Intelligent Fault Diagnosis of Rotary Machinery by Convolutional Neural Network with Automatic Hyper-Parameters...

Area of Science:

• Engineering
• Computer Science

Background:

• Intelligent fault diagnosis relies on machine learning but often suffers from suboptimal hyper-parameter tuning.
• Manual hyper-parameter configuration by experts is time-consuming and limits model performance.

Purpose of the Study:

• To develop a data-driven intelligent fault diagnosis technique for rotary machinery using a convolutional neural network (CNN) with optimized hyper-parameters and structure.
• To address the gap in hyper-parameter optimization for improved machine learning model performance in fault diagnosis.

Main Methods:

• A deep learning approach using a 19-layer CNN to process raw three-axis accelerometer signals (12,800 × 1 × 3 matrix).
• Bayesian optimization was employed during the model learning phase to optimize CNN hyper-parameters.
• The model was trained and evaluated on vibration data from rotary machinery, including tests for overfitting with altered bearings.

Main Results:

• Achieved an overall classification accuracy of 99.94% on the evaluation set for classifying 8 machine states and 2 rotational speeds.
• Attained 100% classification accuracy on a second evaluation set, demonstrating robustness and potential for real-world application.
• Successfully classified different rotary machinery states and speeds using optimized CNN with raw vibration signal input.

Conclusions:

• The proposed optimized CNN technique effectively diagnoses faults in rotary machinery using raw vibration data.
• The data-driven approach with hyper-parameter optimization significantly enhances diagnostic accuracy and robustness.
• This method shows strong potential for practical implementation in intelligent fault diagnosis systems.