The Maintenance Manager should be able to keep all the machines under her or his responsibility running smoothly. The effort covers the domain that ranges from making sure that engineers arrive on the site in time, carrying the right tools and replacement parts when a unit breaks down, up to creating and maintaining a smooth program of preventive maintenance, resulting in a 100% availability of the line. This job comes with a number of specific challenges:
- How can I predict that a machine is going to fail?
- What is happening when I’m not around?
- Are the ‘production’ people not abusing the machine?
- Can I know from remote what is broken, such that we don’t have to wait to start troubleshooting till when we arrive on the spot?
- Do I know what causes the most common breakdowns without having to spend 3 days on a pareto analysis?
What you read in blogs, magazines and online advertisements is that Neural Networks are a great tool to help you with coming up with an answer for all of these issues. When facing such a question today, you however don’t have a big Data Science team at your disposal, just sitting about and waiting to be given a data problem to crunch: your team is busy performing the actual maintenance or fixing a machine. What isn’t going to help either in providing an answer today is following a 1-year course on Python programming and building Neural Networks. There, one undoubtedly will learn a lot about some of the concepts behind this type of data crunching and see a number of examples. Learning on how to make it applicable to one of the specific situations described above and figuring out which of all the possible network types is relevant for your specific issue will easily cost a few months extra. Getting the model trained will take a few weeks more, but first you need to be able to identify and collect all the data required to even apply the concept of a Neural Network to the specific issue.
Without even thinking about deployment and the scale and internal resources (time as well as people) required for this, there’s many years ahead during which you’ll still be facing the stress and, in the meantime, getting your head around a sometimes very complex mathematical concept.
For most people that doesn’t sound like fun and not at all like something you’d immediately want to start with. Reality is even worse: Neural Networks alone aren’t going to solve these problems. What’s more: Neural Networks might not even be the most appropriate and efficient tool to solve them. The issues may be too low in frequency, the data set required to build a solid model may be too long for it to be practically possible, the data you have may be too incomplete or inconsistent…
Not all is lost however, as today a new breed of solutions is arising.
You can even rest assured, because these solutions are not made up out of services offered by consultants. What is available today are vertically integrated solutions with an open structure and a modular architecture, offered as a full product:
‘Vertically integrated’ means that the solution is provided as a product that is pre-made for the specific machine, line or problem and containing all the individual components required to get it running, smoothly: from sensors and/or connection to existing data sources, including integrated, automatic analysis packages up to the interfacing with the end user through tailored warnings, dashboards, reports or integrations with existing operational platforms such as PTC or Maximo.
The ‘open’ structure is reflected in the fact that the data remains the property of the client, isn’t shared with anyone else, nor is it stored in proprietary file types. Next to that the products are designed such that they are intrinsically made for interfacing with existing software packages or databases: (re)using as much as possible the relevant existing applications and don’t disturb ongoing operations and processes. As an expert you can always take a look at the raw data to understand more profoundly the context of an issue that has occurred or is predicted.
The ‘modular’ aspect refers to the structure of the products being, well, modular… This modularity is found in a horizontal as well as in a vertical dimension. Horizontal modularity means that the data the product uses as a ‘fuel’ for generating insights can come from additional sensors, from existing data sources such as historians, or from a mix of both. Communication can be taken up by a 4G network, LoRa, or your own internal ethernet on a separate section of network. Vertical modularity means that the product can take care of multiple aspects, but not all have to be used (thus paid for). Only these modules related to the selected machine type and issue are used, and these modules in turn are making use of the most relevant mathematical tool for the specific issue at hand. Basing itself on a Neural Network when required, but in many cases relying on plain physics, statistics or correlation analysis when appropriate.
This new breed of products has a sole focus on delivering results to its users. Each product is built on proven experience in monitoring the specific asset type, it takes away the burden of finding the most appropriate platform to deploy a solution on, it avoids losing time on selecting, trying and testing models and it helps to avoid the frustration building up in the company when the solution turns out not to deliver what was promised. Each of them provides actionable insights, without hassle, and does it fast.
What’s there not to like?
Image: JJ Ying @unsplash.com