False positives from BSEC

[Gatis]

Hi!

Our company is preliminary researching the opportunity to use BME688 for recognition of cigarette smoke and perfume (two classes: Smoke and Perfume). We made recordings of signals for classification task in laboratory using the Bosch BME688 Gas Sensor Developer Kit and BME688 Software, exported the trained AI configuration data and performed the classification AI training validation using Adafruit ESP32-S3 TFT Feather and Adafruit BME688. During validation in laboratory we found that in time fading perfume switches the classification result (100%) from Perfume to Smoke and vice versa resulting in large count of False Positives (far from 5%) for both classes.

What we could do in order to improve the classification task accuracy?

[BST_Yubin]

Need captue data and training the sensor before using it

[Valters]

Hi!

We are experiencing similar issue.

We did the data gathering in laboratory environment for cigaratte smoke, vape smoke, fresh air and paper tissue with perfume on it.

We trained the model for two classes, where Vape and cigarette was marked as smoking and the rest as NOT-Smoking. Also the tissue with perfume (high VOC concentration) was marked as NOT smoking.

We split the data 70 - 30 for training and validation data. One of the four thermal profiles used ended up with 99.47% F1 Score.

Then we exported the model to BSEC library and used it on Adafruit ESP32 board + BME688 breakout. Did the tests in the same exact laboratory environment and used the same specimens. Recognized vape and cigarette precisely, however, classified paper tissue with perfume as smoking (false positive).

The question we can't figure out is why there is such difference between our validation data and real life experiment given that they have basically the same environment and specimens?

Thanks for assistance!