A Sensor Fusion Method Using Transfer Learning Models for Equipment Condition Monitoring.

Area of Science:

• Engineering
• Computer Science
• Signal Processing

Background:

• Sensor fusion is crucial for effective condition monitoring and improved classification accuracy.
• Current feature-level fusion methods often require custom deep learning architectures, limiting the use of pre-trained models.
• There is a need for sensor fusion techniques compatible with widely available deep learning architectures.

Purpose of the Study:

• To propose a new sensor fusion method applicable to heterogeneous sensors in the time-frequency domain.
• To enable the use of pre-trained deep learning models for sensor fusion tasks.
• To evaluate the effectiveness of the proposed method using transfer learning.

Main Methods:

• A novel sensor fusion technique inspired by image fusion is developed, focusing on fusing spectrogram images.
• The method fuses multiple and heterogeneous sensors in the time-frequency domain.
• Transfer learning (TL) techniques are applied to four pre-trained convolutional neural network (CNN) architectures.

Main Results:

• The proposed sensor fusion technique effectively classifies device faults.
• The use of pre-trained TL models enhances the model training capabilities.
• The method demonstrates successful fusion of heterogeneous sensor data in the time-frequency domain.

Conclusions:

• The developed sensor fusion method is effective for condition monitoring and fault classification.
• Integrating image fusion concepts with time-frequency domain data is a viable approach.
• Transfer learning with pre-trained CNNs offers a powerful way to leverage existing deep learning models for sensor fusion.