Technological University of the Philippines

Wireless power transfer (WPT) systems have garnered significant market attention owing to their broad applicability in portable electronic devices, electric vehicles, unmanned aerial vehicles, biomedical implants, and related fields. In these systems, operating frequency and efficiency are critical factors affecting both transmission efficiency and transmission distance, making high-frequency operation an important trend for improving overall WPT performance. However, elevating the switching frequency also introduces notable challenges, including increased switching losses in power devices, limited load adaptability, and poor anti-misalignment capability, which in practice often lead to degraded system efficiency and unsatisfactory waveform quality. Accordingly, this paper proposes a high-frequency inverter power supply system capable of operating at a maximum output voltage frequency of 25 KHz. Under conditions of a 10 KHz output frequency and a 20 K Omega load, the system achieves a peak efficiency of 94.01%. A prototype was implemented through the integration of a software algorithm based on ARM Cortex-M3 core control with a hardware architecture consisting of a driving circuit, a full-bridge inverter, and a switchable filtering module. This work offers practical design insights for the development of future high-frequency, high-voltage inverter systems, while also providing valuable experimental data to support further research in this area.

Road infrastructure is essential for economic development and accessibility; however, maintaining road quality in the Philippines remains challenging due to environmental factors and the limitations of manual inspection. This study proposes a road damage detection framework aligned with Department of Public Works and Highways (DPWH) standards using instance segmentation to improve the efficiency and reliability of pavement assessment. A localized dataset of 3,033 images and 3,947 annotated instances was collected under diverse conditions, covering four damage classes: alligator cracks, cracks, potholes, and pumping depression, annotated using polygon-based labeling. Two models, YOLOv8m-Seg and RF-DETR Segmentation, were evaluated to analyze trade-offs between real-time performance and segmentation accuracy. YOLOv8m-Seg achieved a mask mAP50 of 0.511 with an inference speed of 4.8 ms per image, while RF-DETR achieved higher accuracy with a mask mAP50 of 0.533 and mAP50-95 of 0.332. Both models showed lower performance in pothole detection. Results highlight a trade-off between efficiency and precision, supporting scalable, automated road monitoring systems.

Technology-enhanced learning offers a scalable approach for integrating e-vehicle content into engineering education, yet successful implementation in government-owned higher education institutions depends on faculty adoption and classroom-laboratory integration readiness. This study investigated the factors affecting faculty adoption of e-vehicle elearning modules and readiness to implement them across classroom instruction and laboratory-oriented (or lab-equivalent) learning activities. Using purposive sampling method, data were collected from 162 engineering and technology faculty in a Philippine government-owned higher education institution through an online survey. The proposed TAM-based extended model was tested utilizing the Partial Least Square Structural Equation Modelling. Results showed strong explanatory power for perceived usefulness, behavioral intention to integrate, and classroom-laboratory integration readiness. Curriculum compatibility and ICT competency significantly strengthened perceived usefulness and adoption intention, while financial support emerged as the strongest predictor of classroomlaboratory integration readiness. Perceived usefulness significantly increased intention and readiness, and intention significantly translated into readiness to integrate the modules. The findings provide actionable guidance for scaling e-vehicle elearning in public universities by prioritizing curriculum alignment, strengthening faculty ICT capability, and ensuring sustainable financial support to enable classroom-laboratory integration.

In response to the technical requirements for real-time quality control in the hot pressing process of intelligent plywood production, this study proposes a real-time process control framework driven by edge AI. This framework employs a three-layer edge intelligence architecture. This work shows a practical and efficient boundary node model application scheme for defect detection with multi-level lightweight strategies. In particular, this work builds a decision level data fusion approach for visual detection data and process parameters based on rules for defect-process parameter association mapping. Experimental results have shown that this designed scheme can efficiently detect defects in an edge computing environment. Additionally, with more multi-source fusion being considered in the site environment, the overall detection efficiency might be improved while maintaining a stable closed-loop control system. After that, quality enhancement for products and efficiency improvement for detection were realized. The results provide a feasible method for utilizing engineering processes for enhanced online quality detection for the plywood hot-pressing process based on practical experiences for intelligence upgrades in wood processing.

Accurate monitoring of fish fry populations is essential for effective aquaculture management, particularly during early developmental stages when survival rate and stocking density strongly influence production outcomes. However, automated detection of fish fry remains challenging due to their extremely small size, visual similarity, and frequent occlusion in dense rearing environments. This study presents a deep transfer vision–based detection framework for the simultaneous identification of live and dead largemouth bass (Micropterus salmoides) fry under high-density conditions. A large-scale annotated dataset comprising 45,120 images is constructed, containing 325,799 live fry and 22,819 dead fry instances, with individual images including 20–80 fry and exhibiting complex visual scenarios such as overlap, occlusion, and adhesion. Extensive preprocessing and data augmentation are applied to improve robustness to geometric and photometric variability. Experimental results demonstrate stable training convergence and strong generalization performance, achieving precision and recall exceeding 0.99 and a peak mAP@50–95 of 0.9671. Class-wise evaluation confirms reliable discrimination between live and dead fry, while qualitative testing on unseen dense aquaculture scenes validates the robustness of the proposed approach. The developed framework enables automated fry counting, mortality assessment, biomass estimation, and density management, contributing to the advancement of intelligent and data-driven aquaculture systems.