SIGNTALK: A Ghanaian Sign Language Translation System Using Deep Learning       

Project Lead: Dr. Emmanuel Ahene

Communication is an essential aspect of humanity, and we communicate in several ways. One way some of us, specifically the deaf and mute, can communicate with others is through sign language. Sign language, like other languages, is a natural language usually used by the deaf, mute, and hearing impaired for communication. Sign language is non-verbal and visually oriented and uses hand gestures, facial expressions, and posture changes as modes of communication among such individuals. According to statistics, there are at least 111,000 deaf people in Ghana, and sadly there are very few people who understand and can translate sign language. This creates a problem for the deaf and mute in several ways. Like everyone else, they must interact with other people in society daily to meet their physiological and social needs. 

Therefore, there is a need to create an intermediary that can effectively enhance conversation between the deaf/mute and other people. Such a system is very critical and has a wide range of applications. But in this project, we will focus on bridging the communication gap in the health sector, mainly focusing on communication in the doctor-patient relationship. 

In Ghana, there are not enough exceptional hospitals for the disabled where the deaf and mute can readily get access to translators who will effectively communicate their problems to doctors. And the ones present are too expensive to be afforded by most people with disabilities. This project is, therefore, all about solving this problem by producing a system that allows the deaf, mute, and hearing-impaired people to access the vast number of public and private hospitals in Ghana and be able to effectively communicate their ailments to doctors and understand the directions and medical advice given by their doctors. 

Existing solutions have several limitations, requiring specialised cameras and hardware like gloves, depth sensors, and others to achieve this goal. However, our envisioned system uses a simple 0computer’s webcam to capture an individual’s signing gestures and convert them into corresponding text and speech in English in real-time. The text obtained from the signing gestures is then converted into audio. This will be achieved by harnessing the power of machine learning and computer vision and creating a model that can detect a subject’s hand gestures, facial expressions, and posture and predict the desired corresponding word in real-time. This extra capital that must be invested in buying these specialised tools is taken out, thereby making our solution more affordable and convenient.

Design And Optimisation of Hybrid Energy Systems for Electric Aircraft Propulsion

Project Lead: Dr. Mrs. Eunice A. Adjei

Over the last few decades, fossil fuel consumption has increased significantly, drawing concerns regarding energy sustainability. To overcome this challenge, the aeronautics industry’s focus has shifted towards environmentally clean propulsion systems, including electric motors. However, the specific energy and energy density of electrical energy storage sources, like batteries, are much lower than that of fossil fuel. 

Compared with an electric aircraft, the hybrid-powered aircraft based on today’s technology has a more excellent range and offer more flexibility. In such a configuration, the hybrid system’s propulsive unit (engine or electric motor) could enable the engine to operate in an optimal operational envelope with the other unit (i.e., motor) to cover the addition/transient power demands. This flexibility could potentially reduce the size of the engine, however, at the cost of additional complexity and integration challenges. 

The design approach consists of coupling sizing models with an optimisation algorithm that automatically tunes the parameters to optimise system criteria (mass, losses, etc.) while satisfying the technological and operational constraints.

High Throughput Screening of Porous Materials for Photoelectrochemical Reduction Of CO2 into Fuels using a Convolutional Neural Network

Project Lead: Dr. Kwadwo Mensah-Darkwa

Advances in Machine learning (ML) and the development of robust material databases in recent decades have accelerated the application of machine learning in heterogeneous catalysis. Electrochemical reduction of CO2 into valuable products continues to attract the attention of researchers in green chemistry because of its potential to ultimately cut down greenhouse gas (GHG) emissions and produce low-cost fuel, among others. However, the sluggish reactions at the electrode interface limit industrial applications of the CO2 reduction reaction. In this study, an ML-based approach will be utilised to screen different porous materials for photo-catalytic and electrocatalytic reduction of CO2 into fuels.

A deep learning convolutional neural network will be adopted to screen stable porous materials from OC20 and the Materials Project databases by predicting adsorption energies and the turnover frequencies. Surface features, atomic and molecular level information of porous materials relevant for estimating adsorption energies of CO2 on photo- and electro-catalysts with Density Functional Theorem (DFT) level of accuracy will be used as descriptors for the predictive analysis. The performance of the CNN will be compared with kernel-based models such as Kernel Ridge Regression (KRR) and Support Vector Regression (SVR). DFT calculations on the promising material candidates to confirm the ML-predicted adsorption energies and turnover frequencies using Quantum Espresso. This approach is expected to achieve a DFT level of accuracy in estimating adsorption energies and turnover frequencies of CO2 porous materials catalyst surfaces.

At the end of this study, it is also expected that root means square error (RMSE) below -0.1 eV as compared to DFT computed values will be achieved. 

Machine Learning of KNUST’s Renewable Energy Resources for Planning its Sustainable Captive Energy Generation Systems

Project Lead: Dr. Richard Opoku

Electricity supply in public universities in Ghana is challenged by frequent power outages from the national grid, which adversely affects academic work and research activities. An initial assessment has shown that at KNUST alone, there are roughly 2-6 times power outages per day when students are on campus. This has resulted in many Colleges and Faculties buying diesel generators as backup power systems, with substantial operational fuel costs and negative environmental impacts. KNUST has plans to establish its sustainable mini-grid system with electricity generation from renewable energy (RE) resources such as solar, biogas, biomass, municipal solid waste, wind, and geothermal. The availability of these RE resources varies at different times of the day and months of the year and requires proper technical analysis for planning KNUST’s in-house RE generation. In addition, there is insufficient data on the hourly, daily and seasonal energy consumption profiles for KNUST to enable proper planning of KNUST’s renewable energy mini-grid system.

This research project aims to obtain data and analyse the different RE resources at the KNUST campus and KNUST’s hourly and seasonal energy consumption profile to plan its renewable energy min-grid system. In addition, machine learning tools such as SVR, KNN, MLR, and ANN will be used to model and predict KNUST’s RE resources to plan its captive energy.

An Artificial Intelligence-Based Non-Intrusive Load Monitoring of Energy Demand in Electrical Energy Systems using Optimisation Algorithms

Project Lead: Eur-Ing. Benjamin Kommey

Load monitoring is an activity predominantly carried out to ascertain the level of energy or load demand by consumers. In power distribution networks, utility companies undertake this monitoring exercise to give customers quality and reliable power. 

However, in Ghana, most customers on the distribution network oftentimes need better power quality and reliability issues stemming from low voltages, over-voltages, and sags due to load imbalances and other operational issues relating to the electrical system. 

More so, personnel who embark on this exercise are mostly exposed to safety threats since this activity is mainly done manually with a clamp on the ammeter.
This project seeks to resolve power quality and reliability issues by incorporating artificial intelligence to help in optimising load monitoring. 

A non-intrusive approach is used to solve this challenge, and an optimisation algorithm is adopted to design and operate the hardware prototype. Therefore, this would help power utilities with the relevant data required to address facility expansion, renovation, and accurate metering with ease and customers with load and energy management.