20505 research outputs found
Sort by
Mixed-precision federated learning via multi-precision over-the-air aggregation
Get PDFOver-the-Air Federated Learning (OTA-FL) is a privacy-preserving distributed learning mechanism, by aggregating updates in the electromagnetic channel rather than at the server. A critical research gap in existing OTA - FL research is the assumption of homogeneous client computational bit precision. While in real world application, clients with varying hardware resources may exploit approximate computing (AxC) to operate at different bit precisions optimized for energy and computational efficiency. Model updates with varying precisions among clients present a significant challenge for OTA - FL, as they are incompatible with the wireless modulation superposition process. Here, we propose an mixed-precision OTA-FL framework of clients with multiple bit precisions, demonstrating the following innovations: (i) the superior trade-off for both server and clients within the constraints of varying edge computing capabilities, energy efficiency, and learning accuracy requirements compared to homogeneous client bit precision, and (ii) a multi-precision gradient modulation scheme to ensure compatibility with OTA aggregation and eliminate the overheads of precision conversion. Through case study with real world data, we validate our modulation scheme that enables AxC based mixed-precision OTA-FL. In comparison to homogeneous standard precision of 32-bit and 16-bit, our framework presents more than 10% in 4-bit ultra low precision client performance and over 65% and 13% of energy savings respectively. This demonstrates the great potential of our mixed-precision OTA-FL approach in heterogeneous edge computing environments.Engineering and Physical Sciences Research Council (EPSRC)The work is supported by EPSRC CHEDDAR: Communications Hub for Empowering Distributed clouD computing Applications and Research (EP/X040518/1) (EP/Y037421/1).2025 IEEE Wireless Communications and Networking Conference (WCNC
Leveraging animal feed supply chain capabilities through big data analytics: a qualitative study
Get PDFPurpose
Although big data analytics (BDA) has gained widespread interest in supply chain management (SCM) literature in recent years, our understanding of how it contributes to improved animal feed supply chains (SCs) is still underexplored. This study provides a greater understanding of the role of BDA in improving animal feed SC capabilities.
Design/methodology/approach
A qualitative approach was used in this study. Data were collected through 32 semistructured interviews from several actors involved in the production and supply of animal feed concentrates.
Findings
This study provides rich in-description evidence of how BDA enhances performance in the animal feed supply chain through improved logistics capabilities, quality control and information visibility. Our findings also suggest that organizational culture contributes to leveraging BDA capabilities in the feed-processing SCs.
Practical implications
The research provides an in-depth qualitative investigation of implementing big data in the feed processing SCs. The study provides practical implications for SC managers in the agri-food sector.
Originality/value
The study contributes to the growing body of knowledge by providing field evidence of the relevance of BDA to animal feed SCs. Moreover, this study adds to the existing literature by providing an understanding of the role of the internal culture of the organization in leveraging BDA capabilities in the SC.International Journal of Quality & Reliability Managemen
Personalizing driver agent using large language models for driving safety and smarter human–machine interactions
Get PDFDriver assistance systems have been shown to reduce crashes by providing real-time warnings or assistance, with their effectiveness depending on communication with driver. Due to their unique characteristics, human drivers possess varying hazard perception skills and interaction preferences, making personalized assistance crucial to improving the user experience and system acceptance. However, how to leverage multimodal interfaces that dynamically adapt to warning contents and driver characteristics remains an open question. At the same time, large language models (LLMs) have demonstrated advanced capabilities in knowledge acquisition, planning, and human–machine collaboration, offering potential solutions for existing warning systems. Thus, we develop an LLM-based personalized driver agent (PDA), which provides personalized warnings through multimodal interactions (visual, voice, and tactile). The agent’s architecture mimics human cognitive processes via four core modules: memory, perception, control, and action. Results from our experiments indicate that the LLM-PDA effectively customizes warning contents for different drivers in various situations, providing enhanced safety and driver support. This article pioneers the integration of LLMs into automotive human–vehicle interaction and offers novel insights into personalized human–machine interaction in intelligent vehicles.This work was supported by the National Research Founda-tion of Korea (NRF) grant funded by the Korean government (RS-2024-00351865).IEEE Intelligent Transportation Systems Magazin
Exploiting the elastic properties of metallic glass thin foils: the undulatory mechanical response upon loading under geometric confinement
Get PDFSalonitis, Konstantinos - Associate Supervisor
Panagiotopoulos, Nikalaos - Associate External SupervisorA peculiar mechanical response was recently observed for metallic glass thin foils
in confined geometries under normal loading. In particular, when an arc-shaped
metallic glass foil is subjected to normal loading, it deforms elastically, changing
its shape by progressively increasing the number of formed sinusoidal arcs. This
mechanical response type is called undulatory behaviour and results from a
combination of successive elastic bending and buckling events occurring on the
foil. This behaviour is reversible and repeatable as long as the deformation of
metallic glass foils remains within the elastic deformation limit for metallic glasses
(< 2% strain) and can be exploited for developing novel types of nonlinear
springs.
