relayr optimizes elevator maintenance with Azure Machine Learning and IoT

The challenge: Cut the time and costs associated with elevator maintenance

A trailblazer in the field of digital transformation and the Industrial Internet of Things (IIoT), relayr has relied on the power of the cloud to collect and analyze sensor data from connected devices since it was founded in 2013. Today, the former start-up employs around 300 people, including 150 engineers and more than ten data scientists. Around 50,000 connected devices from an assortment of customer businesses transmit approximately 100 million pieces of sensor data to the company every day. This data not only facilitates the continuous monitoring of IIoT environments; it is also used by relayr’s data scientists to train and improve machine learning (ML) models, and ultimately increase the precision of AI analyses and predictions.

But with the volume of orders and data on the rise, the company’s original cloud architecture had reached its limit: the original concept behind relayr was to offer microservices that were written in-house and didn’t rely on a particular provider. “This isn’t sustainable for a company of our size in the long run,” says Dr. Nico Wintergerst, Staff AI Research Engineer and member of the CTO office responsible for multi-project analytics and AI tasks at relayr. It meant the company’s data scientists were left writing huge amounts of code for their self‑developed microservices, besides having to compile the ML models on their notebooks themselves. “Our CTO office and the data scientists were agreed that we had to move away from this approach,” Wintergerst says. “It was taking too long to arrive at the first value creation point in a project. We need scalable solutions, and we have to learn to work more efficiently.”

The solution: Switch to Microsoft Azure services

In 2020, the time had come for a paradigm shift at relayr. The company introduced Microsoft Azure as its new cloud platform and switched from its own microservices to Microsoft’s broad portfolio of Azure AI and ML services: Azure IoT Hub, Event Hubs, Data Lake Storage Gen 2, Azure Machine Learning, Azure Kubernetes Service (AKS), and Azure Stream Analytics. The project team chose to focus on relayr’s Franz software, a cloud solution that monitors the condition of elevators, as the first use case for its new architecture.

At the start of 2021, the project team migrated all the relevant microservices to Azure, before switching them to Azure Services in June. They also set up a data lake for historical data. “We were already using S3 repositories, which are cloud-based object stores,” Wintergerst says, “but we didn’t have a strategy for ‘cold’, or rarely used, data.” It took just one month to draw up an initial proof of concept for the new architecture, and the first production pipeline was ready after two months. “Deployment is quicker than it ever could have been with our old technology,” the engineer says.

Today, relayr takes a slightly different approach to the remote monitoring of elevators. A Franz box installed on the elevator roof continually collects measurement data on aspects such as location, acceleration, speed, vibrational spectrum, magnetic fields, or even sensor data from the door motor. This raw data is aggregated and processed directly in the Franz box. The box documents each ride, recording parameters such as distance and acceleration, besides top and average speeds. If anything appears to be out of place, the box sends the raw data to the cloud for ad-hoc analysis. If, for instance, the door opens faster or slower than normal, maintenance engineers can identify remotely whether this is always the case, or whether it happens just on one floor. If the door opens too slowly on every floor, the motor might need servicing. If it only happens on one floor, then cleaning the door’s guide rail may solve the problem. The Franz box is also able to record new measurement data even after installation: its modular design allows additional sensors to be incorporated as and when required.

The cloud’s role: Determine normal behavior with machine learning

But how is abnormal behavior defined? This is where ML algorithms and the cloud really come into their own. relayr uses an Azure Data Lake to store the necessary operating data separately for each project. It then uses this data to train the models in Azure ML. The algorithms can identify the areas (clusters) where sensor data ought to be located to fall within the normal range. relayr runs the trained models via Azure Kubernetes Service (AKS), and the data is accessed through an Event Hub and visualized interactively with Power BI. The data scientists and domain experts at relayr can look at potentially unusual operating data flagged during automated processing before uploading their verified findings to monitoring dashboards used by the maintenance companies.

The maintenance companies benefit from relayr’s solution in a number of ways. First and most importantly, they can view relevant information straight away without having to be on‑site. Second, they have a central access point for all the elevators they manage, with statistics on the number of rides and other important parameters. Third, they are alerted to deviations, for example when an elevator no longer reaches its top speed. Fourth, they can improve their day-to-day planning by optimizing service routes and, in turn, cutting costs. Fifth, they can increase customer satisfaction by ensuring more effective elevator maintenance.

relayr is already looking at how it can go beyond the scope of condition monitoring and condition alerting and use the dashboard to offer what Wintergerst refers to as the “Holy Grail” of remote monitoring: predictive maintenance with automated, AI-assisted error analysis. 

Working with Azure: Elevators reach new heights of productivity

While a quest for the “Holy Grail” is never going to be easy, relayr was able to get off to a flying start in terms of productivity. “Switching to Azure was a game changer for our data scientists,” Wintergerst says. “It lets us develop and test new models much faster and directly in the cloud, and we can train them with relevant data from the get-go. That allows us to quickly introduce solutions to the cloud that are specifically tailored to the needs of a given project.” This is something the company would have needed several months to do in the past. “We’re now able to get a functioning model with relevant insights up and running in just a couple of weeks thanks to Azure Machine Learning,” Wintergerst says. “We’ve even managed to produce verified models in just four to six weeks.” Wintergerst has received consistently positive feedback: “Our data scientists are extremely satisfied,” he says. “They can release and oversee their models in the cloud themselves without much time or effort.”

The standardized reference architecture that relayr developed for Franz can now be transferred to other projects—to recognize anomalies in emission control systems on ships, or to eliminate sources of error in switchgear assemblies during power grid operation, to give two current examples. At the same time, relayr is broadening its range of services: “We’re expanding our Azure reference architectures to encompass the entire life cycle of connected devices and systems,” Wintergerst says. Together with device manufacturers, relayr wants to enhance condition monitoring to also include billing data, giving companies the option to offer usage-based billing. The goal: equipment as a service (EaaS). According to Wintergerst, cooperations are already in place here. With Azure providing the foundation, not only is relayr set for success in terms of productivity; it is also ready to take its portfolio to all‑new heights.