Influence and Mitigation of Moisture Effects in Percussion-Based Condition Assessment of Concealed Wall Ties

Moisture can alter masonry percussion-acoustic responses and reduce the reliability of concealed wall tie condition assessment. This study investigates the influence of moisture on percussion-based non-destructive testing (NDT) of wall ties and evaluates strategies for mitigating moisture-induced domain shift. Percussion tests were conducted under a dry state and four moisture-affected states using five double-leaf masonry wallettes, each representing a different controlled wall tie condition. Moisture effects were first quantified using a baseline classifier, while the proposed mitigation method combined a 26-to-13 resampled Mel-frequency cepstral coefficient (MFCC) representation with a convolutional neural network incorporating a spatial-attention-only convolutional block attention module (CNN-CBAM-S) and supervised contrastive learning. Leave-one-region-out testing was used to evaluate classification performance on held-out regions within the same specimens. Under fixed wall tie conditions, dry-versus-moisture classification achieved an accuracy of 0.773 and a macro-F1 of 0.735, confirming a recognisable moisture-related acoustic shift. A model trained only on dry data achieved a macro-F1 of 0.715 on dry samples, whereas its macro-F1 decreased to 0.496–0.568 on moisture-affected states. Moisture-diverse training increased the proposed method’s mean macro-F1 from 0.717 to 0.857, and the sample-size-controlled experiment indicated that this improvement was mainly attributable to moisture-state diversity rather than increased data volume alone. Compared with the baseline, the proposed method mainly improved performance in more challenging regions, with a slight reduction in an already well-performing region, resulting in better overall performance. Supplementing dry data with labelled samples from one moisture-affected state also increased the mean macro-F1 on unseen states from 0.657 to 0.770. Ablation studies further confirmed the effectiveness of the resampled MFCC representation and the proposed model. Although the classification results should be interpreted within the present controlled specimen set, the findings consistently indicate that moisture should be explicitly considered in percussion-based wall tie assessment and that moisture-diverse training or single-state calibration can substantially improve classification robustness.