Theses and Dissertations (Mechanical and Aeronautical Engineering)

Permanent URI for this collectionhttp://hdl.handle.net/2263/32359

Browse

Now showing 1 - 20 of 387

Experimental investigation of film-cooling hole performance
(University of Pretoria, 2020-06) Mahmood, Gazi I.; Dlamini, Zimase
Film cooling has, over the years, allowed for the operation of modern gas turbines at temperatures far exceeding the limits of the material properties of the turbine components. This has resulted in increased power output and efficiency of the gas turbines. But over 40+ years of research has not culminated in the goal of achieving ideal cooling films, such as from two-dimensional (2D) continuous slots. This study employed a curvature in the forward diffuser section of the film cooling hole; these holes are referred to as cases 1 to 4 in this study. This was expected to improve the performance of the hole. The performance parameters investigated and reported were the discharge coefficient of the holes, the flowfield downstream of the hole exit trailing edge, the temperature field downstream of the hole exit trailing edge and the effectiveness. The effects of pressure ratio, mainstream crossflow, compound angle, hole geometry, manufacturing method, 3D print build orientation, and inclination angle, on the discharge coefficient were investigated. The effects of blowing ratio, hole geometry, compound angle, turbulence intensity and downstream distance from hole exit trailing edge, on the flowfield, temperature field and effectiveness were also investigated. The hole geometries had a diameter of 8 mm and length to diameter ratio equals to 7.5. The compound angle was varied between zero (0) to sixty (60) degrees. The inclination angles of the holes were either thirty (30) and forty (40) degrees. The effect of the compound angle, manufacturing method and 3D print build orientation was found to be negligible for the discharge coefficient. But the above parameters had a significant effect on the adiabatic film cooling effectiveness. Cases 1 to 4 holes showed higher discharge coefficient values as compared to the cylindrical and the laidback fan-shaped holes. This was a result of the development of the flow inside the hole and the resulting exit coolant jet velocity profile and its interaction with the mainstream crossflow. From the flow structure and temperature field measurements it was determined that employing the curvature and the lateral expansion of the cases 1 to 4 holes decreases the height and trajectory of the jet on exit. The decreased height is due to the decreased vertical momentum content of the coolant jet. The decreased trajectory positions the longitudinal vortices closer to the wall which results in better lateral spread of the coolant. From the effectiveness measurements it was found that increasing the compound angle decreases the lateral averaged effectiveness. And a decrease in the lateral averaged effectiveness was observed as the blowing ratio was increased. The case 2 hole geometry resulted in low jet height when in the mainstream, which means that it was closer to the surface that requires cooling. It also resulted in a relatively good lateral spread of the coolant on the surface. And it resulted in the highest laterally averaged effectiveness at most of the compound angles and blowing ratios tested.
Testing and development of a solar-dish cavity receiver for the melting of zinc metal
(University of Pretoria, 2024-08) Le Roux, Willem G.; u19202343@tuks.co.za; Bezuidenhout, Pieter J.A.
Concentrating solar power technologies can be applied to reduce the cost and carbon footprint of zinc melting processes. This study aims to improve the knowledge related to small-scale solar melting using a dish concentrator. This technology can be applied to zinc production as well as a range of small-scale applications, such as casting, recycling, galvanisation, and thermal storage. An experimental and analytical analysis of a rotating cylindrical cavity receiver for the indirect melting of zinc metal using concentrated solar power is presented. A multi-facet parabolic dish with an incident area of 2.85 m² was considered together with a rotating cylindrical cavity receiver. The receiver had an aperture diameter of 0.2 m and the capacity for housing 17 kg of zinc. Five experimental test runs were executed, during which up to 73.5 % of the zinc inventory could be tapped from the receiver in its molten state, and average thermal efficiencies of up to 42 % were achieved. A predictive analytical model considering wind speed, wind direction, and direct normal irradiance was developed and validated against experimental data. A heat transfer efficiency factor was experimentally determined to account for voids in the zinc feedstock. The model was used to predict that approximately 41 kg of molten zinc could be tapped from the experimental setup throughout a typical day with a peak direct normal irradiance of about 900 W/m² and an average wind speed below 2 m/s. A case study highlighted that energy savings of 0.6 kWh are achievable per kilogram of zinc processed by concentrated solar power rather than the conventional induction furnace.
Handling improvement & rollover mitigation by influencing load transfer
(University of Pretoria, 2024) Els, P.S. (Pieter Schalk); Hamersma, Herman; u15346499@tuks.co.za; Zulu, Bongani
The increasing popularity of SUVs in the global market has led to a corresponding increase in accidents and fatalities associated with these vehicles, particularly rollovers. Such incidents have raised concerns about the safety of SUVs and highlighted the need to develop effective strategies to mitigate their rollover propensity and handling limitations. As such, this dissertation investigates various techniques to enhance SUV safety, with a particular focus on reducing their rollover propensity and improving their handling characteristics. To achieve this goal, a comprehensive literature review was conducted to identify the most effective strategies to improve SUV safety. Based on the findings, a control system was developed and modelled in simulation to alter the suspension characteristics and the ride height of the vehicle to improve its handling and reduce the risk of rollover. The constant radius test was used to evaluate the control system's performance and determine the influence of changes in suspension characteristics and ride height on the vehicle's handling behaviour. The test revealed that changes in suspension characteristics and ride height significantly impact the vehicle's handling behaviour and lateral load transfer, leading to improved manoeuvrability and stability. The reduction in ride height was identified as an effective means of reducing the rollover propensity and improving the handling characteristics of SUVs, with promising results that warrant further investigation and implementation in real-world vehicles. Furthermore, the vehicle's lateral dynamics were tested using the double lane change, where an improvement is shown with a change in the suspension settings and after using anti-roll control. The vehicle was also tested for rollover using the Fishhook 1B test, where the characteristic speed of the vehicle was obtained by investigating the suspension configurations that allowed the vehicle to pass the test. Overall, this research makes an important contribution to developing effective strategies for enhancing SUV safety by influencing vehicle load transfer, which is crucial for reducing the number of fatalities on the road.
