In Eric Siegel’s book, he forecasts how by the year 2020, predictive analytics will impact pretty much all aspects of everyday life. In the chapter titled “Ten predictions for the first hour of 2020”, he details how real time predictive analytics influences your driving habits by helping with collision avoidance and automobile maintenance. The highly readable and informative book does not drill down into too much detail on this particular topic, but we can safely claim that fault detection and predictive maintenance are already important application areas in manufacturing and related activities.
Computer-controlled test and monitor systems provide system status and allow for monitoring of many sub-system parameters in industrial applications, but this data is not always captured and thus not analyzed over time. Preventative maintenance today is not driven by automated system status or performance indicators and trends. Thus, maintenance is performed inefficiently and often fails to predict or prevent component and system failures. Corrective system maintenance usually occurs after a component or system fails, or if component degradation is observed during routine preventative maintenance. Failure to anticipate corrective maintenance requirements increases mean time to repair (MTTR), and decreases operational availability (Ao). Unanticipated corrective maintenance actions also drive up costs due to increased labor costs and expedited shipping costs when parts have to be obtained quickly.
Fault detection and predictive maintenance systems will continuously monitor all the component parametric data streams and then conduct a type of trend analysis. This “smart” system will then combine the trend analysis data with component degradation and failure data reports to improve its prediction. The end result is a system that is capable of providing a report such as, “machine part “A” has a 90% probability of failure within the next 72-96 operating hours” or “the output of component “B” decreased by 10% in the last 7 days with the rate of output decrease accelerating significantly in the last 24 operating hours, indicating there is an 89% probability of component failure in the next 96 operating hours.”
Predictive maintenance is a condition-based maintenance technology (i.e., a predictive condition-based maintenance forecaster) that uses machine learning techniques to “understand” component interdependencies and thus can accurately predict component failures based on all available parametric data.
For someone tasked with the job of overseeing sophisticated production systems, a high level dashboard that captures various alerts and trends would be very valuable. The example below shows one such custom dashboard. Each panel provides a deeper level of understanding than the previous one.
In step 1 above, all the subsystems being monitored are displayed (and sorted by the number of error codes that have occurred since monitoring began). Selecting any subsystem will take us to the second panel which will then sort all the error codes by description which have been recorded for that subsystem.
In step 2, we select the error code which is the most critical one and this will take us to the panel on the right which now shows all the distinct plants or entities which have demonstrated this error code.
In step 3, we simply select the entity of interest (could be a distinct machine or car) and then track its behavior over time where the errors or fault codes appeared and were recorded. In the last step, we can choose to predict the probability that this error or fault code will re-occur in the future and decide to take appropriate action.
The key to building predictive models for fault detection and condition based maintenance is to identify parameters which strongly influence the occurrence of a problematic behavior. Once we short list such parameters, it is possible to build accurate classification models which can predict the likelihood of a similar error or fault condition.