Multi-Variable Transformer-Based Meta-Learning for Few-Shot Fault Diagnosis of Large-Scale Systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial Process Control

Background:

• Large-scale systems face significant fault diagnosis challenges due to data complexity, high dimensionality, and scarce labeled fault data.
• Existing methods struggle with limited fault examples, hindering accurate and timely diagnostics in real-world operational scenarios.

Purpose of the Study:

• To propose a novel fault diagnosis approach, the Multi-Variable Meta-Transformer (MVMT), designed to overcome data scarcity and complexity issues.
• To enhance the Transformer model for effective processing of multi-variable time series data, integrating both continuous and state variables.
• To leverage meta-learning, specifically Model-Agnostic Meta-Learning (MAML), for few-shot fault diagnosis capabilities.

Main Methods:

• Modification of the Transformer model to handle multi-variable time series data.
• Introduction of feature layers to integrate continuous and state variables within the Transformer encoder.
• Application of the Model-Agnostic Meta-Learning (MAML) strategy using the modified Transformer as the base model.

Main Results:

• The Multi-Variable Meta-Transformer (MVMT) demonstrates exceptional performance in few-shot fault diagnosis scenarios.
• Effective diagnosis achieved using continuous-only data and combined continuous and state variables.
• Validation on the Tennessee Eastman Process and Power-Supply System databases confirms the method's robustness.

Conclusions:

• The MVMT approach successfully addresses the challenges of fault diagnosis in large-scale systems with limited labeled data.
• Meta-learning combined with a modified Transformer architecture provides a powerful framework for few-shot fault diagnosis.
• The proposed method offers a promising solution for improving the reliability and safety of industrial processes.