Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Condition-based maintenance relies on effective methods to detect any possible mechanical defect in a timely manner. In this thesis a new vibration-based approach for mechanical fault detection is introduced by incorporating signal processing methods into self-supervised learning. It uses signal processing methods to generate target variables from raw signals, and uses the target variables to guide the direction of self-supervised learning. To demonstrate the performance of the proposed approach, it is used to solve a representative problem, i.e., detect the single point outer-race defects in rolling element bearings. For solving the problem, the proposed approach incorporates the high-frequency resonance technique to the self-supervised learning approach. Tests are performed on both simulated and experimental datasets. Results show that, with the signal processing approach providing the domain knowledge of a certain type of mechanical defects, the proposed approach is able to train a model well capable of detecting the defects despite having only healthy vibration signals for training. Moreover, the proposed approach shows better performance comparing with the high-frequency resonance technique and the unsupervised learning approach. Further tests are also taken to show the performance of the proposed approach under different circumstances. It is found that the proposed approach tends to have better performance than the signal processing method even if the domain knowledge is wrong or incorrectly implemented. Therefore, it is suggested to use the proposed approach for detecting mechanical defects, especially when it is uncertain if the domain knowledge is correct and sufficient.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This work investigated the suppression of vibrations due to the rotor unbalance of a passive magnetically suspended flywheel in a space satellite. For this purpose eight electromagnetic actuators were virtually added to the passive system, subsequently creating an active-passive magnetic bearing. Two types of vibration suppression were investigated. The first type is the rejection of the synchronous bearing reaction force, which aims at minimising the transmitted forces from the flywheel to the surrounding, in this case, the satellite. The rotor regime where this is the case corresponds to making the flywheel rotate around its principal axis of inertia. To achieve this, a control loop was developed, targeting to compensate the forces in the permanent magnetic bearing, thereby creating a free-floating condition of the rotor. Regarding stability, a sensitivity analysis on the estimated stiffness of the permanent magnetic bearing was required and a stable range for the accuracy of this estimate was determined. In this stable range, the control scheme proved to be able to reduce the transmitted forces, with increasing performance for increasing accuracy of the stiffness estimate. A secondary type of vibration suppression aimed at minimising the displacements of the flywheel, thus making it rotate exactly around its geometric axis. This requires much higher forces, as the unbalance force, caused by the rotor unbalance, needs to be compensated here. A control law that enables control of the rotor’s parallel and conical mode separately, called decoupled control, was designed for this concrete system and improved with an iterative learning control scheme. This control scheme was able to reduce the rotor displacements to zero with a zero steady-state error, albeit with a very low convergence rate for high rotor speeds, up to 10000 RPM. An actual electromagnetic actuator was designed and manufactured to validate the two control schemes against reality. A test campaign was set up to perform static and dynamic tests, which led to the finding that the amplitudes of the sinusoidal forces that are required, do not reach the expected values, based on calculations. Probable reasons for this are magnetic hysteresis and electromagnetic inductance. This implied the need for an extra frequency- and amplitude-dependent compensation factor in the control schemes, which could be obtained from further actuator characterisation, including extensive test campaigns. Taking into account this need, two valuable vibration suppression algorithms were developed.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Many advanced machine learning algorithms have already been developed to identify faults in rolling element bearings using acceleration or acoustic measurements. However most of these Fault Detection Algorithms are dependent on failure data for training, to be able to accurately detect if the component is faulty or not. The failure data can come from measurements on a machine that has a failed bearing or it can come from destructive tests of the bearing. The failure data are utilized to train a Fault Detection Algorithm that will determine if there is a fault in the bearing, based on the analysis of the measured signal. This signal is measured on the machine during normal working conditions. The trained machine learning algorithm will evaluate the measured signal and will give a prediction if the bearing is faulty or not. Using this prediction the company can then start the necessary maintenance procedure to replace the bearing or continue working with the machine. From the standpoint of training the Fault Detection Algorithm will more data from destructive tests result in a better performance of the Fault Detection Algorithm. This better performance will reduce the number of false alarms and missed detections and this will result in a cost reduction. But every destructive test will bring with an increase in cost for the company. The industry will require a cost benefit analysis of the investment in destructive test(s) to determine if it is economically feasible to have Condition Based Maintenance with a number of destructive tests. Condition Based Maintenance with destructive tests is one of three maintenance strategies, second is Condition Based Maintenance with no destructive tests and the third is Breakdown Maintenance. Condition Based Maintenance without destructive tests will utilise a Unsupervised Fault Detection Algorithm to determine if the bearing is defective or not. The investment for Condition Based Maintenance without destructive tests is less because there are no cost for the destructive tests, but the performance of the Unsupervised Fault Detection Algorithm will be lower in most cases. So there will be a higher cost due to more false alarms and missed detections. The majority of machines in the industry utilise Breakdown Maintenance, the machine is kept in operations until there is a breakdown, if a breakdown occurs the bearing will be replaced and any other damages repaired. In this scenario has the company no investment cost in Condition Based Maintenance but will have a higher cost due to more downtime because of the breakdown(s). This master thesis presents a framework for quantifying the monetary value associated with a single run-to-failure data point. This framework considers various factors crucial to the decision-making process, including the cost of conducting run-to-failure experiments, the enhanced classification performance achieved by models trained on larger datasets, the potential mismatch between laboratory conditions and real-world operating environments, and the value of early fault detection. One of the main findings of this master thesis is the trade-off found between the added value of a destructive test and the added cost of the destructive test. For different machine setups these trade-offs will be different. These differences are attributable to the cost factors of the machine setups and the prediction performance of the algorithm.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This thesis covers the development of a cavitation detection methodology capable of detection at the visual inception point using accelerometer data. The method- ology is based on a normalised spectral correlation and targets the second order cyclostationary component with cyclic frequencies corresponding to the rotational frequency and blade passing frequency. For this, the proposed detection methodology relies on the concept of a cavitation carrier. Different existing carrier detection algorithms were tested including methods based on the Fast Kurtogram and the TATF. The limitations of these carrier detection algorithms led to the development of two new methodologies, one solely based on spectral correlations and one combining both the spectral correlations with the variational mode decomposition for better detection of the central frequency. Electromagnetic interference, an industrially relevant problem, proved to be challenge for carrier detection, but can be overcome by a significantly high cyclic frequency resolution requiring a sufficiently long time series of data. Finally the detection threshold was set by a maximum of 5% false alert rate in healthy data which provided up to 94% accurate detection as well as detection at the visual inception point.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

The reliability and efficiency of mechanical systems, particularly drivetrains, are critical for industrial applications. Drivetrain mechanical faults can lead to significant operational downtime and costly repairs, making early detection essential for maintaining optimal performance. This thesis explores the dynamic content of torque measurements as a novel approach towards the detection of drivetrain mechanical faults, specifically targeting faults in bearings and gears. Initially, a comprehensive literature review is conducted to understand the various mechanical faults that can occur in bearings and gears, as well as state-of-the-art condition monitoring approaches using dynamic acceleration signals. Advanced signal processing techniques are employed to identify mechanical faults. This foundational review informed the subsequent data analysis techniques used in this research. The experimental phase involved a test setup on which a loaded one-stage test gearbox under four different configurations, combining a damaged/healthy gear with a damaged/healthy bearing, under several different load and speed conditions is tested. The gear fault was simulated by creating artificially induced pitting damage on one tooth surface, while the bearing fault was introduced by carving a notch on the inner race, both using a high-speed multi-tool equipped with a grinding cutter. With an appropriate torque transducer, dynamic torque measurements are acquired for each configuration of the setup. Advance signal processing techniques are employed to analyze the collected torque data. For gear fault detection, it is proven possible to identify a local gear tooth fault starting from the raw torque signals. For bearing fault detection, similar techniques were applied, although the identification of the introduced fault is proven more challenging. While the industrial application of torque sensors within drivetrains may not be widely adopted due to their intrusive nature in the powertrain, the methodologies developed in this research provide a significant step towards analyzing torque signals derived from dedicated load models, for example created with the help of strain-gauge measurements. The findings, in this paper, offer a potential pathway for future research to fault diagnostics and prognostics using torque signals, potentially leading to improved maintenance strategies, reduced downtime, and cost savings in industrial settings.