Alpro protects its ‘lone workers’ with AI technologyAlpro’s plant in Wevelgem (Belgium) towers about 50 meters above the river Lys. Here, all of the soy beans for the Kettering and Wevelgem plants are processed (for soy drinks and other plant-based food products). To further protect the employees inside the tower from injuries, Alpro is exploring the use of machine learning-based fall detection, together with the experts at delaware.
Alone and helplessToday, the risk of a ‘lone worker’ getting hurt inside the Alpro plant and being left helpless is mitigated by automatic calls to their mobile DECT telephone every hour or so. “It has worked so far, but it’s just not good enough,” adds Stijn, whose main concern is that it could mean an injured employee will have to wait for help for a prolonged period of time before he/she gets noticed. The current system also triggers a number of false positives because workers don’t always hear the calls near noisy machines.
“These false alarms seriously harm people’s trust in the system,” says Stijn. “If it happens too often, the sense of urgency disappears. In a previous company, I witnessed how two firemen lost their lives because of inadequate safety systems, so this is indeed a very personal issue for me. We need to do better to make sure that every single employee can leave work safe and sound.”
Hacking workforce safetyWhen Alpro’s long-time IT partner delaware launched its annual DEL20 innovation campaign, Stijn took the opportunity to place worker’s safety front and center. Stijn: “We wanted to deploy all the innovative firepower available to develop a smart system that could ensure lone worker safety. With the help of delaware, we participated in the delaware hackathon: we challenged a group of IT students, supported by delaware experts, to come up with solutions during a 24-hour brainstorm session.”
The hackathon was an eye-opener for Stijn’s team. “Initially, we were convinced that this would be an IoT project, based on vibration sensors. But we soon realized that all the sensor technology we needed for fall detection is readily available in today’s consumer smartphones, in the form of accelerometers and 3-axis gyroscopes.”
Improving accuracy with machine learningIt’s one thing for a smartphone to detect a probable fall through an app. In the case of Alpro, however, the stakes were a lot higher. “There’s no room for error,” explains Stijn. “We need to achieve high levels of accuracy for this to work. For example, we don’t want the system to trigger an alarm every time an employee makes an unexpected movement or jumps. There needs to be a clear difference between a ‘real’ fall and what we’d consider ‘unsafe behavior’, like jumping a few flights of stairs.”
This realization made Alpro and delaware turn to machine learning. “That’s when this became a data project,” says Stijn. To collect the necessary data to get started, Alpro employees started throwing test dummies down the plant’s stairs, smartphone and all. After collecting about 100 data points, the model already achieved an accuracy of 88% in discerning a fall from a no-fall. In a multiclass context – with a ‘safety concern’ category in addition to the ‘fall’ and ‘no-fall’ categories – the accuracy remained at an impressive 84%. This number will surely go up as more data is collected.
Safety, compliance, and privacy
Stijn’s – and Alpro’s – long-term dream is to create a factory where safety is handled preventively, instead of reactively or even pro-actively. “There’s still a lot of potential that goes beyond compliance with existing safety regulations,” he adds. “With the right amount of high-quality data, we can effectively spot potential hazards before they occur.”
Of course, there is still the issue of employee privacy. “If people feel monitored all the time, they will resist the adoption of new safety technology. The last thing we want to do, is create a ‘big brother’ atmosphere on the work floor by tracking employees’ individual behavior. That’s why, with delaware’s help, we’re setting up a two-part system that effectively de-individualizes the data. In system A, data is stored for a very short time: just long enough to analyze whether an alarm should be triggered or not. System B anonymizes the data but keeps it for longer to be able to recognize certain patterns and improve the system’s predictive capabilities.”