The aim of this project is to demonstrate how data and text mining techniques can help wind farm operators to extract unique, quantifiable, site- and asset-specific maintenance information from historic work orders. Understanding how maintenance efforts have been distributed in the past can help develop a more evidence-based maintenance strategy for the future in terms of labour intensity, budgeting and logistics of spare parts. However, work order records – where significant information is entered by a human in the form of free text – can present a particularly complex data source for analysis.
Our approach introduces a novel combination of machine learning techniques supported by a database of domain vocabulary and expert judgement. Significant focus is on term recognition, aided by spelling error correction and semantic matching of synonyms and abbreviations. Task descriptions can thereby be classified by meaning, not just the words present. In the first instance this creates a frequency distribution of all the different tasks carried out. Categorical data can then be extracted about maintenance of different functional locations and subsystems, as well as the occurrence of different failure modes.
Data from major onshore wind farms in Scotland was used to test our approach against undertaking a similar analysis manually. Potential savings were identified on the order of weeks of effort, or £ 9k in labour cost per wind farm, in addition to the benefits of an improved maintenance strategy.
The remaining challenges mainly lie in increasing accuracy and reducing operator input. These are being addressed by our continued research, but also provide opportunities for collaboration and standardisation across the wind energy industry to maximise the value of data.
- work orders
- text mining
- onshore wind