Comparative Evaluation of Multimodal Fusion Architectures for Cyber Physical System Anomaly Classification

Abstract

Smart manufacturing systems generate heterogeneous data from programmable logic controllers (PLCs), industrial networks, and system logs, providing complementary information for cyber-physical system health monitoring. However, few studies have systematically compared multimodal fusion architectures for industrial anomaly classification. This paper evaluates four architectures, Fusion MLP, Fusion 1D CNN, Fusion CNN LSTM, and Cross Modal Attention Fusion, using synchronized PLC telemetry, network traffic, and system log data under an identical experimental protocol. Unimodal and classical machine learning baselines are also evaluated to assess the contribution of multimodal fusion. Results show that all fusion architectures achieve near-perfect classification performance. However, unimodal and temporal evaluations reveal that the benchmark is highly separable, with PLC telemetry alone achieving 99.93% accuracy and 99.68% macro F1 under temporal evaluation. Confusion matrices, t-SNE visualization, and attention allocation analysis further examine model behavior and modality contributions. The findings show that multimodal fusion effectively integrates heterogeneous industrial data, while increased architectural complexity does not necessarily improve classification performance when individual modalities contain highly discriminative signals. These results highlight the importance of unimodal baselines, temporal robustness evaluation, and interpretable multimodal analysis in assessing industrial  anomaly detection systems.