Coarse-to-Fine: Progressive Knowledge Transfer-Based Multitask Convolutional Neural Network for Intelligent...

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Large-scale fault diagnosis in complex industrial systems is critical.
• Existing deep learning methods struggle with numerous fault types, causing intra/inter-class distance imbalance and poor local minima.
• Addressing these limitations is essential for reliable industrial operations.

Purpose of the Study:

• To propose a novel Progressive Knowledge Transfer-based Multitask Convolutional Neural Network (PKT-MCNN) for large-scale fault diagnosis.
• To mitigate intra/inter-class distance imbalance and avoid poor local minima in neural networks.
• To enhance the accuracy and robustness of fault diagnosis in complex systems.

Main Methods:

• A structure learning algorithm clusters fault types into coarse-grained nodes to balance intra/inter-class distances.
• A Multitask Convolutional Neural Network (MCNN) architecture simultaneously learns coarse and fine-grained tasks, extracting general fault information.
• A Progressive Knowledge Transfer (PKT) algorithm transfers knowledge and progressively regulates learning stages via attention weights.

Main Results:

• The proposed PKT-MCNN effectively handles large-scale fault diagnosis with a dataset of 66 fault types from a nuclear power system.
• The method successfully mitigates intra/inter-class distance imbalance in the feature space.
• The approach demonstrates improved performance by avoiding poor local minima compared to traditional methods.

Conclusions:

• The PKT-MCNN is a promising tool for accurate and robust large-scale fault diagnosis in complex industrial systems.
• The progressive knowledge transfer and multitask learning approach effectively addresses the challenges of numerous fault types.
• This method offers a significant advancement in industrial fault diagnosis capabilities.