[4Yin2] Interactive session 2

[4Yin2-27] Evaluation of a Manhole Pump Anomaly Predictor Generated by Neural Architecture Search

〇Michihiro Shonai¹, Katsushi Aso², Masahito Yamamoto³, Takahiko Hattori⁴, Koudai Maekawa⁵ (1.Ecomott Inc., 2.NiX Co., Ltd., 3.Hokkaido Univ., 4.Nihon Suido Consultants Co., Ltd., 5.City of Toyama)

Keywords:Anomaly Sign Detection, Neural Architecture Search, Manhole pump

Many small and medium-sized municipalities that have installed manhole pumps need a way to reduce the waiting time for workers responding to emergency situations at night and to understand the optimal interval for pump inspections. Therefore, we proposed an anomaly predictor that outputs anomaly (emergency) probabilities. The predictor was created by training several neural networks generated by Neural Architecture Search (NAS) using the operating current values of a specific manhole pump with at least three anomaly cases. Subsequently, the best neural network architecture was chosen. As a result, we were able to detect anomalies based on the anomaly probability. Furthermore, in the cases where the predicted anomaly probability was high, the accumulation of dirt inside the manhole pump (a sign of future emergency) was confirmed. These results suggest that NAS could be used to predict and detect unknown signs (characteristics that lead to anomalies) even in the case of a small number of anomalies.