Minerva Elements Records
Permanent URI for this collection
Search Results
1 - 10 of 27
-
ItemWheat grain protein content assessment via plant traits retrieved from airborne hyperspectral and satellite remote sensing imagery( 2023-11)Wheat (Triticum spp.) is crucial to food security. The source of a major proportion of humans’ total dietary carbohydrates and protein, it is among the world’s most widely grown crops and receives concomitantly large quantities of nitrogen (N) fertiliser. Wheat grain protein content (GPC; %) is a key to food quality, determining the baking quality of bread, the cooking quality of pasta, and the nutritional value of food products. For these reasons, wheat is classified and growers are typically paid predominantly on the basis of GPC, setting its farm income value. Global population growth encourages a justified focus on increasing yields. However, because grain proteins are diluted by carbohydrate (CHO) additions in the latter part of growing seasons, GPC is in an inverse relationship with yield: Improved yields are attended by the risk of reducing GPC. Moreover, GPC is influenced by interacting genetic and agronomic factors, soil properties and weather conditions that affect crops’ physiological status and stress levels and can therefore exhibit great spatial variability. Of the vast quantities of nitrogen (N) applied to wheat crops, a variable but substantial proportion is lost, inducing environmental damage and financial costs, which should be averted. Accurate GPC prediction could reduce N losses, assist in crop management decisions, and improve farm incomes. Nitrogen is central to proteins and can be strategically supplied to crops in order to achieve GPC benchmarks a precision agriculture (PA) approach. In such scenarios, estimates of GPC potential in advance of harvest could guide fertiliser dosing, improving fertiliser efficiency and potentially reducing costs. In contrast, where strategic fertiliser applications are not favoured, crop management could benefit from prior knowledge through strategic harvesting aimed at maximising payments per unit of grain at receival. However, GPC is a complex variable, influenced by multiple plant traits, themselves affected by soil and moisture conditions and whose effects change through the growing season. While remote sensing (RS) is likely the only practicable method of estimating GPC during seasons, and shows potential, prediction is complex and success has been limited. To make progress, it is necessary to more robustly identify imaging spectroscopy-based physiological traits closely associated with GPC. Traits with known physiological links to GPC, and which can be retrieved from imaging spectroscopy, include leaf area index (LAI) and chlorophyll (Ca+b). Further inspection of these and other RS traits may advance research relevant to PA. The inverse relationship of GPC to CHO assimilation permits the hypothesis that indicators of plant stress can improve GPC estimation. Such stress indicators, including the pigments anthocyanins and carotenoids, can be accurately retrieved along with other biophysical and biochemical quantities from hyperspectral (HS) remote sensing but their relationship to GPC had yet to be tested. Solar-induced fluorescence (SIF), emitted from the photosystems in proportion to instantaneous photosynthetic rate, was also untested as a GPC predictor. Moreover, in addition to the traits themselves, retrieval of plant traits by inversion of radiative transfer models (RTM) also remained untried for GPC estimation. Finally, the crop water stress indicator (CWSI), a proxy for evapotranspiration and hence carbon assimilation, should also show an association with GPC. Because a large majority of GPC studies have been conducted exclusively in the context of experimental plots, it is appropriate to extend research into the commercial cropping domain, populated to date by only two previous studies. This expansion is facilitated by the recent advent of spatially explicit GPC monitoring during crop harvests. While lacking the ultra-high spectral and spatial resolution of airborne HS sensing, satellite RS, in particular the Sentinel-2 (S2) platforms, possess advantages with respect to broadacre PA. These include a focus on reflectance bands adapted to vegetation sensing, appropriate spatial resolution, and frequent return times. This thesis presents results from piloted HS flights and ground campaigns at two dryland field experiments with divergent water supply and wide-ranging N fertiliser treatments, and from HS flights over 17 commercial fields planted to either bread (T. aestivum L.) or durum (T. turgidum subsp. durum (Desf.) Husn.) wheat, across two years in the southern Australian wheat belt. Imagery was acquired with airborne hyperspectral and thermal sensors, with spatial resolutions of approx. 