ARIMA and Neural Network for Sensors Failure Detection

Sensors play a critical role monitoring processes and control of infrastructure. These together with the adaptation, calibration and signal transmission devices conform measurement devices that operate in a control and monitoring loop, are essential for the correct operation of various industrial production systems. A disarrangement of such devices as a whole and eventual failure, are expressed in deviated measurements of expected values that an operator can detect. Therefore, it is necessary to have algorithms that can offer an early warning when the dynamics of a system does not correspond to the measured values. In this sense, this document proposes two approaches. A forecasting technique based on series of past temporary data, which uses an autoregressive model of integrated moving average (ARIMA) and another based on a neuronal network of the multilayer Perceptron type. For both cases, a 95% prediction interval was established to set a criterion to detect anomalies and issue a warning of failure of the measuring device. Both methods were compared to issue an alert, in industrial temperature measurement systems.