View Proposal

---

Proposer

Chee Een Yap

Title

A Real-Time Waste Bin Monitoring with zero-shot learning

Goal

The goal of this project is to design and implement a portable Internet of Things (IoT) system for real-time waste bin monitoring with zero-shot learning in localized environments, such as university campuses and night markets. The proposed system integrates detachable sensing modules with Zero-Shot Learning (ZSL) techniques to enable the monitoring, forecasting and analysis of waste accumulation patterns in previously unseen deployment environments without the need for extensive data collection or model retraining. To support efficient waste collection operations, the system presents sensor-based fill-level measurements through a mobile-friendly interface. This enables cleaning personnel to make informed, timely, and proactive waste collection decisions, thereby improving operational efficiency and maintaining cleaner public spaces.

Description

The prototype will include two to three ordinary bins fitted with detachable monitoring nodes. Each node will use an ESP32 development board, a calibrated ultrasonic distance sensor, a removable bracket, a protective enclosure and a suitable USB power supply or power bank. A Wi-Fi router or hotspot will connect the nodes to one Raspberry Pi equipped with a microSD card and a suitable power supply. The Raspberry Pi will store readings and host the web interface. The proposed system will use Python for model training and forecasting, Flask for the backend, SQLite for data storage, and HTML, CSS and JavaScript for the interface. Cleaning staff will use an existing smartphone to view bin numbers, location photographs, current fill levels, forecasts and collection recommendations, and confirm completed tasks. Timestamps and connection status will identify outdated readings. The research will investigate zero-shot learning (ZSL) through data collection, model training and cross-scenario evaluation. A forecasting model will be trained on source-scenario data and tested on a separate target deployment excluded from training and model selection. Recent target readings will serve only as prediction inputs, without target-specific retraining or fine-tuning. The model architecture and experimental design will be confirmed with the supervisor. Controlled experiments using clean mock waste will compare forecasts with last-value and recent-trend baselines under different filling patterns and bin-emptying events. Evaluation will cover fill-level estimation error, forecast error, update latency and collection-warning timeliness.

Resources

Background

Url

Difficulty Level

Moderate

Ethical Approval

None

Number Of Students

1

Supervisor

Chee Een Yap

Keywords

internet of things (iot), zero-shot learning (zsl), python

Degrees

Bachelor of Science in Computing Science