Training the BME688 e-Nose

The complexity of designing an electronic nose training process.

I have been working with a client to develop a detection and possibly a monitoring system based on the BME688 electronic nose. He initially collected samples from a lab. But he didn’t collect data from outside the lab. So our class samples and non-class samples, other differences including geographic locational differences. The training results had good accuracy, but we couldn’t be certain that it was really detecting based on the specific gasses we need to detect. It could have been picking up other environmental differences.

Since the device will be used to detect a contaminant in people’s homes, the client had to work out a way to go to several different homes. He contracted someone who would likely use this detector in his job. He is collecting data samples as he does his job. We decided to keep the information very simple and only collect “positive”, “negative”, and “maybe”. The data collection device has 2 buttons so we told him to press button 1 for positive, button 2 for negative, and don’t press a button for maybe.

Some observations regarding data collection:

1. The lab provided good “positive” data, but even the lab room air is likely to contain the smell. We discussed possibly collecting data from someplace else in the building. But we don’t know how far trace levels of the smell could go throughout the building. This would be good for the initial heater profile analysis, but we would need to classify the lab building air as “maybe” for the general device training.
2. Uses for the “maybe” samples. Most of our “maybe” data is collected in the same building as “positive” data. So the “maybe” can have trace levels of the smell. My client and I have discussed whether there is value for detecting “possibly in the building”, but for this use-case we really need yes/no. The “maybe” samples may help with training for trace levels. For now we will not be using the “maybe” samples. We reserve it for testing purposes.
3. The validation results of the training process provides a confusion matrix which includes the number of false positive and false negative results. We want to have as high an accuracy as possible, but we also need to consider whether it is better to fail toward the false positive or false negative.
4. We don’t have good control over how the data is collected. We are using a contractor who is collecting the data as a side job. The only indication we have is the button he presses. We are not collecting information about the severity of the contamination. Severity and proximity would be good for fine-tuning the training process.

Current Status (of this client’s project)

We now have a trained detection model based on real-world data. This can be loaded onto the BME development kit. Through the app interface, the device can be switched to “test” mode.

We are now starting to test the detection process.

Questions my client and I are still working on

Many questions remain regarding how sensitive the device will be. The detection can be further refined with additional data collection.

The data collection has not been well regulated. How can we improve the data collection process?

How long after the source of the smell is eliminated does it take to register as “negative”?

There are several processes that can be used to eliminate the contaminant. The process to eliminate the contaminant may leave additional smells that could interfere with the ability to detect recontamination. We will need to go through a complete secondary data collection and analysis process to evaluate these possibilities and integrate into the final product.

To develop the system into a full commercial product, we need to test the system in several different scenarios.

Our Service Offering

There were many issues we had to work through in the initial training especially since we had to find a good sample of the contaminant to do initial testing. My client found a lab where he could collect some samples. We used that to determine that the electronic nose would be able to detect. After that, we started the heater profile analysis process.

Starter Package

I have streamlined the initial analysis and heater profile analysis. We offer this analysis as a starter package. The results the client gets from completing the starter process:

• Testing and analysis to show whether the BME688 will work for them;
• Heater profile selection for the product training;
• Analysis of the different variables that need to be considered for final data collection;
• 2 function BME development kits that can be run in detection mode or used for ongoing data collection and refinement;
• Data collection plans (including options for additional services from us);
  • Education/support/tools for the client to perform everything;
  • Guidance for the client to complete data collection with our assistance in analysis;
  • Using our lab, we collect and perform all training and analysis;
• Recommedatin for detection devices or monitoring systems;
  • Additional hardware/software and services we can provide.

Data Analysis

After the starter package is complete, they can purchase additional support in continuing to perform data collection and detection system training. We can continue to work with them to refine their process.

E-nose Lab

We can evaluate on a case-by-case basis whether we could offer lab services to help collect data and test the scenarios.