A Study of Non-stationary Signal Processing for Machinery Condition Monitoring

Machinery condition monitoring techniques are carried out based on the knowledge of the characteristics of signals obtained from a machine or plant. These signals are often non stationary signals whose frequency changes over time due to the time-varying natures of machine operations and fault effects. Conventional signal processing techniques are developed based on stationary signals and cannot reveal the time information of the frequency changes. The work undertaken in this research presents a generic study of non-stationary signal processing for machinery condition monitoring.

Starting with examining the concept of non-stationary signals, it can be identified that most condition monitoring signals fall into two main categories: weak non-stationary signals, such as motor electrical current signal and strong non-stationary signal such as machinery vibration and acoustic signals.

For developing techniques to process these two typical non-stationary signals, two experiments were carried out to obtain these them. Firstly, an induction motor drive system was set up based on a two-stage reciprocating compressor; the motor current signals were then acquired for compressor fault detection and diagnosis. Secondly, a set of vibration and acoustic measurement instrumentation was set up based on a diesel engine test system. The engine vibration and acoustic signals were collected for further analysis for engine combustion condition monitoring. The engine was fuelled by different biofuels during data collection allowing a new and efficient method of verifying different sustainable fuels to be developed based non intrusive vibro-acoustic measurements in conjunction with non-stationary signal analysis methods.

A time domain based method, dynamic time warping, was validated and improved for analysing the motor current signal to detect and classify the common faults of reciprocating compressors. Based on the limitations of classical dynamic time warping, a phase estimation and compensation approach is developed to reduce the singularity effect of classical dynamic time warping in order to obtain accurate diagnostic results. A sliding window was also designed to improve computing efficiency. The diagnostic results show that the accuracy and reliability of detection and classification by the proposed dynamic time warping method is higher than that from Fourier transform spectrum and envelope analysis. In addition, the fault detection and classification is based on a root mean square (RMS) linear classifier processes combined with the proposed dynamic time warping method, and is based entirely on time domain analysis which is easier to apply to a real-time condition monitoring system. It was proved that the proposed dynamic time warping is a novel and efficient method for cyclostationary/weak non-stationary analysis.

Various non-stationary signal processing techniques based on time-frequency domain analysis, including Wigner-Ville distribution, fractional Fourier transform and continuous wavelet transform, are investigated to process the engine vibration and the acoustic signals for the condition monitoring of engine combustion. A sound pressure level (SPL) indicator is designed based on the Wigner-Ville distribution (WVD) analysis and the fractional Fourier transform filtering of the engine vibro-acoustic signals. The processing results demonstrate that the combustion induced acoustics can be extracted for the diagnostics of engine combustion process and for condition monitoring.

A root mean square (RMS) linear classifier is developed based on the engine acoustic analysis by time synchronous average and continuous wavelet transform, the classification demonstrates that the root mean square (RMS) values of the continuous wavelet transform coefficients can be used to evaluate the fuel for engine combustion and indicate the engine operating conditions. The analysis results verify that the engine vibro-acoustics have the potential to be used to diagnose the engine combustion process and to monitor the engine operating conditions with the application of suitable non-stationary signal processing techniques. This can be used instead of the cylinder pressure data which is both intrusive and costly to obtain.

Finally, the conclusions and achievements are given based on the entirety of this research work, and suggestions are presented for further research.