Lag-Specific Transfer Entropy for Root Cause Diagnosis and Delay Estimation in Industrial Sensor Networks

Area of Science:

• Process Engineering and Control
• Data Science and Machine Learning
• Industrial Monitoring and Diagnostics

Background:

• Modern industrial plants generate vast amounts of multi-sensor data at high frequencies.
• Variable transport and residence time delays in sensor data complicate accurate disturbance analysis.
• Existing methods struggle to precisely identify the origin and propagation time of process disturbances.

Purpose of the Study:

• To develop and validate a novel method for identifying the origin of process disturbances using historical sensor data.
• To accurately quantify the time delay for disturbances to propagate through industrial systems.
• To improve the reliability of disturbance detection by reducing false links compared to traditional methods.

Main Methods:

• Application of lag-specific transfer entropy (LSTE) to analyze multi-sensor time-series data.
• Inclusion of a self-prediction optimization step to remove self-information from sensor data.
• Benchmarking LSTE on diverse industrial case studies: a nonlinear simulation, the Tennessee Eastman plant, a three-phase separator, and a blast furnace.

Main Results:

• LSTE successfully identified the initial sensor deviating from normal operation in all tested scenarios.
• The method accurately reported the propagation times of disturbances, aligning with known process physics.
• LSTE demonstrated a significant reduction in false causal links compared to classical transfer entropy.

Conclusions:

• Lag-specific transfer entropy is an effective tool for pinpointing disturbance origins and quantifying propagation delays in complex industrial processes.
• The self-prediction optimization enhances the accuracy and reliability of transfer entropy for industrial sensor data analysis.
• LSTE offers a robust advancement for real-time monitoring, fault detection, and process optimization in industrial settings.