| Title |
TinyML-based Low Power Smart Metering System |
| Authors |
이승헌(Seungheon Lee) ; 최경택(Kyongtaek Choi) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.2.89 |
| Keywords |
TinyML; Otsu threshold; Blob filtering; Smart metering; nRF52840 |
| Abstract |
There are three main types of meters, which are divided into mechanical, digital, and cloud server types. Each meter has its advantages and disadvantages: mechanical requires a person to check the number directly, digital requires that the charging be impossible when the battery is discharged, and cloud server type has problems that increase communication costs and replace the existing charging system because it transmits video data. Therefore, in order to solve the problems of the basic methods, this paper proposes a method of attaching an edge device for TinyML equipped with a camera to a mechanical meter to automatically recognize the number of the meter at the edge device and transmit only the recognized number through the existing charging system. The algorithm created through transfer learning shows high accuracy (99.07%). In addition, the efficiency of DNN is important as the proposed method operates on low-capacity edge devices. Therefore, a very light network consisting of two convolutional layers and two FC layers was applied post-quantization to the edge device. In addition, in order to recognize the state in which the number of meters is converted (e.g., from 0 to 1), existing studies have designated a new class for each state of conversion, whereas the proposed method was recognized using a traditional image processing algorithm, which simplified the network. The accuracy of the algorithm for recognizing the state in which numbers are transformed was 100%. The proposed algorithm ported to the nRF52840 micro controller, and if the number of meters is recognized every 6 hours, the current consumption for one year is estimated to be about 121mAh. |