A few weeks ago, I met with an investor to offer an assessment and guidance concerning a potential investment in a software company developing a predictive maintenance software: a system designed to use sensor data to assess the condition of a piece of equipment, detect an impending failure, and prescribe a remedial action.
As is often the case, there was a profound difference between the ideal world as described by the technologists and the harsh reality of equipment maintenance as experienced by field-service personnel.
Writing the Complete Guide to Predictive Maintenance Technology is obviously a grand goal for a short blog post. But now that I have got your attention, I’d like to focus on just one aspect: while diagnostic algorithms can be extremely powerful, implementing an application that can tackle complex, real-world maintenance problems requires much more than the ability to detect an anomaly in a time-series data stream.
Let’s look at a simple example: a reading from a temperature sensor is above the specified threshold. The problem and the repair action seem to be sufficiently straightforward, don’t they? Well, while the symptom is clear, the root cause isn’t: the over-temperature condition could have been caused by a coolant leak or by a stuck thermostat (and theoretically, while less likely, by both). The service technician should take additional troubleshooting steps to pinpoint the culprit before attempting to repair the problem.
But we are not done yet. Can the service technician trust that the temperature sensor reading is accurate? A bad sensor, a malfunctioning analog-to-digital converter circuit, or a wiring harness problem will generate a false temperature reading. Again, the experienced technician may choose to verify the temperature before engaging in a lengthy and expensive repair operation.
And there’s more. As it turned out, some of systems in the field have been modified to allow safe operation at a higher temperature. Is the system under repair one of these systems? Was a software update installed?
Obviously, a data analytic tool is useful, but is only one component in the field service diagnostics and repair workflow.
Another example from the favorite domain of many Industrial IoT predictive maintenance applications: vibration data as a predictor of failures in rotary equipment. Consider two identical pieces of equipment, for example pumps, installed in two locations. One installation followed the vendor-prescribed installation procedure of bolting the pump directly onto a cement slab. The other installation, due to site limitations (or, perhaps, the installer’s creativity), has a longer bracket welded to the base of the pump and secured to a side wall. As you can imagine, the vibration signatures of the two installations will be markedly different, yet they are both normal. The analytic engine must be able to account for these differences, as well as natural changes in the vibration signature that result from normal wear and tear.
While many proof of concept demonstrations appear to work well in detecting data anomalies under lab conditions, some statistical methods to model, trend and predict data streams will have difficulties dealing with multiple variant and configurations, and self-correct to reflect the ongoing changes in the installed equipment.
The purpose of this short writeup is not to criticize and dismiss data analysis and predictive algorithms. It is to highlight the fact that real-world diagnosis and repair workflows are much more intricate than a simplistic (not to say naïve) technology-centric demonstration reveals.
Whether you are an investor assessing a predictive maintenance technology, or a service organization looking to implement it, you must consider that building a real-world application takes much more than a clever signal analysis algorithm. Real-world situations of the type that service technicians face daily can easily outstrip the value offered by predictive algorithms.
Image: M.C. Escher and the crystal ball used for his self portrait (1935)