University of Strathclyde1, Scottish Water2
Proactive maintenance of assets is a much sought after goal in the water and wastewater industry, where substantial savings could be made by identifying impending failures in pumps and other essential components of the system. A detailed analysis of the operational behaviour of the monitored assets can be used as the foundation to generate estimations on the likelihood of a failure or malfunction in a particular component based on knowledge of previous behavioural patterns. Preventative maintenance or component replacement can then be optimally scheduled based on need, as opposed to traditional reactive maintenance strategies. In most current condition monitoring software, an alarm is normally raised once a fault has occurred, therefore often requiring immediate action. On the other hand, combining the condition monitoring and fault log data that is normally acquired with expert knowledge of the meaning and causes of faults embedded in the software allows predictive maintenance to be implemented. The paper reports on a number of advanced machine learning techniques that have been applied to operational data acquired over a significant period of water pump operation. Results from a representative site within Scottish Water's water network will be presented that demonstrate the application of such software techniques can indeed surface changes in parameters, for example flow and pump power drawn, forming the basis to infer the state of components and the onset of changes in the health of the asset.