Although diagnosing anomalies in construction equipment is essential to ensure operational safety, real working environments impose significant constraints in terms of inspection times, accessibility, and data acquisition. Thus, supervised learning-based diagnostic methods are impractical because collecting sufficient data is difficult owing to operational limitations. To overcome these challenges, we propose a noncontact unsupervised anomaly detection approach that learns the normal operating boundary using only normal data. Thermal data are employed to extract statistical features from automatically defined regions of interest, and a One-Class Support Vector Machine is used to learn the boundary of the normal condition. To evaluate the proposed framework under limited anomaly data conditions, synthetic anomalies are generated and incorporated into the evaluation process. The results demonstrate the feasibility of the proposed approach. Thus, our findings show that thermal statistical features combined with normal boundary learning can be utilized to detect abnormal conditions.
Kim et al. (Fri,) studied this question.