Saving innocent machines from certain failure: how IoT empowers predictive maintenance
Most machinery comes fitted with a range of sensors required to control it. IoT unlocks this large volume of sensor data, enabling further processing, either on the edge or once uploaded to a cloud. Our team’s objective is to analyse this data to search for early symptoms of failure. Through this lens, we cover the use of automatically-generated models versus expert models; Tried-and-tested statistical models versus “black-box” models; and how the knowledge in-house experts can be combined with machine learning to deliver results faster.
We also cover architectural aspects: from using the IoT devices merely to upload sensor readings to the cloud for centralised processing, versus performing some pre-processing on the edge device. Our application (mining) makes these architecture decisions particularly critical, in the presence of a generally unreliable network.
The storyline of this talk follows a 4-year timeline from the very start of this project to a somewhat mature stage, and each topic is discussed through “lessons learned” from mini case-studies of models we implemented, and where to next, in terms of edge processing, remote deployment of containerised analytics.
Outline/Structure of the Case Study
1. Introduction and context: the mining industry, general challenges, motivation behind high-reliability
2. The past: simple dataloggers harvesting sensor readings from the control system, and uploading them
3. Analytics: from simple rule-based analytics, to advances statistical processing, and into Google’s TensorFlow AI library.
How we went from a one-man band, to skunk-works and into running state-of-the-art analytics on an high-performance IT platform.
What to do when you’ve already got plenty of data and you need to prove the value of analytics quickly: automated analytics, dealing with the false positives, etc.
How to process the results of analytics: taking ideas from the social media to deal with the tidal wave of information.
4. The IoT future: moving beyond “upload-only” IoT devices, and into edge processing, centralised control of IoT devices using an IoT hub, and “closing the loop” on the edge: edge data acquisition, edge processing, and displaying on the edge, for a more robust solution and shorter “round-trip-time” from the event detection to the notification.
5. Conclusion: where we came from, where we are now, and where IoT is going to take us next.
An understanding of the particular challenges and needs of the mining industry with regards to IoT, in particular, advantages linked to edge processing, and direct feedback without going through the cloud.
An example of analytics strategy, in particular for large volume of industrial data and for predictive maintenance/anomaly detection
The place of AI versus more conventional analytics techniques, large-scale and automated deployment of analytics, adapting to fleets of heterogeneous machines
Managers in charge of elaborating strategies for an IoT analytics, Practitioners in the field of IoT data analytics (focussing on time-series data) People interested in machine learning/AI