Development of an approach to incorporate proportional hazard modelling into a risk-based inspection methodology for an optimal inspection policy
(University of Pretoria, 2024-07-26) Heyns, P.S. (Philippus Stephanus); Wannenburg, Johann; lelosud2013@gmail.com; Lelo, Nzita Alain
The health and safety of pressure vessels are a concern in the power, chemical, petrochemical, and other industries handling gases or liquids at high temperatures. The content of a pressure vessel usually is at a substantially different pressure than the ambient pressure, and if not handled carefully it can lead to fatal accidents such as an explosion. Therefore, industry decision makers rely on a risk-based approach to perform inspection and maintenance on the pressure vessel. According to the Risk-Based Inspection and Maintenance Procedures project (RIMAP) for the European industry, risk has two main components: the probability of failure (PoF) and the consequence of failure (CoF). As one of these risk drivers, a more accurate estimation of the PoF will contribute to a more accurate risk assessment. Current methods to estimate the probability of failure are either time-based or founded on expert judgement. This work proposes enhancements to the quantitative risk assessment for the probability of failure ( PoF) and the consequence of failure (CoF) through the utilization of a newly proposed methodology. The proposed methodology consists of incorporating the proportional hazard model (PHM), which is a statistical procedure to estimate the risk of failure for a component subject to condition monitoring, into the risk-based inspection (RBI) methodology so that the PoF estimation can be enhanced to optimize inspection policies. To achieve the overall goal of this work, case studies applying the PHM to determine the PoF for real-time condition data components, are discussed. Also, considering the consequences of failure due to accidents which can occur in pressure vessels using steam and water as reference material, boiling expanding vapour explosions (BLEVEs) are especially important due to their severity and diverse effects such as overpressure, thermal radiation and missile ejection. By way of example this work considers only the overpressure due to BLEVE to model the CoF. The first benefit of this work is that by incorporating PHM with the RBI approach, the PHM uses real-time condition data, to allow dynamic decision-making on inspection and maintenance planning. An additional advantage of the PHM is that where traditional techniques might not give an accurate estimation of the remaining useful life to plan an inspection, the PHM method can consider the condition as well as the age of the component. Another benefit of this work is that risk-based inspection is presently one of the best methodologies to provide an inspection schedule and ensure the mechanical integrity of pressure vessels. RBI usually provides an inspection schedule based on calendar or usage time intervals. This work however optimizes the inspection schedule on pressure vessels, by incorporating proportional hazard modelling (PHM) into RBI methodology as stated above. The work presented here comprises the application of the newly proposed methodology in the context of pressure vessels, considering the important challenge of possible explosion accidents due to boiling liquid expanding vapour explosion (BLEVE) as the consequence of failure calculations. The proposed risk management methodology incorporates a quantitative assessment of the Probability of Failure (PoF), based on Proportional Hazard Modelling (PHM), and the Consequence of Failure (CoF), of an explosion event. The unmitigated risk is thereby quantified by means of a risk matrix, which enables evaluating and deciding on suitable risk mitigation strategies.
Development of low-cost focus control systems for vacuum-membrane solar dish facets
(University of Pretoria, 2024-07-25) Le Roux, Willem G.; duncan.mcgee@tuks.co.za; McGee, Duncan Sean
Concentrating solar power (CSP) is a growing method of harnessing energy from the sun for generating electricity and process heat, especially in South Africa which boasts one of the most plentiful solar resources globally. A small-scale CSP system, consisting of a multi-faceted concentrator that employs vacuum-membrane technology, is actively being developed at the University of Pretoria. The facets constituting this innovative design are constructed from a reflective polymer-based membrane adhered to the rims of readily available and cost-effective elliptical television antennas. A crucial step involves creating a vacuum within each facet, forming a near-parabolic membrane shape. Previous studies found that the membrane depth shifts slightly due to varying ambient conditions throughout an operational day. These slight depth shifts lead to major focal point shifts, reducing the CSP system’s overall efficiency and performance. The first goal of this research was to examine in more detail how static ambient conditions impact the displacement of membranes used on vacuum-membrane solar-dishes. A controlled-environment enclosure was employed to achieve this, allowing for the independent manipulation of a facet’s ambient pressure and temperature. The second goal was to investigate methods to mitigate membrane displacement. Various manufacturing techniques were investigated within the controlled-environment enclosure, which included alterations in pretension, changes in membrane thickness by removing the removable plastic layer on the EverBright mirror film, and adjustments to overall facet sizes. Results revealed that ambient temperature impacted the membrane displacement significantly more than ambient pressure. It was also determined that opting for a small facet with a thin membrane and high pretension will effectively minimise membrane displacement. This, however, would not suffice to mitigate membrane displacement. The outdoor test results of a facet without a focus control system indicated that solar radiation, specifically global horizontal irradiance (GHI), affected the internal temperature (depending on the wind velocity), and therefore also affected the membrane depth. Furthermore, to further reduce membrane displacement, low-cost focus control systems were investigated. A focus control system for USD 29.34 maintained a constant differential pressure for a vacuum-membrane facet within the required accuracy of ±2 mm membrane displacement. An attempt was made to further mitigate membrane displacement by incorporating the effects of temperature on membrane stiffness, which demonstrated slight improvements. A focus control system consisting of a low-cost Hall effect module actively monitoring membrane depth emerged as the most effective in eliminating membrane displacement, with an increase of about 0.09 mm and a decrease of approximately 0.02 mm from an initial depth of 10 mm. This level of stability will ensure that the facet maintains a consistent optical performance, ultimately advancing the reliability and efficiency of low-cost vacuum-membrane technology.
The development of a linear cutting machine used to characterise fem modelling parameters for cutting ug2 reef
(University of Pretoria, 2024-05) Heyns, P.S. (Philippus Stephanus); Malan, Daniel Francois; u18018832@tuks.co.za; Du Preez, Ulrich
Mining is still an important industry in South Africa. Traditional mining methods involving drilling and blasting have steadily been replaced by mechanised mining systems in soft rock environments but not in hard rock environments. Mechanised mining systems can lead to continuous mining, which lead to improved rates of face advance and better utilization of the invested capital. A fundamental understanding of the tool-rock interaction, for rock found in gold and platinum mines in South Africa, and potential solutions to problems in mechanised mining methods in narrow reef hard rock mines, are required. South Africa has two main platinum reef deposits namely the Merensky reef and the UG2 reef. A renewed effort is required to study the problem of mechanical mining, and develop numerical models, that take the rock properties into account. This will allow optimisation of the mechanised cutting in hard rock environments. In this dissertation a linear cutting machine (LCM) was designed and manufactured to conduct laboratory scale cutting tests on both sandstone and UG2 reef samples. Firstly sandstone was cut to ensure that the LCM functions as expected. By conducting tests on sandstone, it ensured that all the functions of the LCM could be optimized. The comparison between the samples showed that there are similarities between the results from the different rock types, but some inconsistencies were found. The key difference is that the sandstone considered here has little to no variance in strength on a millimetre scale whereas the UG2 reef sample has large variance in strength on a millimetre scale. This introduces uncertainty in the results due to added variance. Another problem is the inconsistency in rock properties of the UG2 reef. The rock properties of the UG2 reef changes a lot from reef to reef as well as different areas in the mine. The results showed that the optimal cutting parameters are similar for sandstone and UG2, but there are some differences. The depth of cut has a larger influence on the results of UG2 reef samples than for the sandstone samples. Therefore if the sandstone data was used to make design decisions for new mining equipment the decision might have been incorrect due to the assumption that sandstone i and UG2 cut similarly. An important difference between cutting sandstone and UG2 reef is the size of the chips formed. At 2 mm cutting depth, for both samples, the force signals were impulsive and the material produced was fine fragmentations. At a cutting depth of 4 mm, for both samples, the force signals had a saw tooth shape. This implies larger fragment sizes were formed. The sandstone produced large fragments whereas the UG2 still produced fine fragmentations. This fine fragmentations is undesirable in underground mining conditions as this causes that material can not be easily cleaned and removed from the stopes. A fast Fourier transform (FFT) analysis on the cutting signal showed that the sandstone had a peri- odic cutting force signal whereas the UG2 does not have a periodic cutting force signal. Also for the sandstone a good relationship was present between the size of the chips formed and the dominant frequencies of the FFT. The numerical simulations showed that there are various model parameters that influence the results and while other have little effect. Thus, there are many choices that need to be made about model parameters, such as element size, element type, boundary conditions, contact parameters and model parameters. Some are based on material properties and other are obtained through trial and error. It is possible to model rock cutting of UG2 reef samples using the Ansys LS-DYNA multi-physics simulation software and the continuous surface cap model (CSCM). But this is only possible by editing the model parameters through trial and error for one set of cutting parameters. When the cutting parameters are changed, the model does not give acceptable results. Future work is required to improve the ability of models to generalise when the cutting parameters change.
