If predictive maintenance is the cherry, where is the cake?

Predictive maintenance programs to increase plant productivity are a current fashion trend. However, doubts about this approach are growing. It is undoubtedly correct to have plant and process data centrally available. However, there is a better and more comprehensive approach to capitalizing on investment in information technology.

The promise of predictive maintenance is tempting: Using machine learning technologies to find links between faults and causes in historical performance, downtime and sensor data in order to predict when a plant component will fail and proactively replace it shortly before this defect occurs. The aim is to turn unplanned outages into planned downtimes and thus increase plant productivity.

At first glance, the procedure appears to be a worthwhile innovation in almost all manufacturing machines. But can this promise be fulfilled?

If you analyse the situation in more detail, the optimization potential through predictive maintenance is reduced depending on 4 factors:

• Too few sensors: It is only possible to predict which detailed data is also available. However, many technical defects can also occur on components that are not subject to sensor monitoring.
• Too little data: Predicting defects is more difficult than you might think. Defects in individual plant components are only very rare events, which means that there is typically not enough data available to train a reliable machine learning model.
• Too little impact: Although system defects reduce OEE, the downtime of a plant in the event of a technical defect can also be kept short with physical or planning redundancy, a clever spare parts strategy and quickly available and qualified personnel.
• Too strong focus: If you look at a distribution of all plant losses, you see not only the technical defects but also other availability losses such as organizational disruptions (no personnel, no material, waiting for Q approval...), performance losses such as slow speeds or quality losses as periods of sub-optimal productivity. These losses generally represent the clear majority of productivity losses, meaning that the significance of technical defects fades into the background.

For one or more of these four reasons, expectations for predictive maintenance projects often fall short of reality.

Does this mean that the manufacturing industry is not a field of application for data-driven productivity initiatives? No, quite the opposite. However, the narrow focus on predicting future technical defects is usually* not the right approach. Instead, the comprehensive approach of analyzing and optimizing OEE losses in the sense of targeted OEE management appears promising.

For this purpose, the root causes of disturbance as defined by the deviation from the ideal running of the system must be classified in the categories of availability, performance and quality losses. A small part of the data can be delivered from the system control system, but the majority of the classifications result from the operator's domain knowledge. The following figure provides information about the structure:

With transparent reporting on the data collected in this way, a major contribution can be made to focusing improvement activities and thus increasing plant productivity. In addition, artificial intelligence, which is usually machine learning, or even statistics in this data stream can be used to gain interesting insights into patterns and anomalies for ways to increase plant productivity:

• Grouping of frequency/duration of loss causes
• Identification of periods with accumulation/absence of losses
• Identification of anomalous (e.g. particularly long or short) losses
• Identification of cyclical presence of losses
• Determining the set-up quality of a conversion
• Calculation of a key figure for running stability or identification of periods of stable or unstable plant operation
• Identification of a temporary deviation from longer-term disruption trends

If selected analyses are carried out on-line, operations can even be intervened while the fault still exists. For example, according to the algorithm, an unusually long standstill can be identified and reported to a manager so that he gets a personal impression and takes measures depending on the situation.

If necessary, this can be enriched with sensor data, which are usually collected for predictive maintenance: Set vibrations, temperature, pressures, etc. in connection with availability, performance and quality losses. However, this data is then only an addition to the further refinement of the methodology. In other words, the cherry on the cake rather than the cake itself.

What all analyses have in common is that the technology provides an indication of the possibility of optimization, but people must decide and act in a targeted manner on the basis of this data.

The potential of data-driven process improvement therefore goes far beyond predictive maintenance. For manufacturing companies, OEE management provides digital optimization approaches that are easier to implement and offer greater added value. So why limit yourself to your options with an exclusive focus on predictive maintenance?

* To avoid misunderstandings: Of course, there are also worthwhile applications for predictive maintenance technologies, e.g. in safety-relevant environments (aircraft engines) or during shutdowns with enormous follow-up costs (offshore wind turbines, chemical plants).