There is quite a buzz around Unsupervised Machine Learning for Asset Maintenance. For good reason: Unsupervised Machine Learning provides the benefits of the Digital Twin (real time alerts to emerging asset degradation and failure based on analysis of sensor data). At the same time, Unsupervised machine learning does not come with a hefty price tag because the advanced AI algorithm can analyze vast amounts of data in the cloud and is sensor agnostic.
Big Data Machine Learning (both Supervised and Unsupervised) are at the forefront of the move to the Smart Factory. Industrial plants now recognize that only a fraction of their existing SCADA data is ever used for Condition Based Monitoring. Furthermore, because controls are manually set, only breaches of control limits are monitored. The result is that emerging failure are typically ignored until it is too late and the CBM control limits have been exceeded.
Unsupervised Big Data Machine Learning for Asset Maintenance.
There are five discrete steps required for using Unsupervised Machine Learning for Asset Maintenance:
Sensor Data Generated and Stored: Raw Sensor data is generated and transmitted to the Historian database.
Data Streamed to Presenso Cloud: All raw sensor data is continuously streamed from the local Historian database to the Presenso Cloud. Whereas Conditioned Based Monitoring only monitors a few critical sensors, there are no limits to the number of sensors that can monitored by Presenso.
AI Driven Anomaly Detection: Whereas CBM tracks breaches of manually set control limits, the Presenso Advanced AI Algorithm detect anomalies in data patterns. The algorithms are sensor agnostic – they are looking for abnormal data patterns to identify emerging degradations and machine failure.
Contextualization of Machine Behavior: Detecting abnormal sensor data behavior is insufficient. Presenso’s algorithm searches for correlations and patterns between sensors, even if they are seemingly unrelated.
Once emerging patterns are detected, the facility owner receives an alert of potential failure. Included in an alert is information that provides insights into the failure origin.
Which Historian Databases can be used for Unsupervised Machine Learning?
The following is a partial list of the Historian databases that can be accessed by the Presenso Unsupervised Machine Learning cloud: OSIsoft PI Historian, Aspen Tech InfoPlus Historian, Emerson Ovation/DeltaV, Schneider Electric – Wonderware/Citect Historian, GE Digital (formerly Proficy), ABB Decathlon, SAP HANA, Siemens SIMATIC, Honeywell Historian, MS-SQL and ORACLE.
Comparing Supervised versus Unsupervised Machine Learning
There has been much discussion about Supervised Machine Learning, especially given the significant investments in the GE Predix platform and Siemens MindSphere. Supervised Machine Learning is used in by the Digital Twin. It is called “Supervised” because the algorithm needs access to the data labels of historic downtime records. In the Digital Twin, a complete and accurate 3D modelling using the asset’s blueprints is needed. Once the Digital Twin is created, the learning algorithm can “learn” the workings of the machine to be able to duplicate its behavior.
Unsupervised Machine learning is sensor and asset agonistic. With Unsupervised learning, there is no need to understand the physical working of the machine, because the algorithm seeks to detect abnormal data patterns. Because there are few limitations to the amount of sensor data analyzed, an entire factory can be monitored.
Ultimately, the differences between Supervised and Unsupervised models impact price, speed to market, and the number of assets that can be monitored.