Evaluating the Influence of Normalisation Procedures on a Negative Selection Algorithm to Enhance Damage Detection

Damage detection is a critical component of Structural Health Monitoring (SHM) strategies. Data-driven machine learning methods are widely employed for this purpose. However, their effective application to the management of structures and infrastructure requires addressing key challenges, in particular the limited knowledge during training of how damage affects monitored features. Although one-class classification algorithms may be adopted, their performance strongly depends on the appropriate definition of their components and the tuning of their parameters. When this optimisation is performed using undamaged data only, it may result in classifiers that are insensitive to small-scale damage. Within this context, the present study investigates the sensitivity of damage detection performance of a Deterministically Generated Negative Selection Algorithm to feature scaling and two intrinsic algorithm parameters. A novel strategy for generating artificial damaged data to support parameter tuning is proposed and evaluated against alternative approaches. A wide range of parameter values is explored, considering multiple pairs of monitored features as detection spaces, and four feature scaling methods are compared. Feature scaling is a fundamental aspect of classification problems. Thus, the main findings may be generalised to other machine learning algorithms for damage detection. To ensure full control over the monitoring data and underlying phenomena, a numerical case study is adopted. A replicable framework for generating controlled yet realistic structural monitoring data is presented. The simulated monitoring captures the natural frequencies of a bridge, accounting for temperature effects, and incorporates three damage scenarios: one diffuse and two localised at critical locations. The analysis highlights the importance of conducting anomaly detection in feature spaces where damage affects each feature differently. Pairing a damage-sensitive feature with one unaffected by damage (e.g., temperature) may be advantageous when damage effects are not known a priori. Although suboptimal, the proposed strategy for generating damaged data for parameter tuning outperforms approaches based solely on undamaged data. The identified parameter trends suggest that small detector radii and relatively short censoring distances improve algorithm performance across all normalisation strategies and damage scenarios. The results show that feature scaling has limited influence for large damage extents but becomes critical for the early detection of minor damage. Z-score normalisation provides the best overall balance between false negatives and false positives, whereas methods using denominator multipliers smaller than one provide higher True Positive Rates for small damage extents.