0.3 m and 0.5 m for experimental plots and 1 m / 1.7 m in commercial fields. Leaf clip measurements, leaf and grain samples were collected from plots and through a transect in one field. In commercial fields, ~40,000 records obtained from harvester-mounted protein monitors. CWSI, SIF, vegetation indices and PRO4SAIL RTM inverted parameters were retrieved for each plot and GPC record location. Sentinel-2 (S2) timeseries (TS) were subsequently acquired for > 6,000 ha of commercial dryland wheat fields, inclusive of those included in HS campaigns, also in south-east Australia and through two consecutive years of dissimilar rainfall. In this case, growers provided ~92,000 GPC data points from harvester-mounted protein monitors. For each, Ca+b, leaf dry matter, leaf water content (Cw) and LAI were retrieved from the S2 images by radiative transfer model inversion. A gradient boosted machine learning algorithm was applied to analyse these traits’ importance to GPC and to predict GPC in 30% of samples unseen by the algorithm in training. From HS analyses, the photochemical reflectance index (PRI) related to xanthophyll pigments was consistently associated with GPC at both leaf and canopy scale in the plots and transect. In the commercial crops, a gradient boosted machine learning algorithm (GBM) ranked CWSI as the strongest indicator of GPC under severe water stress, while SIF, PRI and inverted biochemical constituents anthocyanins and carotenoids were consistently important under more moderate growing conditions. Structural parameters inverted from HS were not prominent except under severe drought when CWSI was omitted from models. Statistically significant results were obtained for GPC estimation in unseen samples, with best relationships between predicted and observed GPC of R2 = 0.80 and RMSE = 0.62 in a model built with thermal and physiological traits obtained from the HS and thermal imagery. Trait importance in S2 analyses was consistent with that seen from HS, in that the rankings of physiological, structural and water stress indicators were aligned: severe drought increased the importance of water stress measures relative to other traits, but in milder conditions physiological traits were emphasised. Airborne SIF added substantially to model skill from single-image S2, particularly in moderate conditions. While coefficients of determination varied substantially according to water stress, error metrics invariably sat within a tight range, under 1 % GPC. Overall, these predictive modelling results, obtained at within-field scale and in challenging conditions, place the current study among others in the same research domain, most of which consider either plot or regional scales. The strongest relationships between predicted and observed GPC (R2 = 0.86, RMSE = 0.56 %), in a model built from five S2 images across a season, were better than those from single-date hyperspectral (HS). In severe water stress, LAI was the main predictor of GPC early in the season, but this switched to Cw later. In milder conditions, importance was more evenly distributed both through the years and between traits, and predictive skill was lower. S2 TS had a clear accuracy advantage over single-date S2, and approached that of HS, especially in benign conditions, emphasising its previously unexplored potential for large-scale GPC monitoring. The methods developed are a novel contribution and can be proposed as a useful basis for future research.
-
ItemUnderstanding protein variants with high-throughput mutagenesis and machine learning( 2023-10)Genetic variations in protein-coding genes may cause amino acid substitutions in the matured proteins. These variants can potentially change the properties and functions of a protein. To evaluate the effects of these protein variants, multiple experimental and computational approaches have been utilised. Within these approaches, deep mutational scanning (DMS), a recently developed high-throughput mutagenesis method, enables the measurement of thousands of protein variant effects in a single experiment. To fully investigate the rich information in DMS results and have a better understanding of protein variant effects, here, I leveraged machine learning algorithms to build advanced computational models for DMS data. First, I reviewed that there are missing variant effect data in most DMS results, and I developed imputation models to fill in the missing values. I started by investigating the correlations between the variant effects measured within a DMS experiment and used these correlations to build imputation models. To understand the strengths and weaknesses of these models, I benchmarked them with previously published DMS imputation methods. At the end of this study, I built an ensemble imputation model by combining these novel and previously published methods to further improve the imputation accuracy. Many of the state-of-the-art variant effect predictors are built with DMS data, and I then managed to improve these predictors by further integrating variant effect data from alanine scanning (AS), a low-throughput mutagenesis approach. In this study, I established a rule-based classification tree to evaluate the compatibility between DMS and AS studies according to the similarity of their experimental assays. I showed that an improved variant effect predictor could be built only by modelling with high compatibility DMS and AS data. Finally, experimental measurements of protein variant effects may conflate protein stability and function. Here, I explored this relationship using DMS-measured variant effects and computed variant stability. I demonstrated that the correlation between variant effect and stability data differed on distinct protein regions and properties measured. Analysing these data with a dimensional reduction algorithm, I was able to automatically distinguish protein residues with different scales of fitness–stability association. Further investigation showed that this approach might be applied to discover protein functional sites and explain the mechanisms of loss-of-function variants.
