Flow pattern recognition model based on deep neural networks using only an in-service dual-frequency Coriolis flowmeter

• Proposes a novel gas-liquid two-phase flow pattern recognition method that uses only an in-service dual-frequency CMF—no additional sensors required—to reduce industrial hardware costs and capture more flow-pattern-related variables from the CMF’s multifrequency resonance responses. • Adopts an industry-suitable “time-series curve images + well-proven pre-trained CNNs” approach: it converts selected variable curves into visual images, enabling direct utilization of well-proven, off-the-shelf pre-trained deep neural networks (DNNs: AlexNet, GoogLeNet) for feature extraction from variable curve images and flow pattern recognition. • The proposed GoogLeNet-based and AlexNet-based deep learning models both achieve effective identification of different flow patterns. • Relies solely on an in-service dual-frequency CMF, ensuring the method is cost-effective and easily integrable into existing industrial systems. Given the dual-frequency Coriolis Mass Flowmeter (CMF)’s advantages of multi-variable output and high-precision measurement, this study proposed a novel online flow pattern recognition method based on deep neural networks (DNNs), which only utilizes an in-service dual-frequency CMF. By using only an in-service dual-frequency CMF (no additional sensors), this method reduced hardware costs for industrial applicability, and collected more flow-pattern-related variables and features from multi-frequency resonance responses of the CMF. Specifically, the same two-phase fluid in the measuring tube exerts distinct effects on the 1 st-order and 3 rd-order vibration modes of CMF’s measuring tube, and these effects can provide richer features for flow pattern recognition, including 1st -order damping ( D 1 ) ,1st -order resonance frequency ( f 1 ) , and 3rd -order resonance frequency ( f 3 ). Subsequently, variables capable of characterizing different flow patterns were chosen, and their time-series curves (marked with different colors) were converted into visual curves’ images. Two DNN-based (based on GoogLeNet and AlexNet) flow pattern recognition models were established respectively, which were well-proven to excel in image feature extraction and classification. demonstrate that both models can effectively identify different flow patterns, with overall recognition accuracies of 96.9 % and 98.7 %, respectively; the AlexNet-based model exhibits higher recognition accuracy. The method uses only an in-service dual-frequency CMF with no additional sensors, making it cost-effective and easy to integrate into existing industrial systems. It also lays a foundation for developing flow pattern-based metering correction models for in-service dual-frequency CMFs.