Machine maintenance may sound like a dull subject but getting it right could propel maintenance strategies straight into the heart of state of the art technologies. I’m talking about predictive maintenance. If you consider that a single unplanned shutdown in the semiconductor industry can cost as much as $3 million, you can see how getting it right can save millions.
Preventative maintenance works on the basis of swapping out a component whether it needs it or not. You may be advised to change the cam belt on your car at 50,000 miles; it may be that the average failure happens at 100,000 miles but the distribution of failures span 70,000 to 130,000 miles. So 50,000 miles is deemed to have a safe margin. This works but errs of the side of safety – you may still have another 50,000 miles life in your cam belt, but you don’t know.
Many industries need to sweat their assets to get the best returns where margin are low, so they are turning to predictive maintenance, or machine condition monitoring (MCM). Sensors such as accelerometers are deployed to monitor machine operation, and by their nature they are likely to be distributed and equally likely to be wireless. And so we quickly move into solving problems of a different kind.
Managing data from multiple sensors, choosing the right wireless infrastructure, factoring in battery life for standalone sensors and crunching all that data – that is quite a balancing act. In the White Paper Edge Intelligence, Douglas O’Rourke looks at the limiting factors that govern this as a system and how this can be addressed by moving some of the intelligence from the ‘centre’ to the ‘edge’, i.e. in the sensors.
Modern low-power microcontrollers have the power to reduce raw vibration data by a factor of something like 100. Think of it as MP3 for IoT data. The data is reduced in size but does not lose the ability to represent differences between ‘normal’ and failure modes. In this way our smart sensors use less power and use less bandwidth. But our server at the centre of the cloud now has to be smarter as it is working with greatly simplified data.
We will still have to ‘teach’ the system what is normal and what is a failure mode but the acquired learning is being accumulated in the MCM which will improve over time. Ultimately the system will be alerting an engineer to a potential anomaly and based on operator feedback will become more accurate.
MCM is not about taking the human out of the loop but about providing the humans with better, actionable information. Gathering a large body of training data over time will lead to better predictions. Moving some intelligence to the edge will enable scaling without hitting bandwidth and battery life constraints.
If you’d like to learn more about Edge Intelligence you can download a copy of the White Paper – Edge Intelligence: the implications of the ‘Internet of Things’ on the architecture of preventative maintenance systems