Predictive Maintenance – it promises greater levels of efficiency, higher overall equipment efficiency (OEE) and a longer service life for the machinery used. But what exactly does this definition mean? That’s the topic for the first part of this blog post series.
In principle, predictive maintenance is about successfully and accurately predicting future events with the help of statistical methods. Events refer to the failure of machines or their components. The key to success lies in the patterns of past events.
The rise of predictive maintenance in mechanical engineering has much to do with the fact that the maintenance of machines has been a popular area for outsourcing over recent decades, which has led to the steadily growing importance of the industrial services industry. A more recent development is that the condition monitoring and maintenance of machines is increasingly perceived as part of the production processes and is thus treated as such. For external service providers from the industrial services sector, this is not always an easy situation. After all, their tasks are often closely linked to the core competencies of their customers while at the same time also forming a central prerequisite for their commercial success.
Generally speaking, there are various different options and models for maintaining machines and replacing wear parts.
- Reactive maintenance: a component is replaced when it is broken or as soon as it is noticed that it no longer fulfils its function according to the requirements.
- Preventive maintenance: parts are replaced at regular intervals based on the running time or the number of units produced. This reduces the downtime of machines – nevertheless it can, of course, be the case that components are replaced that still have a long service life ahead of them.
- Proactive maintenance: this involves attempting to counteract the failure of components and machines as a whole by various measures, such as better training of machine operators. This can make it possible to tackle the causes of machine failure itself.
In addition, there are various ways in which the time of maintenance or replacement can be determined more precisely, such as through the use of sensors.
- Condition-based maintenance: damaged or faulty components are replaced as soon as a sensor indicates that they are damaged. This can also be done in the context of quality control, for example, if it is determined that the manufactured products no longer meet the requirements.
- Predictive maintenance: the aim here is to predict the time of failure of a component as precisely as possible.
- Prescriptive analytics: machines try to correct problems themselves, for example by adjusting individual parameters and thereby “fixing themselves”. This model is currently still a long way off, although the technology for it already exists.
Accurate predictions using predictive and prescriptive maintenance
Both predictive and prescriptive maintenance work with empirical values. Algorithms classify the condition of plant or machine parts as “OK”, i.e. operational, or “faulty”, depending on data from the manufacturing process, such as pressure curves, vibrations or temperature changes.
The empirical values can then be subjected to an event history analysis – also known as an event time analysis, historical data analysis, survival analysis or reliability analysis. This is a type of statistical analysis in which the time period up to a specific event is compared between data sets to estimate the effect of prognostic factors or adverse influences. This analysis can be applied to all statistically recorded measurement objects that are successively discarded in the course of the data collection period.
The knowledge gained in this way can be used to create models for the expected wear and tear and train them based on the automatically recorded sensor data. The quality of these models can be continuously improved by means of an iterative process. The result is a predictive model that uses current sensor data to indicate the condition of the wear parts and when they need to be replaced. The great advantage of this is that the learning process used to create the model includes all the influencing variables, so it also shows previously overlooked relationships.
Data quality is crucial
The following example illustrates just how much potential there is in predictive maintenance for companies working in mechanical and plant engineering: a model predicts the failure of an important ball bearing on a production machine in four days’ time with a 90% probability. Replacing the ball bearing means eight hours of machine downtime. On the basis of the prediction, there are possibilities for automatic business decisions:
- Requesting the spare part from the central warehouse
- Allocation of a pending rush order to another location
- Replacement of the component scheduled for the day after next, since the current batch will then be finished.
- Booking of service technicians for the planned replacement period
- Adaptation of the production sequence for the following order to provide leeway in respect of the planned machine downtime
All these decisions are associated with great savings potential as they reduce the downtime of the machine and the number of rejects, as well as helping to ensure that agreed delivery deadlines are met.
The advantages of predictive maintenance are therefore clear. However, it relies on various crucial factors, the most important point being the quality of the data used as well as the data collection and processing processes. In addition, the challenges of predictive maintenance must be made just as transparent to stakeholders as the successes. Above all, the company management must be able to clearly understand the value contribution of the new maintenance programme so that it is further promoted.