Reconstruction Assisted Semi-Supervised Conformer for Bolt Looseness Diagnosis in Large-Scale Bolted Mechanical...

Bolt looseness diagnosis in power transmission towers is complicated by costly state labels and changing acquisition conditions. This study evaluates a reconstruction-assisted semi-supervised conformer using triaxial vibration-derived MFCCs. Labeled samples supervise binary classification, while labeled and unlabeled samples support denoising and masked reconstruction without state pseudo-labels. A retrospective audit retains 70,409 labeled and 154,394 distinct unlabeled segments and separates 42,421 training, 13,934 validation, and 14,054 test segments by identifiable source-record, inferred-acquisition, and exact-feature groups. Across five seeds, the full model obtains test macro-F1 0.9124 ± 0.0033 and ROC-AUC 0.9692 ± 0.0025, compared with 0.9032 ± 0.0135 and 0.9644 ± 0.0074 for labeled reconstruction. The paired macro-F1 difference is 0.00915 with a 95% seed interval of -0.00503 to 0.02333; reduced-label experiments show no consistent benefit. Performance is substantially weaker on held-out tower types and a later cohort, and waveform-level perturbations cause marked degradation. At clean inference, one encoder call exactly preserves the original duplicated-feature predictions. The results support useful grouped within-archive classification but do not establish physical-tower-independent generalization, annotation savings, or validated field-noise tolerance.