
What is preventive maintenance?
- Von Björn Piepenburg
- Failure, Forecast quality, Intervals, Maintenance, Prevention, Preventive maintenance
Share post:
The term “preventive maintenance” is often used in connection with the optimization of maintenance plans. This article explains what is behind this and how to implement preventive maintenance based on historical failure data.
A very well-known and old example of preventive maintenance is the V-belt in a car, which transmits the rotation of the engine crankshaft to various auxiliary units. It is particularly important to supply the alternator, which provides power to many of the car’s electrical components, as well as the hydraulic pump, water pump, power steering, ventilation and air conditioning.
Back to preventive maintenance: Experience has shown that a V-belt lasts an average of around 100,000 kilometers. However, the running time depends on various factors, such as the power of the engine, the age of the V-belt and the number of start/stop operations. Without knowing these influences precisely, a safety factor is applied to the average service life of the V-belt. The safety factor must also take into account that a broken V-belt can lead to significant engine damage. For example, we set the safety factor to γ=0.8 and replace the V-belt approximately every γ⋅100,000=80,000 kilometers, using the inspection dates recommended by the manufacturer. Of course, individual V-belts can also last 120,000 or 150,000 kilometers, but proactive replacement or preventive maintenance minimizes the risk of engine damage without completely eliminating it. When determining the safety factor, the potential engine damage must be taken into account on the one hand; on the other hand, the effort or workshop costs for replacing the V-belt.
This example can be applied to all mechanical (and electronic) devices. Typically, the expected load cycles of ball bearings, the average burning time of light bulbs or the service life of sealing rings are known. These components are used in trains, airplanes, generators, robots, elevators, power plants, bicycles, etc. in every country in the world. These many millions of components make effective preventive maintenance important in order to avoid damage and system failures.
So how does preventive maintenance work? The example of the V-belt has already shown this quite well: you need historical failure data for the system type under consideration or, as in the example of the V-belt, historical failure data for the component under consideration. If you want to carry out the calculation for a new type of system, you can aggregate findings from the installed components or test the system in various practical tests in the laboratory.
As with V-belts, in practice there are several time dimensions that describe the failure intervals of the system. Let’s take the example of a car that is rarely driven. In this case, the V-belt will eventually become porous and may only last 50,000 kilometers or less. In another example, the car is moved several times a day (e.g. in a postal delivery vehicle), which can also cause the V-belt to break after 50,000 kilometers. A distinction is made between the Mean Time Between Removal (MTBR), the Mean Duration Between Removal (MDBR) and the Mean Cycles Between Removal (MCBR). Depending on the application, further time dimensions are conceivable. A specific failure interval results as a linear combination of the individual time dimensions, the coefficients of which are derived from an optimization algorithm. Typically, we use a meta-heuristic such as Simulated Annealing for this task, which converges to a good (local) optimum in finite time.
The next step is to determine which situations can have an influence on the failure of the system/component. To this end, it is advisable to consult experts from development and production and, if necessary, users. If there are too many influences, these can be reduced using statistical methods for feature selection. Alternatively, the information content of the influences can be condensed using a principal component or factor analysis. In the case of V-belts, these are typically the average speed of the belt, the quality of the material, the ambient temperatures (perhaps a V-belt lasts longer in Norway than in Saudi Arabia? or shorter?) and others.
A set of rules must now be derived from the influencing variables in order to be able to predict specific failure behavior. Statistical decision trees or the Random Forest machine learning method are usually used for this purpose. Both methods arrange the influences in descending order depending on their information content and separate the characteristics in each tree level as shown schematically in the following figure. In order to achieve maximum flexibility in the consideration and combination of all influences and time dimensions, we have also implemented individual algorithms for clamping the trees for our customers.

Figure 2: Schematic representation of a decision tree for the implementation of preventive maintenance for V-belts
Each node of the tree contains average times (for three time dimensions) from all historical failures that fulfill the criteria of the branch under consideration. The tree shown was not fully formed (this is known as prunning) because, for example, only a small amount of data was available from manufacturer 5. The difference between the actual distances between two changes and the calculated times in the corresponding node of the tree determine the quality of the tree.
A typical service life can now be read from the tree for a specific application situation of the part or component. As described above, it is advisable to apply a safety factor to the times, which should take into account the costs of failure and the costs of replacement. The forecast model should also be presented with current outage and replacement information in order to be able to take changed environmental conditions into account in the forecast.
To increase the forecast quality, either further statistical characteristics can be taken into account or data from the specific V-belt (the specific serial) can be recorded with the aid of sensors. For example, an acoustic sensor could be used to measure the smooth running of the V-belt. If sensors are used to measure the current condition of the component, this is referred to as “predictive maintenance”. How predictive maintenance works will be explained in one of our next blog posts.