-
ItemAddressing domain shift in deeply-learned jet tagging at the LHC( 2023-09)Over the last fifteen years, deep learning has emerged as an extremely powerful tool for exploiting large datasets. At the Large Hadron Collider, which has been in operation over the same time span, an important use case is to identify the initiating particles of hadronic jets. Due to the complexity of the radiation patterns within jets, neural network-based classifiers are able to out-perform traditional techniques for jet tagging. While these approaches are powerful, neural networks must be applied carefully to avoid performance losses in the presence of domain shift—where the data on which a model is evaluated follows different statistics to the training dataset. This thesis presents studies of possible strategies to mitigate domain shift in the application of deep learning to jet tagging. Firstly, we develop a deep generative model that can separately learn the distribution of quark and gluon jets from mixed samples. Building on the jet topics framework, this model provides the ability to sample quark and gluon jets in high dimension without taking input from Monte Carlo simulations. We demonstrate the advantage of the model over a conventional approach in terms of estimating the performance of a quark/gluon classifier on experimental data. One can also use likelihoods under the model to perform classification that is robust to outliers. We go on to evaluate fully- and weakly-supervised classifiers using real datasets collected at the CMS experiment. Two measurements of the quark/gluon mixture proportions of the datasets are made under different assumptions. Compared to the predictions based on simulation, we either over- or under-estimate the quark fractions of each sample depending on which assumption is made. When estimating the discrimination power of the classifiers in real data we find that while the absolute performance depends on the choice of fractions, the rankings among the models are stable. In particular, weakly-supervised models trained on real jets outperform both simulation-trained models. Our generative networks yield competitive classification and provide a better model for the quark and gluon jet topic distributions in data than the simulation. Finally, we investigate the performance of a number of methods for training mass-generalised jet taggers, with a focus on algorithms that leverage meta-learning. We study the discrimination of jets from boosted Z' bosons against a QCD background and evaluate the networks' performance at masses distant from those used in training. We find that a simple data augmentation strategy that standardises the angular scale of jets with different masses is sufficient to produce strong generalisation. The meta-learning algorithms provide only a small improvement in generalisation when combined with this augmentation.
-
ItemA mathematical, digital and experimental investigation of the freezing, thawing and storage of Mozzarella cheese( 2023-09)Mozzarella cheese is a popular cheese product, resulting in widespread markets and demand for cheese distribution. The limited shelf-life of this cheese makes long-distance shipping more challenging. This short shelf-life is due to relatively high moisture content, which promotes bacterial and enzymatic activity that accelerates proteolysis, impacting functional attributes, such as stretchability and meltability. Managing proteolysis is therefore important for maintaining Mozzarella cheese quality throughout storage and shelf-life. Freezing has been proposed as a potential way to retain the functional shelf-life of Mozzarella cheese by arresting proteolysis. Yet there are limited tools available for prediction of freezing and thawing times as a function of composition and processing variables. Specifically, thawing of Mozzarella cheese has not been investigated in detail and there is lack of a model that can predict thawing of Mozzarella cheese. This thesis aimed to develop robust models for prediction of freezing and thawing processes for Mozzarella cheese. To do this, six different Mozzarella cheese blocks, differing significantly in block size and composition, were assessed using freezing and thawing experiments. Temperature profiles of freezing and thawing were obtained for these different Mozzarella cheese samples. The enthalpy method was then used to develop a novel robust heat and mass transfer model for freezing and thawing in Mozzarella cheese. The presence of salt (NaCl) affected both freezing and thawing, depressing the freezing point. The freezing point depression was predicted for Mozzarella cheese with different salt contents, providing useful information that was not available elsewhere in the literature. Salt content had a significant impact on freezing, with a decrease in salt from 1.34% w/w to 0.07% altering the temperature of phase change from ~-4.5 C to -3 C. Additionally, simulations showed salt migration was restricted to the first ~1-2 centimeters from the surface during freezing, with a slight increase of 8-10% salt in free moisture at the block center. Another aim was to develop a response surface methodology (RSM) model based on the simulated data to create a computationally efficient tool for predicting freezing and thawing times for single cheese blocks, using a range of industrially relevant conditions. The proposed models can increase the current understanding of impact of cheese composition and size, as well as storage settings in the case of thawing, on freezing and thawing of Mozzarella cheese. The approach presented here can also be used to determine freezing and thawing times, which are important information for manufacturers in the design of storage facilities, the