Carnegie Mellon University Africa

Despite increasing instability and recurrent blackouts, advanced monitoring technologies remain underutilized in Rwanda's power grid. This paper presents the first comprehensive study on strategic Phasor Measurement Unit (PMU) deployment in Rwanda, integrating optimal placement, field testing, and stability analysis. The Optimal PMU Placement problem was formulated using a mixed-integer linear programming framework and solved through a customized combinatorial setcover search refined with observability redundancy metrics. A constrained variant of the algorithm was further developed to determine optimal PMU siting under limited deployment budgets, identifying the top five critical substations that maximize network observability. Comparative benchmarking with Depth-First Search and commercial-solver-based MILP approaches verified the method's accuracy, achieving full topological observability with 21 PMUs. A pilot PMU installation provided 50 frames-per-second synchrophasor data, which analyzed, revealed interarea oscillations. These results demonstrate the value of highresolution synchrophasor analytics and establish a scalable, datadriven framework for resilient grid modernization in emerging power systems.

Abstract The Ethiopic script used in the Amharic Language presents persistent challengesfor Optical Character Recognition (OCR) due to its large character vocabulary,diacritics, and high variability in handwriting. Although Amharic is spoken byover 58 million people, progress in OCR has been constrained by the lack of large,diverse, sentence-level datasets. Existing datasets are small, synthetic-only, orlimited to character- or word-level annotations, preventing models from captur-ing the complexity of real documents. We introduce Fidel, the first large-scaleAmharic OCR dataset spanning handwritten, typed, and synthetic text. Fidelcontains 40k handwritten and 28k typed line images collected from 411 nativewriters, providing broad coverage of handwriting styles and modern vocabulary.We further formalize our approach as a scalable data acquisition and preprocess-ing pipeline deskewing, line extraction, and alignment designed to guide futuredataset creation for low-resource scripts. To complement the real data, we gener-ate high-quality synthetic Amharic text images to support robust model training.Using Fidel, we construct the first comprehensive benchmark for Amharic OCR,evaluating seven deep learning based OCR models. These models spanCNN, CTC, transformer and hybrid architectures, enabling a robust assessmentof domain transfer and modality-specific performance across handwritten, typed,and synthetic text. The best-performing model trained on Fidel achieves state-of-the-art results, with a CER of 2.64% and WER of 7.29%, demonstrating thesubstantial impact of our dataset on advancing practical, high-accuracy AmharicOCR.

While face recognition systems are increasingly deployed in critical domains, they remain vulnerable to presentation attacks and exhibit significant demographic bias, particularly affecting African populations. This paper presents a fairness-aware Presentation Attack Detection (PAD) system using Local Binary Patterns (LBP) with novel ethnicity-aware processing techniques specifically designed for African contexts. Our approach introduces three key technical innovations: (1) adaptive preprocessing with differentiated CLAHE parameters and gamma correction optimized for different skin tones, (2) group-specific decision threshold optimization using Equal Error Rate minimization for each ethnic group, and (3) three novel statistical methods for PAD fairness evaluation such as Coefficient of Variation analysis, McNemar's significance testing, and Bootstrap confidence intervals representing the first application of these techniques in presentation attack detection. Comprehensive evaluation on the CASIA-SURF CeFA dataset demonstrates significant bias reduction achievements: 75.6% reduction in accuracy gap between African and East Asian subjects (from 3.07% to 0.75%), elimination of statistically significant bias across all ethnic group comparisons, and strong overall performance with 95.12% accuracy and 98.55% AUC. Our work establishes a comprehensive methodology for measuring and mitigating demographic bias in PAD systems while maintaining security effectiveness, contributing both technical innovations and statistical frameworks for inclusive biometric security research.

Audio annotation is a crucial, yet time-consuming process, mostly due to the effort it takes human listeners to label sound data. We present an automated system for audio annotation that leverages a camera, a microphone and an off-the-shelf computer vision model to detect objects on a video track and label the corresponding audio track. The labeled audio data is then used to train a deep learning model, which enables a microphone based system to detect objects correctly without the need for a camera. To prove the concept, we developed a simple traffic monitoring system, taking advantage of the YOLOv8 model to annotate the audio data automatically and developing a ResNet-12 model to match road sounds to vehicle types, e.g., cars and motorcycles. Our automated solution achieves a 26x improvement in labeling speed compared to the manual approach. The resulting audio model for vehicle sounds achieves 0.91 precision and 0.9 recall. The proposed system is also able to count vehicles with less than 5% error. The generic methodology for automated audio annotation using off-the-shelf computer vision models extends naturally to a broad range of use cases for example in surveillance, wildlife protection or industrial monitoring.

