The baggage handling systems at airports are some of the most complex conveyor systems. They are the bloodstream of the airports, which makes them a vital component of airport operations.
The systems consist of thousands of moving parts that work together in a perfectly synchronised ecosystem that enables travellers’ baggage to arrive at the right destination on time. This is why maintenance of these baggage handling systems is a fundamental part of the life of the complex baggage conveyor system.
Lasse Vilhelmsen, a Former Data Scientist at Beumer Group, talked about how they maintain, and monitor baggage handling systems relying on predictive maintenance at the Maintenance Analytics Summit 2019.
Challenges with baggage handling systems at airports
One prevalent challenge with the system is its size, as we’ve mentioned. Lasse says that the conveying elements can spread more than 25 kilometres, including upward of 6,500 motors and around a million configured parameters. Considering the number of constituents and the size, we can imagine that there is much stuff that can go wrong in these baggage handling systems, explains Lasse. And as the elements function in a loop, if one element breaks down, it affects the entire line. Which is why predictive maintenance aims to keep the error rate of each element as low as possible to avoid forced shutdowns.
These systems are like lego bricks that you can stick together. They are made up of different kinds of elements which we want to work as impeccably as possible. What we see below is the simplest element.
The predictive maintenance goal
“When starting to work with predictive maintenance, it’s always good to take a step back and review what we have today and what we want to improve,” highlights Lasse.
The driver behind their predictive maintenance efforts is that they want to build high-quality products. As part of their vision, Beumer Group also includes a maintenance contract for the systems they build.
In terms of maintenance, Lasse relates that they do several kinds of maintenance:
- Planned/scheduled maintenance – scheduled inspections
- Condition-based maintenance – taking into account running hours, starts and stops
- Predictive maintenance – relying on the system data to tell them when to take actions.
Some metrics to consider
There are several metrics to consider when implementing a predictive maintenance model, Lasse states. Some of them are:
- Hit rate – which related the quality of predictions and how often they find an issue. However, they experience a low number of failures as the systems are quite reliable. So it’s basically looking for the needle in the haystack, as Lasse says, among 6,000 or more elements.
- Cost to install and operate – although the cost of a single element is not high, installing, e.g. a vibrational sensor of every one of them amounts to a considerable figure.
- The barrier to getting started – Having new installations, it’s easy to manage, but they also want to retrofit their large install base.
- Amount of direct insight provided – As they use several approaches, some of them give indirect insights, only showing that there is an anomaly and not how to proceed with it.
The three approaches to maintenance
Beumer Group have taken three different approaches to the maintenance of their baggage handling systems. Lasse gave an overview of all three and presented the advantages and disadvantages weighing them against the four metrics in all three stages of collecting, analysing and visualising the data.
Using existing data
The first approach relies on direct “hot-line” contracts with customers and VPN connection to their sites that allow them to log into customers’ systems and assess performance. They use this existent data for predictive maintenance instead of retrofitting the systems with sensors.
They opted for a local data pipeline instead of going to the cloud for the data collection process. The data from customers’ sites is streamed to their master data drive and driven to several tools for data processing and visualisation.
For the data analysis, they have taken a more model-based and physics-based approach where they count the number of alarms and timeouts of elements and look for changing trends in the stops. The system highlights the problematic element and a technician goes on the site to perform maintenance. Additionally, based on the logs, they can measure travel time between elements and monitor change in the histograms to identify elements that have growing slippage overtime.
They present the findings to service managers in a visualisation application that shows elements that require attention, taking into account data from the past day and past two months. Lasse says they are trying to work out how they can label the data based on technicians’ feedback which can serve to improve the model, so it also states the precise defect and not only trend changes.
A moving observer
The second approach is more hardware-based which monitors the data that passes through the system and gives insight into the performance. They build as Lasse names it, an “intelligent tray” which consists of accelerometers, Raspberry Pi, tag readers and photocells.
This tray moves around on the system at a speed of 2-3 meters per second and detects vibrations of elements. It collects accelerometer data and measures impact of sidecars, element misalignment and micro stops. These micro stops indicate that some elements are not up-to-speed, which is shown in speed change of the intelligent tray. They measure this speed changes and take actions to increase the lifetime of the system.
The results are visualised in a digital twin that represents each element in the system and real-time movement of the intelligent trays and baggage. The digital twin enables them to have a 3D visual representation of the kilometres-long system and make maintenance decisions based on it.
Individual vibrational sensors
The third approach also involves installing IoT hardware in the form of vibrational sensors on system elements. To be able to collect sensor data, Beumer Group created an infrastructure with an edge device – Raspberry Pi, that forwards the data from the sensors to the cloud.
The vibrational sensors generate FFTs, which are transmitted to the data channel for analysis. Doing data analytics on the FFT spectrum allows them to classify the data into three characteristic clusters corresponding to the three different operational modes of an element, and identify outliers or deviations.
For this presenting results, they are using the same visualisation platform with a digital twin to monitor the system and crucial components with a higher error rate.
Takeaways from implementing the three approaches to predictive maintenance
Summing up all three approaches, Lasse states that for some of them, they are already out of the PoC phase and have started offering them to customers. The main learnings he highlights are:
- The smallest changes can be observed, even from operational data.
- There is no silver bullet solution to monitor all systems, but they need a combination of different approaches.
- The business case is not always easy to establish. Instead, the business case is better established in reducing the maintenance costs with fewer replacements, inspections, and not because stops are avoided.