Steve Matthews | ACOEM USA

| IMAGE 1: Pump with AI analysis diagnostics (Image credit: Munny - stock.adobe.com)

There is no question that AI is a big topic of discussion in the machinery reliability realm and the world at large. This article will discuss general observations about AI in industry and will offer advice on how to evaluate the growing myriad of options available, focusing on reliability systems for rotating machinery.

The most basic definition of AI is a computational system built to do tasks that usually need a human mind. Machine learning is part of AI; it is a general term describing the use of math and statistics in pattern recognition. Machine learning is not specific to the application of AI to learn machines, although in effect, that is the case. Other AI terms include: supervised versus unsupervised learning, Bayesian and neural networks, deep learning, algorithm, narrow versus general AI, and the list goes on. There is no need to define all these terms here, just to mention some of the more common buzzwords one may encounter while navigating this subject.

AI has been around in the machine reliability world for several decades. The earliest versions were called smart, expert or knowledge-based. The difference in modern systems is mainly the addition of machine learning to detect patterns without explicit instruction versus being programmed rules-based only.

The current market has exploded with offerings from every corner of industry, responding to increasing demand for the ability to eliminate unplanned downtime with no expertise needed. A once-high barrier to entry (research and development and manufacturing expense) is now significantly lowered by sensor and battery technology coupled with declining cost and wide availability. Additionally, market growth is currently fueled by risk tolerant, software as a service (SaaS)-model-loving investments and well-funded marketing efforts.

More choice is generally a good thing for markets. Unlimited choice, in this case, seems to be creating confusion, and in many cases frustrations for those who have chosen a model that has not delivered on expectations or was supplied by vendors that were either limited in application expertise or have since shut down, merged or been sold. With the fundamentals and background laid, what can AI do for machine reliability?

Fundamentally, start by defining the objectives. AI should be viewed as just another tool in the toolbox and applied as any other tool is applied—the right one for the right job. The benefits of condition monitoring are widely known and accepted, yet this simple, very important first step is still overlooked by many organizations. In many cases, objectives are opposing, disconnected, vaguely defined or not defined at all. A few examples include corporate programs or purchases that are pushed out without any notice, input or “buy-in” from the plants; deployment of sensors (vibration, ultrasonic, temperature, etc.) by automation or controls teams with no maintenance/reliability collaboration; deployment of the wrong sensor type for an application; or investment in predictive technologies without plans or personnel to remediate issues, so machines fail despite ample warning.

The point is simple: Look for AI that will do what it is being asked to do, and have a plan to act on the information it provides. This starts with knowing what exactly the AI will need to do.

The simplest AI models seem to rely strictly on machine learning. That is to say, the AI trains itself to recognize changes in patterns and alert users on those changes. This type is pretty basic in that the alert may indicate a problem, but there is little diagnostic value beyond answering the question: What is the component or fault causing the change? This type of AI model is probably the most widely available and lowest cost. Some in this category explicitly state the model needs to be trained using the machine’s data for some period before any reliable results can be obtained.

More sophisticated condition monitoring AI models will both detect change and diagnose the cause. For example, bearing lubrication, bearing fault, misalignment, unbalance, belt issues, looseness and cavitation, just to name a few. Obviously, more in-depth information has additional value. The most sophisticated models incorporate larger datasets from which the models are trained, and a wider variety of application types. Development of AI systems has been targeted first to the most common machinery such as pumps, compressors, fans, electric motors and simple gearboxes. As an AI model matures, so should its accuracy and breadth of applications.

Within the condition monitoring world, there are two general delivery models available: AI only, and AI that is backed by some type of human expertise service. There are multiple variations on the second model. A key advantage of a human-expertise-backed model is the availability for interaction in the recommendations/action phase, and post-maintenance follow-up.

As a side note, it is also possible to use common generative AI programs to perform a single machine analysis using only screenshots of collected data. For a vibration enthusiast, it can be both fun and impressive to see what these models turn out in a matter of seconds. However, this is more of a novelty application of AI for machinery analysis.

Finally, consider the old adage, “garbage in, garbage out.” This term originated in computer science, and it is also ubiquitous in vibration analysis. Good data quality and data resolution are as important as ever in the AI era. Make sure the data type and quality is supportive of the objective as well. Sensor technology and applications is a whole topic in itself, but just as an example, installing a battery-powered sensor on a 185 F surface is going to cook the battery in a matter of weeks to months. Even the most sophisticated AI model cannot fix that.

AI is here to stay, and it keeps getting better while simultaneously producing undesirable results along the way. The keys are understanding the objectives and choosing the right applications. Regardless of the current status or sophistication of those objectives, the most value can be gained by working with people and companies with specific expertise and experience in machine condition monitoring.

Steve Matthews is regional manager, training programs manager and corporate reliability programs specialist at ACOEM USA. For more information, visit acoem.us.

In This Issue

Cover Series Chemical Processing