NAERINGSPH-Nærings-phd

With the aim of utilizing recent advances in Deep Learning (DL) and UAV technologies for facilitating automatic autonomous vision-based inspection of power lines, we propose a novel automatic autonomous vision-based power line inspection concept that uses UAV inspection as the main inspection method, optical images as the primary data source, and deep learning as the backbone of data analysis. To facilitate the implementation of the concept, we first identify six main challenges of DL vision-based UAV inspection: (i) the lack of training data; (ii) class imbalance; (iii) the detection of small power line components and defects; (iv) the detection of power lines in cluttered backgrounds; (v) the detection of previously unseen power line components and defects; and (vi) the lack of metrics for evaluating inspection performance. Next, we address the first three challenges by creating four medium-sized datasets for training component detection and classification models, by applying a series of effective data augmentation techniques to balance out the imbalanced classes, and by proposing a multi-stage component detection and classification approach based on SSD (Single Shot MultiBox Detector) and ResNets (Residual Networks) to detect small power line components and defects. Then, we address the fourth challenge by proposing LS-Net, a fast single-shot line-segment detector, for then to apply it to power line detection. Finally, we propose few-shot learning as a potential solution to the fifth challenge. Specifically, to pave the way for addressing the challenge, we propose a novel dissimilarity measure for distance metric learning-based few-shot learning, called SEN, to address the existing issues of the traditional Euclidean distance in high dimensional spaces.

Prosjektet har hatt stor virkning i utviklingen av ny teknologi som vil støtte en overgang fra manuelle inspeksjoner av kritisk infrastruktur som transmisjon og distribusjonsnett til en automatisert inspeksjon som vil redusere risiko for utfall, redusere inspeksjonskostnadene, og redusere HMS risiko knyttet til de manuelle inspeksjonene.

For most power grid operators, it is challenging to have an accurate and up to date overview of the condition of their grid, equipment, and power lines. In fault situations and for preventive and corrective maintenance of power grid infrastructure, power grid companies have traditionally relied on manual ground based inspections (crews walking the lines) and helicopter based inspections. Recently the use of RPAS (Remote Piloted Aircraft Systems), popularly called drones, has started to be considered by several grid operators as a low cost and low risk alternative to helicopter inspections, and as an additional support to ground crews, greatly reducing the need for them to climb up power pylons with the associated HSE risks. This industry PhD project aims at investigating the use of the new generation of information technologies based on big data, machine learning, and real-time processing, to support the analysis of data acquired through RPAS based inspections. One of the main sources of data from RPAS inspections will be still images and video. Other sources of information, such as power-line noise measurements, IR (infra-red) imaging, multispectral and hyperspectral imaging, and LIDAR (light detection and ranging), are being evaluated for their suitability, and might be included later on in the scope of this project. The industry PhD project will primarily focus on image analysis and object recognition. The results of the PhD will be used directly in an ongoing R&D project at eSmart Systems aiming at developing a system to support and as much as possible automate infrastructure inspections for power grid companies.