Zero Flow Rate Detection of Ultrasonic Water Meter Using Machine Learning Techniques

Abstract

This paper presents a classification-based method to detect false-positive ultrasonic sensor measurements when the flow rate is zero while the pipeline is full of water caused by the physical disturbances in flow metering. Due to the outlined disturbances, these false-positive readings of the meters may cause the wrong billing. To overcome this problem, ultrasonic sensor measurements are collected as timeseries data at variously different water flow rates. Numerical and statistical measures of the timerseries are computed to construct an input-output relation. Hence, the feature extraction process is performed. For the modeling phase, both zero flow rate and non-zero flow rate parts of the dataset, and its corresponding features are labeled. Logistic Regression (LR), Support Vector Machine (SVM), and Linear Discriminant Analysis (LDA) algorithms are employed to classify the flow status in the MATLAB environment. Model performances are compared in terms of accuracy, sensitivity, specificity, and precision. For the investigation of the availability of the implementation of an embedded system, model complexities are discussed. Selected model parameters are embedded in a water meter, and consumption values are compared to a water meter without the detection algorithm in the same test bench underfilled pipeline with zero flow rate condition. To simulate the physical disturbance conditions, and observe the effect of the false-positive detection algorithm on flow metering, the flow pipeline is vibrated periodically. Consumption loggings of the water meters are tabled, model performance results are discussed, and test results are shared. According to the results, the radial basis kernel SVM algorithm performs better in terms of all metrics. LR algorithm is employed for the real plant experiment when its model complexity is considered.