Physically Constrained Generative Adversarial Network Data Augmentation Method for Multichannel Ultrasonic Flowmeters of Natural Gas

Abstract

During the performance evaluation process of ultrasonic flowmeters, there are issues such as high cost of acquiring real flow data and severe scarcity of data. Traditional data augmentation methods suffer from problems such as low accuracy and poor diversity, making it difficult to effectively model ultrasonic flowmeter data. This paper proposes a method that incorporates physical constraints into a conditional Wasserstein Generative Adversarial Network with gradient penalties (CWGAN-GP) for ultrasonic flowmeters data augmentation. Experimental results demonstrate that the data generated by our model meets the requirements in terms of accuracy and diversity, and closely aligns with the distribution of the original data. In the performance evaluation of ultrasonic flowmeters, the evaluation model derived from data generated by our method demonstrates a reduction in the average percentage error by 1.495% compared to the assessment using original data without data augmentation.