![]() Each aeration cycle could then be labelled as OK or as a fault type. To do this, an interactive software tool for systematic labelling of each aeration cycle was made. Therefore, it was sought to obtain a labelled data set for drift by manually labelling the data in the PCT with alternating operation, as this was the easiest PCT to assess. Such labelled data sets are rarely available for normal WWTPs, which is also the case for the data available for this study. used fuzzy swarm intelligence and chaos theory to detect faults in a real data set from 1993 available at the UCI Machine Learning Repository however, details on the fault types detected are not described.Ī characteristic for almost all the methods presented in the literature is that they have been developed and tested on data sets where the faults are already known, either because the faults have been simulated or because the data come from well monitored WWTPs. The suggested approach was able to detect the faults and the ANN prediction could be used for process control when a fault was detected. For testing, three types of faults were introduced in real data. The faults considered in the study were sensor faults caused by wrong calibration, process anomaly and drift. The sensors were cleaned on a weekly basis and calibrated if there was a difference detected of more than 15% between the sensor and the reference. Six sensors were installed in the plant including two NH 4 sensors. This study was based on more than one year of data from a real plant. Cecconi and Rosso used ANN to predict the NH 4 concentration and used PCA along with Shewhart monitoring charts for detection of the variation between measured values and predicted values. The data were classified according to faulty NH 4 data by an expert and a Long Short-Term Memory Network was developed and outperformed PCA-SVM. Mamandipoor and Majd possessed 11 months of data from 12 sensors in a real plant. The method was trained on an in-control data set and tested on a simulated data set with introduced faults and data from a real plant, which contained one fault that they were able to detect. proposed a method based on auto correlation and Fused Lasso. The authors stated that this method could obtain more accurate, intuitive and efficient fault detection. In the real case, the authors had 213 samples of which 45 were from normal operation, and these were used for training however, these samples were also included in the test set. The method was both evaluated for simulated data and for data from a real plant. proposed a version of ICA called complex-valued ICA. Mali and Laskar proposed an optimized Monte Carlo deep neural network and were able to detect faults of low magnitude in simulated data. ![]() applied PCA and statistic for fault detection in DO sensors in simulated data and stated that the method was successful in detecting the faults. investigated a number of technics including Support Vector Machine, Ensemble Neural Network and Extreme Learning and found that they performed better than a PCA based method after testing on simulated data. showed that incremental PCA was able to distinguish between time varying events and faults in simulated data, while Kazemi et al. The method was evaluated on simulated dry weather data and the authors state that the method was superior to existing methods and can reduce operating costs and improve the monitoring of the influent. used stacked denoising autoencoders for detection of drift, bias, precision degradation and complete failure. In 20, several methods for fault detection in WWTPs were proposed in the literature. Furthermore, the data should be accompanied by sufficient logging of factors affecting the patterns of the data, such as changes in control settings. These include storing data and select data parameters at resolutions which positively contribute to this purpose. Several recommendations are suggested for utilities that wish to bridge the gap between academia and practice regarding drift detection. The challenges related to data quality raise the question of whether the data-driven approach for drift detection is the best solution, as this requires a high-quality data set. The results show that it is difficult to detect drift in the data to a sufficient level due to missing and imprecise logs, ad hoc changes in control settings, low data quality and the equality in the patterns of some fault types and optimal operation. In this study, the gap between academia and practice is investigated by applying suggested algorithms on real WWTP data. However, these solutions often remain as academic projects. Several studies have investigated anomaly detection and fault detection in WWTPs. Sensor drift in Wastewater Treatment Plants (WWTPs) reduces the efficiency of the plants and needs to be handled.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |