[3E2-GS-2g] Machine learning

11:15 AM - 11:30 AM JST(2:15 AM - 2:30 AM UTC)

[3E2-GS-2g-02] Anomaly Detection for Electromagnetic Flowmeters Using a BiLSTM-Transformer Hybrid Model

〇Seita Okamoto^1,2, Ryoichi Koga², Satoshi Oyama² (1. Aichi Tokei Denki Co.,Ltd., 2. Graduate School of Data Science, Nagoya City University)

Keywords:

Electromagnetic flowmeter,Hybrid-model,Time-series analysis

This study aims to accurately detect sudden noise events, such as air entrainment, in electromagnetic flowmeters by constructing a hybrid model that combines a Bidirectional Long Short-Term Memory (BiLSTM) network and a Transformer in a sequential architecture. The effect of layer ordering on anomaly detection performance was systematically evaluated. Experimental data were collected from a pipeline equipped with an electromagnetic flowmeter under three flow conditions: 20, 40, and 60 m³/h. The results showed that under the low-flow condition (20 m³/h), the configuration in which the BiLSTM processes the data first, followed by the Transformer, achieved the highest performance, reaching an accuracy of 99.7% and demonstrating superior generalization capability. In contrast, under medium- and high-flow conditions (40 and 60 m³/h), performance differences between model configurations were relatively small. These findings suggest that the optimal model architecture depends on the flow rate, highlighting the importance of flow-dependent model design for robust anomaly detection in electromagnetic flow measurement systems.