Is the value of Machine Learning and Prognostics being overestimated?

Machine Learning is at the Peak of Inflated Expectations according to the Gartner hype cycle for emerging technologies 2016, the topic dominates industrial / maintenance conferences and a wide range of companies are defining their take on machine learning and associated offerings, all part of the Industrial Internet of Things. Once at the Peak of Inflated Expectations the next stage of the Gartner hype cycle is Trough of Disillusionment, so it is expected that the mood on Machine Learning could change in the foreseen future, from being hyped to being questioned due to lack of initial results. Are we overestimating the impact of Machine Learning within the Industrial Internet of Things? Time will tell but keep in mind Amara’s law: “We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”

Machine Learning vs Prognostics

In 1959 (!) Arthur Samuel, an American pioneer in the field of computer gaming and artificial intelligence, defined Machine Learning as “field of study that gives computers the ability to learn without being explicitly programmed.” Prognostics is the ability to predict the time when a system or a component will no longer perform its intended function. The predicted time is classified as Remaining Useful Life which is crucial input for planning the (maintenance) activities to mitigate the risk of losing the required function. Data driven prognostics are based on failure pattern recognition based on historical failure knowledge and Machine Learning for detecting changes in current behaviour. Model based prognostics are based on a physical understanding of the systems into the estimation of Remaining Useful Life.

Data and Knowledge

Applying Machine Learning and Prognostics in an industrial manufacturing environment requires data sets of a system or component’s performance and a good understanding of potential failure modes, effects, degradation patterns and a physical understanding. Data and Knowledge. The challenge is huge, historical data is not always stored and if available stored in various legacy systems. Furthermore, there a strong push to eliminate failures from repeating by implementing modifications, this will change the behaviour of the system or component in the data set, often without notice. On the knowledge side the challenges are similar, understanding failure modes, effects and criticality requires technical know-how of the system or component and a structured approach gathering and documenting the knowledge. This data is not always stored and if available stored in various legacy or maintenance management systems.

Organisational Maturity

When considering implementing Machine Learning and / or Prognostic systems it’s important to understand the level of organisational readiness and maturity. It not only requires the technology but also the processes and the people. A pure technology driven approach is imminent to fail to deliver the desired outcome. Once the technology provides a prediction of Remaining Useful Life the organisation needs to act. In the past, we have seen similar challenges when implementing condition monitoring systems, once installed and the first analytics were made, the organisation didn’t act on the advice due to other priorities or lack of trust in the system. As a consequence, an unplanned failure did occur and the system was blamed. Often there was no process in place for handling the condition monitoring advise and the technicians involved were seen as whiz kids not understanding operational targets and requirements. Other elements that need consideration when implementing Machine Learning and / or Prognostic systems are security, legacy systems, IT requirements for interfacing, etc.

Value based approach

Equipment reliability (the probability that a component or system will perform a required function for a given time when used under stated operating conditions) is fundamental for any reliable operations. Reliability Centered Maintenance (RCM) is best practice for defining a risk based, cost effective maintenance strategy. The maintenance strategy takes into consideration the required function, potential failure, likelihood of failure, consequence of failure and risk mitigation of the failure by defining a maintenance plan; a value based approach that is clearly linked to cost, output and compliance. Machine Learning and / or Prognostics should be linked to the risk based approach, they are part of Condition Based Maintenance (CBM) program that should, ultimately, improve shareholder value.

Is the potential value of Machine Learning and Prognostics being overestimated? I don’t think so, there is a clear advantage when implemented correctly. Is the ease of implementing Machine Learning and Prognostics being overestimated? I truly believe so, there is a strong focus on the technology but and what it could bring but there is limited focus on all the elements a company needs to take into account when going on this digital journey. 

And...

Industrial Internet of Things success relies on combining technology, processes and people. It is certainly not just a matter of measuring everything, everywhere. Any initiative should be aligned with your risk based approach driving long-term value to your company; reducing cost, improving output and ensuring compliance.