Using machine learning to interpret turbidity readings during drinking water processes

Abstract

To ensure good health in a population, it is necessary that the drinking water provided is of a certain quality. Different laws and regulations have been implemented across the world to make drinking water safe and accessible. This thesis have looked at a drinking water treatment plant in a large Norwegian city, and one of their subprocessess. Several parameters are measured as part of this subprocess, and one such parameter is turbidity. To ensure that the turbidity is at sufficiently low levels, the water has to be filtered. These filters have to be flushed to avoid clogging, and it is between these flushings that the problem occur. The plant expects these readings to show a consistent behaviour, but instead, some readings show an unexpected behaviour. The goal of this thesis was to gain an understanding as to why this happened. This was done by first creating methods of quantifying the instability before using these to divide the segments into stable and unstable segments respectively. This part of the work was considered a success as the segments deemed unstable by the methods were the same as those deemed unstable by the client. The second part of the work was to develop models that could accurately classify these segments. The best performing model was a multilayer perceptron model and had an F1 score of 0.743\%. The third an final part was to interpret the results and provide an explanation as to what caused the differences in stability. For this, a series of shapelet models were used. The shapelets produced by these models were not able to provide further insight to the client. Further analysis is needed to determine the value of these shapelets. The multilayer perceptron models and shapelet models had similar rankings in what variable provided the best result. Having discussed with the client, this aligned with some of their findings as well. This would suggest that amount of drinking water produced during the process, amount of iron chloride dosed and the filter maturation could help in explaining why these segments are irregular. Lastly, some suggestions for how to further improve the methods and models in this was provided.