TOP CHALLENGES OF PROCESS MONITORING
JANUARY 9, 2018 - AARON SHELLY, DATA SCIENTIST
Over the past few decades, industries have been moving away from traditionally reactive, time-based preventive maintenance practices, toward a more predictive, condition-based maintenance philosophy.
This shift is no coincidence – it has been accompanied by significant advances in production automation, as well as an expanding availability of process-monitoring data. As a result, significant resources have been put into the development of health prediction models that prevent unscheduled downtime and unexpected breakdowns, ensure healthy conditions for operating equipment and maintain product quality.
Despite the noticeable success of different predictive maintenance strategies across manufacturing sectors, many challenges remain when developing robust health models for different machines.
Machines Operating under Multiple Regimes
Monitoring applications that operate under multiple loads, temperatures and speeds can be challenging, as different settings may change during use. Applications that exhibit this multi-regime characteristic include wind turbines, excavators, machine tools and aircraft engines, spanning across numerous industrial sectors such as manufacturing, aerospace, automotive and semiconductor.
When interpreting condition-monitoring signals and features, which are highly influenced by varying operating conditions, one approach has been to divide the operating space into smaller sub-regimes, developing health models for each individual operational state. Although various clustering methods are available, regime separation is typically done using one’s engineering experience and knowledge of the system. By modeling each individual regime independently, varying operational effects are mitigated and degradation signatures for each operational state can be observed.
While local regime health modeling has been demonstrated successfully for monitoring assets with multi-regime behavior, it has its limitations. In some instances, the degradation pattern may be more pronounced in one operational regime as opposed to another. Therefore, additional modeling techniques are needed to fuse the health values from each regime and infer the health state of the whole system. Additionally, the number of varying operational states for some assets may be too large, making the local modeling approach unrealistic or unfeasible for each unique operating regime.
The high level of machine-to-machine variability between different assets is also a common challenge in process-monitoring applications. Machine variation can be a product of many things, including similar assets with varying operational states, different assets with similar functionalities and older models of a tool. Even in situations where one would expect two identical machines to deliver the same capability when manufacturing the same part, machine-to-machine variation can still happen.
Therefore, it is paramount modeling techniques used to monitor different assets consider the variation between them. The most common approach is to use individual models for each new machine. While this method works well for small-scale monitoring applications, however, it is far less feasible for entire production lines with many individual assets, especially if each machine needs unique models. New machines also tend to lack historical data required by some monitoring methods.
Significant machine-to-machine variation and changing operational regimes remain two of the several challenges to condition-based monitoring in industrial applications. To fully address these challenges and deploy new technologies, we must think creatively and solve problems effectively. Although some case-by-case solutions already exist, predictive analytics companies such as Predictronics work toward developing more general solutions for a variety of industrial use cases.
For more information and weekly updates, follow Predictronics: