The Story Behind the Self-Learning Factory Concept

It is called SLF: Self-Learning Factory

• It is more than predictive maintenance, as we rely on machine-aging algorithms and machine utilization history (if you are always driving in the red rpm, then you will be notified every week to change the oil)
• Forget about any UI, simply follow the insights, which are useful, reliable and just in timeI
• Intelligent ETA for your orders, like a car navigation system with in-depth knowledge about everyday traffic jam patterns

The concept

It all started last year during the COVID pandemic. We did some brainstorming in our team and dug through 10 years of our implementations at factories, contrasting it all with the capabilities of the cloud technology we are using as the base of our LogiX platform. LogiX – which we’ve been developing since 2018 – is a solution for monitoring the production processes at factories and increasing their efficiency.

That is how we came up with the concept of extending the platform with intelligent recommendations about the ongoing production process and the condition of the machines being used. The recommendations are intended to minimize unplanned downtime and increase the number of manufactured goods, while also minimizing the energy consumption and making sure that the people involved in the production process are working in a hospitable environment.

We want LogiX to help factories become eco-factories that do not waste energy by idle work, but instead, we want them to work with maximum effectiveness. It means that every unit produced has lower energy cost while workers receive recommendations given “on a plate” which allows them to focus on their tasks rather than on looking for solutions to typical problems.

We want the implementation process of a platform with such intelligence and a self-learning recommendation system to be very simple for the end-user. Our goal is for users to simply connect their machines to our platform and immediately start benefiting from the self-learning system, without the time-consuming creation of rules, algorithms or predictive models.

We kept on talking to large companies involved in big projects related to predictive maintenance. For example, one of those companies was in the airline industry, and their project dealt with detecting possible engine failures when airplanes were parked in hangars. This meant collecting hundreds of measuring points, together with hundreds of thousands of measurements – and with a short time frame for making a go/no-go decision. Those projects were always huge and expensive, and when they were brought over to the manufacturing industry, they simply didn’t add up financially.

We reached the conclusion that a key measure would be building a library of machines, types of machines, and types of production processes. Next, we had to catalog it all and create failure prevention models. With that in place, it should be possible to connect such ready models to the definitions of machines in the library and, a crucial step, generalize them to a specific class of machine, not only to a single machine in the library. We approached the rules for detecting bottlenecks in the production process in a similar way: by cataloging them in the library of processes (for example, the packing process or the number of machines connected on the production line).

Readily available! Prepared! Waiting to be used in your factory.

Using a cycling analogy, it is just as if you wanted to create a model showing the wear on a bicycle that depended on the total distance the bike traveled and the type of road surface (as tracked by a GPS system during training sessions) as well as the weather when the bicycle was being used. Then you would apply that model to specific bicycles built by various manufacturers and belonging to various sub-categories of bicycles (for example, MTB, road, or gravel bikes). It only requires some tuning up and it will work just fine. After all, we all know that riding in rain and mud has a much more “destructive” impact on your bike than when you do it in the sun along an even road.

Using such a ready-made model, you could receive recommendations on when you should replace the chain on your bicycle, because otherwise the chain may break, or the level of wear may negatively impact the cogs in the drivetrain. You could also receive suggestions on how to use your bike to make it last longer without needing to service it so often. This could be done without actually measuring the level of wear (by measuring the size of the links in the chain, for example), but estimating it based on the wear model (failure prediction).

Our goals

We want to be able to:

• Determine the aging of a machine’s functions and their elements (the health factor), so that any servicing can be planned for the time after the wear warning point has been reached (when it may start to impact the quality of products). The aging functions will also show how to use such a machine to increase failure-free work between servicing sessions, which will be done at the optimal time, i.e., not too early (as it increases servicing costs) and not too late (as you risk unplanned downtimes).
• Establish failure prediction models based on anomaly detection in the process parameters (process time, energy consumption and temperature) in order to warn the user that there is a possibility of failure.
• Establish simulation models in order to provide the user with suggestions to adjust the parameters when atypical problems arise (for example, problems with the speed at which fillers are moved up and bottles positioned in fillers: if done too fast, foam starts flowing out of bottles, flooding the conveyor system and disrupting the approaching bottles).
• Define what data we collect for a specific machine or class of machines.
• Define what states we use to describe a machine from the monitoring viewpoint.

This all is completed with a supporting vision system that is set to automatically recognize problems (for example, it notices when bottles are knocked over and start blocking the conveyor).

All those features, once defined, set, and linked to a machine definition in the machine library, are readily available for you to use on the LogiX platform to which you connect your factory.

