A lot has been written on the use of predictive analytics and its application for industrial equipment. The first stage in data analytics for industrial equipment revolves around efficiency metric construction and its improvement. The rules for metric constructions is often part of the contracting stage for many large industrial assets. In the power generation space equipment availability is often well defined and covers details including alarms that are part of OEM/O&M downtime vs alarms that are part of environmental downtime. Therefore a strong focus on predictive analytics is essential for a solution provider in this space to add value.
We started our journey in predictive analytics in 2015. After addressing the early challenges around data quality we developed scores to identify anomalies. Our approach to predictive analytics is at a simplistic level like a body mass index. We model parameter behaviour for say gearbox oil temperature or drive train vibration by
After tuning the model to ensure that the control limits are appropriate, we work with a beta customer to test the validity of the model. The output of the model cannot often be a clear Boolean operator (Successful vs Not successful). Even when model indicates deviation from the set control limits we often have to contend with a range of findings
We have had to slowly sift through various findings to improve the model accuracy over the last year. We have provided below two cases which highlight our experience with predictive notifications.
One of our customer having over 1 GW of assets across Asia wanted to reduce breakdowns. To help them address the issue, we deployed our predictive analytics model to identify component level failures. Our Predictive models work with existing SCADA data and take thermal, electrical and vibration based signals to identify nearness to failure of individual components. In one such case, as can be seen in the graph below, we identified an anomaly in generator ring (Depicted in blue), which was operating at a higher temperature range than the one predicted by our model (Depicted in red). Our platform sent out a predictive notification on July 16 and the site maintenance team took action and resolved the issue on July 19.
The O&M inspection report mentioned that the bush in the ring was damaged and needed replacement. Our model helped our client reduce downtime and a larger failure had it continued to operate.
Our customer recently commissioned a 200 MW wind farm in North America with turbines from a leading OEM. The OEM had also installed a Condition Monitoring System (CMS). Since we had worked with particular make of turbine before, we were able to implement our predictive notifications within a month of integration. In our very first notification, our model detected an anomaly in the generator bearing of a particular turbine but since the same was not detected by the CMS, the OEM did not take any action. However, a month later, the same issue was flagged by the CMS and the OEM replaced the bearing. During the interim period the turbine output had to be curtailed since operating under normal conditions could have increased the extent of damage. Our predictive model was able to provide notification prior to the CMS and with a good level of accuracy.
These are two among quite a few of our success stories, since we started predictive notifications. As we now move to next phase by providing prescriptive analytics to our customers, we envision quite a few challenges, but look forward to conquering the same.