The undulatory behaviour of metallic glasses was first reported in 2016 and has
been very little studied. This project aims to explore the limitations of studying the
mechanism and the underlying phenomena associated with this behaviour and
exploit its potential for engineering applications. More specifically, the undulatory
behaviour for three different metallic glass systems (i.e. Fe-Cr-Si-B and Ni-Fe-Si-
B-Mo and Ni-B-Si) with varying thicknesses of foil in the range from 19 – 40 μm
was studied, and various initial set-up geometries were investigated. Specific
focus was given to understanding the mechanism of the undulatory behaviour,
establishing some spring design principles and exploring the cycling fatigue life
of the glassy foils.
The undulatory behaviour involves elastic deformation, buckling and post-
buckling phenomena occurring in a repeatable sequence. Buckling plays a
significant role in the mechanism of undulatory behaviour. The required load for
buckling was approached using the Euler’s for buckling. The standing wave was
shown to be able to describe the sequence of the undulatory response but only
for the waveforms with perfect sinusoidal arcs. The parameters of the initial arc’s
geometry set-up, such as the boundary length (chord), amplitude, and foil
thickness enable the tuning of the undulatory behaviour and offer capabilities for
designing spring-type devices with desirable characteristics that can efficiently
operate in different load ranges.
Under cyclic loading the glassy foils endured between 19 x 10³to 225 x 10³
cycles, depending on load/strain parameters, foil thickness and alloy
compositions, typically higher than conventional low cycle springs. Crack
initiation during cyclic loading was found to be associated with surface defects
and side edges of the foils. This unique mechanical response of metallic glass
foils makes them a promising material for a wide range of applications,
contributing to being exploited for the enormous potential for designing novel
types of micro-flat springs using the exceptional mechanical and elastic
properties of metallic glass foils.PhD in Manufacturin
Detection of Fusarium spp. and T-2 and HT-2 toxins contamination in oats using visible and near-infrared spectroscopy
Get PDFFusarium langsethiae (FL) is one of the major contaminants in oats in the United Kingdom (UK) and is a significant producer of T-2 and HT-2 toxins, among the most prevalent mycotoxins in oats. Visible and near-infrared (Vis-NIR) (350–2500 nm) spectroscopy was explored as a non-invasive, rapid method for detecting FL, Fusarium species that produce T-2 and HT-2 toxins, and T-2 and HT-2 toxins content. Oat grains were artificially inoculated with FL and other Fusarium species under controlled water activity (aw) conditions (0.98, 0.90, and 0.80). FL was found to be particularly responsible for producing T-2 and HT-2 toxins. Classification models were developed to distinguish oat grains based on the presence of FL. The best performance was achieved with all the Vis-NIR spectra, with a classification accuracy of 76.2 %. The Vis region (350–995 nm) emerged as the most important range for classification. Additionally, oat grains were classified by T-2 and HT-2 toxin content, distinguishing oats above and below the European Union (EU) threshold with 93.3 % accuracy. For mycotoxin quantification, the best performance was obtained using the Vis region with a coefficient of determination (R2) of 0.875. Key wavelengths such as 464, 568, 575 and 636 nm were relevant for toxin detection. The NIR region (1005–1795 nm) also played a significant role in the models. This study shows that Vis-NIR spectroscopy is a promising, non-destructive tool for detecting Fusarium and type A trichothecenes in oats, though further research is needed to improve model robustness and support food safety monitoring.Biotechnology and Biological Sciences Research Council (BBSRC)This research is supported by a BBSRC-SFI research grant (BB/P001432/1) between the Applied Mycology Group at Cranfield University and the School of Biology and Environmental Science, University College Dublin, Ireland. This work was also supported by the Spanish Ministry of Universities (predoctoral grant FPU21/00073).International Journal of Food Microbiolog
Trailblazing specific generative models (SGMs) for early-stage engineering design concepts
