COMPARISON OF DIFFERENT METHODS FOR SENSOR FAULT DIAGNOSIS IN A MUNICIPAL WASTEWATER TREATMENT PLANT

Wastewater plants play an important role in removing the contaminants from wastewater and converting it into an effluent that can be reused for various purposes such as irrigation, among others. There are tightening treatment regulations on the quality of effluent regarding the values of the primary variables such as concentrations of ammonia, nitrates and total nitrogen, phosphates and total phosphorus, suspended solids, biochemical and chemical oxygen demand, as well as other process variables like the sludge blanket level. The conventional monitoring of these parameters contains on-line (via the field-installed probes) and off-line (implementing the laboratory experiments) analysis. The real-time monitoring is hard-to-measure, costly, time-consuming, and the field instruments need frequent maintenance that makes the field measurements reliability challenging. In other words, these sensors produce several anomalies during the recording process. Accordingly, to have accurate values of the effluent quality, it is of great importance to detect the outliers among the probe recorded data points. Before final utilization of data derived from the sensors, it is recommended to remove the observations that are not in line with the general data trend. During the last two decades, different methods for data analyses and outlier detection have been presented by many authors. Among the proposed approaches, artificial neural networks (ANNs), principal component analysis (PCA), fuzzy logic, clustering, fisher discriminant analysis (FDA), independent component analysis, and partial least squares regression (PLS) have attracted much attention. We examined the performance of two different methods including PCA analyzes with T-square (T2) and the Modified Z-score in detecting the outliers and cleaning the data obtained from sensors at the effluent of the Peschiera Borromeo wastewater treatment plant in Milan. The robustness of the proposed methods is evaluated using the benchmark data derived from the laboratory measurements. The MATLAB scripts written based on the T-square and Z-score approaches are implemented to analyze the results and compare the performance of the methods. The results suggested that both methods showed satisfactory results in the recognition of anomalies. However, the PCA outperforms the modified Z-score on the detection of clustered outliers. The number of outliers detected by the PCA outnumbers those of the moving filter method since the T-squared method is based on the applied radius for the dataset, the effect of the alpha (radius) and the selected interval is highly important. Moving median filter shows acceptable results close to the PCA with less number of outliers detected. The moving median IV limit plays a significant role in the number of outliers. Different window sizes and moving median limits are tested to find the best fit where minimum bias is obtained with the maximum number of outliers. For the future investigation, it is recommended to apply an optimization algorithm to adjust the alpha value in the PCA method that may lead to even better results in the outlier detection.