*4.3. Discussion*

The framework presented is built upon several questions determined by the authors to distinguish diagnostic models by two key metrics. The first decision tree (Figure 6) is designed to split models and features by which channels and sensors are similar across both turbine configurations. This decision is of course influenced by the selection of features used in each paper, which can lead to some variation in what channels were deemed relevant. The papers chosen for this study have been peer reviewed; therefore, it is assumed that the channels were selected either through some expert domain knowledge, or through a data-driven technique to select variables relevant to the component examined. For this reason it can also be assumed that these features would prove to be related to the component and how transferable these features are will help decide how transferable each particular failure mode is.

The second decision tree (Figure 7) is designed to compare faults across different components, as well as the manner in which a particular fault is likely to develop through time. Although the decision tree itself is designed to be as general as possible, the specific failure modes related to the diagnostic models presented here do not represent all possible failure modes, or even components, within the wind turbine. In Figure 7, faults to be diagnosed are first split at a component level, as this is the fundamental aspect of how transferable a fault is. From here, faults are split by the manner they are likely to progress and present themselves, which is more specific to the fault and failure mode itself. This requires some expert knowledge of wind turbines and mechanical or electrical systems depending on the individual fault.

Some papers presented in Table 3 were less specific about the fault than others. For example, Dhiman [6] presents a gearbox fault, whereas Zeng [54] presents a gearbox oil temperature over limit fault. The more specific a fault, the easier it can be scored using the decision trees; however, it is more sensitive. Conversely, as a more general fault is harder to score without further information about the failure mode from the authors, more general faults will typically score lower than a specific fault. Similarly, those papers which utilise a wider range of SCADA channels, compared to ones that use channels picked for their relevance to the fault, will typically score lower. By using a wider range of data, it is less likely that all of the features used will still be transferable to a direct-drive machine. For example, while a paper may examine a generator fault which is fairly transferable, the authors may incorporate many features from the gearbox. Obviously, most SCADA channels in the gearbox will not be transferable to those in a direct-drive turbine.

The transferability of each diagnostic model is shown in Figure 8. Each quadrant relates to a different level of transferability. The bottom left quadrant is the least transferable section, with both the required sensors and physics of failure relating to a particular component fault being unsuitable. The top left quadrant is where the sensors required to create the model features are available, and hence more transferable, but the fault is not relevant. An example of this could be a gearbox failure sensed through other means such as power curve analysis. The bottom right quandrant represents models which scored a high score on the component metric, but a low score on sensor metric. An example of this could be in detecting a high-speed bearing fault using features primarily created from sensors associated with the gearbox. Although the sensors are not available across both configurations, high-speed bearing faults could still be relevant to a lower-speed shafts on the direct-drive machine. The top right quandrant consists of the most transferable fault cases, with both the sensors and components being relevant and transferable.

The most transferable components in this upper right quadrant appear to be the components more ubiquitous across turbine configurations, such as pitch and yaw systems. The channels relevant to these faults are usually quite ubiquitous as well. The generator is also a fairly transferable component; however, due to its size compared to that of a direct-drive turbine, some processing may be required to the channels to engineer more appropriate features. As expected, the least transferable component is the gearbox; however, as can be seen, some failure modes and sensors are slightly more transferable to a direct drive. For example, a gearbox bearing may exhibit similar behaviour to other bearings within a direct-drive turbine. One interesting component group is general shafts and bearings of the turbine. These are typically transferable faults; however, the choice of sensors may need examined. Many of these failures focus on the high-speed shaft, which is not present in a direct-drive turbine, but could be transferred with some feature engineering.