No full textRecent advancements in large language models (LLMs) have the potential to revolutionise engineering design practices. LLM-powered tools such as ChatGPT and DeepSeek have shown significant promise in enhancing various aspects of engineering design. They leverage emergent data-driven technologies to deliver powerful capabilities that challenge the long-held belief that generating new concepts and automating engineering design activities are beyond the reach of artificial intelligence (AI). The capabilities demonstrated by these technologies are gradually overcoming initial resistance and fostering a more receptive attitude towards their adoption. However, a substantial gap remains between the generic capabilities of AI technologies and their specific application in engineering design. This study aims to leverage a combination of AI technologies to support early-stage engineering design activities. Specifically, we created a generative Markov Chain model adjusted as a specific generative model (SGM) to support the complex and time-consuming conceptual early-stage design activities. The SGM does not require extensive and expensive data training like the generic LLM, but produces comparably more natural concepts. The SGM used in this study effectively addresses the privacy concerns associated with generic LLMs. This pioneering research on SGM is cost-effective and holds promise for applications beyond engineering design, including fields such as medicine.Design and Artificial Intelligenc
Evaluating biological realism in ecological modelling: application of a novel framework to compare mechanistic and process-based earthworm and wild pollinator population models
No full textEcological models can support land management decisions and optimisation schemes that need to account for invertebrate population responses at the field to landscape level. However, models that incorporate greater biological detail (e.g. individual-level physiological and behavioural responses) often become computationally intractable at larger spatial extents. Such trade-offs in model development lead to ad hoc model design for different species and management questions, hindering generalisable insights needed to advance predictive ecological models for decision support. To facilitate model comparison, we developed and applied a novel approach to quantify the biological realism of models for two functionally important invertebrate groups commonly targeted by management interventions. Mechanistic and process-based population models for earthworms (n = 23) and wild pollinators (n = 24) were identified through a structured review. We find that earthworm models are predominantly non-spatial or micro-scale (<10 m extent) and often incorporate detailed physiological mechanisms. Pollinator models frequently simulate landscape-scale scenarios (≥1 km extent) and typically rely on aggregated processes to predict population dynamics or crop visitation rates, although some include detailed individual-level movement behaviours. Species- and scale-specific model structures highlight the need for greater integration of physiological and behavioural mechanisms across broader spatial extents. We recommend systematic strategies to build on the progress made by existing models, aiming to resolve the trade-off between realism and tractability for more informed population predictions at management-relevant spatial scales. Our framework complements existing efforts towards greater transparency in model development, communication, and application for robust environmental decision support.This work was supported by UKRI BBSRC FoodBioSystems Doctoral Training Partnership (DTP), grant number BB/T008776/1 and CASE support from Syngenta.Ecological Modellin
Atmospheric pressure plasma surface engineering to improve polymer matrix composite bonding
No full textNichols, John - Associate Supervisor
See, Tian-Long - Associate SupervisorThis thesis explores the improvement of adhesive properties on carbon fibre
reinforced epoxy composites, specifically MTM44-1, through the application of
atmospheric pressure plasma (APP) treatments. The study aims to address the
critical challenge of enhancing the adhesion of composite surfaces by
investigating both thermal and non-thermal plasma techniques. A comprehensive
analysis of plasma treatment uniformity, temperature effects, and treatment
speeds has been conducted.
