00 · Orientation
Your gas-leak project, ready to defend.
This site explains one document: Real-Time Gas Leak Detection in Logistics and Manufacturing Environments Using Support Vector Machine on IoT Gas Sensor Data, your M.S. project report. Every number here comes from the real experiment. Where the work has limits, this site says so plainly, so you hear it here first and not in the viva.
The whole idea of a Support Vector Machine in one picture. Teal circles are one class, orange squares the other. The solid line separates them; the dashed lines mark the margin, the empty gap on each side. The SVM picks the line with the widest gap. The circled points touch the margin: they are the support vectors, and only they decide where the line goes.
In plain words — what you actually did
Factories, warehouses and cold stores use gases such as LPG and ammonia. Leaks kill people. Most cheap gas alarms just watch one sensor and ring when it passes a fixed level. That alarm cannot tell a dangerous leak from a harmless smell like perfume, so it gives false alarms and people learn to ignore it.
You used seven cheap gas sensors at once. Each sensor reacts a little differently to each gas, so together they give a pattern, like a fingerprint. You then trained a Support Vector Machine (SVM) to read that fingerprint and say which of four situations it is: no gas, perfume, smoke, or a mixture.
A plain straight-line SVM got 80.31% right. After putting all sensors on the same scale and choosing two settings carefully, the curved (RBF) SVM got 98.23% right on readings it had never seen. It decides in about 1 millisecond, while a new reading arrives only every 2 seconds. So it is fast enough for real-time use.
Every page opens with a box like this one. Read those boxes in order and you will have the whole project in about twenty minutes, without a single equation. The mathematics sits underneath for when you need it.
The project in one paragraph
Using the public MultimodalGasData set (6,400 readings from 7 MQ metal-oxide sensors, 4 balanced classes), you derived the soft-margin SVM from first principles (margin, Lagrangian dual, KKT conditions, RBF kernel, one-vs-one voting), then trained it. Standardisation plus a 5-fold cross-validated grid search chose \(C = 100\) and \(\gamma = 10\). The model reached 98.23% test accuracy, macro F1 0.9823, macro ROC-AUC 0.9986, 99.03% leak sensitivity and about 1.1 ms per prediction. A time-ordered test then showed the main limit honestly: sensor drift.
How to use this site
-
Learn the words first
Page 02 defines every key term in one or two sentences: margin, kernel, support vector, recall and so on. If a word on another page is unclear, it is there.
-
Rebuild the report in your own words
Pages 03–08 walk through the problem, the data, the SVM idea, the full derivation, the tuning and the results. Read them once, end to end.
-
Answer out loud
The Q&A pages (09–11) have a Quiz mode. Turn it on, read the question, answer aloud, then click to reveal. Each answer starts with one bold sentence: that is what you say first.
-
Run the model yourself
Page 13 runs real Python and scikit-learn inside the browser. It loads the same data, trains the same SVM and gets the same 98.23%. Being able to show that live is very convincing in a viva.
Where you actually stand
This is an applied mathematics + machine learning project on a public laboratory dataset. You did not build hardware and the gases are deodorant and incense smoke, not LPG or ammonia. Say this early and calmly. The method is the contribution; the data is the test bed. Page 11 prepares you for exactly these questions.
The facts, in case you are asked flatly
| Title | Real-Time Gas Leak Detection in Logistics and Manufacturing Environments Using Support Vector Machine on IoT Gas Sensor Data |
|---|---|
| Degree | M.S. in Mathematics (Applied Mathematics) — course code AMAT-527 |
| Institution | Department of Mathematics, University of Chittagong |
| Student | Md. Ashfaqur Rahman · ID 18203096 · Session MS 2023-2024 |
| Supervisor | Dr. Mohammad Jalal Ahammad, Professor |
| Submitted | 25 October 2026 |
| Length | 70 pages · 5 chapters · 22 figures · 19 tables · 34 numbered equations · 100 references |
| Data | MultimodalGasData (Narkhede et al., 2022), sensor CSV only, CC BY 4.0 |
| Tools | Python 3, scikit-learn 1.9.1 (LIBSVM), NumPy, pandas, Matplotlib, Seaborn |
| Hardware | Ordinary laptop, Intel Core i3-6006U, 8 GB RAM, no GPU |
One line to open with
"My project classifies gas conditions from a seven-sensor array with a Support Vector Machine. I derived the SVM mathematically, tuned an RBF kernel, and reached 98.23% test accuracy at about one millisecond per reading, so it works in real time. I also tested it on later readings and showed that sensor drift is its main limitation."