Federated Learning Based on Fuzzy Fusion Rules for Chemical Production Process Fault Diagnosis.

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Process data is crucial for fault diagnosis in chemical production but is often sensitive, posing privacy risks.
• Federated learning (FL) addresses data privacy, but standard FL struggles with accuracy on non-independent and identically distributed (non-IID) data.
• Existing FL methods often use simple averaging, which is insufficient for complex, heterogeneous industrial datasets.

Purpose of the Study:

• To develop a personalized federated learning framework for accurate industrial fault diagnosis.
• To address the limitations of conventional federated learning strategies with non-IID process data.
• To improve the precision of fault identification in chemical production while preserving data privacy.

Main Methods:

• A personalized federated learning framework incorporating a Takagi-Sugeno (T-S) fuzzy fusion rule.
• A structured procedure for personalized model construction: fuzzification, fuzzy rule definition, fuzzy inference, and defuzzification.
• Layer-wise fusion strategy to enhance the aggregation precision for improved fault diagnosis.

Main Results:

• The proposed Fuzzy Rule-Based Federated Layer-wise Fusion (FedFZ) framework demonstrated superior fault identification accuracy.
• Effective handling of heterogeneous data distributions inherent in industrial chemical processes.
• Validation of the framework's efficacy on the challenging Tennessee Eastman (TE) process dataset.

Conclusions:

• The FedFZ framework offers a robust solution for industrial fault diagnosis using sensitive process data.
• Personalized federated learning with fuzzy fusion significantly improves accuracy compared to conventional methods.
• The approach is effective in real-world industrial scenarios with non-IID data, enhancing safety and efficiency.