Assets need to be maintained so that they keep working as intended. There have been several methodologies that have been developed to help with maintenance, and the one of the latest to come forth has been #Predictive #Maintenance based on Industrial #IoT sensors (#IIoT). A lot of statements have been made about it, but what’s true and what’s not? Will it do as much as everyone says? This article intends to help you understand the true possibilities of the technology, and how we’ve gotten to this point, so that you can come up with the answer yourself. First things first, though… How much do you know about the technology, and how did you learn about it? For many readers, the main avenue of learning has been the media and the manufacturers themselves, who have come up with interesting articles and case studies about a wonderful new technology. This is how hype is created.
“Hype” y expectativas
Hype is everywhere in the industry. It helps create awareness and market when a new technology comes out. Sometimes, however, it does it to a certain extent that it starts to hinder the actual progress of technology in the short run. Let me give you an example:
“Technology X comes out to the market. After some proof-of-concept runs, it shows a lot of potential and is a promising answer to some hard problems in several industries. Venture Capitalists become more and more interested in it and start investing. This is when things really kick off: sales and marketing teams are created and push new products into the mainstream market. Early successes start to oil the hype machine, which in turn keeps bringing new customers in with new demands. After a stream of early successes, high expectations are set, and people start to believe Technology X is going to solve many problems, including outlandish claims that come with the hype cycle. When Tech X fails to deliver in these crazy demands, negative press starts to unfold and people lose interest. VCs stop investing in it and people start to drop the tech. If the tech survives this, a second wave of more mature implementations comes, having learnt from past mistakes.”
Sounds familiar? This is basically a paraphrasing of Gartner’s Hype Cycle methodology, illustrated in the following chart.
What I believe is important to keep in mind about this version of the Hype Cycle is the place where #MachineLearning, #DeepLearning and #IIoT platforms are (all on the “Peak of Inflated Expectations”). I concur with this assessment, in that the maintenance industry is starting to realize that these technologies by themselves won’t solve their problems as hype has made it believe. Users are expecting magical tools that identify patterns and recommend actions without much effort from their part.
I believe the issue is not in the technology. I believe it’s in the mindset of technology adopters, in that #AI is not a silver bullet and it doesn’t come close to human intelligence (yet?), and even if it did, it wouldn’t provide the answers expected from it as-is. Let me give you another example to illustrate this: think of an AI tool as a freshman starting college. Even to the student with the highest potential, if you drop years and years of IoT sensor information to him or her without any context, previous knowledge or expertise, there’s only so much they can do with the data. The lack of experience and knowledge of the processes and the involved equipment hampers understanding, and with it, the ability to infer future states (or, as we like to call it now, predictive capabilities).
Context is King
To understand plant maintenance and reliability, the analyst first needs to understand the assets to be maintained. It is crucial to know how a piece of equipment works and should work to be able to infer how it will (or won’t) work under different circumstances.
How do you obtain this knowledge? Well, unlike many other fields in which machine learning is being implemented, plant maintenance is firmly based on engineering, which means that there’s a huge body of knowledge to be used in the form of Physical laws. Not only are mechanical systems well characterized, they are deterministic and, in many occasions, linear. Data scientists don’t need to reinvent the wheel to be successful in plant maintenance. A brute-force approach to this problem would be like asking a student to understand how a plant works by feeding her years of sensor data, without having studied her engineering courses first. The knowledge of a seasoned engineer is of immense help in a predictive maintenance scenario. Many times, intuition and expertise are more accurate than your models because the maintenance engineer understands the context in which the machine operates, and understands how both internal and external factors affect the plant’s operations.
Trough of Disillusionment
A student without context can, nonetheless, give interesting results when aided by a knowledgeable peer. There are cases in which machine learning can help you, by letting the algorithm know when a failure happened. If the machine fails in a predictable way, and there’s enough data to let the student find a pattern with confidence, a good result comes out for this failure.
Now, when the new model is deployed, the rate at which the machine fails lowers, because you start to avoid the failure that your model identified. As failures start being averted, a new dataset is created where this failure happens so few times that there’s not enough data to detect a pattern that lets you avoid further other occurrences of the failure. As years go by, this new dataset will be your data source to re-train the models, and since the failure is not there, the #AI by itself won’t find a failure to avert. Predictive maintenance falls victim of its own success. If you’re not careful and don’t include the previous pattern into your new model, you will reduce the effectiveness of your own model!
Even if you include previous patterns and improve the model, insights will become scarcer. This will make belief and trust in the project dwindle, which will in turn make it lose sponsorship and fall in a downwards spiral, and there’s not much you can do to revive it.
Business knowledge to the rescue
The failure of your project to deliver new insights as time goes by is because it was entirely based in sensor data processing. We believe the best way to fix this problem is to combine business knowledge, engineering expertise, expert experience and data science in one team and one single discipline. In this way, your maintenance strategy gets the best from all worlds, having data science and predictive analytics as tools that warn you of impending predictable failures and help you make better decisions instead of an all-knowing oracle that doesn’t help you as much. Four pillars that sustain your operation, instead of one.
Through the combination of our skills, the use of our unique equipment model, and our pattern recognition engine for sensor information, we at Uptime Analytics understand the importance of having a 360 view of maintenance and predicting failures. Having a combined-arms approach, we can help you achieve and surpass your maintenance goals, decreasing your downtime as well as your maintenance costs.
Written by:Daniel Muñoz – Senior Analytics Engineer en Uptime Analytics