A Fault Detection Framework Using Recurrent Neural Networks for Condition Monitoring of Wind Turbines
Time: Thu 2021-01-28 13.00
Location: zoom link for online defense (English)
Subject area: Electrical Engineering
Doctoral student: Yue Cui , Elektroteknik
Opponent: Professor Bikash C Pal, Imperial College London
Supervisor: Professor Lina Bertling, Elektrotekniska system, Elkraftteknik, Elektroteknisk teori och konstruktion; Dr Pramod Bangalore, Greenbyte AB
The global energy system is experiencing a transition to a sustainable system with ambitious targets for increased use of renewable energy. One key trend for this transition has been the large introduction of wind power and integration into the electricity grid. In order to succeed in this transition, there is a need to develop efficient tools to support the handling of the assets. Asset management is a coordinated activity for the organization to get value from an asset. As the main part of asset management, maintenance includes all the technical and corresponding administrative actions to keep or restore the asset to the desired state in which it can perform its required functions. Traditional maintenance is usually based on scheduled monitoring and physical inspections. However, with new access to data and information about condition-based maintenance shows to be an efficient solution for asset management. This thesis explores data-driven solutions for electrical equipment to generate alerts towards potential operation risks, which targets digital, efficient, and cost-effective asset management. Specifically, the thesis investigates wind turbines.
This thesis proposes a fault detection framework for cost-effective preventive maintenance of wind turbines by using condition monitoring systems. The thesis utilizes the data from supervisory control and data acquisition systems as the main input. For log events, each event is mapped to corresponding components based on the Reliawind taxonomy. For operation data, recurrent neural networks are applied to model normal behaviors, which can learn the long-time temporal dependencies between various time series. Based on the estimation results, a two-stage threshold method is proposed as post-processing to determine operation conditions. The method evaluates the shift values deviating from the estimated behaviors and their duration time to attenuate minor fluctuations. A two-level condition monitoring system is constructed to apply the proposed fault detection framework, which targets to detect possible faults of components and conduct performance analysis of turbines. The fault detection framework is tested with the experience data from onshore wind farms. The results demonstrate that the framework can detect operational risks and reduce false alarms.