High-quality, high-frequency data is essential for successful process modeling and process control. In this paper, a data-driven approach which makes use of deep learning techniques was taken to solve a real-world ammonia concentration dataset. Two rule-based algorithms serve as the benchmark, where both algorithms detect anomalies based on the statistical features. The LSTM approach considers periodicity to distinguish the normal with the abnormal behaviors, with the predicted anomalous data flagged and qualitatively ranked based on the severity and likelihood that the data are faulty (i.e., good, maybe faulty, probably faulty, definitely faulty).The results show that the LSTM based algorithm outperform the rule-based algorithm, where ten out of 11 anomalies can be detected with only one false positive. Both "real" anomalies were successfully detected. Further elimination of the "real" anomalies was then attempted with the flow and temperature datasets. The results show that temperature is not a perfect substitute for flow data. In practice, some water quality datasets may be needed to fully eliminate the impact of precipitation. The algorithms have been successfully applied to well-maintained sensor signals and are now being tested with poorly maintained sensors to judge their suitability in a real-world application.
展开▼
