How the ability to detect defects has evolved and continues to advance today.
ATEK Companies

Adding condition monitoring and predictive maintenance for legacy or aftermarket industrial pumps and equipment is touted today as a significant source of return on investment (ROI) for the industrial internet of things (IIoT). For some, predictive maintenance is a new term, one many confuse with condition monitoring. The main difference in predictive maintenance and condition monitoring is the timing. While both monitor the health and condition of a rotating asset like a pump, fan, compressor, mixer, agitator, conveyor, etc., condition monitoring focuses on here-and-now conditions. Predictive maintenance focuses on the early detection of defects, 60 or 90 days before the defect causes collateral damage or impacts production. Here is information highlighting the history and transgression of condition monitoring.

Condition Monitoring 1.0—the 1980s & Earlier

Much like the red alerts on the dash of your car, legacy condition monitoring in industry has included lagging indicators such as:
  • low lube oil pressure
  • high temperature
  • low or high pump discharge pressure
  • low or high seal pressure
  • low or high seal pot level
An alert condition on these measurements means a failure has already or is currently taking place and a timely response is required. In predictive maintenance, this is referred to as “condition-based-reactive maintenance.” Although useful, the indicator does not give enough time to strategize. Production does not have enough time to plan and maintenance does not have enough time to line up the right parts or skills.

Condition Monitoring 2.0—the 1990s to 2000s

A second wave of measurements has been adopted and has dramatically improved the detection of defects. Motor current, speed and power are the results of variable speed drives that have been deployed to improve efficiencies in electrical energy consumption. Additional vibration and bearing temperature measurements are more reachable thanks to cost reductions, reliability improvement, input/ output systems infrastructure and easy magnetic or epoxy mounting of sensors. This second wave of measurements includes motor current, speed, power, overall vibration and bearing temperature. A variance in any of these measurements can indicate a condition of the pump or pumping system that needs attention. Using these measurements has proved fruitful for diagnosing problems, but setting the alert thresholds for use with automated alerting has proved challenging. The varying nature of the process, product recipe or season of the year has made nuisance alarms common and has challenged the simplicity and clarity of the approach, thus requiring in-house or third-party expertise to realize success. Motor current, flow and pressure can vary with process conditions and require human analysis to identify a fault or anomaly in the normally varying measurements. One example is distinguishing a normal inrush current from an abnormal high current during steady state operations. Novice users have tried to baseline and set statistical alerts. Unfortunately with traditional methods and systems, this can be time consuming and has resulted in misses for pre-existing faults. Overall vibration deployed with knowledge of the International Organization for Standardization (ISO) 10186 alert standards has helped identify pre-existing conditions. Clarification of the failure modes detected by overall vibration has helped explain misses. Failure modes detected by overall vibration include imbalance, misalignment, looseness and late-stage bearing failure. Overall vibration is a direct measurement for detecting and monitoring imbalance, misalignment and looseness of rotating assets. The units for overall vibration are inches per second (ips)-peak, which is a velocity measure. Overall vibration is typically calculated from an acceleration reading measured using a $100 to $200 accelerometer. This ISO standard measurement has been around for decades. ISO 10816 defines how to measure and set alert thresholds. For example, the ISO 10816 standard calls for a 2-1000 hertz (Hz) frequency range and recommends alert levels for typical machines at 0.2, 0.5 and 1.0 ips-peak for minor, warning and critical alert levels. While overall vibration is excellent at detecting the presence and severity of imbalance, misalignment and looseness, many would argue it is not predictive and that overall vibration is a lagging indicator, as the problem or defect already exists. Yet, finding an imbalance or looseness defect when it is small has significant benefit if operations and maintenance have enough time to take the machine down to fix the problem while it is still small. It is best to repair the problem early, when it is a small cost, in comparison to waiting too long and fixing the problem when it has caused additional collateral damage. One example is waiting until the pump shaft is broken, versus aligning the motor, pump, and inlet and outlet piping.