Development and application of a photogrammetry based statistical shape analysis technique for condition monitoring of rotating structures
(University of Pretoria, 2024-05) Oberholster, Abrie; Heyns, P.S. (Philippus Stephanus); benji.gwashavanhu@hotmail.com; Gwashavanhu, Benjamin
Large rotating structures such as wind turbine blades require specialized measurement techniques for the purpose of online condition monitoring and assessment. Contact transducers such as accelerometers and strain gauges are traditionally used to capture vibrational data that can be analysed to understand the dynamics of a system. These are however intrusive in the sense that they must be physically attached to the structure under investigation. In addition, they are point-wise in nature, implying that measurements are only captured for those specific locations where the transducer is attached. They may also alter the local structural properties at the point of attachment, including additional mass loading effects of the sensor on light structures. Optical techniques such as photogrammetry and laser vibrometry are promising alternatives that have been receiving much attention. 3D Point Tracking (3DPT) and Digital Image Correlation (DIC) constitute photogrammetric-based optical measurement techniques that have proven to be efficient for the vibration analysis of rotating machinery. In addition to complex image processing software and tracking algorithms, these two approaches typically require surface preparation in the form of markers and speckle patterns. The surface preparation typically requires a system shutdown which can be complicated and costly. Applied surface treatments also do not last throughout the lifespan of the structure and often must be reapplied. In order to track specific pixels for 3DPT and DIC, the lighting on the surface of the structure needs to be closely monitored since the tracking is based on pixel gray scale values. These requirements limit the applicability of photogrammetry as a condition monitoring tool, especially when it comes to field or outdoor full-scale testing. Photogrammetric shape-based analysis is an alternative approach that does not require prior surface preparation. By focusing on the boundary outline of a structure, the technique is a suitable candidate for outdoor investigations where consistent uniform lighting on an entire structural surface may be impossible. It can also be applied to large structures with significant levels of rigid body motions. To date, this approach has not yet been employed for dynamic analysis of machines. The concept of shape analysis is typically applied for object recognition or shape matching in applications such as Content Based Image Retrieval (CBIR). Thus a single image is captured and then analysed to be matched to another image stored in a database, for instance. This research focuses on the development and application of a shape based photogrammetric technique that can be used to capture dynamics of rotating structures without the requirement for surface preparation. The goal of the study is to develop an approach that can be used to distinguish faults in the system and classify machine behaviour for condition monitoring purposes. In this type of application, sequences of images of a structure in operation are captured, and boundary contours of an object in the images extracted. Through defining parameters that characterise contours extracted from each of these images, and then monitoring the variation of these parameters in time, the idea of shape analysis can be adopted for condition monitoring of machines as an optical non-contact measurement technique. Shape Principal Component Descriptors (SPCDs) determined by performing Shape Principal Component Analysis (SPCA) of Fourier descriptors calculated from shape signatures of the extracted contours are the parameters investigated in this study. Condition monitoring strategies for rotating structures are discussed in a literature review that highlights the importance of structural health maintenance and the shortcomings and limitations of conventional measurement techniques. Advanced optical measurement techniques that include photogrammetry and laser vibrometry are discussed to describe and illustrate the evolution of recent noncontact technologies as viable tools for condition monitoring purposes. Typical applications of photogrammetric techniques such as surface strain measurement are highlighted. Instances that demonstrate the successful use of optical approaches to capture dynamics of rotating structures are discussed. This lays out a foundation onto which the necessity of advancing optical based strategies into more suited techniques for industrial application purposes is built. The concept of shape analysis is introduced and its SPCDs investigated for a 2D shape that has in-plane form variations associated with it. On application to a physical rotor system, it is illustrated that different dynamics of the rotor resulting from different faults of unbalance, rotor-stator rub and hydrodynamic bearing oil instabilities can be detected and classified using the shape-based approach. It is clearly illustrated that the multi-dimensional measurement technique provides insights into the behaviour of a rotor system, as confirmed by uniaxial conventional proximity probe measurements. The proposed approach complemented the widely used proximity probe sensing technique in terms of investigating rotor systems. An extension of the approach from 2D to 3D is also presented, starting with analysing how different shape descriptors influence the form of contours representing blade shapes in 3D. A detailed numerical investigation in which a Finite Element (FE) model of a physical rotor is analysed for changes in dynamic behaviour resulting from introduced damage in the blades, is conducted. The FE environment provides a platform in which the procedures for 3D shape analysis can be developed and tested before the proposed approach can be implemented experimentally. An experimental study that involves the use of a calibrated system of high-speed cameras to synchronously capture stereoscopic images of a rotating turbomachine is then presented. Variations in the dynamics of rotating blades are investigated and through a revolution based Principal Component Analysis (PCA) of SPCDs, the feasibility of a shape-based condition monitoring approach for turbine blades is illustrated. A comparative study to investigate the performance of PCA of SPCDs in relation to Kernel Principal Component Analysis (KPCA) is also conducted, and it is shown that KPCA outperforms PCA in terms of classifying different blade faults. The feasibility of using Multi-domain Statistical Features (MSFs) as feature vectors to which PCA or KPCA is applied for classification purposes is also presented. Results indicating how well different blade damage modes can be distinguished are provided, and it is clearly illustrated that MSFs are more robust to noise contamination in the signals compared to using the raw SPCDs time data.
Sensitivity study of dynamic features to monitor the condition of vibrating screen isolators
(University of Pretoria, 2023-11-07) Heyns, P.S. (Philippus Stephanus); u12019420@tuks.co.za; Pienaar, Simeon
Vibrating screens form an integral part in the processing of bulk materials. The major components of a vibrating screen are purposefully designed and selected to contribute to the operational parameters required. The monitoring of each components function is essential to guarantee optimal performance and reduce downtime through predictive maintenance. This is the premise of a condition-based maintenance strategy which is becoming increasingly popular with the advent of new, affordable technology and advanced signal processing and classification algorithms. For vibrating screen components such as motors, gearboxes, and bearings there exist several validated monitoring strategies. However, for components such as vibrating screen isolators very few working strategies exist. The isolators are a notoriously difficult component to monitor due to the operating environment, limited access, isolator geometry and material composition. Each of these factors restrict the use of conventional non-destructive testing (NDT) and visual inspection. Another monitoring strategy is a static compression or displacement technique which suffers practical relevance for very large vibrating screens with numerous isolators. The most promising techniques for isolator condition monitoring are vibration-based approaches. One such approach is the evaluation of modal parameters by experimental modal analysis (EMA) and modal parameter extraction. The premise for this approach is that a deterioration of isolator condition will manifest as a change in stiffness which directly influences modal natural frequencies, particularly for the rigid body modes (RBM). Another approach is the use of signal processing to extract features directly from operational vibration measurements. These features need to be sensitive to faults such that their change will be indicative of a fault developing. The difficulty of the latter approach is that the evaluation of feature sensitivity is expensive when done experimentally. It is therefore common to use a model-based approach for feature evaluation. This implies the use of simulated measurements from which the feature sensitivity can be established. However, this does not excuse the use of some experiments to validate both the numerical model as well as the feature sensitivity results. The purpose of this dissertation is to evaluate the sensitivity of identified features to known faults in isolators using both numerically simulated and experimentally obtained measurements. The same premise as for an EMA approach is used to evaluate isolator deterioration (i.e. as a change in stiffness only). A numerical model of a vibrating screen is developed which uses linear approximations for the exciters and isolators. However, the experiments were performed on a vibrating screen with different isolator types considered (spring steel and rubber-based isolators). This was done to demonstrate the generic use of the features for isolators condition monitoring in that they are not coupled to only one ii isolator type. The features identified for this study are exploratory as research has yet to identify the most appropriate features for isolator condition monitoring. The entire operating envelope (startup, steady operation and coast down) of a vibrating screen, with no material on the deck, is considered. An EMA is also conducted to evaluate the behaviour of modal parameters for changes in isolator condition. The sensitivity results from both the simulated and experimental measurements are compared to one another. It was found that the features that undergo the highest percentage change and are the most sensitive to changes in isolator stiffness are those obtained from the transient startup and coast down envelopes. The steady operating orbit features underwent considerably smaller changes. The rigid body mode natural frequencies obtained by EMA and those extracted from the transient startup and coast down envelopes are comparable in magnitude and behaviour. However, from the experimental results the features and how they change are dependent on the type of isolator used, the temperatures of the isolators and exciters as well as the sensor locations.
