A Time-Frequency Domain Diagnosis Network for ICE Fault Detection

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Internal combustion engines (ICEs) face challenges in real-time fault diagnosis due to limitations in traditional methods for feature extraction and data requirements.
• Existing deep learning models like CNN and LSTM struggle to efficiently capture both time and frequency domain features for comprehensive ICE fault analysis.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and efficient fault diagnosis in internal combustion engines.
• To address the limitations of existing methods in feature extraction and data dependency for real-time ICE condition monitoring.

Main Methods:

• Proposed a novel Time-Frequency Domain Diagnosis Network (TFDN) integrating parallel time-domain (ResNet, self-attention) and frequency-domain (CNN) feature extraction paths.
• Utilized Swish activation functions and batch normalization for improved training efficiency.
• Validated the TFDN model on a six-cylinder diesel engine dataset encompassing 12 distinct fault types.

Main Results:

• TFDN achieved high diagnostic accuracy (98.12%~99.79%) under full-load conditions, surpassing baseline models (CNN, ResNet, LSTM).
• The model demonstrated robust performance under mixed operating conditions, maintaining high accuracy, precision, and recall.
• TFDN showed significant robustness with limited data, achieving 60%~70% accuracy with only 5 samples per fault.

Conclusions:

• The TFDN effectively integrates time-frequency features, significantly improving diagnostic accuracy and stability for ICEs.
• The proposed network offers a practical and data-efficient solution for real-time fault detection and condition monitoring in internal combustion engines.