| Popis: |
Water pipes have exceeded their service life, and the number of leakage and burst accidents is increasing. These accidents cause economic losses due to the interruption of operations and supply. However, replacing all the deteriorated pipes is difficult, as this would involve a huge cost and long time. Therefore, the condition of water pipes must be quantified, and an efficient maintenance management is required. Current sensors for detecting water leakage have many problems because their measurement principle is a passive system, and thus they are susceptible to external noise. In our previous study, we proposed an active vibration sensing-actuation device to improve the measurement reliability for a water pipe deterioration diagnosis. The principle of deterioration diagnosis was based on the detection of the frequency change of the in-plane bending mode. In that study, the thresholds of discrimination were determined by the amplitude of the response based on empirical knowledge, which did not allow for theoretical or systematic discrimination. In this study, we focused on a simplified discrimination method of deterioration of a water pipe and its implementation with an IoT sensor module. The following three methods were considered in terms of implementation cost: an absolute threshold method based on theoretical random vibration analysis (based on a physical model), a linear discrimination method, and a support vector machine (based on a machine learning model), and the most appropriate discrimination method was determined. Furthermore, we considered the discrimination accuracy of models based on physics and machine learning. The results show that the support vector machine has the best accuracy, followed by the absolute threshold method, which has good accuracy. However, when compared in terms of computational complexity, the absolute threshold method is superior from both perspectives in terms of implementation. However, the application of the absolute threshold method is limited to the case where the physical characteristics of the sensing target are well known. In the case of unknown physical characteristics of the target, the method based on machine learning can be applied. |
