Pushing the Boundaries of AI Through Constant Innovation
Adrian Kostrz, Global Innovation Manager at NTT DATA Business Solutions, leads a team that is working on teaching artificial intelligence to perceive smells, evaluate them and use the results to control processes. Making technology that can compete with natural senses is no easy task, but he is confident it’s started well thanks to a strict training program, and a good dose of coffee.
Since the very early days of IT, and as technology continues to advance, we are discovering new and exciting ways in which machines can assist and enhance human capabilities. Technical advancements are intended to support our abilities and ideally, make tasks easier to perform. But instead of being replaced, we can look forward to a future where machines can support us in various areas, ultimately benefiting both humans and technology.
This idea can be applied to our human senses and capabilities, such as sight, hearing and touch, as much as it can to physical tasks and processes. Artificial sensors today are far ahead of ours as humans: They can screen 12,000 beer bottles an hour, hear whether pumps are suffering from wear and tear and control CO2, humidity and temperature levels in vast agricultural facilities . But there are areas where we continue to have the advantage over technology.
When it comes to smell the situation is different. As humans, we are not known for being especially good at sniffing things out, but what we can do is use our nose and tongue to associate smells with objects, feelings and events in our memory. That is where the technical challenge lies.
Technology has yet to develop the skill to accurately record and interpret every scent it encounters. As the interpretation part happens in the brain, however, that is where artificial intelligence (AI) and the Internet of Things (IoT) come into play. This is where we saw an opportunity.
With the aim of successfully interpreting different smells through the use of AI, we set ourselves a challenging goal: To develop a solution that will provide process support, forming part of production quality control. We may still be some way from achieving a perfect smelling machine, but we are on an exciting path.
To do this, we set a target of harnessing the power of AI to adopt human senses, and thereby eliminate the human error element that we inevitably bring to any workflow.
Despite having well defined ingredients and established protocols, natural fluctuations can still occur in any process, resulting in highly qualified human testers making mistakes. Is it possible that AI can help overcome this situation and reduce errors?
We aimed to explore the possibility of using an AI nose that knows the ideal state, so we set up an AI smell experiment.
Various business applications would benefit from an “artificial nose” whose capabilities exceed that of a specialized gas sensor. A classic example that could solve an issue we all encounter from time to time: Should a public toilet be cleaned on a regular cycle or as required in order to make quality assurance more efficient?
In another case, a food producer may want to sniff out the actual expiry date of a batch rather than relying on strict best before dates, so that good food is not unnecessarily thrown out in the name of rigid safety regulations.
To train the AI’s regression model, we spent three days placing coffee samples under the AI nose and labeling them. There were three potential responses: “good coffee powder”, “bad coffee powder” (mixed with vinegar), and “not coffee powder”.
If the coffee meets the standard, the sample gets released as part of the process flow, recorded in the ERP system and placed into storage.
If the nose reports a quality issue, the ERP system submits a standardized quality notification and the coffee becomes blocked from further use. If the system detects something other than coffee powder, it will stop the process, trigger an alarm and notify a human to check it.
Despite the complexity involved in the creation and detection of smells, we successfully distinguished the samples using the trained AI.
However, there were some real challenges along the way. For example, fine tolerances were required when it came to the distance between the sensor and the object being measured. And when the room was ventilated, the detection rate fell.
The AI also turned out to have some surprisingly human qualities: The AI nose got “tired” and had to be recalibrated every few hours. And it didn’t like mornings – it typically achieved better hit rates later in the day.
SAP’s #Hack2Build competition provided the ideal scenario for a rigorous test of our AI nose. The technical basis was provided by SAP as part of their SAP BTP services, which we are accustomed to utilizing in the teaching of machine learning models to record and analyze data and deliver results.
The test process was called “incoming goods” and the nose was tasked with testing coffee samples. It used four sensors to measure different gas values, and the specific combination of these four values represents the “fingerprint” of a smell. The concept proved successful as we won the hackathon.
Quality Control: An AI nose can be used to ensure the freshness and quality of food products in supermarkets. It can detect odors that indicate spoilage or contamination, allowing supermarket staff to remove the affected products from the shelves and prevent them from being sold to customers.
Inventory Management and Food Waste Prevention: An AI nose can be used to monitor the inventory levels of fresh produce in supermarkets. It can detect the ripeness of fruits and vegetables and alert supermarket staff when they are reaching the end of their shelf life, allowing them to be sold or used before they go bad.
Customer Experience: A more pleasant shopping experience for customers can be created by detecting and neutralizing unpleasant odors in the store. Custom scents that enhance the shopping experience and create a welcoming atmosphere can also be created.
Quality Control: Quality and freshness of crops during harvesting and storage can be controlled more closely. Detection of specific odors released by crops when they are ripe or beginning to spoil can allow farmers to harvest or sell them at the optimal time.
Livestock Management: An AI nose can be used to monitor the health and wellbeing of livestock by detecting specific odors associated with disease or stress. This can help farmers identify and treat sick animals before the disease spreads.
Quality Control: The quality and safety of food products during production can be better ensured through the detection of specific odors released by food products when they are beginning to spoil or are contaminated.
Process Optimization: Inefficiencies or potential quality issues can be identified early on through the analysis of odors, allowing food production staff to make adjustments and improve and optimize the overall process.
Quality Control: Unearthing specific odors released by materials or chemicals, e.g. formaldehyde vapors, during production can allow furniture production staff to take action to prevent any safety issues or defects.
Consumer Testing: An AI nose could be used to test fragrances on consumers by analyzing the specific odors associated with different fragrances. This can help fragrance developers identify the scents that are most appealing to consumers or help in the development of personalized perfumes for individual customers.
Success also depends on connecting the platform to the production IT and ERP systems in order to trigger follow-up processes and ensure continuity in the process. This ensures that action is always taken when the nose sends back a notification on what it has sniffed out – whether that means adding salt to a soup, lowering the temperature or calling a human.
The IoT and AI can transform a business in many ways , but underpinning any efficient application is a solid logistics and management infrastructure.
Although “artificial noses” already exist, they are veritable specialists. From talking to various sensor manufacturers and scientists, we have established that you can develop dedicated sensors if you know which molecule or aroma you are looking for. However, it is not possible to pack thousands of different sensors into an artificial nose in order to pick up a wide spectrum of olfactory impressions.
We are now looking for new exciting use cases for a scalable smell detection.
Cover image: ©Andreas Kusy Photography
You prefer watching to reading? In this video Gherdì Glaser, Innovation Manager at NTT DATA Business Solutions, presents how the AI Nose works and how it can improve quality control.
Steve Niesman
aug 21, 2024
Thomas Nørmark
jun 12, 2024
Adrian Kostrz
maj 23, 2024
Adrian Kostrz
maj 08, 2024
Jonas Winther Kvist
maj 07, 2024
NTT DATA Business Solutions
mar 18, 2024
