What if you leveraged the power of artificial intelligence (AI) to predict more accurately when your equipment requires maintenance? That’s not a pipedream. The technology is available today. It’s called AI-based predictive maintenance, and it goes beyond the manual threshold setting and simple alerting commonly found in condition monitoring solutions today and provides detection weeks to months further in advance. Even more, with the right tools, you can leverage your existing experts to enable this capability in short order.

We’ve put together this ultimate guide to help you learn more about this emerging space. Here you’ll find everything you need to get started with building your predictive maintenance strategy, including:

As you well know, time is money. Let’s get started.

What is predictive maintenance?

Predictive maintenance is the next evolution in how companies approach equipment maintenance. It’s built on the idea that all equipment has a normal pattern of behavior and that deviations from a normal state can be quickly identified and triaged to a reliability engineer or maintenance team member for further analysis and action days to months before functional failure occurs. 

Reliability personnel attempt to derive these patterns by studying trend information from sensor data produced when monitoring this equipment. The idea is that failure modes will consistently produce patterns leading up to a failure event. If these patterns can be profiled, operators can intercede before their equipment or process reaches a failed state.

This methodology is helpful even when equipment isn’t failing, and it’s simply on its way out of a compliant state and into another state. For example, say you have a gearbox that you know takes a long time to repair. Using predictive maintenance, you can choose when it’s best to shut the gearbox down, make preventative repairs and get it back online. 

Any equipment that’s used in a production setting – and has the potential to break down and require repairs over the life of the asset – can benefit from predictive maintenance technology. The most common industries using this AI-driven software include:

  • Oil and gas
  • Manufacturing
  • Power and energy
  • Metals and mining
  • Food and beverage

How does AI-based predictive maintenance work?

Most high-value equipment has sensors that monitor the asset’s performance. Unfortunately, sifting through each sensor’s data is time-consuming and requires a level of expertise not all companies have.

ai-based predictive maintenance flash gas compressorPredictive maintenance primarily uses software-driven, physics-based and/or statistical models – and more recently, attempts to implement AI-based modeling techniques – that most often train using data from prior equipment failures and are heavily dependent on experts for configuration and data interpretation.

Predictive maintenance software ingests data from equipment sensors and compliments the equipment’s native controllers and monitoring software with business intelligence metrics and insights. The equipment insights are then elevated to a person who can take action and respond to the data, significantly reducing and even eliminating unplanned equipment outages.

Predictive maintenance trends 

Predictive maintenance that uses artificial intelligence is usually divided into these two categories: 

Supervised AI-Based predictive maintenance

This category represents the main approach to AI-driven predictive maintenance that has come into currency in the past 10 years. Supervised AI, most commonly referred to as neural networks, starts with a human identification of the “what” you would like to identify in the future. This task can be straightforward in applications such as object detection or in autonomous driving, where a human is training a navigation system to identify a class of objects such as a “stop sign.” It starts with thousands to hundreds of thousands of images where a human labels the images with the class, in this case, “stop sign.” This is the “supervision” component of the process. 

The next step is for the neural network to find the common traits in the training images which associate each image into the common class of “stop sign.” This process is most often conducted using specialized compute resources in the cloud. Once complete, the model is ready to be deployed to “inference” or compare new, untrained images against the model to determine if there is a match. This process is often iterative, and the supervisor must perform a lot of hands-on data labeling and other pre-work before deriving any benefit. It is time intensive and requires the same process for each new kind of object you are identifying. 

Supervised AI becomes difficult when the “what you are trying to predict” is difficult to identify by a human, when the training set is small, and when there is a large amount of variation in the training set. When we think about equipment data, the number of sensors, baseline variation in signals, various operating modes, dynamic operating conditions, and infrequent occurrence of consistent failure modes (unlike the millions of available images of stop signs) we can see a challenge emerging. Moreover, what happens when you encounter a state of operation you have never seen before? What label should the neural net assign then?

 

Unsupervised AI-Based predictive maintenance

This category is a more advanced solution and represents the future of predictive maintenance.

Unsupervised AI includes approaches like “clustering” (a.k.a. segmentation), and begins when the algorithm groups information together based on similar characteristics without a human directing it. In other words, this approach utilizes the algorithm first to organize and identify what is consistent or different, and a human second to assign the meaning to the data or respond to the result. 

With the right unsupervised tools, equipment, and processes producing large amounts of data with some amount of normal variation can have a unique model built automatically for each. With this model, a monitoring state can be used to quickly identify deviations from “normal operation”…this is anomaly detection. Once identified, a root cause can be elevated to a reliability engineer or maintenance team to determine the source of the issue and schedule repairs or replacement. 

Using unsupervised AI, companies can deploy this solution more quickly and with far greater scale than neural networks and can leverage existing equipment experts within your organization. This approach dramatically improves the lead-time of notifications prior to machine failure and is the space where Boon Logic operates.

