The industrial Internet of Things and its Digital Twin surrogate are fueling exciting conversations about business process innovation on the factory floor and in industrial equipment manufacturing. One hot area in particular is the broad and often loosely-defined practice of Predictive Maintenance (PdM) of complex machinery. As is frequently the case when technology innovation is trying to penetrate (should I have said “disrupt”?) an established business practice, there’s a good dose of hype and optimism on the side of technology pundits, countered by skepticism and resistance to change from maintenance organizations and experienced field service technicians.
A business associate sent me an article written by a technology vendor seeking to debunk what the authors deem to be myths about PdM (try searching for “predictive maintenance myths” and you will find a few on the topic). That article and recent vendor presentations I attended suggest that further examination of misunderstandings and perhaps exaggerated expectations from PdM technology is in order.
My goal is not to debunk myths (or propagate alternative ones), but rather give you a pragmatic perspective, so you know what to expect and how to prepare for your PdM initiative. As you will see, some of the observations stem from similar technical and process roots and lead to similar conclusions. This is intentional.
One Model to Rule Them All
The implied assumption that similar devices generate identical sensor data patterns and therefore a machine learning algorithm can handle multiple devices in the installed base is flawed. In reality, no two mechanical systems are alike even as they roll off the production line. And they begin changing as soon as they are put into duty.
The baseline data generated by rotating and reciprocating equipment is highly configuration and application specific, and changes continually throughout the life of the asset due to wear and tear, different duty cycles, and operation and maintenance practices. Machine learning algorithms must adapt to these changes without compromising detection precision or suffering from an increasing rate of false positives.
The Digital Twin provides the ability to track and analyze each asset individually, allowing the machine learning apply context beyond machine-generated sensor data, such as configuration and maintenance history.
Although the digital twin is essential to implementing a PdM system, the focus on individual as-maintained unit configurations also highlights a potential concern: if, over time, assets drift to the degree they are no longer similar, the ability to conduct any type of broad installed base analysis is impeded.
It’s Magic: Machines Learning Algorithms Do All the Work
Machine learning enthusiasts seem to propose something little short of magic. Just feed the software with a wealth of machine-generated data, and AI-based algorithms do all the work on their own. They remove signal noise and data outliers, and smooth data just enough so no key features are lost; they identify the best-suited analytic algorithm; and provide highly accurate data trending and failure prediction.
The reality can prove to be more complex, to say the least. Early adopters find that while building a proof-of-concept model is a manageable effort and the results can be very impressive indeed, these initial models can be difficult to scale, as the models must be validated for a much broader range of product configurations and applications, and be able to adopt to changes induced by cyclical changes and wear and tear.
The effort to test and validate machine learning algorithms cannot be underestimated. Some types of artificial intelligence systems require regression testing every time a change is made. Ohers, black box type systems, such as neural networks, cannot be trusted blindly based on test results from limited training data.
You Don’t Need to Be an Expert to Make Sense of the Data
Can clever data scientists equipped with powerful analytics and machine learning algorithms build and test a PdM system, or do you also need to recruit subject matter experts and maintenance specialists?
Because operating signatures of rotating and reciprocating machinery varied by configuration and duty cycle, and changes over time due to wear and tear, making sense of data patterns and prescribing the appropriate response require more than mere statistical data analytics. Whether you employ data science PhDs or believe that out-of-the-box machine learning algorithms can eliminate the need for them, machine information viewed outside the exact context of a machine’s configuration, task, and maintenance history is likely to be ineffective and possibly dangerously erroneous. Don’t fall victim to spurious correlations!
Good Vibrations
Using vibration analysis to predict remaining life and early failure detection in rotating machinery has been the subject of research for decades. For example, this report was published more than 30 years old and its bibliography references research from the early 1980s!
Novel MEMS sensor technology and increased power are getting remaining-life prediction systems that vibration analysis and similar technologies from academic research to pragmatic applications faster than ever before. Earlier this year, I saw a demonstration of a vibration anomaly detection system using audio signatures, which reduces the need to mount and wire accelerometers.
However, before we rush to install sensors and implement complex signature analysis algorithms, we need to acknowledge, once again, that the signal signature will vary from one unit to another and will keep changing over time. And it’s easy to overlook the details that separate the real world from the lab. For instance, identical units installed using different mounting techniques on different foundations are likely to exhibit different waveforms.
If You Build It, They Will Come
This one isn’t about implementing autonomous failure detection algorithms, but rather using artificial intelligence to augment human intelligence: provide a probabilistic evaluation of failure modes and make recommendations for further troubleshooting and corrective actions.
The proponents of these systems focus on making these systems as powerful as they know how and assume that service technicians are just going to love them. After all, these systems are so smart, so why wouldn’t they?
Tepid acceptance of AI-based decision-support systems have been a challenge since the dawn of the concept. Service technicians often find these systems tedious, inflexible, and overly authoritative, and not sufficiently tuned to their needs, habits, and workflows.
If predictive and prescriptive maintenance support systems are not designed with service technicians in mind, they will not get much use, losing a critical element of constructing machine-learning algorithms that improve over time though feedback about the accuracy of the diagnosis and efficiency of the repair process.
Image: The Conjurer (School of Hieronymus Bosch, after 1500)