Data-Driven Fault Detection in a Thermocouple Network Using Neighboring Redundancy, XGBoost Classifier, and Up–Down Counter

Diego a. Velandia cárdenas, Erwin jose lópez Pulgarín, Jorge iván Sofrony

Research output: Contribution to journalArticlepeer-review

32 Downloads (Pure)

Abstract

Fault detection and isolation (FDI) is of great interest for the control community since it can drive improved performance in a system by allowing predictive maintenance/repairing and catering for improved operational safety. FDI in large-scale smelting furnaces presents several challenges, as it requires the understanding of complex thermal and chemical reactions occurring inside the structure. Furthermore, the impossibility of having full operational information about the process makes the use of model-based methods very complex or unfeasible. This article introduces a methodology to develop a data-driven FDI system for the detection of incipient and intermittent failures in a network made out of 322 thermocouples located on the shell of the furnace. Statistical metrics over fault counter time windows (FTCWs) were used to identify different types of sensor failures, which led to establishing a baseline of known failure events and to create a dataset to train the machine learning (ML) classification models. A data-driven approach was proposed based on the sensors’ physical (neighboring) redundancy, which led to some type of physical redundancy. A postprocessing stage was used to stabilize the model’s response in time, determining that the proposed FDI system successfully detects faults while reducing reported false negatives.
Original languageEnglish
Pages (from-to)5215-5223
JournalIEEE Sensors Journal
Volume24
Issue number4
Early online date25 Dec 2023
DOIs
Publication statusPublished - 15 Feb 2024

Keywords

  • FDI
  • machine learning
  • parameter variation
  • redundancy
  • sensors network
  • thermocouple
  • up-down counter

Fingerprint

Dive into the research topics of 'Data-Driven Fault Detection in a Thermocouple Network Using Neighboring Redundancy, XGBoost Classifier, and Up–Down Counter'. Together they form a unique fingerprint.

Cite this