Predictive Maintenance and modern Machine Learning meets Pragmatism

With now 30 years in various positions in baggage handling and assignments in project in over 10 countries, I have always found time for research together with Universities. It was quite an experience to write academical papers or supervise students with their final projects. The research target was and is operational excellence in Baggage Handling Systems (BHS) applying modern technology. Mainly focusing to find solutions that are beneficial for operation and maintenance (O&M) focusing on predictive maintenance , machine learning and the data presentation in a way that is understandable for the staff that do the daily work.

The actual final project is about the efficiency increase for baggage transport trays. For readers that are not familiar with BHS technology, here a brief description about the baggage transport trays systems and the known problems. The easiest description is that a tray system looks like a roller-coaster. Bags are loaded onto trays and are transported on tracks. The advantage that such systems have is the speed of over 10 m/S as compared to belt conveyors which are limited to 2.5m/S. There are various tray systems on the market. They all have different technology. For instance linear motor driven carts on tracks. Here the carts have caster wheels. Twin belt conveyor systems with trays (no wheels). Twin belt conveyor with center Z-rail where trays are driven by belts and centered by 2 rollers. No matter the system, they are all mechanical robust and highly reliable if serviced well. To simplify the following let’s use the word ‘trays’ even if this is not correct as some system have carts with wheels.

Hub airports BHS, will soon be built with conveyors that all together have a length of over 80 km and operate with several ten thousand trays. The trays, no matter the system, are subject to wear and tear. Depending on the manufacturer, such trays can lose material and become as thin as paper and break. Others have bearing failure, start wobbling under high speed and break. Others have guide roller that wear over time. A good and well organised maintenance program is the key to keep such critical systems operational.

Typically tray maintenance is organised using a round robin principle. That means, following a sequence every tray is routed to a service station where a manual check is done. The service technician follows a procedure on what to do and what part to replace under certain conditions. The remaining risk is that there are trays in bad condition in the system that are scheduled for service. Depending on the type and manufacturer they may break, get blocked or derail.?

Typically such systems have a service station where checks and repairs are done. Depending on the size of the BHS, it can be that service is performed 24/7/ and 365 days/a. For many O&M organisations, the cost to service their fleet of trays following a round-robin principal is very costly. Furthermore, it cannot guarantee that there are trays in bad condition which are operating causing damages on the conveyor system itself.

This now sounds like a dilemma, right? The answer is a switch to predictive maintenance makes perfectly sense.

Why?

Such trays circulate in the system and with modern technology such as cameras, vibration and laser distance sensors, whereby data can be captured and used to make the right service calls. The target is to always service the tray with the worst condition in the system, first. This ensures that the remaining trays in operation are in good condition.

Regarding data capturing, camera systems are known to provide a good outcome. Image processing data, combined with vibration captured on the track, distanced captured from distance sensors?enables data scientists to develop machine learning models that deliver a prediction on every tray′s condition. With a sortation from worst to best condition, a round robin could be replaced to a data driven principle. Trays in good condition will remain in the system. Such systems were developed for the most common tray systems and are now in operation.

All good so far and a success story, right? It is in the nature of many organisations to be demanding. Here is where the title of this article ‘Predictive Maintenance meets Pragmatism’ kicks in.

Let me quickly explain some basic terms in machine learning. If you ask your data scientist about the accuracy of the prediction he will refer to a confusion matrix. The word ‘confusion’ may a bit?misleading as there is nothing really confusing on such a matrix. Essential is that the values of this matrix are used to calculate the model accuracy in %. Perfect would be 100%. In reality a good model would come close to 90% especially if the available data are from image processing , vibration, distance and other data captured on moving objects. Reflections, minor signs of wear or vibration generated from sources close by may influence the outcome. Prediction scores way under 90% are not fit for purpose.?

However, in simple terms, a confusion matrix shows where the predictions were right or wrong when comparing with the real condition. The terms used in the ML literature about confusion matrix seems to have its origin in ML applied in medicine. There ‘positive’ means ‘sick’ and ‘negative’ means healthy. However, and agreed this is maybe confusing, for technicians ‘positive’ and ‘negative’ have a different meaning when it becomes to the condition of parts.

What this all means for targeting operational excellence?

In short, nobody complains when the predictions are right. If the model predicts a part as defect or healthy and it fits with the reality. For the tray PdM solutions, the savings were significant as the predictions are in the high 90’s %. This means that trays in good condition remain in the system and the trays in bad condition, what are commonly in lower quantities, are assigned for service. The time waste inspecting tray in good condition disappears. This is duly noted by the commercial department confirming savings in efforts of 90+% concurrent with a reduction in spare parts of 70+%.

So what is the problem? All is great and quality is increased and cost reduced, right?

The answer is yes and no. And now please be careful as the terms used in ML may?be confusing if you don’t have a medical background! The ‘false-negative’, attention reflected on part condition are the parts in good condition but in reality the part was defect. In technical systems where the reduction of unplanned downtime is of the essence, this value needs to be very low. Best 0. In BHS unplanned downtime can be extremely costly and in worst case scenario, the baggage remains on the ground, causing all kinds of additional efforts and cost.

Interesting, right? But what to do?

Motivate the data scientist to tweak the model to eliminate the ‘false negative’. The impact of such efforts are typically that the ‘false positive’ counts raise.

Is this a problem?

Yes but maybe a minor one as the impact for this use case is an increase of trays that show up at the service station but which are in fact in good condition. Even if the number of ‘false positive’ are very low, the technician in charge of service will complain and he is right. Every tray that shows up at the service station in good condition is a waste of time.

How can we get a grip on this?

Operational excellence includes continues improvement and if we take this serious, an active dialogue between the staff at the shop floor and the management is key for success. In this case the solution was that the images obtained by the camera system were made visible on the dashboard. Having images of parts was very useful. It left the final call of trays that are predicted with bad condition to the technician.?The effort providing such images are relatively low. However, the acceptance, using digitalisation as a way to reach operational excellence within the service crew increased significant. Complaints, as well as time wasting, disappeared entirely.

Conclusion: Machine Learning models for complex applications with many different data types can deliver predictions in the high 90%. It needs to be well defined if the focus is on the ‘fails negative’ or ‘fails positive’ and that depends on the individual use case and O&M organization target. With the increasing number of use cases using image processing it is worth adding the image of the part in question to the dashboard. This will enable the technician in charge to further increase efficiency.?Especially for parts like rollers, belts, chains (just to list a few) where dust or reflections caused the image processing algorithm result in a the wrong prediction. As said, this false predictions, if the model is fit for purpose, are in low numbers but investigating them could be a huge effort. A final view on an image can limit the efforts needed and bring the service closer to operational excellence.

This means that the technology is available and that PdM has found its way into BHS. Furthermore, it is good to have students that are interested in final projects. That reminds me of when I was a student and PLC technology became rapidly common and replaced panels full of contactors. That’s now 3 decades ago but I still feel young. This seems to be a ‘true positive’ or a ‘false negative’.

Recommended article about PdM market growth:

https://iot-analytics.com/numbers-of-predictive-maintenance-vendors-surges/