*4.1. Framework*

In the previous section, reliability rates (wind turbine stop rates and downtimes) for both direct and geared wind turbines were presented, where several key differences were highlighted. Building on these results, a framework for assessing the transferability of failure modes and associated sensors will now be presented. The aim for this framework is to help determine how well a particular fault could be diagnosed in modern direct-drive wind turbines using data and diagnostic models demonstrated on geared wind turbines. This framework does not make any attempt to predict future reliability rates of larger direct-drive machines. Stop rates and downtimes presented in Section 3 will be used in conjunction with the framework to assess which components need to be the focus of future diagnostic research. An example of how this framework can be used is presented later in this section, which examines previous fault detection papers by assessing their failure mode and input data to examine how transferable the fault case from each paper was. By doing this, we can see which components in particular are suitable to assessment in the future with direct-drive machines, and whether the data inputs used previously would be suitable or would require some level of processing. The transferability of each paper was assessed over two dimensions to examine overall transferability. The first dimension examined how transferable the sensors, or data channels, used by each paper to predict, or diagnose, the fault. The second dimension assessed how well the specific failure mode transferred between geared and direct-drive turbines based on fundamental understanding of the physics of failure.

Two small-scale decision trees were drawn up to assess each paper examined for their transferability, one for each dimension. The first tree, Figure 6, is used to assess the transferability of the sensors, or channels, used in each paper. So this assesses the paper's selection of features for modelling a particular failure mode—and how well these features, and required sensors, transfer from a geared to a direct-drive machine. The first question, in the top diamond, asks if the direct-drive would have all the sensors required for the features used by the paper. The second question, in the middle two diamonds, asks if the majority of the sensors used in the paper would be in the same location within the turbine. The third question, the bottom four diamonds, asks if these sensors are of the same specification as those on an arbitrary direct-drive machine. This question is essentially used to assess if the general expected data range you would ge<sup>t</sup> from this sensor would be the same as one on a geared machine. An example could be bearing temperature, you could consider if a similar thermistor could be used to cover the expected range and resolution of recorded temperatures. This helps to assess scale of component as well, for example a generator bearing in a direct-drive machine will be much larger than one in a similarly rated geared machine. For all of these questions, it was assumed that the direct-drive and geared machines were of the same arbitrary turbine model; however, the only difference being that

the the direct-drive machine had no gearbox and the generator was of an appropriate size for the same power rating as the geared machine.

**Figure 6.** Decision tree for sensor/channel group transferability.

The second decision tree, Figure 7, assesses how transferable the physics of failure mode of each paper is. This tree examines the particular component that failed, and any failure mode information provided by each paper. First the papers are split by if the fault, or damage, can be found anywhere across either turbine configuration. Next it is split by whether it can be found on a specific component—if it can then it asks if the fault progresses in the same physical manner, and if not then it asks if the model corresponds to a specific failure mode. If it cannot be found on a specific component, then the questions determine if it is on an assembly within the turbine, and then if it follows the same physical manner, and lastly if it is specific to the failure mode examined. Again these questions were answered under the assumption that the turbines would be of similar models, with the exception of the gearbox and generator.

To test these decision trees, a database of past fault detection papers were collected and their fault and input data were collated. Table 3 shows all the past papers examined, and their scores based on the decision trees referenced earlier. The majority of these faults occurred on the drivetrain, with some in the blades. SCADA data were the focus of this study; however, some papers were assessed that used images as their input data. The input data that each paper utilised were assessed with the first decision tree, and the fault itself was assessed using the second decision tree. The data ranged both in terms of years assessed, but also in the turbine rating. All the turbines within the dataset of papers examined were geared turbines.


#### **Table 3.** Database of papers examined for their transferability.

**Figure 7.** Decision tree for component failure mode transferability.

This technique could potentially be expanded to papers that utilised high-frequency conditon monitoring systems (CMS) data, which is commonly used in the literature. SCADA data, which is 10 min aggregate data, is what is examined here. High-frequency CMS data may be of interest in transfer learning for two reasons, the first being the lack of direct-drive data for the larger machines, but also because CMS data are much more difficult to acquire. SCADA systems are commonly installed on wind turbines; however, CMS sensors are less so (particularly on older machines). Typically, CMS data are specifically used to target an area of interest; however, it is becoming cheaper to install relative to power output and therefore more common.

#### *4.2. Framework Application Results*

Results are presented based on the framework explained in the above section for the literature stated in Table 3. Figure 8 shows the component tranferability scores on the x-axis and the sensor tranferability scores on the y-axis. Each single point represents a diagnostic model presented in the literature (see Table 3) with the fault grouped into the corresponding component or wider assembly. The mean value per component for each axis was calculated and plotted as a larger cross on Figure 8. Note that this does not correspond to any particular model, but simply represents the average transferability of that component or assembly. All 35 models from Table 3 are shown, although some points have identical coordinates and cannot be easily distinguished. The most transferable diagnostic models relate to the hydraulic, pitch and yaw systems, which makes sense due to them being mostly universal to different wind turbine drive train configurations. For this same reason, sensor faults and electrical issues also scored highly. The least transferable models were, as expected given the drive train topology, associated with the gearbox. The components with the most variability in scoring were the the generator, shafts and bearings. For these, the overall transferability was largely dependent on both the sensors required to create model features and specific failure mode being diagnosed.

**Figure 8.** Results showing transferability of diagnostic model.