Automated learning rates in machine learning for dynamic mini-batch sub-sampled losses
(University of Pretoria, 2020-12) Wilke, Daniel Nicolas; u11207312@tuks.co.za; Kafka, Dominic
Learning rate schedule parameters remain some of the most sensitive hyperparameters in machine learning, as well as being challenging to resolve, in particular when mini-batch subsampling is considered. Mini-batch sub-sampling (MBSS) can be conducted in a number of ways, each with their own implications on the smoothness and continuity of the underlying loss function. In this study, dynamic MBSS, often applied in approximate optimization, is considered for neural network training. For dynamic MBSS, the mini-batch is updated for every function and gradient evaluation of the loss and gradient functions. The implication is that the sampling error between mini-batches changes abruptly, resulting in non-smooth and discontinuous loss functions. This study proposes an approach to automatically resolve learning rates for dynamic MBSS loss functions using gradient-only line searches (GOLS) over fifteen orders of magnitude. A systematic study is performed, which investigates the characteristics and the influence of training algorithms, neural network architectures and activation functions on the ability of GOLS to resolve learning rates. GOLS are shown to compare favourably against the state-ofthe-art probabilistic line search for dynamic MBSS loss functions. Matlab and PyTorch 1.0 implementations of GOLS are available for both practical training of neural networks as well as a research tool to investigate dynamic MBSS loss functions.
Modelling and multi-objective optimisation of heat transfer characteristics and pressure drop of nanofluids in microtubes
(University of Pretoria, 2020-02) Mehrabi, Mehdi; Meyer, Josua P.; u14011809@tuks.co.za; Meyer, Marcel
A literature study was performed on the inner mechanisms of nanofluids and flow in microchannels. With ever changing technology, the need for smaller and more efficient devices has come about in the last couple of years. With the shrinking in size of components in electronics, an increase in heat has become a notable problem. With conventional heat transfer fluids not being able to handle the required heat removal rates, research into fluid enhancing has been of great interest. A nanofluid is a fluid with enhanced heat transfer potential, which can solve the problem of extracting enough of the added heat of new-age components. This will allow electronics to work with increased power and accomplish tasks faster. Nanofluids have been a very controversial method of heat transfer as problems with stability were keeping the fluid from replacing traditional heat transfer fluids. Some research has been done on the models used for simulating and defining the thermal properties of nanofluids. Added accuracy of the models has been seen in recent years. However, no optimal setup for nanofluids has been found in terms of combining parameters like the base fluid and nanoparticle, as well as the concentration and diameter of the nanoparticle. An optimal setup of this kind would produce the best heat transfer rates at the lowest pressure drop. The simulation of nanofluids was done in Ansys CFD. The validation was done with previous literature that had experimental and numerical results. The validation had a very good outcome as some of the temperature data inside the microchannel presented a good correlation to previous work. The setup of the model for simulation and duplication to create a design study was also described and shown. This was done to ensure that the model can be used again if further investigation is needed. This will enable one to determine the effect of a new nanoparticle on the field of study to continuously improve on the model. The results indicated the best nanoparticle to use with the best base fluid to ensure the lowest pressure drop and highest heat transfer. This was done with a multi-objective optimisation general algorithm. The outcome of the optimisation was that silicon dioxide, as nanoparticle, and water, as base fluid, would give the optimal setup. The diameter also appeared to have a very small effect on the outcome.
The shape optimisation of compliant structures to produce a desired snap-through load-placement path
(University of Pretoria, 2023) Kok, Schalk; Wilke, Daniel Nicolas; bouwer2608@gmail.com; Bouwer, Johann Mynhardt
The research problem completed in this thesis is to develop an efficient procedure to design the optimal shape of compliant mechanisms for specified load-deflection paths and snap-through behaviour. Here, computationally intensive simulations are required to approximate the entire load path as a function of the shape or spatial variables. Solving this problem efficiently, as will be demonstrated in this thesis, requires unconventional multidisciplinary strategies, as conventional modern techniques are ineffective or impractical. In the case of simulation in the loop, where the numerical simulation is evaluated directly in the optimisation loop, unavoidable numerical discontinuities are present in the objective function. These discontinuities grow in number and size as the complexity and dimensionality of the problem increases. Modern gradient-based optimisers are incapable of bypassing these discontinuities and terminate prematurely, misrepresenting these discontinuities as local minima. Therefore, this research advocates for the use of non-negative gradient projection points, utilised by gradient-only optimisation techniques, to define meaningful shape optimisation solutions. These techniques ignore the discontinuous changes in function value to find non-negative gradient projection points. Simulation in the loop is limited by the sequential nature of iterating from design to design, incurring the time penalty of having to wait for time-consuming simulations. Surrogate-based optimisation parallelises the time-consuming computational simulations, enabling computationally efficient surrogate models to be constructed instead. As the load paths evolve with systematic load application, these models evolve not only spatially but also temporally as a function of a pseudo-time variable. Spatial and temporal variables result in two sources of anisotropy. Firstly, the response anisotropy of the function as the temporal evolution of load-deflection curves is distinct from the spatial evolution as a function of shape variables. Secondly, sampling anisotropy as spatial variables are sampled distinctly from the usually densely sampled temporal variable resulting from iteratively evolving the load-deflection path. Response and sampling anisotropies can result in significant mismatches between the model forms from sampled data and typical isotropic kernels used in surrogate construction. This study develops two solutions that address the response and sampling anisotropies, respectively. The results definitively demonstrate that both sources of anisotropy need to be addressed to construct accurate surrogate models that are meaningful for the shape optimisation problem. First, a novel coordinate transformation scheme is developed to transform the function response to be more isotropic as a data pre-processing step. The key here is to utilise gradient information to estimate an updated isotropic reference frame, which also makes the strategy more tractable for higher dimensions. Secondly, sampling anisotropy is addressed by redistributing the surrogate kernels over the spatio-temporal domain and relying on regression to fit the surrogate response as opposed to limiting the centres to the sampling points. These improved surrogate models require significantly fewer computational resources to complete the optimisation problem as compared to placing the simulations directly in the optimisation loop.
