A Lightweight BERT-Based Framework for Automated Severity Classification of Power Equipment Defects

The rapid digitalization of power systems has generated vast unstructured defect logs, but inconsistent terminology, missing labels, and narrative-label mismatches hinder automated analysis. We introduce a lightweight domain-adapted BERT framework featuring structural simplifications (8 layers, 512 -dim hidden size, pruned heads), power industry pre-training on GB-T 33601/IEC corpora with a custom tokenizer, and multi-task learning (severity classification, NER, cause effect prediction, textual entailment) via a two-stage transfer learning protocol with dynamic cosine annealing learning rates and regularization. On the experimental dataset, our model yields 0.98 F1 for binary defect detection, 0.94 F1 for four class severity tagging, and 0.94 overall F1 on 12 category classification with less than 100 M parameters. This compact BERT variant embeds domain standards and multi-task objectives within an edge-friendly design, enabling scalable, real-time severity classification that reduces manual audit costs and enhances data driven maintenance planning.