The useless machine

A useless machine is a device which has a function but no direct purpose. It may be intended to make a philosophical point, as an amusing engineering "hack", or as an intellectual joke. Devices which have no function or which malfunction are not considered to be "useless machines". - Wikipedia -

Imagine your maintenance team or the whole organization is challenged with excess in OpEx, unplanned downtime impacting the OEE (Over all Equipment Efficiency) or other unplanned work at the the sites. Suddenly, you find yourself on a vendor selection team to evaluate predictive analytics that will help your asset management team be more proactive in transition of asset maintenance planning for your assets and systems.

NOTE: On purpose I do not write "predictive maintenance" while that would only confuse many of the subject matter experts. and to clarify it once more predictive analytics is the practice of extracting information from existing data sets in order to determine patterns and predict future outcomes and trends. Predictive analytics does not tell you what will happen in the future; rather, they forecast what might happen in the future with an acceptable level of certainty. 

Anyway, back to the useless box. And before I enter into the business outcome erana allow me to take the opportunity to determine "Useless" in relationship with data and information management. First imagine that you have decided to install a tech IoT vibration sensor that provides you with an alarm trough mail or sms or even on a fancy dashboard. Sales and especially the IoT Hype has made you to expect to receive an alarm when an action has to be taken to, for example, predict a failure before it happend...how did that go? Most likely the "box" / application did generate alarms, alarms and even more alarms and perhaps so many that the alarm messaging was turned off and the key valuable feature becomes useless. Yeah, yeah you can change the trash hold, feedback loop and many other alternations to learn the useless box to become useful. Allow me to provide ML/AI 101 to help you our the next time to evaluate a vendors solution;

The false negative / positive alarms is the #1 error that drives the value from a asset condition monitoring solution.

Sensitivity and specificity are two statistical measures used to evaluate the performance of predictive analytics when the condition being predicted is either “true” or “false.” This is called a (binary) classification model, because there are only two possible outcomes. Sensitivity can be used to evaluate the ability of component health to correctly identify those with an issue like degradation by wear or failure (true positive rate). Specificity, on the other hand, is the ability to correctly identify those without an issue (true negative rate).

In a perfect world, both sensitivity (the accuracy of predicting yeses) and specificity (the accuracy of predicting nos) would both be 100% (expressed as 1.0). In the real world, however, there is no such thing as a perfect predictive model. Therefore, it is important to interpret both the sensitivity and specificity together when deciding on the power of a predictive analytic. We leave it for now at this stage before entering into the Confusion Matrix (also known as Contingency Table) assuming you got the concept.

Perhaps the most important formula to consider when deciding if predictive analytics are right for your organization is: IF you could know X, then you could DO Y. It isn’t enough to just implement predictive analytics. The stakeholders in your organization need to be clear and in agreement about the actions that will be taken when a predictive analytic is generated for a given team, asset or system. Planning for the "Y" (expected actions) and evaluating the feasibility of potential interventions prior to deploying the "X" (predictive analytics) is an important step in the selection process and will add to the overall success of your investment to ensure that you don't end-up with a Useless Box.