Conference Agenda

Verifying engineering requirements is a critical step in railway Signalling projects because delayed or inconsistent compliance assessment can slow certification, postpone system handover, and affect operational readiness. Engineers use a Requirements Verification Matrix to compare stated requirements with submitted evidence and confirm whether each requirement has been satisfied. This process remains labor-intensive, error-prone, and difficult to scale across large projects. The challenge increases when expert-labeled compliance records are scarce, especially for non-compliant cases that require specialist judgment.

This study proposes a semi-supervised adversarial transformer framework for railway compliance verification when only a small number of records have been checked and labeled by compliance experts. The framework uses 636 confirmed compliant records from a railway signalling project and applies six controlled perturbation strategies to construct 636 synthetic non-compliant records. These perturbations include negation insertion, evidence mismatch, scope substitution, meaning inversion, numerical alteration, and standard reference swapping. The resulting 1,272-record dataset serves as a proof-of-concept evaluation setting for testing whether artificial intelligence can distinguish compliant evidence from non-compliant evidence under this controlled setting. Each record is represented in a natural language inference format that combines verification context, the stated requirement, and the submitted evidence.

The proposed model follows a GAN-BERT-inspired training strategy. A transformer encoder learns representations from real compliance records, while a generator produces artificial feature vectors. The discriminator learns from three types of training signals, including expert-labeled real records for compliance classification, real records without expert labels for distribution learning, and generated feature vectors for adversarial regularization. This design enables learning from both labeled and unlabeled records without requiring full expert annotation.

Five transformer architectures are evaluated within this pipeline, namely BERT-base, RoBERTa, DeBERTa, SciBERT, and DistilBERT. The experiment tests nine training settings, where only 8 to 64 records are provided with expert labels. The results show that BERT-base and SciBERT provide stable performance across different training settings, while SciBERT performs strongly when more expert-labeled records are available. DistilBERT reaches a practical safety-oriented threshold with only 14 expert-labeled records, achieving non-compliant recall of 0.901 without training collapse. Within the controlled perturbation-based setting, DistilBERT achieved very high internal performance at 48 expert-labeled records, with macro F1 of 0.998 and non-compliant recall of 1.000. RoBERTa and DeBERTa show higher sensitivity to instability under extreme expert-label scarcity.

The findings suggest that semi-supervised adversarial transformer learning can reduce the number of records that compliance experts need to label while supporting digital assurance workflows for railway Signalling projects. The study contributes an AI-based verification pipeline for scarce-label environments and highlights the importance of matching model capacity to the number of available expert-labeled records. Because the non-compliant cases in this experiment are synthetically constructed, future validation with real project non-compliance records is needed before operational deployment.