As you’ll have seen from previous blog posts, we’re enthusiastic about the applications of using Machine Learning with industrial data.
Machine Learning identifies features more reliably than other methods with fewer false positives/negatives; ideal when trying to find recurring behaviour which needs early action, e.g., asset failures, golden batches, repeated maintenance, non-conformances, etc.
As Sabisu Machine Learning is nearly ready for general release, here’s an update on what you can expect in the first release.
Training made easy
Machine learning is an ‘umbrella term’ that covers a myriad of Sabisu algorithms and techniques. They’re all statistical methods which “learn” from data; the more data you have, the better they will perform.
Sabisu makes this easy with Events, which allow users to mark a section of data, e.g., decoke, heat exchanger failure, pump failure, etc.
This provides a library of identified features in your data. It’s exactly what the algorithms need to identify thousands of events at the touch of a button.
An Event shown on a Sabisu chart.
After marking the Event, Sabisu will ask whether it should find similar occurrences of this behaviour.
Sabisu will identify every single similar event in the date range you choose. Machine learning ensures that the match is good – we’ve optimised our machine learning algorithms to outperform other statistical techniques.
After an Event has been tagged, you are given the option to find similar behaviour in your data.
Each of these matches is marked as a new Event so you can quickly move between them, looking for similarities or root causes.
All the usual collaboration capabilities are in play; highlight similarities in Notes, allocate an Action to investigate or analyse further, upload files and so on.
How does Sabisu ML work?
To analyse large quantities of data Sabisu utilises cloud technology, seamlessly passing data to our cloud data-lake; a highly versatile & scalable structured and unstructured data store.
Distributed processing is used to to build a model using unsupervised learning (see our previous blog post). This is used to identify past occurrences of the chosen event throughout the entirety of the supplied data. The results of this operation are stored in our distributed cloud data warehouse, allowing Sabisu to display them to you through pipelines, widgets, and events.
This schematic shows how your event and historical data are passed from your Sabisu unit to our data lake, through a Sabisu analytics cluster where machine learning is used to identify similar events in historical data, with the results stored in a parallelised distributed data warehouse for rapid access.
This is just the first step.
In beta testing we have real-time event handling; using machine learning to detect when an identified event is about to occur (‘incipient events’) or has just started to occur (‘triggering events’) so that you can take early or pre-emptive action.
An obvious application of this is in the case of asset failures where early warning can be invaluable in reducing downtime and optimising maintenance.
Also in beta is image recognition, so that any images captured during a process can be automatically tagged, e.g., EHSS audit photographs taken on an intrinsically safe mobile device.
Machine learning is also being used to de-noise and clean data intelligently before it’s committed to the data-lake.
Very soon, Sabisu will provide users greater control over how machine learning models are constructed and used with customised data classification to further tune model behaviour.
Machine learning is becoming an invaluable tool to help keep processes running in the best possible way, reducing wastage and maximising quality through the identification of recurring good or bad behaviour as soon as possible.
We’re always interested in hearing from you with any comments or suggestions, feel free to get in touch.