Reliable sensor values are important for resource-efficient control and operations of wastewater treatment processes. Automatic fault detection methods are necessary to monitor the increasing amount of data produced in any modern water resource recovery facility (WRRF). Most on-line measurements exhibit large variations under normal conditions, due to considerable variations in the influent flow. The work reported in this licentiate thesis deals with fault detection in WRRFs. In the first paper, we studied how Gaussian process regression (GPR), a probabilistic machine learning method, could be applied for fault detection in WRRFs. The results showed that the standard parameter estimation method for GPR suffered from local optima which could be solved by instead estimating the distribution of the parameters with a sequential Monte Carlo algorithm (GPR-SMC). The GPR-SMC allowed for automatic estimation of missing data in a simulated influent flow signal with high noise, which is a representative signal for on-line sensors in WRRFs. In addition, the GPR-SMC provided uncertainty predictions for the estimated data and accurate sensor noise estimates. Care should be taken in selecting a suitable kernel for GPR, since the results were in contrast to the general assumption that prior knowledge can easily be encoded by means of selecting a proper kernel. Here, the autocorrelation graph was found useful as diagnostic tool for selecting a proper kernel. In the second paper, we studied how active fault detection (AFD) could be used to reveal information about the sensor status. The AFD was implemented by evaluating the change in a dissolved oxygen (DO)-signal caused by the sensor's automatic cleaning system. Fault signatures were obtained for fouling and several other sensor faults such as a worn out or mechanically damaged membrane. This demonstrates the potential of AFD, not only for fault detection, but also for fault diagnosis. Interestingly, the progression of the sensor bias due to organic biofilm fouling differed depending on the measurement technique used within the DO-sensor. This is new knowledge that is valuable for process control and should be further studied. The AFD was implemented on a full scale system to demonstrate its applicability, which is rarely done in research papers in the field of WRRFs.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:uu-329777 |
| Date | January 2017 |
| Creators | Samuelsson, Oscar |
| Publisher | Uppsala universitet, Avdelningen för systemteknik, Uppsala universitet, Reglerteknik |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Licentiate thesis, comprehensive summary, info:eu-repo/semantics/masterThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | IT licentiate theses / Uppsala University, Department of Information Technology, 1404-5117 ; 2017-003 |
Page generated in 0.0019 seconds