E-nose that guesses a fragrance like Shazam

Smell of all the senses has a more mysterious way of accessing memories beyond the conscious mind. The idea of Oler is to identify fragrances around you.

When I started working on this peculiar project, I never had much expectations per se. I've had this idea from 2016 at campus after a friend of mine introduced me to the Raspberry Pi. I mean for most of the friends I talked to about this project it sounded like a gimmick. Like how on earth will a machine smell and later on tell what it is. I mean currently phones don’t have inbuilt gas sensors but will in very near updates.

I chose to employ the MQ-138, MQ-3, and AQ-3 gas sensors after determining the common elements of synthetic musks. Tests are conducted in a closed environment because this prototype is solely intended to serve as proof of concept.

The test

By measuring Air, versace eros, Dior Sauvage and Creed Aventus, a training data set is produced. Over 500 sample readings were obtained per test from each sample, which is typically taken for 8 minutes at one-second intervals, and the data is converted into CSV files. An additional column is manually added to the sample for classification purposes (i.e., Dior, Eros, Creed, Air). In MATLAB, the four datasets are imported and merged. With the use of trial and error, the value of k was found to be 5, and this data is utilized to build a k-nearest neighbor model. A Principal Component Analysis (PCA) is employed as an exploratory tool to confirm the model and the results, comparable to the 10-fold cross-validation.

Accuracy naturally suffers as a data set's variables are reduced, but the answer to dimensionality reduction is to trade a little accuracy for simplicity. Machine learning algorithms can analyze data much more quickly and easily with smaller data sets because there are less unnecessary factors to process.

We randomly rearrange the data imported using matlab

y = ingredients;
rng('default'); % for reproducibility
ix = random('unif',0,1,size(y))<0.30
y(ix) = NaN

The testing data was classified using the classification model developed from the training dataset, and the results show 95.2% accuracy.

PCA