Strategies for Removing Non-stationary Effects in Machine Diagnostics using Model-driven and Artificial Intelligence Techniques

Abstract

Condition-based health diagnostics of critical machine elements like planetary gears and bearings via processing of measured vibration signals play an important role in improving the reliability of industrial machines. Gear and bearing faults can be diagnosed using different frequency and time-frequency domain approaches, as they occur at specific fault characteristic frequencies. However, in practice almost all industrial machines experience different levels of speed variations during operation that challenge the practical implementation of these approaches because of problems like smearing of the measured signal spectrum. To address the issues arising due to operating speed fluctuations, this thesis develops (1) a spectral smearing alleviation technique, (2) an adaptive diagnostic technique under speed fluctuations, and (3) a deep learning based time-frequency image classification and enhancement technique for automatic fault diagnostics. For the smearing alleviation technique, a framework is proposed that utilises fast dynamic time warping (FDTW) to optimally align a filtered shaft vibration signal and a simple sinusoidal reference signal of constant frequency to resample the measured vibration signal to the order domain by using the points of the optimal warping path. As the reference signal is built using a constant frequency, it is observed that the application of the FDTW based approach squeezes the time-dependent shaft rotational frequency and its harmonics towards their corresponding constant peaks through the optimal alignment of corresponding similar data points, which results in a clearer vibration spectrum. The FDTW based smearing alleviation technique is demonstrated for both small and large speed variation cases. This technique is then extended to propose a fault diagnostic approach by adaptively extracting the shaft vibration signal from the measured vibration signal using the variational mode decomposition, instead of the manually intensive band-pass filter. Finally, an automatic fault diagnostic approach based on time-frequency (TF) image classification using a transfer learning based CNN architecture is proposed. A TF image enhancement approach using generative adversarial networks (GANs) is also proposed to analyse short-time Fourier transform (STFT) based TF images of non-stationary machine vibration signals. The techniques developed in this thesis do not require the use of auxiliary speed measuring sensors for their implementation, as such sensors are generally not available in an industrial setting. The effectiveness of the techniques developed in this thesis has been validated through simulation analyses and experimental investigations using measurements from a mechanical fault simulator, an electromechanical drivetrain test-rig, and different wind turbine drivetrains.