Title: Enhancing Robustness in Multi-Product Cyber-Physical Systems via Product-Aware Deep Autoencoders

Abstract:

As Industry 4.0 drives the deeper integration of Cyber-Physical Systems (CPS) into manufacturing landscapes, ensuring process safety and security through robust anomaly detection has become paramount. Standard data-driven methodologies typically rely on "product-agnostic" or global models, which are trained on aggregated normal operating data. However, contemporary industrial facilities often manage a variety of product grades. Although computationally efficient, these global models tend to broaden their decision boundaries to encompass the variance associated with multiple operational modes. This expansion creates a "blind spot," potentially allowing subtle anomalies or targeted cyber-physical attacks to go undetected within the model’s wide acceptance region.

This study first confirms that global-agnostic models are indeed vulnerable to such issues when operating across multiple product grades. To address this, we introduce a Product-Aware Autoencoder, a principled solution that constrains the learning domain to grade-specific distributions. While this method effectively mitigates the identified blind-spot risk, we acknowledge that it may not represent the single optimal solution among all potential alternatives. We rigorously evaluate this approach against a Global Agnostic baseline using the Extended Tennessee Eastman Process (TEP) benchmark.

Empirical findings show that the Product-Aware framework matches the global baseline in standard detection metrics while providing enhanced robustness to operating modes specific to product grades. Most significantly, stress tests modeling hypothetical attack scenarios demonstrate a stark contrast in performance: the global model failed to identify operational deviations in 77.8% of cases, whereas the product-aware system achieved 100% detection accuracy. These results indicate that in flexible manufacturing settings, generalized anomaly detectors can introduce significant security vulnerabilities, highlighting the need to transition toward mode-aware diagnostic architectures.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC