Neural networks used to monitor an experimental test workbench

Abstract

This work presents the application of neural networks in an experimental workbench. This bench was developed with the purpose of conducting real time tests and data acquisition. The method applied for this work allowed to generate faulty data in a gradual and controlled way through the binary combination of double action valves. Using the SCADA application (Supervisory Control and Data Acquisition), it became possible to acquire data for analysis in Matlab / Simulink software. This bench has two reservoirs: a reservoir that has sensors for recording pressure and temperature variables for later analysis, and another reservoir that has level sensors. Four models were used to develop the respective practical experiments. In the first model, it was possible to perform all practical tests of the plant, as well as mechanical changes like repositioning of some mechanical components, piping, sensors and electrovalves. In the second model, it was noticed that the positioning of the flow meter, located after the pump output, prevented a good measurement of the flow variable. In the third model, it was perceived that the number of failures initially adopted, made the data too confusing for the neural network analysis. In the last model, it was possible to obtain a performance of 96.6% of hits after the reconfiguration for 4 controlled faults.