Imene Mitiche

PhD Student
Project title: Machine Learning for High Voltage Condition Monitoring

High voltage power supply sites are compound of electrical and mechanical assets such as generators, transformers, motors etc. Such systems are susceptible to insulation defects that arise in insulation or conduction medium. When faults develop, Electro-Magnetic Interference (EMI) radiation or conduction may be produced from discharge sources such as arcing, corona and Partial Discharge (PD). EMI signals may be exploited to gain information on power plant assets’ condition. This may help in early fault detection and to take further actions on the operating assets, whether for repair or shut down in extreme case. It also benefits the plant owner from safety enhancement, low maintenance or replacement cost, and reduced system's down time. Overall, it will enable the power plant companies to maximize return of investment, revenue and business profits.

Traditionally, experts in EMI investigate the data through audio spectrum analysis in order to identify any occurring defect. However, this method can be time consuming, particularly for large dataset, as it is manual. Moreover, it is not practical for continuous monitoring.

The aim of this project is to develop an intelligent and automatic condition monitoring system inspired by EMI experts’ knowledge for pattern recognition and classification of the different faults. Our approach is to employ the-state-of-the-art signal processing algorithms based on feature extraction and machine learning classification.

The nature of such signals is complex and non-stationary which makes their analysis challenging. The study involves an investigation of suitable features which are useful to reveal unique patterns related to the specific faults. Here, we employ real-world signals where noise is almost unavoidable. Noise is a common issue in signal processing applications as it masks the meaningful information contained in the signal, which hinders the accuracy of signals’ analysis and interpretation of the useful information. Thus, we also aim to develop an appropriate Denoising model for pre-processing.