An interpretable hybrid framework combining convolution latent vectors with transformer based attention mechanism for...

Area of Science:

• Engineering
• Artificial Intelligence
• Data Science

Background:

• Industrial asset failures pose significant financial, operational, and safety risks.
• Effective condition monitoring is essential for maintaining industrial operations.
• Large volumes of sensor data necessitate advanced analytical techniques for real-time monitoring.

Purpose of the Study:

• To develop a novel hybrid framework for industrial asset condition monitoring.
• To leverage machine learning, specifically CNNs and ViTs, for accurate fault detection and classification.
• To enable timely predictive maintenance through advanced data analysis.

Main Methods:

• A hybrid framework integrating CNNs for local feature extraction and ViTs for global context understanding.
• Utilizing data augmentation techniques for improved model generalization and computational efficiency.
• Training and validation on the Case Western Reserve University (CWRU) and MFPT fault datasets.

Main Results:

• Achieved an average fault classification accuracy of 99.62% across three fault classes on the CWRU dataset.
• Demonstrated an average time-to-fault detection of 38.4 ms.
• Validated with 99.17% accuracy for outer race and 99.26% for inner race faults on the MFPT dataset.

Conclusions:

• The proposed hybrid CNN-ViT framework offers a robust solution for industrial asset fault detection and classification.
• The method shows high accuracy and efficiency, applicable to various dataset sizes.
• The framework's adaptability allows for integration with alternative convolutional models for enhanced predictive maintenance.