Research On In-Use Inspection Method Of Gas Ultrasonic Flowmeter Based On Supervised Learning

Ultrasonic gas flow meter is one of the widely used measuring instruments in natural gas trade at present.In order to ensure the accuracy of measurement,the ultrasonic gas flow meter is verified periodically in the verification agency with the flow standard device in general.The ultrasonic gas flow meter can be verified with in-use inspection method,which can be treated as a more efficient and fast way to ensure the ultrasonic flowmeter measurement accuracy and prolong the verification period.This thesis proposes to apply "digital" metrology to the inspection of ultrasonic flowmeter in use.Based on the real flow measurement data of ultrasonic flowmeter and supervised learning algorithm,the prediction model of flow measurement deviation of ultrasonic flowmeter in use is constructed,and the uncertainty of flow deviation prediction is estimated at different flow points,which provides a new idea for the in-use inspection method.There were mainly conclusions gotten in this thesis:(1)The experimental platform for real-flow testing of ultrasonic flowmeters was built.For the size deviation between the inner diameter of the flowmeter and the installation pipe,the effects of deviations on measurement results were quantitatively analyzed by numerical simulation and real flow test.To avoid the extra influence of the installation effect of size deviation on the measurement results,the upstream and downstream straight pipe sections with no size deviation from the inner diameter of the flowmeter were used to carry out the research.(2)There were 6 flow rates points selected to carry out two sets of tests,while there were 15 indicators determined as the characteristic input of the sample.With consideration the influence of system noise on the ultrasonic signal,the Fourier transform and Gaussian window function was used to denoise the test data.The flow rates deviation prediction models of ultrasonic flowmeter were developed with random forest and BP neural network learning algorithms respectively.The results showed that the prediction results of the two models were not much different,and both models could effectively predict the deviation of flow measurement.(3)The optimal approximate distribution function of the input probability density function of each feature was obtained by calculating the KL divergence,and the uncertainty of the two prediction models was evaluated by using the probability density distribution of Monte Carlo propagation.The uncertainty was within(0.02 %～ 0.12 %)for the result from the random forest prediction model under each flow rates point,while it was within(0.14 % ～ 0.28 %)for the result from the BP neural network prediction model.The research results showed that both prediction models could achieve the measurement requirements of the accuracy level of the ultrasonic flowmeter,and the prediction model of ultrasonic flowmeter based on random forest algorithm showed better performance.