choice of transportation and process optimization to improve Mozzarella shelf-life and quality during storage. For instance, the interaction effects provided by the RSM model can be used to find the optimum thawing conditions including the air velocity, thawing temperature and thawing time for the frozen cheeses. The next aim was to predict the level of proteolysis in Mozzarella cheese during the storage at 4 C, using data available in the literature from extensive prior studies on Mozzarella cheese proteolysis. Considering the complex nature of cheese proteolysis (i.e., presence of correlated variables, such as cheese composition, and nonlinear relationships between these variables and proteolysis levels) there are no generalizable tools that can predict proteolysis accurately. The predictive performance of a multilinear regression model and three different machine learning techniques, namely, gradient boosting, support vector regression and random forest, were explored using various input features including chemical composition, storage time and well-known manufacturing factors, such as milling pH, acidification type and coagulating enzyme. Machine learning techniques outperformed the multilinear regression due to their ability to handle the presence of correlated input features and nonlinearity between input features and the level of proteolysis. The collected data from the literature had not been previously used for predicting cheese proteolysis and the performance of machine learning techniques for this purpose has not yet been assessed. The proposed approach was also applied to predict the proteolysis in Cheddar, which shares some similar manufacturing steps with Mozzarella cheese, using a relevant set of data collected from the literature. The gradient boosting method could more accurately predict proteolysis than the other two machine learning algorithms for both Mozzarella (R2 = 92%) and Cheddar (R2 = 97%) cheese, possibly due to the iterative sequential training of the gradient boosting algorithm. Storage time was the most important input feature for both cheese types, followed by coagulating enzyme concentration and calcium content for Mozzarella cheese as well as fat and moisture content for Cheddar cheese. Manufacturers can use this information for distribution of Mozzarella cheese within an optimal window of functionality. Information on proteolysis is also important for Cheddar cheese, as it can be correlated with Cheddars quality attributes, such as flavor and texture. Finally, freezing and thawing of Mozzarella cheese was investigated under industrial conditions using a pallet containing 96 cheese blocks, each weighing 10 kg. This was the first time that effects of freezing and thawing processes under industrial conditions were evaluated on properties of Mozzarella cheese during a storage period of five months post-thaw. Additionally, the effects of the placement of cheese blocks at inner and outer positions of the pallet were investigated, as placement could lead to different properties during storage, which has not previously been examined. Heat and mass transfer simulations predicted temperature profiles at different pallet locations, achieving a maximum root mean square error of 3.6 C for both freezing and thawing processes. Significant differences were observed in the cooling rates between blocks located at inner (0.7 C day-1) and outer (0.87 C day-1) positions during freezing, as well as in the warming rates for inner (0.81 C day-1) and outer (5.9 C day-1) locations during thawing. Freezing simulations showed salt migration during the freezing of the samples, with more migration observed on the sides of the outer blocks. While the frozen-thawed samples had slightly lower urea-PAGE intact casein levels after 5 months of storage at 4 C compared to the control unfrozen sample, the confocal microstructure, expressible serum obtained by centrifugation and select functional properties including the hardness, stretchability and meltability measured by the transition temperature method were only different between frozen-thawed and control samples during the first month. Despite varied heat transfer rates, pallet block locations during freezing and thawing did not significantly affect most properties, ensuring that freezing and thawing pallet-stacked products under industrial conditions of this study led to a consistent good quality (i.e., stretchability measured via texture profile analysis and meltability measured via rheological analysis) after thawing and during the subsequent storage. These results can help manufacturers make informed decisions about inventory management and distribution of their products. Additionally, a gradient boosting algorithm, which was shown to outperform other algorithms in earlier sections, successfully predicted proteolysis levels in the frozen-thawed samples (R2 = 88%), which can be potentially linked to the functionality of industrially frozen-thawed products. This thesis addressed the challenges posed by the short shelf-life of Mozzarella cheese and freezing and thawing processes. Well-established principles, previously employed in other disciplines, have been applied to offer innovative insights into the freezing and thawing processes of Mozzarella cheese as well as its proteolysis. A series of recommendations were also made for future studies in the field, including steps that could be taken to increase the accuracy and the applicability of the current simulations and models to a wider range of other dairy products. This thesis not only addressed the challenges in the dairy manufacturing industry but also improved research on dairy storage and quality control by introducing a range of new modelling approaches.
