In order to remain competitive, wind turbines must be reliable machines with efficient and effective maintenance strategies. However, thus far, wind turbine reliability information has been closely guarded by the original equipment manufacturers (OEMs), and turbine reliability studies often rely on data that are not always in a usable or consistent format. In addition, issues with turbine maintenance logs and alarm system data can make it hard to identify historical periods of faulty operation. This means that building new and effective data-driven condition monitoring techniques and methods can be challenging, especially those that rely on supervisory control and data acquisition (SCADA) system data. Such data are rarely standardised, resulting in challenges for researchers in contextualising these data. This work aims to summarise some of the issues seen in previous studies, highlighting the common problems seen by researchers working in the areas of condition monitoring and reliability analysis. Standards and policy initiatives that aim to alleviate some of these problems are given, and a summary of their recommendations is presented. The main finding from this work is that industry would benefit hugely from unified standards for turbine taxonomies, alarm codes, SCADA operational data and maintenance and fault reporting.

Using 10-minute wind turbine supervisory control and data acquisition (SCADA) system data to predict faults can be an attractive way of working toward a predictive maintenance strategy without needing to invest in extra hardware. Classification methods have been shown to be effective in this regard, but there have been some common issues in their application within the literature. To use these data-driven methods effectively, historical SCADA data must be accurately labelled with the periods when turbines were down due to faults, as well as with the reason for the fault. This can be manually achieved using maintenance logs, but can be highly tedious and time-consuming due to the often unstructured format in which this information is stored. Alarm systems can also help, but the sheer volume of alarms and false positives generated complicate efforts. Furthermore, a way to implement and evaluate the field deployed system beyond simple classification metrics is needed. In this work, we present a prescribed and reproducible framework for: (i) automatically identifying periods of faulty operation using rules applied to the turbine alarm system; (ii) using this information to perform classification which avoids some of the common pitfalls observed in literature; and (iii) generating alerts based on a sliding window metric to evaluate the performance of the system in a real-world scenario. The framework was applied to a dataset from an operating wind farm and the results show that the system can automatically and accurately label historical stoppages from the alarms data. For fault prediction, classification scores are quite low, with precision of 0.16 and recall of 0.49, but it is envisaged that this can be greatly improved with more training data. Nonetheless, the sliding window metric compensates for the low raw classification scores and shows that 71% of faults can be predicted with an average of 30 h notice, with false alarms being active for 122 h of the year. By adjusting some of the parameters of the fault prediction alerts, the duration of false alarms can be drastically reduced to 2 h, but this also reduces the number of predicted faults to 8%.

Using 10-minute wind turbine SCADA data for fault prediction offers an attractive way of gaining additional prognostic capabilities without needing to invest in extra hardware. To use these data-driven methods effectively, the historical SCADA data must be labelled with the periods when the turbine was in faulty operation as well the sub-system the fault was attributed to. Manually identifying faults using maintenance logs can be effective, but is also highly time consuming and tedious due to the disparate nature of these logs across manufacturers, operators and even individual maintenance events. Turbine alarm systems can help to identify these periods, but the sheer volume of alarms and false positives generated makes analysing them on an individual basis ineffective. In this work, we present a new method for automatically identifying historical stoppages on the turbine using SCADA and alarms data. Each stoppage is associated with either a fault in one of the turbine's sub-systems, a routine maintenance activity, a grid-related event or a number of other categories. This is then checked against maintenance logs for accuracy and the labelled data fed into a classifier for predicting when these stoppages will occur. Results show that the automated labelling process correctly identifies each type of stoppage, and can be effectively used for SCADA-based prediction of turbine faults.

Turbine alarm systems can give useful information to remote technicians on the cause of a fault or stoppage. However, alarms are generally generated at much too high a rate to gain meaningful insight from on their own, so generally require extensive domain knowledge to interpret. By grouping together commonly occurring alarm sequences, the burden of analysis can be reduced. Instead of analysing many individual alarms that occur during a stoppage, the stoppage can be linked to a commonly occurring sequence of alarms. Hence, maintenance technicians can be given information about the shared characteristics or root causes of stoppages where that particular alarm sequence appeared in the past. This research presents a methodology to identify relevant alarms from specific turbine assemblies and group together similar alarm sequences as they appear during stoppages. Batches of sequences associated with 456 different stoppages are created, and features are extracted from these batches representing the order the alarms appeared in. The batches are grouped together using clustering techniques, and evaluated using silhouette analysis and manual inspection. Results show that almost half of all stoppages can be attributed to one of 15 different clusters of alarm sequences.