Latent approaches for unsupervised learning-based methodologies in gearbox fault diagnosis applications
(University of Pretoria, 2024-02-28) Heyns, P.S. (Philippus Stephanus); Wilke, Daniel Nicolas; Schmidt, Stephan; u15020658@tuks.co.za; Balshaw, Ryan Cameron
The monitoring of critical industrial assets is of fundamental importance to the longevity and sustainability of many industrial sectors. Vibration-based fault diagnosis methodologies seek to enhance the monitoring process by providing rich diagnostic information related to the instantaneous asset health state to support maintenance decisions. While conventional signal analysis techniques applied to vibration data have been demonstrated to provide information related to the asset health state, their utility is often impeded by other vibration sources and time-varying operating conditions. Hence, powerful data-driven techniques are used to extract asset health state information to support the decision-making process. Learning-based methodologies offer the ability to perform effective fault diagnostics by extracting informative features for the fault diagnosis problem. Unsupervised learning latent-based approaches, e.g., latent variable models (LVMs), offer an effective strategy to bypass dataset-dependent feature engineering practices and the requirement for labelled fault conditions inherent to many learning-based methodologies. Temporal preservation is an important approach which advocates for the preservation of time structure to produce an informative latent manifold, and new possibilities and opportunities for condition inference are potentially available from a latent variable perspective. However, this is only the first step to unlocking the potential of LVMs for asset condition monitoring. Improving the application of temporal-preserving LVMs is crucial to advance unsupervised learning-based methodologies for effective utilisation and adoption in industrial applications. The approaches considered in this thesis address current issues in gearbox condition monitoring, and areas for improvement are identified for temporal-preserving LVMs. Three research directions are established to improve the application of temporal-preserving LVMs for gearbox fault diagnostics. The first direction focuses on leveraging the latent manifold to drive effective condition inference, the second direction explores different LVM training approaches, and the third direction investigates latent health indicators for LVMs developed using different training approaches. First, a latent indicator framework for temporal-preserving LVMs is proposed, which identifies unique classes to capture several manifold perspectives for latent health indicator derivation. The benefits of an ensemble-based latent condition inference approach are identified, and the latent manifold is demonstrated to be a fruitful source for fault diagnostic information. This enhances LVM-based methodologies for the fault diagnosis task. As the models are fault condition agnostic, it is demonstrated that a diverse set of latent health indicators is necessary to obtain a robust and insightful condition inference approach. Second, the parameter estimation process for temporal-preserving LVMs is studied to explore methods used to construct the latent manifold. For data-driven models in condition monitoring, two strategies exist to estimate the model parameters: offline and online training. Albeit equally applicable for LVMs, conventional research applications disregard online strategies for LVMs. Online training suits scenarios which lack historical data from a healthy asset. Thus, understanding the implications of using online training for temporal-preserving LVMs is considered. Moreover, two main methods are identified for online training: reconstruction-focused and interpretation-focused methods. Both online and offline models are demonstrated to be effective using basic methods from the proposed latent indicator framework, which reveals their utility in condition monitoring and confirms their ability to perform effective fault diagnostics. Finally, the proposed latent indicator framework pre-supposes an offline training approach and was only assessed using offline models. Therefore, the latent indicator framework bounds are studied using online models. Implicit assumptions in the proposed latent indicator framework are identified to develop a new category of latent health indicators, and a set of online latent indicators is proposed for online methods. The applicability of the two indicator categories is studied. Diversity is highlighted to be a prominent factor in the indicator categories, and the study advocates for the use of a plethora of diverse indicators for both offline and online methods. The research carried out in this thesis indicates that the considered temporal-preserving latent variable approaches are advantageous to the fault diagnosis task. The identified research directions ensure that the latent manifold of offline and online temporal-preserving LVMs can be effectively applied to drive meaningful condition monitoring. By combining offline and online models with a diverse, categorised latent indicator framework comprising of heterogeneous collection latent health indicators, a holistic fault diagnosis strategy for temporal-preserving LVMs can be developed. Temporal-preserving LVMs are established as an effective tool for unsupervised learning-based condition monitoring using gearbox vibration data.
Investigation of the effects of welding parameters on the tensile strength and fatigue life of the structural welded joints S355J2+N steel plate
(University of Pretoria, 2023) Heyns, P.S. (Philippus Stephanus); vusimuzi87@gmail.com; Moloi, Vusimuzi Patrick
The railway industry wants to obtain structural welded joints that have optimum weld strength and fatigue life. Since the variance of welding parameters producing welding defects and weak structural integrity, there are uncertainties in the selecting the welding parameters that produce a good structural welded joint with the necessary weld quality. This necessitates the optimisation of the welding parameters to produce welded joints with improved mechanical properties. The impact of various welding parameters (i.e. wire-feed speed (WFS), voltage and travel speed) on the ultimate tensile strength (UTS) and fatigue life of S355J2+N single V butt weld produced by metal inert gas (MIG) robotic welding was experimentally investigated. The design of experiments (DOE) approach was used to optimise the welding parameters to ensure the reliability of the experimental results. A removable ceramic weld backing bar was used to improve weld root penetration and minimise the risk of lack of fusion. To ensure weld quality and reliability of the experimental results the flush ground welded joints were used to minimise the geometric notch effect. A minimum number of two specimens for each number of experiments were tested to ensure the correct evaluation of welds. The magnetic particle testing (MT) technique was used to detect the welding defects that might have an impact on the material properties of the welded joints. Analysis of variance (ANOVA) was used to precisely demonstrate which welding parameters had the greatest impact on the performance output of the welded joint and to determine interactions between welding parameters. It was observed that varying the welding parameters had an impact on the weld quality. An increase in voltage and travel speed at lower WFS are the primary contributing factors to weld defects. Only the defect-free specimens were tested to avoid making the experimental results inconclusive for statistical analysis. According to the ANOVA results, voltage and travel speed interact to affect the welded joint’s UTS. Increasing voltage increases the UTS of the welded joints at the higher ranges of travel speed, while decreasing the UTS in the lower and medium range of travel speed. The most influential welding parameter that affects the UTS of the welded joint is travel speed. The fatigue life of the welded joint is affected by interactions between WFS and travel speed, as well as the voltage and travel speed. Increasing WFS increases the fatigue life at the medium range of travel speed. When welding at lower speed, the fatigue life duration becomes longer as the voltage increases. The fatigue life of the welded joint is significantly influenced by the WFS. The optimal welding parameters for the welded joint is A2B1C1 (i.e. WFS at level 2, voltage at level 1 and travel speed at level 1) for better UTS and fatigue life. This research reduces uncertainties in the selection of optimum settings of welding parameters of a MIG welded joint. The welding parameters that significantly affect the welded joint mechanical properties performance were identified. The optimum welding parameters selection for UTS and fatigue life can be developed. Undesirable welding defects that affect the structural integrity of the welded joint can be minimised by an improved selection of welding parameters.
