Keeping the world's elevators running smoothly with machine learning and IoT
Chances are you've ridden in a ThyssenKrupp elevator.
The German-based firm looks after more than 1.1 million elevators worldwide and runs lifts in iconic buildings such as the One World Trade Center in New York and the 1263-foot CMA Tower in Riyadh, Saudi Arabia -- the country's tallest skyscraper.
Keeping those elevators running is a full-time business, and demand is increasing every year as new buildings spring up in the fast-growing nations of China and India.
"The elevator business is a maintenance operation. You install an elevator once and you maintain it for 100 years," said Dr Rory Smith, director of Strategic Development Americas for ThyssenKrupp.
"Our customers don't want us to say 'We know it broke'; they want 'We fixed it before it broke'. All they care about is never hearing about their elevator."
Making sure an elevator never breaks down requires a lot of data, and ThyssenKrupp turned its attention to the large amounts of untapped information it generates each day.
Since last year the company has been working with Microsoft to build a monitoring system that feeds data from its elevators and escalators to the Azure cloud platform, using Microsoft's Intelligent Systems Service to help capture that information and its Machine Learning service to make sense of it.
The aim is to develop a system that knows what repairs need to be carried out before anything breaks and which can advise engineers on what work needs doing during call-outs.
ThyssenKrupp has trialled a proof-of-concept version of the system on less than 50 elevators and plans to begin rolling it out in earnest this year, expecting to extend it to about 600,000 elevators and escalators in 12 months' time. Eventually the company intends to use the system to monitor about 60 percent of its elevators worldwide.
How the system works
Modern elevators -- those fitted in the past 30 years -- generally use multiple embedded computer systems to help run the lift. These systems generate an array of data, indicating when floor buttons are pressed, when the door opens and closes, how often the motors driving the lift are running and the weight being carried by the elevator car.
The lift's systems also generates error or event codes, which can be read by a maintenance engineer during the next routine service and help them work out what needs attention.
ThyssenKrupp plans to attach a device that collects these codes, alongside other data about the operation of the elevator, and sends them to the Azure platform every day.
By monitoring usage in this way ThyssenKrupp plans to target when and where it carries out maintenance. Rather than scheduling a routine service every x number of months, the frequency and nature of these services would instead be based on how each elevator is functioning. Keeping tabs on their workings will be the Azure machine learning service, which will monitor details such as how often a lift door opens or the energy expended to drive the elevator.
"We can really adjust our maintenance and learn just how much is needed. What are the tasks we need to perform on our next visit based upon the usage, the environment, the weather and the building," said Smith.
"Let's suppose we know the load of a car and the motor current; if we see the current going up at the same load over time, it means some friction element is increasing -- such as maybe a bearing going bad."
"With machine learning we're going to be given a list saying 'On your next maintenance these are the tasks that you should perform'," said Smith.
In addition, the rules used by the machine-learning service to determine when a service is needed, and what work should be done, will be automatically updated based on feedback from engineers. For example, a lift door might be scheduled for a service every 10,000 times it opens, but that rule could be altered if experience dictates that door generally needs attention every 5,000 times.
"For instance, elevators that are in harsh environments. I spent five years in Dubai, where you have the problem of sand getting into everything. Maybe you need to do more maintenance there than you do in a very clean environment."
An expert in your pocket
Smith hopes the system will help ThyssenKrupp handle the explosive growth in demand for elevators in China, India and other rapidly developing nations.
"In the US there are one million elevators currently operating. In China each year 500,000 elevators are added. So every two years in China the number of elevators installed and running equals the number that have been running for about 150 years in the US," he said.
"How do you keep up with that? The training problem is huge because you have to be training all these people that have no experience with elevators because there hasn't been an elevator industry of that magnitude."
"We're looking at this to help training, to control maintenance and keep up with the growth," said Smith.
ThyssenKrupp and Microsoft have built an expert system based on Azure Machine Learning that can tell an engineer what's likely to be wrong with an elevator during a service visit -- listing the four most probable causes of an error code and ways to test for each problem.
Feedback from engineers on what the actual problem was and how they fixed it will help the system learn and refine how it interprets such error codes and advises other engineers in future.
Challenges
Working out what information the company has available and what to do with it accounted for the lion's share of work on developing the system over the past year, according to Smith.
Beyond this work, the two companies have spent most of their time devising software to handle data flowing from its equipment.
"It's very complex because we have multiple types of elevators and escalators," Smith said.
"We've had to put a great deal of effort into building the rules engines and software to support this."
Working with Microsoft, the firm had to devise software to prioritise handling of elevator error codes so the system understood the best course of action.
First, the device gathering the data has been configured so any error codes that suggest an issue requiring urgent attention are immediately reported to the company .
However, the bulk of error codes will be sent once per day, and Smith said ThyssenKrupp and Microsoft had developed a "very complex" rules engine for interpreting those error codes -- one that takes into account both general and specific factors -- such as the country an elevator is based in. Creating that rules engine is "one of the biggest challenges we have", according to Smith.
The other main focus of ThyssenKrupp has been integrating data and systems on Microsoft's Azure cloud with its ERP systems -- Oracle in the US and SAP in Europe.
Although it's still early days for the project, Smith is hopeful that giving its engineers a pocket expert on its elevators and using targeted servicing schedules will realise ThyssenKrupp's goal of cutting breakdowns.
"Generally speaking, lift companies sell maintenance based upon 'Whatever happens, we fix it for you', every breakdown is paid for. All we have to do is eliminate one breakdown [per elevator] and this system pays for itself," Smith said.