ABSTRACTAir pollution spikes pose significant health risks and environmental challenges that demand innovative solutions for effective analysis and mitigation. This paper introduces a groundbreaking approach to revolutionize air pollution spikes analysis using a blockchain‐driven machine learning framework. Leveraging the transparency and immutability of blockchain technology, coupled with the predictive power of machine learning algorithms, our framework offers real‐time monitoring, accurate prediction, and proactive management of air pollution spikes. Our framework provides comprehensive insights into air quality dynamics by integrating data from diverse sources, including IoT sensors. Furthermore, the decentralized nature of blockchain ensures data integrity and enhances trust among stakeholders, including regulatory authorities, industries, and communities. Through case studies and simulations, we demonstrated the efficacy and scalability of our framework in addressing air pollution spikes across diverse geographical regions. The Machine learning techniques for the time series model (RNNs, ARIMA, and Exponential Smoothing) were analyzed and compared using statistical metrics (Mean Absolute Error [MAE], Mean Squared Error [MSE], and R‐squared [R2]). The exponential Smoothing model performed well compared to the other two models for all parameters, while both ARIMA and RNNNN models showed negative R2 values for certain pollutants, particularly SO2. For example, the PM10 scored 82.4% for R2. This research signifies a paradigm shift in air quality management, empowering stakeholders to make informed decisions and mitigate the adverse impacts of air pollution spikes on public health and the environment. This research demonstrated that machine learning and blockchain can be integrated to analyze data on air pollution spikes and predict pollutant emissions. This solution will help prevent harmful exposure to pollutants, protecting human health and the environment.

Despite significant technological advancements, there continues to be a considerable communication gap between individuals with hearing disabilities and the rest of society. This gap is exacerbated by the fact that the development and research of technologies, such as caption glasses, aimed at bridging this divide, primarily focus on sign languages used in countries with prominent tech industries, including European countries and USA. Consequently, there is a lack of resources and attention devoted to sign language recognition and translation systems for languages spoken in Africa. This research addresses this issue by concentrating on twenty-two common gestures in Kinyarwanda sign language. Through extensive exploration and evaluation of various machine learning algorithms, the study identifies the most effective approach for recognizing and translating these gestures. To validate the effectiveness of the developed system, real-world Kinyarwanda sign language video data is utilized for thorough training and testing. The research successfully culminates in the creation of a functional web application capable of accurately recognizing the 22 Kinyarwanda sign language gestures, both in live video feeds and recorded videos. This achievement represents a significant outcome of the research, as it addresses the specific needs of the Kinyarwanda signing community. By providing a reliable and accessible tool for gesture recognition and translation, the research contributes to narrowing the communication gap between individuals with hearing disabilities who use Kinyarwanda sign language and the wider society.

Proton exchange membrane electrolyzers are an attractive technology for hydrogen production due to their high efficiency, low maintenance cost, and scalability. To receive these benefits, however, electrolyzers require high power reliability and have relatively high demand. Due to their intermittent nature, integrating renewable energy sources like solar and wind has traditionally resulted in a supply too sporadic to consistently power a proton exchange membrane electrolyzer. This study develops an electrolyzer model operating with renewable energy sources at a highly instrumented university site. The simulation uses dynamic models of photovoltaic solar and wind systems to develop models capable of responding to changing climatic and seasonal conditions. The aim therefore is to observe the feasibility of operating a proton exchange membrane system fuel cell year-round at optimal efficiency. To address the problem of feasibility with dynamic renewable generation, a case study demonstrates the proposed energy management system. A site with a river onsite is chosen to ensure sufficient wind resources. Aside from assessing the feasibility of pairing renewable generation with proton exchange membrane systems, this project shows a reduction in the intermittency plaguing previous designs. Finally, the study quantifies the performance and effectiveness of the PEM energy management system design. Overall, this study highlights the potential of proton exchange membrane electrolysis as a critical technology for sustainable hydrogen production and the importance of modeling and simulation techniques in achieving its full potential.

