Issues Creating Algorithm Using Adafruit BME688 in Conjunction with BME688 Development Kit

[sophmatheson]

Hi everyone! A few of my classmates and I have been working on a project involving the Bosch development kit and AI Studio. The project at hand is a smell detection between cleaning solutions (bleach and Lysol) and normal air using the BME688 development kit (8 sensors) for collecting specimens to train in AI Studio and tested using an Adafruit single BME688 to test the algorithms. The only algorithm that we got to be successful was when we sprayed the cleaning solution in a Tupperware box for training and testing, but this is not ideal for our application. When trying to spray the cleaning supplies on a table, most of the algorithms that we have created are unambiguously detecting high levels of cleaning solution when sampling normal air. We were able to mitigate this issue slightly by using more concentrated trials and able to detect upwards of 70% air when sampling normal air but there is minimal change when the sensor is exposed to Lysol. This algorithm did not work originally and only improved when we changed out the BME688 sensor, which brings about the concern that we’ve seen on some of the other forum posts.

We’ve read that training and testing should be done on the same development kit since the single BME688 could be at a different stage of aging/power consumption, but the scope of our project requires that we use the Adafruit single BME688. Could this be a cause of some of our issues? We know about the live testing in the BME mobile app, but is there another way to be able to test on the development kit by wiring it to a raspberry pi? Would this mess with the BSEC library or other bme dependencies?

Another concern of ours relates to the Tupperware scenario. Should the sensors be sensitive enough to handle a relative distance? Originally, we tried 50-minute samples of cleaning supplies sprayed on a table (using multiple development kits at varying distances, max 3ft) and then decreased to 10 minutes (using six development kits in a square surrounding the sprayed solution). We did this because we were concerned about the dissipation of the cleaning solutions and their samples being too similar to normal air. Is this more likely to be the cause of our issues?

We are open to all suggestions and thank you in advance!

[BSTRobin · Accepted solution]

Hi sophmatheson,

We checked the issue from customer, regarding this issue, it may be due to sensor-to-sensor deviation, however, the percentage BSEC output is a probability, not the proportion of test gas.
You can place the training items and sensors in the same sealed container and try more time to train. If they want, they can use the tested single sensor for training themselves which is worth noting that when using this method, the obtained data can be replaced in the format of the 'datablock', and training algorithm, which can avoid the mistake of sensor-to-sensor deviation.

Refer to the format, you can change all 'sensor index' to 0, and keep 'sensor id' from original, replace some valid output, others like 'Scanning enabled' etc. keep original.