Abstract: A photovoltaic thermal spot fault classification detection method is proposed to resolve the issue of small target feature loss in the thermal spot detection method for aerial photovoltaic infrared images. Firstly, the multi-head self-attention mechanism is integrated with the CSPNet structure for improvement, resulting in the proposed CSPMAT network. Subsequently, this network is introduced into the New CSP-Darknet architecture, leading to the construction of the CSPMAT-Darknet model, achieving both localization and classification of photovoltaic component thermal spots. Experimental results demonstrate that the model enhances performance in small target detection tasks significantly. Moreover, in fault classification detection tasks with substantial target size variations, the achieved mean average precision（mAP） reaches 82.92%, an increase of 13.97 percentage points, thereby showcasing commendable detection accuracy and generalization capability.