Automatic analog meter reading for plant inspection using a deep neural network
- PDF / 2,174,677 Bytes
- 11 Pages / 595.276 x 790.866 pts Page_size
- 44 Downloads / 206 Views
ORIGINAL ARTICLE
Automatic analog meter reading for plant inspection using a deep neural network Yuki Funayama1 · Keita Nakamura1 · Kenta Tohashi2 · Taku Matsumoto1 · Akira Sato3 · Shigeki Kobayashi4 · Yutaka Watanobe1 Received: 13 April 2020 / Accepted: 25 September 2020 © International Society of Artificial Life and Robotics (ISAROB) 2020
Abstract The high demand for automatic plant inspections by robots has been because a plant has many dangerous locations to cover during daily inspections. In this study, we propose automatically reading an analog meter using a Deep Neural Network (DNN), because reading an analog meter is included in the daily inspections. In particular, we artificially generate training data including shooting noise and apply these data for training the DNN. The learned DNN is robust against readable angles, meter contamination, and meter light reflection, and achieves a reading absolute error within 0.05. Additionally, we verify the effectiveness of the proposed method using cross-validation and accuracy comparison with conventional methods. Keywords Plant inspection · Robot vision · Image processing · Machine learning · Deep neural network
1 Introduction Daily inspections at plants include reading pressures and thermometers, opening and closing valves, reading water levels, detecting rust, loose bolts, bad piping, and pumps, and measuring oxygen concentration in tanks. Therefore, the demand for automatic plant inspections by robots has been high, because a plant has many dangerous locations to cover during daily inspections. However, there is a major difference between the inspection ability required in a plant and the ability of the current robot. Many robot competitions [1] This work was presented in part at the 25th International Symposium on Artificial Life and Robotics (Beppu, Oita, January 22–24, 2020)
have been held to address this difference; the World Robot Challenge Disaster Robotics Category [2, 3] and RoboCupRescue Robot League [4] are typical examples of these competitions. In particular, meter reading is one of the main tasks for a plant inspection, because that reading can then be applied to other inspection tasks. For this reason, it is one of the tasks included in the robot competitions. In these competitions, it is required that the absolute error for reading the analog meter is within 0.05 considering environmental noise. In this study, we developed the automatic reading of the analog meter using a Deep Neural Network (DNN) that satisfies this accuracy. There are primarily two methods for reading the analog meter. One method is to calculate the angle of the pointer
* Yuki Funayama m5231152@u‑aizu.ac.jp
Yutaka Watanobe yutaka@u‑aizu.ac.jp
Keita Nakamura keita‑n@u‑aizu.ac.jp
1
Kenta Tohashi [email protected]
The University of Aizu, Tsuruga, Ikki‑machi, Aizu‑Wakamatsu, Fukushima 965‑8580, Japan
2
Taku Matsumoto d8201105@u‑aizu.ac.jp
Makino Milling Machine Co., Ltd., 4007 Nakatsu, Aikawa‑machi, Aiko‑gun, Kanagawa 243‑0303, Japan
3
Akira Sato aqu
Data Loading...
