Today's industry executives are discovering new ways to maximize the reliability and value of their assets. With asset performance management powered by the Industrial Internet of Things (IIoT) and machine learning, companies can leverage both equipment and process data to extend the life of their assets and achieve optimum reliability.
On average, up to 15 percent of gross margin is eaten by unplanned downtime. By comparison, best-in-class performance is estimated at 5 percent. Eliminating these losses will require maintenance and production to work together in new ways.
The traditional approach to reliability has been to build a first principles model of the asset, tune the model with real-time data, implement corrective factors or create rules for accuracy, compare model outputs with real-time data and highlight statistical deviations from normal conditions. However, these models only look at asset data. They cannot "see" upstream in the process to identify causal behaviors that degrade assets and can only signal when the onset of damage becomes evident — when the damage is already done.
This conventional method for predicting performance was developed 40 years ago in models based on engineering equations, statistical techniques and rules engines, but many still rely on it. Machine learning has emerged only recently. Both techniques often appear to solve the same problems but differ in areas of human involvement and accuracy of prediction. Modeling techniques, which require extensive experience and skills with appropriate calibration techniques, have been and continue to be very successful. With first principles, specific behavior must be understood. Real-time, dynamic models offer predictions of forecast behavior at any points in time, providing an in-depth understanding of expected performance.
What makes solving the problem of unplanned disruptions and downtime so challenging is the dynamic nature of production processes. With thousands of variations occurring simultaneously within the process, it is difficult for models to predict exactly which patterns or trends will lead to unplanned events.
First principles (engineering) models show only estimated, expected or perceived behavior based on hygienically clean, best-case performance. How often does the mechanical equipment run this way? Is it the same at 30, 50, 100 or 110 percent throughput? In contrast, machine learning can learn based on the actual real-world behavior of the equipment under all conditions, including seasonal variations, different operating campaigns, startup/shutdown and changing duty cycles. It can also take into account the deteriorating process and mechanical performance.
Machine learning mines the process and asset data for early warning. It does the heavy lifting of finding the patterns in the process that signal future asset problems. By identifying the process behaviors that are the root cause of degradation, issues are identified much earlier. With this approach, risk analysis and machine learning work together to continually and accurately predict asset failures weeks or months in advance. This can provide time to plan, coordinate and take action rather than just react. This time is what allows maintenance and production to work together in new ways.
Machine learning applications do not build models in the traditional sense of heat/material balance and thermodynamic polytropic equations, logic and rules, and statistical interpretation. They measure failure signatures rather than model machines.
Applied with skill and domain knowledge, machine learning absorbs hard, measured sensor and maintenance data collected over long time periods to identify minuscule, multi-variate and temporal patterns that humans cannot see. Discovered patterns are the exact signatures that define both normal behavior and the excursions leading to degradation and failure. For the sake of conformity, we can call these signatures models, but they are conceptually far away from the ideas of engineering or mathematical models.
Signatures of failure developed with machine learning do not know or care about the type of machine, the industry where it is used or the engineering principles behind its operation. Signatures only care that there are sufficient sensors supplying enough data that contain learnable relationships between the sensors to accurately declare the operating behavior of the asset through normal and degradation/failure circumstances.
Even a library of 125 models cannot approach the hundreds of thousands of unique assets that need protection. However, machine learning can rapidly assess patterns and deploy on assets it has never seen before in hours or minutes without intense engineering skills. A best-in-class approach can do this without data science skills, run automatically in-line and in real-time, and present actionable results in seconds.
If you're still relying solely on first principles models, it's time to modernize. Using a combination of models and machine learning is the most powerful way to detect and avoid risky process operating conditions. This combination can explain the explicit conditions at any time using the model, with machine learning calibrating and fine-tuning the model automatically without much human guidance or programming rules. It's the best of both worlds — timely, accurate process status along with simpler calibration. It also gives your maintenance and operations teams the insight needed to work together for the best possible performance.
Michael Brooks is a senior advisory consultant in asset performance management for AspenTech.