Predictive maintenance: understand data and predict failures
Starting point
Theoretically we can define predictive maintenance as the science that helps us predict failures by combining operational information generated by critical assets, maintenance histories and context information to improve reliability and performance, reducing unscheduled downtime and costs.
I remind you of a simple definition of preventive maintenance: a set of actions to be carried out on an asset to extend its useful life, and not aimed at predicting when this asset is going to suffer a failure or breakdown.
Data as the key to predictive maintenance
As extracted from the definition, we need information and data. Before that, the performance of predictive maintenance has two prerequisites: to know well which are our critical assets and to understand the data generated by these assets in order to perform the necessary analytics.
In the first step, our clients are now opting to use the ABC classification model for different asset classes in manufacturing operations. Thanks to this classification, we can define different strategies for each type of asset:
- Custom models: the most critical equipment, with unique functionality, unscheduled shutdowns cause a large impact on production, very significant repair costs.
- Standard Models: critical equipment, many assets of similar types or class, unscheduled shutdowns impact production, cumulative maintenance cost is significant.
- Less critical equipment: easy replacement, run to failure or time-based maintenance.
Logically, cataloguing the assets of your company is always a difficult task. Do we do it for its initial cost? For the time it has been in operation in my company? Is it a critical machine in my production process? For direct cost in case of production stoppage?
The second step is to understand what data we are able to extract from the machine. To have clear the necessary data map and the way in which those data are generated falls on the manufacturing companies, nevertheless, to extract those data and to join them with the rest of data that are generated is the great challenge of a production company. On the other hand, as can be seen in the following table, we speak of predictive analytics when a large number of companies do not perform descriptive or diagnostic analytics:
Are we ready to use IoT and Big Data? Do we want to predict future failures when, on the other hand, we do not analyze failures that have occurred in the past? Are we talking about Industry 4.0 when we mean digitization of the company?
To help you reflect, I recommend the following article by my colleague Ivan Toda: The level of maturity of organizations in data analysis.
Without a doubt, the great challenge for everyone is to design a roadmap that helps companies to understand their assets and data in a simple way.
As a very simple roadmap, predictive maintenance requires that the company has defined a strategy that integrates the following concepts at the technical, technological and operational levels:
- Assets + Sensors + Data + Connectivity + Analytics + Monitoring + Reporting. If we define a unified strategy with the above terms it is because we have an initial objective:
- Understand in real time the facts that affect the performance and use of an asset. And therefore if we manage to fulfill our objective in a direct way we will be able to find the following results:
- Expected results:
- Reduce unscheduled stops
- Reduce maintenance costs
- Improve product warranty
- Reduce inventory cost
- Extend the life of assets
- Improve the production and quality of products
- Improve maintenance planning
If we want to achieve a predictive maintenance status, we must always consider two more parts: specific indicators to control our production line and the methodologies of predictive analytics or advanced analytics.
In relation to indicators within production companies, it is common to use tools or techniques such as SPC (Statistic Process Control), QEWS (Quality Early Warning System patented by IBM), 3Sigma, Negative Margin (expected time that the equipment fails before its scheduled maintenance has to be carried out), etc.
As far as predictive analytics are concerned, we identified 4 types of algorithm families:
- Classification and prediction algorithms: Quest, Chaid, C.5.0, decision list, linear regression, etc.
- Algorithms of association: a priori, charma, rules of association, etc.
- Segmentation algorithms: K-medias, Kohonen, TwoStep, etc.
- Auto modeling algorithms: Autoclassifier, Autonumeric, Autocluster, etc.
Finally, to introduce the methodological and technological supports necessary to be able to deploy this advanced analytics:
- Technology Support: Oracle Data Miner, IBM DB infoSphere Warehouse, Microsfot Analysis Services.
- Predictive analysis methodology: CRISP-DM. Cross Industry Standard Process for Data Mining
- Programming environments and languages: R and Python
Without a doubt, predictive maintenance is no longer a fashion but a reality. Many companies, especially in the industrial and manufacturing sectors, are defining their strategies to obtain specific results. Can we already say that predictive maintenance is in the hype phase for technological gurus?
This opens other debates such as the professional profiles that are necessary to develop these projects, what performance or profitability we want to achieve and, above all, how to use all this to transform our business model and therefore the sector where we work.
The range of possibilities opened up is almost unmanageable, also very encouraging, but this new challenge does not have to make us forget that many companies are still in the digitalisation phase, so entering industry 4.0 requires completing many stages beforehand.
Time will tell if predictive maintenance remains a reality only for large companies or becomes a reality that reaches end consumers. Why not dream that our mobile warns us that it is going to spoil in 2 months and 3 days? Or almost better not?
