Abstract: The inspection of photovoltaic modules for faults using drones is typically conducted by processing and detecting in both visible light and infrared light scenarios separately. This paper proposes a fault detection method based on the residual neural network ResNet50 and improved YOLOv5, achieving high-precision automatic classification and fault detection of two types of image. For infrared data, chromaticity transformation is used to remove sun reflection and retain hot spots, while for visible light data, sharpening is used to highlight small targets such as foreign objects and cracks. Different YOLOv5 object detection algorithms are used to achieve fast detection and positioning of small foreign object faults under visible light and hot spot faults under infrared light.