Using Deep Neural Networks to calculate air pollution

We hired Dataspectrum to build a machine-learning-based solution for calculating gas concentration. They quickly grasped what our business problem were so communication was very pleasant and productive - we used our familiar domain language. All truly important ML-aspects were sanely translated to human language. At every step they were trying to solve our problem, and not just write code and get money. And they managed to find a solution - smart & concise.

Client: an environment monitoring devices manufacturer

R-NOX is an environmental monitoring company, developing devices, solutions and services for people, industries and smart cities.

Problem: having data from numerous raw sensors calculate single human-understandable air pollution indicator

Some sensors can’t tell what is the exact concentration of SO2 (for example) around. Same raw values (voltage) may stand for pretty different PPM (parts-per-notation - a human-understandable indicator). The reason is in the physics of gas sensors - they react on temperature, humidity, etc. In order to obtain PPM, a correction have to be made. But was is the formula for correction? That’s a tough question: scientists have to be hired and researches conducted. And even after that each particular sensor would have it’s own unique characteristics so calibration (search for unique initial parameters) would be required.

Industry: Environmental Monitoring

Location: USA, Belarus, Lithuania

Technologies: Python, Pandas, Jupyther, Tensorflow

Solution: machine learning

Machine learning is a technique for restoring (finding) relations between some observable values (X) and outcome (Y). And what’s important - no prior assumptions about that relationship have to be made. Linear? Exponential? Are inputs relevant? What are coefficients and parameters? - machine learning will find all the answers itself. Ideal fit to simultaneously solve both problems of sensors - correction formula and calibration. X=voltage, temperature, humidity; Y=PPM. To generate training set R-NOX installed their sensors at the local meteorological station where target PPM were available. After conducting many experiments with different topologies, hyperparameters and optimization methods we stopped at a certain variation of feed-forward deep neural network (DNN).


DNN performance overcame expectancies: it reacted precisely even on slight raw values changes.