Cross-Attention Fusion of Time-Frequency Acoustic Features for Insulation Fault Recognition in Power Equipment Using...

Area of Science:

• Electrical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Insulation faults in power equipment generate complex acoustic signals.
• Conventional single-domain analysis methods struggle with nonlinear, non-Gaussian acoustic characteristics.
• Accurate fault diagnosis is crucial for reliable power system operation.

Purpose of the Study:

• To develop a non-contact method for recognizing insulation faults in power equipment.
• To enhance automated fault classification by fusing time- and frequency-domain acoustic features.
• To address limitations of conventional single-domain feature extraction.

Main Methods:

• A time-frequency feature extraction and fusion framework using a cross-attention (CA) mechanism.
• Extraction of time-domain features via a temporal convolutional network and autoencoder.
• Extraction of frequency-domain features from Mel spectrograms using a convolutional block attention module.
• Adaptive fusion of temporal and spectral features using CA.
• Fault classification using a 1D CNN optimized with the Cuckoo Search algorithm.

Main Results:

• The proposed framework effectively characterizes complex acoustic fingerprints of insulation faults.
• High classification performance was achieved for four insulation fault types.
• The method demonstrated superior performance compared to conventional feature extraction approaches.

Conclusions:

• The developed framework offers a robust non-contact strategy for insulation fault diagnosis.
• This approach improves the condition monitoring of electrical power equipment.
• Fusing time- and frequency-domain acoustic features enhances fault recognition capabilities.