Special emphasis is placed on the removal of silicone contamination, from mould
release agents, a common barrier to effective adhesion. Current industrial
approaches often struggle to fully eliminate silicone contamination, which
severely limits adhesive bonding on composite surfaces. Plasma has the
potential to remove silicone contamination, while also improving the surface
energy, prior to adhesion.
Another consideration in this study has been to the speed of plasma treatment.
One of the barriers to using adhesion to bond composites in the industry is the
time it takes to pre-treat surfaces prior to adhesion.
This study aims to also investigate the relationship between treatment time and
surface energy, especially as APP can be easily integrated into a manufacturing
processes, compared to traditional plasma surface treatment. This is due to the
lack of vacuum which saves time, space and often energy in the process.
Various plasma generation methods were explored, including microwave, radio
frequency (RF), and Dielectric Barrier Discharge (DBD), to treat the composite
surfaces. This allowed for a variety of different plasma feed gases to be used
including nitrogen, argon and a mixture of fluorinated gas and argon.
Surface characterisation was conducted using X-ray photoelectron spectroscopy
(XPS) to assess the chemical composition of the treated surfaces, while water
contact angle (WCA), surface energy measurements and lap shear testing were
used to evaluate the effectiveness of the treatments in promoting adhesion. The
results demonstrate that plasma treatments significantly enhance the surfaceii
energy of the composites by removing surface contaminants, particularly silicon
containing release agents, while also altering the surface chemistry to promote
better adhesive bonding.
The work also evaluates the impact of processing parameters such as treatment
speed and gas flow rates on surface modification efficiency and surface
temperature, particularly when considering thermal radio frequency (RF) plasma.PhD in Manufacturin
A distributed digital twin framework for effective asset management in industry 4.0
No full textSamie, Mohammad - Associate SupervisorWind energy is crucial for achieving net-zero emissions by 2050. However, unexpected wind
turbine failures, including fires and abrupt breakdowns, pose significant challenges to the
reliability of wind energy. Predictive Maintenance (PdM) is an Asset Management (AM)
strategy used to forecast equipment fault and resolve them before they occur, while Condition
Monitoring (CM) is another asset management strategy used to monitor the condition of
equipment at any given time. This work focuses on utilizing a technology called Digital Twin
(DT) – a digital replica of wind turbines’ physical components in the digital domain, to support
asset (equipment) management using predictive maintenance. This DT is used to achieve
condition monitoring of wind turbine components and Machine Learning (ML) is applied to
achieve predictive maintenance (PdM). The DT framework achieved in this work shows the
use of streaming sensor data from the Supervisory Control and Data Acquisition (SCADA)
system of wind turbines to constantly monitor the asset condition and use ML to anticipate
component failures well in advance so as to provide operational teams with a cost-effective
solution to simulate configurations that can help in managing turbine fault. While there have
been successes in using digital twins and predictive maintenance in manufacturing systems,
DT itself is an emerging technology that is still being explored as a concept. The explorations
of the application of DT in literature highlighted various misconceptions, limitations and
challenges that were necessary to look at in this study before achieving a wholistic framework
that proposes a valuable and cost-effective DT implementation for the wind turbine predictive
maintenance case study. To achieve the DT framework, multiple enabling technologies such as
Industrial Internet of Things (IIoT), cloud computing and machine learning were considered
and broken down into layers and applications. To standardize the DT framework, the ISO
23247 DT standard was used to guide the implementation of multiple DT architectures while
evaluating various technology options. The methodology brought together and evaluated
different machine learning techniques, software and hardware configurations, data streaming
and storage technologies as well as architectural design patterns towards clarifying and
validating the proposed DT framework. Experiments and simulations with open data, and
results from real operational wind turbine SCADA data provided by an industry partner helped
in validation.
The PhD contribution is a functional architectural framework that is effective in the adoption
of digital twins for predictive maintenance in wind turbines. Thus, the framework serves
operational, resource utilization, scalability, standardization and cost benefits using a
distributed digital twin approach for predictive maintenance. The work highlights and covers
gaps from the associated challenges and limitations such as the lack of standardization found
in earlier approaches in DT research within the domain of predictive maintenance in IIoT
applications.PhD in Transport System