31 Mar
14:00 Master's Defense Auditorium 1 - IC 3
Using Machine Learning to Predict Overflow in Sewer Pumps
Heber Augusto Scachetti
Advisor / Teacher
Guido Costa Souza de Araújo
Brief summary
Sewage elevators are elements responsible for transporting sewage to treatment plants through one or more pumps that are driven by a controller. The sewage pump controller uses the sewage level information in the suction well to decide between turning the pumps on and off. When the sewage is not pumped correctly, the sewage level rises until it spills out. The leakage event is nothing more than a sewage leak, which can mean the occurrence of ecological disasters and serious financial losses to the company responsible for the maintenance and operation of the sewage pumping stations. This master's work presents a study of thirty sewage elevators monitored by the supervision system of the DVME (Sewage Operation Macro Division) at COPASA in Minas Gerais. As far as we know, this dissertation proposes the first solution, based on Machine Learning Models, for predicting leakage in sewage pumping stations. The machine learning models applied in this work are based on recurrent neural networks of the type LSTM (Long Short Term Memory or, in Portuguese, Long and Short Term Memory), neural networks of the type CNN (Convolutional Neural Network or, in Portuguese, convolutional neural network) and ConvLSTM neural networks (LSTM with convolution-based transformations). For the scenario evaluated in this work, the ConvLSTM model was able to predict with 95% accuracy the occurrence of sewage overflow at the station within the next 1 hour.
Examination Board
Guido Costa Souza de Araújo IC / UNICAMP
Niederauer Mastelar FEM / UNICAMP
Rodrigo Dias Arruda Senra Work & Co
Marcio Machado Pereira IC / UNICAMP
Fabio Augusto Faria ICT / UNIFESP