基于小样本量的超声流量计使用中检验模型构建与优化OA

To address the problems of large sample size,long test cycle and low efficiency in the model construction of existing machine learning-based in-service inspection models for ultrasonic flowmeters,backpropagation(BP)neural network and random forest models under different sample sizes were established based on experimental data from the national urban gas flow standard device(uncertainty 0.26%,k=2).A small sample set was obtained by halving the initial sample set(time interval 6 s,30 s)using the arithmetic mean method.The results show that the prediction performance(evaluated by R² and root mean square erro)of both models under small sample size is lower than that of the initial sample size model,and the BP neural network has an overall better performance.To improve the performance of the small sample model,three feature optimization algorithms(ReliefF,Recursive Feature Elimination,and random forest feature algorithm)were adopted.The results indicate that the random forest feature algorithm achieves the optimal optimization effect,with the maximum improvement of prediction accuracy reaching 48.65%(at 884 m³/h flow point).The optimized small sample model retains 9～14 key features,and its prediction performance is comparable to that of the initial sample size model,with significantly improved modeling efficiency.Field application verification shows that the predicted indication error of the model is stable within±1%,which can meet the in-service inspection requirements of ultrasonic flowmeters in natural gas transmission stations.