Condition-Adaptive CNN with Spatiotemporal Fusion for Enhanced Motor Fault Diagnosis

Area of Science:

• Industrial Automation
• Machine Learning
• Vibration Analysis

Background:

• Electric motors are crucial in industrial systems but prone to faults under demanding conditions.
• Accurate fault diagnosis is challenging due to varying signal characteristics and noise.
• Existing methods struggle with complex operating states and feature extraction.

Purpose of the Study:

• To develop a robust fault diagnosis framework for electric motors.
• To enhance feature representation and adaptive parameter optimization for CNN models.
• To improve diagnostic accuracy and generalization capabilities under diverse operating conditions.

Main Methods:

• A novel convolutional neural network (CNN) framework integrating the bee colony algorithm (BCA) for adaptive parameter optimization.
• A spatiotemporal fusion architecture with large-kernel convolution, bottleneck layers, and an improved self-attention (ISA) mechanism for enhanced feature extraction.
• An engineering-oriented data augmentation strategy using multi-scale window offset and noise superposition on 1D vibration signals.

Main Results:

• The proposed CNN-BCA-ISA framework achieved a diagnostic accuracy of 96.4% on a mixed dataset.
• The model demonstrated stable performance across varying noise levels, indicating strong generalization.
• A real-time fault diagnosis system was successfully implemented and validated in industrial environments.

Conclusions:

• The developed CNN-BCA-ISA framework offers an effective solution for electric motor fault diagnosis.
• The adaptive optimization and enhanced feature extraction significantly improve diagnostic accuracy and robustness.
• The framework is feasible for practical state monitoring applications in industrial settings.