Deep Learning and Transformer Based Models For Power Transformer Nameplate Information Extraction: A Systematic Comparative Review

Abstract

Power-transformer nameplates record the ratings and identification data required for asset registers, loading studies, protection settings and electrical audits, yet field photographs of these plates are frequently degraded by specular glare, corrosion, low contrast, perspective distortion and small or engraved characters. This paper presents a systematic comparative review of deep-learning methods for extracting structured information from power-transformer and related electrical-equipment nameplates. Following a documented search of six bibliographic databases with explicit inclusion and exclusion criteria, 53 sources were retained and classified against a five-stage pipeline: image enhancement, text detection, text recognition, structured field extraction and engineering validation. Detection models (CTPN, EAST, CRAFT, DBNet), recognition models spanning CNN–RNN–CTC, attention-based, Transformer-based (e.g., PARSeq, TrOCR) and multimodal document-understanding architectures (LayoutLM family, Donut) are compared using source-reported results, each stated together with its dataset, test-set size, protocol and metric definition. Three principal findings emerge. First, domain-specific nameplate studies use small private datasets and non-standard metric definitions, and some headline figures were not obtained on nameplates at all (e.g., the 98.2% recognition rate of a CTPN–CRNN nameplate pipeline was measured on ICDAR 2017), so these figures cannot be ranked against one another or against public benchmarks. Second, public-benchmark results indicate architectural trends but do not transfer directly to nameplates. Third, the reviewed nameplate studies concentrate on detection and recognition; field association, unit normalization and engineering plausibility are rarely, if ever, evaluated. Because a single-character error such as “11 kV” read as “1.1 kV”, or “Dyn11” read as “Dyn1”, can alter an engineering decision, an engineering-aware evaluation framework is synthesized, comprising field-level exact match, an all-critical-fields-correct rate, normalized numeric error, cross-field consistency checks and confidence-based abstention. Research gaps and a literature-derived evaluation protocol for future studies are identified.