Grafted Composite Decision Tree: Adaptive Online Fault Diagnosis with Automated Robot Measurements

TL;DR

This paper introduces a new adaptive fault diagnosis framework using grafted composite decision trees to improve online monitoring in industrial systems.

Contribution

The novel grafted composite decision tree framework combines monitored and robot-measured variables for adaptive fault diagnosis.

Findings

The proposed framework improves measurement selection and inference confidence in fault diagnosis.

Performance comparisons show the new method outperforms conventional decision trees in utilizing information for inference.

Abstract

In many industrial facilities, online monitoring systems have improved the reliability of key equipment, reducing the cost of operation and maintenance over recent decades. However, it often requires additional on-site inspection of target facilities due to limited information from installed sensors. To systematically automate such processes, an adaptive online fault diagnosis framework is required, which consecutively selects variables to measure and updates its inference with additional information at each measurement step. In this paper, adaptive online fault detection models—grafted composite decision trees—are proposed for such a framework. While conventional decision trees themselves can serve two required objectives of the framework, information from monitored variables can be less utilized because decision trees do not consider if required input variables are always monitored when the models are trained. On the other hand, the proposed grafted composite decision tree models are designed to fully utilize both monitored and robot-measured variables at any stage in a given measurement sequence by grafting two types of trees together: a prior-tree trained only with observed variables and sub-trees trained with robot-measurable variables. The proposed method was validated on a cooling water system in a nuclear power plant with multiple leak scenarios, in which improved measurement selection and increase in inference confidence in each measurement step are demonstrated. The performance comparison between the proposed models and the conventional decision tree model clearly illustrates how the acquired information is fully utilized for the best inference while providing the best choice of the next variable to measure, maximizing information gain at the same time.

Figures