Gas Detection and Classification

This repository contains information about Gas Detection Classification with Traditional Classification Machine Learning Techniques.

Relying on a single sensor for gas detection may not yield accurate results, as various real-world applications demand robust and reliable detection methods. Implementing a modern ML techniques also requires a lot computational resource, meaning not efficient for an emergency system. Thus, there is a need for developing efficient gas detection systems that utilize simple traditional classification machine learning techniques to enhance the accuracy and reliability of gas detection.

Dataset Information

• Source paper: "MultimodalGasData: Multimodal Dataset for Gas Detection and Classification"
• DOI: 10.17632/zkwgkjkjn9.2
• Dataset downloaded from: https://data.mendeley.com/datasets/zkwgkjkjn9/2 (accessed on 12/03/23)

Instrument Used

The dataset was generated using the following instruments:

1. Metal Oxide Semiconductor (MQ) sensors (consisting of 7 sensors, with each one sensitive to certain gases)
2. Seek Compact Thermal Imaging Camera (UW-AAA)

Model Stages

The dataset can be used for the following stages of machine learning:

1. Basic ML - Only using numerical data to classify the current environmental condition (Decision Tree, kNN) -> this is what we are planning to do for the moment.
2. Intermediate ML - Using OpenCV to add thermal camera images as a model feature (increasing accuracy on low concentration gas)
3. Adv ML - Implementing deep learning to create a self-learning model

Gas Sensors and Sensitive Gases

The following table lists the gas sensors used in the dataset and their corresponding sensitive gases:

Sensor	Sensitive Gas
MQ2	LPG, Butane, Methane, Smoke
MQ3	Smoke, Ethanol, Alcohol
MQ5	LPG, Natural Gas
MQ6	LPG, Butane
MQ7	Carbon Monoxide
MQ8	Hydrogen
MQ135	Air Quality (Smoke, Benzene)

Steps Block Diagram

This block diagram outlines the steps involved in our process, which include:

Data Preparation

Data Pre-processing

Model Training

Model Evaluation

Report Analysis

For a detailed analysis of our research report, please refer to the PDF file Report Analysis.pdf. This report provides an in-depth discussion of our data analysis, model evaluation, and conclusions. We encourage you to read this report to gain a better understanding of our methodology and results.

Predict Gas Type API (FastAPI)

Endpoint

POST /predict_gas

Description

This API endpoint accepts sensor values as input and returns the predicted gas type and its probability for both Random Forest and k-Nearest Neighbors models.

Request

Body

Parameter	Type	Description
MQ2	int	Sensor value for MQ2
MQ3	int	Sensor value for MQ3
MQ5	int	Sensor value for MQ5
MQ6	int	Sensor value for MQ6
MQ7	int	Sensor value for MQ7
MQ8	int	Sensor value for MQ8
MQ135	int	Sensor value for MQ135

Example

{
  "MQ2": 2202,
  "MQ3": 799,
  "MQ5": 529,
  "MQ6": 515,
  "MQ7": 507,
  "MQ8": 696,
  "MQ135": 768
}

Response

Body

Parameter	Type	Description
knn_prediction	int	Predicted gas type using k-Nearest Neighbors model
knn_probability	float	Probability of the predicted gas type (k-Nearest Neighbors)

Example

{
  "knn_prediction": 'Perfume',
  "knn_probability": 0.83
}

Deploy Model Locally

Prerequisites

• Make sure Python and Pip installed in your system (this project use Python 3.10.9 and Pip 23.0.1)
• Install the required packages from requirements.txt (you may use virtual environment for this)

pip install -r requirements.txt

Running FastAPI app locally

1. Open a terminal or command prompt
2. Navigate to the API project folder

cd Gas-Detection-Using-Traditional-ML/src/api

3. Start the FastAPI app by running the following command:

uvicorn api:app --reload

This will start the FastAPI server on http://localhost:8000.

Running Streamlit app locally

1. Open a new terminal or command prompt
2. Navigate to the Streamlit project folder

cd Gas-Detection-Using-Traditional-ML/src/streamlit

3. Start the Streamlit app by running the following command:

streamlit run src/streamlit/streamlit.py

This will start Streamlit app and open it in your default web browser. By default, Streamlit run on http://localhost:8051

Now, both FastAPI and Streamlit apps are running locally, and you can interact with the Streamlit app to test and make prediction using the FastAPI and our trained model.

Retraining Model

You can re-train the model by executing these python source code sequentially:

python src/data/preparation.py
python src/features/preprocessing.py
python src/models/model_training.py

These scripts will re-load the source dataset, pre-processed it and re-train models based on the newly loaded dataset. NB: You can check the hyperparameter tuning and model training logs at logs folder.

Docker Build and Run

Or if you want to use Docker engine, you can simply build a Docker image of this project and get it run with the following commands:

# make sure you are currently on the project folder
docker build -t awdawd gas-type-ML .
docker run -p 8000:8000 -p 8501:8501 gas-type-ML 

PS: tested on WSL2

Deployment Report

If you want to read my report about this ML model solution deployment process on FastAPI and Streamlit, you can this Medium's article