Toward automatic condition assessment of high-voltage transmission infrastructure using deep learning techniques

Abstract Electrical Transmission System Operators (TSO) are trusted with ensuring the safety and reliability of transmission infrastructure which can span thousands of kilometers. Maintenance of such a geographically expansive system is naturally a matter of concern and companies invest heavily in tracking infrastructure state which still relies predominantly on visual inspection. This paper presents an automated condition assessment methodology for concrete poles supporting overhead conductors based on deep learning object detection networks. Nine defect conditions ranging from incipient to severe are automatically detected from infrastructure photographs and mapped onto established Health Indices used by maintenance personnel. Three different deep learning networks are tested and new metrics, specific to this problem, are defined to evaluate their performance based on asset Health Index (HI) values. Results indicate that deep learning object detection networks hold promise for significantly reducing manual labour associated with visual inspection, especially when combining with automatic asset identification based on image geotag. This paper shows acceptable performance on more severe defect types.