A neural network approach for on-line fault detection of nitrogen sensors in alternated active sludge treatment plants

In this paper, a technique for fault detection of nitrogen sensors in alternating active sludge treatment plants is presented, based on two predictive neural networks used as residuals generators. The neural networks are trained in a simulation environment using a historical set of data collected during a fault-free operation of the plant. They are, then, used to predict the concentrations of reduced (ammonium) and oxidized (nitrates and nitrites) nitrogen with an advance of 1 minute. The neural networks appear characterized by good generalization ability and robustness with respect to the influent variability with time and weather conditions. A second simulation set is carried out in the presence of different kinds of fault on both sensors, as isolated spikes, abrupt bias and increased measurement noise. Analysis of residuals, computed as difference between measured concentration values and neural networks predictions, allows a quick revealing of the fault as well as the isolation of the corrupted sensor. Indeed, the occurrence of the three classes of fault can be easily characterized by the peculiar shape of the corresponding residuals: a double peak for the isolated spike, a single peak for the abrupt bias and an increased standard deviation of residuals for the increased measurement noise. In a real contest other unmodelled disturbances can affect residuals rather than sensor faults; the present approach is unable to individuate these cases, but could give a prompt warning calling for further investigation.