Imagine artificial intelligence (AI) could automate data analysis to create prescriptive and prioritized maintenance tasks. What if AI could automatically create a work order in existing computerized maintenance management systems (CMMS)? Imagine automating data collection through the internet of things (IoT) so the user experience is cost-effective and scalable.
Sound too good to be true? Believe it or not, it is possible.
Where to Start
Typically, there are three major categories of conditions that lead to machinery failure:
- Late-stage conditions: imbalance, misalignment, bent shaft, soft foot, looseness, etc.
- Early-stage conditions: lubrication, early-stage bearing failures
- Intermittent conditions: cavitation, resonance, speed related, process related
For late-stage conditions, an AI model views the frequency band from 2 to 1,000 Hertz (Hz). The AI compares the amplitude of the vibration to a set of ISO 10816 standard alert limits and generates an appropriate alert based on the severity (how much amplitude is in the vibration). Using the ISO-defined and industry-accepted alerts of 0.2, 0.5 and 1 inch per second (IPS) peak, the AI rules engine sends an alert via text message or email to the user.
For example, when a 0.2 IPS threshold is crossed, the automated prescriptive maintenance task is to “field inspect” the machine using human experience and senses of sight, sound, touch and smell. When an alert condition is present, human senses are often successful in finding the root cause of equipment issues. Remote monitoring of spectrum data, although valuable, is no substitute for physical presence at the asset to inspect equipment and communicate with operations about its history and context.
Diagnosing Early Stage Conditions
For early stage conditions, an AI model watches the frequency band from 1,000 to 30,000 Hz. This range includes ultrasonic frequencies and compares the amplitude of the ultrasonic measurement to a set of established alert limits. Ultrasonic or high-frequency monitoring is relatively new and ISO or industry standards are limited. Using experience and industry-proven alerts of 6, 12, and 18 Gs (force of gravity), the AI model sends an alert via text message or email.
For example, when plant assets are operating between 600 and 6,000 rpm and a 6 Gs threshold is crossed, the automated prescriptive maintenance task is to field inspect and lubricate. Again, human experience and senses are used to collect information relating to the cause of the alert condition.
Tips for the human inspection include:
- asking operations about any changes or insights
- looking or smelling for smoke or hot sensations or unusual odors
- looking for leaks, spills or broken or loose supports
- listening for rubbing or grinding
Some maintenance technicians use a stethoscope to listen to bearings to compare a good-sounding bearing with a suspect one. Using an infrared (IR) temperature gun or camera to inspect for hot spots is another helpful technology. The user may also check lubrication levels and color in sight glasses, or for signs of recent greasing. Then, they determine if it is a greaseable bearing, if it is under- or over-greased, if the wrong grease was used or if it was auto-greased. If it is an oil bath bearing, they will also check the oil bath level. They will determine if the machine is a chronic problem, if it is running at design or if it has been modified, sped up, load increased, belt tension increased, etc.
As always, employing an experienced vibration analyst is a safe bet. However, in midtier plant and/or with midcritical assets, a hybrid human and AI program can be successful.
This is where continuous monitoring with permanent or temporary monitoring stands apart from a monthly route-based program. With wireless sensors and the IoT, coupled with the AI models described above, intermittent conditions can be monitored and root issues can be quickly identified and resolved.
A common reality of vibration analysis on midcritical assets is they are typically monitored monthly, quarterly or not at all. In the route-based predictive maintenance program, the analyst gathers data monthly, taking 10 or more points per machine, and conducts a brief inspection of the machine. In contrast, continuous monitoring gathers fewer measurement points, and the data is logged and reviewed remote from the machine, in a control room, plant office, by remote third party, or all of the above.
Through continuous monitoring with low-installed-cost IoT sensors, many issues that have perplexed the monthly routine are resolved. Examples vary from boiler feedwater pumps that blow a seal on the night shift, to kraft pulpers that crater a gearbox over the weekend, failed pump seals related to misalignment from thermal growth, or inconsistent lubrication practices. Intermittent issues are best solved with continuous data coupled with AI to set alert limits.
These systems typically include a sensor, cellular connectivity and AI in the cloud.
A 45-Day Predictive Maintenance Case Study
On May 15, a combination ultrasonic/vibration sensor was deployed to monitor a 300-horsepower (hp) plant air compressor at a Midwest industrial manufacturing plant. The sensors were battery-powered and connected to the bearings using a two-rail magnetic mount. With existing cellular, cloud and smartphone infrastructure, the hardware and system were in place. An alert system was in place with preset alert thresholds using a proven AI model. The last step was adding instruction on what to do when an alert was generated.
A second AI model was used to deliver specific and customized instructions for each alert and each severity level of the alerts. This AI model translated the alert to a prescriptive maintenance task with an understanding of severity and prioritization, allowing the work to be appropriately planned and scheduled.
Upon connecting the magnet to the bearing of the machine, a critical (red) alert was issued based on the ultrasonic measurement (green trend line), and a work order was emailed stating: “Field inspect and lubricate within 10 days.” Within seven days, the compressor was inspected, but no visible or audible symptoms were noticed. Moving the ultrasonic/vibration sensor to each bearing of the compressor, it was clear the motor inboard bearing was reading high—approximately 36 Gs.
A closer look revealed the grease Zerk had not been greased in a while. The plant maintenance technician was new to the plant and was unaware of the plant lubrication practices—pointing to the need for proper training. It was unclear if or when the motor was last greased. The plant manager decided to add grease, resulting in a temporary reduction in the ultrasonic measurement and then a return to 36 Gs.
Recently, the plant had a failure on the motor of the backup air compressor. They knew the subject motor had been in-service for more than 15 years, so they weighed the impact of an unplanned motor failure versus the cost of repairing the motor. They decided to change out the motor during the July Fourth outage, approximately 45 days from the initial discovery of the fault.
On June 13, they switched to the backup compressor. The new motor was installed and a month later on July 13 they switched back to the main compressor. The ultrasonic reading was much lower, hovering between 4 and 6 Gs, but still occasionally setting off a minor alert. On July 20, grease was added, and the ultrasonic reading was reduced to the normal range, hovering between 2 and 4 Gs.
A second fault based on overall vibration (blue trend line) was also found during initial connection of the sensor. In this case, the AI rules engine for overall vibration is based on the ISO 10816 standard. This alert condition was minor, and the emailed prescriptive task was to field inspect for looseness, imbalance or misalignment. Again, no visual or audible evidence was present. The new plant maintenance tech had little history on this compressor. The overall vibration levels were compared to the backup compressors and confirmed it was higher.
The compressor was inspected carefully for looseness. One technique was to walk the sensor down the motor from top to bottom looking for a change in vibration levels at the different interfaces between motor and sole plate, sole plate and foundation. No evidence of looseness was found. Imbalance—a common fault on fans—was temporarily ruled out. However, this was a motor/compressor combination. Plant alignment practices were unknown and left up to the mechanical contractors. A contractor was brought in to check alignment and a misalignment was confirmed. After aligning the motor, it was determined that the motor was bolt bound (not enough slot) to correct the alignment. The team decided the alignment would be corrected when the motor was replaced at the Fourth of July outage.
On startup July 13, the overall vibration returned to its alert condition and it did not pass the acceptance test. The contractor was called back to correct the alignment. On July 14 the compressor was started and the overall vibration hovered around 0.15 IPS-peak (Pk)—a normal reading.