Blade Rub-Impact Fault Identification Using Autoencoder-Based Nonlinear Function Approximation and a Deep Neural...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Blade rub-impact faults are common and complex issues in turbines.
• Analyzing nonlinear and nonstationary vibration signals requires computationally intensive methods.
• Existing linear-based fault diagnosis techniques may not fully capture the complexity of these faults.

Purpose of the Study:

• To propose a novel and computationally efficient method for diagnosing blade rub-impact faults in rotor systems.
• To accurately assess different severity levels of blade rub-impact faults.
• To compare the proposed nonlinear-based method with traditional linear-based techniques.

Main Methods:

• Utilizing a deep undercomplete denoising autoencoder to estimate the system's nonlinear function under normal conditions.
• Computing residual signals by subtracting the estimated normal signals from the original vibration data.
• Employing a deep neural network to classify the rotor system's state based on the computed residual signals.

Main Results:

• The amplitudes of the residual signals effectively indicate changes in the rotor system's state and fault severity.
• The combination of residual signals and deep neural network achieved promising results in identifying complex blade-rubbing faults.
• The proposed nonlinear-based fault diagnosis algorithm demonstrated superior performance compared to a linear-based observer.

Conclusions:

• The developed deep learning approach provides an effective and efficient means for diagnosing blade rub-impact faults.
• Residual signals derived from autoencoder estimations are valuable indicators of fault presence and severity.
• This nonlinear-based fault diagnosis method shows significant potential for real-world turbine health monitoring.