Electrical resistance tomography-based prediction for solid phase fraction by CNN with inaccurate samples

Electrical resistance tomography (ERT) faces critical challenges in solid-liquid two-phase flow measurement, particularly in mud monitoring, due to unreliable samples caused by uncertain mud conductivity and nonstationary working cycles. This paper proposes a hybrid data-model framework to enhance ERT-based solid-phase fraction (SPF) prediction. First, a kernel-based similarity metric with dynamic time warping (DTW) partitions time-series ERT data, achieving high accuracy in valid sample identification through double-linear weighting averaging (DWLA) that incorporates boundary distance and neighborhood density. Second, a bidirectional long/short-term memory (BiLSTM) network replaces conventional CNNs to capture temporal dependencies in conductivity fluctuations, integrating physical constraints from Maxwell's equations. Experimental results demonstrate significant improvements in both accuracy and stability. The proposed DWLA-BiLSTM framework provides a systematic solution for industry application under complex flow patterns.