Heat transfer of nanostructure coating on commercially micro-enhanced refrigerant tubes under pool boiling and falling film boiling conditions
(University of Pretoria, 2023) Bock, Bradley D.; Thome, John R.; u17008515@tuks.co.za; Dickson, Dian
Refrigeration tube manufacturers commercially produce micro-enhanced tubes for the refrigeration industry with the aim to enhance outside heat transfer coefficients. A new engagement on commercially micro-enhanced tubes is to nanocoat these tubes in an inexpensive attempt to further enhance the heat transfer coefficients. Decreased nucleation site dryout because of increased wettability imposed by a hydrophilic copper oxide nanocoating is hypothesised to enhance heat transfer. In this study, commercially micro-enhanced tubes were nanocoated with copper oxide and tested in R134a refrigerant in pool boiling, and falling film boiling conditions; and additionally in condensation and dryout performance tests. Relevant literature indicated high surface wettability from hydrophilic nanocoatings increasing heat transfer coefficients under pool boiling conditions. This was said to be due to less surface dryout because of the surface’s liquid affinity, however, numerous studies resulted in heat transfer coefficient degradation because of nucleation site flooding and high surface energy requirement. This study proved through scanning electron microscopy that copper oxide nanocoatings successfully coated all microstructured tubes evenly by using a coating procedure and a dedicated tube coating machine without impeding the surface features. For the uncoated tubes in pool boiling at 5°C, the EHPII and GEWA-B5 tubes were the most independent from heat flux with flat heat transfer coefficient curves. They performed the best with heat transfer coefficients of 299% and 318% higher than a plain roughened tube, whereas the low-finned GEWA-KS performed moderately well with heat transfer coefficients 57% higher than the heat transfer coefficients of the plain roughened tube. An increase in saturation temperature to 25°C decreased the EHPII’s heat transfer coefficients by 10%, whereas the GEWA-KS’s heat transfer coefficients increased by 25%. The copper oxide nanocoated tubes in pool boiling at 5°C performed similar to the uncoated tubes in pool boiling. However, the copper oxide nanocoating generally decreased the heat transfer coefficients from the uncoated case, where the EHPII, GEWA-KS and plain roughened tube had heat transfer coefficients approximately 89%, 91% and 85% of the uncoated heat transfer coefficients respectively. The GEWA-B5 was affected the most with heat transfer coefficients approximately 60% of the uncoated heat transfer coefficients. For the uncoated tubes in falling film boiling at 5°C, the EHPII and GEWA-B5 tubes were the most independent from heat flux with flat heat transfer coefficient curves. They had a similar high heat transfer performance, whereas the GEWA-KS performed moderately well. The plain roughened tube’s heat transfer coefficients had an average heat transfer coefficient of 8.6 kW/m2 · K. The increase in saturation temperature to 25°C decreased the EHPII’s heat transfer coefficients with 7%, whereas the GEWA-KS’s heat transfer coefficients increased with 7% on average. The copper oxide nanocoated tubes in falling film boiling at 5°C performed with marginal improvement compared to the uncoated tubes in falling film boiling. The copper oxide nanocoating generally had a degrading effect on the GEWA-B5 and plain roughened tube achieving about 66% of the uncoated heat transfer coefficients. Moderate enhancement for the EHPII tube with a peak enhancement of 110% at 100 kW/m2was seen, however remained steady at achieving 99% of the uncoated heat transfer coefficients on average. The greatest enhancement was achieved by the GEWA-KS with an average of 119% of the uncoated heat transfer coefficients, and a peak enhancement of about 160% at 20 kW/m2, similarly seen at 25°C saturation temperature. The dryout performance tests showed no improvement for all tubes through the addition of the nanocoating and further experimental research is required to deduce an optimal multiscale enhancement to increase dryout performance. The addition of the copper oxide nanocoating is therefore not a reliable option to enhance heat transfer coefficients except for the GEWA-KS tube under falling film conditions. Degradation of the heat transfer coefficients are thought to be due to the flooding of the nucleation sites and the degradation in the hydraulic bubble pumping action of the microstructure capillary channels facilitating the sensible and latent heat transfer. The condensation tests showed a consistent degradation in heat transfer coefficients and is likely due to the inefficient dry surface exposure because of inadequate liquid expulsion from microstructure cavities and general surface liquid retention by the hydrophilic copper oxide nanocoating.
Predicting and analyzing geometric and morphing wing characteristics of the Grey-Headed Albatross
(University of Pretoria, 2023) Smith, Lelanie; Schoombie, Janine; u17001430@tuks.co.za; Winter, Alexander Ernest
Geometric and aerodynamic properties of various avian species allow engineers and biologists to gain valuable insight into the evolutionary honing of the capabilities of natural flyers. Very little research has been done to establish reliable 3D models and detailed descriptions of the aerodynamic characteristics of the Grey-headed Albatross. Therefore, the purpose of this work was to determine the static and passively morphed geometric and aerodynamic characteristics of the Grey-headed Albatross. A laser scanned 3D point cloud of a Grey-headed Albatross wing specimen was used to obtain spanwise airfoils using the PARSEC method, a novel method to the field of avian wings. A single objective optimization study using a pseudo 2D computational fluid dynamics model was done on an averaged airfoil of the arm section of the Grey-headed Albatross to determine the maximum potential aerodynamic efficiency (lift-to-drag-ratio) at a Reynolds number of 2 × 105 . This delivered the first reliable estimate of the passive morphing that an avian wing undergoes. The optimized Grey-headed Albatross airfoil decreased in camber creating a more streamlined body when compared to the highly cambered static airfoil. The optimized airfoil exhibited a maximum lift-to-drag ratio of 44 (αactual = −0.5 ∘ , αgeometric = −11.5 ∘ ) when compared to the baseline airfoil with a lift-to-drag ratio of 3 (α = 16∘ ). The increase in lift-to-drag ratio was partly due to the drastic decrease in pressure drag from 0.395 to 0.029 between the static and optimized airfoil, a decrease by a factor of 13.6. The optimized airfoil geometry was similar to that of a 3D laser scan which was done on a GHA wing in the presence of airflow. The increase of the aerodynamic efficiency is consistent with the notion that Grey-headed Albatrosses are efficient flyers.
The effect of large values of relative surface roughness on heat transfer and pressure drop characteristics in the laminar, transitional, quasi-turbulent, and turbulent flow regimes
(University of Pretoria, 2023) Everts, Marilize; u15000380@tuks.co.za; Mahomed, Faiyaad
Numerous studies experimentally investigated the heat transfer and pressure drop characteristics of laminar, transitional, quasi-turbulent, and turbulent flow through smooth tubes, however, studies that investigate the effect of surface roughness on the heat transfer and pressure drop characteristics in macrotubes are sparse. This study experimentally investigated the effect of large values of relative surface roughness on the heat transfer and pressure drop characteristics using simultaneously measured heat transfer and pressure drop data. Experiments were conducted using a horizontal circular tube with a base inner diameter of 5 mm, a length of 4 m, and a square-edged inlet. The constricted diameter was used for the rough tubes. One smooth and two rough tubes, with relative roughnesses of 0.04 and 0.11, were tested at different constant heat fluxes between Reynolds numbers of 100 and 8 500. Water was used as the test fluid and the Prandtl number varied between 3 and 7. The smooth tube was used for validation purposes, as well as a reference to compare the rough tube results. The heat transfer and pressure drop results were plotted and discussed using the average Nusselt numbers, friction factors, and Reynolds numbers. Contrary to the trend in the Moody Chart, a significant increase in friction factors with increasing surface roughness was observed in the laminar flow regime. Free convection effects of both Nusselt numbers and friction factors were suppressed by the velocity of the fluid caused by the large roughness elements, even so at low Reynolds numbers. It was found that for a rough tube with a relative roughness of 0.04 at a constant heat flux of 3 kW/m2, the transitional flow regime occurred at a Reynolds number of 560, and the quasi-turbulent flow regime at a Reynolds number of 760. For a tube with relative roughness of 0.11, the critical Reynolds number was below 390 and the quasi-turbulent flow regime occurred as early as at a Reynolds number of 490. In general, for both the friction factors and Nusselt numbers as functions of Reynolds number, there was a clear upward and leftward shift with increasing surface roughness across the different flow regimes in comparison to a smooth tube. The transitional flow regime for friction factors and Nusselt numbers were narrower and had a differing profile in comparison to smooth tubes. The relative roughnesses of both rough tubes were in the saturating region and the influence of heat flux and thus the Grashof number had little effect on the critical Reynolds number. The quasi-turbulent and turbulent flow regimes occurred at lower Reynolds number for increasing roughness. Trends of the friction factors and Colburn j-factors were similar in all the flow regimes for the smooth and rough tubes and the boundaries between the flow regimes were the same for both the pressure drop and heat transfer results. When comparing the relationship between heat transfer and pressure drop, it was found that an increase in surface roughness favoured heat transfer in the quasi-turbulent flow regime. This is useful for rough tubes as the quasi-turbulent flow regime onsets early with regards to the Reynolds number in tubes with large roughnesses.