Large Language models (LLMs) have brought about substantial advancements in the field of Question Answering (QA) systems. These models do remarkably well in addressing intricate inquiries in a variety of disciplines. However, because of domain-specific vocabulary, complex technological concepts, and the requirement for exact responses, applying LLMs to specialized sectors like telecommunications presents additional obstacles. GPT-3.5 has been used in recent work, to obtain noteworthy accuracy for telecom-related questions in a Retrieval Augmented Generation (RAG) framework. Notwithstanding these developments, the practical use of models such as GPT-3.5 is restricted by their proprietary nature and high computing demands. This paper introduces QMOS, an innovative approach which uses a Question-Masked loss and Option Shuffling trick to enhance the performance of LLMs in answering Multiple-Choice Questions in the telecommunications domain. Our focus was on using open-source, smaller language models (Phi-2 and Falcon-7B) within an enhanced RAG framework. Our multi-faceted approach involves several enhancements to the whole LLM-RAG pipeline of finetuning, retrieval, prompt engineering and inference. Our approaches significantly outperform existing results, achieving accuracy improvements from baselines of 24.70% to 49.30% with Falcon-7B and from 42.07% to 84.65% with Phi-2.

The rise of digital technology has dramatically increased the potential for cyberbullying and online abuse, necessitating enhanced measures for detection and prevention, especially among children. This study focuses on detecting abusive obfuscated language in Swahili, a low-resource language that poses unique challenges due to its limited linguistic resources and technological support. Swahili is chosen due to its popularity and being the most widely spoken language in Africa, with over 16 million native speakers and upwards of 100 million speakers in total, spanning regions in East Africa and some parts of the Middle East. We employed machine learning models including Support Vector Machines (SVM), Logistic Regression, and Decision Trees, optimized through rigorous parameter tuning and techniques like Synthetic Minority Over-sampling Technique (SMOTE) to handle data imbalance. Our analysis revealed that, while these models perform well in high-dimensional textual data, our dataset’s small size and imbalance limit our findings’ generalizability. The precision, recall, and F1 scores were thoroughly analyzed, highlighting the nuanced performance of each model in detecting obfuscated language. This research contributes to the broader discourse on ensuring safer online environments for children, advocating for expanded datasets and advanced machine-learning techniques to improve the effectiveness of cyberbullying detection systems. Future work will focus on enhancing data robustness, exploring transfer learning, and integrating multimodal data to create more comprehensive and culturally sensitive detection mechanisms.

Leveraging weather radar technology for environmental monitoring, particularly the detection of biometeors like birds, bats, and insects, presents a significant challenge due to the dynamic nature of their behavior. Unlike hydrometeor targets, biometeor targets exhibit arbitrary changes in direction and position, which significantly alter radar wave polarization upon scattering. This study addresses this challenge by introducing a novel methodology utilizing Rwanda’s C-Band Polarization Radar. Our approach exploits the capabilities of dual-polarization radar by analyzing parameters such as differential reflectivity (ZDR) and correlation coefficient (RHOHV) to derive the Depolarization Ratio (DR). While existing radar metrics offer valuable insights, they have limitations in fully capturing depolarization effects. To address this, we propose an advanced fuzzy logic algorithm (FL_DR) integrating the DR parameter. The FL_DR’s performance was rigorously evaluated against a standard FL algorithm. Leveraging a substantial dataset comprising nocturnal clear air radar echoes collected during a Fall Armyworm (FAW) outbreak in maize fields from September 2020 to January 2021, the FL_DR demonstrated a notable improvement in accuracy compared to the existing FL algorithm. This improvement is evident in the Fraction of Echoes Correctly Identified (FEI), which increased from 98.42% to 98.93% for biological radar echoes and from 87.02% to 95.81% for meteorological radar echoes. This enhanced detection capability positions FL_DR as a valuable system for monitoring, identification, and warning of environmental phenomena in regions similar to tropical areas facing FAW outbreaks. Additionally, it could be tested and further refined for other migrating biological targets such as birds, insects, or bats.