The fascination of creating digital products is strongly connected to the computational power of today's computers.
With their immense power to process large amounts of data in a short time, we have powerful tools and ways to design digital products.
Validation results from an automated neural net hyperparameter optimization
Algorithms are like a recipe for the computer to follow in order to achieve certain goals. These goals can be manyfold, like detecting a fire from gas data or using vital data for monitoring personal well-being and consumer health application. To develop these algorithms, there are two opposing approaches...
Bottom-Up Approach
Breaking down complex problems into their most basic and fundamental aspects – very often down to the physics – helps us understand the core principles of how a systems works (first principle thinking). Based on this understanding, we can explain how the signals in the data are formed and exactly how individual features in the data are to be interpreted (e.g. the specific aspects in the shape of an EV charging curve can tell you a lot about the battery health). Data analytics and algorithms can benefit from such fundamental understanding. Mathematical tools help us to extract useful information, that is hidden within the data.
Top-Down Approach
However, sometimes the mechanisms of cause and effect are very complex and it can be difficult to gain a basic understanding. A complementary approach, that is based on a large amount of labelled data, uses generic machine learning architectures (often called AI) that can be trained with such data. The system will then learn to mimic a certain intended behavior. Surprisingly even very complex behavior can be achieved with such an approach. The key to successful development of AI algorithms is to carefully train those algorithms in a way that they generalize very well, so their application in a product is robust and reliable.
We have multi-year, multi-domain and multi-project experience in collecting and curating relevant data, analyzing and understanding the data, and developing algorithms — using both bottom-up and top-down approaches.
Overheat and fire detection on multiple complexity levels.
In the case of our algorithm development for the Bosch Sensortec BME688 gas sensor we use both bottom-up and top-down approaches. We develop conventional algorithms for air monitoring and more complex machine learning approaches for specific fire detection.
Handling reliability was key to the development of these security related use cases:
- Overheat protection for vans, campers and boats (based on conventional algorithms from temperature and humidity sensors).
- Fire detection algorithms with an advanced threshold mechanism that can detect sudden changes in the air composition.
- A detection for smouldering fires based on a neural network that would detect carbon monoxide at concentrations that are way below any danger for people.
12-dimensional neural net classfication results projected into 3-dimensional space
It was important that we designed the user interface of BME AI-Studio in an intuitive way, so users can easily create and configure their own AI algorithms.
If you don't know what your customers can do with it, empower them to find out on their own.
Together with Bosch Sensortec we are investigating the possibilities of AI-driven gas sensors since 2019. And it turned out, there are simply too many interessting use-cases that can be explored. That's why we transfered our process and knowledge on how to explore gas sensor use cases into a software application, that lets customers build their own algorithms in a simple to use and easily accessible way. A challenge from both design and computational perspectives – harmonizing multiple years of gas sensor experience and various data science approaches into a coherent hardware/software ecosystem: BME AI-Studio.
We stripped down our complex machine learning pipeline in a way that even non data scientists can make use of it. Therefore, we needed to ...
- ... generalize and abstract the process of developing AI-based gas sensor algorithms, since initially we did not know which type of use cases would be relevant for users.
- ... find suitable presets and defaults that work well for the every type of use case.
Consumer health applications as a first step into fully digitized health care systems.
Providing widely available access for healthcare services through remote monitoring and automated data analytics is key for tomorrow's health care systems. The aging global population and rising healthcare costs necessitate efficient, accessible, and cost-effective healthcare solutions, which digital technologies can provide. Consumer health products are an important step, and in close collaboration with Bosch Sensortec, Universitätsklinikum Freiburg (Medical Center Freiburg), Saarland University and others, this is a constant effort at Intervall. In addition, the collected data and knowledge can later lead into medical grade products.
For example, monitoring and analyzing indoor air quality can have multiple benefits:
- Bad air quality warnings connected with notifications for necessary venting, resulting in better living and working environments.
- Analyzing the micro dynamics of indoor situations can provide useful information for the spread of viral infections. Humidity, temperature, activity tracking and the detection of specific volatile organic compounds can help estimate the lifetime and mobility of virus-containing aerosols and thus predicting the risk of viral infections.
Sleep tracking is another field, where we have been working for years now to detect and optimize parameters for a healthy night. Therefore we like to collaborate internationally in public funded projects.
With a built-in gas sensor, the smartphone can be used on the bedside table for AI-assisted sleep monitoring.
Distributed temperature, humidity and gas sensors within office buildings enable air quality monitoring for a better working environment.
Enriching people's lives through digital products is now enabled by the power of ubiquitous computing. At Intervall, we use algorithms based on first-principle understanding and AI-powered technology in a variety of areas, such as charging infrastructure, smart home, personal wellbeing and many more.