Computational investigation into jet impingement boiling on pin-fin surfaces
(University of Pretoria, 2023) Craig, K.J. (Kenneth); Meyer, Josua P.; Valluri, Prashant; u17057419@tuks.co.za; Ludick, Luwan
Thermal management of densely packed chips is critical for developing prevailing chips. For years, conventional air-cooling techniques have been utilised for numerous microsystems where fans and heat sinks were used in high-power computing systems due to their low cost and high reliability. Unfortunately, recent developments have exceeded the heat dissipation capability of these conventional techniques, leading to a shift towards liquid-to-vapour phase-change cooling techniques. Various multiphase cooling techniques have been reported in the literature. Over the last few decades, jet impingement has been shown to be an effective and attractive way to transfer energy from high heat flux components by the substantial amount of thermal energy transferred between the solid and the liquid. Surface enhancement techniques have also gained traction due to the increased average surface heat transfer coefficient and surface area by disrupting boundary layer growth and improving turbulent transport. This research combined jet impingement with phase change or boiling and surface area enhancement to improve heat transfer from a surface. Different boiling types arise in boiling jet impingement on pin-fin surfaces due to the various flow patterns caused by the pin-fin layout, thereby increasing the numerical complexity. All relevant numerical studies documented in the literature focused on boiling jet impingement on flat surfaces, whereas no studies were found on pin-fin surfaces. Therefore, conducting a well-documented numerical study of pin-fin surfaces formed an essential part of the current work. The complex flow patterns and boiling types between the pin fins had to be better understood before they could be widely implemented in electronic cooling applications. In this study, the heat transfer effect of pin-fin surfaces in boiling jet impingement was investigated using the RPI boiling model embedded in the Eulerian multiphase framework, which is an option in ANSYS Fluent. The numerical results of wall surface temperature in the stagnation area of the jet correlated well with experimental data reported in the literature. Not measured in the reference experiment, the pool-boiling areas caused by flow obstruction led to the cyclic behaviour of vapour formation and condensation. Furthermore, the cyclic behaviour was linked to the dry-out behaviour in the pool-boiling regions. An automatic mesh adaption tool allowed cell refinement at cells experiencing unrealistically high vapour velocities and increased numerical stability. The temperature distribution over the pin-fins formed cool regions corresponding to the flow-boiling regions; and warmer pockets corresponding to the pool-boiling regions. The turbulent kinetic energy followed the formation and condensation of the vapour column in the pool-boiling regions. The highest turbulent kinetic energy was produced as the liquid was forced into the staggered-facing pin-fins. These results highlighted the advantage of a validated numerical study to understand the detailed jet impingement boiling behaviour. Finally, a parametric study was conducted on a single jet impinging on a pin-fin surface to comprehend the effect of the inlet Reynolds number, pin-fin height, spacing and distribution on the heat transfer characteristics. The study of the inlet Reynolds number considered a lower and higher inlet velocity than for the validation case. An increase in jet velocity increased heat transfer at the stagnation region but had a limited effect on eliminating the dry-out areas at the outer regions of the domain. The study of pin-fin height and spacing suggested that heat transfer was mainly linked to surface augmentation. However, the decrease in pin-fin height allowed the liquid to spread to the outer regions of the domain and eliminated dry-out. The height and spacing study also suggested that the pressure drop over the domain was mainly linked to the stagnation pressure drop of the jet, while the pin-fin height and spacing had a negligible influence on the pressure drop for the parameter variation considered. The change in pin-fin configuration allowed the liquid to reach the outer regions of the domain while keeping the surface augmentation factor at a maximum. A star arrangement eliminated dry-out at 23.2 𝑊��/𝑐��𝑚��^2 and increased the average surface heat transfer. Therefore, the RPI boiling model, along with the use of a 𝑦��+ insensitive near-wall treatment model could accurately predict the heat transfer of a single jet boiling on pin-fin surfaces. The findings of the parametric study aligned well with expectations to eliminate dry-out at the outer regions of the domain while increasing the overall surface heat transfer. The CFD model suggested that researchers would have to measure local dry-out if pin-fins were used in boiling jet impingement. Furthermore, the influence of pin-fin shape, distributions and the working fluid needs further investigation to allow for heat transfer at higher heat fluxes, which align with modern-day electronic applications.
