There has been some debate about the efficiency of ballast water treatment systems that use filters. Clogging is a factor, but is it predictable?
In a paper entitled A Machine Learning Approach to Predictively Determine Filter Clogging in a Ballast Water Treatment System, Kristoffer Sliwinski, a Master’s student at Sweden’s KTV Industrial Engineering and Management school looks at how statistics could be used to predict clogging in a ballast water treatment system filter.
In the first stage, he looked at the different grades of clogging using data from a ballast water treatment system. In the second stage he used regression analysis on the data through two neural networks for parameter prediction, one long/short term memory (LSTM) neural network and one convolutional neural network.
The first stage of identifying the grades of clogging is based on established scientific work on the rate and intensity of clogging impact on the differential pressure in the filter. Both simple and back-flushing filters were analysed.
The ﬁltration data was obtained from ﬁlter tests by Alfa Laval at Lake Malmasjon over two weeks. Each data point was sampled every five seconds and contains sensor data for the diﬀerential pressure over the ﬁlter, the ﬂuid ﬂow in the entire system, the ﬂuid pressure in the entire system and the ﬂuid ﬂow in the backﬂush mechanism. The data was uploaded to Alfa Laval’s connectivity cloud service.
The two neural networks were trained and then used to analyse the data. The conclusion of the paper is that neural networks have a place in predicting clogging, with the LSTM achieving a 93.3% accuracy on the state of clogging up to 30 seconds ahead.
With more research it would seem that neural networks could have a commercial application in predicting filter clogging in ballast water treatment systems.