Sensor Fault Detection with Bayesian Networks

EasyChair Preprint 4592

Abstract

Several sensors are installed in the majority of chemical reactors and storage tanks to monitor temperature profiles for safety and decision-making processes such as heat demand or flow rate calculations. These sensors fail occasionally and generate erroneous measurement data that need to be detected and excluded from the calculations. However, due to the high number of process variables displayed in the chemical plants, this task is not trivial. In this work, a Bayesian network approach to detect faulty temperature sensors is proposed. By comparing the sensor measurements with each other, the faulty sensor is detected. A modular approach is preferred, and networks are created for 10 K temperature intervals to increase flexibility and sensitivity. Created networks can be adjusted for the operating temperature ranges; hence, they can be used for any catalyst and entire life cycle. The developed method is demonstrated on an industrial scale hydrocracker unit with 92 sensor couples installed in a series of reactors. From the investigated sensors, 16 of them showed a greater difference than the 2 K threshold chosen for the fault. In addition to that, 13 sensors showed an increasing temperature difference that may lead to a fault. Two scenarios were created to calculate the energy loss due to a faulty measurement, and a 5.5 K offset error was found to cause a 5.79 TJ energy loss every year for a small scale hydrocracker.

Keyphrases: Bayesian network, diagnostics, fault detection, sensor fault

@booklet{EasyChair:4592,
  author    = {Esin Iplik and Ioanna Aslanidou and Konstantinos Kyprianidis},
  title     = {Sensor Fault Detection with Bayesian Networks},
  howpublished = {EasyChair Preprint 4592},
  year      = {EasyChair, 2020}}