Human-informed Artificial Intelligence for improved wind turbine health monitoring

Project Description:

Offshore wind farms are entering a new era with much greater adoption of sensing systems thereby presenting huge opportunities for the health-monitoring of wind turbines (WTs) and entire wind farms. The wealth of data, both in terms of high sampling frequency and density of sensor coverage, leads to the potential of health monitoring systems that are capable of predicting the onset of damage in a quicker and more robust fashion than previous campaigns. That said, this explosion in big data also presents challenges within the data-to-decision chain.

Arguably, the key step in this chain is that of Feature Selection whereby damage sensitive features, drawn from raw sensor data, are identified the level of success in a health-monitoring campaign is inextricably linked to the quality of the selected features. When the amount and type of raw data is limited, the task of selecting optimal or near-optimal features is one that may be conducted manually but, when the amount and types of data increase significantly, the task becomes one that requires automated computational support.

Artificial Intelligence (AI) algorithms may be developed to address such problems and barely a day goes by without a new AI success story appearing in the press. Equally well-documented are stories where AI algorithms demonstrate bias and a lack of intelligence leading to potentially very harmful consequences. AI will inevitably become a hugely powerful tool in the field of Structural Health Monitoring (SHM) but, in order to avoid misleading or harmful outcomes where AI identifies features from within the data that initially appear promising but later transpire to be detrimental, it will be essential to incorporate the views of human experts.

Consequently, this project proposes to develop a Human-plus-AI approach for SHM of offshore wind farms. A number of machine learning tools ranging from relatively simple techniques such as Outlier Analysis through to more advanced techniques such as Gaussian Processes (GPs) and Convolutional Neural Networks (CNNs) shall be investigated to develop architectures that allow for the incorporation of human input. The project will provide state-of-the-art AI and Machine Learning technologies that incorporate human expertise throughout their development thereby resulting in robust SHM for wind turbines that truly leverages the power of big data without the risk of harmful outcomes from rogue AI.

We will develop Bayesian technologies and tools which process the collection of data and provide advanced data features that are rich in diagnostic information. This feature identification phase will incorporate human expertise at various points to ensure that unwanted biases be removed and to give the expert the opportunity to approve, from a physical viewpoint, the AI identified features. Depending upon the particular dataset being investigated, the source of the human expert input may come from within the Dynamics Research Group (DRG) or knowledge from the industrial partners.

This project will consider a broad range of tools such as active, semi-supervised and transfer learning. To build such models, the project will make use of the latest developments in machine learning and statistics; in particular, techniques like Gaussian Processes (GPs). Furthermore, tools like outlier analysis will also be considered. Offshore wind farm SCADA data will be used from the team’s connections with Vattenfall and Siemens-Gamesa. The DRG and the supervisor team have parallel projects in offshore wind with industrial partners that will be used to create links for advice and data (https://npow.group.shef.ac.uk/).

References:

[1] Bull LA, Worden K, Fuentes R, Manson G, Cross EJ, Dervilis N. Outlier ensembles: A robust method for damage detection and unsupervised feature extraction from high-dimensional data. Journal of Sound and Vibration. 2019 Aug 4;453:126-50.

[2] Bull LA, Rogers TJ, Wickramarachchi C, Cross EJ, Worden K, Dervilis N. Probabilistic active learning: An online framework for structural health monitoring. Mechanical Systems and Signal Processing. 2019 Dec 1;134:106294.

[3] Fuentes R, Gardner P, Mineo C, Rogers TJ, Pierce SG, Worden K, Dervilis N, Cross EJ. Autonomous ultrasonic inspection using Bayesian optimisation and robust out-
generated October 5, 2022 2 lier analysis. Mechanical Systems and Signal Processing. 2020 Nov 1;145:106897.

[4] Papatheou E, Dervilis N, Maguire AE, Antoniadou I, Worden K. A performance monitoring approach for the novel Lillgrund offshore wind farm. IEEE Transactions on Industrial Electronics 2015; 62(10):66366644.

[5] Bull LA, Gardner PA,
Gosliga J, Rogers TJ, Dervilis N, Cross EJ, Papatheou E, Maguire AE, Campos C, Worden K. Foundations of population-based SHM, Part I: Homogeneous populations and forms. Mechanical Systems and Signal Processing. 2020;148:107141.