A novel multi-scale Bayesian-KAN model for PV defect detection

The use of AI techniques to detect defects in photovoltaic electroluminescence (EL) images is becoming increasingly important. However, most existing AI-based models fail to fully capture the diversity of defect types present in EL images, resulting in suboptimal detection performance for small defects such as linear cracks, star cracks, and so on. In this work, a novel defect segmentation framework is proposed, which integrates Bayesian Neural Networks (BNN) and Kolmogorov–Arnold Networks (KAN) into an enhanced U-shaped architecture. The proposed model is an encoder-decoder architecture. First, a multi-scale Bayesian module (MSBM) is designed to allow the encoder to learn weight distributions instead of deterministic convolution kernels, thereby enhancing robustness and predictive reliability. Second, a KAN-based Feature Mining (KFM) module is introduced to learn complex nonlinear mappings in place of conventional linear combinations, providing more effective representations of diverse defect patterns. Furthermore, a specialized decoder is incorporated to efficiently fuse multi-scale features, balancing accuracy and inference speed for practical deployment. Extensive experiments on two public PV defect datasets show that the proposed method outperforms baselines in five metrics, achieving a 5.1% and 7.1% improvement in the IOU metric on two datasets. Overall, the proposed method combines deep uncertainty modeling with photovoltaic defect detection, which is crucial for reliable large-scale solar deployment. The dataset and code are available at https://github.com/limiaoair/PVdefect.