We have been experimenting with the app board 3.1 + bme690 x8 board the past few weeks but we can’t manage to get satisfactory results and we don’t know why. We were hoping you could help us.
Our goal is to try to train the sensor with the least possible consumption profile since our intended final product will be a battery operated Iot device, in a room, which needs to detect mild smoke from cigarettes and nothing else. My question is, what is the best approach training-wise?
We have tried a combination of different heater profiles and duty cycles but in the end, we cannot get consistent results. At our office, under certain circumstances, we get reliable results but when we test for smoke at a different site it has a very hard time detecting the smoke. In another training session, although it reliably detects smoke in the beginning, after a while of running, it starts to have false detections in a perfectly clean room.
Please give us an analytical training procedure and please answer the following questions.
1) In the beginning, we used a 3D printer chamber as the recording environment for the smoke specimen. But we then assumed that it may be too rich in smoke concentration for our case. What we actually want is, just like the human nose, to detect in small concentrations that someone has been smoking in a room. So, our question is, is it better to perform a couple of recordings (in the same class) in dense smoke and then train using a combination of classification and regression algorithms or should we record in actual conditions, i.e. a room with little smoke?
2) How long should we record in smoke and how long in clean air given the default config profile (HP-354, RDC-5-10)?
3) How many samples of each specimen are enough?
4) When training, should I use all 4 channels, some or only the gas data channel?
5) Since our intended application needs to have a narrow spectrum of cigarette smoke only, would it help to have multiple brands of cigarettes in the same class? It would not be practical to train for every cigarette brand and mildness level in the market, do you agree?
6) If we wanted, for example, to exclude smokes like cooking smoke, would we have to train the sensor with a couple specimens from cooking smoke and then add it to the clean air class during training?
7) Finally, during the live-test, even though all sensors have been trained with the same heater profiles and duty cycles, we do not see unanimous results. Sometimes some give false results, sometimes some have a hard time detecting and some flicker a lot. Is this normal?
Waiting for your response,
Peter