Industries working with chemicals are always subject to leaks that could harm workers. Sometimes those leaks are formed by a specific combination of gases not suitable by off the shelf sensors but a machine learning model could be trained to identify subtle relationships between all the gas readings and calculations.

The device will read the multigas sensor x times for y seconds and then calculate min, max and average. Those values will be forwarded to the Machine Learning model for inference and then a score will be obtained for “harmful” gas 1 or “regular” gas 2.

Connections

Circuit schema will work as a reference. In detail you need to connect Buzzer to D2 and GND. Oled Screen VCC to Arduino 3.3v, GND to Arduino GND, SDA to A4, SCL to A5. Gravity Gas Sensor to Arduino 3.3v, GND, SDA to A4, SCL to A5

Data acquisition

If you are going to train a model to detect harmful gases take all the precautions like wearing a mask and ventilate the area. Several minutes are required to gather data and any person could become ill during that time.

Download aquisition script

Upload to the Arduino BLE 33 sense. Place the sensor unit close to the gas or substance and open serial monitor. You should see there multi gas sensor values as CSV records. Uncheck timestamp, then copy and paste the serial monitor screen into a text file. Save that file has harmful.csv and add this header:

timestamp CO2avg C2H5OHavg H2avg NH3avg CO2min C2H5OHmin H2min NH3min CO2max C2H5OHmax H2max NH3max

Repeat the procedure for all the gases to be detected.

Cut 20% of your data for testing and save into a different CSV file. This is useful to check later whether the machine learning model is actually doing good predictions. We know that all data came from the same gas. We train the model with some data and then we feed data not used to train the model just to make sure that those values are detected. 

Model training

First check that you have all the data for training and testing uploaded with correct labels. Then go to Impulse Design, Create Impulse. What is an Impulse? An impulse takes raw data, uses signal processing to extract features, and then uses a learning block to classify new data.

In Times Series Data, we will use 1500ms Windows Size and 0,6 frecuency. Window increase is not important here, since samples are taken at exactly X seconds.

Processing block will be raw data with all axis checked. For classification we will use Keras with 2 output features: regular and harmful.

In Raw Data you can see all values for regular and harmful inside every windows size. Then you have to click Generate Features.

For NN Classifier we will use 60 training cycles, 0.0005 Learning Rate, Validation 20 and Autobalance dataset. It worked for me adding an extra layer Droput Rate 0.1 Click Start Training and check if you get good accuracy. In my case 86% good harmful detection and 96% regular gas detection.

If you are ok with results you can go to Model Testing and check the performance with new data. If there are lots of readings with wrong classification you should check again data acquisition procedure. 

Using Bhopal 84

After displaying Bhopal and Edge Impulse logos, the unit will start the calibration phase. During this phase do not place any substance or gas under the sensor unit. It should read normal air conditions. As soon as this step is finished you can put the leaked substance under the sensor unit and it should be detected in 1.5 seconds. Why 1.5 seconds? 4 readings will be made to obtain mix, max and average for all gases. That information will be forwarded to the model and a classification will be returned.

The prototype is able to detect normal air conditions, a regular gas and a harmful gas. 

Demo