As was highlighted in the previous post, “Alarms & Events” often serve as lagging indicators of failure. They identify problems that have already occurred. In many cases other components in the sub-system may have been damaged due to the failure. In this post, we will identify approaches to developing leading indicators of failure. A leading indicator is one which can inform the operator of a wind turbine or solar plant that trouble is on the horizon. In this post we will discuss two simple approaches and then look at adding machine learning to strengthen them further.
a. Multi-parameters analysis
b. Relative value analysis
Multi-parameter analysis – This approach requires that we first cluster together the sensors that are part of the same sub system or are inputs to a given sub system. A sub-system in case of a wind turbine can be the blades, main bearing, gearbox etc. In case of a solar plant the sub-systems are modules, inverter, transformer etc. Analyzing the data produced by sub-system components in tandem is essential for identifying leading trends. There are also some generic performance parameters (power output) which associate with each cluster. We then derive correlations between various parameters within the cluster.
Lets develop a correlation matrix between the various sensor parameters in a gearbox – rotor speed, gearbox temperature, gearbox acceleration, RPM and so on, we will end up with a correlation matrix which looks like this:
We can now compare values at various times to see if there are outliers; construct limits and bands for outliers to ensure that large un-explained deviations are highlighted. Such deviations can then be analysed to see if there is cause of concern in any part of the sub system.
Relative value analysis – The approach is based on the comparison of values between sections of a solar plant or turbines in the farm. This comparison can help us identify deviations that need to be analysed.For example, at any given time of the day, the measured irradiance on a solar farm will result in a certain range of typical solar panel surface temperature. Thermal issues in any panel or section can then be identified by relative deviation of temperature as compared to other farm sections.
Addition of machine learning to the above approaches can make the approaches more robust. Some methods that can be applied are logistic regression to predict values of parameters, classification techniques to identify outliers and so on.
Predictive / Machine learning approaches – In this approach, there is a continuous monitoring of difference between actual & estimated value of any particular sensor value. This is inherently a machine learning approach. The estimated value is generated by a model which is trained based on normal operating data of the entire turbine/section. Limits of permissible deviations from estimate are applied and the operator is notified on breach to investigate further.
|Architecture of Predictive Approach|
Issues with lubricating oil, exhaust fan, generator exhaust ducting in a wind turbine can be quickly detected using this type of analysis. Thermal issues in various sections of a solar plant can also be monitored.
Algo Engines leverages correlations between system parameters to give health & condition related insight to farm operators and maintenance teams. Early detection of failures leads to focused maintenance directed to resolution of specific issues leading to lower downtime, extended component life and lower O&M costs.