The Attention Enhanced Autoencoder for Balise Anomaly Detection in Potting Process

Balise is a critical trackside component in railway systems, operating in harsh environments where silicone gel potting enhances reliability. However, the potting process is irreversible; defects lead to costly scrap and operational delays. For this reason, this study introduces an anomaly detection method for the transponder potting process for the first time. We propose an attention-enhanced autoencoder (AEA) for anomaly detection. Our approach integrates three core innovations: max-min normalisation to handle scaling discrepancies between input attributes, an attention mechanism that dynamically highlights key features, and empirically derived adaptive reconstruction error thresholds to discriminate anomalies. Experiments on real production data show that the AEA model achieves Precision of 97.10%, Recall of 100.00%, weighted F1-Score of 98.53%, Accuracy of 99.90%, ROC-AUC of 99.96%, and G-means of 98.54%, outperforming all baselines. Compared to the best baseline AE model, the AEA model improves Precision by 1.45%, Recall by 1.49%, weighted F1-Score by 1.47%, Accuracy by 0.09%, ROC-AUC by 0.03%, and G-means by 1.47%. These results demonstrated that our model significantly outperforms existing well-regarded models in anomaly detection performance.