Roller bearing fault detection using rotary encoder

Abstract

For many industrial complex machines, there are various challenging issues which include reducing machine downtime, managing repairs and maximising operating times. Any problem or fault in machines can cause failures and downtimes which in turn can lead to significant economic losses. Therefore, industrial companies need to plan organized maintenance strategies for optimum productivity. Condition based monitoring stands out as a highly effective and dependable method widely utilized in the field of maintenance. For rotating systems, rolling bearings are one of the commonly used essential machine elements that are prone to unexpected failures. Traditional monitoring methods predominantly rely on conventional vibration measurements. In recent years, a novel approach to monitoring the condition of bearings using torsional vibration signals via encoder has attracted great attention by scholars. Encoder signals offer notable benefits over standard vibration signals. For instance, encoders have higher signal to noise ratio than accelerometers because they are located close to the rotary components while accelerometers suffer from long and complicated transfer paths. Moreover, encoders are usually built-in type sensors which make them part of the available systems, and this brings additional economic advantages for condition monitoring. However, captured encoder signals are impacted by adverse factors like speed uncertainties due to random load fluctuations and variations in electric supply. These factors predominantly affect low-level signals, where diagnostic information is frequently masked by noise. In order to overcome this challenging problem, researchers continuously strive to create sophisticated signal processing strategies for the effective extraction of crucial diagnostic insights from signals with significant noise interference. In this thesis, conventional and relatively well-established signal processing methods typically employed in vibration-based fault detection are examined and their implementations in encoder-based fault diagnosis are investigated. Particular attention is paid to signal de-noising and enhancement of the measured signals to improve fault detection performance of proposed method. In the first chapter, the problem addressed in this thesis is introduced in detail and the existing literature is thoroughly reviewed. In the second chapter, encoder specific details and employed signal processing methods are described. Briefly, working principle of encoders and Instantaneous Angular Speed (IAS) measurement concept are examined. Theoretical background of the the signal processing methods used in this thesis are also presented in this chapter. The subsequent chapter details the experimental setup and outlines the specifics of the measurement campaign. For the experimental part of the study, an existing Bosch test bench, designed for endurance validation of high-pressure pumps, is employed. For the experimental validation of the fault detection methods used in this thesis, artificial faults are created on the inner rings of cylindrical roller bearings. Due to the complicated design of the setup and the adverse effects encountered during the signal acquisition, measured data inherently contained significant amount of background noise. Chapter four focuses on the signal processing of the measured raw data, aiming to extract hidden information which is critical for detecting bearing faults. An open-source software, Python, along with its signal processing libraries, are employed to process the measured signal and apply various signal processing methods for extracting diagnosic information from measured data. This software choice is based on the diverse range of available techniques and exponential growth observed in this area. In this chapter, three different methodologies for fault detection are introduced. The first employs envelope analysis and spectral kurtosis for detection of faults on the bearing's inner ring. In this context, different fault sizes are examined, and the effectiveness of a hybrid approach is investigated. The results clearly indicate that successful identification of the fault frequency of the bearing's inner ring can be captured via the envelope spectra. In the second method, signal de-noising is the main focus of the investigation. Empirical mode decomposition and singular value decomposition-based bearing fault detection methodology is proposed and proposed method is compared with direct empirical mode decomposition applied signal without prior signal de-noising. The findings reveal that the proposed methodology effectively identifies the bearing inner ring fault frequency in the presence of considerable amount of background noise. In contrast, approaches relying solely on spectrum analysis and the direct application of empirical mode decomposition demonstrate limited effectiveness under similar conditions. When analyzing instantaneous angular speed variations captured by an encoder, directly detecting fault-indicative frequency components is challenging since the bearing fault carries low energy in the signal. Therefore, the third method focuses on removing the most deterministic components from the signal. After filtering, fault frequencies and harmonics were distinguishable in the signal spectra at various speeds, yielding consistent results. Modulation-related sidebands were also observed in the signal. Upon examining the effect of speed, it was found that in our case, detecting bearing frequencies at relatively lower rpms was easier due to the increase in noise content with rising speed. As a result, findings in this thesis leads to the conclusion that encoder signal-based fault detection methods offer an important alternative in bearing condition monitoring. Besides, bearing fault detection capability of the existing methods can be significantly improved by the use of signal de-noising.