Condition Monitoring 3.0—Predictive Maintenance—2010s

IIoT measurements for predictive maintenance is much akin to the discussion around business key performance indicators (KPI)-leading versus lagging. The monitoring described earlier, although good, is still lagging or condition based. For some, predictive maintenance is synonymous with technologies such as:
  • infrared thermography (IR)
  • ultrasonic
  • partial discharge testing
  • monthly vibration routes and analysis of vibration spectrum by trained and experienced professionals
With IIoT and its seven elements of data structure (see January 2019 Pumps & Systems article “7 Elements of IoT for Predictive Pump Maintenance) more intelligent sensors, coupled with more intelligent processing, communication, storage, alerting and translating, has emerged. The failure modes targeted by this new intelligence includes 60- and 90-day advance detection of lubrication defects, bearing defects, cavitation and pump seal failure. Multiple industry studies agree lubrication is the root cause of failure on 50 to 80 percent of rotating assets. Past technologies of overall vibration or bearing temperature were simply too late or too difficult to establish meaningful alerts. Pump seal failures are being predicted based on the understanding that the common root cause of a pump seal failure is shaft deflection. A leading indicator of shaft deflection is the ultrasonic or high-frequency vibration that is transferred through the bearing housing to a sensor. In “Condition Monitoring 3.0,” overall vibration in combination with high frequency or ultrasonic provides the opportunity to realize predictive maintenance, where a fault condition is identified 60 or 90 days in advance. This allows operations and maintenance to plan and schedule a repair, with the right parts, tools and skills at the right time. Ultrasonic vibration [1,000 to 25,000 hertz (Hz)] is the measurement for detecting lubrication, bearing fault, gearbox defects and cavitation. The units for high-frequency vibration or contact-based ultrasonic is Gs (units of gravity), which is an acceleration measurement. Ultrasonic is a less-known measurement and does not have any ISO standards. The one requirement of an ultrasonic measurement that is common across all sensors is for the sample rate to be equal to or greater than 64,000 samples per second. Thankfully, in recent years the processing power of computer chips for this demanding processing has become affordable and readily available in formats that can hold up in the temperature ranges of industrial applications.

Condition Monitoring 4.0—Industrial Artificial Intelligence—Current

The current advancement in predictive maintenance is to further automate the analysis process using artificial intelligence (AI) models. With the new and rich data stream from ultrasonic sensors, edge processing of the data and connectivity to the cloud, proven AI models have been developed that detect pre-existing conditions without the time, cost and frustration of baselining. Established AI models allow 90 percent or more of industrial assets to have alerts accurately established upon initial deployment. It also can amass the leading indicators mentioned above with sensors and connectivity, but also make the leap to automatic pattern detection driving quantitative patterns with various weak indicators or the specifics of the plant or line itself. The final step to realize success in any predictive maintenance program is to translate the alert or condition to a specific or “prescriptive” maintenance task. For this application, a second AI model is deployed that translates the alert or machine learning classification and severity to a prioritized maintenance task, which is then emailed or integrated with existing computerized management systems (CMMS) for planning, scheduling and managing maintenance work orders. This AI model relies on metadata about the asset type and context of its service criticality to establish the intelligent and prioritized work order. Using an application programming interface (API), the intelligence can be directed into existing systems, which limits any additional training. As a result, the catch phrase can be: "It is not rocket science, we're just automating the mundane tasks.” This frees human resources for higher-level thinking. The goal is to focus on the right work or assets and do fewer things with a higher standard by examining criticality and workflow.
Sixty-day predictive maintenance on legacy equipmentSixty-day predictive maintenance on legacy equipmentImages 1 & 2. Sixty-day predictive maintenance on legacy equipment (Images courtesy of ATEK)

Case Study

Images 1 and 2 show an example of a 60-day predictive maintenance case study on a 17-year-old legacy piece of equipment. On July 9, 2018, ultrasonic/vibration sensors were deployed to monitor a rooftop make-up air fan on the top of a Midwest industrial manufacturing plant. The sensors were battery powered and quickly and easily connected to the bearings using a two-rail magnetic mount. With existing cellular, cloud and smartphone infrastructure, the hardware and system were all in place. An alert system was in place with alert thresholds that were preset using an existing and proven AI model. The last step was clear: concise instruction on what to do when an alert is generated. For this, a second AI engine was used to deliver specific and customized instructions for each alert and each severity level of the alerts. This tested AI model translates the alert to a prescriptive maintenance task with an understanding of severity and prioritization, allowing the work to be appropriately planned and scheduled. After connecting the magnet to the bearing of the machine, a minor alert was issued and an intelligent work order was emailed stating, “Field inspect and lubricate within 10 days.” Within one day, the bearing was inspected and a 1/4-inch gap between the shaft and the inner bearing race was identified. The bearing had been slipping on the shaft and had worn a 1/2 inch of the diameter off the shaft. The overall vibration levels remained normal and the plant was able to plan a repair for 60 days out in mid-September. The plan included a non-OEM repair, with plans to use a split bearing and move the bearing location to an undamaged location of the shaft. This completed the full-value chain of predictive maintenance from monitoring to alert to prescriptive task, and ultimately, a planned and scheduled repair that avoided any collateral damage, loss of production and reactive maintenance.

Issue