-
ItemA predictive neural model of visual information processing( 2023-08)Recent understanding of how the brain processes sensory input has moved away from understanding sensory processing as just being the passive processing of input to a framework that employs a more active role of higher-level expectations in sensory processing. One such theory is predictive coding in which the brain generates a prediction of the sensory input that it will receive and compares this prediction with the actual sensory input. Because the transmission of visual information from the eyes to the brain takes time, for the brain to accurately respond to the real-time location of a moving object, the prediction mechanism has to take into account the change in object’s location during the period of transmission latency. Understanding such temporal prediction mechanism will extend our understanding of the way by which the brain actively interacts with the environment we live in. This research project investigated the predictive signal transmission pathways of the mammalian visual system and focused on the early stages of the visual pathway, including the retina, the Lateral Geniculate Nucleus (LGN) and the primary visual cortex (V1). These structures play critical roles in visual signal gathering and integration. Mathematical and computational models were constructed based on predictive coding strategies and spike-based neural coding principles, where neurons with specific firing timings are arranged into hierarchical areas, and upper areas predict the neuronal behaviours of lower areas that receive sensory stimuli. The first goal of the project is to investigate the encoding of visual information in precise neuron spike timings and neuronal interactions, because the temporal prediction mechanism involves small time scales and detailed object motions. We intend to show that results obtained via spike-based neural principles, which involves cumulative computations in small time scales, do not contradict with the results from classical rate-based neural networks that operate based on longer time scales, and results from physiological recordings. The second goal is to investigate the mechanism by which temporal prediction can be achieved using the spike-based neural network, given moving input stimuli. Through the project, we validated that a predictive coding network can be built based on spike-based neural principles, and it has the potential to encode moving stimuli with less error compared with rate-based approaches. Based on the model developed, the next step is to study the specific mechanism by which the alignment between the real and the perceived locations of a moving object can be achieved, i.e., a mechanism that compensates for the signal transmission delay from the eyes to the brain. Outcomes of the research are expected to advance our understanding about human visual system and provide new insights into the development of neural implants, prostheses and machine learning algorithms. The principles investigated are hypothesised to apply throughout the cerebral cortex. Consequently, the results are anticipated to have application to the processing of other modes of sensory stimulus, such as auditory and olfactory inputs, applications can also be expanded to the research areas of memory, motor control, cognition and decision making.
-
ItemMeasuring lower-back injury risk in repetitive stooped work: a population-data driven approach( 2023-05)More than 80% of people will experience back pain in their lifetime. Recent data indicates that lower back pain is the leading cause of disability globally, and the problem is getting worse, not better. Lower back injuries leading to disabling pain are especially common in stooped work occupations, which are characterised by workers spending long periods of time with a fully flexed torso and mostly straight legs. Dysfunction in neuromuscular control could predict injury, however as these kinds of injuries take a long time to develop, they are almost never observed alongside this data. Advances in wearable sensors allow for the collection of this data outside the lab, but due to the lack of labelled injury data, there's no easy way to connect these measurements to injury. Additionally, neuromuscular control is task-specific, which means any dysfunction is also task-specific. So different measurements may be required depending on the task, and when it comes to real occupations it's not clear which measurements are the most useful indicators of injury risk. In this work, I propose a novel unsupervised feature selection framework to identify key indicators of lower back injury in a task specific way. It is noted that the vast majority of lower back injuries occur over time via repetitive stress. The lower back injury risk continually 'gets worse', and workers doing the same tasks are exposed to similar stresses. The proposed framework analyses which features contribute to this underlying shared monotonic trend of injury development. The framework is evaluated with data collected from Australian sheep shearers, who experience extreme rates of lower back injury. Features identified by the framework are consistent with the literature on lower back pain, and outperform commonly used unsupervised feature selection techniques when evaluated using a pseudo-ground-truth. Utilising the identified features, a prototype wearable sensor was developed to obtain these measurements from sheep shearers in real-time.
-
