Sensor Fault Detection and Process Monitoring in Water Resource Recovery Facilities

Abstract: Water resource recovery facilities (WRRFs) operate 24/7 to reduce the environmental impact from wastewater on receiving waters. Inaccurate measurements hinder the improvement of operations, limits the performance of automatic control, and deteriorate data quality for decision support and other purposes. This thesis studied how faults can be detected in sensors and impact the treatment process, including aeration diffusers. Simulation studies as well as three 6-18 months long pilot and full-scale experiments were conducted. Evidence was given for the commonplace problem with biofilm formation, and the consequence of biased measurements in two types of dissolved oxygen (DO) sensors. The condition of the energy critical aeration diffusers was monitored by combining process models and a tailored process disturbance, which subsequently improved the information content in existing measurements. The deliberate disturbance approach was also successful in predicting fouling and other faults in DO sensors, and further enabled separation of sensor faults from process variations. The practicability of several machine learning methods was studied for both sensor and process monitoring applications. Probabilistic one-class classification methods showed promising for automatically tuning the alarm threshold, although simple methods produced similarly good results in many situations. Lack of annotated data limited the applicability of the classification methods. For sensor fault detection, this was mitigated by using data from sensor maintenance routines. The need for overall good data quality to identify deviating measurements was underscored when data reconciliation was applied for process monitoring. Reaching a balance between theoretical and practical limitations was further pinpointed as a success factor for data reconciliation. Many previously unknown disturbances in the sensors and the treatment process were revealed during the experiments and resulted in improvement opportunities. A major negative impact from biased sensor signals on treatment efficiency was quantified and analysed in simulations, where the drift direction appeared to be vital. Knowledge gaps related to current sensor data quality were identified and studies were proposed to mitigate the identified shortcomings. Ultimately, the findings in this thesis underline the significance of analysing data using fault detection methods, which can enable a better overall system understanding and decision support.