Prediction of alarms in a pump station using neural networks

Abstract: Prediction of pump station alarms based on data is an interesting future service for companies as it allows them to both offer a new service and reduce the downtime of the pumps. In order to predict the alarms a study of the available data was done in order to examine feasibility and identify problems. The chosen solution was to use two neural networks which were connected in a chain to create a complete solution. The first neural network used long short term memory (LSTM) neurons in order to recursively predict time-series data from sensors, such as sump water level and pump electric current, this was then used by the second LSTM neural network in order to determine if these parameter values would trigger an alarm. The second LSTM network was unable to determine if an alarm would happen and thus the whole solution did not work. There were two main reasons for this, the first being that the alarms stop time did not correlate to the time-series sensor data which created an uncertainty of which parameter levels actually belonged to an alarm, making the LSTM network unable to identify what an alarm is. The second reason was that the data was downsampled too much making it even harder for the LSTM network to identify what an alarm was. This thesis has helped Xylem further understand the use and needs of machine learning, which will help Xylem progress further into the area of predictive and smart services.