We are doing everything we can to ensure that our clients can easily connect their factories to LogiX to start using the self-learned system and benefit from the recommendations it provides. It’s as easy as using a car navigation system: you simply enter the destination and follow the directions (“turn left”, “turn right” and so on). At a factory, those directions will be recommendations on what you should do to manage production in the most efficient way, which means avoiding unplanned downtime, optimally planning servicing sessions, and reducing energy consumption. The LogiX suggestions are so trustworthy that it’s enough for operators and shift foremen to follow them and “drive” the optimal route.

Everything is done without frantically clicking from screen to screen, digging through piles of reports or looking for vaguely described clues.

Quick and easy configuration

It will take you 10 minutes to configure a ready-to-use monitoring system with anomaly detection and failure prediction – a system that can also recognize hidden causes of downtimes on your production lines.

• Without data scientists.
• Without months of collecting data.
• Without model learning.

Simply sit back, fire up LogiX and log in to it just as you log in to your email account. Open the digital version of your factory and drag a machine from the machine library, for example, a machine called: Unilogo Futureproof 120. If you can’t find it in the library, it’s not a problem. Just select an appropriate machine category.

Once the machine has been selected, drag your production process – for example, packaging – and you will see a complete chain of connected machines: a filler connected to a capper; a cap sorter to a capper; a capper to a labeler, and so on.

Your Factory’s Digital Twin is ready to work and support you in your decisions by giving you intelligent recommendations.

All this should take no more than 10 minutes and offers the full potential of the AI that we built into the digital twins of your machines and processes in our library.

The suggestions that generated for you will include warnings about specific failures, changes in machine setpoints, finding connections between human actions and problems with machines, determining non-obvious causes of problems (for example, problems resulting from servicing sessions in the past), and will automatically find root causes based on the graph of interconnected machines.

Some examples of recommendation that you may receive:

• Replace the capper element that holds caps because, based on the time of the process, we detected that it must be seriously worn.
• This week, John has been taking, on average, 8 minutes longer than others to deal with this kind of failure. Maybe he is overworked or needs additional training.
• We noticed that you have been experiencing downtime due to several failures of a lid press that occurred after John serviced the machine. Perhaps you could offer John additional training.
• If you reduce the speed at which fillers are lifted, you will avoid downtime resulting from the foam overflowing onto the conveyor.
• If you lower the temperature by 10 degrees while mixing the cosmetic mass, you will save up to 10% of energy while retaining the same quality of products.
• The manufacturing order on Production Line 1 will be finished in 6 hours and 30 minutes. We have calculated this time based on the historical data from the past 6 months related to the manufacturing of such products.
• You are again experiencing a problem with cartons for the case packer from one of your suppliers. Talk to them, because you are losing 50% of productivity because of this.
• Reduce the operational speed of one of the robots by 5%. This will lower the number of incidental holdups by 20%.
• The supporting vision system has notified us that many problems related to conveyor blockages result from bottles being knocked over because they are not held in a stable position.

With our project defined the way set out above, we went to The National Centre for Research and Development (NCBR) and we were granted EUR 2 million in funding. We simply fascinated the NCBR experts with our project.

However, to develop it further, we are in need of partners:

• Factories manufacturing machines, because they can help us define their machines in our library (which is how we’ve been collaborating with Unilogo, for example).
• Factories manufacturing goods, because we need to catalog manufacturing processes, identify bottlenecks, and establish rules of recognizing them and warning about them. We also need to prepare definitions of machines in our library and to try them out in action.

If you are thinking of becoming a LogiX partner, please contact us at slf@ilabo.com

How does the partner program work?

We will connect your machines to LogiX – our execution platform.

Together, we will decide which of those interconnected machines should be included in the project and in the partner program.

This will allow us to figure out:

• Prediction models
• Anomaly detection
• Algorithms that evaluate the health of machines
• Determining hidden root causes

We will provide all this for your LogiX instance and for your machines for free, though we reserve the right to generalize the results and include them in our library (of course, the data will be collected anonymously, which we guarantee in the partners’ agreement).

Ilabo Sp. z o.o since 01.09.2020 r. has been implementing a project titled Self Learning Factory – Recommendation Engine built based on artificial intelligence with video based production events recognition and utilizing global production machines libraries.

Project aims to create an innovative software that works in a cloud and which will allow to automatically generate hints to maximize the effectiveness of manufacturing processes. As a supporting tool, Ilabo will build an innovative hardware-programmed system for visual event recognition and a digital library – a catalog of machinery and production equipment with built-in predictive models.

Project’s main result will be an innovative recommendation engine – LogiX.RE which Ilabo will use as a part of LogiX platform operating in the Microsoft AZURE cloud. In this way Ilabo’s platform will support intelligent hints that aim to eliminate inefficiency, start corrective actions before the breakdown occurs (breakdown prediction model) or change the machine set-up to eliminate problems (based on simulation models).

Project value: 11 252 962,26 zł
The contribution of European Funds: 7 817 291,53 zł