Unsupervised Online Learning for Substation Equipment Defect Identification Based on SoftHebb

With the advancement of artificial intelligence, deep learning algorithms have increasingly been adopted for defect detection in substation equipment. However, a widely recognised limitation of such models is their inability to learn autonomously over time, resulting in performance degradation when faced with changing data distributions—a phenomenon commonly referred to as ‘concept drift’ in continual learning contexts. This issue is particularly critical in industrial applications such as power systems, where operational environments evolve continuously. Recently, Hebbian learning has gained renewed interest within the machine learning community due to its unsupervised and localised nature. Several studies have explored its integration with deep neural networks trained via backpropagation (BP), yet combining Hebbian and BP learning remains challenging, especially within complex vision tasks beyond simple classification. In this paper, we propose an object detection framework that integrates Hebbian learning with BP in a layered architecture, supporting unsupervised online learning through a two‐stage training strategy. This approach mitigates dependency on manual annotation and enhances adaptability in non‐stationary environments. Our method demonstrates improved robustness compared to standard BP‐based networks on both a perturbed defect dataset of substation equipment and the COCO2017 benchmark perturbed dataset.