Case Study: Setting Multi-Attribute Alarms
A significant challenge in operating a wind turbine is to know when there is trouble on the horizon. A primary method is to take measurements of various parameters. Turbines are constantly logging attributes such as wind speed and temperatures. Alongside, there are periodic measurements taken, such as gearbox oil samples. These attributes as a whole can effectively define the operational performance, health, and likelihood of issues.
In the following we will focus on the limitations of individual attribute thresholds, as proposed by some original equipment manufacturers, and discuss how we develop multi-attribute thresholds. The discussion is centered on gearbox oil analysis, as it provides a large number of attributes to consider, and the individual attribute approach is particularly flawed. However, the fundamentals of our recommendations should be exercised on all turbine level data: temperatures, speeds, direction, etc.
When utilizing gearbox oil analysis to assess risk, the current paradigm has been to set thresholds on an individual attribute basis. As examples, these thresholds state for Oil Brand C: the acceptable range for Viscosity at 40oC should be greater than 256 and below 384 mm2/s and/or Iron should be less than 100 ppm. One is then expected to regularly measure viscosity and iron, and take actions, such as change oil, if either of the values is out of the acceptable range. In our analysis of >25,000 oil samples, the occurrence of viscosity values outside of the “acceptable range” is <0.1%. Moreover, the average value of Viscosity at 40oC, in the oil sample pulled just prior to a gearbox failure, was 318 (based on a sample size of >200). The average value of Iron, in the oil sample, pulled just prior to a gearbox failure, was 11 (based on a sample size of >200). These examples show that by simply maintaining values within an acceptable range, one can neither predict the health of the gearbox and oil nor mitigate the likelihood to fail.
To further corroborate the above, the following table notes the acceptable ranges prescribed by the manufacturer of Oil Brand C alongside the percentile rank of the value (the percent of oil samples that are below that value). As you can see, nearly 99% of the time, the values received from oil analysis are below the critical limits.
A limiting factor with data, and specifically oil data, is that it can be unreliable. The methodology to pull and analyze oil is highly variable. Therefore making a decision based on a single attribute is statistically more risky than making a decision based on many attributes. An everyday example is within crime fighting: trying to catch a criminal only based on height leaves the detective with a poor crime fighting tool; adding weight, eye color, hair style, etc, significantly increases the odds that the culprit will be properly identified. This simple logic should be applied when analyzing oil analysis. Rather than view attributes as independent values, view them as a profile, or fingerprint.
Importantly, one must set the thresholds by examining previous failures. Rather than utilize theoretical ranges, we can devise ranges based on what has led to issues previously. For instance, we can know within what range of Silicon we generally see failure. To complete the crime fighting analogy above, the thresholds for identifying a culprit is based on known criminals’ height, weight, etc, not on theoretic values. Accordingly, we establish thresholds based on known failures.
The use of multi-attribute thresholds set by failure history is an extremely robust method to avoid unplanned outages. We now go into more detail on the specifics of creating multi-attribute fingerprints. We will focus on 3 aspects of multi-attribute fingerprints, in order of complexity:
- Genetic attributes
- Rate of change attributes
The first set of attributes to add into a data fingerprint is the genetic attributes of the turbine’s unique equipment: make, model, and year. This means that thresholds should be set based on what oil is being used within which model gearbox. This is specifically important when utilizing wear and ingression attributes, such as Cu, Fe, Si, water, etc. The values of these attributes are dependent on the design of the gearbox. It is surprising that most wind operators track acceptable ranges of oil analysis based only on the oil brand. Simply tuning this value to the gearbox model significantly increases the value of threshold ranges. Tuning to more attributes, such as Turbine model, climate zone, etc, improves accuracy, and is possible via big data modeling.
In creating multi-attribute thresholds one has to first define which attributes to utilize. This is a complex process that is the subject of a future paper. Suffice to say, it is function of what attributes show a large deviation between good and bad: e.g. attributes which show very different values within an oil sample prior to a failed gearbox versus a non-failed gearbox. For oil analysis these attributes are:
Categorical: OilType, OilBrand, OilStandardName, WtgModel, Component, ComponentModel, ComponentManufacturer
Continuous: Iron, Visc40, OilAge, TAN, Water, Copper, Silicon, PQIndex, Zinc, Boron, ComponentAge
Fluitec Wind’s advanced data analytics models take these values to create a fingerprint, then measure the distance (using log-likelihood similarity) of this fingerprint to all other samples. If a given fingerprint is significantly similar to a known poor instance, there is high likelihood that it will follow its failure path. There are many complex moving parts to this process. To help simplify and visualize we have created the following charts:
You can see below for the profile of gearbox brand M for a random set of data samples.
The next chart shows data profiles just prior to a catastrophic gearbox failure for the same genetic permutation. Note the difference to the control chart above. The dotted line is a sample turbine with a Brand M Gearbox and Brand C Oil.
To help explain the impact of genetic permutations, the final chart shows profiles of Brand C Oil data samples in a mixture of gearboxes, just prior to a gearbox failure.
The first takeaway is that, again, in the majority of instances just prior to failure, oil attributes were within the “acceptable range” provided by OEMs. The second takeaway is the pronounced difference in the profile of a Brand M gearbox prior to failure, versus in general. We see lower Viscosity, higher Silicon, moderate Iron, and higher PQ Index. The final takeaway is the difference between the failure bands when isolating Oil or Gearbox: what indicates a poor brand M gearbox is not the same as a generic gearbox using a specific oil brand.
Such findings are made by the machine learning model, and are the basis for creating thresholds for a given turbine. Therefore in the future, the model is able to flag a turbine based on whether the multi-attribute fingerprints are nearing known critical zones.
The above should help show both the value in thresholds based on specific genetic permutations and the use of multiple attributes to best define thresholds. If we simply look for high Iron or “abnormal” viscosity, to trigger a failure, we would miss virtually every failure.
Rate of Change Attributes
A further improvement to the use of multi-attribute thresholds is to incorporate “latent attributes”: attributes that are a function of attributes. Rate of change attributes fall into this category. They are based on the difference between a current attribute value, and the previous value of that attribute. For example, it is much more interesting that a gearbox bearing temperature increased from 30oC to 60oC in 10 minutes, than is the fact that the bearing temperature is currently 60oC. Moreover, if we take into account the change in wind speed, power produced, and ambient temperature, the information extracted improves: i.e. if the increase in bearing temperature happened while the power produced and nacelle temperature dropped, it is more relevant. By having a massive historical database we are able to define such rate of change thresholds in the same manner discussed in the previous section. Creating such multi-dimensional thresholds is too complex for human creation; however it is fundamental to valuable big data analysis.
We strongly encourage all operators to develop multi-attribute thresholds tuned to genetic permutation and historical failures. At Fluitec Wind our predictive analytics are based on such thresholds, and as a customer of the service you would receive both the predictive recommendations and unique threshold values where relevant. Please contact your Fluitec Wind representative for more details.