Incorporating sensor measurements using data assimilation and machine learning to improve the accuracy of thermal finite element models
(University of Pretoria, 2022) Heyns, P.S. (Philippus Stephanus); Wannenburg, Johann; karl.hellberg@tuks.co.za; Hellberg, Karl Gustav
Complex systems are commonly encountered in engineering. Such systems have a degree of unpredictability, which can lead to undesirable results. The digital twin concept has been proposed as a method to obtain more information about the system so that its behaviour can be better predicted. To construct a digital twin of a system, a high-fidelity model that predicts the evolution of the system over time is required. Classically, such high fidelity models would be either physics-based or data-driven, though both of these approaches have disadvantages that may make them unsuitable for application in a digital twin. A hybrid model is a combination of physics-based and data-driven models that seeks to exploit the advantages of both approaches, and is a promising candidate for producing a model that is suitable for application in a digital twin. This work investigates the training of hybrid models of real engineering systems. It considers systems that are dynamic and that are partially observed. Physics-based models of these systems are available in the form of partial differential equations that are solved numerically using the finite element method. To reduce the computational cost associated with such models, surrogate models are constructed. The construction of surrogate models involves a dimensionality reduction step in the form of proper orthogonal decomposition (POD), as well as a prediction step that involves the training of a data-driven model to predict the evolution of the POD coefficients over time. Hybrid models are then trained using the surrogate models as their physics-based component. The training of hybrid models takes place using a combination of data assimilation and machine learning. The machine learning-data assimilation (ML-DA) algorithm that is documented in the literature is used, together with two algorithms proposed in this work, called the data assimilation-observation (DA-O) and per-step DA-O algorithms. It is the nature of the application of this methodology of training hybrid models that allows this work to make a contribution relative to the published literature. Previous uses of data assimilation in the training of hybrid models perform investigations using simplified problems that allow direct use of the physics-based model in the hybrid model. Meanwhile, the increased computational demands of the real engineering problem considered in this work necessitates the use of a surrogate model of the physics-based component of the hybrid model. Surrogate models have been previously applied in hybrid models constructed to solve engineering problems. However, these approaches do not apply naturally to applications such as digital twins where observations are continuously available. The use of data assimilation in this work allows it to address this shortcoming. The proposed methodology of training hybrid models is evaluated using two case studies. The first case study involves a thermal simulation model of a small section of the freeboard of a process converter. Surrogate models of the simulation model are constructed using different data-driven function approximation techniques, such as Gaussian process regression, Support Vector Machines (SVMs) and neural networks. These surrogate models are then used to train hybrid models using simulated observations and the DA-O, per-step DA-O and ML-DA algorithms. When 30 of the 29077 nodes of the simulation model are observed, the per-step DA-O algorithm produces the best hybrid model in terms of a root mean square error (RMSE) metric calculated using the analysis states estimated during data assimilation. The ML-DA algorithm which performed next best is, however, easier to apply to different numbers of observed nodes and is potentially less sensitive to observation noise. The ML-DA algorithm is subsequently used to investigate the effect of using different numbers of observed nodes. While there is a benefit to using a greater number of observed nodes, the trained hybrid models still outperform the physics-based model when as few as two of the 29077 nodes of the simulation model are observed. These results indicate that the training of hybrid models for sparsely observed systems is feasible. The second case study considered in this work involves a thermal half model of the process converter freeboard for which actual sensor observations are available. Surrogate models are again constructed using different data-driven function approximation techniques, and these are used in the training of hybrid models. Only the ML-DA algorithm is now used for the training of hybrid models. Simulated sensor observations are used at first to understand whether improvements in predictive performance that hybrid models make relative to physics-based models on the observed nodes extend to larger subsets of the nodes of the system. It is found that when the performance of the hybrid models is evaluated in terms of the RMSE calculated using analysis states estimated during data assimilation, it is possible that improvements in performance are made on the observed nodes but not on larger subsets of the nodes. When the RMSE is instead calculated using predictions of the evolution of the system over 30 time steps, the performance of the hybrid models on the observed nodes correlates with their performance on larger subsets of the nodes. When real observations are used to train hybrid models, the trained hybrid models improve on the performance of physics-based models on the observed nodes in terms of the RMSE calculated using analysis states and in terms of the RMSE calculated using 30 time step predictions. The improvement in performance on the latter could indicate that this improvement on the observed nodes extends to larger subsets of nodes of the system. There are, however, other possible explanations for this improvement.
Experimental Investigation of Factors that Influence Bevameter Terrain Characterisation
(University of Pretoria, 2022) Els, P.S. (Pieter Schalk); Hamersma, Herman; u14349592@tuks.co.za; Kruger, Ray
The Bekker‐Wong soil‐wheel interaction model has been widely adopted in the terramechanics field. This model traditionally requires the soil to be characterised using a Bevameter, which entails performing in situ plate sinkage and shear stress tests. Bevameter soil characterisation is not a standardised test procedure, and the test setup may influence the identified soil model parameters. This study investigates the influence of the following five factors for partially saturated sandy soil: i) soil preparation on pressure‐sinkage, ii) soil preparation on shear stress, iii) torsional vs. translational shear mechanism, iv) shear contact area, and v) shear velocity. Literature indicates in situ soil mechanical properties exhibit stochastic behaviour; however, the uncertainty of the identified soil parameters is rarely taken into consideration. This study employs the Bayesian statistical framework for probabilistic parameter estimation and formal hypothesis testing. The results indicated that the influence of soil preparation on pressure‐sinkage response is substantial, exhibiting an order o magnitude influence. The influence of soil preparation on shear tests is notable, but less significant. The shear mechanism, shear contact area and shear velocity all exhibited a statistically significant influence (Bayes Factor >10) with a maximum absolute shear stress difference of 18%, 20% and 10%, respectively. Moreover, depending on the test setup configuration and data processing decisions, the estimated internal soil friction angles ranged from 16.5 to 37.5 degrees for the same soil. The findings are expected to have significant implications for the prediction of vehicle drawbar pull using the Bekker‐Wong model. Further investigation into which Bevameter test configuration is more representative of the shear stress‐displacement curve of an actual wheel is recommended.
Utilizing available time-varying operating condition information in learning-based methods for fault diagnostics of rotating machinery
(University of Pretoria, 2022) Heyns, P.S. (Philippus Stephanus); Wilke, Daniel Nicolas; abelnortje@gmail.com; Nortje, Abel Hermanus
Historical failure datasets for critical assets are rarely available. This makes it difficult to integrate condition monitoring techniques found in literature into industry, as these techniques are often either equipment-specific or highly dependent on the availability of a historical failure database. Recent research in the vibration-based condition monitoring field addressed this problem, by focusing on using unsupervised latent variable models that require only healthy data for training. By learning the distribution of the healthy data, faults can be detected and tracked when new data starts to deviate from the learned healthy distribution. The complexity of the healthy data distribution drastically increases when machines are operated at varying rotational speeds, resulting in smearing in the frequency spectrum, amplitude modulation, and heteroscedastic noise in the data. This makes it more difficult to accurately model the healthy data distribution. Signal processing methods that incorporate available shaft speed and phase information, have been applied extensively to vibration data from varying speed conditions, to make it easier to analyse. Order tracking has been performed to convert the signals to the angle domain, and regression methods have been applied to normalize the effect of amplitude modulation. This work systematically investigates to what extent the availability of operating condition information can help to simplify the learning process for latent variable models in a semi-supervised setting. For a baseline, the operating conditions are mapped to the vibration data in a supervised setting. This is done to see how much of the variance in the vibration data can be explained by the operating conditions, and to highlight the importance of using latent variable models when the data is influenced by generative factors that cannot easily be measured or extracted from the data. Unsupervised models (that take raw unprocessed vibration data as inputs) are compared with semi-supervised models that incorporate operating condition information during data pre-processing (order tracking), and during modelling (learning distributions conditioned on the associated operating conditions). Two latent variable frameworks that are often used for anomaly detection are investigated: Principal Component Analysis (PCA) and Variational Auto-Encoders (VAEs). This work highlights practical limitations in the PCA and VAE frameworks for condition monitoring in an unsupervised setting. It is shown that PCA can't accurately capture the heteroscedastic noise in the data, since it is formulated by assuming constant variance across the whole dataset. It is proposed that the correct variance can be learned using linear regression between either the latent space representations (unsupervised) or the available operating condition information (semi-supervised). Using the available operating condition information yielded the most robust results. For the VAE framework, it is shown that the latent space prior commonly used to facilitate disentanglement in $\beta$-VAEs, can in this case lead to posterior collapse, which leads to poor discrepancy analysis results. This occurs because of circular structures present in the underlying latent space manifold, caused by the periodic nature of vibration signals. The isotropic variance Gaussian usually used as a latent space prior is not well suited to capture these manifolds. This challenge is overcome by conditioning the prior on the associated operating condition information. This work also provides clear insights into how linear and nonlinear models capture these distributions differently with lower and higher dimensional latent spaces, which improves the interpretability of the latent space and the associated model performance. It is shown how the size of the latent space affects whether the damage is detected in the latent space or the reconstruction space. In addition, the latent space representations can successfully be monitored for anomalies when conditioned on the available operating condition information. The latent variable models' performances are compared to traditional signal processing approaches (envelope analysis). The semi-supervised models return promising results on a dataset that is known to be difficult to analyse with traditional signal processing methods.

Browse

Recent Submissions