Everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it too! What are we talking about? Predictive maintenance, of course! As a result of the digital transformation and the increasing amount of available data, predictive maintenance is today emerging as one of the most successful use cases in Industry 4.0. But what is it exactly? What are its advantages? And, above all, is it always possible to adopt it or are there any limitations to bear in mind? We tried to put together a brief, (non) ultimate guide.
Let’s start with a simple definition: predictive maintenance refers to a system able to detect upcoming failures in a production plant, a single machine or one of its components, and notify the maintenance team of a timely intervention. In fact, a predictive maintenance system is able to identify where the problem is, why it is happening, if it just refers to an error or to a real breakdown and, in the latter case, when it could eventually occur in the near future.
Predictive maintenance is not science fiction; on the contrary, it has already been adopted in various sectors, especially in those industries where reliability is a crucial aspect. Think for example about power plants, utilities, transport systems, communication and emergency services.
Markets & Markets estimates that the value of its market will reach a value of $4.9 million by the end of 2021 - that’s almost three times the value in 2016! - and that it will increase up to $6.3 billion in 2022 ( Market Research Future ) and to $12.3 billion in 2025 ( Capgemini ). So it is easy to understand why there is so much interest (and noise!) around the topic.
According to Capgemini , a French multinational IT services and consulting company, predictive maintenance has the potential to decrease:
Similarly, the study Anticipate the Future published in 2019 by Porsche consulting, claims that predictive management of industrial plants can extend asset lifetime by 30%. The study further reports an estimation of the potential savings on maintenance costs thanks to the implementation of predictive systems across different sectors: 30% in the aircraft field, 20% in the rail transport, wind power and automotive industries, and up to 40% in the mining sector.
These benefits can bring a significant competitive advantage to businesses, and this is the reason why the topic is a hot issue. However, while on the one hand demand for innovation is certainly a positive sign, on the other, this interest has also led to communicative misunderstandings. Going straight to the point: predictive maintenance is not black magic, but it is a complex system based on specific requirements and which, to be correctly implemented, needs specialized skills.
Predictive maintenance is not the only type of maintenance system a business can adopt; actually, it is just an evolution of former types of maintenance which was born after the ever-increasing diffusion of IoT technologies (in particular advanced sensors) and the development of new models and methods of Machine Learning.
So, let’s take a step back and identify the top 3 macro categories of maintenance: manutenzione correttiva, manutenzione preventiva and manutenzione predittiva.
Corrective maintenance consists in replacing a device (machinery or other system components) when it is broken. This approach does not require upfront costs nor any additional time and money than repairs. On the other hand, this leads to a) a shorter lifetime of the asset, b) risks in terms of security, and c) a potential loss of earnings due to technical downtimes. Given these premises, it may be spontaneous to wonder whether it is worth considering valid this type of maintenance; well, if costs and time required for replacement can be considered negligible, corrective maintenance may be sufficient.
Preventive maintenance, instead, consists in scheduling the replacement of certain devices or components based on their lifetime estimation. This means preventing breakdowns without waiting for something to actually break - it is like when we do periodic check-ups to our car and replace certain components at a specific mileage. Although this type of maintenance allows to reduce the number of breakdowns, there is yet a major disadvantage: replacements take place regardless of the real status of the device, therefore they can occur sometimes too soon or too late.
While the first two methods depend on the status of the device, predictive maintenance exploits both historical data and real-time data generated by IoT sensors connected to a device for providing an accurate snapshot of its health situation. In this way, as it happens for preventive maintenance, it is possible to reduce the number of failures and breakdowns; the difference, however, is that predictive maintenance only occurs when needed, avoiding acting too early or too late. This leads to a reduction in costs and downtimes, while it drives an increase in revenues and in the safety level of the workplace.
At a conceptual level, a predictive maintenance system can be based on two different approaches: model-based or data-driven.
The model-based approach involves an equipe of engineers and plant architects who design and implement an observation model to monitor the health status of a device according to specific values measured by IoT sensors installed.
In the data-driven approach, instead, data generated by sensors are not used as input variables of the model, but to build the model itself. In this case, the only experts involved are data scientists, whose “super-power” consists in handling large volumes of data and conducting analysis in order to get valuable information.
The idea behind this approach is that real-time data can provide insights about the health of the device and when it will break, without worrying about what the device does. Nuclear power plants? Inverter? In this approach, it makes conceptually no difference what the device actually is: the only thing that matters is data.
However, a clarification here is required: although this approach is extremely powerful, in practice it brings along a few limitations, which may inhibit its application.
In fact, a predictive maintenance system based on a data-driven approach, and in particular on Artificial Intelligence algorithms, requires the fulfillment of a series of prerequisites which does not always take place.
First of all, a device must be equipped with sensors capable of detecting those variables that are functional to determine the health status of the device itself. Differently said, it is not enough to have sensors connected to the device; we need sensors collecting those parameters that are useful to understand whether the device is healthy or not. And, damn, that’s not always the case.
Let’s take an electric motor as an example: typically, we will have sensors to monitor the number of revolutions per unit of time (that is a functional aspect of this device). However, a useful parameter to understand its health status would be the current absorption, which in turn may require the installation of additional sensors.
However, it is not enough to dispose of sensors measuring specific parameters; there must also be the possibility of collecting such data, saving it and accessing it - ideally in real time. In other words, the device must be connected.
Once the hardware and software infrastructure for data collection are all set, it is then necessary to dispose of a sufficiently large database reporting different scenarios, i.e. data relating to normal situations as well as breakdowns. Moreover, it is important that such data is machine-readable and annotated - that is categorized and labelled in such a way that it can be “understood” by AI applications.
At this point, you may wonder what volume of readable and annotated data we are talking about. Of course, it depends on the type of device and possible breakdowns, but to give a rough idea we need at least a couple of years of history and a few hundreds breakdowns to be able to effectively train an AI model.
But what if this data is not available because the device is new or it has not broken “often enough”? A possible solution is to carry out stress tests with the aim of identifying the breakdowns that could arise during the device’s lifetime. This approach, however, can partially or even irremediably damage the available asset - so it may not be worth the risk.
To overcome this problem, a virtual simulation - that is a software cloning a real behavior - can be the answer. Even in this case, though, it is difficult to envisage all the possible scenarios and understand the natural (real) causes underlying failures and breakdowns. In short, as you may notice, we are greatly restricting the situations in which AI can be used cum grano salis.
If, at this point, you have more doubts than certainties about the feasibility of adopting a predictive maintenance solution - well, don’t worry, because that was exactly our goal! At U-Hopper we know how these things work and we want to warn you of the belief that predictive maintenance is an easy and quick strategy to implement and adopt.
To sum up, we would like to use a last simile: an AI-based solution is like a recipe for baking a cake; it needs the right ingredients, in the right quantity, and it requires you to respect the right baking time. This is the only way to get a perfect cake! Similarly, only if the “recipe” for best-in-class predictive maintenance solution is followed to the letter, you’ll be able to collect tangible benefits.