to unrealistic bubble fragmentation, while a too low threshold provokes unrealistic unification of bubbles. In previous papers an optimum approach was found by introducing a differential threshold determined for each bubble individually from the maximum gas fraction inside the given bubble to identify the cores of the bubbles. This is combined with an agglomeration step, where elements with a gas fraction below the threshold are assigned to already existing bubble cores, if these elements were not yet assigned to a bubble in the recursive fill step.
In the present paper, the algorithm will be analysed by applying it to synthetic measuring data, which is generated by placing bubbles of known geometrical parameters into the discretization measuring grid defined by the wire-mesh geometry. The discretization is performed with due care to reflect the fact that the gas-liquid interface at the periphery of the bubbles crosses control volumes of the measuring grid, producing local instantaneous void fractions of intermediate values. The synthetic signals are furthermore superposed by synthetic white noise. It is shown that the original bubble recognition algorithm needs to be improved in certain cases: (a) Unrealistic fragmentation in axial (time) direction may occur, when bubbles cross the sensor plane with a low velocity. (b) non-realistic coalescence becomes dominating at very high gas fractions, where bubbles of complex shape occur. Different improvements were tested. Unrealistic coalescence is avoided by introducing a criterion to terminate the recursive fill algorithm, which are based on the local gradient of the gas fraction. A second approach relies on a repair algorithm, which unites unrealistic fragments in an additional processing step after the preliminary bubble recognition. This repair algorithm takes benefit from a-priory knowledge to distinguish realistic close neighbour bubbles from unrealistic bubble fragments divided by an early truncation of the bubble recognition. After the check against synthetic bubbles, the algorithms are applied to real measuring data from gas-liquid flows in vertical pipes and the results obtained by the different algorithms are compared.