Predictive maintenance is positioned as one of the undisputed top trends for 2022 when it comes to Industry 4.0, by Forbes and many comparable reports on the near future of manufacturing – already seeing an emerging demand for advanced analytics and AI-based forecasting. Despite the undoubted value behind it, few companies ultimately decide to invest in it or simply give up at an early stage. The main reason is the lack of expected results.
In this article, I try to define the most realistic form of predictive maintenance. I focus on the reasons why the success rate is so low today. And finally, I suggest ways to increase the chances of adopting this exciting technology. I show you the way that will yield valuable returns and bring organizations closer to the Smart Factory transformation.
What is Predictive Maintenance?
The definition of predictive maintenance is understood in many ways depending on the area in which it is used. For production maintenance professionals, it may mean predicting machine failures in advance. For others, it may mean detecting anomalies in equipment. I have even come across an opinion that predictive maintenance is the optimization of the entire production plant.
However meaningless the definition itself may be in various ways. Such a common understanding of what lies behind it often leads to serious misunderstandings and missed expectations. That ultimately makes further progress in implementing this technology impossible.
The following general classification of analytics is a very good entry point for an overview of the topic.
The graphic clearly shows us the four main types of analytics:
1) Explanatory Analytics – that allows for a more comprehensive understanding of the data. It answers questions about what happened in the past, when and why. Focuses on values that can be extracted directly from historical data without requiring detailed calculations:
a) Data visualization and advanced reporting (Business Intelligence).
b) Statistical analysis (discovering covert correlations, finding anomalies across specific signals).
2) Predictive Analytics – that makes it possible to predict future outcomes. Responds to questions about what, when, and why things will happen in the future. Focuses on the value that can be gained from historical data (typically using machine learning). Generally includes data science activities related to:
a) Analyze in-depth and prepare historical data for modeling using machine learning.
b) Finding the right model and training it in just the right use case context that we are actually trying to solve.
3) Prescriptive Analytics – enables you to optimize future performance. Answers what should be done and when. Focuses on the value that can be extracted from existing predictive models to maximize or minimize expected outcomes.
4) Cognitive Analytics – which allows for real-time decision-making based on the ongoing state of the environment. Unlike the previous types, here historical data is not analyzed to discover patterns or hidden correlations. Rather as a source of simulation data when training the model. The goal of cognitive analytics is to accumulate knowledge and wisdom, by taking specific actions. Those operations modify the artificial environment and improve decisions over time based on feedback.
Rarely is predictive maintenance classified into a different category than predictive analytics (not just because of the similarity of names), so some conclusions can be drawn. It is not likely to be about anomaly detection, which fits more into the explanatory cluster. It’s certainly not about any optimization, we’re talking about prescriptive maintenance then. Nor is it about failure prevention, which is called preventive maintenance, by the way.
Using the knowledge and experience I have gained, I would describe predictive maintenance in one sentence:
Predictive maintenance is the use of Artificial Intelligence (AI) based on historical data collected from production lines to detect patterns in the production process and predict the probability, location, and timing of downtime that can be used to schedule equipment service in a timely manner before breakdowns occur.
The most important thing, however, is not to come up with a perfect definition, but to develop a common understanding. So that all parties involved in a potential project can communicate in the same “language”. The best way to focus on business problems and manageable and deliverable goals.
Why is Predictive Maintenance so hard?
However, a major hallmark (but also a massive challenge) of predictive maintenance and industrial artificial intelligence analytics, in general, is the fact that it is also not very replicable. Every manufacturing process is unique in some way. Even the same type of equipment used on multiple production lines may look similar at first glance. Nonetheless, they also typically differ in terms of internal physics and mechanical details. Therefore, literally, all predictive maintenance projects require extensive expertise in the nature of the process they are intended to address.
Data scientists, who simultaneously combine the knowledge of computer science and mathematics required to model a certain problem. They are expected to have the necessary specialized domain experience and know which key elements of machines or processes they should focus on.
Of course, it is impossible to be an expert in every field. Researchers can gain wisdom and specialize in certain narrow areas. By continually working on similar business problems, for example, the pumps, and milling machines. But it is still likely that this experience will be worthless if they are faced with some other forecasting problem that involves other processes or types of equipment. Process and machine understanding are quite simply essential.
How to Overcome Predictive Maintenance Implementation?
The only way to overcome this challenge is to engage an expert. The person who is close to the entire process and comes directly from the shop floor. Specialist with hands-on knowledge of machine behavior, who can exchange crucial insights with a data scientist.
Defining the business problem with precision
Another equally important factor of a successful predictive maintenance project is a very meticulous definition of the business problem. Discovering the value that solving the business problem should bring will be the start of an effective Smart Factory implementation. Companies that are willing to invest in advanced analytics must be told by service providers that this initial step is absolutely critical to their success. “We would like to implement predictive maintenance in our factory” doesn’t tell us about your organization’s needs and problems. That’s really not enough detail.
Haste breeds risk
After the predictive maintenance use case is finally defined (best via a discovery workshop with the customer), another common mistake occurs. It is hurrying into machine learning modeling without properly exploring the data. It’s always a good idea to take the time to gather more information.
Use business analytics and statistics to look for trends or correlations. Decide what features we want to focus on while training our models. There is still time at this early stage to assess whether we have quality enough data and can conduct a feasibility study. And answer whether the chosen business pain point is even possible to solve.
When to consider Predictive Maintenance?
Predictive maintenance enables you to take corrective action ahead of time. Helps detect patterns in your manufacturing process, optimize failure thresholds, extend asset life, and comply with safety regulations. It provides the ability to make precise determinations using artificial intelligence (AI). Based on historical data collected from production lines, fits perfectly with the idea of Industry 4.0.
The core technologies of Industry 4.0, which includes smart factories, are IoT, cloud computing, and big data analytics. To become a smart factory organization and implement cost-effective solutions and zero-waste production through predictive maintenance, companies need to collect data, huge amounts of historical data, therefore, successful IIoT implementation and digitization of factory operational information is usually the first step on the road to advanced analytics.
Beyond that, facilities need to deal with some kind of repeatable situations. Such as infrequent, unexpected breakdowns of a particular machine for which there is enough historical data. If we want to predict the failure of an asset that has broken down once in the last two or three years, we probably don’t have enough information to model the pattern of what led to that breakdown. Therefore, there is no way to implement predictive maintenance in that case.
Smart factories and Predictive Maintenance are the future. Organizations that want to leverage their power must be ready to invest time and money in the initial research of the problem domain they want to solve. It might not bring immediate value to their business, but is a necessary step to achieve ultimate success.
Explore the predictive maintenance capabilities of our Industrial Analytics Accelerator solution.
Watch the demo.
In the industrial world, maintenance is a critical success factor. Predictive maintenance is the best way to ensure that you are maximizing your efficiency and ensuring top-quality production. Our predictive maintenance solution rewards you with resources that are running at their peak performance and reliability, saving you money in the long run. If it’s time to think about predictive maintenance in your organization, feel free to contact me.