Everyone knows that maintenance is critical for running a well-oiled machine (almost by definition here). The cost of broken equipment goes beyond just parts and labor, especially when it impacts production throughput. At the same time, equipment maintenance has an associated opportunity cost. As far as your operation is concerned, the impact on output is the same whether a machine is down because it’s broken versus being down for repairs.
There’s clearly a tradeoff here, and different strategies have developed to balance maintenance vs. production. One approach, predictive maintenance attempts to avoid unnecessary downtime by predicting when a machine will go down. But how?
What is predictive maintenance?
Predictive maintenance, or PdM is defined as a set of techniques and tools designed to continuously monitor equipment condition in order to predict when maintenance should be performed. This differs from preventative maintenance where equipment is periodically inspected for damage. The ability to minimize inspection time by focusing on equipment with a higher chance of breaking down is what makes predictive maintenance so compelling.
When you think about it, predictive maintenance is a very common-sense approach. After all, if you have limited resources you should focus your energy on the things that look like they’re broken. The challenge is in understanding which equipment needs repair just before it needs the repair.
How does predictive maintenance work?
The requirements for predictive maintenance are relatively simple: you need equipment data, a method of interpreting that data, and a way to report results to the maintenance team.
In order to identify equipment issues, we first need equipment data. Typically, predictive maintenance systems monitor data from sensors that measure vibration, temperature, noise and pressure. Increasingly, equipment manufacturers are embedding these sensors in their new machines as the cost of sensors drops. And since many of these sensors are so small, the effort of installing them in existing facilities has become easier.
Once you have sensor data, you need to compare it against something in order to make predictions. One approach is to compare sensor data against maintenance logs in order to look for the correlation between when damage occurs and the associated threshold that the measurement has crossed. For example, you might notice that motors running above a certain temperature are very likely to break down, so whenever you detect the motor hitting that temperature you immediately inspect it for issues.
Another approach is to look for anomalies. This involves analysis of data over time to establish a baseline against which you can compare current results. For example, vibration data for our motor might indicate a natural operating frequency of 1 kHz with noticeable harmonics at 2, 3 and 4 kHz. Any deviation from this frequency pattern would immediately signal that something is wrong with the motor.
What are the challenges associated with predictive maintenance?
While the concept of predictive maintenance is simple, the implementation can be challenging. For starters, companies that lack solid data and IT infrastructure may struggle to deal with high-volume sensor data feeds that pass through their network and into their databases.
The interpretation of sensor data can also be challenging, leading many predictive maintenance companies to focus on creating dashboards with graphs, heat maps, etc., and offloading the interpretation of the data to members of the maintenance team. While not the most sophisticated approach, it can be very cost-effective for companies with relatively simple operations. After all, why buy a Ferrari if all you use it for is to pick up groceries?
In complex environments involving multiple stages of processing, a more sophisticated predictive model is justified, especially if multiple pieces of equipment are critical to the operation.
What are we trying to optimize, maintenance or throughput?
The underlying assumption behind predictive maintenance is that by optimizing equipment maintenance we can improve our operations. And for simple operations, this is obviously true. For example, a wind turbine generating electricity is relatively self-contained. There is a clear one-to-one correlation between electrical output and downtime.
Complex operations on the other hand are, well, complex. Environments with multiple machines feeding each other in an assembly-line setup require consideration of more than just downtime. Strictly speaking, a piece of equipment operating at 70% capacity may not need to be repaired until it reaches 40% or 50%, but if it’s a bottleneck then the impact in lost throughput could be many times the associated replacement cost.
A strict focus on optimizing maintenance costs will help to optimize maintenance costs and streamline your maintenance team. This is a very good goal to strive for. But an even more audacious goal is to maximize throughput for the entire operation by integrating predictive maintenance into broader decision making.
What are your thoughts on this issue? Leave a note in our comments section below.