Linear Chain Conditional Random Field for Operating Mode Identification and Multimode Process Monitoring.

Area of Science:

• Process engineering
• Machine learning
• Fault detection

Background:

• Conditional random fields (CRFs) are primarily used for fault classification, but struggle with novel faults.
• Existing CRF applications do not address faulty variable location.
• Multimode industrial processes exhibit complex transitions requiring advanced modeling.

Purpose of the Study:

• To adapt linear chain conditional random fields (LC-CRFs) for multimode process modeling with transitions.
• To develop a method for fault detection and faulty variable location using LC-CRFs.
• To validate the proposed approach on benchmark industrial processes.

Main Methods:

• A linear chain conditional random field model was trained using normal operational data with mode labels.
• The model was designed to differentiate between stable operational modes and process transitions.
• An expectation of state feature function was developed for fault identification and localization.

Main Results:

• The LC-CRF model effectively distinguished between stable modes and transitions in multimode processes.
• The developed method successfully detected faults and located faulty variables.
• Validated effectiveness on the Tennessee Eastman process and a continuous stirred tank reactor (CSTR).

Conclusions:

• Linear chain conditional random fields provide a robust framework for modeling and analyzing multimode industrial processes.
• The proposed method enhances fault detection capabilities by identifying unknown faults and pinpointing their origins.
• This approach offers significant potential for improving industrial process safety and efficiency.