MetaRes-DMT-AS: A Meta-Learning Approach for Few-Shot Fault Diagnosis in Elevator Systems.

Area of Science:

• Deep learning applications in industrial systems.
• Machine learning for predictive maintenance.
• Fault diagnosis in complex machinery.

Background:

• Deep learning for elevator fault diagnosis faces challenges due to the need for extensive labeled data.
• Real-world industrial settings often lack sufficient data for training robust diagnostic models.
• Data scarcity hinders the development of reliable fault diagnosis systems.

Purpose of the Study:

• To propose MetaRes-DMT-AS, a novel meta-learning framework for few-shot fault diagnosis in elevators.
• To address the limitations of data-scarce conditions in industrial fault diagnosis.
• To enhance the reliability and accuracy of elevator diagnostic models.

Main Methods:

• Utilized Gramian Angular Fields to convert 1D sensor data into 2D image representations.
• Employed episodic task construction via stochastic sampling for meta-training.
• Implemented an adaptive scheduling module for dynamic support/query set configuration and prototype network regularization.

Main Results:

• MetaRes-DMT-AS achieved state-of-the-art few-shot classification performance on bearing and elevator acceleration datasets.
• The framework surpassed benchmark models by 0.94-1.78% in overall accuracy.
• Significant accuracy improvements were observed for critical faults: 3-16% for emergency stops and 17-29% for severe vibrations.

Conclusions:

• MetaRes-DMT-AS effectively addresses data scarcity in elevator fault diagnosis using a meta-learning approach.
• The proposed framework demonstrates superior performance, particularly for rare but critical fault types.
• This methodology offers a robust solution for developing reliable diagnostic systems in data-limited industrial environments.