There are a variety of practices and regimens in asset maintenance available to plant operators, including run-to-failure, calendar or cycle-based maintenance, and predictive maintenance. Of these, predictive maintenance is the most cost-effective and best ways to reduce downtime.
Applying patterns found in historical sensor data to anticipate future maintenance and servicing needs, predictive maintenance is able to identify asset maintenance requirements months in advance based on small variations in sensor data. This allows for manufacturing downtime to be scheduled intelligently, avoiding costly unplanned downtime from emergency maintenance.
Predictive maintenance compares data compiled from sensors to records of asset maintenance in order to find patterns and signatures of failure. Thus, one important requirement for predictive maintenance is high-quality historical data. This data must be correctly aligned with current sensing capabilities in order to increase the effectiveness of predictive maintenance. Even a slight change in sensor calibration can mask errors or produce false positives of failure signatures.
Another requirement for predictive maintenance is asset management software that is able to apply machine learning to the historical data. This type of software uses artificial intelligence (AI) to make predictions about asset maintenance, continually learning and adapting according to real-time sensor data.
The final requirement for predictive maintenance is responsive plant operators who are willing and able to act on the recommendations of the machine learning algorithm. Plant operators may feel constrained by production targets and be unwilling to take assets offline based on the recommendations coming from a software program. Good communication with your maintenance team and proper technology education can go a long way towards building faith in your asset maintenance software systems.
How does the predictive maintenance process work? First, historical data from sensors are gathered and cleaned. Data cleaning may be as simple as ensuring time format compatibility, or as complicated as applying calibration offsets to mismatched sensing data.
Next, the cleaned historical data are fed to a machine-learning algorithm that identifies patterns and conditions in which assets are more likely to experience failures. The algorithm can even find counterintuitive signatures of failure. For example, a valve failure from over-pressuring may be preceded by a drop-in pressure. Or, a distillation stack that overheats may show reduced temperatures due to sensor fouling.
After training, the algorithm is applied to current sensor data. The machine-learning software searches the current data for patterns found in the historical data. Asset maintenance requirements are flagged by the predictive maintenance software for review by plant operators.
Finally, plant operators use the software’s recommendations to plan asset maintenance. Intelligent anticipation of maintenance needs provides greater flexibility to reduce downtime.
Halted production in response to an emergency failure is costlier and more dangerous than planned manufacturing downtime. Predictive maintenance tools provide advance warning of maintenance needs, allowing plant operators to intelligently reduce downtime by proactively scheduling asset maintenance. Multiple assets that have been flagged for maintenance can be serviced simultaneously, and production can be shifted to other lines, leading to higher reliability.
Without an effective predictive maintenance tool, plant operators are severely limited in their ability to reduce downtime. Prior to the development of predictive maintenance, the gold standard of reliability management practices had been preventive maintenance. Unfortunately, preventive maintenance can only reduce emergencies and inefficiencies caused by normal wear and tear, which is estimated to represent approximately only 20% of all failures.
Predictive maintenance also reduces unnecessary asset maintenance. It may be prudent to overhaul equipment according to a set schedule, but manufacturing downtime is not without costs. Predictive maintenance provides a method by which only necessary maintenance is performed, avoiding expensive loss of output.
How does predictive maintenance work?
Predictive maintenance finds patterns in sensor data that are associated with reduced asset optimization. Predictive maintenance uses machine learning to identify maintenance and service requirements with greater accuracy and longer lead times than are otherwise possible. Historical asset performance monitoring data are made available to machine learning algorithms to train a model that will predict plant maintenance needs going forward. This model will then be deployed on site through asset management software to provide real-time predictions.
What is the difference between predictive and preventive maintenance?
The goal of preventive maintenance is to prevent or reduce emergency failures and unplanned downtime by servicing assets prior to their point of failure. Preventive maintenance uses rules-of-thumb and calendar scheduling to set servicing schedules, regardless of asset conditions.
Predictive maintenance also seeks to reduce unplanned downtime by anticipating maintenance needs. However, predictive maintenance determines when to perform asset maintenance by applying machine learning models to asset performance monitoring data.
What are predictive maintenance tools?
Falling under the asset management software category, predictive maintenance tools use machine learning to help schedule planned downtime based on sensor data and production schedules.
How do you implement predictive maintenance?
Predictive maintenance is typically implemented with a machine learning tool such as Aspen Mtell®. Historical sensor and failure data are fed into an AI algorithm that teaches a software model to anticipate future maintenance needs. This predictive maintenance model can then be integrated with traditional asset management software.
Predictive Maintenance Takes on Operational Risk