http://theses.ncl.ac.uk/jspui/handle/10443/1522 Wed, 04 Feb 2026 22:06:29 GMT 2026-02-04T22:06:29Z http://theses.ncl.ac.uk/jspui/handle/10443/6299 Title: Advancing rehabilitative robotics through signal processing and machine learning algorithms Authors: Suppiah, Ravi Abstract: Rehabilitative robotics holds tremendous promise in improving the quality of life for individuals with motor impairments. The integration of signal processing and machine learning algorithms into rehabilitative robotics systems has emerged as a powerful approach to enhance the effectiveness and efficiency of rehabilitation therapies. This thesis aims to explore and contribute to the advancements in this exciting field. The first part of this research focuses on signal processing techniques applied to the analysis and interpretation of sensor data in rehabilitative robotics. Various signal processing methods such as filtering, feature extraction, and time-frequency analysis are investigated to extract relevant information from sensory signals captured by robotic devices. These processed signals serve as valuable inputs for subsequent machine learning algorithms. The second part of the thesis delves into the application of machine learning algorithms in rehabilitative robotics. Supervised, unsupervised, and reinforcement learning techniques are studied to model and predict user intent, adapt robot behaviour, and optimize rehabilitation exercises. These algorithms play a pivotal role in personalizing the rehabilitation process, enabling tailored interventions based on individual needs and progress. The integration of signal processing and machine learning presents unique opportunities for real-time adaptation and closed-loop control in rehabilitative robotics. The combination of sensor data processing and machine learning enables the creation of intelligent robotic systems that can dynamically adjust therapy parameters, ensuring optimal engagement and challenging the user at an appropriate level. In addition to technological advancements, this research also addresses practical challenges in the implementation of signal processing and machine learning algorithms in real-world rehabilitative robotics applications. Considerations such as computational efficiency, robustness to noise and variability, and user acceptance are carefully examined to ensure the feasibility and effectiveness of the proposed approaches. Overall, this thesis aims to contribute to the field of rehabilitative robotics by advancing the integration of signal processing and machine learning algorithms. Description: PhD Thesis Mon, 01 Jan 2024 00:00:00 GMT http://theses.ncl.ac.uk/jspui/handle/10443/6299 2024-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/6261 Title: Improved Finite Control Set of Fast Model Predictive Control for Modular Multilevel Inverter Authors: Di, Kexin Abstract: The rising demand of energy is currently an inexorable trend. According to the database, a number of consulting firms concurred with the prediction of a potential energy crisis in the near future due to extreme climate and natural disasters, import and export regulations under Covid, and the rapid development of industries such as electric vehicles. In the meantime, the cost of renewable energy is lower than it was in the past, and in light of mounting environmental concerns, accelerating the building of the renewable energy industry has become the optimal solution. On the other hand, the challenges of grid connection become an impetus for the evolution of the power transmission system. Nowadays, modular multilevel converter based HVDC system has become more popular and has the most potential. In addition, MMC (modular multilevel converter) is normally used in medium/high voltage motor drive systems as well. Modular multilevel converter inherits the advantages of low filter cost, low output distortion, and easy redundancy, from the multilevel converter family. Additionally, due to modular design, MMCs have the widest range of operating voltage and the broadest applications. The commercialization of MMC is hindered by two factors: the difficulty in developing high-voltage DC circuit breakers and the extremely complex and difficult control schemes. This research project is focused on the challenges in control schemes: design constraints, sub-module capacitor voltage control, and circulating current control. A comparison has been made between the classical control scheme and the model predictive control scheme. A new MPC (model predictive control) approach has been presented, which releases the heavy computational complexity, and improves the performance of circulating current minimization while smoothing the capacitor voltage ripples to lower the switching frequency. In addition, throughout the hardware implementation procedure, the realization of the general sorting algorithm has been improved, accelerating the sorting time while decreasing hardware resource use. Both simulation and hardware evaluations have been performed on this approach. In this thesis, specific outcomes and results are presented. Description: Ph. D. Thesis. Mon, 01 Jan 2024 00:00:00 GMT http://theses.ncl.ac.uk/jspui/handle/10443/6261 2024-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/5981 Title: Towards Automatic Photo-Identification of Cetaceans: A Fine-Grained, Few-Shot Problem in Marine Ecology Authors: Trotter, Cameron Abstract: Understanding the health of Earth’s ecosystems is