ItemOptimising Preoperative Decision-Making in Total Knee Arthroplasty Using a Machine Learning Approach: Development, internal validation, and clinical acceptability evaluation of a clinician-informed machine learning model for the prediction of 30-day readmission following total knee arthroplasty( 2023-06)Background: Total knee arthroplasty is an effective treatment for advanced osteoarthritis of the knee joint, leading to reduced pain, improved function, and better quality of life for affected patients. Following a total knee arthroplasty (TKA) procedure, 30-day readmissions indicate a suboptimal postoperative course which negatively impacts upon the patient’s recovery and poses a significant burden to the healthcare system. Machine learning techniques can be used to predict readmission risk for individual patients and therefore can be implemented in tools to support shared clinical decision-making between patient and orthopaedic surgeon. Objectives: 1. To utilise the experience and expertise of clinicians involved in the care of TKA patients in the identification and appraisal of risk factors for 30-day day readmission. 2. To develop a statistical model to predict 30-day readmission in TKA patients, utilising machine learning techniques and clinical insight for use in shared clinical decision-making. 3. To evaluate the performance of clinicians regularly involved in the care of TKA patients on predicting 30-day readmission following TKA for individual patients then compare the predictive performance of a risk prediction model with that of clinicians. 4. To explore the understanding of TKA patients regarding what AI is and what are its perceived benefits and potential pitfalls in the context of shared clinical decision-making. Methods: Mixed methods approach involving five stages, adapted from literature pertaining to the development and implementation of complex interventions. Stakeholder involvement was utilised throughout the project to engage clinicians, hospital administrative staff, and patients themselves. Patient involvement was embedded throughout the project by means of a research buddy program, and this was detailed in a perspective piece included in the Methods. Stage 1 involved risk factor identification and evaluation, comprising two stages: first, a narrative review, systematic review protocol, and systematic review and meta-analysis on patient-related risk factors for 30-day readmission following TKA; second, a modified Delphi survey and focus group study based on systematic review findings. Stage 2 involved dataset acquisition and description, comprising a cohort profile for the institutional arthroplasty registry and a narrative description of the process of accessing and utilising hospital administrative data. Stage 3 involved a multivariable predictive model development study based utilising machine learning techniques as well as clinical insight gained in Stage 1. Stage 4 involved clinical acceptability evaluation in the form of a computer vs clinician comparison study. Finally, Stage 5 involved clinical acceptability evaluation, capturing the patient perspective in a qualitative semi-structured interview study. Findings: Clinicians provided insight into the complexity of predicting readmission on account of the diverse range of risk factors. Together with machine learning and statistical techniques, this insight was applied to arthroplasty registry and hospital administrative data to develop a predictive model which i) outperformed clinicians’ predictive capabilities and ii) was adequately calibrated to facilitate implementation in the clinical setting. The qualitative study, co-designed with a consumer advocate, found that TKA patients were open to the use of AI in shared clinical decision-making, and these findings were contextualised in prior literature to generate recommendations for future implementation. Conclusions: This thesis demonstrated the development of a bespoke readmission risk prediction model for TKA patients in a process involving broad stakeholder involvement in recognition of the intrinsic value of involving stakeholders in research and development initiatives that impact upon them, and in recognition of the responsibility of researchers to do so. This process primed the model for future implementation to enhance shared clinical decision-making.
-
ItemRELIABILITY-BASED ASSESSMENT OF CONCRETE TUNNEL LINING EXPOSED TO FIRE( 2023-02)The use of underground space has been increasing and gaining importance in the development of the cities and urbanized areas in order to reduce surface congestion and environmental related issues. Because of this rapid development, road tunneling has become one of the major construction processes all around the world. However, the consequences in an event of a road tunnel fire can be extremely serious than a fire in an open road. This can be a far more serious issue in future due to the greater demand for more and longer tunnels as well as the increasing traffic densities. The requirements for fire safety critically affect the overall design of concrete tunnel lining. Fire design guidelines which are based on prescriptive design guidelines developed few decades ago hinder the effective use of concrete tunnel lining as they do not account for uncertainties involved in the design. Since then, material properties of concrete have significantly changed, and construction methods have evolved considerably. Therefore, new research is needed to assess the structural fire performance of concrete tunnel lining and provide a research base to improve the current fire safety design guidelines by introducing precise reliability-based assessment for concrete tunnel lining exposed to fire. Generally, tunnel fires are complex in nature. Because of that proper understanding about fire behaviour in concrete tunnel lining is vital but only a limited number of full-scale experiments have been carried out to evaluate fire performance of concrete tunnel lining in fire. Since fire tests are often time-consuming and costly, it is important to investigate the use of numerical approaches to assess the fire behaviour. This research, therefore, carried out a parametric study using Fire Dynamic simulator (FDS) modelling approach to study the temperature distribution of the tunnel lining surface. Further, the effect on several factors such as tunnel geometry, fire source and the location of fire were also explored using the developed models. Also, this study developed a finite element (FE) modelling approach to conduct a heat transfer analysis of tunnel lining. The developed heat transfer model was validated using the results from the full-scale experimental investigation conducted for the Melbourne Metro Tunnel lining specimens at the fire testing facility of Victoria University. In general, the validation of heat transfer analysis of tunnel