Anomaly Detection Based on Data Super-Resolution in Industrial Cyber–Physical Systems With Multirate Sampling

In industrial cyber-physical control systems (ICPSs), different types of variables are typically sampled at diverse intervals, ranging from subseconds to minutes or even longer, resulting in multirate sampling data. The variation in sampling rates can lead to data gaps, which impede the performance of traditional anomaly detection techniques. To address this challenge, a new anomaly detection approach is presented for ICPSs with multirate data. Employing a mechanism model (MM), it performs data super-resolution in both temporal and spatial domains, thus filling in the missing data across various sampling rates and expanding additional virtual spatial states. Moreover, the approach integrates an anomaly-forced autoencoder (AFA). The autoencoder is trained to improve anomaly detection by injecting anomalies randomly during the training phase, thereby enhancing its capacity to accurately reconstruct normal behavior under anomalous conditions. Experimental studies are conducted on a petrochemical fractionation unit simulation testbed and also a pubic benchmark dataset Battle of the Attack Detection Algorithms (BATADALs) to demonstrate the proposed approach. Experimental results from the testbed show that the proposed approach consistently maintains stable detection accuracy (ACC), ranging between 95.7% and 96.5%. On BATADAL dataset, the detection ACC varies between 87.5% and 91.5% across different multirate sampling scenarios. This performance not only outperforms state-of-the-art baseline methods but also exhibits similar superior results when applied to other datasets.