imperative for the future safeguarding of our planet and its inhabitants. One of the most common tools utilised by researchers to develop their understanding of an area’s health is indicator species, organisms whose abundance or absence in a system reflects overall environmental health. Cetaceans such as dolphins, porpoises, and odontocetes (toothed whales) are excellent indicator species given their status as top predators, allowing for the monitoring of risks to marine environments, such as offshore wind farm development or commercial fishing activity. Cetacean monitoring is frequently performed using capture-recapture surveys, through which researchers record the presence of individual animals to produce population estimates. Photo-identification (photo-id) is one of the main non-invasive capture-recapture methods, whereby image data containing the animals’ individually identifiable prominent markings are captured. Upon survey completion these data are curated to produce a photo-id catalogue, allowing for an abundance estimate to be generated and ecosystem health to be determined. Catalogues are updated over time as more surveys are undertaken, new individuals are encountered, and prominent markings change. Photo-id catalogue curation is traditionally performed manually and can be extremely labour and cost intensive, especially for large resident populations. This thesis details a framework for automatic photo-id catalogue matching based on unprocessed field imagery via a pipeline of computer vision models. The creation of a photo-id catalogue containing cetaceans resident in the waters of Northumberland, UK, is first outlined. The development of a coarse-grained dorsal fin detector and the use of post-processing techniques to aid downstream identification is then examined. Next, the created photo-id catalogue is utilised to facilitate the development of a model capable of finegrained, few-shot catalogue matching via latent space similarity, allowing for the flagging of potentially uncatalogued individuals. At all stages, the developed techniques’ robustness to spatio-temporal changes is evaluated, including their generalisability to multiple cetacean species. The automation of photo-id data curation outlined in this thesis affords researchers more time to work on application of their data, for example to inform mitigation and policy change, rather than administration. Description: Ph. D. Thesis Sun, 01 Jan 2023 00:00:00 GMT http://theses.ncl.ac.uk/jspui/handle/10443/5981 2023-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/5634 Title: Multi-objective torque control of switched reluctance machine Authors: Dankadai, Najib Kabir Abstract: The recent growing interest in Switched Reluctance Drives (SRD) is due to the electrification of many products in industries including electric/hybrid electric vehicles, more-electric aircrafts, white-goods, and healthcare, in which the Switched Reluctance Machine (SRM) has potential prospects in satisfying the respective requirements of these applications. Its main merits are robust structure, suitability for harsh environments, fault-tolerance, low cost, and ability to operate over a wide speed range. Nevertheless, the SRM has limitations such as large torque ripple, high acoustic noise, and low torque density. This research focuses on the torque control of the SRD with the objectives of achieving zero torque error, minimal torque ripple, high reliability and robustness, and lower size, weight, and cost of implementation. Direct Torque Control and Direct Instantaneous Torque Control are the most common methods used to obtain desired torque characteristics including optimal torque density and minimized torque ripple in SRD. However, these torque control methods, compared to conventional hysteresis current control, require the use of power devices with a higher rating of about 150% to achieve the desired superior performance. These requirements add extra cost, conduction loss, and stress on the drive’s semiconductors and machine winding. To overcome these drawbacks, a simple and intuitive torque control method based on a novel adaptive quasi sliding mode control is developed in this study. The proposed torque control approach is designed considering the findings of an investigation performed in this thesis of the existing widely used control techniques for SRD based on information flow complexity. A test rig comprising a magnet assisted SRM driven by an asymmetric converter is constructed to validate the proposed torque control method and to compare its performance with that of direct instantaneous torque control, and current hysteresis control methods. The simulation and experimental results show that the proposed torque control reduces the torque ripple over a wide speed range without demanding a high current and/or a high switching frequency. In addition, It has been shown that the proposed method is superior to current hysteresis control method in the sensorless operation of the machine. Furthermore, the sensorless performance of the proposed method is investigated with the lower component count R-Dump converter. The simulation results have also demonstrated the excellent controller response using the standard R-Dump converter and also with its novel version developed in this thesis that needs only one current sensor. Description: PhD Thesis Sat, 01 Jan 2022 00:00:00 GMT http://theses.ncl.ac.uk/jspui/handle/10443/5634 2022-01-01T00:00:00Z