lining achieved reasonable accuracy compared to the experimental results. Fire-induced spalling in concrete is a serious issue in tunnel lining design because it can reduce the load bearing capacity of the tunnel and the cross-section area of the tunnel lining. The adverse consequences of concrete spalling can cause serious damage to the tunnel lining or even failure occasionally. Hence, concrete spalling at elevated temperatures particularly explosive spalling must be properly assessed by considering it as a crucial factor for fire resistance in concrete tunnel lining designs. In the last several years, there has been a surge of scientific studies aimed at explaining why concrete spalls when exposed to fire. Despite these attempts, a current evaluation method that can reliably forecast the average depth of spalling of concrete tunnel lining has not yet been developed, and a comprehensive analysis of this phenomena has not been completed. Many areas of structural engineering have benefited from the use of machine learning, but no one has yet attempted to use it to predict the spalling depth of concrete tunnel lining. Most sophisticated techniques in machine learning such as ensemble learning approaches have not been adopted. This thesis also addressed this issue by developing a database of 415 spalling test results to provide predictions about the spalling depth of concrete tunnel lining using ensemble learning approaches such as Random Forest (RF), Categorical gradient boosting algorithm (Catboost), Light gradient boosting algorithm (LightGBM) and Extreme gradient boosting algorithm (XGBoost). This research developed a novel machine learning-based framework to predict the spalling behaviour in tunnel lining exposed to fire. XGBoost ML model demonstrated high accuracy in predicting spalling depth in concrete tunnel lining. XGBoost showed the best performance against the testing set which indicates less overfitting against the training set. A parametric investigation and a feature importance analysis were conducted which highlighted the most influential parameters on the spalling depth. Based on the feature importance analysis, the amount of polypropylene fibre is found to be the most important input variable. The interest in structural fire safety of tunnels has increased as a result of disastrous tunnel fire events occurred around the world. Due to ignorance of any uncertainties involved, the structural fire safety levels providing from prescriptive designs are considered to be inconsistent. This might provide unexpected results, leading to erroneous conclusions about the reliability of a structure. As a result, the requirements of fire safety are transitioning from a prescriptive design standard to a performance-based design standard. Though the area of structural reliability is gaining interest among research community, there is a lack of precise reliability-based assessment in particular for tunnel lining exposed to elevated temperatures with regard to the bending moment capacity. A probabilistic computation approach was used in this study to evaluate the level of safety provided by concrete tunnel linings that are subjected to bending and exposed to standard fire (ISO 834 curve). This probabilistic evaluation enabled a detailed examination of the relationship between reliability index with increased temperature as well as a study of the effect of concrete spalling on reliability. Finally, a reliability framework for concrete spalling was developed by combining crude Monte Carlo simulation with the First Order Reliability Method. In general, a reduction of reliability index was observed as the temperature increases for concrete tunnel lining without considering the spalling effect. A higher reduction of reliability index was observed as the fire exposure time increases for concrete tunnel lining with consideration of the spalling effect compared to the reliability index reduction without considering the spalling effect. This observation further highlights the effect of spalling on reliability indices of tunnel lining is significant and must be considered for tunnel fire designs.
-
ItemPathogenicity and early detection of Verticillium dahliae and V. albo-atrum in potatoes in Australia( 2023-04)Verticillium wilt is a vascular wilt disease of potatoes primarily caused by the soilborne ascomycete fungi, Verticillium dahliae Kleb. and V. albo-atrum Reinke & Berthold. Verticillium wilts are amongst the most economically devastating crop diseases worldwide as they detrimentally impact crop yield and quality as well as exponentially increase pre- and post- planting management costs. Verticillium dahliae is widespread across potato growing fields in south-eastern Australia whilst, to date, V. albo-atrum has only been detected in Victoria and Tasmania. Verticillium wilt is an exceedingly challenging disease to control due to the broad host range of the causal pathogens and their long-term persistence in soil in the absence of a susceptible host. There are no curative measures available once Verticillium spp. have invaded plant tissues. As such, early detection is a critical component of an effective management strategy for Verticillium wilt, as it can facilitate growers to take actions that can prevent or significantly restrict disease expansion. Non-destructive early detection methods may involve machine- based technologies coupled with machine learning modelling, but initial validation trials are yet to be conducted in potato in controlled conditions. Glasshouse trials were conducted to assess early (2 to 5 weeks after inoculation) infection and colonisation by V. dahliae using visual symptom assessment, tissue colonisation assays, and RT-qPCR DNA quantification of the amount of V. dahliae present in the infected stems. Visible symptoms of Verticillium wilt first occurred in plants of the moderately resistant cultivar Denali and the susceptible cultivar Russet Burbank at two weeks after infection, and tissue colonisation assays showed that the lower stems, crowns, and upper root tissues of plants of both cultivars were extensively colonised within these two weeks. RT-qPCR also showed the quantity of V. dahliae