Information
Table Of Contents
- A. Overview
- 1. Getting started
- 1.1 The BME688 Gas Sensor
- 1.2 Example: Coffee vs. Normal Air
- 1.3 A Few Things To Keep In Mind
- 1.4 Step 1: Record Normal Air
- 1.5 Step 2: Record Espresso Coffee
- 1.6 Step 3: Record Normal Air Again
- 1.7 Step 4: Record Filter Coffee
- 1.8 Step 5: Import & Label The Data
- 1.9 Step 6: Create New Algorithm and Classes
- 1.10 Step 7: Train And Evaluate The Algorithm
- 1.11 Step 8: Export The Algorithm
- 1.12 Conclusion
- 2. Introduction
- 2.1 What is it about? – An analogy
- 2.2 Why the BME688?
- 2.3 What is a use case for a gas sensor?
- 2.4 What is special about the BME688 gas sensor?
- 2.5 How can I evaluate BME688 performance for a specific use case?
- 2.6 How can I use the results for my product development?
- 3. Glossary
- 3.1 Sensor Board
- 3.2 Measurement Session
- 3.3 Algorithm
- B. Process Steps
- 1. Configure Board
- 1.1 Overview
- 1.2 Board Type
- 1.3 Board Mode
- 1.4 Heater Profile
- 1.5 Duty Cycle
- 1.6 Board Layout
- 2. Record Data
- 2.1 Overview
- 2.2 Start recording
- 2.3 During recording
- 2.4 End recording
- 3. Import Data
- 3.1 Overview
- 3.2 Data Overview
- 3.3 Board ID
- 3.4 Board Type
- 3.5 Board Mode
- 3.6 Session Name
- 3.7 Session Date
- 3.8 Specimen Data
- 4. Collect Specimens
- 4.1 Overview
- 4.2 Label
- 4.3 Comment
- 4.4 Session
- 4.5 Start & End Time
- 4.6 Duration
- 4.7 Cycles Total
- 4.8 Cycles Dropped
- 4.9 Remaining Cycles
- 4.10 Board Configuration
- 4.11 Board ID
- 4.12 Board Type
- 4.13 Board Mode
- 4.14 Show Configuration
- 5. Train Algorithms
- 5.1 Overview
- 5.2 Name
- 5.3 Created
- 5.4 Classes
- 5.5 Class Name & Color
- 5.6 Common Data
- 5.7 Data Balance
- 5.8 Data Channels
- 5.9 Neural Net
- 5.10 Training Method
- 5.11 Max. Training Rounds
- 5.12 Data Splitting
- 6. Evaluate Algorithms
- 6.1 Overview
- 6.2 Confusion Matrix
- 6.3 Accuracy
- 6.4 Macro-averaged F1 Score
- 6.5 Macro-averaged False Positive Rate
- 6.6 Training Data
- 6.7 Test Data
- 6.8 Additional Testing
- 2.1
Bosch Sensortec | BME AI-Studio Documentation 42 | 49
Modifications reserved | Data subject to change
without notice Document number: BST-BME688-AN001-00
5.7 Data Balance
This indicates how the data for your algorithm training is distributed over your classes. Ideally, the distribution should be
balanced. If there is a little checkmark, everything is OK and the data is balanced enough for training.
How data balance is calculated
Each class consists of multiple specimens and each specimen has a duration. The application checks if the total duration
(sum of all specimen durations) of any of the classes is below the following threshold:
threshold = 24% / number of classes
E.g.
5.8 Data Channels
Specimen data includes four data channels. You can choose which data channel of each specimen should be part of
the training:
Gas Data Channel (10 data points)
Humidity Data Channel (1 data point)
Temperature Data Channel (1 data point)
Barometric Pressure Data Channel (1 data point)
By default, only the gas data channel is used for training. If other channels play a key role in your use case, you can try
to include additional channels in the training and compare the training results.
Please note
Be careful with using additional channels for your training. Using additional data channels does not automatically
mean better training results. The algorithm might focus only on one of the additional data channels during training,
which might not be what you actually want. E.g. if one of your recorded specimens has a strong impact on humidity,
the algorithm might only focus on the correlation between the specimen and the corresponding rise in humidity,
ignoring all other data. Your training results may look very good, but the algorithm only “learned” to distinguish your
specimen by looking at the humidity, completely ignoring the respective gas data.
threshold for 2 classes: 12%
threshold for 3 classes: 8%
threshold for 4 classes: 6%