Deep residual neural-network-based robot joint fault diagnosis method.

Area of Science:

• Robotics and Automation
• Artificial Intelligence
• Machine Learning

Background:

• Robot joint failures, including sensor and actuator issues like gain errors and malfunctions, pose significant challenges to system reliability.
• Existing fault diagnosis methods often lack the accuracy or efficiency required for complex robotic systems.

Purpose of the Study:

• To propose and evaluate a data-driven fault diagnosis method for robot joints utilizing a Deep Residual Neural Network (DRNN).
• To assess the DRNN's performance against other machine learning models in terms of accuracy and training time.
• To verify the noise immunity and effectiveness of the DRNN for diagnosing various robot joint faults.

Main Methods:

• A Deep Residual Neural Network (DRNN) model was developed by stacking convolutional layers.
• Gaussian white noise was injected into the dataset to test the DRNN's noise immunity.
• The DRNN was compared with Support Vector Machine (SVM), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long-Term Memory Network (LTMN) through simulations.

Main Results:

• The DRNN achieved higher accuracy in diagnosing robot joint faults (gain error, offset error, malfunction) compared to SVM, ANN, CNN, and LTMN.
• The DRNN required significantly less model training time.
• Visualization analysis confirmed the feasibility and effectiveness of the DRNN method.

Conclusions:

• The proposed DRNN-based method offers a superior approach for robot joint fault diagnosis, excelling in accuracy and efficiency.
• The DRNN demonstrates robustness against noise, making it suitable for real-world applications.
• This method provides a reliable solution for diagnosing sensor and actuator faults in robotic systems.