Machine learningMar — Jul 2026Final Year Project

LSTM Gas Detection

Detecting hazardous gas leaks with 99.84% accuracy.

99.84%

accuracy

Gas detection prototype showing live sensor feeds and hazard predictions — Live monitoring · sensor feeds and hazard predictions during simulated environmental events.

Smoke diagnostics comparing zero and increased sensor noise — Smoke diagnostics · model confidence with and without added sensor noise.

Non-hazardous perfume diagnostics comparing zero and increased sensor noise — Perfume diagnostics · distinguishing a non-hazardous anomaly under different noise conditions.

Training and validation accuracy and loss across 13 epochs — Training curves · accuracy and loss across training and validation runs.

Confusion matrices at 20 percent and 40 percent noise for mixture, clean air, perfume and smoke — Noise robustness · confusion matrices at 20% and 40% noise.

Confusion matrices at 60 percent and 80 percent noise for mixture, clean air, perfume and smoke — Noise robustness · confusion matrices at 60% and 80% noise.

Model accuracy plotted against noise levels from 10 to 100 percent — Stress testing · accuracy across increasing noise levels, from 10% to 100%.

01

The problem

Industrial sites need early warning of hazardous gas, and sensor data is noisy time-series that simple thresholds miss.

02

How I approached it

1.Preprocessed and analysed time-series sensor data in Python.
2.Built a full TensorFlow/Keras LSTM training and validation pipeline.
3.Integrated predictions into a prototype monitoring system.

03

What shipped

LSTM classification modelTraining & validation pipelinePrototype hazard monitor

Wine Quality ML →