Benefits of leveraging predictive maintenance

Every industry wants or needs different benefits, but these are some of the common benefits we see across industries: 

  • Gain a competitive advantage by using cutting-edge technology that delivers valuable operational insights
  • Reduce maintenance costs and improve overall equipment effectiveness by catching minor issues before they become major headaches
  • Maintain productivity and reduce the likelihood of unexpected operational failures
  • Extend the life of high-value equipment assets and improve your equipment ROI by avoiding expensive damages
  • Leverage deeper insights by adding AI to the monitoring dashboard you’re already using (no need to learn a new system or tap into separate data sources)
  • Engage talent and elevate human intelligence with higher value, more meaningful work that combines their unique expertise with technology insights (rather than performing monotonous tasks or driving from location to location – just in case)

Benefits in action: In a recent deployment, our AMBER solution for condition monitoring and predictive maintenance detected an anomaly in a specialty gas compressor for a customer in the semiconductor industry. Amber used months of historical data to build the model in a matter of seconds. For three consecutive months, Amber sent anomaly warnings to the customer that indicated not only that the equipment was deviating from its historic operating condition, but also which sensors on which components were driving the anomalies. The customer’s existing solution had no indication of any deviation, as the severity of the change hadn’t reached a critical point for their threshold-based approach to detect. Finally, after 90 days of notifications, the customer scheduled an inspection. To their surprise, Amber had identified a small component failure inside the cooling bundle. The customer quickly replaced the $65k component and notified us that a total failure would have resulted in $800k in damages and over 15 months of downtime due to supply-chain constraints. 

Read about how Boon Logic’s Amber helped AIONT provide unsupervised anomaly detection and saved them time and money.

 

AI-based predictive maintenance process 

At this point, you might say: “This all sounds great, but how does it work? And how can I convince my boss to try a proof of concept?” Here’s the exact process we follow:

  • Telemetry – we start with telemetry, which means recording instrument readings and analyzing the data with AI and machine learning.
  • Visualization – next, we create visualizations that turn telemetry data into more valuable insights (and make raw data easier to read than a boring Excel spreadsheet or sifting through dozens of data points. You just look at a few simple values that are visualized in green, amber, and red.
  • Condition monitoring – that’s performed by software rather than people. This makes it more accurate and allows you to monitor multiple pieces of equipment simultaneously.
  • Anomaly detection – our algorithm helps you identify when equipment deviates from its normal behavior. Once normal operating conditions have been factored into the model, our algorithm helps to identify very subtle deviations, not just from individual sensors monitoring an asset, but from the abnormal variation that may occur in the relationship between sensors. This is far more sensitive and insightful than traditional thresholding approaches most often used in industry solutions today.
  • Root cause analysis – uses telemetry to tell you not only that something’s failing but what part of the asset is failing and what’s causing the anomaly. This helps your boots on the ground (what we call your domain expert) more accurately find the root cause and address it more quickly.
  • Asset Maintenance – once you have identified an anomalous state is occurring and have isolated the likely source(s) for this anomaly, you can optimize your maintenance schedule and fix your asset when it works best with your operation, ideally just in time. This maximizes asset life and reduces component replacement with usable life remaining.

With the right unsupervised tools, equipment, and processes producing large amounts of data with some amount of normal variation can have a unique model built automatically for each. With this model, a monitoring state can be used to quickly identify deviations from “normal operation”…this is anomaly detection. Once identified, a root cause can be elevated to a reliability engineer or maintenance team to determine the source of the issue and schedule repairs or replacement. 

Using unsupervised AI, companies can deploy this solution more quickly and with far greater scale than neural networks and can leverage existing equipment experts within your organization. This approach dramatically improves the lead-time of notifications prior to machine failure and is the space where Boon Logic operates. 

AI-based predictive maintenance tools

The most fundamental tool you receive when using our predictive maintenance solution is called the Anomaly Index. This index gives you highly detailed measurements of your equipment and helps you quickly identify when it needs attention. Let’s look at how this works.

When an asset is operating in a normal or compliant state and doing its job, Amber assigns an index of zero to 500. When your equipment operates within that range, you know it’s performing as it should and can rest assured that the likelihood of an unplanned failure is significantly reduced.

When you see the Anomaly Index climb to 600 or 700, you know immediately that some equipment condition is changing. In the past, you’d go out onto the manufacturing floor or rig site, take a look at or listen to the equipment, and it would probably look and sound fine. But internally, something’s changing. The RCA (root cause analysis) can point you to what that may be. 

When the Anomaly Index hits 900 or 1,000, you know you’re in a critical range. The index is telling you something within the asset has changed dramatically since it was put into operation, and you need to act soon.

Readiness Assessment CTA

 

Ready to get operational insights in a matter of days?

Today’s AI-based predictive maintenance solutions can help you identify equipment issues in a fraction of the time and at a fraction of the cost. In fact, Boon Logic’s solution can be trained and deployed in hours instead of months or years. That’s because it doesn’t require highly skilled, hard-to-find talent resources like traditional AI development methods.

Your subject matter experts can use our predictive maintenance condition monitoring tool to produce rich operational insights in a matter of days.

If you’re looking for a way to guard against unplanned downtime, costly repairs, and productivity loss, let’s start a conversation about a pilot program. Contact us today.