How AWS used ML to help Amazon fulfillment centers reduce downtime by 70%

Over the years, Amazon customers have gotten used to — and have high expectations for — ultrafast delivery.

But it doesn’t happen by magic, of course. Instead, packages at the company’s hundreds of fulfillment centers traverse miles of conveyor and sorter systems every day, so Amazon needs its equipment to operate reliably if it hopes to deliver packages to customers quickly.

To take on this challenge, the retail leader has announced it uses Amazon Monitron, an end-to-end machine learning (ML) system to detect abnormal behavior in industrial machinery — that launched in December 2020 — to provide predictive maintenance.

Monitron includes:

• Sensors to capture vibration and temperature data.
• A gateway to securely transfer data to the AWS Cloud.
• A service that analyzes the data for abnormal machine patterns using machine learning.
• A companion mobile app to set up the devices and track potential failures in your machinery.

As a result, Amazon has reduced unplanned downtime at the fulfillment centers by nearly 70%, which helps deliver more customer orders on time.

Amazon Monitron solves real-world industrial problems

“One of the key things that Amazon does is they take technologies like machine learning and they apply them at scale to solve real world problems,” Vasi Philomin, VP of AI services at AWS, told VentureBeat. “That’s really what drew me to this company in the first place.”

According to Amazon, up to 80 engineers are responsible for maintaining the equipment at each fulfillment center. Before implementing Amazon Monitron, technicians walked around the site, taking readings and manually analyzing the measurements to determine the condition of the equipment, including ultrasound, thermography and oil analysis.

Unplanned downtime, the company notes, can be costly and delay customer deliveries. For example, if a critical sorter fails for three hours during the peak Christmas period, it can lead to the late delivery of more than 30,000 orders. 

Monitron receives automatic temperature and vibration measurements every hour, detecting potential failures within hours, compared with 4 weeks for the previous manual techniques. In the year and a half since the fulfillment centers began using it, they have helped avoid about 7,300 confirmed issues across 88 fulfillment center sites across the world, said Philomin.

Allowing technicians to use ML for predictive maintenance on-site

“We learned that the persona using this isn’t the developer, they’re technicians in those manufacturing sites,” he explained. With Monitron, the cost per sensor is $100 and they can be bought on Amazon.com. “So it’s disruptive in terms of the cost, and the setup is super-simple — it comes with an app on the phone that helps you get permission in five minutes. A technician can do it and doesn’t have to be an expert on any AI or even predictive maintenance.”

Finally, there is the machine learning piece: “The ML learns a customized behavior for every individual sensor that’s being installed, so it learns the default behavior for vibration and temperature for that part of the machine and is able to quickly figure out when there’s a deviation,” Philomin said. “All three of those aspects are really what makes Monitron very disruptive.”

Amazon plans to expand use of Monitron

According to Amazon Customer Fulfillment, the company originally anticipated that it would take about two years to realize cost savings to pay for implementing Monitron. But the company analyzed 25 live sites and calculated that it had saved enough money to achieve an ROI in under one year.

As a result, according to Amazon, it plans to scale the use of Monitron to new fulfillment centers across the North America, Europe and Asia Pacific regions. Amazon Customer Fulfillment also plans to fine-tune the thresholds that invoke alarms and expand into other areas like monitoring control equipment.

The bottom line, said Philomin, is about democratizing AI and ML.

“You can have technology that only caters to advanced machine learning guys — of course, we have multiple layers of the stack that are more focused on data scientists,” he said. “But if you truly want to democratize machine learning and put it into use every day, technology needs to become invisible. What matters is you fully understand the person that’s going to be using it and you build in such a way that that person can actually use it.”