Air quality is a critical determinant of human health, with severe consequences resulting from air pollution. The growing necessity for air quality monitoring has led to the adoption of IoT sensor networks, which provide real-time data for forecasting, issuing warnings, and informing public health interventions. In this context, machine learning (ML) algorithms have proven to be powerful tools for enhancing air quality prediction and addressing monitoring challenges. However, a comprehensive review compiling the research space of ML for air quality is seldom available. This review analyzes over 70 recent studies that apply ML techniques to air quality monitoring, categorizing them based on the type of learning approach employed, with a focus on identifying the most effective algorithms in each category. The findings demonstrate that ensemble models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) consistently achieve high accuracy in structured datasets, while deep learning (DL) approaches like Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) excel in capturing temporal dependencies and spatial patterns in pollution forecasting. Unsupervised approaches like clustering and anomaly detection effectively enhance data quality and sensor calibration, whereas reinforcement learning shows promise in adaptive control scenarios, despite challenges related to computational intensity and interpretability. This review is highly significant, offering valuable insights for policymakers and researchers in developing strategies to mitigate air pollution and improve public health using advanced ML techniques.

Ensuring the reliability and operational efficiency of critical scientific equipment is a central challenge in high-stakes research environments such as nuclear physics laboratories and particle accelerator facilities. Unexpected failures entail significant financial cost and prolonged interruptions to experimental programmes. We present a predictive maintenance (PdM) framework built around a two-layer Vanilla Long Short-Term Memory (LSTM) network trained on multivariate sensor streams collected at NRF-iThemba LABS between January 2021 and December 2023. Four channels, namely supply voltage, vibration velocity, differential pressure, and rotational speed, were recorded at 5 $\min$ intervals using a suite of industrial-grade transducers (power quality analyser, IEPE accelerometers, differential pressure transmitters, and proximity encoders) feeding a multi-channel data-acquisition chassis via OPC-UA, yielding a time-synchronised dataset of 315,360 observations. A normalised failure score converts the binary classifier output into a continuous, interpretable health indicator that supports tiered scheduling of maintenance. The Vanilla LSTM achieved a test-set F1-score of $75\%$ and an area under the receiver-operating-characteristic curve (AUC) of $0.856$, outperforming five competing architectures (PCA/$T^2$, Random Forest, Deep Neural Network, LSTM Autoencoder, and Bidirectional LSTM Autoencoder), and delivered a mean failure lead time of $(42.3\pm 7.2)$$\mathrm{h}$, exceeding the 36 $\mathrm{h}$ engineering requirement for proactive maintenance scheduling.

Transfer learning has been gaining significant traction as a means of reusing representations learned in one domain to accelerate and improve performance in another. While it first rose to prominence in Natural Language Processing and Computer Vision, it is increasingly being recognized as a powerful framework for other domains, including time-series analysis. In this work, we investigate transfer learning for medium to long horizon time-series forecasting of Wi-Fi traffic across building floors. We train a seasonality-aware model on one floor (source domain) and transfer them to other floors (target domains) to reduce data and compute requirements while preserving accuracy. Our approach decouples seasonal structure by first modeling recurring seasonal patterns using the Seasonal Auto-Regressive Integrated Moving Average (SARIMA) model then learn the trend and residuals with an LSTM. We implement the transfer model with partial fine-tuning strategies, and evaluate performance using accuracy and number of model parameters as key metrics. In addition, we compare the fine-tuned models against baseline models trained from scratch to assess how closely transfer learning approaches approximate full training. The proposed transfer learning framework is easily scalable to larger sensor networks, particularly in air-quality monitoring systems, requiring only minimal re-training. Moreover, it offers greater efficiency than knowledge distillation, which depends on first training a large, high-parameter teacher model.