What to look for in IOT based Predictive Maintenance Solutions

Predictive maintenance solutions promise to revolutionise plant productivity, uptimes and availability. However it is early days. It is still difficult to find solution providers who can deliver the full solution by themselves or as a consortium. There are many providers but few who have the full story stitched together seamlessly.

If you are looking for predictive maintenance solution, you will need to grapple with the following challenges:

1. What do you want to predict?

Accurate predictions of failures are not enough. They need to be far out enough in time for management folks to act on the prediction. Ultimately the plant benefits if the failure can be managed without incurring the same cost as the failure actually occurring. Which means that the equipment in question can be attended to as part of a planned activity or opportunistically before it actually fails. Typically companies want several weeks to factor the failure prediction into their work plan.

2. What parameters do you need to capture?

Selecting the right set of variables is key. You will have historical digital data on some of them, some are being monitored but not recorded digitally, and others are simply not being monitored. Since voluminous data is required to train models and make accurate predictions, balancing what is available with what is on your wish list is important. Otherwise you will end up spending time, money and effort on poor predictions and / or waiting for a long period of time before you have enough data to make good predictions.

3. How does it all come together?

1. Data acquisition (via IOT sensors placed on equipment) - needless to say sometimes you will need ruggedized sensors that can withstand high temperatures, vibrations etc. or other hazardous conditions in the plant

2. Transmission from the sensors to an IOT Gateway which connects up both with the control room systems as well as the cloud based solution – with sparse mobile and wi-fi connectivity in large plant sites, you may have to invest in a dedicated network which can be an expensive proposition

3. Incoming data fed into the prediction model and alerts generated which are fed back to the control room – if you have gotten so far, you are doing well!

4. Handling alerts as part of operations and maintenance planning – this will require modification of planning and execution SOPs and change management with plant personnel.

A few months ago, we met a big provider of predictive analytics solutions for manufacturing plants. A global brand, they had a bunch of impressive case studies from multi nationals who had deployed their solutions in various plants all over the world. It came down to this – the solution provider had no generalised datasets that they could deploy for new implementations. They were expecting to use the data from the implementation to train their models for that plant and make predictions. Apparently their client contracts prevented them from any sort of data reuse across scenarios.

Soon after that we met a startup operating in a niche segment with far fewer but nevertheless impressive case studies, who told us that every contract they signed had a clause allowing them to strip away all identifying information of the client, and then use that data to train their models and deploy the solutions in any other client situation as they saw fit. Every implementation would start up with an initial model that was trained on “generalised” data available from other clients, and then fine-tuned / sharpened with the data from that specific implementation.

What did we do? That is a story for another day.

Meanwhile, choose well!