DNA generally increased in the lower stems and crown tissue each week after infection in Russet Burbank plants. This confirmed that V. dahliae was pathogenic and highly virulent in potato plants within two weeks of infection. Furthermore, this study showed that quantitative disease resistance or tolerance is likely to occur in Denali and further studies are required to confirm the type of resistance. Visual symptom severity progressed more slowly in Denali and by the conclusion of the trials, the mean symptom severity was 39% and 75% for Denali and Russet Burbank respectively. Despite this, both cultivars were observed to be extensively colonised within two weeks. During Trial 1, the mean quantity of V. dahliae DNA was significantly different (P = 0.004) and was 0.56 pg/ul and 2.23 pg/ul in Denali and Russet Burbank respectively. Conversely, during Trial 2, the mean quantity of V. dahliae DNA was not significantly different (P = 0.106) and was 1.51 pg/ul and 2.25 pg/ul respectively. Glasshouse trials to assess early infection and colonisation by V. albo-atrum were also conducted using visual symptom assessment, tissue colonisation assays, plant physiological measurements, and visible near-infrared (Vis-NIR) spectroscopy. Symptoms of Verticillium wilt were not observed during at any time during the trials. Although roots, stolons, and crowns were observed to be extensively colonised in inoculated plants, stem colonisation was not observed in 69 out of 140 plants. Multifactorial analyses showed that infected and uninfected plants could be visually grouped at two weeks after inoculation using Vis-NIR. Vis- NIR measurements alongside stem, root, and crown infection, were the best indicators of infection within plants. These results showed that V. albo-atrum is pathogenic in potato but has low virulence within five weeks of infection. These trials justified the merit of conducting further investigations into the use of NIR spectroscopy and NIR modelling for the early detection of infection of potatoes by the more pathogenic V. dahliae. Two artificial neural network (ANN) models were developed using the raw absorbance values within the 1596-2396 nm light spectral range as inputs to predict photosynthetic rate, transpiration rate, and stomatal conductance (Model 1), as well as predict whether Denali and Russet Burbank plants were infected or not infected by V. dahliae (Model 2). The results showed high accuracy for Model 1 (R = 0.89 using all samples from Trial 1) with a deployment accuracy of R = 0.76 using all samples from Trial 2. High accuracy was also obtained for Model 2 using all samples from Trial 1 (94%) with a deployment accuracy of 84% using samples from Trial 2. The results showed that V. dahliae presence could correctly be identified in potato plants two days after infection without any visible symptoms present in plants. These novel ANN model types show high potential in introducing a cost-effective, efficient, and user- friendly means for growers to detect Verticillium wilt before symptoms occur. Overall, this study demonstrated that pathogenic and virulent strains of Verticillium spp. are active within plant tissues within 2 weeks of infection and that machine-based technologies and machine learning modelling can be used for the early detection of Verticillium wilt of potatoes in Australia.
-
ItemInformation Theory and Machine Learning: A Coding Approach( 2022-11)This thesis investigates the principles of using information theory to analyze and design machine learning algorithms. Despite recent successes, deep (machine) learning algorithms are still heuristic, vulnerable, and black-box. For example, it is still not clear why and how deep learning works so well, and it is observed that neural networks are very vulnerable to adversarial attacks. On the other hand, information theory is a well-established scientific study with a strong foundation in mathematical tools and theorems. Both machine learning and information theory are data orientated, and their inextricable connections motivate this thesis. Focusing on data compression and representation, we first present a novel, lightweight supervised dictionary learning framework for text classification. Our two-stage algorithm emphasizes the conceptual meaning of dictionary elements in addition to classification performance. A novel metric, information plane area rank (IPAR), is defined to quantify the information-theoretic performance. The classification accuracy of our algorithm is promising following extensive experiments conducted on six benchmark text datasets, where its classification performance is compared to multiple other state-of-the-art algorithms. The resulting dictionary elements (atoms) with conceptual meanings are displayed to provide insights into the decision processes of the learning system. Our algorithm achieves competitive results on certain datasets and with up to ten times fewer parameters. Motivated by the similarity between communication systems and adversarial learning, we secondly investigate a coding-theoretic approach to increase adversarial robustness. Specifically, we develop two novel defense methods (eECOC and NNEC) based on error-correcting code. The first method uses efficient error-correcting output codes (ECOCs), which encode the labels in a structured way to increase adversarial robustness. The second method is an encoding structure that increases the adversarial robustness of neural networks by encoding the latent features. Codes based on Fibonacci lattices and variational autoencoders are used in the encoding process. Both methods are validated on three benchmark datasets, MNIST, FashionMNIST, and CIFAR-10. An ablation study is conducted to compare the effectiveness of different encoding components. Several distance metrics and t-SNE visualization are used to give further insights into how these coding-theoretic methods increase adversarial robustness. Our work indicates the effectiveness of using information theory to analyze and design machine learning algorithms. The strong foundation of information theory provides opportunities for future research in data compression and adversarial robustness areas.