TabNet Interpretable Deep Structure for Adaptive Recognition of Audit Voucher Anomalies

This study proposes an interpretable deep learning framework for adaptive recognition of audit voucher anomalies based on the TabNet architecture. The framework addresses the dual challenges of decision transparency and environmental adaptability in intelligent auditing systems. Structured audit voucher data are processed through sequential attention layers to generate traceable feature importance masks and contribution paths. A sliding timewindow mechanism is introduced to accumulate historical feature importance distributions, while a dynamic threshold calibration strategy based on statistical characteristics enables adaptive anomaly boundary adjustment under concept drift conditions. The proposed framework integrates feature selection, temporal distribution analysis, and anomaly scoring to achieve transparent and adaptive anomaly recognition. Experimental results demonstrate that the detection accuracy reaches 92.3% for amount logic conflicts and 93.1% for supplier-related anomalies, while all anomaly categories achieve F1-scores above 87%. In addition, the verification time for supplier anomalies is reduced from 19.6 min to 8.4 min, resulting in a 57.1% improvement in audit verification efficiency. The average importance of transaction amount features reaches 0.883 in financial audit scenarios, confirming the model’s ability to adaptively focus on critical business characteristics. The proposed framework can be integrated with wireless data acquisition infrastructures, intelligent financial information systems, and edge-computing-enabled audit platforms to support realtime anomaly monitoring and interpretable decision-making in large-scale financial auditing environments.