Beyond pdM — A Continuously Learning Factory

Progressive organizations have started adopting predictive maintenance as one of the enablers for their overall Asset Performance Management strategy at various maturity level.

Predictive maintenance can be as simple as statistics based anomaly detection to as complex as leveraging Neural Networks to determine when an equipment can fail or what is the remaining useful life of the asset is.

Organisations has achieved success and ROI at various levels when adopting the predictive maintenance as their APM strategy. PdM works very well when predicting the failure of assets or the system as a whole for those equipment which are independent of the SKUs they process. For example — In a metals industry the load on the processing equipment is a combination of the chemistry of the metal, the thickness and the width of metals blocks being processed. In many cases, it also depends on the pre-heating temperature of the metal block before the actual metal processing is carried out on the machine. In such scenarios having a pdM strategy that solely depends on the machine operating condition might not yield the expected ROI or results.

The maintenance organization of any factory leverages pdM algorithms to answer fundamental questions related to Asset up keep

1) Is the machine likely to fail

2) Why the machine will fail

3) Till when I can run the machine without a breakdown

However, many a times pdM algorithms strives to answer the above questions without a context of “manufacturing production plans” or the “order book” situation of the factory / enterprise as a whole.

Let me explain this scenario with my past experience working as a production planner in one of the largest integrated steel making company in India. We had a situation where the steel mills were not able to produce certain SKUs of the material because it was causing a lot of stress on the bearing of the steel mill motor. So as a production planner, it was my job to ensure the mill was not getting loaded while coordinating with sales team and other downstream plants to ensure alternate SKUs are provisioned to this mill with problems. This was to ensure the mill is still running under constraints while the mill OEM was able to provide a fix.

Back in time (a decade back) there was no pdM and no algorithms were in place to detect the failures and provide an ETF ( Earliest Time to Fail ) and the plant was dependent on that “one expert” who uses his / her experience and could advise the plant when the mill is likely to break down and what material processing should be avoided in order not to push the mill to break down / stop completely.

Now in this scenario, as you could imagine pdM comes handy where the algorithms can continuously monitor various condition parameters of the mill and in real time they could predict “if this continues then the mill is likely to fail in next 45 days”.

“if this continues”

The pdM algorithms have no idea what is planned to be processed on the machine in next 1 day, 1 week or 1 month or 3 months while it predicted the failure of the mill in next 90 days, This is left to the maintenance planners, production planners, Supply Chain and sales team now to be collaboratively decide what future action to be taken given a “warning” is in place by the pdM algorithms.

As you would imagine, this requires an inter departmental consensus for arriving at a supply plan trading off between the the asset health warning from the pdM algorithms, the Delivery commitments to the customers, the contract obligations and finally the time to “fix” and make the asset ready as usual.

Optimization and Sequencing

With this problem of finding the right supply plan in place, there is a need to create a supply plan that is optimized looking into various inter departmental constraints and commitments.

This is where organizations must look beyond pdM. There must be some means and methods to understand given the “warning” by the pdM algorithms based on current conditions,

• How my asset is going to behave if we change the SKUs be processed . What impact it will have of the ETF if I continue to run some SKUs while is less stressful for the machine in question
• How my assets is going to behave if I plan a combination of SKUs which are stressful and easy to be processed on the “to be broken” assets. Is there a sequence I can follow to maximize the asset life before it cease to operate
• How my asset is going to behave if I have the same SKUs with alternate processing conditions ( for example- increasing the pre-heat temperature of the metal block by 15 degrees more ). What is the additional benefits I am going to get

This is clearly beyond the scope of current generation of pdM algorithms which are not “plan aware” at the moment. To make this shift happen, high performing organizations are getting into data platform which enable “continuous learning”.

Towards a continuously Learning Factory

The continuously learning factory leverages a unified data model that encompasses all dimensions of manufacturing (production, quality, maintenance, inventory, methods & SOPs, energy, human skill) and the analytics behind the screen enables exploring the relationship between various elements involved within the manufacturing setup. With data available at one single data model, the impact of one on the other entity can be created significantly faster and with much more ease.

In next blogs, we will see some use cases of the continuously learning factory leveraging a unified manufacturing data model.

[Views expressed in this blog is author’s own views and it does not necessarily reflects the view of his employer, A&M]