Reliable operation of large-scale photovoltaic (PV) installations requires inspection methods capable of detecting degradation and failures at an early stage. Conventional ground-based inspections are often impractical for extensive PV farms, leading to the use of unmanned aerial vehicles (UAVs) equipped with infrared thermography for condition monitoring. This paper presents a two-stage inspection methodology supporting reliability-oriented diagnostics of PV modules under real conditions. In the first stage, a convolutional neural network (CNN) with dedicated post-processing extracts individual PV module geometries from radiometric images, including installations with non-uniform orientation and high module density. In the second stage, lightweight machine learning models identify reliability-critical defects at module and string levels. The approach was validated on novel, annotated and publicly accessible thermographic dataset collected during five years of UAV inspections, demonstrating high detection accuracy with low computational complexity suitable for large-scale deployment.